xref: /freebsd/sys/geom/raid/tr_raid1.c (revision b07549f3c15637af6064d2f52b572f239f5ba870)
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/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/bio.h>
34 #include <sys/endian.h>
35 #include <sys/kernel.h>
36 #include <sys/kobj.h>
37 #include <sys/limits.h>
38 #include <sys/lock.h>
39 #include <sys/malloc.h>
40 #include <sys/mutex.h>
41 #include <sys/sysctl.h>
42 #include <sys/systm.h>
43 #include <geom/geom.h>
44 #include <geom/geom_dbg.h>
45 #include "geom/raid/g_raid.h"
46 #include "g_raid_tr_if.h"
47 
48 SYSCTL_DECL(_kern_geom_raid_raid1);
49 
50 #define RAID1_REBUILD_SLAB	(1 << 20) /* One transation in a rebuild */
51 static int g_raid1_rebuild_slab = RAID1_REBUILD_SLAB;
52 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_slab_size, CTLFLAG_RWTUN,
53     &g_raid1_rebuild_slab, 0,
54     "Amount of the disk to rebuild each read/write cycle of the rebuild.");
55 
56 #define RAID1_REBUILD_FAIR_IO 20 /* use 1/x of the available I/O */
57 static int g_raid1_rebuild_fair_io = RAID1_REBUILD_FAIR_IO;
58 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_fair_io, CTLFLAG_RWTUN,
59     &g_raid1_rebuild_fair_io, 0,
60     "Fraction of the I/O bandwidth to use when disk busy for rebuild.");
61 
62 #define RAID1_REBUILD_CLUSTER_IDLE 100
63 static int g_raid1_rebuild_cluster_idle = RAID1_REBUILD_CLUSTER_IDLE;
64 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_cluster_idle, CTLFLAG_RWTUN,
65     &g_raid1_rebuild_cluster_idle, 0,
66     "Number of slabs to do each time we trigger a rebuild cycle");
67 
68 #define RAID1_REBUILD_META_UPDATE 1024 /* update meta data every 1GB or so */
69 static int g_raid1_rebuild_meta_update = RAID1_REBUILD_META_UPDATE;
70 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_meta_update, CTLFLAG_RWTUN,
71     &g_raid1_rebuild_meta_update, 0,
72     "When to update the meta data.");
73 
74 static MALLOC_DEFINE(M_TR_RAID1, "tr_raid1_data", "GEOM_RAID RAID1 data");
75 
76 #define TR_RAID1_NONE 0
77 #define TR_RAID1_REBUILD 1
78 #define TR_RAID1_RESYNC 2
79 
80 #define TR_RAID1_F_DOING_SOME	0x1
81 #define TR_RAID1_F_LOCKED	0x2
82 #define TR_RAID1_F_ABORT	0x4
83 
84 struct g_raid_tr_raid1_object {
85 	struct g_raid_tr_object	 trso_base;
86 	int			 trso_starting;
87 	int			 trso_stopping;
88 	int			 trso_type;
89 	int			 trso_recover_slabs; /* slabs before rest */
90 	int			 trso_fair_io;
91 	int			 trso_meta_update;
92 	int			 trso_flags;
93 	struct g_raid_subdisk	*trso_failed_sd; /* like per volume */
94 	void			*trso_buffer;	 /* Buffer space */
95 	struct bio		 trso_bio;
96 };
97 
98 static g_raid_tr_taste_t g_raid_tr_taste_raid1;
99 static g_raid_tr_event_t g_raid_tr_event_raid1;
100 static g_raid_tr_start_t g_raid_tr_start_raid1;
101 static g_raid_tr_stop_t g_raid_tr_stop_raid1;
102 static g_raid_tr_iostart_t g_raid_tr_iostart_raid1;
103 static g_raid_tr_iodone_t g_raid_tr_iodone_raid1;
104 static g_raid_tr_kerneldump_t g_raid_tr_kerneldump_raid1;
105 static g_raid_tr_locked_t g_raid_tr_locked_raid1;
106 static g_raid_tr_idle_t g_raid_tr_idle_raid1;
107 static g_raid_tr_free_t g_raid_tr_free_raid1;
108 
109 static kobj_method_t g_raid_tr_raid1_methods[] = {
110 	KOBJMETHOD(g_raid_tr_taste,	g_raid_tr_taste_raid1),
111 	KOBJMETHOD(g_raid_tr_event,	g_raid_tr_event_raid1),
112 	KOBJMETHOD(g_raid_tr_start,	g_raid_tr_start_raid1),
113 	KOBJMETHOD(g_raid_tr_stop,	g_raid_tr_stop_raid1),
114 	KOBJMETHOD(g_raid_tr_iostart,	g_raid_tr_iostart_raid1),
115 	KOBJMETHOD(g_raid_tr_iodone,	g_raid_tr_iodone_raid1),
116 	KOBJMETHOD(g_raid_tr_kerneldump, g_raid_tr_kerneldump_raid1),
117 	KOBJMETHOD(g_raid_tr_locked,	g_raid_tr_locked_raid1),
118 	KOBJMETHOD(g_raid_tr_idle,	g_raid_tr_idle_raid1),
119 	KOBJMETHOD(g_raid_tr_free,	g_raid_tr_free_raid1),
120 	{ 0, 0 }
121 };
122 
123 static struct g_raid_tr_class g_raid_tr_raid1_class = {
