xref: /freebsd/sys/geom/raid/tr_raid1.c (revision 9a14aa017b21c292740c00ee098195cd46642730)
1 /*-
2  * Copyright (c) 2010 Alexander Motin <mav@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/bio.h>
32 #include <sys/endian.h>
33 #include <sys/kernel.h>
34 #include <sys/kobj.h>
35 #include <sys/limits.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/mutex.h>
39 #include <sys/sysctl.h>
40 #include <sys/systm.h>
41 #include <geom/geom.h>
42 #include "geom/raid/g_raid.h"
43 #include "g_raid_tr_if.h"
44 
45 SYSCTL_DECL(_kern_geom_raid);
46 static SYSCTL_NODE(_kern_geom_raid, OID_AUTO, raid1, CTLFLAG_RW, 0,
47     "RAID1 parameters");
48 
49 #define RAID1_REBUILD_SLAB	(1 << 20) /* One transation in a rebuild */
50 static int g_raid1_rebuild_slab = RAID1_REBUILD_SLAB;
51 TUNABLE_INT("kern.geom.raid.raid1.rebuild_slab_size",
52     &g_raid1_rebuild_slab);
53 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_slab_size, CTLFLAG_RW,
54     &g_raid1_rebuild_slab, 0,
55     "Amount of the disk to rebuild each read/write cycle of the rebuild.");
56 
57 #define RAID1_REBUILD_FAIR_IO 20 /* use 1/x of the available I/O */
58 static int g_raid1_rebuild_fair_io = RAID1_REBUILD_FAIR_IO;
59 TUNABLE_INT("kern.geom.raid.raid1.rebuild_fair_io",
60     &g_raid1_rebuild_fair_io);
61 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_fair_io, CTLFLAG_RW,
62     &g_raid1_rebuild_fair_io, 0,
63     "Fraction of the I/O bandwidth to use when disk busy for rebuild.");
64 
65 #define RAID1_REBUILD_CLUSTER_IDLE 100
66 static int g_raid1_rebuild_cluster_idle = RAID1_REBUILD_CLUSTER_IDLE;
67 TUNABLE_INT("kern.geom.raid.raid1.rebuild_cluster_idle",
68     &g_raid1_rebuild_cluster_idle);
69 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_cluster_idle, CTLFLAG_RW,
70     &g_raid1_rebuild_cluster_idle, 0,
71     "Number of slabs to do each time we trigger a rebuild cycle");
72 
73 #define RAID1_REBUILD_META_UPDATE 1024 /* update meta data every 1GB or so */
74 static int g_raid1_rebuild_meta_update = RAID1_REBUILD_META_UPDATE;
75 TUNABLE_INT("kern.geom.raid.raid1.rebuild_meta_update",
76     &g_raid1_rebuild_meta_update);
77 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_meta_update, CTLFLAG_RW,
78     &g_raid1_rebuild_meta_update, 0,
79     "When to update the meta data.");
80 
81 static MALLOC_DEFINE(M_TR_RAID1, "tr_raid1_data", "GEOM_RAID RAID1 data");
82 
83 #define TR_RAID1_NONE 0
84 #define TR_RAID1_REBUILD 1
85 #define TR_RAID1_RESYNC 2
86 
87 #define TR_RAID1_F_DOING_SOME	0x1
88 #define TR_RAID1_F_LOCKED	0x2
89 #define TR_RAID1_F_ABORT	0x4
90 
91 struct g_raid_tr_raid1_object {
92 	struct g_raid_tr_object	 trso_base;
93 	int			 trso_starting;
94 	int			 trso_stopping;
95 	int			 trso_type;
96 	int			 trso_recover_slabs; /* slabs before rest */
97 	int			 trso_fair_io;
98 	int			 trso_meta_update;
99 	int			 trso_flags;
100 	struct g_raid_subdisk	*trso_failed_sd; /* like per volume */
