xref: /freebsd/sys/geom/raid3/g_raid3.c (revision d056fa046c6a91b90cd98165face0e42a33a5173)
1 /*-
2  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/limits.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/bio.h>
38 #include <sys/sysctl.h>
39 #include <sys/malloc.h>
40 #include <sys/eventhandler.h>
41 #include <vm/uma.h>
42 #include <geom/geom.h>
43 #include <sys/proc.h>
44 #include <sys/kthread.h>
45 #include <sys/sched.h>
46 #include <geom/raid3/g_raid3.h>
47 
48 
49 static MALLOC_DEFINE(M_RAID3, "raid3_data", "GEOM_RAID3 Data");
50 
51 SYSCTL_DECL(_kern_geom);
52 SYSCTL_NODE(_kern_geom, OID_AUTO, raid3, CTLFLAG_RW, 0, "GEOM_RAID3 stuff");
53 u_int g_raid3_debug = 0;
54 TUNABLE_INT("kern.geom.raid3.debug", &g_raid3_debug);
55 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, debug, CTLFLAG_RW, &g_raid3_debug, 0,
56     "Debug level");
57 static u_int g_raid3_timeout = 4;
58 TUNABLE_INT("kern.geom.raid3.timeout", &g_raid3_timeout);
59 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, timeout, CTLFLAG_RW, &g_raid3_timeout,
60     0, "Time to wait on all raid3 components");
61 static u_int g_raid3_idletime = 5;
62 TUNABLE_INT("kern.geom.raid3.idletime", &g_raid3_idletime);
63 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, idletime, CTLFLAG_RW,
64     &g_raid3_idletime, 0, "Mark components as clean when idling");
65 static u_int g_raid3_disconnect_on_failure = 1;
66 TUNABLE_INT("kern.geom.raid3.disconnect_on_failure",
67     &g_raid3_disconnect_on_failure);
68 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, disconnect_on_failure, CTLFLAG_RW,
69     &g_raid3_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
70 static u_int g_raid3_syncreqs = 2;
71 TUNABLE_INT("kern.geom.raid3.sync_requests", &g_raid3_syncreqs);
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 TUNABLE_INT("kern.geom.raid3.use_malloc", &g_raid3_use_malloc);
76 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, use_malloc, CTLFLAG_RDTUN,
77     &g_raid3_use_malloc, 0, "Use malloc(9) instead of uma(9).");
78 
79 static u_int g_raid3_n64k = 50;
80 TUNABLE_INT("kern.geom.raid3.n64k", &g_raid3_n64k);
81 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n64k, CTLFLAG_RD, &g_raid3_n64k, 0,
82     "Maximum number of 64kB allocations");
83 static u_int g_raid3_n16k = 200;
84 TUNABLE_INT("kern.geom.raid3.n16k", &g_raid3_n16k);
85 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n16k, CTLFLAG_RD, &g_raid3_n16k, 0,
86     "Maximum number of 16kB allocations");
87 static u_int g_raid3_n4k = 1200;
88 TUNABLE_INT("kern.geom.raid3.n4k", &g_raid3_n4k);
89 SYSCTL_UINT(_kern_geom_raid3, OID_AUTO, n4k, CTLFLAG_RD, &g_raid3_n4k, 0,
90     "Maximum number of 4kB allocations");
91 
92 SYSCTL_NODE(_kern_geom_raid3, OID_AUTO, stat, CTLFLAG_RW, 0,
93     "GEOM_RAID3 statistics");
94 static u_int g_raid3_parity_mismatch = 0;
95 SYSCTL_UINT(_kern_geom_raid3_stat, OID_AUTO, parity_mismatch, CTLFLAG_RD,
96     &g_raid3_parity_mismatch, 0, "Number of failures in VERIFY mode");
97 
98 #define	MSLEEP(ident, mtx, priority, wmesg, timeout)	do {		\
99 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));	\
100 	msleep((ident), (mtx), (priority), (wmesg), (timeout));		\
101 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, (ident));	\
102 } while (0)
103 
104 static eventhandler_tag g_raid3_pre_sync = NULL;
105 
106 static int g_raid3_destroy_geom(struct gctl_req *req, struct g_class *mp,
107     struct g_geom *gp);
108 static g_taste_t g_raid3_taste;
109 static void g_raid3_init(struct g_class *mp);
110 static void g_raid3_fini(struct g_class *mp);
111 
112 struct g_class g_raid3_class = {
113 	.name = G_RAID3_CLASS_NAME,
114 	.version = G_VERSION,
115 	.ctlreq = g_raid3_config,
116 	.taste = g_raid3_taste,
117 	.destroy_geom = g_raid3_destroy_geom,
118 	.init = g_raid3_init,
119 	.fini = g_raid3_fini
120 };
121 
122 
123 static void g_raid3_destroy_provider(struct g_raid3_softc *sc);
124 static int g_raid3_update_disk(struct g_raid3_disk *disk, u_int state);
125 static void g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force);
126 static void g_raid3_dumpconf(struct sbuf *sb, const char *indent,
127     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
128 static void g_raid3_sync_stop(struct g_raid3_softc *sc, int type);
129 static int g_raid3_register_request(struct bio *pbp);
130 static void g_raid3_sync_release(struct g_raid3_softc *sc);
131 
132 
133 static const char *
134 g_raid3_disk_state2str(int state)
135 {
136 
137 	switch (state) {
138 	case G_RAID3_DISK_STATE_NODISK:
139 		return ("NODISK");
140 	case G_RAID3_DISK_STATE_NONE:
141 		return ("NONE");
142 	case G_RAID3_DISK_STATE_NEW:
143 		return ("NEW");
144 	case G_RAID3_DISK_STATE_ACTIVE:
145 		return ("ACTIVE");
146 	case G_RAID3_DISK_STATE_STALE:
147 		return ("STALE");
148 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
149 		return ("SYNCHRONIZING");
150 	case G_RAID3_DISK_STATE_DISCONNECTED:
151 		return ("DISCONNECTED");
152 	default:
153 		return ("INVALID");
154 	}
155 }
156 
157 static const char *
158 g_raid3_device_state2str(int state)
159 {
160 
161 	switch (state) {
162 	case G_RAID3_DEVICE_STATE_STARTING:
163 		return ("STARTING");
164 	case G_RAID3_DEVICE_STATE_DEGRADED:
165 		return ("DEGRADED");
166 	case G_RAID3_DEVICE_STATE_COMPLETE:
167 		return ("COMPLETE");
168 	default:
169 		return ("INVALID");
170 	}
171 }
172 
173 const char *
174 g_raid3_get_diskname(struct g_raid3_disk *disk)
175 {
176 
177 	if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
178 		return ("[unknown]");
179 	return (disk->d_name);
180 }
181 
182 static void *
183 g_raid3_alloc(struct g_raid3_softc *sc, size_t size, int flags)
184 {
185 	void *ptr;
186 
187 	if (g_raid3_use_malloc)
188 		ptr = malloc(size, M_RAID3, flags);
189 	else {
190 		ptr = uma_zalloc_arg(sc->sc_zones[g_raid3_zone(size)].sz_zone,
191 		   &sc->sc_zones[g_raid3_zone(size)], flags);
192 		sc->sc_zones[g_raid3_zone(size)].sz_requested++;
193 		if (ptr == NULL)
194 			sc->sc_zones[g_raid3_zone(size)].sz_failed++;
195 	}
196 	return (ptr);
197 }
198 
199 static void
200 g_raid3_free(struct g_raid3_softc *sc, void *ptr, size_t size)
201 {
202 
203 	if (g_raid3_use_malloc)
204 		free(ptr, M_RAID3);
205 	else {
206 		uma_zfree_arg(sc->sc_zones[g_raid3_zone(size)].sz_zone,
207 		    ptr, &sc->sc_zones[g_raid3_zone(size)]);
208 	}
209 }
210 
211 static int
212 g_raid3_uma_ctor(void *mem, int size, void *arg, int flags)
213 {
214 	struct g_raid3_zone *sz = arg;
215 
216 	if (sz->sz_max > 0 && sz->sz_inuse == sz->sz_max)
217 		return (ENOMEM);
218 	sz->sz_inuse++;
219 	return (0);
220 }
221 
222 static void
223 g_raid3_uma_dtor(void *mem, int size, void *arg)
224 {
225 	struct g_raid3_zone *sz = arg;
226 
227 	sz->sz_inuse--;
228 }
229 
230 #define	g_raid3_xor(src1, src2, dst, size)				\
231 	_g_raid3_xor((uint64_t *)(src1), (uint64_t *)(src2),		\
232 	    (uint64_t *)(dst), (size_t)size)
233 static void
234 _g_raid3_xor(uint64_t *src1, uint64_t *src2, uint64_t *dst, size_t size)
235 {
236 
237 	KASSERT((size % 128) == 0, ("Invalid size: %zu.", size));
238 	for (; size > 0; size -= 128) {
239 		*dst++ = (*src1++) ^ (*src2++);
240 		*dst++ = (*src1++) ^ (*src2++);
241 		*dst++ = (*src1++) ^ (*src2++);
242 		*dst++ = (*src1++) ^ (*src2++);
243 		*dst++ = (*src1++) ^ (*src2++);
244 		*dst++ = (*src1++) ^ (*src2++);
245 		*dst++ = (*src1++) ^ (*src2++);
246 		*dst++ = (*src1++) ^ (*src2++);
247 		*dst++ = (*src1++) ^ (*src2++);
248 		*dst++ = (*src1++) ^ (*src2++);
249 		*dst++ = (*src1++) ^ (*src2++);
250 		*dst++ = (*src1++) ^ (*src2++);
251 		*dst++ = (*src1++) ^ (*src2++);
252 		*dst++ = (*src1++) ^ (*src2++);
253 		*dst++ = (*src1++) ^ (*src2++);
254 		*dst++ = (*src1++) ^ (*src2++);
255 	}
256 }
257 
258 static int
259 g_raid3_is_zero(struct bio *bp)
260 {
261 	static const uint64_t zeros[] = {
262 	    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
263 	};
264 	u_char *addr;
265 	ssize_t size;
266 
267 	size = bp->bio_length;
268 	addr = (u_char *)bp->bio_data;
269 	for (; size > 0; size -= sizeof(zeros), addr += sizeof(zeros)) {
270 		if (bcmp(addr, zeros, sizeof(zeros)) != 0)
271 			return (0);
272 	}
273 	return (1);
274 }
275 
276 /*
277  * --- Events handling functions ---
278  * Events in geom_raid3 are used to maintain disks and device status
279  * from one thread to simplify locking.
280  */
281 static void
282 g_raid3_event_free(struct g_raid3_event *ep)
283 {
284 
285 	free(ep, M_RAID3);
286 }
287 
288 int
289 g_raid3_event_send(void *arg, int state, int flags)
290 {
291 	struct g_raid3_softc *sc;
292 	struct g_raid3_disk *disk;
293 	struct g_raid3_event *ep;
294 	int error;
295 
296 	ep = malloc(sizeof(*ep), M_RAID3, M_WAITOK);
297 	G_RAID3_DEBUG(4, "%s: Sending event %p.", __func__, ep);
298 	if ((flags & G_RAID3_EVENT_DEVICE) != 0) {
299 		disk = NULL;
300 		sc = arg;
301 	} else {
302 		disk = arg;
303 		sc = disk->d_softc;
304 	}
305 	ep->e_disk = disk;
306 	ep->e_state = state;
307 	ep->e_flags = flags;
308 	ep->e_error = 0;
309 	mtx_lock(&sc->sc_events_mtx);
310 	TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
311 	mtx_unlock(&sc->sc_events_mtx);
312 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
313 	mtx_lock(&sc->sc_queue_mtx);
314 	wakeup(sc);
315 	wakeup(&sc->sc_queue);
316 	mtx_unlock(&sc->sc_queue_mtx);
317 	if ((flags & G_RAID3_EVENT_DONTWAIT) != 0)
318 		return (0);
319 	sx_assert(&sc->sc_lock, SX_XLOCKED);
320 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
321 	sx_xunlock(&sc->sc_lock);
322 	while ((ep->e_flags & G_RAID3_EVENT_DONE) == 0) {
323 		mtx_lock(&sc->sc_events_mtx);
324 		MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "r3:event",
325 		    hz * 5);
326 	}
327 	error = ep->e_error;
328 	g_raid3_event_free(ep);
329 	sx_xlock(&sc->sc_lock);
330 	return (error);
331 }
332 
333 static struct g_raid3_event *
334 g_raid3_event_get(struct g_raid3_softc *sc)
335 {
336 	struct g_raid3_event *ep;
337 
338 	mtx_lock(&sc->sc_events_mtx);
339 	ep = TAILQ_FIRST(&sc->sc_events);
340 	mtx_unlock(&sc->sc_events_mtx);
341 	return (ep);
342 }
343 
344 static void
345 g_raid3_event_remove(struct g_raid3_softc *sc, struct g_raid3_event *ep)
346 {
347 
348 	mtx_lock(&sc->sc_events_mtx);
349 	TAILQ_REMOVE(&sc->sc_events, ep, e_next);
350 	mtx_unlock(&sc->sc_events_mtx);
351 }
352 
353 static void
354 g_raid3_event_cancel(struct g_raid3_disk *disk)
355 {
356 	struct g_raid3_softc *sc;
357 	struct g_raid3_event *ep, *tmpep;
358 
359 	sc = disk->d_softc;
360 	sx_assert(&sc->sc_lock, SX_XLOCKED);
361 
362 	mtx_lock(&sc->sc_events_mtx);
363 	TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
364 		if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0)
365 			continue;
366 		if (ep->e_disk != disk)
367 			continue;
368 		TAILQ_REMOVE(&sc->sc_events, ep, e_next);
369 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
370 			g_raid3_event_free(ep);
371 		else {
372 			ep->e_error = ECANCELED;
373 			wakeup(ep);
374 		}
375 	}
376 	mtx_unlock(&sc->sc_events_mtx);
377 }
378 
379 /*
380  * Return the number of disks in the given state.
381  * If state is equal to -1, count all connected disks.
382  */
383 u_int
384 g_raid3_ndisks(struct g_raid3_softc *sc, int state)
385 {
386 	struct g_raid3_disk *disk;
387 	u_int n, ndisks;
388 
389 	sx_assert(&sc->sc_lock, SX_LOCKED);
390 
391 	for (n = ndisks = 0; n < sc->sc_ndisks; n++) {
392 		disk = &sc->sc_disks[n];
393 		if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
394 			continue;
395 		if (state == -1 || disk->d_state == state)
396 			ndisks++;
397 	}
398 	return (ndisks);
399 }
400 
401 static u_int
402 g_raid3_nrequests(struct g_raid3_softc *sc, struct g_consumer *cp)
403 {
404 	struct bio *bp;
405 	u_int nreqs = 0;
406 
407 	mtx_lock(&sc->sc_queue_mtx);
408 	TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
409 		if (bp->bio_from == cp)
410 			nreqs++;
411 	}
412 	mtx_unlock(&sc->sc_queue_mtx);
413 	return (nreqs);
414 }
415 
416 static int
417 g_raid3_is_busy(struct g_raid3_softc *sc, struct g_consumer *cp)
418 {
419 
420 	if (cp->index > 0) {
421 		G_RAID3_DEBUG(2,
422 		    "I/O requests for %s exist, can't destroy it now.",
423 		    cp->provider->name);
424 		return (1);
425 	}
426 	if (g_raid3_nrequests(sc, cp) > 0) {
427 		G_RAID3_DEBUG(2,
428 		    "I/O requests for %s in queue, can't destroy it now.",
429 		    cp->provider->name);
430 		return (1);
431 	}
432 	return (0);
433 }
434 
435 static void
436 g_raid3_destroy_consumer(void *arg, int flags __unused)
437 {
438 	struct g_consumer *cp;
439 
440 	g_topology_assert();
441 
442 	cp = arg;
443 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
444 	g_detach(cp);
445 	g_destroy_consumer(cp);
446 }
447 
448 static void
449 g_raid3_kill_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
450 {
451 	struct g_provider *pp;
452 	int retaste_wait;
453 
454 	g_topology_assert();
455 
456 	cp->private = NULL;
457 	if (g_raid3_is_busy(sc, cp))
458 		return;
459 	G_RAID3_DEBUG(2, "Consumer %s destroyed.", cp->provider->name);
460 	pp = cp->provider;
461 	retaste_wait = 0;
462 	if (cp->acw == 1) {
463 		if ((pp->geom->flags & G_GEOM_WITHER) == 0)
464 			retaste_wait = 1;
465 	}
466 	G_RAID3_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
467 	    -cp->acw, -cp->ace, 0);
468 	if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
469 		g_access(cp, -cp->acr, -cp->acw, -cp->ace);
470 	if (retaste_wait) {
471 		/*
472 		 * After retaste event was send (inside g_access()), we can send
473 		 * event to detach and destroy consumer.
474 		 * A class, which has consumer to the given provider connected
475 		 * will not receive retaste event for the provider.