124 	"RAID1",
125 	g_raid_tr_raid1_methods,
126 	sizeof(struct g_raid_tr_raid1_object),
127 	.trc_enable = 1,
128 	.trc_priority = 100,
129 	.trc_accept_unmapped = 1
130 };
131 
132 static void g_raid_tr_raid1_rebuild_abort(struct g_raid_tr_object *tr);
133 static void g_raid_tr_raid1_maybe_rebuild(struct g_raid_tr_object *tr,
134     struct g_raid_subdisk *sd);
135 
136 static int
137 g_raid_tr_taste_raid1(struct g_raid_tr_object *tr, struct g_raid_volume *vol)
138 {
139 	struct g_raid_tr_raid1_object *trs;
140 
141 	trs = (struct g_raid_tr_raid1_object *)tr;
142 	if (tr->tro_volume->v_raid_level != G_RAID_VOLUME_RL_RAID1 ||
143 	    (tr->tro_volume->v_raid_level_qualifier != G_RAID_VOLUME_RLQ_R1SM &&
144 	     tr->tro_volume->v_raid_level_qualifier != G_RAID_VOLUME_RLQ_R1MM))
145 		return (G_RAID_TR_TASTE_FAIL);
146 	trs->trso_starting = 1;
147 	return (G_RAID_TR_TASTE_SUCCEED);
148 }
149 
150 static int
151 g_raid_tr_update_state_raid1(struct g_raid_volume *vol,
152     struct g_raid_subdisk *sd)
153 {
154 	struct g_raid_tr_raid1_object *trs;
155 	struct g_raid_softc *sc;
156 	struct g_raid_subdisk *tsd, *bestsd;
157 	u_int s;
158 	int i, na, ns;
159 
160 	sc = vol->v_softc;
161 	trs = (struct g_raid_tr_raid1_object *)vol->v_tr;
162 	if (trs->trso_stopping &&
163 	    (trs->trso_flags & TR_RAID1_F_DOING_SOME) == 0)
164 		s = G_RAID_VOLUME_S_STOPPED;
165 	else if (trs->trso_starting)
166 		s = G_RAID_VOLUME_S_STARTING;
167 	else {
168 		/* Make sure we have at least one ACTIVE disk. */
169 		na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
170 		if (na == 0) {
171 			/*
172 			 * Critical situation! We have no any active disk!
173 			 * Choose the best disk we have to make it active.
174 			 */
175 			bestsd = &vol->v_subdisks[0];
176 			for (i = 1; i < vol->v_disks_count; i++) {
177 				tsd = &vol->v_subdisks[i];
178 				if (tsd->sd_state > bestsd->sd_state)
179 					bestsd = tsd;
180 				else if (tsd->sd_state == bestsd->sd_state &&
181 				    (tsd->sd_state == G_RAID_SUBDISK_S_REBUILD ||
182 				     tsd->sd_state == G_RAID_SUBDISK_S_RESYNC) &&
183 				    tsd->sd_rebuild_pos > bestsd->sd_rebuild_pos)
184 					bestsd = tsd;
185 			}
186 			if (bestsd->sd_state >= G_RAID_SUBDISK_S_UNINITIALIZED) {
187 				/* We found reasonable candidate. */
188 				G_RAID_DEBUG1(1, sc,
189 				    "Promote subdisk %s:%d from %s to ACTIVE.",
190 				    vol->v_name, bestsd->sd_pos,
191 				    g_raid_subdisk_state2str(bestsd->sd_state));
192 				g_raid_change_subdisk_state(bestsd,
193 				    G_RAID_SUBDISK_S_ACTIVE);
194 				g_raid_write_metadata(sc,
195 				    vol, bestsd, bestsd->sd_disk);
196 			}
197 		}
198 		na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
199 		ns = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
200 		     g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
201 		if (na == vol->v_disks_count)
202 			s = G_RAID_VOLUME_S_OPTIMAL;
203 		else if (na + ns == vol->v_disks_count)
204 			s = G_RAID_VOLUME_S_SUBOPTIMAL;
205 		else if (na > 0)
206 			s = G_RAID_VOLUME_S_DEGRADED;
207 		else
208 			s = G_RAID_VOLUME_S_BROKEN;
209 		g_raid_tr_raid1_maybe_rebuild(vol->v_tr, sd);
210 	}
211 	if (s != vol->v_state) {
212 		g_raid_event_send(vol, G_RAID_VOLUME_S_ALIVE(s) ?
213 		    G_RAID_VOLUME_E_UP : G_RAID_VOLUME_E_DOWN,
214 		    G_RAID_EVENT_VOLUME);
215 		g_raid_change_volume_state(vol, s);
216 		if (!trs->trso_starting && !trs->trso_stopping)
217 			g_raid_write_metadata(sc, vol, NULL, NULL);
218 	}
219 	return (0);
220 }
221 
222 static void
223 g_raid_tr_raid1_fail_disk(struct g_raid_softc *sc, struct g_raid_subdisk *sd,
224     struct g_raid_disk *disk)
225 {
226 	/*
227 	 * We don't fail the last disk in the pack, since it still has decent
228 	 * data on it and that's better than failing the disk if it is the root
229 	 * file system.