101 	void			*trso_buffer;	 /* Buffer space */
102 	struct bio		 trso_bio;
103 };
104 
105 static g_raid_tr_taste_t g_raid_tr_taste_raid1;
106 static g_raid_tr_event_t g_raid_tr_event_raid1;
107 static g_raid_tr_start_t g_raid_tr_start_raid1;
108 static g_raid_tr_stop_t g_raid_tr_stop_raid1;
109 static g_raid_tr_iostart_t g_raid_tr_iostart_raid1;
110 static g_raid_tr_iodone_t g_raid_tr_iodone_raid1;
111 static g_raid_tr_kerneldump_t g_raid_tr_kerneldump_raid1;
112 static g_raid_tr_locked_t g_raid_tr_locked_raid1;
113 static g_raid_tr_idle_t g_raid_tr_idle_raid1;
114 static g_raid_tr_free_t g_raid_tr_free_raid1;
115 
116 static kobj_method_t g_raid_tr_raid1_methods[] = {
117 	KOBJMETHOD(g_raid_tr_taste,	g_raid_tr_taste_raid1),
118 	KOBJMETHOD(g_raid_tr_event,	g_raid_tr_event_raid1),
119 	KOBJMETHOD(g_raid_tr_start,	g_raid_tr_start_raid1),
120 	KOBJMETHOD(g_raid_tr_stop,	g_raid_tr_stop_raid1),
121 	KOBJMETHOD(g_raid_tr_iostart,	g_raid_tr_iostart_raid1),
122 	KOBJMETHOD(g_raid_tr_iodone,	g_raid_tr_iodone_raid1),
123 	KOBJMETHOD(g_raid_tr_kerneldump, g_raid_tr_kerneldump_raid1),
124 	KOBJMETHOD(g_raid_tr_locked,	g_raid_tr_locked_raid1),
125 	KOBJMETHOD(g_raid_tr_idle,	g_raid_tr_idle_raid1),
126 	KOBJMETHOD(g_raid_tr_free,	g_raid_tr_free_raid1),
127 	{ 0, 0 }
128 };
129 
130 static struct g_raid_tr_class g_raid_tr_raid1_class = {
131 	"RAID1",
132 	g_raid_tr_raid1_methods,
133 	sizeof(struct g_raid_tr_raid1_object),
134 	.trc_priority = 100
135 };
136 
137 static void g_raid_tr_raid1_rebuild_abort(struct g_raid_tr_object *tr);
138 static void g_raid_tr_raid1_maybe_rebuild(struct g_raid_tr_object *tr,
139     struct g_raid_subdisk *sd);
140 
141 static int
142 g_raid_tr_taste_raid1(struct g_raid_tr_object *tr, struct g_raid_volume *vol)
143 {
144 	struct g_raid_tr_raid1_object *trs;
145 
146 	trs = (struct g_raid_tr_raid1_object *)tr;
147 	if (tr->tro_volume->v_raid_level != G_RAID_VOLUME_RL_RAID1 ||
148 	    tr->tro_volume->v_raid_level_qualifier != G_RAID_VOLUME_RLQ_NONE)
149 		return (G_RAID_TR_TASTE_FAIL);
150 	trs->trso_starting = 1;
151 	return (G_RAID_TR_TASTE_SUCCEED);
152 }
153 
154 static int
155 g_raid_tr_update_state_raid1(struct g_raid_volume *vol,
156     struct g_raid_subdisk *sd)
157 {
158 	struct g_raid_tr_raid1_object *trs;
159 	struct g_raid_softc *sc;
160 	struct g_raid_subdisk *tsd, *bestsd;
161 	u_int s;
162 	int i, na, ns;
163 
164 	sc = vol->v_softc;
165 	trs = (struct g_raid_tr_raid1_object *)vol->v_tr;
166 	if (trs->trso_stopping &&
167 	    (trs->trso_flags & TR_RAID1_F_DOING_SOME) == 0)
168 		s = G_RAID_VOLUME_S_STOPPED;
169 	else if (trs->trso_starting)
170 		s = G_RAID_VOLUME_S_STARTING;
171 	else {
172 		/* Make sure we have at least one ACTIVE disk. */
173 		na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
174 		if (na == 0) {
175 			/*
176 			 * Critical situation! We have no any active disk!