476 		 * This is the way how I ignore retaste events when I close
477 		 * consumers opened for write: I detach and destroy consumer
478 		 * after retaste event is sent.
479 		 */
480 		g_post_event(g_raid3_destroy_consumer, cp, M_WAITOK, NULL);
481 		return;
482 	}
483 	G_RAID3_DEBUG(1, "Consumer %s destroyed.", pp->name);
484 	g_detach(cp);
485 	g_destroy_consumer(cp);
486 }
487 
488 static int
489 g_raid3_connect_disk(struct g_raid3_disk *disk, struct g_provider *pp)
490 {
491 	struct g_consumer *cp;
492 	int error;
493 
494 	g_topology_assert_not();
495 	KASSERT(disk->d_consumer == NULL,
496 	    ("Disk already connected (device %s).", disk->d_softc->sc_name));
497 
498 	g_topology_lock();
499 	cp = g_new_consumer(disk->d_softc->sc_geom);
500 	error = g_attach(cp, pp);
501 	if (error != 0) {
502 		g_destroy_consumer(cp);
503 		g_topology_unlock();
504 		return (error);
505 	}
506 	error = g_access(cp, 1, 1, 1);
507 		g_topology_unlock();
508 	if (error != 0) {
509 		g_detach(cp);
510 		g_destroy_consumer(cp);
511 		G_RAID3_DEBUG(0, "Cannot open consumer %s (error=%d).",
512 		    pp->name, error);
513 		return (error);
514 	}
515 	disk->d_consumer = cp;
516 	disk->d_consumer->private = disk;
517 	disk->d_consumer->index = 0;
518 	G_RAID3_DEBUG(2, "Disk %s connected.", g_raid3_get_diskname(disk));
519 	return (0);
520 }
521 
522 static void
523 g_raid3_disconnect_consumer(struct g_raid3_softc *sc, struct g_consumer *cp)
524 {
525 
526 	g_topology_assert();
527 
528 	if (cp == NULL)
529 		return;
530 	if (cp->provider != NULL)
531 		g_raid3_kill_consumer(sc, cp);
532 	else
533 		g_destroy_consumer(cp);
534 }
535 
536 /*
537  * Initialize disk. This means allocate memory, create consumer, attach it
538  * to the provider and open access (r1w1e1) to it.
539  */
540 static struct g_raid3_disk *
541 g_raid3_init_disk(struct g_raid3_softc *sc, struct g_provider *pp,
542     struct g_raid3_metadata *md, int *errorp)
543 {
544 	struct g_raid3_disk *disk;
545 	int error;
546 
547 	disk = &sc->sc_disks[md->md_no];
548 	error = g_raid3_connect_disk(disk, pp);
549 	if (error != 0) {
550 		if (errorp != NULL)
551 			*errorp = error;
552 		return (NULL);
553 	}
554 	disk->d_state = G_RAID3_DISK_STATE_NONE;
555 	disk->d_flags = md->md_dflags;
556 	if (md->md_provider[0] != '\0')
557 		disk->d_flags |= G_RAID3_DISK_FLAG_HARDCODED;
558 	disk->d_sync.ds_consumer = NULL;
559 	disk->d_sync.ds_offset = md->md_sync_offset;
560 	disk->d_sync.ds_offset_done = md->md_sync_offset;
561 	disk->d_genid = md->md_genid;
562 	disk->d_sync.ds_syncid = md->md_syncid;
563 	if (errorp != NULL)
564 		*errorp = 0;
565 	return (disk);
566 }
567 
568 static void
569 g_raid3_destroy_disk(struct g_raid3_disk *disk)
570 {
571 	struct g_raid3_softc *sc;
572 
573 	g_topology_assert_not();
574 	sc = disk->d_softc;
575 	sx_assert(&sc->sc_lock, SX_XLOCKED);
576 
577 	if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
578 		return;
579 	g_raid3_event_cancel(disk);
580 	switch (disk->d_state) {
581 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
582 		if (sc->sc_syncdisk != NULL)
583 			g_raid3_sync_stop(sc, 1);
584 		/* FALLTHROUGH */
585 	case G_RAID3_DISK_STATE_NEW:
586 	case G_RAID3_DISK_STATE_STALE:
587 	case G_RAID3_DISK_STATE_ACTIVE:
588 		g_topology_lock();
589 		g_raid3_disconnect_consumer(sc, disk->d_consumer);
590 		g_topology_unlock();
591 		disk->d_consumer = NULL;
592 		break;
593 	default:
594 		KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
595 		    g_raid3_get_diskname(disk),
596 		    g_raid3_disk_state2str(disk->d_state)));
597 	}
598 	disk->d_state = G_RAID3_DISK_STATE_NODISK;
599 }
600 
601 static void
602 g_raid3_destroy_device(struct g_raid3_softc *sc)
603 {
604 	struct g_raid3_event *ep;
605 	struct g_raid3_disk *disk;
606 	struct g_geom *gp;
607 	struct g_consumer *cp;
608 	u_int n;
609 
610 	g_topology_assert_not();
611 	sx_assert(&sc->sc_lock, SX_XLOCKED);
612 
613 	gp = sc->sc_geom;
614 	if (sc->sc_provider != NULL)
615 		g_raid3_destroy_provider(sc);
616 	for (n = 0; n < sc->sc_ndisks; n++) {
617 		disk = &sc->sc_disks[n];
618 		if (disk->d_state != G_RAID3_DISK_STATE_NODISK) {
619 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
620 			g_raid3_update_metadata(disk);
621 			g_raid3_destroy_disk(disk);
622 		}
623 	}
624 	while ((ep = g_raid3_event_get(sc)) != NULL) {
625 		g_raid3_event_remove(sc, ep);
626 		if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0)
627 			g_raid3_event_free(ep);
628 		else {
629 			ep->e_error = ECANCELED;
630 			ep->e_flags |= G_RAID3_EVENT_DONE;
631 			G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, ep);
632 			mtx_lock(&sc->sc_events_mtx);
633 			wakeup(ep);
634 			mtx_unlock(&sc->sc_events_mtx);
635 		}
636 	}
637 	callout_drain(&sc->sc_callout);
638 	cp = LIST_FIRST(&sc->sc_sync.ds_geom->consumer);
639 	g_topology_lock();
640 	if (cp != NULL)
641 		g_raid3_disconnect_consumer(sc, cp);
642 	g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
643 	G_RAID3_DEBUG(0, "Device %s destroyed.", gp->name);
644 	g_wither_geom(gp, ENXIO);
645 	g_topology_unlock();
646 	if (!g_raid3_use_malloc) {
647 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_64K].sz_zone);
648 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_16K].sz_zone);
649 		uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_4K].sz_zone);
650 	}
651 	mtx_destroy(&sc->sc_queue_mtx);
652 	mtx_destroy(&sc->sc_events_mtx);
653 	sx_xunlock(&sc->sc_lock);
654 	sx_destroy(&sc->sc_lock);
655 }
656 
657 static void
658 g_raid3_orphan(struct g_consumer *cp)
659 {
660 	struct g_raid3_disk *disk;
661 
662 	g_topology_assert();
663 
664 	disk = cp->private;
665 	if (disk == NULL)
666 		return;
667 	disk->d_softc->sc_bump_id = G_RAID3_BUMP_SYNCID;
668 	g_raid3_event_send(disk, G_RAID3_DISK_STATE_DISCONNECTED,
669 	    G_RAID3_EVENT_DONTWAIT);
670 }
671 
672 static int
673 g_raid3_write_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
674 {
675 	struct g_raid3_softc *sc;
676 	struct g_consumer *cp;
677 	off_t offset, length;
678 	u_char *sector;
679 	int error = 0;
680 
681 	g_topology_assert_not();
682 	sc = disk->d_softc;
683 	sx_assert(&sc->sc_lock, SX_LOCKED);
684 
685 	cp = disk->d_consumer;
686 	KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
687 	KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
688 	KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
689 	    ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
690 	    cp->acw, cp->ace));
691 	length = cp->provider->sectorsize;
692 	offset = cp->provider->mediasize - length;
693 	sector = malloc((size_t)length, M_RAID3, M_WAITOK | M_ZERO);
694 	if (md != NULL)
695 		raid3_metadata_encode(md, sector);
696 	error = g_write_data(cp, offset, sector, length);
697 	free(sector, M_RAID3);
698 	if (error != 0) {
699 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
700 			G_RAID3_DEBUG(0, "Cannot write metadata on %s "
701 			    "(device=%s, error=%d).",
702 			    g_raid3_get_diskname(disk), sc->sc_name, error);
703 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
704 		} else {
705 			G_RAID3_DEBUG(1, "Cannot write metadata on %s "
706 			    "(device=%s, error=%d).",
707 			    g_raid3_get_diskname(disk), sc->sc_name, error);
708 		}
709 		if (g_raid3_disconnect_on_failure &&
710 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
711 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
712 			g_raid3_event_send(disk,
713 			    G_RAID3_DISK_STATE_DISCONNECTED,
714 			    G_RAID3_EVENT_DONTWAIT);
715 		}
716 	}
717 	return (error);
718 }
719 
720 int
721 g_raid3_clear_metadata(struct g_raid3_disk *disk)
722 {
723 	int error;
724 
725 	g_topology_assert_not();
726 	sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
727 
728 	error = g_raid3_write_metadata(disk, NULL);
729 	if (error == 0) {
730 		G_RAID3_DEBUG(2, "Metadata on %s cleared.",
731 		    g_raid3_get_diskname(disk));
732 	} else {
733 		G_RAID3_DEBUG(0,
734 		    "Cannot clear metadata on disk %s (error=%d).",
735 		    g_raid3_get_diskname(disk), error);
736 	}
737 	return (error);
738 }
739 
740 void
741 g_raid3_fill_metadata(struct g_raid3_disk *disk, struct g_raid3_metadata *md)
742 {
743 	struct g_raid3_softc *sc;
744 	struct g_provider *pp;
745 
746 	sc = disk->d_softc;
747 	strlcpy(md->md_magic, G_RAID3_MAGIC, sizeof(md->md_magic));
748 	md->md_version = G_RAID3_VERSION;
749 	strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
750 	md->md_id = sc->sc_id;
751 	md->md_all = sc->sc_ndisks;
752 	md->md_genid = sc->sc_genid;
753 	md->md_mediasize = sc->sc_mediasize;
754 	md->md_sectorsize = sc->sc_sectorsize;
755 	md->md_mflags = (sc->sc_flags & G_RAID3_DEVICE_FLAG_MASK);
756 	md->md_no = disk->d_no;
757 	md->md_syncid = disk->d_sync.ds_syncid;
758 	md->md_dflags = (disk->d_flags & G_RAID3_DISK_FLAG_MASK);
759 	if (disk->d_state != G_RAID3_DISK_STATE_SYNCHRONIZING)
760 		md->md_sync_offset = 0;
761 	else {
762 		md->md_sync_offset =
763 		    disk->d_sync.ds_offset_done / (sc->sc_ndisks - 1);
764 	}
765 	if (disk->d_consumer != NULL && disk->d_consumer->provider != NULL)
766 		pp = disk->d_consumer->provider;
767 	else
768 		pp = NULL;
769 	if ((disk->d_flags & G_RAID3_DISK_FLAG_HARDCODED) != 0 && pp != NULL)
770 		strlcpy(md->md_provider, pp->name, sizeof(md->md_provider));
771 	else
772 		bzero(md->md_provider, sizeof(md->md_provider));
773 	if (pp != NULL)
774 		md->md_provsize = pp->mediasize;
775 	else
776 		md->md_provsize = 0;
777 }
778 
779 void
780 g_raid3_update_metadata(struct g_raid3_disk *disk)
781 {
782 	struct g_raid3_softc *sc;
783 	struct g_raid3_metadata md;
784 	int error;
785 
786 	g_topology_assert_not();
787 	sc = disk->d_softc;
788 	sx_assert(&sc->sc_lock, SX_LOCKED);
789 
790 	g_raid3_fill_metadata(disk, &md);
791 	error = g_raid3_write_metadata(disk, &md);
792 	if (error == 0) {
793 		G_RAID3_DEBUG(2, "Metadata on %s updated.",
794 		    g_raid3_get_diskname(disk));
795 	} else {
796 		G_RAID3_DEBUG(0,
797 		    "Cannot update metadata on disk %s (error=%d).",
798 		    g_raid3_get_diskname(disk), error);
799 	}
800 }
801 
802 static void
803 g_raid3_bump_syncid(struct g_raid3_softc *sc)
804 {
805 	struct g_raid3_disk *disk;
806 	u_int n;
807 
808 	g_topology_assert_not();
809 	sx_assert(&sc->sc_lock, SX_XLOCKED);
810 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
811 	    ("%s called with no active disks (device=%s).", __func__,
812 	    sc->sc_name));
813 
814 	sc->sc_syncid++;
815 	G_RAID3_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
816 	    sc->sc_syncid);
817 	for (n = 0; n < sc->sc_ndisks; n++) {
818 		disk = &sc->sc_disks[n];
819 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
820 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
821 			disk->d_sync.ds_syncid = sc->sc_syncid;
822 			g_raid3_update_metadata(disk);
823 		}
824 	}
825 }
826 
827 static void
828 g_raid3_bump_genid(struct g_raid3_softc *sc)
829 {
830 	struct g_raid3_disk *disk;
831 	u_int n;
832 
833 	g_topology_assert_not();
834 	sx_assert(&sc->sc_lock, SX_XLOCKED);
835 	KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) > 0,
836 	    ("%s called with no active disks (device=%s).", __func__,
837 	    sc->sc_name));
838 
839 	sc->sc_genid++;
840 	G_RAID3_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
841 	    sc->sc_genid);
842 	for (n = 0; n < sc->sc_ndisks; n++) {
843 		disk = &sc->sc_disks[n];
844 		if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
845 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
846 			disk->d_genid = sc->sc_genid;
847 			g_raid3_update_metadata(disk);
848 		}
849 	}
850 }
851 
852 static int
853 g_raid3_idle(struct g_raid3_softc *sc, int acw)
854 {
855 	struct g_raid3_disk *disk;
856 	u_int i;
857 	int timeout;
858 
859 	g_topology_assert_not();
860 	sx_assert(&sc->sc_lock, SX_XLOCKED);
861 
862 	if (sc->sc_provider == NULL)
863 		return (0);
864 	if (sc->sc_idle)
865 		return (0);
866 	if (sc->sc_writes > 0)
867 		return (0);
868 	if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
869 		timeout = g_raid3_idletime - (time_uptime - sc->sc_last_write);
870 		if (timeout > 0)
871 			return (timeout);
872 	}
873 	sc->sc_idle = 1;
874 	for (i = 0; i < sc->sc_ndisks; i++) {
875 		disk = &sc->sc_disks[i];
876 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
877 			continue;
878 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
879 		    g_raid3_get_diskname(disk), sc->sc_name);
880 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
881 		g_raid3_update_metadata(disk);
882 	}
883 	return (0);
884 }
885 
886 static void
887 g_raid3_unidle(struct g_raid3_softc *sc)
888 {
889 	struct g_raid3_disk *disk;
890 	u_int i;
891 
892 	g_topology_assert_not();
893 	sx_assert(&sc->sc_lock, SX_XLOCKED);
894 
895 	sc->sc_idle = 0;
896 	sc->sc_last_write = time_uptime;
897 	for (i = 0; i < sc->sc_ndisks; i++) {
898 		disk = &sc->sc_disks[i];
899 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE)
900 			continue;
901 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
902 		    g_raid3_get_diskname(disk), sc->sc_name);
903 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
904 		g_raid3_update_metadata(disk);
905 	}
906 }
907 
908 /*
909  * Treat bio_driver1 field in parent bio as list head and field bio_caller1
910  * in child bio as pointer to the next element on the list.