230 	 *
231 	 * XXX should this be controlled via a tunable?  It makes sense for
232 	 * the volume that has / on it.  I can't think of a case where we'd
233 	 * want the volume to go away on this kind of event.
234 	 */
235 	if (g_raid_nsubdisks(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == 1 &&
236 	    g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == sd)
237 		return;
238 	g_raid_fail_disk(sc, sd, disk);
239 }
240 
241 static void
242 g_raid_tr_raid1_rebuild_some(struct g_raid_tr_object *tr)
243 {
244 	struct g_raid_tr_raid1_object *trs;
245 	struct g_raid_subdisk *sd, *good_sd;
246 	struct bio *bp;
247 
248 	trs = (struct g_raid_tr_raid1_object *)tr;
249 	if (trs->trso_flags & TR_RAID1_F_DOING_SOME)
250 		return;
251 	sd = trs->trso_failed_sd;
252 	good_sd = g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE);
253 	if (good_sd == NULL) {
254 		g_raid_tr_raid1_rebuild_abort(tr);
255 		return;
256 	}
257 	bp = &trs->trso_bio;
258 	memset(bp, 0, sizeof(*bp));
259 	bp->bio_offset = sd->sd_rebuild_pos;
260 	bp->bio_length = MIN(g_raid1_rebuild_slab,
261 	    sd->sd_size - sd->sd_rebuild_pos);
262 	bp->bio_data = trs->trso_buffer;
263 	bp->bio_cmd = BIO_READ;
264 	bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
265 	bp->bio_caller1 = good_sd;
266 	trs->trso_flags |= TR_RAID1_F_DOING_SOME;
267 	trs->trso_flags |= TR_RAID1_F_LOCKED;
268 	g_raid_lock_range(sd->sd_volume,	/* Lock callback starts I/O */
269 	   bp->bio_offset, bp->bio_length, NULL, bp);
270 }
271 
272 static void
273 g_raid_tr_raid1_rebuild_done(struct g_raid_tr_raid1_object *trs)
274 {
275 	struct g_raid_volume *vol;
276 	struct g_raid_subdisk *sd;
277 
278 	vol = trs->trso_base.tro_volume;
279 	sd = trs->trso_failed_sd;
280 	g_raid_write_metadata(vol->v_softc, vol, sd, sd->sd_disk);
281 	free(trs->trso_buffer, M_TR_RAID1);
282 	trs->trso_buffer = NULL;
283 	trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
284 	trs->trso_type = TR_RAID1_NONE;
285 	trs->trso_recover_slabs = 0;
286 	trs->trso_failed_sd = NULL;
287 	g_raid_tr_update_state_raid1(vol, NULL);
288 }
289 
290 static void
291 g_raid_tr_raid1_rebuild_finish(struct g_raid_tr_object *tr)
292 {
293 	struct g_raid_tr_raid1_object *trs;
294 	struct g_raid_subdisk *sd;
295 
296 	trs = (struct g_raid_tr_raid1_object *)tr;
297 	sd = trs->trso_failed_sd;
298 	G_RAID_DEBUG1(0, tr->tro_volume->v_softc,
299 	    "Subdisk %s:%d-%s rebuild completed.",
300 	    sd->sd_volume->v_name, sd->sd_pos,
301 	    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
302 	g_raid_change_subdisk_state(sd, G_RAID_SUBDISK_S_ACTIVE);
303 	sd->sd_rebuild_pos = 0;
304 	g_raid_tr_raid1_rebuild_done(trs);
305 }
306 
307 static void
308 g_raid_tr_raid1_rebuild_abort(struct g_raid_tr_object *tr)
309 {
310 	struct g_raid_tr_raid1_object *trs;
311 	struct g_raid_subdisk *sd;
312 	struct g_raid_volume *vol;
313 	off_t len;
314 
315 	vol = tr->tro_volume;
316 	trs = (struct g_raid_tr_raid1_object *)tr;
317 	sd = trs->trso_failed_sd;
318 	if (trs->trso_flags & TR_RAID1_F_DOING_SOME) {
319 		G_RAID_DEBUG1(1, vol->v_softc,
320 		    "Subdisk %s:%d-%s rebuild is aborting.",
321 		    sd->sd_volume->v_name, sd->sd_pos,
322 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
323 		trs->trso_flags |= TR_RAID1_F_ABORT;
324 	} else {
325 		G_RAID_DEBUG1(0, vol->v_softc,
326 		    "Subdisk %s:%d-%s rebuild aborted.",
327 		    sd->sd_volume->v_name, sd->sd_pos,
328 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
329 		trs->trso_flags &= ~TR_RAID1_F_ABORT;
330 		if (trs->trso_flags & TR_RAID1_F_LOCKED) {
331 			trs->trso_flags &= ~TR_RAID1_F_LOCKED;
332 			len = MIN(g_raid1_rebuild_slab,
333 			    sd->sd_size - sd->sd_rebuild_pos);
334 			g_raid_unlock_range(tr->tro_volume,
335 			    sd->sd_rebuild_pos, len);
336 		}
337 		g_raid_tr_raid1_rebuild_done(trs);
338 	}
339 }
340 
341 static void
342 g_raid_tr_raid1_rebuild_start(struct g_raid_tr_object *tr)
343 {
344 	struct g_raid_volume *vol;
345 	struct g_raid_tr_raid1_object *trs;
346 	struct g_raid_subdisk *sd, *fsd;
347 
348 	vol = tr->tro_volume;
349 	trs = (struct g_raid_tr_raid1_object *)tr;
350 	if (trs->trso_failed_sd) {
351 		G_RAID_DEBUG1(1, vol->v_softc,
352 		    "Already rebuild in start rebuild. pos %jd\n",
353 		    (intmax_t)trs->trso_failed_sd->sd_rebuild_pos);