177 			 * Choose the best disk we have to make it active.
178 			 */
179 			bestsd = &vol->v_subdisks[0];
180 			for (i = 1; i < vol->v_disks_count; i++) {
181 				tsd = &vol->v_subdisks[i];
182 				if (tsd->sd_state > bestsd->sd_state)
183 					bestsd = tsd;
184 				else if (tsd->sd_state == bestsd->sd_state &&
185 				    (tsd->sd_state == G_RAID_SUBDISK_S_REBUILD ||
186 				     tsd->sd_state == G_RAID_SUBDISK_S_RESYNC) &&
187 				    tsd->sd_rebuild_pos > bestsd->sd_rebuild_pos)
188 					bestsd = tsd;
189 			}
190 			if (bestsd->sd_state >= G_RAID_SUBDISK_S_UNINITIALIZED) {
191 				/* We found reasonable candidate. */
192 				G_RAID_DEBUG1(1, sc,
193 				    "Promote subdisk %s:%d from %s to ACTIVE.",
194 				    vol->v_name, bestsd->sd_pos,
195 				    g_raid_subdisk_state2str(bestsd->sd_state));
196 				g_raid_change_subdisk_state(bestsd,
197 				    G_RAID_SUBDISK_S_ACTIVE);
198 				g_raid_write_metadata(sc,
199 				    vol, bestsd, bestsd->sd_disk);
200 			}
201 		}
202 		na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
203 		ns = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
204 		     g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
205 		if (na == vol->v_disks_count)
206 			s = G_RAID_VOLUME_S_OPTIMAL;
207 		else if (na + ns == vol->v_disks_count)
208 			s = G_RAID_VOLUME_S_SUBOPTIMAL;
209 		else if (na > 0)
210 			s = G_RAID_VOLUME_S_DEGRADED;
211 		else
212 			s = G_RAID_VOLUME_S_BROKEN;
213 		g_raid_tr_raid1_maybe_rebuild(vol->v_tr, sd);
214 	}
215 	if (s != vol->v_state) {
216 		g_raid_event_send(vol, G_RAID_VOLUME_S_ALIVE(s) ?
217 		    G_RAID_VOLUME_E_UP : G_RAID_VOLUME_E_DOWN,
218 		    G_RAID_EVENT_VOLUME);
219 		g_raid_change_volume_state(vol, s);
220 		if (!trs->trso_starting && !trs->trso_stopping)
221 			g_raid_write_metadata(sc, vol, NULL, NULL);
222 	}
223 	return (0);
224 }
225 
226 static void
227 g_raid_tr_raid1_fail_disk(struct g_raid_softc *sc, struct g_raid_subdisk *sd,
228     struct g_raid_disk *disk)
229 {
230 	/*
231 	 * We don't fail the last disk in the pack, since it still has decent
232 	 * data on it and that's better than failing the disk if it is the root
233 	 * file system.
234 	 *
235 	 * XXX should this be controlled via a tunable?  It makes sense for
236 	 * the volume that has / on it.  I can't think of a case where we'd
237 	 * want the volume to go away on this kind of event.
