xref: /freebsd/sys/geom/raid/tr_raid1.c (revision 2e3f49888ec8851bafb22011533217487764fdb0)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2010 Alexander Motin <mav@FreeBSD.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/param.h>
30 #include <sys/bio.h>
31 #include <sys/endian.h>
32 #include <sys/kernel.h>
33 #include <sys/kobj.h>
34 #include <sys/limits.h>
35 #include <sys/lock.h>
36 #include <sys/malloc.h>
37 #include <sys/mutex.h>
38 #include <sys/sysctl.h>
39 #include <sys/systm.h>
40 #include <geom/geom.h>
41 #include <geom/geom_dbg.h>
42 #include "geom/raid/g_raid.h"
43 #include "g_raid_tr_if.h"
44 
45 SYSCTL_DECL(_kern_geom_raid_raid1);
46 
47 #define RAID1_REBUILD_SLAB	(1 << 20) /* One transation in a rebuild */
48 static int g_raid1_rebuild_slab = RAID1_REBUILD_SLAB;
49 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_slab_size, CTLFLAG_RWTUN,
50     &g_raid1_rebuild_slab, 0,
51     "Amount of the disk to rebuild each read/write cycle of the rebuild.");
52 
53 #define RAID1_REBUILD_FAIR_IO 20 /* use 1/x of the available I/O */
54 static int g_raid1_rebuild_fair_io = RAID1_REBUILD_FAIR_IO;
55 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_fair_io, CTLFLAG_RWTUN,
56     &g_raid1_rebuild_fair_io, 0,
57     "Fraction of the I/O bandwidth to use when disk busy for rebuild.");
58 
59 #define RAID1_REBUILD_CLUSTER_IDLE 100
60 static int g_raid1_rebuild_cluster_idle = RAID1_REBUILD_CLUSTER_IDLE;
61 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_cluster_idle, CTLFLAG_RWTUN,
62     &g_raid1_rebuild_cluster_idle, 0,
63     "Number of slabs to do each time we trigger a rebuild cycle");
64 
65 #define RAID1_REBUILD_META_UPDATE 1024 /* update meta data every 1GB or so */
66 static int g_raid1_rebuild_meta_update = RAID1_REBUILD_META_UPDATE;
67 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_meta_update, CTLFLAG_RWTUN,
68     &g_raid1_rebuild_meta_update, 0,
69     "When to update the meta data.");
70 
71 static MALLOC_DEFINE(M_TR_RAID1, "tr_raid1_data", "GEOM_RAID RAID1 data");
72 
73 #define TR_RAID1_NONE 0
74 #define TR_RAID1_REBUILD 1
75 #define TR_RAID1_RESYNC 2
76 
77 #define TR_RAID1_F_DOING_SOME	0x1
78 #define TR_RAID1_F_LOCKED	0x2
79 #define TR_RAID1_F_ABORT	0x4
80 
81 struct g_raid_tr_raid1_object {
82 	struct g_raid_tr_object	 trso_base;
83 	int			 trso_starting;
84 	int			 trso_stopping;
85 	int			 trso_type;
86 	int			 trso_recover_slabs; /* slabs before rest */
87 	int			 trso_fair_io;
88 	int			 trso_meta_update;
89 	int			 trso_flags;
90 	struct g_raid_subdisk	*trso_failed_sd; /* like per volume */
91 	void			*trso_buffer;	 /* Buffer space */
92 	struct bio		 trso_bio;
93 };
94 
95 static g_raid_tr_taste_t g_raid_tr_taste_raid1;
96 static g_raid_tr_event_t g_raid_tr_event_raid1;
97 static g_raid_tr_start_t g_raid_tr_start_raid1;
98 static g_raid_tr_stop_t g_raid_tr_stop_raid1;
99 static g_raid_tr_iostart_t g_raid_tr_iostart_raid1;
100 static g_raid_tr_iodone_t g_raid_tr_iodone_raid1;
101 static g_raid_tr_kerneldump_t g_raid_tr_kerneldump_raid1;
102 static g_raid_tr_locked_t g_raid_tr_locked_raid1;
103 static g_raid_tr_idle_t g_raid_tr_idle_raid1;
104 static g_raid_tr_free_t g_raid_tr_free_raid1;
105 
106 static kobj_method_t g_raid_tr_raid1_methods[] = {
107 	KOBJMETHOD(g_raid_tr_taste,	g_raid_tr_taste_raid1),
108 	KOBJMETHOD(g_raid_tr_event,	g_raid_tr_event_raid1),
109 	KOBJMETHOD(g_raid_tr_start,	g_raid_tr_start_raid1),
110 	KOBJMETHOD(g_raid_tr_stop,	g_raid_tr_stop_raid1),
111 	KOBJMETHOD(g_raid_tr_iostart,	g_raid_tr_iostart_raid1),
112 	KOBJMETHOD(g_raid_tr_iodone,	g_raid_tr_iodone_raid1),
113 	KOBJMETHOD(g_raid_tr_kerneldump, g_raid_tr_kerneldump_raid1),
114 	KOBJMETHOD(g_raid_tr_locked,	g_raid_tr_locked_raid1),
115 	KOBJMETHOD(g_raid_tr_idle,	g_raid_tr_idle_raid1),
116 	KOBJMETHOD(g_raid_tr_free,	g_raid_tr_free_raid1),
117 	{ 0, 0 }
118 };
119 