354 		return;
355 	}
356 	sd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_ACTIVE);
357 	if (sd == NULL) {
358 		G_RAID_DEBUG1(1, vol->v_softc,
359 		    "No active disk to rebuild.  night night.");
360 		return;
361 	}
362 	fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_RESYNC);
363 	if (fsd == NULL)
364 		fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_REBUILD);
365 	if (fsd == NULL) {
366 		fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_STALE);
367 		if (fsd != NULL) {
368 			fsd->sd_rebuild_pos = 0;
369 			g_raid_change_subdisk_state(fsd,
370 			    G_RAID_SUBDISK_S_RESYNC);
371 			g_raid_write_metadata(vol->v_softc, vol, fsd, NULL);
372 		} else {
373 			fsd = g_raid_get_subdisk(vol,
374 			    G_RAID_SUBDISK_S_UNINITIALIZED);
375 			if (fsd == NULL)
376 				fsd = g_raid_get_subdisk(vol,
377 				    G_RAID_SUBDISK_S_NEW);
378 			if (fsd != NULL) {
379 				fsd->sd_rebuild_pos = 0;
380 				g_raid_change_subdisk_state(fsd,
381 				    G_RAID_SUBDISK_S_REBUILD);
382 				g_raid_write_metadata(vol->v_softc,
383 				    vol, fsd, NULL);
384 			}
385 		}
386 	}
387 	if (fsd == NULL) {
388 		G_RAID_DEBUG1(1, vol->v_softc,
389 		    "No failed disk to rebuild.  night night.");
390 		return;
391 	}
392 	trs->trso_failed_sd = fsd;
393 	G_RAID_DEBUG1(0, vol->v_softc,
394 	    "Subdisk %s:%d-%s rebuild start at %jd.",
395 	    fsd->sd_volume->v_name, fsd->sd_pos,
396 	    fsd->sd_disk ? g_raid_get_diskname(fsd->sd_disk) : "[none]",
397 	    trs->trso_failed_sd->sd_rebuild_pos);
398 	trs->trso_type = TR_RAID1_REBUILD;
399 	trs->trso_buffer = malloc(g_raid1_rebuild_slab, M_TR_RAID1, M_WAITOK);
400 	trs->trso_meta_update = g_raid1_rebuild_meta_update;
401 	g_raid_tr_raid1_rebuild_some(tr);
402 }
403 
404 static void
405 g_raid_tr_raid1_maybe_rebuild(struct g_raid_tr_object *tr,
406     struct g_raid_subdisk *sd)
407 {
408 	struct g_raid_volume *vol;
409 	struct g_raid_tr_raid1_object *trs;
410 	int na, nr;
411 
412 	/*
413 	 * If we're stopping, don't do anything.  If we don't have at least one
414 	 * good disk and one bad disk, we don't do anything.  And if there's a
415 	 * 'good disk' stored in the trs, then we're in progress and we punt.
416 	 * If we make it past all these checks, we need to rebuild.
417 	 */
418 	vol = tr->tro_volume;
419 	trs = (struct g_raid_tr_raid1_object *)tr;
420 	if (trs->trso_stopping)
421 		return;
422 	na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
423 	nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_REBUILD) +
424 	    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
425 	switch(trs->trso_type) {
426 	case TR_RAID1_NONE:
427 		if (na == 0)
428 			return;
429 		if (nr == 0) {
430 			nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_NEW) +
431 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
432 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_UNINITIALIZED);
433 			if (nr == 0)
434 				return;
435 		}
436 		g_raid_tr_raid1_rebuild_start(tr);
437 		break;
438 	case TR_RAID1_REBUILD:
439 		if (na == 0 || nr == 0 || trs->trso_failed_sd == sd)
440 			g_raid_tr_raid1_rebuild_abort(tr);
441 		break;
442 	case TR_RAID1_RESYNC:
443 		break;
444 	}
445 }
446 
447 static int
448 g_raid_tr_event_raid1(struct g_raid_tr_object *tr,
449     struct g_raid_subdisk *sd, u_int event)
450 {
451 
452 	g_raid_tr_update_state_raid1(tr->tro_volume, sd);
453 	return (0);
454 }
455 
456 static int
457 g_raid_tr_start_raid1(struct g_raid_tr_object *tr)
458 {
459 	struct g_raid_tr_raid1_object *trs;
460 	struct g_raid_volume *vol;
461 
462 	trs = (struct g_raid_tr_raid1_object *)tr;
463 	vol = tr->tro_volume;
464 	trs->trso_starting = 0;
465 	g_raid_tr_update_state_raid1(vol, NULL);
466 	return (0);
467 }
468 
469 static int
470 g_raid_tr_stop_raid1(struct g_raid_tr_object *tr)
471 {
472 	struct g_raid_tr_raid1_object *trs;
473 	struct g_raid_volume *vol;
474 
475 	trs = (struct g_raid_tr_raid1_object *)tr;
476 	vol = tr->tro_volume;
477 	trs->trso_starting = 0;
478 	trs->trso_stopping = 1;
479 	g_raid_tr_update_state_raid1(vol, NULL);
480 	return (0);
481 }
482 
483 /*
484  * Select the disk to read from.  Take into account: subdisk state, running
485  * error recovery, average disk load, head position and possible cache hits.