238 	 */
239 	if (g_raid_nsubdisks(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == 1 &&
240 	    g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == sd)
241 		return;
242 	g_raid_fail_disk(sc, sd, disk);
243 }
244 
245 static void
246 g_raid_tr_raid1_rebuild_some(struct g_raid_tr_object *tr)
247 {
248 	struct g_raid_tr_raid1_object *trs;
249 	struct g_raid_subdisk *sd, *good_sd;
250 	struct bio *bp;
251 
252 	trs = (struct g_raid_tr_raid1_object *)tr;
253 	if (trs->trso_flags & TR_RAID1_F_DOING_SOME)
254 		return;
255 	sd = trs->trso_failed_sd;
256 	good_sd = g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE);
257 	if (good_sd == NULL) {
258 		g_raid_tr_raid1_rebuild_abort(tr);
259 		return;
260 	}
261 	bp = &trs->trso_bio;
262 	memset(bp, 0, sizeof(*bp));
263 	bp->bio_offset = sd->sd_rebuild_pos;
264 	bp->bio_length = MIN(g_raid1_rebuild_slab,
265 	    sd->sd_size - sd->sd_rebuild_pos);
266 	bp->bio_data = trs->trso_buffer;
267 	bp->bio_cmd = BIO_READ;
268 	bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
269 	bp->bio_caller1 = good_sd;
270 	trs->trso_flags |= TR_RAID1_F_DOING_SOME;
271 	trs->trso_flags |= TR_RAID1_F_LOCKED;
272 	g_raid_lock_range(sd->sd_volume,	/* Lock callback starts I/O */
273 	   bp->bio_offset, bp->bio_length, NULL, bp);
274 }
275 
276 static void
277 g_raid_tr_raid1_rebuild_done(struct g_raid_tr_raid1_object *trs)
278 {
279 	struct g_raid_volume *vol;
280 	struct g_raid_subdisk *sd;
281 
282 	vol = trs->trso_base.tro_volume;
283 	sd = trs->trso_failed_sd;
284 	g_raid_write_metadata(vol->v_softc, vol, sd, sd->sd_disk);
285 	free(trs->trso_buffer, M_TR_RAID1);
286 	trs->trso_buffer = NULL;
287 	trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
288 	trs->trso_type = TR_RAID1_NONE;
289 	trs->trso_recover_slabs = 0;
290 	trs->trso_failed_sd = NULL;
291 	g_raid_tr_update_state_raid1(vol, NULL);
292 }
293 
294 static void
295 g_raid_tr_raid1_rebuild_finish(struct g_raid_tr_object *tr)
296 {
297 	struct g_raid_tr_raid1_object *trs;
298 	struct g_raid_subdisk *sd;
299 
300 	trs = (struct g_raid_tr_raid1_object *)tr;
301 	sd = trs->trso_failed_sd;
302 	G_RAID_DEBUG1(0, tr->tro_volume->v_softc,
303 	    "Subdisk %s:%d-%s rebuild completed.",
304 	    sd->sd_volume->v_name, sd->sd_pos,
305 	    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
306 	g_raid_change_subdisk_state(sd, G_RAID_SUBDISK_S_ACTIVE);
307 	sd->sd_rebuild_pos = 0;
308 	g_raid_tr_raid1_rebuild_done(trs);
309 }
310 
311 static void
312 g_raid_tr_raid1_rebuild_abort(struct g_raid_tr_object *tr)
313 {
314 	struct g_raid_tr_raid1_object *trs;
315 	struct g_raid_subdisk *sd;
316 	struct g_raid_volume *vol;
317 	off_t len;
318 
319 	vol = tr->tro_volume;
320 	trs = (struct g_raid_tr_raid1_object *)tr;
321 	sd = trs->trso_failed_sd;
322 	if (trs->trso_flags & TR_RAID1_F_DOING_SOME) {
323 		G_RAID_DEBUG1(1, vol->v_softc,
324 		    "Subdisk %s:%d-%s rebuild is aborting.",
325 		    sd->sd_volume->v_name, sd->sd_pos,
326 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
327 		trs->trso_flags |= TR_RAID1_F_ABORT;
328 	} else {
329 		G_RAID_DEBUG1(0, vol->v_softc,
330 		    "Subdisk %s:%d-%s rebuild aborted.",