120 static struct g_raid_tr_class g_raid_tr_raid1_class = {
121 	"RAID1",
122 	g_raid_tr_raid1_methods,
123 	sizeof(struct g_raid_tr_raid1_object),
124 	.trc_enable = 1,
125 	.trc_priority = 100,
126 	.trc_accept_unmapped = 1
127 };
128 
129 static void g_raid_tr_raid1_rebuild_abort(struct g_raid_tr_object *tr);
130 static void g_raid_tr_raid1_maybe_rebuild(struct g_raid_tr_object *tr,
131     struct g_raid_subdisk *sd);
132 
133 static int
134 g_raid_tr_taste_raid1(struct g_raid_tr_object *tr, struct g_raid_volume *vol)
135 {
136 	struct g_raid_tr_raid1_object *trs;
137 
138 	trs = (struct g_raid_tr_raid1_object *)tr;
139 	if (tr->tro_volume->v_raid_level != G_RAID_VOLUME_RL_RAID1 ||
140 	    (tr->tro_volume->v_raid_level_qualifier != G_RAID_VOLUME_RLQ_R1SM &&
141 	     tr->tro_volume->v_raid_level_qualifier != G_RAID_VOLUME_RLQ_R1MM))
142 		return (G_RAID_TR_TASTE_FAIL);
143 	trs->trso_starting = 1;
144 	return (G_RAID_TR_TASTE_SUCCEED);
145 }
146 
147 static int
148 g_raid_tr_update_state_raid1(struct g_raid_volume *vol,
149     struct g_raid_subdisk *sd)
150 {
151 	struct g_raid_tr_raid1_object *trs;
152 	struct g_raid_softc *sc;
153 	struct g_raid_subdisk *tsd, *bestsd;
154 	u_int s;
155 	int i, na, ns;
156 
157 	sc = vol->v_softc;
158 	trs = (struct g_raid_tr_raid1_object *)vol->v_tr;
159 	if (trs->trso_stopping &&
160 	    (trs->trso_flags & TR_RAID1_F_DOING_SOME) == 0)
161 		s = G_RAID_VOLUME_S_STOPPED;
162 	else if (trs->trso_starting)
163 		s = G_RAID_VOLUME_S_STARTING;
164 	else {
165 		/* Make sure we have at least one ACTIVE disk. */
166 		na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
167 		if (na == 0) {
168 			/*
169 			 * Critical situation! We have no any active disk!
170 			 * Choose the best disk we have to make it active.
171 			 */
172 			bestsd = &vol->v_subdisks[0];
173 			for (i = 1; i < vol->v_disks_count; i++) {
174 				tsd = &vol->v_subdisks[i];
175 				if (tsd->sd_state > bestsd->sd_state)
176 					bestsd = tsd;
177 				else if (tsd->sd_state == bestsd->sd_state &&
178 				    (tsd->sd_state == G_RAID_SUBDISK_S_REBUILD ||
179 				     tsd->sd_state == G_RAID_SUBDISK_S_RESYNC) &&
180 				    tsd->sd_rebuild_pos > bestsd->sd_rebuild_pos)
181 					bestsd = tsd;
182 			}
183 			if (bestsd->sd_state >= G_RAID_SUBDISK_S_UNINITIALIZED) {
184 				/* We found reasonable candidate. */
185 				G_RAID_DEBUG1(1, sc,
186 				    "Promote subdisk %s:%d from %s to ACTIVE.",
187 				    vol->v_name, bestsd->sd_pos,
188 				    g_raid_subdisk_state2str(bestsd->sd_state));
189 				g_raid_change_subdisk_state(bestsd,
190 				    G_RAID_SUBDISK_S_ACTIVE);
191 				g_raid_write_metadata(sc,
192 				    vol, bestsd, bestsd->sd_disk);
193 			}
194 		}
195 		na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
196 		ns = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
197 		     g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
198 		if (na == vol->v_disks_count)
199 			s = G_RAID_VOLUME_S_OPTIMAL;
200 		else if (na + ns == vol->v_disks_count)
201 			s = G_RAID_VOLUME_S_SUBOPTIMAL;
202 		else if (na > 0)
203 			s = G_RAID_VOLUME_S_DEGRADED;
204 		else
205 			s = G_RAID_VOLUME_S_BROKEN;
206 		g_raid_tr_raid1_maybe_rebuild(vol->v_tr, sd);
207 	}
208 	if (s != vol->v_state) {
209 		g_raid_event_send(vol, G_RAID_VOLUME_S_ALIVE(s) ?
210 		    G_RAID_VOLUME_E_UP : G_RAID_VOLUME_E_DOWN,
211 		    G_RAID_EVENT_VOLUME);
212 		g_raid_change_volume_state(vol, s);
213 		if (!trs->trso_starting && !trs->trso_stopping)
214 			g_raid_write_metadata(sc, vol, NULL, NULL);
215 	}
216 	return (0);
217 }
218 
219 static void
220 g_raid_tr_raid1_fail_disk(struct g_raid_softc *sc, struct g_raid_subdisk *sd,
221     struct g_raid_disk *disk)
222 {
223 	/*
224 	 * We don't fail the last disk in the pack, since it still has decent
225 	 * data on it and that's better than failing the disk if it is the root
226 	 * file system.