911  */
912 #define	G_RAID3_HEAD_BIO(pbp)	(pbp)->bio_driver1
913 
914 #define	G_RAID3_NEXT_BIO(cbp)	(cbp)->bio_caller1
915 
916 #define	G_RAID3_FOREACH_BIO(pbp, bp)					\
917 	for ((bp) = G_RAID3_HEAD_BIO(pbp); (bp) != NULL;		\
918 	    (bp) = G_RAID3_NEXT_BIO(bp))
919 
920 #define	G_RAID3_FOREACH_SAFE_BIO(pbp, bp, tmpbp)			\
921 	for ((bp) = G_RAID3_HEAD_BIO(pbp);				\
922 	    (bp) != NULL && ((tmpbp) = G_RAID3_NEXT_BIO(bp), 1);	\
923 	    (bp) = (tmpbp))
924 
925 static void
926 g_raid3_init_bio(struct bio *pbp)
927 {
928 
929 	G_RAID3_HEAD_BIO(pbp) = NULL;
930 }
931 
932 static void
933 g_raid3_remove_bio(struct bio *cbp)
934 {
935 	struct bio *pbp, *bp;
936 
937 	pbp = cbp->bio_parent;
938 	if (G_RAID3_HEAD_BIO(pbp) == cbp)
939 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
940 	else {
941 		G_RAID3_FOREACH_BIO(pbp, bp) {
942 			if (G_RAID3_NEXT_BIO(bp) == cbp) {
943 				G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
944 				break;
945 			}
946 		}
947 	}
948 	G_RAID3_NEXT_BIO(cbp) = NULL;
949 }
950 
951 static void
952 g_raid3_replace_bio(struct bio *sbp, struct bio *dbp)
953 {
954 	struct bio *pbp, *bp;
955 
956 	g_raid3_remove_bio(sbp);
957 	pbp = dbp->bio_parent;
958 	G_RAID3_NEXT_BIO(sbp) = G_RAID3_NEXT_BIO(dbp);
959 	if (G_RAID3_HEAD_BIO(pbp) == dbp)
960 		G_RAID3_HEAD_BIO(pbp) = sbp;
961 	else {
962 		G_RAID3_FOREACH_BIO(pbp, bp) {
963 			if (G_RAID3_NEXT_BIO(bp) == dbp) {
964 				G_RAID3_NEXT_BIO(bp) = sbp;
965 				break;
966 			}
967 		}
968 	}
969 	G_RAID3_NEXT_BIO(dbp) = NULL;
970 }
971 
972 static void
973 g_raid3_destroy_bio(struct g_raid3_softc *sc, struct bio *cbp)
974 {
975 	struct bio *bp, *pbp;
976 	size_t size;
977 
978 	pbp = cbp->bio_parent;
979 	pbp->bio_children--;
980 	KASSERT(cbp->bio_data != NULL, ("NULL bio_data"));
981 	size = pbp->bio_length / (sc->sc_ndisks - 1);
982 	g_raid3_free(sc, cbp->bio_data, size);
983 	if (G_RAID3_HEAD_BIO(pbp) == cbp) {
984 		G_RAID3_HEAD_BIO(pbp) = G_RAID3_NEXT_BIO(cbp);
985 		G_RAID3_NEXT_BIO(cbp) = NULL;
986 		g_destroy_bio(cbp);
987 	} else {
988 		G_RAID3_FOREACH_BIO(pbp, bp) {
989 			if (G_RAID3_NEXT_BIO(bp) == cbp)
990 				break;
991 		}
992 		if (bp != NULL) {
993 			KASSERT(G_RAID3_NEXT_BIO(bp) != NULL,
994 			    ("NULL bp->bio_driver1"));
995 			G_RAID3_NEXT_BIO(bp) = G_RAID3_NEXT_BIO(cbp);
996 			G_RAID3_NEXT_BIO(cbp) = NULL;
997 		}
998 		g_destroy_bio(cbp);
999 	}
1000 }
1001 
1002 static struct bio *
1003 g_raid3_clone_bio(struct g_raid3_softc *sc, struct bio *pbp)
1004 {
1005 	struct bio *bp, *cbp;
1006 	size_t size;
1007 	int memflag;
1008 
1009 	cbp = g_clone_bio(pbp);
1010 	if (cbp == NULL)
1011 		return (NULL);
1012 	size = pbp->bio_length / (sc->sc_ndisks - 1);
1013 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
1014 		memflag = M_WAITOK;
1015 	else
1016 		memflag = M_NOWAIT;
1017 	cbp->bio_data = g_raid3_alloc(sc, size, memflag);
1018 	if (cbp->bio_data == NULL) {
1019 		pbp->bio_children--;
1020 		g_destroy_bio(cbp);
1021 		return (NULL);
1022 	}
1023 	G_RAID3_NEXT_BIO(cbp) = NULL;
1024 	if (G_RAID3_HEAD_BIO(pbp) == NULL)
1025 		G_RAID3_HEAD_BIO(pbp) = cbp;
1026 	else {
1027 		G_RAID3_FOREACH_BIO(pbp, bp) {
1028 			if (G_RAID3_NEXT_BIO(bp) == NULL) {
1029 				G_RAID3_NEXT_BIO(bp) = cbp;
1030 				break;
1031 			}
1032 		}
1033 	}
1034 	return (cbp);
1035 }
1036 
1037 static void
1038 g_raid3_scatter(struct bio *pbp)
1039 {
1040 	struct g_raid3_softc *sc;
1041 	struct g_raid3_disk *disk;
1042 	struct bio *bp, *cbp, *tmpbp;
1043 	off_t atom, cadd, padd, left;
1044 
1045 	sc = pbp->bio_to->geom->softc;
1046 	bp = NULL;
1047 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1048 		/*
1049 		 * Find bio for which we should calculate data.
1050 		 */
1051 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1052 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1053 				bp = cbp;
1054 				break;
1055 			}
1056 		}
1057 		KASSERT(bp != NULL, ("NULL parity bio."));
1058 	}
1059 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1060 	cadd = padd = 0;
1061 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1062 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1063 			if (cbp == bp)
1064 				continue;
1065 			bcopy(pbp->bio_data + padd, cbp->bio_data + cadd, atom);
1066 			padd += atom;
1067 		}
1068 		cadd += atom;
1069 	}
1070 	if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_NOPARITY) == 0) {
1071 		/*
1072 		 * Calculate parity.
1073 		 */
1074 		bzero(bp->bio_data, bp->bio_length);
1075 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1076 			if (cbp == bp)
1077 				continue;
1078 			g_raid3_xor(cbp->bio_data, bp->bio_data, bp->bio_data,
1079 			    bp->bio_length);
1080 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_NODISK) != 0)
1081 				g_raid3_destroy_bio(sc, cbp);
1082 		}
1083 	}
1084 	G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1085 		struct g_consumer *cp;
1086 
1087 		disk = cbp->bio_caller2;
1088 		cp = disk->d_consumer;
1089 		cbp->bio_to = cp->provider;
1090 		G_RAID3_LOGREQ(3, cbp, "Sending request.");
1091 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1092 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1093 		    cp->acr, cp->acw, cp->ace));
1094 		cp->index++;
1095 		sc->sc_writes++;
1096 		g_io_request(cbp, cp);
1097 	}
1098 }
1099 
1100 static void
1101 g_raid3_gather(struct bio *pbp)
1102 {
1103 	struct g_raid3_softc *sc;
1104 	struct g_raid3_disk *disk;
1105 	struct bio *xbp, *fbp, *cbp;
1106 	off_t atom, cadd, padd, left;
1107 
1108 	sc = pbp->bio_to->geom->softc;
1109 	/*
1110 	 * Find bio for which we have to calculate data.
1111 	 * While going through this path, check if all requests
1112 	 * succeeded, if not, deny whole request.
1113 	 * If we're in COMPLETE mode, we allow one request to fail,
1114 	 * so if we find one, we're sending it to the parity consumer.
1115 	 * If there are more failed requests, we deny whole request.
1116 	 */
1117 	xbp = fbp = NULL;
1118 	G_RAID3_FOREACH_BIO(pbp, cbp) {
1119 		if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0) {
1120 			KASSERT(xbp == NULL, ("More than one parity bio."));
1121 			xbp = cbp;
1122 		}
1123 		if (cbp->bio_error == 0)
1124 			continue;
1125 		/*
1126 		 * Found failed request.
1127 		 */
1128 		if (fbp == NULL) {
1129 			if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_DEGRADED) != 0) {
1130 				/*
1131 				 * We are already in degraded mode, so we can't
1132 				 * accept any failures.
1133 				 */
1134 				if (pbp->bio_error == 0)
1135 					pbp->bio_error = cbp->bio_error;
1136 			} else {
1137 				fbp = cbp;
1138 			}
1139 		} else {
1140 			/*
1141 			 * Next failed request, that's too many.
1142 			 */
1143 			if (pbp->bio_error == 0)
1144 				pbp->bio_error = fbp->bio_error;
1145 		}
1146 		disk = cbp->bio_caller2;
1147 		if (disk == NULL)
1148 			continue;
1149 		if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1150 			disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1151 			G_RAID3_LOGREQ(0, cbp, "Request failed (error=%d).",
1152 			    cbp->bio_error);
1153 		} else {
1154 			G_RAID3_LOGREQ(1, cbp, "Request failed (error=%d).",
1155 			    cbp->bio_error);
1156 		}
1157 		if (g_raid3_disconnect_on_failure &&
1158 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1159 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1160 			g_raid3_event_send(disk,
1161 			    G_RAID3_DISK_STATE_DISCONNECTED,
1162 			    G_RAID3_EVENT_DONTWAIT);
1163 		}
1164 	}
1165 	if (pbp->bio_error != 0)
1166 		goto finish;
1167 	if (fbp != NULL && (pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1168 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_VERIFY;
1169 		if (xbp != fbp)
1170 			g_raid3_replace_bio(xbp, fbp);
1171 		g_raid3_destroy_bio(sc, fbp);
1172 	} else if (fbp != NULL) {
1173 		struct g_consumer *cp;
1174 
1175 		/*
1176 		 * One request failed, so send the same request to
1177 		 * the parity consumer.
1178 		 */
1179 		disk = pbp->bio_driver2;
1180 		if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1181 			pbp->bio_error = fbp->bio_error;
1182 			goto finish;
1183 		}
1184 		pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1185 		pbp->bio_inbed--;
1186 		fbp->bio_flags &= ~(BIO_DONE | BIO_ERROR);
1187 		if (disk->d_no == sc->sc_ndisks - 1)
1188 			fbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1189 		fbp->bio_error = 0;
1190 		fbp->bio_completed = 0;
1191 		fbp->bio_children = 0;
1192 		fbp->bio_inbed = 0;
1193 		cp = disk->d_consumer;
1194 		fbp->bio_caller2 = disk;
1195 		fbp->bio_to = cp->provider;
1196 		G_RAID3_LOGREQ(3, fbp, "Sending request (recover).");
1197 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1198 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1199 		    cp->acr, cp->acw, cp->ace));
1200 		cp->index++;
1201 		g_io_request(fbp, cp);
1202 		return;
1203 	}
1204 	if (xbp != NULL) {
1205 		/*
1206 		 * Calculate parity.
1207 		 */
1208 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1209 			if ((cbp->bio_cflags & G_RAID3_BIO_CFLAG_PARITY) != 0)
1210 				continue;
1211 			g_raid3_xor(cbp->bio_data, xbp->bio_data, xbp->bio_data,
1212 			    xbp->bio_length);
1213 		}
1214 		xbp->bio_cflags &= ~G_RAID3_BIO_CFLAG_PARITY;
1215 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0) {
1216 			if (!g_raid3_is_zero(xbp)) {
1217 				g_raid3_parity_mismatch++;
1218 				pbp->bio_error = EIO;
1219 				goto finish;
1220 			}
1221 			g_raid3_destroy_bio(sc, xbp);
1222 		}
1223 	}
1224 	atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1225 	cadd = padd = 0;
1226 	for (left = pbp->bio_length; left > 0; left -= sc->sc_sectorsize) {
1227 		G_RAID3_FOREACH_BIO(pbp, cbp) {
1228 			bcopy(cbp->bio_data + cadd, pbp->bio_data + padd, atom);
1229 			pbp->bio_completed += atom;
1230 			padd += atom;
1231 		}
1232 		cadd += atom;
1233 	}
1234 finish:
1235 	if (pbp->bio_error == 0)
1236 		G_RAID3_LOGREQ(3, pbp, "Request finished.");
1237 	else {
1238 		if ((pbp->bio_pflags & G_RAID3_BIO_PFLAG_VERIFY) != 0)
1239 			G_RAID3_LOGREQ(1, pbp, "Verification error.");
1240 		else
1241 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1242 	}
1243 	pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_MASK;
1244 	while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1245 		g_raid3_destroy_bio(sc, cbp);
1246 	g_io_deliver(pbp, pbp->bio_error);
1247 }
1248 
1249 static void
1250 g_raid3_done(struct bio *bp)
1251 {
1252 	struct g_raid3_softc *sc;
1253 
1254 	sc = bp->bio_from->geom->softc;
1255 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_REGULAR;
1256 	G_RAID3_LOGREQ(3, bp, "Regular request done (error=%d).", bp->bio_error);
1257 	mtx_lock(&sc->sc_queue_mtx);
1258 	bioq_insert_head(&sc->sc_queue, bp);
1259 	wakeup(sc);
1260 	wakeup(&sc->sc_queue);
1261 	mtx_unlock(&sc->sc_queue_mtx);
1262 }
1263 
1264 static void
1265 g_raid3_regular_request(struct bio *cbp)
1266 {
1267 	struct g_raid3_softc *sc;
1268 	struct g_raid3_disk *disk;
1269 	struct bio *pbp;
1270 
1271 	g_topology_assert_not();
1272 
1273 	pbp = cbp->bio_parent;
1274 	sc = pbp->bio_to->geom->softc;
1275 	cbp->bio_from->index--;
1276 	if (cbp->bio_cmd == BIO_WRITE)
1277 		sc->sc_writes--;
1278 	disk = cbp->bio_from->private;
1279 	if (disk == NULL) {
1280 		g_topology_lock();
1281 		g_raid3_kill_consumer(sc, cbp->bio_from);
1282 		g_topology_unlock();
1283 	}
1284 
1285 	G_RAID3_LOGREQ(3, cbp, "Request finished.");
1286 	pbp->bio_inbed++;
1287 	KASSERT(pbp->bio_inbed <= pbp->bio_children,
1288 	    ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
1289 	    pbp->bio_children));
1290 	if (pbp->bio_inbed != pbp->bio_children)
1291 		return;
1292 	switch (pbp->bio_cmd) {
1293 	case BIO_READ:
1294 		g_raid3_gather(pbp);
1295 		break;
1296 	case BIO_WRITE:
1297 	case BIO_DELETE:
1298 	    {
1299 		int error = 0;
1300 
1301 		pbp->bio_completed = pbp->bio_length;
1302 		while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL) {
1303 			if (cbp->bio_error == 0) {
1304 				g_raid3_destroy_bio(sc, cbp);
1305 				continue;
1306 			}
1307 
1308 			if (error == 0)
1309 				error = cbp->bio_error;
1310 			else if (pbp->bio_error == 0) {
1311 				/*
1312 				 * Next failed request, that's too many.
1313 				 */
1314 				pbp->bio_error = error;
1315 			}
1316 
1317 			disk = cbp->bio_caller2;
1318 			if (disk == NULL) {
1319 				g_raid3_destroy_bio(sc, cbp);
1320 				continue;
1321 			}
1322 
1323 			if ((disk->d_flags & G_RAID3_DISK_FLAG_BROKEN) == 0) {
1324 				disk->d_flags |= G_RAID3_DISK_FLAG_BROKEN;
1325 				G_RAID3_LOGREQ(0, cbp,
1326 				    "Request failed (error=%d).",
1327 				    cbp->bio_error);
1328 			} else {
1329 				G_RAID3_LOGREQ(1, cbp,
1330 				    "Request failed (error=%d).",
1331 				    cbp->bio_error);
1332 			}
1333 			if (g_raid3_disconnect_on_failure &&
1334 			    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1335 				sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1336 				g_raid3_event_send(disk,
1337 				    G_RAID3_DISK_STATE_DISCONNECTED,
1338 				    G_RAID3_EVENT_DONTWAIT);
1339 			}
1340 			g_raid3_destroy_bio(sc, cbp);
1341 		}
1342 		if (pbp->bio_error == 0)
1343 			G_RAID3_LOGREQ(3, pbp, "Request finished.");
1344 		else
1345 			G_RAID3_LOGREQ(0, pbp, "Request failed.");
1346 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_DEGRADED;
1347 		pbp->bio_pflags &= ~G_RAID3_BIO_PFLAG_NOPARITY;
1348 		bioq_remove(&sc->sc_inflight, pbp);
1349 		/* Release delayed sync requests if possible. */
1350 		g_raid3_sync_release(sc);
1351 		g_io_deliver(pbp, pbp->bio_error);
1352 		break;
1353 	    }
1354 	}
1355 }
1356 
1357 static void
1358 g_raid3_sync_done(struct bio *bp)
1359 {
1360 	struct g_raid3_softc *sc;
1361 
1362 	G_RAID3_LOGREQ(3, bp, "Synchronization request delivered.");
1363 	sc = bp->bio_from->geom->softc;
1364 	bp->bio_cflags |= G_RAID3_BIO_CFLAG_SYNC;
1365 	mtx_lock(&sc->sc_queue_mtx);
1366 	bioq_insert_head(&sc->sc_queue, bp);
1367 	wakeup(sc);
1368 	wakeup(&sc->sc_queue);
1369 	mtx_unlock(&sc->sc_queue_mtx);
1370 }
1371 
1372 static void
1373 g_raid3_start(struct bio *bp)
1374 {
1375 	struct g_raid3_softc *sc;
1376 
1377 	sc = bp->bio_to->geom->softc;
1378 	/*
1379 	 * If sc == NULL or there are no valid disks, provider's error
1380 	 * should be set and g_raid3_start() should not be called at all.