331 		    sd->sd_volume->v_name, sd->sd_pos,
332 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
333 		trs->trso_flags &= ~TR_RAID1_F_ABORT;
334 		if (trs->trso_flags & TR_RAID1_F_LOCKED) {
335 			trs->trso_flags &= ~TR_RAID1_F_LOCKED;
336 			len = MIN(g_raid1_rebuild_slab,
337 			    sd->sd_size - sd->sd_rebuild_pos);
338 			g_raid_unlock_range(tr->tro_volume,
339 			    sd->sd_rebuild_pos, len);
340 		}
341 		g_raid_tr_raid1_rebuild_done(trs);
342 	}
343 }
344 
345 static void
346 g_raid_tr_raid1_rebuild_start(struct g_raid_tr_object *tr)
347 {
348 	struct g_raid_volume *vol;
349 	struct g_raid_tr_raid1_object *trs;
350 	struct g_raid_subdisk *sd, *fsd;
351 
352 	vol = tr->tro_volume;
353 	trs = (struct g_raid_tr_raid1_object *)tr;
354 	if (trs->trso_failed_sd) {
355 		G_RAID_DEBUG1(1, vol->v_softc,
356 		    "Already rebuild in start rebuild. pos %jd\n",
357 		    (intmax_t)trs->trso_failed_sd->sd_rebuild_pos);
358 		return;
359 	}
360 	sd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_ACTIVE);
361 	if (sd == NULL) {
362 		G_RAID_DEBUG1(1, vol->v_softc,
363 		    "No active disk to rebuild.  night night.");
364 		return;
365 	}
366 	fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_RESYNC);
367 	if (fsd == NULL)
368 		fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_REBUILD);
369 	if (fsd == NULL) {
370 		fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_STALE);
371 		if (fsd != NULL) {
372 			fsd->sd_rebuild_pos = 0;
373 			g_raid_change_subdisk_state(fsd,
374 			    G_RAID_SUBDISK_S_RESYNC);
375 			g_raid_write_metadata(vol->v_softc, vol, fsd, NULL);
376 		} else {
377 			fsd = g_raid_get_subdisk(vol,
378 			    G_RAID_SUBDISK_S_UNINITIALIZED);
379 			if (fsd == NULL)
380 				fsd = g_raid_get_subdisk(vol,
381 				    G_RAID_SUBDISK_S_NEW);
382 			if (fsd != NULL) {
383 				fsd->sd_rebuild_pos = 0;
384 				g_raid_change_subdisk_state(fsd,
385 				    G_RAID_SUBDISK_S_REBUILD);
386 				g_raid_write_metadata(vol->v_softc,
387 				    vol, fsd, NULL);
388 			}
389 		}
390 	}
391 	if (fsd == NULL) {
392 		G_RAID_DEBUG1(1, vol->v_softc,
393 		    "No failed disk to rebuild.  night night.");
394 		return;
395 	}
396 	trs->trso_failed_sd = fsd;
397 	G_RAID_DEBUG1(0, vol->v_softc,
398 	    "Subdisk %s:%d-%s rebuild start at %jd.",
399 	    fsd->sd_volume->v_name, fsd->sd_pos,
400 	    fsd->sd_disk ? g_raid_get_diskname(fsd->sd_disk) : "[none]",
401 	    trs->trso_failed_sd->sd_rebuild_pos);
402 	trs->trso_type = TR_RAID1_REBUILD;
403 	trs->trso_buffer = malloc(g_raid1_rebuild_slab, M_TR_RAID1, M_WAITOK);
404 	trs->trso_meta_update = g_raid1_rebuild_meta_update;
405 	g_raid_tr_raid1_rebuild_some(tr);
406 }
407 
408 
409 static void
410 g_raid_tr_raid1_maybe_rebuild(struct g_raid_tr_object *tr,
411     struct g_raid_subdisk *sd)
412 {
413 	struct g_raid_volume *vol;
414 	struct g_raid_tr_raid1_object *trs;
415 	int na, nr;
416 
417 	/*
418 	 * If we're stopping, don't do anything.  If we don't have at least one
419 	 * good disk and one bad disk, we don't do anything.  And if there's a
420 	 * 'good disk' stored in the trs, then we're in progress and we punt.
421 	 * If we make it past all these checks, we need to rebuild.