227 	 *
228 	 * XXX should this be controlled via a tunable?  It makes sense for
229 	 * the volume that has / on it.  I can't think of a case where we'd
230 	 * want the volume to go away on this kind of event.
231 	 */
232 	if (g_raid_nsubdisks(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == 1 &&
233 	    g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == sd)
234 		return;
235 	g_raid_fail_disk(sc, sd, disk);
236 }
237 
238 static void
239 g_raid_tr_raid1_rebuild_some(struct g_raid_tr_object *tr)
240 {
241 	struct g_raid_tr_raid1_object *trs;
242 	struct g_raid_subdisk *sd, *good_sd;
243 	struct bio *bp;
244 
245 	trs = (struct g_raid_tr_raid1_object *)tr;
246 	if (trs->trso_flags & TR_RAID1_F_DOING_SOME)
247 		return;
248 	sd = trs->trso_failed_sd;
249 	good_sd = g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE);
250 	if (good_sd == NULL) {
251 		g_raid_tr_raid1_rebuild_abort(tr);
252 		return;
253 	}
254 	bp = &trs->trso_bio;
255 	memset(bp, 0, sizeof(*bp));
256 	bp->bio_offset = sd->sd_rebuild_pos;
257 	bp->bio_length = MIN(g_raid1_rebuild_slab,
258 	    sd->sd_size - sd->sd_rebuild_pos);
259 	bp->bio_data = trs->trso_buffer;
260 	bp->bio_cmd = BIO_READ;
261 	bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
262 	bp->bio_caller1 = good_sd;
263 	trs->trso_flags |= TR_RAID1_F_DOING_SOME;
264 	trs->trso_flags |= TR_RAID1_F_LOCKED;
265 	g_raid_lock_range(sd->sd_volume,	/* Lock callback starts I/O */
266 	   bp->bio_offset, bp->bio_length, NULL, bp);
267 }
268 
269 static void
270 g_raid_tr_raid1_rebuild_done(struct g_raid_tr_raid1_object *trs)
271 {
272 	struct g_raid_volume *vol;
273 	struct g_raid_subdisk *sd;
274 
275 	vol = trs->trso_base.tro_volume;
276 	sd = trs->trso_failed_sd;
277 	g_raid_write_metadata(vol->v_softc, vol, sd, sd->sd_disk);
278 	free(trs->trso_buffer, M_TR_RAID1);
279 	trs->trso_buffer = NULL;
280 	trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
281 	trs->trso_type = TR_RAID1_NONE;
282 	trs->trso_recover_slabs = 0;
283 	trs->trso_failed_sd = NULL;
284 	g_raid_tr_update_state_raid1(vol, NULL);
285 }
286 
287 static void
288 g_raid_tr_raid1_rebuild_finish(struct g_raid_tr_object *tr)
289 {
290 	struct g_raid_tr_raid1_object *trs;
291 	struct g_raid_subdisk *sd;
292 
293 	trs = (struct g_raid_tr_raid1_object *)tr;
294 	sd = trs->trso_failed_sd;
295 	G_RAID_DEBUG1(0, tr->tro_volume->v_softc,
296 	    "Subdisk %s:%d-%s rebuild completed.",
297 	    sd->sd_volume->v_name, sd->sd_pos,
298 	    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
299 	g_raid_change_subdisk_state(sd, G_RAID_SUBDISK_S_ACTIVE);
300 	sd->sd_rebuild_pos = 0;
301 	g_raid_tr_raid1_rebuild_done(trs);
302 }
303 
304 static void
305 g_raid_tr_raid1_rebuild_abort(struct g_raid_tr_object *tr)
306 {
307 	struct g_raid_tr_raid1_object *trs;
308 	struct g_raid_subdisk *sd;
309 	struct g_raid_volume *vol;
310 	off_t len;
311 
312 	vol = tr->tro_volume;
313 	trs = (struct g_raid_tr_raid1_object *)tr;
314 	sd = trs->trso_failed_sd;
315 	if (trs->trso_flags & TR_RAID1_F_DOING_SOME) {
316 		G_RAID_DEBUG1(1, vol->v_softc,
317 		    "Subdisk %s:%d-%s rebuild is aborting.",
318 		    sd->sd_volume->v_name, sd->sd_pos,
319 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
320 		trs->trso_flags |= TR_RAID1_F_ABORT;
321 	} else {
322 		G_RAID_DEBUG1(0, vol->v_softc,
323 		    "Subdisk %s:%d-%s rebuild aborted.",
324 		    sd->sd_volume->v_name, sd->sd_pos,
325 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
326 		trs->trso_flags &= ~TR_RAID1_F_ABORT;
327 		if (trs->trso_flags & TR_RAID1_F_LOCKED) {
328 			trs->trso_flags &= ~TR_RAID1_F_LOCKED;
329 			len = MIN(g_raid1_rebuild_slab,
330 			    sd->sd_size - sd->sd_rebuild_pos);
331 			g_raid_unlock_range(tr->tro_volume,
332 			    sd->sd_rebuild_pos, len);
333 		}
334 		g_raid_tr_raid1_rebuild_done(trs);
335 	}
336 }
337 
338 static void
339 g_raid_tr_raid1_rebuild_start(struct g_raid_tr_object *tr)
340 {
341 	struct g_raid_volume *vol;
342 	struct g_raid_tr_raid1_object *trs;
343 	struct g_raid_subdisk *sd, *fsd;
344 
345 	vol = tr->tro_volume;
346 	trs = (struct g_raid_tr_raid1_object *)tr;
347 	if (trs->trso_failed_sd) {
348 		G_RAID_DEBUG1(1, vol->v_softc,
349 		    "Already rebuild in start rebuild. pos %jd\n",
350 		    (intmax_t)trs->trso_failed_sd->sd_rebuild_pos);