1381 	 */
1382 	KASSERT(sc != NULL && (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
1383 	    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE),
1384 	    ("Provider's error should be set (error=%d)(device=%s).",
1385 	    bp->bio_to->error, bp->bio_to->name));
1386 	G_RAID3_LOGREQ(3, bp, "Request received.");
1387 
1388 	switch (bp->bio_cmd) {
1389 	case BIO_READ:
1390 	case BIO_WRITE:
1391 	case BIO_DELETE:
1392 		break;
1393 	case BIO_GETATTR:
1394 	default:
1395 		g_io_deliver(bp, EOPNOTSUPP);
1396 		return;
1397 	}
1398 	mtx_lock(&sc->sc_queue_mtx);
1399 	bioq_insert_tail(&sc->sc_queue, bp);
1400 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1401 	wakeup(sc);
1402 	mtx_unlock(&sc->sc_queue_mtx);
1403 }
1404 
1405 /*
1406  * Return TRUE if the given request is colliding with a in-progress
1407  * synchronization request.
1408  */
1409 static int
1410 g_raid3_sync_collision(struct g_raid3_softc *sc, struct bio *bp)
1411 {
1412 	struct g_raid3_disk *disk;
1413 	struct bio *sbp;
1414 	off_t rstart, rend, sstart, send;
1415 	int i;
1416 
1417 	disk = sc->sc_syncdisk;
1418 	if (disk == NULL)
1419 		return (0);
1420 	rstart = bp->bio_offset;
1421 	rend = bp->bio_offset + bp->bio_length;
1422 	for (i = 0; i < g_raid3_syncreqs; i++) {
1423 		sbp = disk->d_sync.ds_bios[i];
1424 		if (sbp == NULL)
1425 			continue;
1426 		sstart = sbp->bio_offset;
1427 		send = sbp->bio_length;
1428 		if (sbp->bio_cmd == BIO_WRITE) {
1429 			sstart *= sc->sc_ndisks - 1;
1430 			send *= sc->sc_ndisks - 1;
1431 		}
1432 		send += sstart;
1433 		if (rend > sstart && rstart < send)
1434 			return (1);
1435 	}
1436 	return (0);
1437 }
1438 
1439 /*
1440  * Return TRUE if the given sync request is colliding with a in-progress regular
1441  * request.
1442  */
1443 static int
1444 g_raid3_regular_collision(struct g_raid3_softc *sc, struct bio *sbp)
1445 {
1446 	off_t rstart, rend, sstart, send;
1447 	struct bio *bp;
1448 
1449 	if (sc->sc_syncdisk == NULL)
1450 		return (0);
1451 	sstart = sbp->bio_offset;
1452 	send = sstart + sbp->bio_length;
1453 	TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1454 		rstart = bp->bio_offset;
1455 		rend = bp->bio_offset + bp->bio_length;
1456 		if (rend > sstart && rstart < send)
1457 			return (1);
1458 	}
1459 	return (0);
1460 }
1461 
1462 /*
1463  * Puts request onto delayed queue.
1464  */
1465 static void
1466 g_raid3_regular_delay(struct g_raid3_softc *sc, struct bio *bp)
1467 {
1468 
1469         G_RAID3_LOGREQ(2, bp, "Delaying request.");
1470         bioq_insert_head(&sc->sc_regular_delayed, bp);
1471 }
1472 
1473 /*
1474  * Puts synchronization request onto delayed queue.
1475  */
1476 static void
1477 g_raid3_sync_delay(struct g_raid3_softc *sc, struct bio *bp)
1478 {
1479 
1480         G_RAID3_LOGREQ(2, bp, "Delaying synchronization request.");
1481         bioq_insert_tail(&sc->sc_sync_delayed, bp);
1482 }
1483 
1484 /*
1485  * Releases delayed regular requests which don't collide anymore with sync
1486  * requests.
1487  */
1488 static void
1489 g_raid3_regular_release(struct g_raid3_softc *sc)
1490 {
1491         struct bio *bp, *bp2;
1492 
1493         TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1494                 if (g_raid3_sync_collision(sc, bp))
1495                         continue;
1496                 bioq_remove(&sc->sc_regular_delayed, bp);
1497                 G_RAID3_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1498 		mtx_lock(&sc->sc_queue_mtx);
1499 		bioq_insert_head(&sc->sc_queue, bp);
1500 #if 0
1501 		/*
1502 		 * wakeup() is not needed, because this function is called from
1503 		 * the worker thread.
1504 		 */
1505 		wakeup(&sc->sc_queue);
1506 #endif
1507 		mtx_unlock(&sc->sc_queue_mtx);
1508         }
1509 }
1510 
1511 /*
1512  * Releases delayed sync requests which don't collide anymore with regular
1513  * requests.
1514  */
1515 static void
1516 g_raid3_sync_release(struct g_raid3_softc *sc)
1517 {
1518         struct bio *bp, *bp2;
1519 
1520         TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1521                 if (g_raid3_regular_collision(sc, bp))
1522                         continue;
1523                 bioq_remove(&sc->sc_sync_delayed, bp);
1524                 G_RAID3_LOGREQ(2, bp,
1525                     "Releasing delayed synchronization request.");
1526                 g_io_request(bp, bp->bio_from);
1527         }
1528 }
1529 
1530 /*
1531  * Handle synchronization requests.
1532  * Every synchronization request is two-steps process: first, READ request is
1533  * send to active provider and then WRITE request (with read data) to the provider
1534  * beeing synchronized. When WRITE is finished, new synchronization request is
1535  * send.
1536  */
1537 static void
1538 g_raid3_sync_request(struct bio *bp)
1539 {
1540 	struct g_raid3_softc *sc;
1541 	struct g_raid3_disk *disk;
1542 
1543 	bp->bio_from->index--;
1544 	sc = bp->bio_from->geom->softc;
1545 	disk = bp->bio_from->private;
1546 	if (disk == NULL) {
1547 		sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1548 		g_topology_lock();
1549 		g_raid3_kill_consumer(sc, bp->bio_from);
1550 		g_topology_unlock();
1551 		free(bp->bio_data, M_RAID3);
1552 		g_destroy_bio(bp);
1553 		sx_xlock(&sc->sc_lock);
1554 		return;
1555 	}
1556 
1557 	/*
1558 	 * Synchronization request.
1559 	 */
1560 	switch (bp->bio_cmd) {
1561 	case BIO_READ:
1562 	    {
1563 		struct g_consumer *cp;
1564 		u_char *dst, *src;
1565 		off_t left;
1566 		u_int atom;
1567 
1568 		if (bp->bio_error != 0) {
1569 			G_RAID3_LOGREQ(0, bp,
1570 			    "Synchronization request failed (error=%d).",
1571 			    bp->bio_error);
1572 			g_destroy_bio(bp);
1573 			return;
1574 		}
1575 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1576 		atom = sc->sc_sectorsize / (sc->sc_ndisks - 1);
1577 		dst = src = bp->bio_data;
1578 		if (disk->d_no == sc->sc_ndisks - 1) {
1579 			u_int n;
1580 
1581 			/* Parity component. */
1582 			for (left = bp->bio_length; left > 0;
1583 			    left -= sc->sc_sectorsize) {
1584 				bcopy(src, dst, atom);
1585 				src += atom;
1586 				for (n = 1; n < sc->sc_ndisks - 1; n++) {
1587 					g_raid3_xor(src, dst, dst, atom);
1588 					src += atom;
1589 				}
1590 				dst += atom;
1591 			}
1592 		} else {
1593 			/* Regular component. */
1594 			src += atom * disk->d_no;
1595 			for (left = bp->bio_length; left > 0;
1596 			    left -= sc->sc_sectorsize) {
1597 				bcopy(src, dst, atom);
1598 				src += sc->sc_sectorsize;
1599 				dst += atom;
1600 			}
1601 		}
1602 		bp->bio_driver1 = bp->bio_driver2 = NULL;
1603 		bp->bio_pflags = 0;
1604 		bp->bio_offset /= sc->sc_ndisks - 1;
1605 		bp->bio_length /= sc->sc_ndisks - 1;
1606 		bp->bio_cmd = BIO_WRITE;
1607 		bp->bio_cflags = 0;
1608 		bp->bio_children = bp->bio_inbed = 0;
1609 		cp = disk->d_consumer;
1610 		KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1611 		    ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1612 		    cp->acr, cp->acw, cp->ace));
1613 		cp->index++;
1614 		g_io_request(bp, cp);
1615 		return;
1616 	    }
1617 	case BIO_WRITE:
1618 	    {
1619 		struct g_raid3_disk_sync *sync;
1620 		off_t boffset, moffset;
1621 		void *data;
1622 		int i;
1623 
1624 		if (bp->bio_error != 0) {
1625 			G_RAID3_LOGREQ(0, bp,
1626 			    "Synchronization request failed (error=%d).",
1627 			    bp->bio_error);
1628 			g_destroy_bio(bp);
1629 			sc->sc_bump_id |= G_RAID3_BUMP_GENID;
1630 			g_raid3_event_send(disk,
1631 			    G_RAID3_DISK_STATE_DISCONNECTED,
1632 			    G_RAID3_EVENT_DONTWAIT);
1633 			return;
1634 		}
1635 		G_RAID3_LOGREQ(3, bp, "Synchronization request finished.");
1636 		sync = &disk->d_sync;
1637 		if (sync->ds_offset == sc->sc_mediasize / (sc->sc_ndisks - 1) ||
1638 		    sync->ds_consumer == NULL ||
1639 		    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1640 			/* Don't send more synchronization requests. */
1641 			sync->ds_inflight--;
1642 			if (sync->ds_bios != NULL) {
1643 				i = (int)(uintptr_t)bp->bio_caller1;
1644 				sync->ds_bios[i] = NULL;
1645 			}
1646 			free(bp->bio_data, M_RAID3);
1647 			g_destroy_bio(bp);
1648 			if (sync->ds_inflight > 0)
1649 				return;
1650 			if (sync->ds_consumer == NULL ||
1651 			    (sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
1652 				return;
1653 			}
1654 			/*
1655 			 * Disk up-to-date, activate it.
1656 			 */
1657 			g_raid3_event_send(disk, G_RAID3_DISK_STATE_ACTIVE,
1658 			    G_RAID3_EVENT_DONTWAIT);
1659 			return;
1660 		}
1661 
1662 		/* Send next synchronization request. */
1663 		data = bp->bio_data;
1664 		bzero(bp, sizeof(*bp));
1665 		bp->bio_cmd = BIO_READ;
1666 		bp->bio_offset = sync->ds_offset * (sc->sc_ndisks - 1);
1667 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1668 		sync->ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
1669 		bp->bio_done = g_raid3_sync_done;
1670 		bp->bio_data = data;
1671 		bp->bio_from = sync->ds_consumer;
1672 		bp->bio_to = sc->sc_provider;
1673 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
1674 		sync->ds_consumer->index++;
1675 		/*
1676 		 * Delay the request if it is colliding with a regular request.
1677 		 */
1678 		if (g_raid3_regular_collision(sc, bp))
1679 			g_raid3_sync_delay(sc, bp);
1680 		else
1681 			g_io_request(bp, sync->ds_consumer);
1682 
1683 		/* Release delayed requests if possible. */
1684 		g_raid3_regular_release(sc);
1685 
1686 		/* Find the smallest offset. */
1687 		moffset = sc->sc_mediasize;
1688 		for (i = 0; i < g_raid3_syncreqs; i++) {
1689 			bp = sync->ds_bios[i];
1690 			boffset = bp->bio_offset;
1691 			if (bp->bio_cmd == BIO_WRITE)
1692 				boffset *= sc->sc_ndisks - 1;
1693 			if (boffset < moffset)
1694 				moffset = boffset;
1695 		}
1696 		if (sync->ds_offset_done + (MAXPHYS * 100) < moffset) {
1697 			/* Update offset_done on every 100 blocks. */
1698 			sync->ds_offset_done = moffset;
1699 			g_raid3_update_metadata(disk);
1700 		}
1701 		return;
1702 	    }
1703 	default:
1704 		KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1705 		    bp->bio_cmd, sc->sc_name));
1706 		break;
1707 	}
1708 }
1709 
1710 static int
1711 g_raid3_register_request(struct bio *pbp)
1712 {
1713 	struct g_raid3_softc *sc;
1714 	struct g_raid3_disk *disk;
1715 	struct g_consumer *cp;
1716 	struct bio *cbp, *tmpbp;
1717 	off_t offset, length;
1718 	u_int n, ndisks;
1719 	int round_robin, verify;
1720 
1721 	ndisks = 0;
1722 	sc = pbp->bio_to->geom->softc;
1723 	if ((pbp->bio_cflags & G_RAID3_BIO_CFLAG_REGSYNC) != 0 &&
1724 	    sc->sc_syncdisk == NULL) {
1725 		g_io_deliver(pbp, EIO);
1726 		return (0);
1727 	}
1728 	g_raid3_init_bio(pbp);
1729 	length = pbp->bio_length / (sc->sc_ndisks - 1);
1730 	offset = pbp->bio_offset / (sc->sc_ndisks - 1);
1731 	round_robin = verify = 0;
1732 	switch (pbp->bio_cmd) {
1733 	case BIO_READ:
1734 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
1735 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1736 			pbp->bio_pflags |= G_RAID3_BIO_PFLAG_VERIFY;
1737 			verify = 1;
1738 			ndisks = sc->sc_ndisks;
1739 		} else {
1740 			verify = 0;
1741 			ndisks = sc->sc_ndisks - 1;
1742 		}
1743 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0 &&
1744 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
1745 			round_robin = 1;
1746 		} else {
1747 			round_robin = 0;
1748 		}
1749 		KASSERT(!round_robin || !verify,
1750 		    ("ROUND-ROBIN and VERIFY are mutually exclusive."));
1751 		pbp->bio_driver2 = &sc->sc_disks[sc->sc_ndisks - 1];
1752 		break;
1753 	case BIO_WRITE:
1754 	case BIO_DELETE:
1755 		/*
1756 		 * Delay the request if it is colliding with a synchronization
1757 		 * request.
1758 		 */
1759 		if (g_raid3_sync_collision(sc, pbp)) {
1760 			g_raid3_regular_delay(sc, pbp);
1761 			return (0);
1762 		}
1763 
1764 		if (sc->sc_idle)
1765 			g_raid3_unidle(sc);
1766 		else
1767 			sc->sc_last_write = time_uptime;
1768 
1769 		ndisks = sc->sc_ndisks;
1770 		break;
1771 	}
1772 	for (n = 0; n < ndisks; n++) {
1773 		disk = &sc->sc_disks[n];
1774 		cbp = g_raid3_clone_bio(sc, pbp);
1775 		if (cbp == NULL) {
1776 			while ((cbp = G_RAID3_HEAD_BIO(pbp)) != NULL)
1777 				g_raid3_destroy_bio(sc, cbp);
1778 			/*
1779 			 * To prevent deadlock, we must run back up
1780 			 * with the ENOMEM for failed requests of any
1781 			 * of our consumers.  Our own sync requests
1782 			 * can stick around, as they are finite.
1783 			 */
1784 			if ((pbp->bio_cflags &
1785 			    G_RAID3_BIO_CFLAG_REGULAR) != 0) {
1786 				g_io_deliver(pbp, ENOMEM);
1787 				return (0);
1788 			}
1789 			return (ENOMEM);
1790 		}
1791 		cbp->bio_offset = offset;
1792 		cbp->bio_length = length;
1793 		cbp->bio_done = g_raid3_done;
1794 		switch (pbp->bio_cmd) {
1795 		case BIO_READ:
1796 			if (disk->d_state != G_RAID3_DISK_STATE_ACTIVE) {
1797 				/*
1798 				 * Replace invalid component with the parity
1799 				 * component.
1800 				 */
1801 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1802 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1803 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1804 			} else if (round_robin &&
1805 			    disk->d_no == sc->sc_round_robin) {
1806 				/*
1807 				 * In round-robin mode skip one data component
1808 				 * and use parity component when reading.
1809 				 */
1810 				pbp->bio_driver2 = disk;
1811 				disk = &sc->sc_disks[sc->sc_ndisks - 1];
1812 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1813 				sc->sc_round_robin++;
1814 				round_robin = 0;
1815 			} else if (verify && disk->d_no == sc->sc_ndisks - 1) {
1816 				cbp->bio_cflags |= G_RAID3_BIO_CFLAG_PARITY;
1817 			}
1818 			break;
1819 		case BIO_WRITE:
1820 		case BIO_DELETE:
1821 			if (disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
1822 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
1823 				if (n == ndisks - 1) {
1824 					/*
1825 					 * Active parity component, mark it as such.
1826 					 */
1827 					cbp->bio_cflags |=
1828 					    G_RAID3_BIO_CFLAG_PARITY;
1829 				}
1830 			} else {
1831 				pbp->bio_pflags |= G_RAID3_BIO_PFLAG_DEGRADED;
1832 				if (n == ndisks - 1) {
1833 					/*
1834 					 * Parity component is not connected,
1835 					 * so destroy its request.