422 	 */
423 	vol = tr->tro_volume;
424 	trs = (struct g_raid_tr_raid1_object *)tr;
425 	if (trs->trso_stopping)
426 		return;
427 	na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
428 	nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_REBUILD) +
429 	    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
430 	switch(trs->trso_type) {
431 	case TR_RAID1_NONE:
432 		if (na == 0)
433 			return;
434 		if (nr == 0) {
435 			nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_NEW) +
436 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
437 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_UNINITIALIZED);
438 			if (nr == 0)
439 				return;
440 		}
441 		g_raid_tr_raid1_rebuild_start(tr);
442 		break;
443 	case TR_RAID1_REBUILD:
444 		if (na == 0 || nr == 0 || trs->trso_failed_sd == sd)
445 			g_raid_tr_raid1_rebuild_abort(tr);
446 		break;
447 	case TR_RAID1_RESYNC:
448 		break;
449 	}
450 }
451 
452 static int
453 g_raid_tr_event_raid1(struct g_raid_tr_object *tr,
454     struct g_raid_subdisk *sd, u_int event)
455 {
456 
457 	g_raid_tr_update_state_raid1(tr->tro_volume, sd);
458 	return (0);
459 }
460 
461 static int
462 g_raid_tr_start_raid1(struct g_raid_tr_object *tr)
463 {
464 	struct g_raid_tr_raid1_object *trs;
465 	struct g_raid_volume *vol;
466 
467 	trs = (struct g_raid_tr_raid1_object *)tr;
468 	vol = tr->tro_volume;
469 	trs->trso_starting = 0;
470 	g_raid_tr_update_state_raid1(vol, NULL);
471 	return (0);
472 }
473 
474 static int
475 g_raid_tr_stop_raid1(struct g_raid_tr_object *tr)
476 {
477 	struct g_raid_tr_raid1_object *trs;
478 	struct g_raid_volume *vol;
479 
480 	trs = (struct g_raid_tr_raid1_object *)tr;
481 	vol = tr->tro_volume;
482 	trs->trso_starting = 0;
483 	trs->trso_stopping = 1;
484 	g_raid_tr_update_state_raid1(vol, NULL);
485 	return (0);
486 }
487 
488 /*
489  * Select the disk to read from.  Take into account: subdisk state, running
490  * error recovery, average disk load, head position and possible cache hits.
491  */
492 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
493 static struct g_raid_subdisk *
494 g_raid_tr_raid1_select_read_disk(struct g_raid_volume *vol, struct bio *bp,
495     u_int mask)
496 {
497 	struct g_raid_subdisk *sd, *best;
498 	int i, prio, bestprio;
499 
500 	best = NULL;
501 	bestprio = INT_MAX;
502 	for (i = 0; i < vol->v_disks_count; i++) {
503 		sd = &vol->v_subdisks[i];
504 		if (sd->sd_state != G_RAID_SUBDISK_S_ACTIVE &&
505 		    ((sd->sd_state != G_RAID_SUBDISK_S_REBUILD &&
506 		      sd->sd_state != G_RAID_SUBDISK_S_RESYNC) ||
507 		     bp->bio_offset + bp->bio_length > sd->sd_rebuild_pos))
508 			continue;
509 		if ((mask & (1 << i)) != 0)
510 			continue;
511 		prio = G_RAID_SUBDISK_LOAD(sd);
512 		prio += min(sd->sd_recovery, 255) << 22;
513 		prio += (G_RAID_SUBDISK_S_ACTIVE - sd->sd_state) << 16;
514 		/* If disk head is precisely in position - highly prefer it. */
515 		if (G_RAID_SUBDISK_POS(sd) == bp->bio_offset)
516 			prio -= 2 * G_RAID_SUBDISK_LOAD_SCALE;
517 		else
518 		/* If disk head is close to position - prefer it. */
519 		if (ABS(G_RAID_SUBDISK_POS(sd) - bp->bio_offset) <
520 		    G_RAID_SUBDISK_TRACK_SIZE)
521 			prio -= 1 * G_RAID_SUBDISK_LOAD_SCALE;
522 		if (prio < bestprio) {
523 			best = sd;
524 			bestprio = prio;
525 		}
526 	}
527 	return (best);
528 }
529 
530 static void
531 g_raid_tr_iostart_raid1_read(struct g_raid_tr_object *tr, struct bio *bp)
532 {
533 	struct g_raid_subdisk *sd;
534 	struct bio *cbp;
535 
536 	sd = g_raid_tr_raid1_select_read_disk(tr->tro_volume, bp, 0);
537 	KASSERT(sd != NULL, ("No active disks in volume %s.",
538 		tr->tro_volume->v_name));
539 
540 	cbp = g_clone_bio(bp);
541 	if (cbp == NULL) {
542 		g_raid_iodone(bp, ENOMEM);
543 		return;
544 	}
545 
546 	g_raid_subdisk_iostart(sd, cbp);
547 }
548 
549 static void
550 g_raid_tr_iostart_raid1_write(struct g_raid_tr_object *tr, struct bio *bp)
551 {
552 	struct g_raid_volume *vol;
553 	struct g_raid_subdisk *sd;
554 	struct bio_queue_head queue;
555 	struct bio *cbp;
556 	int i;
557 
558 	vol = tr->tro_volume;
559 
560 	/*
561 	 * Allocate all bios before sending any request, so we can return
562 	 * ENOMEM in nice and clean way.