486  */
487 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
488 static struct g_raid_subdisk *
489 g_raid_tr_raid1_select_read_disk(struct g_raid_volume *vol, struct bio *bp,
490     u_int mask)
491 {
492 	struct g_raid_subdisk *sd, *best;
493 	int i, prio, bestprio;
494 
495 	best = NULL;
496 	bestprio = INT_MAX;
497 	for (i = 0; i < vol->v_disks_count; i++) {
498 		sd = &vol->v_subdisks[i];
499 		if (sd->sd_state != G_RAID_SUBDISK_S_ACTIVE &&
500 		    ((sd->sd_state != G_RAID_SUBDISK_S_REBUILD &&
501 		      sd->sd_state != G_RAID_SUBDISK_S_RESYNC) ||
502 		     bp->bio_offset + bp->bio_length > sd->sd_rebuild_pos))
503 			continue;
504 		if ((mask & (1 << i)) != 0)
505 			continue;
506 		prio = G_RAID_SUBDISK_LOAD(sd);
507 		prio += min(sd->sd_recovery, 255) << 22;
508 		prio += (G_RAID_SUBDISK_S_ACTIVE - sd->sd_state) << 16;
509 		/* If disk head is precisely in position - highly prefer it. */
510 		if (G_RAID_SUBDISK_POS(sd) == bp->bio_offset)
511 			prio -= 2 * G_RAID_SUBDISK_LOAD_SCALE;
512 		else
513 		/* If disk head is close to position - prefer it. */
514 		if (ABS(G_RAID_SUBDISK_POS(sd) - bp->bio_offset) <
515 		    G_RAID_SUBDISK_TRACK_SIZE)
516 			prio -= 1 * G_RAID_SUBDISK_LOAD_SCALE;
517 		if (prio < bestprio) {
518 			best = sd;
519 			bestprio = prio;
520 		}
521 	}
522 	return (best);
523 }
524 
525 static void
526 g_raid_tr_iostart_raid1_read(struct g_raid_tr_object *tr, struct bio *bp)
527 {
528 	struct g_raid_subdisk *sd;
529 	struct bio *cbp;
530 
531 	sd = g_raid_tr_raid1_select_read_disk(tr->tro_volume, bp, 0);
532 	KASSERT(sd != NULL, ("No active disks in volume %s.",
533 		tr->tro_volume->v_name));
534 
535 	cbp = g_clone_bio(bp);
536 	if (cbp == NULL) {
537 		g_raid_iodone(bp, ENOMEM);
538 		return;
539 	}
540 
541 	g_raid_subdisk_iostart(sd, cbp);
542 }
543 
544 static void
545 g_raid_tr_iostart_raid1_write(struct g_raid_tr_object *tr, struct bio *bp)
546 {
547 	struct g_raid_volume *vol;
548 	struct g_raid_subdisk *sd;
549 	struct bio_queue_head queue;
550 	struct bio *cbp;
551 	int i;
552 
553 	vol = tr->tro_volume;
554 
555 	/*
556 	 * Allocate all bios before sending any request, so we can return
557 	 * ENOMEM in nice and clean way.
558 	 */
559 	bioq_init(&queue);
560 	for (i = 0; i < vol->v_disks_count; i++) {
561 		sd = &vol->v_subdisks[i];
562 		switch (sd->sd_state) {
563 		case G_RAID_SUBDISK_S_ACTIVE:
564 			break;
565 		case G_RAID_SUBDISK_S_REBUILD:
566 			/*
567 			 * When rebuilding, only part of this subdisk is
568 			 * writable, the rest will be written as part of the
569 			 * that process.
570 			 */
571 			if (bp->bio_offset >= sd->sd_rebuild_pos)
572 				continue;
573 			break;
574 		case G_RAID_SUBDISK_S_STALE:
575 		case G_RAID_SUBDISK_S_RESYNC:
576 			/*
577 			 * Resyncing still writes on the theory that the
578 			 * resync'd disk is very close and writing it will
579 			 * keep it that way better if we keep up while
580 			 * resyncing.
581 			 */
582 			break;
583 		default:
584 			continue;
585 		}
586 		cbp = g_clone_bio(bp);
587 		if (cbp == NULL)
588 			goto failure;
589 		cbp->bio_caller1 = sd;
590 		bioq_insert_tail(&queue, cbp);
591 	}
592 	while ((cbp = bioq_takefirst(&queue)) != NULL) {
593 		sd = cbp->bio_caller1;
594 		cbp->bio_caller1 = NULL;
595 		g_raid_subdisk_iostart(sd, cbp);
596 	}
597 	return;
598 failure:
599 	while ((cbp = bioq_takefirst(&queue)) != NULL)
600 		g_destroy_bio(cbp);
601 	if (bp->bio_error == 0)
602 		bp->bio_error = ENOMEM;
603 	g_raid_iodone(bp, bp->bio_error);
604 }
605 
606 static void
607 g_raid_tr_iostart_raid1(struct g_raid_tr_object *tr, struct bio *bp)
608 {
609 	struct g_raid_volume *vol;
610 	struct g_raid_tr_raid1_object *trs;
611 
612 	vol = tr->tro_volume;
613 	trs = (struct g_raid_tr_raid1_object *)tr;
614 	if (vol->v_state != G_RAID_VOLUME_S_OPTIMAL &&
615 	    vol->v_state != G_RAID_VOLUME_S_SUBOPTIMAL &&
616 	    vol->v_state != G_RAID_VOLUME_S_DEGRADED) {
617 		g_raid_iodone(bp, EIO);
618 		return;
619 	}
620 	/*
621 	 * If we're rebuilding, squeeze in rebuild activity every so often,
622 	 * even when the disk is busy.  Be sure to only count real I/O
623 	 * to the disk.  All 'SPECIAL' I/O is traffic generated to the disk
624 	 * by this module.