1836 					 */
1837 					pbp->bio_pflags |=
1838 					    G_RAID3_BIO_PFLAG_NOPARITY;
1839 					g_raid3_destroy_bio(sc, cbp);
1840 					cbp = NULL;
1841 				} else {
1842 					cbp->bio_cflags |=
1843 					    G_RAID3_BIO_CFLAG_NODISK;
1844 					disk = NULL;
1845 				}
1846 			}
1847 			break;
1848 		}
1849 		if (cbp != NULL)
1850 			cbp->bio_caller2 = disk;
1851 	}
1852 	switch (pbp->bio_cmd) {
1853 	case BIO_READ:
1854 		if (round_robin) {
1855 			/*
1856 			 * If we are in round-robin mode and 'round_robin' is
1857 			 * still 1, it means, that we skipped parity component
1858 			 * for this read and must reset sc_round_robin field.
1859 			 */
1860 			sc->sc_round_robin = 0;
1861 		}
1862 		G_RAID3_FOREACH_SAFE_BIO(pbp, cbp, tmpbp) {
1863 			disk = cbp->bio_caller2;
1864 			cp = disk->d_consumer;
1865 			cbp->bio_to = cp->provider;
1866 			G_RAID3_LOGREQ(3, cbp, "Sending request.");
1867 			KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1868 			    ("Consumer %s not opened (r%dw%de%d).",
1869 			    cp->provider->name, cp->acr, cp->acw, cp->ace));
1870 			cp->index++;
1871 			g_io_request(cbp, cp);
1872 		}
1873 		break;
1874 	case BIO_WRITE:
1875 	case BIO_DELETE:
1876 		/*
1877 		 * Put request onto inflight queue, so we can check if new
1878 		 * synchronization requests don't collide with it.
1879 		 */
1880 		bioq_insert_tail(&sc->sc_inflight, pbp);
1881 
1882 		/*
1883 		 * Bump syncid on first write.
1884 		 */
1885 		if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0) {
1886 			sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
1887 			g_raid3_bump_syncid(sc);
1888 		}
1889 		g_raid3_scatter(pbp);
1890 		break;
1891 	}
1892 	return (0);
1893 }
1894 
1895 static int
1896 g_raid3_can_destroy(struct g_raid3_softc *sc)
1897 {
1898 	struct g_geom *gp;
1899 	struct g_consumer *cp;
1900 
1901 	g_topology_assert();
1902 	gp = sc->sc_geom;
1903 	if (gp->softc == NULL)
1904 		return (1);
1905 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1906 		if (g_raid3_is_busy(sc, cp))
1907 			return (0);
1908 	}
1909 	gp = sc->sc_sync.ds_geom;
1910 	LIST_FOREACH(cp, &gp->consumer, consumer) {
1911 		if (g_raid3_is_busy(sc, cp))
1912 			return (0);
1913 	}
1914 	G_RAID3_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1915 	    sc->sc_name);
1916 	return (1);
1917 }
1918 
1919 static int
1920 g_raid3_try_destroy(struct g_raid3_softc *sc)
1921 {
1922 
1923 	g_topology_assert_not();
1924 	sx_assert(&sc->sc_lock, SX_XLOCKED);
1925 
1926 	if (sc->sc_rootmount != NULL) {
1927 		G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1928 		    sc->sc_rootmount);
1929 		root_mount_rel(sc->sc_rootmount);
1930 		sc->sc_rootmount = NULL;
1931 	}
1932 
1933 	g_topology_lock();
1934 	if (!g_raid3_can_destroy(sc)) {
1935 		g_topology_unlock();
1936 		return (0);
1937 	}
1938 	sc->sc_geom->softc = NULL;
1939 	sc->sc_sync.ds_geom->softc = NULL;
1940 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_WAIT) != 0) {
1941 		g_topology_unlock();
1942 		G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
1943 		    &sc->sc_worker);
1944 		/* Unlock sc_lock here, as it can be destroyed after wakeup. */
1945 		sx_xunlock(&sc->sc_lock);
1946 		wakeup(&sc->sc_worker);
1947 		sc->sc_worker = NULL;
1948 	} else {
1949 		g_topology_unlock();
1950 		g_raid3_destroy_device(sc);
1951 		free(sc->sc_disks, M_RAID3);
1952 		free(sc, M_RAID3);
1953 	}
1954 	return (1);
1955 }
1956 
1957 /*
1958  * Worker thread.
1959  */
1960 static void
1961 g_raid3_worker(void *arg)
1962 {
1963 	struct g_raid3_softc *sc;
1964 	struct g_raid3_event *ep;
1965 	struct bio *bp;
1966 	int timeout;
1967 
1968 	sc = arg;
1969 	mtx_lock_spin(&sched_lock);
1970 	sched_prio(curthread, PRIBIO);
1971 	mtx_unlock_spin(&sched_lock);
1972 
1973 	sx_xlock(&sc->sc_lock);
1974 	for (;;) {
1975 		G_RAID3_DEBUG(5, "%s: Let's see...", __func__);
1976 		/*
1977 		 * First take a look at events.
1978 		 * This is important to handle events before any I/O requests.
1979 		 */
1980 		ep = g_raid3_event_get(sc);
1981 		if (ep != NULL) {
1982 			g_raid3_event_remove(sc, ep);
1983 			if ((ep->e_flags & G_RAID3_EVENT_DEVICE) != 0) {
1984 				/* Update only device status. */
1985 				G_RAID3_DEBUG(3,
1986 				    "Running event for device %s.",
1987 				    sc->sc_name);
1988 				ep->e_error = 0;
1989 				g_raid3_update_device(sc, 1);
1990 			} else {
1991 				/* Update disk status. */
1992 				G_RAID3_DEBUG(3, "Running event for disk %s.",
1993 				     g_raid3_get_diskname(ep->e_disk));
1994 				ep->e_error = g_raid3_update_disk(ep->e_disk,
1995 				    ep->e_state);
1996 				if (ep->e_error == 0)
1997 					g_raid3_update_device(sc, 0);
1998 			}
1999 			if ((ep->e_flags & G_RAID3_EVENT_DONTWAIT) != 0) {
2000 				KASSERT(ep->e_error == 0,
2001 				    ("Error cannot be handled."));
2002 				g_raid3_event_free(ep);
2003 			} else {
2004 				ep->e_flags |= G_RAID3_EVENT_DONE;
2005 				G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__,
2006 				    ep);
2007 				mtx_lock(&sc->sc_events_mtx);
2008 				wakeup(ep);
2009 				mtx_unlock(&sc->sc_events_mtx);
2010 			}
2011 			if ((sc->sc_flags &
2012 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2013 				if (g_raid3_try_destroy(sc)) {
2014 					curthread->td_pflags &= ~TDP_GEOM;
2015 					G_RAID3_DEBUG(1, "Thread exiting.");
2016 					kthread_exit(0);
2017 				}
2018 			}
2019 			G_RAID3_DEBUG(5, "%s: I'm here 1.", __func__);
2020 			continue;
2021 		}
2022 		/*
2023 		 * Check if we can mark array as CLEAN and if we can't take
2024 		 * how much seconds should we wait.
2025 		 */
2026 		timeout = g_raid3_idle(sc, -1);
2027 		/*
2028 		 * Now I/O requests.
2029 		 */
2030 		/* Get first request from the queue. */
2031 		mtx_lock(&sc->sc_queue_mtx);
2032 		bp = bioq_first(&sc->sc_queue);
2033 		if (bp == NULL) {
2034 			if ((sc->sc_flags &
2035 			    G_RAID3_DEVICE_FLAG_DESTROY) != 0) {
2036 				mtx_unlock(&sc->sc_queue_mtx);
2037 				if (g_raid3_try_destroy(sc)) {
2038 					curthread->td_pflags &= ~TDP_GEOM;
2039 					G_RAID3_DEBUG(1, "Thread exiting.");
2040 					kthread_exit(0);
2041 				}
2042 				mtx_lock(&sc->sc_queue_mtx);
2043 			}
2044 			sx_xunlock(&sc->sc_lock);
2045 			/*
2046 			 * XXX: We can miss an event here, because an event
2047 			 *      can be added without sx-device-lock and without
2048 			 *      mtx-queue-lock. Maybe I should just stop using
2049 			 *      dedicated mutex for events synchronization and
2050 			 *      stick with the queue lock?
2051 			 *      The event will hang here until next I/O request
2052 			 *      or next event is received.
2053 			 */
2054 			MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "r3:w1",
2055 			    timeout * hz);
2056 			sx_xlock(&sc->sc_lock);
2057 			G_RAID3_DEBUG(5, "%s: I'm here 4.", __func__);
2058 			continue;
2059 		}
2060 process:
2061 		bioq_remove(&sc->sc_queue, bp);
2062 		mtx_unlock(&sc->sc_queue_mtx);
2063 
2064 		if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR) != 0)
2065 			g_raid3_regular_request(bp);
2066 		else if ((bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC) != 0)
2067 			g_raid3_sync_request(bp);
2068 		else if (g_raid3_register_request(bp) != 0) {
2069 			mtx_lock(&sc->sc_queue_mtx);
2070 			bioq_insert_head(&sc->sc_queue, bp);
2071 			/*
2072 			 * We are short in memory, let see if there are finished
2073 			 * request we can free.
2074 			 */
2075 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2076 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_REGULAR)
2077 					goto process;
2078 			}
2079 			/*
2080 			 * No finished regular request, so at least keep
2081 			 * synchronization running.
2082 			 */
2083 			TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
2084 				if (bp->bio_cflags & G_RAID3_BIO_CFLAG_SYNC)
2085 					goto process;
2086 			}
2087 			sx_xunlock(&sc->sc_lock);
2088 			MSLEEP(&sc->sc_queue, &sc->sc_queue_mtx, PRIBIO | PDROP,
2089 			    "r3:lowmem", hz / 10);
2090 			sx_xlock(&sc->sc_lock);
2091 		}
2092 		G_RAID3_DEBUG(5, "%s: I'm here 9.", __func__);
2093 	}
2094 }
2095 
2096 static void
2097 g_raid3_update_idle(struct g_raid3_softc *sc, struct g_raid3_disk *disk)
2098 {
2099 
2100 	sx_assert(&sc->sc_lock, SX_LOCKED);
2101 	if (!sc->sc_idle && (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) == 0) {
2102 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as dirty.",
2103 		    g_raid3_get_diskname(disk), sc->sc_name);
2104 		disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2105 	} else if (sc->sc_idle &&
2106 	    (disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0) {
2107 		G_RAID3_DEBUG(1, "Disk %s (device %s) marked as clean.",
2108 		    g_raid3_get_diskname(disk), sc->sc_name);
2109 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2110 	}
2111 }
2112 
2113 static void
2114 g_raid3_sync_start(struct g_raid3_softc *sc)
2115 {
2116 	struct g_raid3_disk *disk;
2117 	struct g_consumer *cp;
2118 	struct bio *bp;
2119 	int error;
2120 	u_int n;
2121 
2122 	g_topology_assert_not();
2123 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2124 
2125 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2126 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2127 	    sc->sc_state));
2128 	KASSERT(sc->sc_syncdisk == NULL, ("Syncdisk is not NULL (%s, %u).",
2129 	    sc->sc_name, sc->sc_state));
2130 	disk = NULL;
2131 	for (n = 0; n < sc->sc_ndisks; n++) {
2132 		if (sc->sc_disks[n].d_state != G_RAID3_DISK_STATE_SYNCHRONIZING)
2133 			continue;
2134 		disk = &sc->sc_disks[n];
2135 		break;
2136 	}
2137 	if (disk == NULL)
2138 		return;
2139 
2140 	sx_xunlock(&sc->sc_lock);
2141 	g_topology_lock();
2142 	cp = g_new_consumer(sc->sc_sync.ds_geom);
2143 	error = g_attach(cp, sc->sc_provider);
2144 	KASSERT(error == 0,
2145 	    ("Cannot attach to %s (error=%d).", sc->sc_name, error));
2146 	error = g_access(cp, 1, 0, 0);
2147 	KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
2148 	g_topology_unlock();
2149 	sx_xlock(&sc->sc_lock);
2150 
2151 	G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
2152 	    g_raid3_get_diskname(disk));
2153 	disk->d_flags |= G_RAID3_DISK_FLAG_DIRTY;
2154 	KASSERT(disk->d_sync.ds_consumer == NULL,
2155 	    ("Sync consumer already exists (device=%s, disk=%s).",
2156 	    sc->sc_name, g_raid3_get_diskname(disk)));
2157 
2158 	disk->d_sync.ds_consumer = cp;
2159 	disk->d_sync.ds_consumer->private = disk;
2160 	disk->d_sync.ds_consumer->index = 0;
2161 	sc->sc_syncdisk = disk;
2162 
2163 	/*
2164 	 * Allocate memory for synchronization bios and initialize them.
2165 	 */
2166 	disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_raid3_syncreqs,
2167 	    M_RAID3, M_WAITOK);
2168 	for (n = 0; n < g_raid3_syncreqs; n++) {
2169 		bp = g_alloc_bio();
2170 		disk->d_sync.ds_bios[n] = bp;
2171 		bp->bio_parent = NULL;
2172 		bp->bio_cmd = BIO_READ;
2173 		bp->bio_data = malloc(MAXPHYS, M_RAID3, M_WAITOK);
2174 		bp->bio_cflags = 0;
2175 		bp->bio_offset = disk->d_sync.ds_offset * (sc->sc_ndisks - 1);
2176 		bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
2177 		disk->d_sync.ds_offset += bp->bio_length / (sc->sc_ndisks - 1);
2178 		bp->bio_done = g_raid3_sync_done;
2179 		bp->bio_from = disk->d_sync.ds_consumer;
2180 		bp->bio_to = sc->sc_provider;
2181 		bp->bio_caller1 = (void *)(uintptr_t)n;
2182 	}
2183 
2184 	/* Set the number of in-flight synchronization requests. */
2185 	disk->d_sync.ds_inflight = g_raid3_syncreqs;
2186 
2187 	/*
2188 	 * Fire off first synchronization requests.
2189 	 */
2190 	for (n = 0; n < g_raid3_syncreqs; n++) {
2191 		bp = disk->d_sync.ds_bios[n];
2192 		G_RAID3_LOGREQ(3, bp, "Sending synchronization request.");
2193 		disk->d_sync.ds_consumer->index++;
2194 		/*
2195 		 * Delay the request if it is colliding with a regular request.
2196 		 */
2197 		if (g_raid3_regular_collision(sc, bp))
2198 			g_raid3_sync_delay(sc, bp);
2199 		else
2200 			g_io_request(bp, disk->d_sync.ds_consumer);
2201 	}
2202 }
2203 
2204 /*
2205  * Stop synchronization process.