351 		return;
352 	}
353 	sd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_ACTIVE);
354 	if (sd == NULL) {
355 		G_RAID_DEBUG1(1, vol->v_softc,
356 		    "No active disk to rebuild.  night night.");
357 		return;
358 	}
359 	fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_RESYNC);
360 	if (fsd == NULL)
361 		fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_REBUILD);
362 	if (fsd == NULL) {
363 		fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_STALE);
364 		if (fsd != NULL) {
365 			fsd->sd_rebuild_pos = 0;
366 			g_raid_change_subdisk_state(fsd,
367 			    G_RAID_SUBDISK_S_RESYNC);
368 			g_raid_write_metadata(vol->v_softc, vol, fsd, NULL);
369 		} else {
370 			fsd = g_raid_get_subdisk(vol,
371 			    G_RAID_SUBDISK_S_UNINITIALIZED);
372 			if (fsd == NULL)
373 				fsd = g_raid_get_subdisk(vol,
374 				    G_RAID_SUBDISK_S_NEW);
375 			if (fsd != NULL) {
376 				fsd->sd_rebuild_pos = 0;
377 				g_raid_change_subdisk_state(fsd,
378 				    G_RAID_SUBDISK_S_REBUILD);
379 				g_raid_write_metadata(vol->v_softc,
380 				    vol, fsd, NULL);
381 			}
382 		}
383 	}
384 	if (fsd == NULL) {
385 		G_RAID_DEBUG1(1, vol->v_softc,
386 		    "No failed disk to rebuild.  night night.");
387 		return;
388 	}
389 	trs->trso_failed_sd = fsd;
390 	G_RAID_DEBUG1(0, vol->v_softc,
391 	    "Subdisk %s:%d-%s rebuild start at %jd.",
392 	    fsd->sd_volume->v_name, fsd->sd_pos,
393 	    fsd->sd_disk ? g_raid_get_diskname(fsd->sd_disk) : "[none]",
394 	    trs->trso_failed_sd->sd_rebuild_pos);
395 	trs->trso_type = TR_RAID1_REBUILD;
396 	trs->trso_buffer = malloc(g_raid1_rebuild_slab, M_TR_RAID1, M_WAITOK);
397 	trs->trso_meta_update = g_raid1_rebuild_meta_update;
398 	g_raid_tr_raid1_rebuild_some(tr);
399 }
400 
401 static void
402 g_raid_tr_raid1_maybe_rebuild(struct g_raid_tr_object *tr,
403     struct g_raid_subdisk *sd)
404 {
405 	struct g_raid_volume *vol;
406 	struct g_raid_tr_raid1_object *trs;
407 	int na, nr;
408 
409 	/*
410 	 * If we're stopping, don't do anything.  If we don't have at least one
411 	 * good disk and one bad disk, we don't do anything.  And if there's a
412 	 * 'good disk' stored in the trs, then we're in progress and we punt.
413 	 * If we make it past all these checks, we need to rebuild.
414 	 */
415 	vol = tr->tro_volume;
416 	trs = (struct g_raid_tr_raid1_object *)tr;
417 	if (trs->trso_stopping)
418 		return;
419 	na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
420 	nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_REBUILD) +
421 	    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
422 	switch(trs->trso_type) {
423 	case TR_RAID1_NONE:
424 		if (na == 0)
425 			return;
426 		if (nr == 0) {
427 			nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_NEW) +
428 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
429 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_UNINITIALIZED);
430 			if (nr == 0)
431 				return;
432 		}
433 		g_raid_tr_raid1_rebuild_start(tr);
434 		break;
435 	case TR_RAID1_REBUILD:
436 		if (na == 0 || nr == 0 || trs->trso_failed_sd == sd)
437 			g_raid_tr_raid1_rebuild_abort(tr);
438 		break;
439 	case TR_RAID1_RESYNC:
440 		break;
441 	}
442 }
443 
444 static int
445 g_raid_tr_event_raid1(struct g_raid_tr_object *tr,
446     struct g_raid_subdisk *sd, u_int event)
447 {
448 
449 	g_raid_tr_update_state_raid1(tr->tro_volume, sd);
450 	return (0);
451 }
452 
453 static int
454 g_raid_tr_start_raid1(struct g_raid_tr_object *tr)
455 {
456 	struct g_raid_tr_raid1_object *trs;
457 	struct g_raid_volume *vol;
458 
459 	trs = (struct g_raid_tr_raid1_object *)tr;
460 	vol = tr->tro_volume;
461 	trs->trso_starting = 0;
462 	g_raid_tr_update_state_raid1(vol, NULL);
463 	return (0);
464 }
465 
466 static int
467 g_raid_tr_stop_raid1(struct g_raid_tr_object *tr)
468 {
469 	struct g_raid_tr_raid1_object *trs;
470 	struct g_raid_volume *vol;
471 
472 	trs = (struct g_raid_tr_raid1_object *)tr;
473 	vol = tr->tro_volume;
474 	trs->trso_starting = 0;
475 	trs->trso_stopping = 1;
476 	g_raid_tr_update_state_raid1(vol, NULL);
477 	return (0);
478 }
479 
480 /*
481  * Select the disk to read from.  Take into account: subdisk state, running
482  * error recovery, average disk load, head position and possible cache hits.