625 	 */
626 	if (trs->trso_failed_sd != NULL &&
627 	    !(bp->bio_cflags & G_RAID_BIO_FLAG_SPECIAL)) {
628 		/* Make this new or running now round short. */
629 		trs->trso_recover_slabs = 0;
630 		if (--trs->trso_fair_io <= 0) {
631 			trs->trso_fair_io = g_raid1_rebuild_fair_io;
632 			g_raid_tr_raid1_rebuild_some(tr);
633 		}
634 	}
635 	switch (bp->bio_cmd) {
636 	case BIO_READ:
637 		g_raid_tr_iostart_raid1_read(tr, bp);
638 		break;
639 	case BIO_WRITE:
640 	case BIO_DELETE:
641 		g_raid_tr_iostart_raid1_write(tr, bp);
642 		break;
643 	case BIO_SPEEDUP:
644 	case BIO_FLUSH:
645 		g_raid_tr_flush_common(tr, bp);
646 		break;
647 	default:
648 		KASSERT(1 == 0, ("Invalid command here: %u (volume=%s)",
649 		    bp->bio_cmd, vol->v_name));
650 		break;
651 	}
652 }
653 
654 static void
655 g_raid_tr_iodone_raid1(struct g_raid_tr_object *tr,
656     struct g_raid_subdisk *sd, struct bio *bp)
657 {
658 	struct bio *cbp;
659 	struct g_raid_subdisk *nsd;
660 	struct g_raid_volume *vol;
661 	struct bio *pbp;
662 	struct g_raid_tr_raid1_object *trs;
663 	uintptr_t *mask;
664 	int error, do_write;
665 
666 	trs = (struct g_raid_tr_raid1_object *)tr;
667 	vol = tr->tro_volume;
668 	if (bp->bio_cflags & G_RAID_BIO_FLAG_SYNC) {
669 		/*
670 		 * This operation is part of a rebuild or resync operation.
671 		 * See what work just got done, then schedule the next bit of
672 		 * work, if any.  Rebuild/resync is done a little bit at a
673 		 * time.  Either when a timeout happens, or after we get a
674 		 * bunch of I/Os to the disk (to make sure an active system
675 		 * will complete in a sane amount of time).
676 		 *
677 		 * We are setup to do differing amounts of work for each of
678 		 * these cases.  so long as the slabs is smallish (less than
679 		 * 50 or so, I'd guess, but that's just a WAG), we shouldn't
680 		 * have any bio starvation issues.  For active disks, we do
681 		 * 5MB of data, for inactive ones, we do 50MB.
682 		 */
683 		if (trs->trso_type == TR_RAID1_REBUILD) {
684 			if (bp->bio_cmd == BIO_READ) {
685 				/* Immediately abort rebuild, if requested. */
686 				if (trs->trso_flags & TR_RAID1_F_ABORT) {
687 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
688 					g_raid_tr_raid1_rebuild_abort(tr);
689 					return;
690 				}
691 
692 				/* On read error, skip and cross fingers. */
693 				if (bp->bio_error != 0) {
694 					G_RAID_LOGREQ(0, bp,
695 					    "Read error during rebuild (%d), "
696 					    "possible data loss!",
697 					    bp->bio_error);
698 					goto rebuild_round_done;
699 				}
700 
701 				/*
702 				 * The read operation finished, queue the
703 				 * write and get out.
704 				 */
705 				G_RAID_LOGREQ(4, bp, "rebuild read done. %d",
706 				    bp->bio_error);
707 				bp->bio_cmd = BIO_WRITE;
708 				bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
709 				G_RAID_LOGREQ(4, bp, "Queueing rebuild write.");
710 				g_raid_subdisk_iostart(trs->trso_failed_sd, bp);
711 			} else {
712 				/*
713 				 * The write operation just finished.  Do
714 				 * another.  We keep cloning the master bio
715 				 * since it has the right buffers allocated to
716 				 * it.