2206  * type: 0 - synchronization finished
2207  *       1 - synchronization stopped
2208  */
2209 static void
2210 g_raid3_sync_stop(struct g_raid3_softc *sc, int type)
2211 {
2212 	struct g_raid3_disk *disk;
2213 	struct g_consumer *cp;
2214 
2215 	g_topology_assert_not();
2216 	sx_assert(&sc->sc_lock, SX_LOCKED);
2217 
2218 	KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED,
2219 	    ("Device not in DEGRADED state (%s, %u).", sc->sc_name,
2220 	    sc->sc_state));
2221 	disk = sc->sc_syncdisk;
2222 	sc->sc_syncdisk = NULL;
2223 	KASSERT(disk != NULL, ("No disk was synchronized (%s).", sc->sc_name));
2224 	KASSERT(disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2225 	    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2226 	    g_raid3_disk_state2str(disk->d_state)));
2227 	if (disk->d_sync.ds_consumer == NULL)
2228 		return;
2229 
2230 	if (type == 0) {
2231 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2232 		    sc->sc_name, g_raid3_get_diskname(disk));
2233 	} else /* if (type == 1) */ {
2234 		G_RAID3_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2235 		    sc->sc_name, g_raid3_get_diskname(disk));
2236 	}
2237 	free(disk->d_sync.ds_bios, M_RAID3);
2238 	disk->d_sync.ds_bios = NULL;
2239 	cp = disk->d_sync.ds_consumer;
2240 	disk->d_sync.ds_consumer = NULL;
2241 	disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2242 	sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2243 	g_topology_lock();
2244 	g_raid3_kill_consumer(sc, cp);
2245 	g_topology_unlock();
2246 	sx_xlock(&sc->sc_lock);
2247 }
2248 
2249 static void
2250 g_raid3_launch_provider(struct g_raid3_softc *sc)
2251 {
2252 	struct g_provider *pp;
2253 
2254 	sx_assert(&sc->sc_lock, SX_LOCKED);
2255 
2256 	g_topology_lock();
2257 	pp = g_new_providerf(sc->sc_geom, "raid3/%s", sc->sc_name);
2258 	pp->mediasize = sc->sc_mediasize;
2259 	pp->sectorsize = sc->sc_sectorsize;
2260 	sc->sc_provider = pp;
2261 	g_error_provider(pp, 0);
2262 	g_topology_unlock();
2263 	G_RAID3_DEBUG(0, "Device %s: provider %s launched.", sc->sc_name,
2264 	    pp->name);
2265 	if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED)
2266 		g_raid3_sync_start(sc);
2267 }
2268 
2269 static void
2270 g_raid3_destroy_provider(struct g_raid3_softc *sc)
2271 {
2272 	struct bio *bp;
2273 
2274 	g_topology_assert_not();
2275 	KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2276 	    sc->sc_name));
2277 
2278 	g_topology_lock();
2279 	g_error_provider(sc->sc_provider, ENXIO);
2280 	mtx_lock(&sc->sc_queue_mtx);
2281 	while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2282 		bioq_remove(&sc->sc_queue, bp);
2283 		g_io_deliver(bp, ENXIO);
2284 	}
2285 	mtx_unlock(&sc->sc_queue_mtx);
2286 	G_RAID3_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2287 	    sc->sc_provider->name);
2288 	sc->sc_provider->flags |= G_PF_WITHER;
2289 	g_orphan_provider(sc->sc_provider, ENXIO);
2290 	g_topology_unlock();
2291 	sc->sc_provider = NULL;
2292 	if (sc->sc_syncdisk != NULL)
2293 		g_raid3_sync_stop(sc, 1);
2294 }
2295 
2296 static void
2297 g_raid3_go(void *arg)
2298 {
2299 	struct g_raid3_softc *sc;
2300 
2301 	sc = arg;
2302 	G_RAID3_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2303 	g_raid3_event_send(sc, 0,
2304 	    G_RAID3_EVENT_DONTWAIT | G_RAID3_EVENT_DEVICE);
2305 }
2306 
2307 static u_int
2308 g_raid3_determine_state(struct g_raid3_disk *disk)
2309 {
2310 	struct g_raid3_softc *sc;
2311 	u_int state;
2312 
2313 	sc = disk->d_softc;
2314 	if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2315 		if ((disk->d_flags &
2316 		    G_RAID3_DISK_FLAG_SYNCHRONIZING) == 0) {
2317 			/* Disk does not need synchronization. */
2318 			state = G_RAID3_DISK_STATE_ACTIVE;
2319 		} else {
2320 			if ((sc->sc_flags &
2321 			     G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2322 			    (disk->d_flags &
2323 			     G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2324 				/*
2325 				 * We can start synchronization from
2326 				 * the stored offset.
2327 				 */
2328 				state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2329 			} else {
2330 				state = G_RAID3_DISK_STATE_STALE;
2331 			}
2332 		}
2333 	} else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2334 		/*
2335 		 * Reset all synchronization data for this disk,
2336 		 * because if it even was synchronized, it was
2337 		 * synchronized to disks with different syncid.
2338 		 */
2339 		disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2340 		disk->d_sync.ds_offset = 0;
2341 		disk->d_sync.ds_offset_done = 0;
2342 		disk->d_sync.ds_syncid = sc->sc_syncid;
2343 		if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2344 		    (disk->d_flags & G_RAID3_DISK_FLAG_FORCE_SYNC) != 0) {
2345 			state = G_RAID3_DISK_STATE_SYNCHRONIZING;
2346 		} else {
2347 			state = G_RAID3_DISK_STATE_STALE;
2348 		}
2349 	} else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2350 		/*
2351 		 * Not good, NOT GOOD!
2352 		 * It means that device was started on stale disks
2353 		 * and more fresh disk just arrive.
2354 		 * If there were writes, device is fucked up, sorry.
2355 		 * I think the best choice here is don't touch
2356 		 * this disk and inform the user laudly.
2357 		 */
2358 		G_RAID3_DEBUG(0, "Device %s was started before the freshest "
2359 		    "disk (%s) arrives!! It will not be connected to the "
2360 		    "running device.", sc->sc_name,
2361 		    g_raid3_get_diskname(disk));
2362 		g_raid3_destroy_disk(disk);
2363 		state = G_RAID3_DISK_STATE_NONE;
2364 		/* Return immediately, because disk was destroyed. */
2365 		return (state);
2366 	}
2367 	G_RAID3_DEBUG(3, "State for %s disk: %s.",
2368 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(state));
2369 	return (state);
2370 }
2371 
2372 /*
2373  * Update device state.
2374  */
2375 static void
2376 g_raid3_update_device(struct g_raid3_softc *sc, boolean_t force)
2377 {
2378 	struct g_raid3_disk *disk;
2379 	u_int state;
2380 
2381 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2382 
2383 	switch (sc->sc_state) {
2384 	case G_RAID3_DEVICE_STATE_STARTING:
2385 	    {
2386 		u_int n, ndirty, ndisks, genid, syncid;
2387 
2388 		KASSERT(sc->sc_provider == NULL,
2389 		    ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2390 		/*
2391 		 * Are we ready? We are, if all disks are connected or
2392 		 * one disk is missing and 'force' is true.
2393 		 */
2394 		if (g_raid3_ndisks(sc, -1) + force == sc->sc_ndisks) {
2395 			if (!force)
2396 				callout_drain(&sc->sc_callout);
2397 		} else {
2398 			if (force) {
2399 				/*
2400 				 * Timeout expired, so destroy device.
2401 				 */
2402 				sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2403 				G_RAID3_DEBUG(1, "root_mount_rel[%u] %p",
2404 				    __LINE__, sc->sc_rootmount);
2405 				root_mount_rel(sc->sc_rootmount);
2406 				sc->sc_rootmount = NULL;
2407 			}
2408 			return;
2409 		}
2410 
2411 		/*
2412 		 * Find the biggest genid.
2413 		 */
2414 		genid = 0;
2415 		for (n = 0; n < sc->sc_ndisks; n++) {
2416 			disk = &sc->sc_disks[n];
2417 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2418 				continue;
2419 			if (disk->d_genid > genid)
2420 				genid = disk->d_genid;
2421 		}
2422 		sc->sc_genid = genid;
2423 		/*
2424 		 * Remove all disks without the biggest genid.
2425 		 */
2426 		for (n = 0; n < sc->sc_ndisks; n++) {
2427 			disk = &sc->sc_disks[n];
2428 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2429 				continue;
2430 			if (disk->d_genid < genid) {
2431 				G_RAID3_DEBUG(0,
2432 				    "Component %s (device %s) broken, skipping.",
2433 				    g_raid3_get_diskname(disk), sc->sc_name);
2434 				g_raid3_destroy_disk(disk);
2435 			}
2436 		}
2437 
2438 		/*
2439 		 * There must be at least 'sc->sc_ndisks - 1' components
2440 		 * with the same syncid and without SYNCHRONIZING flag.
2441 		 */
2442 
2443 		/*
2444 		 * Find the biggest syncid, number of valid components and
2445 		 * number of dirty components.
2446 		 */
2447 		ndirty = ndisks = syncid = 0;
2448 		for (n = 0; n < sc->sc_ndisks; n++) {
2449 			disk = &sc->sc_disks[n];
2450 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2451 				continue;
2452 			if ((disk->d_flags & G_RAID3_DISK_FLAG_DIRTY) != 0)
2453 				ndirty++;
2454 			if (disk->d_sync.ds_syncid > syncid) {
2455 				syncid = disk->d_sync.ds_syncid;
2456 				ndisks = 0;
2457 			} else if (disk->d_sync.ds_syncid < syncid) {
2458 				continue;
2459 			}
2460 			if ((disk->d_flags &
2461 			    G_RAID3_DISK_FLAG_SYNCHRONIZING) != 0) {
2462 				continue;
2463 			}
2464 			ndisks++;
2465 		}
2466 		/*
2467 		 * Do we have enough valid components?
2468 		 */
2469 		if (ndisks + 1 < sc->sc_ndisks) {
2470 			G_RAID3_DEBUG(0,
2471 			    "Device %s is broken, too few valid components.",
2472 			    sc->sc_name);
2473 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2474 			return;
2475 		}
2476 		/*
2477 		 * If there is one DIRTY component and all disks are present,
2478 		 * mark it for synchronization. If there is more than one DIRTY
2479 		 * component, mark parity component for synchronization.
2480 		 */
2481 		if (ndisks == sc->sc_ndisks && ndirty == 1) {
2482 			for (n = 0; n < sc->sc_ndisks; n++) {
2483 				disk = &sc->sc_disks[n];
2484 				if ((disk->d_flags &
2485 				    G_RAID3_DISK_FLAG_DIRTY) == 0) {
2486 					continue;
2487 				}
2488 				disk->d_flags |=
2489 				    G_RAID3_DISK_FLAG_SYNCHRONIZING;
2490 			}
2491 		} else if (ndisks == sc->sc_ndisks && ndirty > 1) {
2492 			disk = &sc->sc_disks[sc->sc_ndisks - 1];
2493 			disk->d_flags |= G_RAID3_DISK_FLAG_SYNCHRONIZING;
2494 		}
2495 
2496 		sc->sc_syncid = syncid;
2497 		if (force) {
2498 			/* Remember to bump syncid on first write. */
2499 			sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2500 		}
2501 		if (ndisks == sc->sc_ndisks)
2502 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2503 		else /* if (ndisks == sc->sc_ndisks - 1) */
2504 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2505 		G_RAID3_DEBUG(1, "Device %s state changed from %s to %s.",
2506 		    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2507 		    g_raid3_device_state2str(state));
2508 		sc->sc_state = state;
2509 		for (n = 0; n < sc->sc_ndisks; n++) {
2510 			disk = &sc->sc_disks[n];
2511 			if (disk->d_state == G_RAID3_DISK_STATE_NODISK)
2512 				continue;
2513 			state = g_raid3_determine_state(disk);
2514 			g_raid3_event_send(disk, state, G_RAID3_EVENT_DONTWAIT);
2515 			if (state == G_RAID3_DISK_STATE_STALE)
2516 				sc->sc_bump_id |= G_RAID3_BUMP_SYNCID;
2517 		}
2518 		break;
2519 	    }
2520 	case G_RAID3_DEVICE_STATE_DEGRADED:
2521 		/*
2522 		 * Genid need to be bumped immediately, so do it here.
2523 		 */
2524 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2525 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2526 			g_raid3_bump_genid(sc);
2527 		}
2528 
2529 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2530 			return;
2531 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) <
2532 		    sc->sc_ndisks - 1) {
2533 			if (sc->sc_provider != NULL)
2534 				g_raid3_destroy_provider(sc);
2535 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
2536 			return;
2537 		}
2538 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2539 		    sc->sc_ndisks) {
2540 			state = G_RAID3_DEVICE_STATE_COMPLETE;
2541 			G_RAID3_DEBUG(1,
2542 			    "Device %s state changed from %s to %s.",
2543 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2544 			    g_raid3_device_state2str(state));
2545 			sc->sc_state = state;
2546 		}
2547 		if (sc->sc_provider == NULL)
2548 			g_raid3_launch_provider(sc);
2549 		if (sc->sc_rootmount != NULL) {
2550 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2551 			    sc->sc_rootmount);
2552 			root_mount_rel(sc->sc_rootmount);
2553 			sc->sc_rootmount = NULL;
2554 		}
2555 		break;
2556 	case G_RAID3_DEVICE_STATE_COMPLETE:
2557 		/*
2558 		 * Genid need to be bumped immediately, so do it here.
2559 		 */
2560 		if ((sc->sc_bump_id & G_RAID3_BUMP_GENID) != 0) {
2561 			sc->sc_bump_id &= ~G_RAID3_BUMP_GENID;
2562 			g_raid3_bump_genid(sc);
2563 		}
2564 
2565 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NEW) > 0)
2566 			return;
2567 		KASSERT(g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) >=
2568 		    sc->sc_ndisks - 1,
2569 		    ("Too few ACTIVE components in COMPLETE state (device %s).",
2570 		    sc->sc_name));
2571 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) ==
2572 		    sc->sc_ndisks - 1) {
2573 			state = G_RAID3_DEVICE_STATE_DEGRADED;
2574 			G_RAID3_DEBUG(1,
2575 			    "Device %s state changed from %s to %s.",
2576 			    sc->sc_name, g_raid3_device_state2str(sc->sc_state),
2577 			    g_raid3_device_state2str(state));
2578 			sc->sc_state = state;
2579 		}
2580 		if (sc->sc_provider == NULL)
2581 			g_raid3_launch_provider(sc);
2582 		if (sc->sc_rootmount != NULL) {
2583 			G_RAID3_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2584 			    sc->sc_rootmount);
2585 			root_mount_rel(sc->sc_rootmount);
2586 			sc->sc_rootmount = NULL;
2587 		}
2588 		break;
2589 	default:
2590 		KASSERT(1 == 0, ("Wrong device state (%s, %s).", sc->sc_name,
2591 		    g_raid3_device_state2str(sc->sc_state)));
2592 		break;
2593 	}
2594 }
2595 
2596 /*
2597  * Update disk state and device state if needed.
2598  */
2599 #define	DISK_STATE_CHANGED()	G_RAID3_DEBUG(1,			\
2600 	"Disk %s state changed from %s to %s (device %s).",		\
2601 	g_raid3_get_diskname(disk),					\
2602 	g_raid3_disk_state2str(disk->d_state),				\
2603 	g_raid3_disk_state2str(state), sc->sc_name)
2604 static int
2605 g_raid3_update_disk(struct g_raid3_disk *disk, u_int state)
2606 {
2607 	struct g_raid3_softc *sc;
2608 
2609 	sc = disk->d_softc;
2610 	sx_assert(&sc->sc_lock, SX_XLOCKED);
2611 
2612 again:
2613 	G_RAID3_DEBUG(3, "Changing disk %s state from %s to %s.",
2614 	    g_raid3_get_diskname(disk), g_raid3_disk_state2str(disk->d_state),
2615 	    g_raid3_disk_state2str(state));
2616 	switch (state) {
2617 	case G_RAID3_DISK_STATE_NEW:
2618 		/*
2619 		 * Possible scenarios:
2620 		 * 1. New disk arrive.
2621 		 */
2622 		/* Previous state should be NONE. */
2623 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NONE,
2624 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2625 		    g_raid3_disk_state2str(disk->d_state)));
2626 		DISK_STATE_CHANGED();
2627 
2628 		disk->d_state = state;
2629 		G_RAID3_DEBUG(0, "Device %s: provider %s detected.",
2630 		    sc->sc_name, g_raid3_get_diskname(disk));
2631 		if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING)
2632 			break;
2633 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2634 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2635 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2636 		    g_raid3_device_state2str(sc->sc_state),
2637 		    g_raid3_get_diskname(disk),
2638 		    g_raid3_disk_state2str(disk->d_state)));
2639 		state = g_raid3_determine_state(disk);
2640 		if (state != G_RAID3_DISK_STATE_NONE)
2641 			goto again;
2642 		break;
2643 	case G_RAID3_DISK_STATE_ACTIVE:
2644 		/*
2645 		 * Possible scenarios:
2646 		 * 1. New disk does not need synchronization.
2647 		 * 2. Synchronization process finished successfully.
2648 		 */
2649 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2650 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2651 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2652 		    g_raid3_device_state2str(sc->sc_state),
2653 		    g_raid3_get_diskname(disk),
2654 		    g_raid3_disk_state2str(disk->d_state)));
2655 		/* Previous state should be NEW or SYNCHRONIZING. */
2656 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW ||
2657 		    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2658 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2659 		    g_raid3_disk_state2str(disk->d_state)));
2660 		DISK_STATE_CHANGED();
2661 
2662 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
2663 			disk->d_flags &= ~G_RAID3_DISK_FLAG_SYNCHRONIZING;
2664 			disk->d_flags &= ~G_RAID3_DISK_FLAG_FORCE_SYNC;
2665 			g_raid3_sync_stop(sc, 0);
2666 		}
2667 		disk->d_state = state;
2668 		disk->d_sync.ds_offset = 0;
2669 		disk->d_sync.ds_offset_done = 0;
2670 		g_raid3_update_idle(sc, disk);
2671 		g_raid3_update_metadata(disk);
2672 		G_RAID3_DEBUG(0, "Device %s: provider %s activated.",
2673 		    sc->sc_name, g_raid3_get_diskname(disk));
2674 		break;
2675 	case G_RAID3_DISK_STATE_STALE:
2676 		/*
2677 		 * Possible scenarios:
2678 		 * 1. Stale disk was connected.
2679 		 */
2680 		/* Previous state should be NEW. */
2681 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2682 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2683 		    g_raid3_disk_state2str(disk->d_state)));
2684 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2685 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2686 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2687 		    g_raid3_device_state2str(sc->sc_state),
2688 		    g_raid3_get_diskname(disk),
2689 		    g_raid3_disk_state2str(disk->d_state)));
2690 		/*
2691 		 * STALE state is only possible if device is marked
2692 		 * NOAUTOSYNC.