483  */
484 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
485 static struct g_raid_subdisk *
486 g_raid_tr_raid1_select_read_disk(struct g_raid_volume *vol, struct bio *bp,
487     u_int mask)
488 {
489 	struct g_raid_subdisk *sd, *best;
490 	int i, prio, bestprio;
491 
492 	best = NULL;
493 	bestprio = INT_MAX;
494 	for (i = 0; i < vol->v_disks_count; i++) {
495 		sd = &vol->v_subdisks[i];
496 		if (sd->sd_state != G_RAID_SUBDISK_S_ACTIVE &&
497 		    ((sd->sd_state != G_RAID_SUBDISK_S_REBUILD &&
498 		      sd->sd_state != G_RAID_SUBDISK_S_RESYNC) ||
499 		     bp->bio_offset + bp->bio_length > sd->sd_rebuild_pos))
500 			continue;
501 		if ((mask & (1 << i)) != 0)
502 			continue;
503 		prio = G_RAID_SUBDISK_LOAD(sd);
504 		prio += min(sd->sd_recovery, 255) << 22;
505 		prio += (G_RAID_SUBDISK_S_ACTIVE - sd->sd_state) << 16;
506 		/* If disk head is precisely in position - highly prefer it. */
507 		if (G_RAID_SUBDISK_POS(sd) == bp->bio_offset)
508 			prio -= 2 * G_RAID_SUBDISK_LOAD_SCALE;
509 		else
510 		/* If disk head is close to position - prefer it. */
511 		if (ABS(G_RAID_SUBDISK_POS(sd) - bp->bio_offset) <
512 		    G_RAID_SUBDISK_TRACK_SIZE)
513 			prio -= 1 * G_RAID_SUBDISK_LOAD_SCALE;
514 		if (prio < bestprio) {
515 			best = sd;
516 			bestprio = prio;
517 		}
518 	}
519 	return (best);
520 }
521 
522 static void
523 g_raid_tr_iostart_raid1_read(struct g_raid_tr_object *tr, struct bio *bp)
524 {
525 	struct g_raid_subdisk *sd;
526 	struct bio *cbp;
527 
528 	sd = g_raid_tr_raid1_select_read_disk(tr->tro_volume, bp, 0);
529 	KASSERT(sd != NULL, ("No active disks in volume %s.",
530 		tr->tro_volume->v_name));
531 
532 	cbp = g_clone_bio(bp);
533 	if (cbp == NULL) {
534 		g_raid_iodone(bp, ENOMEM);
535 		return;
536 	}
537 
538 	g_raid_subdisk_iostart(sd, cbp);
539 }
540 
541 static void
542 g_raid_tr_iostart_raid1_write(struct g_raid_tr_object *tr, struct bio *bp)
543 {
544 	struct g_raid_volume *vol;
545 	struct g_raid_subdisk *sd;
546 	struct bio_queue_head queue;
547 	struct bio *cbp;
548 	int i;
549 
550 	vol = tr->tro_volume;
551 
552 	/*
553 	 * Allocate all bios before sending any request, so we can return
554 	 * ENOMEM in nice and clean way.
555 	 */
556 	bioq_init(&queue);
557 	for (i = 0; i < vol->v_disks_count; i++) {
558 		sd = &vol->v_subdisks[i];
559 		switch (sd->sd_state) {
560 		case G_RAID_SUBDISK_S_ACTIVE:
561 			break;
562 		case G_RAID_SUBDISK_S_REBUILD:
563 			/*
564 			 * When rebuilding, only part of this subdisk is
565 			 * writable, the rest will be written as part of the
566 			 * that process.
567 			 */
568 			if (bp->bio_offset >= sd->sd_rebuild_pos)
569 				continue;
570 			break;
571 		case G_RAID_SUBDISK_S_STALE:
572 		case G_RAID_SUBDISK_S_RESYNC:
573 			/*
574 			 * Resyncing still writes on the theory that the
575 			 * resync'd disk is very close and writing it will
576 			 * keep it that way better if we keep up while
577 			 * resyncing.
578 			 */
579 			break;
580 		default:
581 			continue;
582 		}
583 		cbp = g_clone_bio(bp);
584 		if (cbp == NULL)
585 			goto failure;
586 		cbp->bio_caller1 = sd;
587 		bioq_insert_tail(&queue, cbp);
588 	}
589 	while ((cbp = bioq_takefirst(&queue)) != NULL) {
590 		sd = cbp->bio_caller1;
591 		cbp->bio_caller1 = NULL;
592 		g_raid_subdisk_iostart(sd, cbp);
593 	}
594 	return;
595 failure:
596 	while ((cbp = bioq_takefirst(&queue)) != NULL)
597 		g_destroy_bio(cbp);
598 	if (bp->bio_error == 0)
599 		bp->bio_error = ENOMEM;
600 	g_raid_iodone(bp, bp->bio_error);
601 }
602 
603 static void
604 g_raid_tr_iostart_raid1(struct g_raid_tr_object *tr, struct bio *bp)
605 {
606 	struct g_raid_volume *vol;
607 	struct g_raid_tr_raid1_object *trs;
608 
609 	vol = tr->tro_volume;
610 	trs = (struct g_raid_tr_raid1_object *)tr;
611 	if (vol->v_state != G_RAID_VOLUME_S_OPTIMAL &&
612 	    vol->v_state != G_RAID_VOLUME_S_SUBOPTIMAL &&
613 	    vol->v_state != G_RAID_VOLUME_S_DEGRADED) {
614 		g_raid_iodone(bp, EIO);
615 		return;
616 	}
617 	/*
618 	 * If we're rebuilding, squeeze in rebuild activity every so often,
619 	 * even when the disk is busy.  Be sure to only count real I/O
620 	 * to the disk.  All 'SPECIAL' I/O is traffic generated to the disk
621 	 * by this module.