717 				 */
718 				G_RAID_LOGREQ(4, bp,
719 				    "rebuild write done. Error %d",
720 				    bp->bio_error);
721 				nsd = trs->trso_failed_sd;
722 				if (bp->bio_error != 0 ||
723 				    trs->trso_flags & TR_RAID1_F_ABORT) {
724 					if ((trs->trso_flags &
725 					    TR_RAID1_F_ABORT) == 0) {
726 						g_raid_tr_raid1_fail_disk(sd->sd_softc,
727 						    nsd, nsd->sd_disk);
728 					}
729 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
730 					g_raid_tr_raid1_rebuild_abort(tr);
731 					return;
732 				}
733 rebuild_round_done:
734 				nsd = trs->trso_failed_sd;
735 				trs->trso_flags &= ~TR_RAID1_F_LOCKED;
736 				g_raid_unlock_range(sd->sd_volume,
737 				    bp->bio_offset, bp->bio_length);
738 				nsd->sd_rebuild_pos += bp->bio_length;
739 				if (nsd->sd_rebuild_pos >= nsd->sd_size) {
740 					g_raid_tr_raid1_rebuild_finish(tr);
741 					return;
742 				}
743 
744 				/* Abort rebuild if we are stopping */
745 				if (trs->trso_stopping) {
746 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
747 					g_raid_tr_raid1_rebuild_abort(tr);
748 					return;
749 				}
750 
751 				if (--trs->trso_meta_update <= 0) {
752 					g_raid_write_metadata(vol->v_softc,
753 					    vol, nsd, nsd->sd_disk);
754 					trs->trso_meta_update =
755 					    g_raid1_rebuild_meta_update;
756 				}
757 				trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
758 				if (--trs->trso_recover_slabs <= 0)
759 					return;
760 				g_raid_tr_raid1_rebuild_some(tr);
761 			}
762 		} else if (trs->trso_type == TR_RAID1_RESYNC) {
763 			/*
764 			 * read good sd, read bad sd in parallel.  when both
765 			 * done, compare the buffers.  write good to the bad
766 			 * if different.  do the next bit of work.
767 			 */
768 			panic("Somehow, we think we're doing a resync");
769 		}
770 		return;
771 	}
772 	pbp = bp->bio_parent;
773 	pbp->bio_inbed++;
774 	if (bp->bio_cmd == BIO_READ && bp->bio_error != 0) {
775 		/*
776 		 * Read failed on first drive.  Retry the read error on
777 		 * another disk drive, if available, before erroring out the
778 		 * read.
779 		 */
780 		sd->sd_disk->d_read_errs++;
781 		G_RAID_LOGREQ(0, bp,
782 		    "Read error (%d), %d read errors total",
783 		    bp->bio_error, sd->sd_disk->d_read_errs);
784 
785 		/*
786 		 * If there are too many read errors, we move to degraded.
787 		 * XXX Do we want to FAIL the drive (eg, make the user redo
788 		 * everything to get it back in sync), or just degrade the
789 		 * drive, which kicks off a resync?
790 		 */
791 		do_write = 1;
792 		if (sd->sd_disk->d_read_errs > g_raid_read_err_thresh) {
793 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
794 			if (pbp->bio_children == 1)
795 				do_write = 0;
796 		}
797 
798 		/*
799 		 * Find the other disk, and try to do the I/O to it.
800 		 */
801 		mask = (uintptr_t *)(&pbp->bio_driver2);
802 		if (pbp->bio_children == 1) {
803 			/* Save original subdisk. */
804 			pbp->bio_driver1 = do_write ? sd : NULL;
805 			*mask = 0;
806 		}
807 		*mask |= 1 << sd->sd_pos;
808 		nsd = g_raid_tr_raid1_select_read_disk(vol, pbp, *mask);
809 		if (nsd != NULL && (cbp = g_clone_bio(pbp)) != NULL) {
810 			g_destroy_bio(bp);
811 			G_RAID_LOGREQ(2, cbp, "Retrying read from %d",
812 			    nsd->sd_pos);
813 			if (pbp->bio_children == 2 && do_write) {
814 				sd->sd_recovery++;
815 				cbp->bio_caller1 = nsd;
816 				pbp->bio_pflags = G_RAID_BIO_FLAG_LOCKED;
817 				/* Lock callback starts I/O */
818 				g_raid_lock_range(sd->sd_volume,
819 				    cbp->bio_offset, cbp->bio_length, pbp, cbp);
820 			} else {
821 				g_raid_subdisk_iostart(nsd, cbp);
822 			}
823 			return;
824 		}
825 		/*
826 		 * We can't retry.  Return the original error by falling
827 		 * through.  This will happen when there's only one good disk.
828 		 * We don't need to fail the raid, since its actual state is
829 		 * based on the state of the subdisks.
830 		 */
831 		G_RAID_LOGREQ(2, bp, "Couldn't retry read, failing it");
832 	}
833 	if (bp->bio_cmd == BIO_READ &&
834 	    bp->bio_error == 0 &&
835 	    pbp->bio_children > 1 &&
836 	    pbp->bio_driver1 != NULL) {
837 		/*
838 		 * If it was a read, and bio_children is >1, then we just
839 		 * recovered the data from the second drive.  We should try to
840 		 * write that data to the first drive if sector remapping is
841 		 * enabled.  A write should put the data in a new place on the
842 		 * disk, remapping the bad sector.  Do we need to do that by
843 		 * queueing a request to the main worker thread?  It doesn't
844 		 * affect the return code of this current read, and can be
845 		 * done at our leisure.  However, to make the code simpler, it
846 		 * is done synchronously.