2693 		 */
2694 		KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_NOAUTOSYNC) != 0,
2695 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2696 		    g_raid3_device_state2str(sc->sc_state),
2697 		    g_raid3_get_diskname(disk),
2698 		    g_raid3_disk_state2str(disk->d_state)));
2699 		DISK_STATE_CHANGED();
2700 
2701 		disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2702 		disk->d_state = state;
2703 		g_raid3_update_metadata(disk);
2704 		G_RAID3_DEBUG(0, "Device %s: provider %s is stale.",
2705 		    sc->sc_name, g_raid3_get_diskname(disk));
2706 		break;
2707 	case G_RAID3_DISK_STATE_SYNCHRONIZING:
2708 		/*
2709 		 * Possible scenarios:
2710 		 * 1. Disk which needs synchronization was connected.
2711 		 */
2712 		/* Previous state should be NEW. */
2713 		KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2714 		    ("Wrong disk state (%s, %s).", g_raid3_get_diskname(disk),
2715 		    g_raid3_disk_state2str(disk->d_state)));
2716 		KASSERT(sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2717 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE,
2718 		    ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2719 		    g_raid3_device_state2str(sc->sc_state),
2720 		    g_raid3_get_diskname(disk),
2721 		    g_raid3_disk_state2str(disk->d_state)));
2722 		DISK_STATE_CHANGED();
2723 
2724 		if (disk->d_state == G_RAID3_DISK_STATE_NEW)
2725 			disk->d_flags &= ~G_RAID3_DISK_FLAG_DIRTY;
2726 		disk->d_state = state;
2727 		if (sc->sc_provider != NULL) {
2728 			g_raid3_sync_start(sc);
2729 			g_raid3_update_metadata(disk);
2730 		}
2731 		break;
2732 	case G_RAID3_DISK_STATE_DISCONNECTED:
2733 		/*
2734 		 * Possible scenarios:
2735 		 * 1. Device wasn't running yet, but disk disappear.
2736 		 * 2. Disk was active and disapppear.
2737 		 * 3. Disk disappear during synchronization process.
2738 		 */
2739 		if (sc->sc_state == G_RAID3_DEVICE_STATE_DEGRADED ||
2740 		    sc->sc_state == G_RAID3_DEVICE_STATE_COMPLETE) {
2741 			/*
2742 			 * Previous state should be ACTIVE, STALE or
2743 			 * SYNCHRONIZING.
2744 			 */
2745 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_ACTIVE ||
2746 			    disk->d_state == G_RAID3_DISK_STATE_STALE ||
2747 			    disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING,
2748 			    ("Wrong disk state (%s, %s).",
2749 			    g_raid3_get_diskname(disk),
2750 			    g_raid3_disk_state2str(disk->d_state)));
2751 		} else if (sc->sc_state == G_RAID3_DEVICE_STATE_STARTING) {
2752 			/* Previous state should be NEW. */
2753 			KASSERT(disk->d_state == G_RAID3_DISK_STATE_NEW,
2754 			    ("Wrong disk state (%s, %s).",
2755 			    g_raid3_get_diskname(disk),
2756 			    g_raid3_disk_state2str(disk->d_state)));
2757 			/*
2758 			 * Reset bumping syncid if disk disappeared in STARTING
2759 			 * state.
2760 			 */
2761 			if ((sc->sc_bump_id & G_RAID3_BUMP_SYNCID) != 0)
2762 				sc->sc_bump_id &= ~G_RAID3_BUMP_SYNCID;
2763 #ifdef	INVARIANTS
2764 		} else {
2765 			KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2766 			    sc->sc_name,
2767 			    g_raid3_device_state2str(sc->sc_state),
2768 			    g_raid3_get_diskname(disk),
2769 			    g_raid3_disk_state2str(disk->d_state)));
2770 #endif
2771 		}
2772 		DISK_STATE_CHANGED();
2773 		G_RAID3_DEBUG(0, "Device %s: provider %s disconnected.",
2774 		    sc->sc_name, g_raid3_get_diskname(disk));
2775 
2776 		g_raid3_destroy_disk(disk);
2777 		break;
2778 	default:
2779 		KASSERT(1 == 0, ("Unknown state (%u).", state));
2780 		break;
2781 	}
2782 	return (0);
2783 }
2784 #undef	DISK_STATE_CHANGED
2785 
2786 int
2787 g_raid3_read_metadata(struct g_consumer *cp, struct g_raid3_metadata *md)
2788 {
2789 	struct g_provider *pp;
2790 	u_char *buf;
2791 	int error;
2792 
2793 	g_topology_assert();
2794 
2795 	error = g_access(cp, 1, 0, 0);
2796 	if (error != 0)
2797 		return (error);
2798 	pp = cp->provider;
2799 	g_topology_unlock();
2800 	/* Metadata are stored on last sector. */
2801 	buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2802 	    &error);
2803 	g_topology_lock();
2804 	g_access(cp, -1, 0, 0);
2805 	if (buf == NULL) {
2806 		G_RAID3_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2807 		    cp->provider->name, error);
2808 		return (error);
2809 	}
2810 
2811 	/* Decode metadata. */
2812 	error = raid3_metadata_decode(buf, md);
2813 	g_free(buf);
2814 	if (strcmp(md->md_magic, G_RAID3_MAGIC) != 0)
2815 		return (EINVAL);
2816 	if (md->md_version > G_RAID3_VERSION) {
2817 		G_RAID3_DEBUG(0,
2818 		    "Kernel module is too old to handle metadata from %s.",
2819 		    cp->provider->name);
2820 		return (EINVAL);
2821 	}
2822 	if (error != 0) {
2823 		G_RAID3_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2824 		    cp->provider->name);
2825 		return (error);
2826 	}
2827 
2828 	return (0);
2829 }
2830 
2831 static int
2832 g_raid3_check_metadata(struct g_raid3_softc *sc, struct g_provider *pp,
2833     struct g_raid3_metadata *md)
2834 {
2835 
2836 	if (md->md_no >= sc->sc_ndisks) {
2837 		G_RAID3_DEBUG(1, "Invalid disk %s number (no=%u), skipping.",
2838 		    pp->name, md->md_no);
2839 		return (EINVAL);
2840 	}
2841 	if (sc->sc_disks[md->md_no].d_state != G_RAID3_DISK_STATE_NODISK) {
2842 		G_RAID3_DEBUG(1, "Disk %s (no=%u) already exists, skipping.",
2843 		    pp->name, md->md_no);
2844 		return (EEXIST);
2845 	}
2846 	if (md->md_all != sc->sc_ndisks) {
2847 		G_RAID3_DEBUG(1,
2848 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2849 		    "md_all", pp->name, sc->sc_name);
2850 		return (EINVAL);
2851 	}
2852 	if (md->md_mediasize != sc->sc_mediasize) {
2853 		G_RAID3_DEBUG(1,
2854 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2855 		    "md_mediasize", pp->name, sc->sc_name);
2856 		return (EINVAL);
2857 	}
2858 	if ((md->md_mediasize % (sc->sc_ndisks - 1)) != 0) {
2859 		G_RAID3_DEBUG(1,
2860 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2861 		    "md_mediasize", pp->name, sc->sc_name);
2862 		return (EINVAL);
2863 	}
2864 	if ((sc->sc_mediasize / (sc->sc_ndisks - 1)) > pp->mediasize) {
2865 		G_RAID3_DEBUG(1,
2866 		    "Invalid size of disk %s (device %s), skipping.", pp->name,
2867 		    sc->sc_name);
2868 		return (EINVAL);
2869 	}
2870 	if ((md->md_sectorsize / pp->sectorsize) < sc->sc_ndisks - 1) {
2871 		G_RAID3_DEBUG(1,
2872 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2873 		    "md_sectorsize", pp->name, sc->sc_name);
2874 		return (EINVAL);
2875 	}
2876 	if (md->md_sectorsize != sc->sc_sectorsize) {
2877 		G_RAID3_DEBUG(1,
2878 		    "Invalid '%s' field on disk %s (device %s), skipping.",
2879 		    "md_sectorsize", pp->name, sc->sc_name);
2880 		return (EINVAL);
2881 	}
2882 	if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2883 		G_RAID3_DEBUG(1,
2884 		    "Invalid sector size of disk %s (device %s), skipping.",
2885 		    pp->name, sc->sc_name);
2886 		return (EINVAL);
2887 	}
2888 	if ((md->md_mflags & ~G_RAID3_DEVICE_FLAG_MASK) != 0) {
2889 		G_RAID3_DEBUG(1,
2890 		    "Invalid device flags on disk %s (device %s), skipping.",
2891 		    pp->name, sc->sc_name);
2892 		return (EINVAL);
2893 	}
2894 	if ((md->md_mflags & G_RAID3_DEVICE_FLAG_VERIFY) != 0 &&
2895 	    (md->md_mflags & G_RAID3_DEVICE_FLAG_ROUND_ROBIN) != 0) {
2896 		/*
2897 		 * VERIFY and ROUND-ROBIN options are mutally exclusive.
2898 		 */
2899 		G_RAID3_DEBUG(1, "Both VERIFY and ROUND-ROBIN flags exist on "
2900 		    "disk %s (device %s), skipping.", pp->name, sc->sc_name);
2901 		return (EINVAL);
2902 	}
2903 	if ((md->md_dflags & ~G_RAID3_DISK_FLAG_MASK) != 0) {
2904 		G_RAID3_DEBUG(1,
2905 		    "Invalid disk flags on disk %s (device %s), skipping.",
2906 		    pp->name, sc->sc_name);
2907 		return (EINVAL);
2908 	}
2909 	return (0);
2910 }
2911 
2912 int
2913 g_raid3_add_disk(struct g_raid3_softc *sc, struct g_provider *pp,
2914     struct g_raid3_metadata *md)
2915 {
2916 	struct g_raid3_disk *disk;
2917 	int error;
2918 
2919 	g_topology_assert_not();
2920 	G_RAID3_DEBUG(2, "Adding disk %s.", pp->name);
2921 
2922 	error = g_raid3_check_metadata(sc, pp, md);
2923 	if (error != 0)
2924 		return (error);
2925 	if (sc->sc_state != G_RAID3_DEVICE_STATE_STARTING &&
2926 	    md->md_genid < sc->sc_genid) {
2927 		G_RAID3_DEBUG(0, "Component %s (device %s) broken, skipping.",
2928 		    pp->name, sc->sc_name);
2929 		return (EINVAL);
2930 	}
2931 	disk = g_raid3_init_disk(sc, pp, md, &error);
2932 	if (disk == NULL)
2933 		return (error);
2934 	error = g_raid3_event_send(disk, G_RAID3_DISK_STATE_NEW,
2935 	    G_RAID3_EVENT_WAIT);
2936 	if (error != 0)
2937 		return (error);
2938 	if (md->md_version < G_RAID3_VERSION) {
2939 		G_RAID3_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
2940 		    pp->name, md->md_version, G_RAID3_VERSION);
2941 		g_raid3_update_metadata(disk);
2942 	}
2943 	return (0);
2944 }
2945 
2946 static void
2947 g_raid3_destroy_delayed(void *arg, int flag)
2948 {
2949 	struct g_raid3_softc *sc;
2950 	int error;
2951 
2952 	if (flag == EV_CANCEL) {
2953 		G_RAID3_DEBUG(1, "Destroying canceled.");
2954 		return;
2955 	}
2956 	sc = arg;
2957 	g_topology_unlock();
2958 	sx_xlock(&sc->sc_lock);
2959 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) == 0,
2960 	    ("DESTROY flag set on %s.", sc->sc_name));
2961 	KASSERT((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0,
2962 	    ("DESTROYING flag not set on %s.", sc->sc_name));
2963 	G_RAID3_DEBUG(0, "Destroying %s (delayed).", sc->sc_name);
2964 	error = g_raid3_destroy(sc, G_RAID3_DESTROY_SOFT);
2965 	if (error != 0) {
2966 		G_RAID3_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
2967 		sx_xunlock(&sc->sc_lock);
2968 	}
2969 	g_topology_lock();
2970 }
2971 
2972 static int
2973 g_raid3_access(struct g_provider *pp, int acr, int acw, int ace)
2974 {
2975 	struct g_raid3_softc *sc;
2976 	int dcr, dcw, dce, error = 0;
2977 
2978 	g_topology_assert();
2979 	G_RAID3_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
2980 	    acw, ace);
2981 
2982 	sc = pp->geom->softc;
2983 	if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
2984 		return (0);
2985 	KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
2986 
2987 	dcr = pp->acr + acr;
2988 	dcw = pp->acw + acw;
2989 	dce = pp->ace + ace;
2990 
2991 	g_topology_unlock();
2992 	sx_xlock(&sc->sc_lock);
2993 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROY) != 0 ||
2994 	    g_raid3_ndisks(sc, G_RAID3_DISK_STATE_ACTIVE) < sc->sc_ndisks - 1) {
2995 		if (acr > 0 || acw > 0 || ace > 0)
2996 			error = ENXIO;
2997 		goto end;
2998 	}
2999 	if (dcw == 0 && !sc->sc_idle)
3000 		g_raid3_idle(sc, dcw);
3001 	if ((sc->sc_flags & G_RAID3_DEVICE_FLAG_DESTROYING) != 0) {
3002 		if (acr > 0 || acw > 0 || ace > 0) {
3003 			error = ENXIO;
3004 			goto end;
3005 		}
3006 		if (dcr == 0 && dcw == 0 && dce == 0) {
3007 			g_post_event(g_raid3_destroy_delayed, sc, M_WAITOK,
3008 			    sc, NULL);
3009 		}
3010 	}
3011 end:
3012 	sx_xunlock(&sc->sc_lock);
3013 	g_topology_lock();
3014 	return (error);
3015 }
3016 
3017 static struct g_geom *
3018 g_raid3_create(struct g_class *mp, const struct g_raid3_metadata *md)
3019 {
3020 	struct g_raid3_softc *sc;
3021 	struct g_geom *gp;
3022 	int error, timeout;
3023 	u_int n;
3024 
3025 	g_topology_assert();
3026 	G_RAID3_DEBUG(1, "Creating device %s (id=%u).", md->md_name, md->md_id);
3027 
3028 	/* One disk is minimum. */
3029 	if (md->md_all < 1)
3030 		return (NULL);
3031 	/*
3032 	 * Action geom.
3033 	 */
3034 	gp = g_new_geomf(mp, "%s", md->md_name);
3035 	sc = malloc(sizeof(*sc), M_RAID3, M_WAITOK | M_ZERO);
3036 	sc->sc_disks = malloc(sizeof(struct g_raid3_disk) * md->md_all, M_RAID3,
3037 	    M_WAITOK | M_ZERO);
3038 	gp->start = g_raid3_start;
3039 	gp->orphan = g_raid3_orphan;
3040 	gp->access = g_raid3_access;
3041 	gp->dumpconf = g_raid3_dumpconf;
3042 
3043 	sc->sc_id = md->md_id;
3044 	sc->sc_mediasize = md->md_mediasize;
3045 	sc->sc_sectorsize = md->md_sectorsize;
3046 	sc->sc_ndisks = md->md_all;
3047 	sc->sc_round_robin = 0;
3048 	sc->sc_flags = md->md_mflags;
3049 	sc->sc_bump_id = 0;
3050 	sc->sc_idle = 1;
3051 	sc->sc_last_write = time_uptime;
3052 	sc->sc_writes = 0;
3053 	for (n = 0; n < sc->sc_ndisks; n++) {
3054 		sc->sc_disks[n].d_softc = sc;
3055 		sc->sc_disks[n].d_no = n;
3056 		sc->sc_disks[n].d_state = G_RAID3_DISK_STATE_NODISK;
3057 	}
3058 	sx_init(&sc->sc_lock, "graid3:lock");
3059 	bioq_init(&sc->sc_queue);
3060 	mtx_init(&sc->sc_queue_mtx, "graid3:queue", NULL, MTX_DEF);
3061 	bioq_init(&sc->sc_regular_delayed);
3062 	bioq_init(&sc->sc_inflight);
3063 	bioq_init(&sc->sc_sync_delayed);
3064 	TAILQ_INIT(&sc->sc_events);
3065 	mtx_init(&sc->sc_events_mtx, "graid3:events", NULL, MTX_DEF);
3066 	callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
3067 	sc->sc_state = G_RAID3_DEVICE_STATE_STARTING;
3068 	gp->softc = sc;
3069 	sc->sc_geom = gp;
3070 	sc->sc_provider = NULL;
3071 	/*
3072 	 * Synchronization geom.