622 	 */
623 	if (trs->trso_failed_sd != NULL &&
624 	    !(bp->bio_cflags & G_RAID_BIO_FLAG_SPECIAL)) {
625 		/* Make this new or running now round short. */
626 		trs->trso_recover_slabs = 0;
627 		if (--trs->trso_fair_io <= 0) {
628 			trs->trso_fair_io = g_raid1_rebuild_fair_io;
629 			g_raid_tr_raid1_rebuild_some(tr);
630 		}
631 	}
632 	switch (bp->bio_cmd) {
633 	case BIO_READ:
634 		g_raid_tr_iostart_raid1_read(tr, bp);
635 		break;
636 	case BIO_WRITE:
637 	case BIO_DELETE:
638 		g_raid_tr_iostart_raid1_write(tr, bp);
639 		break;
640 	case BIO_SPEEDUP:
641 	case BIO_FLUSH:
642 		g_raid_tr_flush_common(tr, bp);
643 		break;
644 	default:
645 		KASSERT(1 == 0, ("Invalid command here: %u (volume=%s)",
646 		    bp->bio_cmd, vol->v_name));
647 		break;
648 	}
649 }
650 
651 static void
652 g_raid_tr_iodone_raid1(struct g_raid_tr_object *tr,
653     struct g_raid_subdisk *sd, struct bio *bp)
654 {
655 	struct bio *cbp;
656 	struct g_raid_subdisk *nsd;
657 	struct g_raid_volume *vol;
658 	struct bio *pbp;
659 	struct g_raid_tr_raid1_object *trs;
660 	uintptr_t *mask;
661 	int error, do_write;
662 
663 	trs = (struct g_raid_tr_raid1_object *)tr;
664 	vol = tr->tro_volume;
665 	if (bp->bio_cflags & G_RAID_BIO_FLAG_SYNC) {
666 		/*
667 		 * This operation is part of a rebuild or resync operation.
668 		 * See what work just got done, then schedule the next bit of
669 		 * work, if any.  Rebuild/resync is done a little bit at a
670 		 * time.  Either when a timeout happens, or after we get a
671 		 * bunch of I/Os to the disk (to make sure an active system
672 		 * will complete in a sane amount of time).
673 		 *
674 		 * We are setup to do differing amounts of work for each of
675 		 * these cases.  so long as the slabs is smallish (less than
676 		 * 50 or so, I'd guess, but that's just a WAG), we shouldn't
677 		 * have any bio starvation issues.  For active disks, we do
678 		 * 5MB of data, for inactive ones, we do 50MB.
679 		 */
680 		if (trs->trso_type == TR_RAID1_REBUILD) {
681 			if (bp->bio_cmd == BIO_READ) {
682 				/* Immediately abort rebuild, if requested. */
683 				if (trs->trso_flags & TR_RAID1_F_ABORT) {
684 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
685 					g_raid_tr_raid1_rebuild_abort(tr);
686 					return;
687 				}
688 
689 				/* On read error, skip and cross fingers. */
690 				if (bp->bio_error != 0) {
691 					G_RAID_LOGREQ(0, bp,
692 					    "Read error during rebuild (%d), "
693 					    "possible data loss!",
694 					    bp->bio_error);
695 					goto rebuild_round_done;
696 				}
697 
698 				/*
699 				 * The read operation finished, queue the
700 				 * write and get out.
701 				 */
702 				G_RAID_LOGREQ(4, bp, "rebuild read done. %d",
703 				    bp->bio_error);
704 				bp->bio_cmd = BIO_WRITE;
705 				bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
706 				G_RAID_LOGREQ(4, bp, "Queueing rebuild write.");
707 				g_raid_subdisk_iostart(trs->trso_failed_sd, bp);
708 			} else {
709 				/*
710 				 * The write operation just finished.  Do
711 				 * another.  We keep cloning the master bio
712 				 * since it has the right buffers allocated to
713 				 * it.