847 		 */
848 		G_RAID_LOGREQ(3, bp, "Recovered data from other drive");
849 		cbp = g_clone_bio(pbp);
850 		if (cbp != NULL) {
851 			g_destroy_bio(bp);
852 			cbp->bio_cmd = BIO_WRITE;
853 			cbp->bio_cflags = G_RAID_BIO_FLAG_REMAP;
854 			G_RAID_LOGREQ(2, cbp,
855 			    "Attempting bad sector remap on failing drive.");
856 			g_raid_subdisk_iostart(pbp->bio_driver1, cbp);
857 			return;
858 		}
859 	}
860 	if (pbp->bio_pflags & G_RAID_BIO_FLAG_LOCKED) {
861 		/*
862 		 * We're done with a recovery, mark the range as unlocked.
863 		 * For any write errors, we aggressively fail the disk since
864 		 * there was both a READ and a WRITE error at this location.
865 		 * Both types of errors generally indicates the drive is on
866 		 * the verge of total failure anyway.  Better to stop trusting
867 		 * it now.  However, we need to reset error to 0 in that case
868 		 * because we're not failing the original I/O which succeeded.
869 		 */
870 		if (bp->bio_cmd == BIO_WRITE && bp->bio_error) {
871 			G_RAID_LOGREQ(0, bp, "Remap write failed: "
872 			    "failing subdisk.");
873 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
874 			bp->bio_error = 0;
875 		}
876 		if (pbp->bio_driver1 != NULL) {
877 			((struct g_raid_subdisk *)pbp->bio_driver1)
878 			    ->sd_recovery--;
879 		}
880 		G_RAID_LOGREQ(2, bp, "REMAP done %d.", bp->bio_error);
881 		g_raid_unlock_range(sd->sd_volume, bp->bio_offset,
882 		    bp->bio_length);
883 	}
884 	if (pbp->bio_cmd != BIO_READ) {
885 		if (pbp->bio_inbed == 1 || pbp->bio_error != 0)
886 			pbp->bio_error = bp->bio_error;
887 		if (pbp->bio_cmd == BIO_WRITE && bp->bio_error != 0) {
888 			G_RAID_LOGREQ(0, bp, "Write failed: failing subdisk.");
889 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
890 		}
891 		error = pbp->bio_error;
892 	} else
893 		error = bp->bio_error;
894 	g_destroy_bio(bp);
895 	if (pbp->bio_children == pbp->bio_inbed) {
896 		pbp->bio_completed = pbp->bio_length;
897 		g_raid_iodone(pbp, error);
898 	}
899 }
900 
901 static int
902 g_raid_tr_kerneldump_raid1(struct g_raid_tr_object *tr, void *virtual,
903     off_t offset, size_t length)
904 {
905 	struct g_raid_volume *vol;
906 	struct g_raid_subdisk *sd;
907 	int error, i, ok;
908 
909 	vol = tr->tro_volume;
910 	error = 0;
911 	ok = 0;
912 	for (i = 0; i < vol->v_disks_count; i++) {
913 		sd = &vol->v_subdisks[i];
914 		switch (sd->sd_state) {
915 		case G_RAID_SUBDISK_S_ACTIVE:
916 			break;
917 		case G_RAID_SUBDISK_S_REBUILD:
918 			/*
919 			 * When rebuilding, only part of this subdisk is
920 			 * writable, the rest will be written as part of the
921 			 * that process.
922 			 */
923 			if (offset >= sd->sd_rebuild_pos)
924 				continue;
925 			break;
926 		case G_RAID_SUBDISK_S_STALE:
927 		case G_RAID_SUBDISK_S_RESYNC:
928 			/*
929 			 * Resyncing still writes on the theory that the
930 			 * resync'd disk is very close and writing it will
931 			 * keep it that way better if we keep up while
932 			 * resyncing.
933 			 */
934 			break;
935 		default:
936 			continue;
937 		}
938 		error = g_raid_subdisk_kerneldump(sd, virtual, offset, length);
939 		if (error == 0)
940 			ok++;
941 	}
942 	return (ok > 0 ? 0 : error);
943 }
944 
945 static int
946 g_raid_tr_locked_raid1(struct g_raid_tr_object *tr, void *argp)
947 {
948 	struct bio *bp;
949 	struct g_raid_subdisk *sd;
950 
951 	bp = (struct bio *)argp;
952 	sd = (struct g_raid_subdisk *)bp->bio_caller1;
953 	g_raid_subdisk_iostart(sd, bp);
954 
955 	return (0);
956 }
957 
958 static int
959 g_raid_tr_idle_raid1(struct g_raid_tr_object *tr)
960 {
961 	struct g_raid_tr_raid1_object *trs;
962 
963 	trs = (struct g_raid_tr_raid1_object *)tr;
964 	trs->trso_fair_io = g_raid1_rebuild_fair_io;
965 	trs->trso_recover_slabs = g_raid1_rebuild_cluster_idle;
966 	if (trs->trso_type == TR_RAID1_REBUILD)
967 		g_raid_tr_raid1_rebuild_some(tr);
968 	return (0);
969 }
970 
971 static int
972 g_raid_tr_free_raid1(struct g_raid_tr_object *tr)
973 {
974 	struct g_raid_tr_raid1_object *trs;
975 
976 	trs = (struct g_raid_tr_raid1_object *)tr;
977 
978 	if (trs->trso_buffer != NULL) {
979 		free(trs->trso_buffer, M_TR_RAID1);
980 		trs->trso_buffer = NULL;
981 	}
982 	return (0);
983 }
984 
985 G_RAID_TR_DECLARE(raid1, "RAID1");
986