3073 	 */
3074 	gp = g_new_geomf(mp, "%s.sync", md->md_name);
3075 	gp->softc = sc;
3076 	gp->orphan = g_raid3_orphan;
3077 	sc->sc_sync.ds_geom = gp;
3078 
3079 	if (!g_raid3_use_malloc) {
3080 		sc->sc_zones[G_RAID3_ZONE_64K].sz_zone = uma_zcreate("gr3:64k",
3081 		    65536, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3082 		    UMA_ALIGN_PTR, 0);
3083 		sc->sc_zones[G_RAID3_ZONE_64K].sz_inuse = 0;
3084 		sc->sc_zones[G_RAID3_ZONE_64K].sz_max = g_raid3_n64k;
3085 		sc->sc_zones[G_RAID3_ZONE_64K].sz_requested =
3086 		    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed = 0;
3087 		sc->sc_zones[G_RAID3_ZONE_16K].sz_zone = uma_zcreate("gr3:16k",
3088 		    16384, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3089 		    UMA_ALIGN_PTR, 0);
3090 		sc->sc_zones[G_RAID3_ZONE_16K].sz_inuse = 0;
3091 		sc->sc_zones[G_RAID3_ZONE_16K].sz_max = g_raid3_n16k;
3092 		sc->sc_zones[G_RAID3_ZONE_16K].sz_requested =
3093 		    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed = 0;
3094 		sc->sc_zones[G_RAID3_ZONE_4K].sz_zone = uma_zcreate("gr3:4k",
3095 		    4096, g_raid3_uma_ctor, g_raid3_uma_dtor, NULL, NULL,
3096 		    UMA_ALIGN_PTR, 0);
3097 		sc->sc_zones[G_RAID3_ZONE_4K].sz_inuse = 0;
3098 		sc->sc_zones[G_RAID3_ZONE_4K].sz_max = g_raid3_n4k;
3099 		sc->sc_zones[G_RAID3_ZONE_4K].sz_requested =
3100 		    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed = 0;
3101 	}
3102 
3103 	error = kthread_create(g_raid3_worker, sc, &sc->sc_worker, 0, 0,
3104 	    "g_raid3 %s", md->md_name);
3105 	if (error != 0) {
3106 		G_RAID3_DEBUG(1, "Cannot create kernel thread for %s.",
3107 		    sc->sc_name);
3108 		if (!g_raid3_use_malloc) {
3109 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_64K].sz_zone);
3110 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_16K].sz_zone);
3111 			uma_zdestroy(sc->sc_zones[G_RAID3_ZONE_4K].sz_zone);
3112 		}
3113 		g_destroy_geom(sc->sc_sync.ds_geom);
3114 		mtx_destroy(&sc->sc_events_mtx);
3115 		mtx_destroy(&sc->sc_queue_mtx);
3116 		sx_destroy(&sc->sc_lock);
3117 		g_destroy_geom(sc->sc_geom);
3118 		free(sc->sc_disks, M_RAID3);
3119 		free(sc, M_RAID3);
3120 		return (NULL);
3121 	}
3122 
3123 	G_RAID3_DEBUG(0, "Device %s created (id=%u).", sc->sc_name, sc->sc_id);
3124 
3125 	sc->sc_rootmount = root_mount_hold("GRAID3");
3126 	G_RAID3_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3127 
3128 	/*
3129 	 * Run timeout.
3130 	 */
3131 	timeout = atomic_load_acq_int(&g_raid3_timeout);
3132 	callout_reset(&sc->sc_callout, timeout * hz, g_raid3_go, sc);
3133 	return (sc->sc_geom);
3134 }
3135 
3136 int
3137 g_raid3_destroy(struct g_raid3_softc *sc, int how)
3138 {
3139 	struct g_provider *pp;
3140 
3141 	g_topology_assert_not();
3142 	if (sc == NULL)
3143 		return (ENXIO);
3144 	sx_assert(&sc->sc_lock, SX_XLOCKED);
3145 
3146 	pp = sc->sc_provider;
3147 	if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
3148 		switch (how) {
3149 		case G_RAID3_DESTROY_SOFT:
3150 			G_RAID3_DEBUG(1,
3151 			    "Device %s is still open (r%dw%de%d).", pp->name,
3152 			    pp->acr, pp->acw, pp->ace);
3153 			return (EBUSY);
3154 		case G_RAID3_DESTROY_DELAYED:
3155 			G_RAID3_DEBUG(1,
3156 			    "Device %s will be destroyed on last close.",
3157 			    pp->name);
3158 			if (sc->sc_syncdisk != NULL)
3159 				g_raid3_sync_stop(sc, 1);
3160 			sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROYING;
3161 			return (EBUSY);
3162 		case G_RAID3_DESTROY_HARD:
3163 			G_RAID3_DEBUG(1, "Device %s is still open, so it "
3164 			    "can't be definitely removed.", pp->name);
3165 			break;
3166 		}
3167 	}
3168 
3169 	g_topology_lock();
3170 	if (sc->sc_geom->softc == NULL) {
3171 		g_topology_unlock();
3172 		return (0);
3173 	}
3174 	sc->sc_geom->softc = NULL;
3175 	sc->sc_sync.ds_geom->softc = NULL;
3176 	g_topology_unlock();
3177 
3178 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_DESTROY;
3179 	sc->sc_flags |= G_RAID3_DEVICE_FLAG_WAIT;
3180 	G_RAID3_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3181 	sx_xunlock(&sc->sc_lock);
3182 	mtx_lock(&sc->sc_queue_mtx);
3183 	wakeup(sc);
3184 	wakeup(&sc->sc_queue);
3185 	mtx_unlock(&sc->sc_queue_mtx);
3186 	G_RAID3_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3187 	while (sc->sc_worker != NULL)
3188 		tsleep(&sc->sc_worker, PRIBIO, "r3:destroy", hz / 5);
3189 	G_RAID3_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3190 	sx_xlock(&sc->sc_lock);
3191 	g_raid3_destroy_device(sc);
3192 	free(sc->sc_disks, M_RAID3);
3193 	free(sc, M_RAID3);
3194 	return (0);
3195 }
3196 
3197 static void
3198 g_raid3_taste_orphan(struct g_consumer *cp)
3199 {
3200 
3201 	KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3202 	    cp->provider->name));
3203 }
3204 
3205 static struct g_geom *
3206 g_raid3_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3207 {
3208 	struct g_raid3_metadata md;
3209 	struct g_raid3_softc *sc;
3210 	struct g_consumer *cp;
3211 	struct g_geom *gp;
3212 	int error;
3213 
3214 	g_topology_assert();
3215 	g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3216 	G_RAID3_DEBUG(2, "Tasting %s.", pp->name);
3217 
3218 	gp = g_new_geomf(mp, "raid3:taste");
3219 	/* This orphan function should be never called. */
3220 	gp->orphan = g_raid3_taste_orphan;
3221 	cp = g_new_consumer(gp);
3222 	g_attach(cp, pp);
3223 	error = g_raid3_read_metadata(cp, &md);
3224 	g_detach(cp);
3225 	g_destroy_consumer(cp);
3226 	g_destroy_geom(gp);
3227 	if (error != 0)
3228 		return (NULL);
3229 	gp = NULL;
3230 
3231 	if (md.md_provider[0] != '\0' && strcmp(md.md_provider, pp->name) != 0)
3232 		return (NULL);
3233 	if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3234 		return (NULL);
3235 	if (g_raid3_debug >= 2)
3236 		raid3_metadata_dump(&md);
3237 
3238 	/*
3239 	 * Let's check if device already exists.
3240 	 */
3241 	sc = NULL;
3242 	LIST_FOREACH(gp, &mp->geom, geom) {
3243 		sc = gp->softc;
3244 		if (sc == NULL)
3245 			continue;
3246 		if (sc->sc_sync.ds_geom == gp)
3247 			continue;
3248 		if (strcmp(md.md_name, sc->sc_name) != 0)
3249 			continue;
3250 		if (md.md_id != sc->sc_id) {
3251 			G_RAID3_DEBUG(0, "Device %s already configured.",
3252 			    sc->sc_name);
3253 			return (NULL);
3254 		}
3255 		break;
3256 	}
3257 	if (gp == NULL) {
3258 		gp = g_raid3_create(mp, &md);
3259 		if (gp == NULL) {
3260 			G_RAID3_DEBUG(0, "Cannot create device %s.",
3261 			    md.md_name);
3262 			return (NULL);
3263 		}
3264 		sc = gp->softc;
3265 	}
3266 	G_RAID3_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3267 	g_topology_unlock();
3268 	sx_xlock(&sc->sc_lock);
3269 	error = g_raid3_add_disk(sc, pp, &md);
3270 	if (error != 0) {
3271 		G_RAID3_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3272 		    pp->name, gp->name, error);
3273 		if (g_raid3_ndisks(sc, G_RAID3_DISK_STATE_NODISK) ==
3274 		    sc->sc_ndisks) {
3275 			g_cancel_event(sc);
3276 			g_raid3_destroy(sc, G_RAID3_DESTROY_HARD);
3277 			g_topology_lock();
3278 			return (NULL);
3279 		}
3280 		gp = NULL;
3281 	}
3282 	sx_xunlock(&sc->sc_lock);
3283 	g_topology_lock();
3284 	return (gp);
3285 }
3286 
3287 static int
3288 g_raid3_destroy_geom(struct gctl_req *req __unused, struct g_class *mp __unused,
3289     struct g_geom *gp)
3290 {
3291 	struct g_raid3_softc *sc;
3292 	int error;
3293 
3294 	g_topology_unlock();
3295 	sc = gp->softc;
3296 	sx_xlock(&sc->sc_lock);
3297 	g_cancel_event(sc);
3298 	error = g_raid3_destroy(gp->softc, G_RAID3_DESTROY_SOFT);
3299 	if (error != 0)
3300 		sx_xunlock(&sc->sc_lock);
3301 	g_topology_lock();
3302 	return (error);
3303 }
3304 
3305 static void
3306 g_raid3_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3307     struct g_consumer *cp, struct g_provider *pp)
3308 {
3309 	struct g_raid3_softc *sc;
3310 
3311 	g_topology_assert();
3312 
3313 	sc = gp->softc;
3314 	if (sc == NULL)
3315 		return;
3316 	/* Skip synchronization geom. */
3317 	if (gp == sc->sc_sync.ds_geom)
3318 		return;
3319 	if (pp != NULL) {
3320 		/* Nothing here. */
3321 	} else if (cp != NULL) {
3322 		struct g_raid3_disk *disk;
3323 
3324 		disk = cp->private;
3325 		if (disk == NULL)
3326 			return;
3327 		g_topology_unlock();
3328 		sx_xlock(&sc->sc_lock);
3329 		sbuf_printf(sb, "%s<Type>", indent);
3330 		if (disk->d_no == sc->sc_ndisks - 1)
3331 			sbuf_printf(sb, "PARITY");
3332 		else
3333 			sbuf_printf(sb, "DATA");
3334 		sbuf_printf(sb, "</Type>\n");
3335 		sbuf_printf(sb, "%s<Number>%u</Number>\n", indent,
3336 		    (u_int)disk->d_no);
3337 		if (disk->d_state == G_RAID3_DISK_STATE_SYNCHRONIZING) {
3338 			sbuf_printf(sb, "%s<Synchronized>", indent);
3339 			if (disk->d_sync.ds_offset == 0)
3340 				sbuf_printf(sb, "0%%");
3341 			else {
3342 				sbuf_printf(sb, "%u%%",
3343 				    (u_int)((disk->d_sync.ds_offset * 100) /
3344 				    (sc->sc_mediasize / (sc->sc_ndisks - 1))));
3345 			}
3346 			sbuf_printf(sb, "</Synchronized>\n");
3347 		}
3348 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3349 		    disk->d_sync.ds_syncid);
3350 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, disk->d_genid);
3351 		sbuf_printf(sb, "%s<Flags>", indent);
3352 		if (disk->d_flags == 0)
3353 			sbuf_printf(sb, "NONE");
3354 		else {
3355 			int first = 1;
3356 
3357 #define	ADD_FLAG(flag, name)	do {					\
3358 	if ((disk->d_flags & (flag)) != 0) {				\
3359 		if (!first)						\
3360 			sbuf_printf(sb, ", ");				\
3361 		else							\
3362 			first = 0;					\
3363 		sbuf_printf(sb, name);					\
3364 	}								\
3365 } while (0)
3366 			ADD_FLAG(G_RAID3_DISK_FLAG_DIRTY, "DIRTY");
3367 			ADD_FLAG(G_RAID3_DISK_FLAG_HARDCODED, "HARDCODED");
3368 			ADD_FLAG(G_RAID3_DISK_FLAG_SYNCHRONIZING,
3369 			    "SYNCHRONIZING");
3370 			ADD_FLAG(G_RAID3_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3371 			ADD_FLAG(G_RAID3_DISK_FLAG_BROKEN, "BROKEN");
3372 #undef	ADD_FLAG
3373 		}
3374 		sbuf_printf(sb, "</Flags>\n");
3375 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3376 		    g_raid3_disk_state2str(disk->d_state));
3377 		sx_xunlock(&sc->sc_lock);
3378 		g_topology_lock();
3379 	} else {
3380 		g_topology_unlock();
3381 		sx_xlock(&sc->sc_lock);
3382 		if (!g_raid3_use_malloc) {
3383 			sbuf_printf(sb,
3384 			    "%s<Zone4kRequested>%u</Zone4kRequested>\n", indent,
3385 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_requested);
3386 			sbuf_printf(sb,
3387 			    "%s<Zone4kFailed>%u</Zone4kFailed>\n", indent,
3388 			    sc->sc_zones[G_RAID3_ZONE_4K].sz_failed);
3389 			sbuf_printf(sb,
3390 			    "%s<Zone16kRequested>%u</Zone16kRequested>\n", indent,
3391 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_requested);
3392 			sbuf_printf(sb,
3393 			    "%s<Zone16kFailed>%u</Zone16kFailed>\n", indent,
3394 			    sc->sc_zones[G_RAID3_ZONE_16K].sz_failed);
3395 			sbuf_printf(sb,
3396 			    "%s<Zone64kRequested>%u</Zone64kRequested>\n", indent,
3397 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_requested);
3398 			sbuf_printf(sb,
3399 			    "%s<Zone64kFailed>%u</Zone64kFailed>\n", indent,
3400 			    sc->sc_zones[G_RAID3_ZONE_64K].sz_failed);
3401 		}
3402 		sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3403 		sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3404 		sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3405 		sbuf_printf(sb, "%s<Flags>", indent);
3406 		if (sc->sc_flags == 0)
3407 			sbuf_printf(sb, "NONE");
3408 		else {
3409 			int first = 1;
3410 
3411 #define	ADD_FLAG(flag, name)	do {					\
3412 	if ((sc->sc_flags & (flag)) != 0) {				\
3413 		if (!first)						\
3414 			sbuf_printf(sb, ", ");				\
3415 		else							\
3416 			first = 0;					\
3417 		sbuf_printf(sb, name);					\
3418 	}								\
3419 } while (0)
3420 			ADD_FLAG(G_RAID3_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3421 			ADD_FLAG(G_RAID3_DEVICE_FLAG_ROUND_ROBIN,
3422 			    "ROUND-ROBIN");
3423 			ADD_FLAG(G_RAID3_DEVICE_FLAG_VERIFY, "VERIFY");
3424 #undef	ADD_FLAG
3425 		}
3426 		sbuf_printf(sb, "</Flags>\n");
3427 		sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3428 		    sc->sc_ndisks);
3429 		sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3430 		    g_raid3_device_state2str(sc->sc_state));
3431 		sx_xunlock(&sc->sc_lock);
3432 		g_topology_lock();
3433 	}
3434 }
3435 
3436 static void
3437 g_raid3_shutdown_pre_sync(void *arg, int howto)
3438 {
3439 	struct g_class *mp;
3440 	struct g_geom *gp, *gp2;
3441 	struct g_raid3_softc *sc;
3442 	int error;
3443 
3444 	mp = arg;
3445 	DROP_GIANT();
3446 	g_topology_lock();
3447 	LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3448 		if ((sc = gp->softc) == NULL)
3449 			continue;
3450 		/* Skip synchronization geom. */
3451 		if (gp == sc->sc_sync.ds_geom)
3452 			continue;
3453 		g_topology_unlock();
3454 		sx_xlock(&sc->sc_lock);
3455 		g_cancel_event(sc);
3456 		error = g_raid3_destroy(sc, G_RAID3_DESTROY_DELAYED);
3457 		if (error != 0)
3458 			sx_xunlock(&sc->sc_lock);
3459 		g_topology_lock();
3460 	}
3461 	g_topology_unlock();
3462 	PICKUP_GIANT();
3463 }
3464 
3465 static void
3466 g_raid3_init(struct g_class *mp)
3467 {
3468 
3469 	g_raid3_pre_sync = EVENTHANDLER_REGISTER(shutdown_pre_sync,
3470 	    g_raid3_shutdown_pre_sync, mp, SHUTDOWN_PRI_FIRST);
3471 	if (g_raid3_pre_sync == NULL)
3472 		G_RAID3_DEBUG(0, "Warning! Cannot register shutdown event.");
3473 }
3474 
3475 static void
3476 g_raid3_fini(struct g_class *mp)
3477 {
3478 
3479 	if (g_raid3_pre_sync != NULL)
3480 		EVENTHANDLER_DEREGISTER(shutdown_pre_sync, g_raid3_pre_sync);
3481 }
3482 
3483 DECLARE_GEOM_CLASS(g_raid3_class, g_raid3);
3484