714 				 */
715 				G_RAID_LOGREQ(4, bp,
716 				    "rebuild write done. Error %d",
717 				    bp->bio_error);
718 				nsd = trs->trso_failed_sd;
719 				if (bp->bio_error != 0 ||
720 				    trs->trso_flags & TR_RAID1_F_ABORT) {
721 					if ((trs->trso_flags &
722 					    TR_RAID1_F_ABORT) == 0) {
723 						g_raid_tr_raid1_fail_disk(sd->sd_softc,
724 						    nsd, nsd->sd_disk);
725 					}
726 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
727 					g_raid_tr_raid1_rebuild_abort(tr);
728 					return;
729 				}
730 rebuild_round_done:
731 				nsd = trs->trso_failed_sd;
732 				trs->trso_flags &= ~TR_RAID1_F_LOCKED;
733 				g_raid_unlock_range(sd->sd_volume,
734 				    bp->bio_offset, bp->bio_length);
735 				nsd->sd_rebuild_pos += bp->bio_length;
736 				if (nsd->sd_rebuild_pos >= nsd->sd_size) {
737 					g_raid_tr_raid1_rebuild_finish(tr);
738 					return;
739 				}
740 
741 				/* Abort rebuild if we are stopping */
742 				if (trs->trso_stopping) {
743 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
744 					g_raid_tr_raid1_rebuild_abort(tr);
745 					return;
746 				}
747 
748 				if (--trs->trso_meta_update <= 0) {
749 					g_raid_write_metadata(vol->v_softc,
750 					    vol, nsd, nsd->sd_disk);
751 					trs->trso_meta_update =
752 					    g_raid1_rebuild_meta_update;
753 				}
754 				trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
755 				if (--trs->trso_recover_slabs <= 0)
756 					return;
757 				g_raid_tr_raid1_rebuild_some(tr);
758 			}
759 		} else if (trs->trso_type == TR_RAID1_RESYNC) {
760 			/*
761 			 * read good sd, read bad sd in parallel.  when both
762 			 * done, compare the buffers.  write good to the bad
763 			 * if different.  do the next bit of work.
764 			 */
765 			panic("Somehow, we think we're doing a resync");
766 		}
767 		return;
768 	}
769 	pbp = bp->bio_parent;
770 	pbp->bio_inbed++;
771 	if (bp->bio_cmd == BIO_READ && bp->bio_error != 0) {
772 		/*
773 		 * Read failed on first drive.  Retry the read error on
774 		 * another disk drive, if available, before erroring out the
775 		 * read.
776 		 */
777 		sd->sd_disk->d_read_errs++;
778 		G_RAID_LOGREQ(0, bp,
779 		    "Read error (%d), %d read errors total",
780 		    bp->bio_error, sd->sd_disk->d_read_errs);
781 
782 		/*
783 		 * If there are too many read errors, we move to degraded.
784 		 * XXX Do we want to FAIL the drive (eg, make the user redo
785 		 * everything to get it back in sync), or just degrade the
786 		 * drive, which kicks off a resync?
787 		 */
788 		do_write = 1;
789 		if (sd->sd_disk->d_read_errs > g_raid_read_err_thresh) {
790 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
791 			if (pbp->bio_children == 1)
792 				do_write = 0;
793 		}
794 
795 		/*
796 		 * Find the other disk, and try to do the I/O to it.
797 		 */
798 		mask = (uintptr_t *)(&pbp->bio_driver2);
799 		if (pbp->bio_children == 1) {
800 			/* Save original subdisk. */
801 			pbp->bio_driver1 = do_write ? sd : NULL;
802 			*mask = 0;
803 		}
804 		*mask |= 1 << sd->sd_pos;
805 		nsd = g_raid_tr_raid1_select_read_disk(vol, pbp, *mask);
806 		if (nsd != NULL && (cbp = g_clone_bio(pbp)) != NULL) {
807 			g_destroy_bio(bp);
808 			G_RAID_LOGREQ(2, cbp, "Retrying read from %d",
809 			    nsd->sd_pos);
810 			if (pbp->bio_children == 2 && do_write) {
811 				sd->sd_recovery++;
812 				cbp->bio_caller1 = nsd;
813 				pbp->bio_pflags = G_RAID_BIO_FLAG_LOCKED;
814 				/* Lock callback starts I/O */
815 				g_raid_lock_range(sd->sd_volume,
816 				    cbp->bio_offset, cbp->bio_length, pbp, cbp);
817 			} else {
818 				g_raid_subdisk_iostart(nsd, cbp);
819 			}
820 			return;
821 		}
822 		/*
823 		 * We can't retry.  Return the original error by falling
824 		 * through.  This will happen when there's only one good disk.
825 		 * We don't need to fail the raid, since its actual state is
826 		 * based on the state of the subdisks.
827 		 */
828 		G_RAID_LOGREQ(2, bp, "Couldn't retry read, failing it");
829 	}
830 	if (bp->bio_cmd == BIO_READ &&
831 	    bp->bio_error == 0 &&
832 	    pbp->bio_children > 1 &&
833 	    pbp->bio_driver1 != NULL) {
834 		/*
835 		 * If it was a read, and bio_children is >1, then we just
836 		 * recovered the data from the second drive.  We should try to
837 		 * write that data to the first drive if sector remapping is
838 		 * enabled.  A write should put the data in a new place on the
839 		 * disk, remapping the bad sector.  Do we need to do that by
840 		 * queueing a request to the main worker thread?  It doesn't
841 		 * affect the return code of this current read, and can be
842 		 * done at our leisure.  However, to make the code simpler, it
843 		 * is done synchronously.
844 		 */
845 		G_RAID_LOGREQ(3, bp, "Recovered data from other drive");
846 		cbp = g_clone_bio(pbp);
847 		if (cbp != NULL) {
848 			g_destroy_bio(bp);
849 			cbp->bio_cmd = BIO_WRITE;
850 			cbp->bio_cflags = G_RAID_BIO_FLAG_REMAP;
851 			G_RAID_LOGREQ(2, cbp,
852 			    "Attempting bad sector remap on failing drive.");
853 			g_raid_subdisk_iostart(pbp->bio_driver1, cbp);
854 			return;
855 		}
856 	}
857 	if (pbp->bio_pflags & G_RAID_BIO_FLAG_LOCKED) {
858 		/*
859 		 * We're done with a recovery, mark the range as unlocked.
860 		 * For any write errors, we aggressively fail the disk since
861 		 * there was both a READ and a WRITE error at this location.
862 		 * Both types of errors generally indicates the drive is on
863 		 * the verge of total failure anyway.  Better to stop trusting
864 		 * it now.  However, we need to reset error to 0 in that case
865 		 * because we're not failing the original I/O which succeeded.
866 		 */
867 		if (bp->bio_cmd == BIO_WRITE && bp->bio_error) {
868 			G_RAID_LOGREQ(0, bp, "Remap write failed: "
869 			    "failing subdisk.");
870 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
871 			bp->bio_error = 0;
872 		}
873 		if (pbp->bio_driver1 != NULL) {
874 			((struct g_raid_subdisk *)pbp->bio_driver1)
875 			    ->sd_recovery--;
876 		}
877 		G_RAID_LOGREQ(2, bp, "REMAP done %d.", bp->bio_error);
878 		g_raid_unlock_range(sd->sd_volume, bp->bio_offset,
879 		    bp->bio_length);
880 	}
881 	if (pbp->bio_cmd != BIO_READ) {
882 		if (pbp->bio_inbed == 1 || pbp->bio_error != 0)
883 			pbp->bio_error = bp->bio_error;
884 		if (pbp->bio_cmd == BIO_WRITE && bp->bio_error != 0) {
885 			G_RAID_LOGREQ(0, bp, "Write failed: failing subdisk.");
886 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
887 		}
888 		error = pbp->bio_error;
889 	} else
890 		error = bp->bio_error;
891 	g_destroy_bio(bp);
892 	if (pbp->bio_children == pbp->bio_inbed) {
893 		pbp->bio_completed = pbp->bio_length;
894 		g_raid_iodone(pbp, error);
895 	}
896 }
897 
898 static int
899 g_raid_tr_kerneldump_raid1(struct g_raid_tr_object *tr, void *virtual,
900     off_t offset, size_t length)
901 {
902 	struct g_raid_volume *vol;
903 	struct g_raid_subdisk *sd;
904 	int error, i, ok;
905 
906 	vol = tr->tro_volume;
907 	error = 0;
908 	ok = 0;
909 	for (i = 0; i < vol->v_disks_count; i++) {
910 		sd = &vol->v_subdisks[i];
911 		switch (sd->sd_state) {
912 		case G_RAID_SUBDISK_S_ACTIVE:
913 			break;
914 		case G_RAID_SUBDISK_S_REBUILD:
915 			/*
916 			 * When rebuilding, only part of this subdisk is
917 			 * writable, the rest will be written as part of the
918 			 * that process.
919 			 */
920 			if (offset >= sd->sd_rebuild_pos)
921 				continue;
922 			break;
923 		case G_RAID_SUBDISK_S_STALE:
924 		case G_RAID_SUBDISK_S_RESYNC:
925 			/*
926 			 * Resyncing still writes on the theory that the
927 			 * resync'd disk is very close and writing it will
928 			 * keep it that way better if we keep up while
929 			 * resyncing.
930 			 */
931 			break;
932 		default:
933 			continue;
934 		}
935 		error = g_raid_subdisk_kerneldump(sd, virtual, offset, length);
936 		if (error == 0)
937 			ok++;
938 	}
939 	return (ok > 0 ? 0 : error);
940 }
941 
942 static int
943 g_raid_tr_locked_raid1(struct g_raid_tr_object *tr, void *argp)
944 {
945 	struct bio *bp;
946 	struct g_raid_subdisk *sd;
947 
948 	bp = (struct bio *)argp;
949 	sd = (struct g_raid_subdisk *)bp->bio_caller1;
950 	g_raid_subdisk_iostart(sd, bp);
951 
952 	return (0);
953 }
954 
955 static int
956 g_raid_tr_idle_raid1(struct g_raid_tr_object *tr)
957 {
958 	struct g_raid_tr_raid1_object *trs;
959 
960 	trs = (struct g_raid_tr_raid1_object *)tr;
961 	trs->trso_fair_io = g_raid1_rebuild_fair_io;
962 	trs->trso_recover_slabs = g_raid1_rebuild_cluster_idle;
963 	if (trs->trso_type == TR_RAID1_REBUILD)
964 		g_raid_tr_raid1_rebuild_some(tr);
965 	return (0);
966 }
967 
968 static int
969 g_raid_tr_free_raid1(struct g_raid_tr_object *tr)
970 {
971 	struct g_raid_tr_raid1_object *trs;
972 
973 	trs = (struct g_raid_tr_raid1_object *)tr;
974 
975 	if (trs->trso_buffer != NULL) {
976 		free(trs->trso_buffer, M_TR_RAID1);
977 		trs->trso_buffer = NULL;
978 	}
979 	return (0);
980 }
981 
982 G_RAID_TR_DECLARE(raid1, "RAID1");
983