563 	 */
564 	bioq_init(&queue);
565 	for (i = 0; i < vol->v_disks_count; i++) {
566 		sd = &vol->v_subdisks[i];
567 		switch (sd->sd_state) {
568 		case G_RAID_SUBDISK_S_ACTIVE:
569 			break;
570 		case G_RAID_SUBDISK_S_REBUILD:
571 			/*
572 			 * When rebuilding, only part of this subdisk is
573 			 * writable, the rest will be written as part of the
574 			 * that process.
575 			 */
576 			if (bp->bio_offset >= sd->sd_rebuild_pos)
577 				continue;
578 			break;
579 		case G_RAID_SUBDISK_S_STALE:
580 		case G_RAID_SUBDISK_S_RESYNC:
581 			/*
582 			 * Resyncing still writes on the theory that the
583 			 * resync'd disk is very close and writing it will
584 			 * keep it that way better if we keep up while
585 			 * resyncing.
586 			 */
587 			break;
588 		default:
589 			continue;
590 		}
591 		cbp = g_clone_bio(bp);
592 		if (cbp == NULL)
593 			goto failure;
594 		cbp->bio_caller1 = sd;
595 		bioq_insert_tail(&queue, cbp);
596 	}
597 	for (cbp = bioq_first(&queue); cbp != NULL;
598 	    cbp = bioq_first(&queue)) {
599 		bioq_remove(&queue, cbp);
600 		sd = cbp->bio_caller1;
601 		cbp->bio_caller1 = NULL;
602 		g_raid_subdisk_iostart(sd, cbp);
603 	}
604 	return;
605 failure:
606 	for (cbp = bioq_first(&queue); cbp != NULL;
607 	    cbp = bioq_first(&queue)) {
608 		bioq_remove(&queue, cbp);
609 		g_destroy_bio(cbp);
610 	}
611 	if (bp->bio_error == 0)
612 		bp->bio_error = ENOMEM;
613 	g_raid_iodone(bp, bp->bio_error);
614 }
615 
616 static void
617 g_raid_tr_iostart_raid1(struct g_raid_tr_object *tr, struct bio *bp)
618 {
619 	struct g_raid_volume *vol;
620 	struct g_raid_tr_raid1_object *trs;
621 
622 	vol = tr->tro_volume;
623 	trs = (struct g_raid_tr_raid1_object *)tr;
624 	if (vol->v_state != G_RAID_VOLUME_S_OPTIMAL &&
625 	    vol->v_state != G_RAID_VOLUME_S_SUBOPTIMAL &&
626 	    vol->v_state != G_RAID_VOLUME_S_DEGRADED) {
627 		g_raid_iodone(bp, EIO);
628 		return;
629 	}
630 	/*
631 	 * If we're rebuilding, squeeze in rebuild activity every so often,
632 	 * even when the disk is busy.  Be sure to only count real I/O
633 	 * to the disk.  All 'SPECIAL' I/O is traffic generated to the disk
634 	 * by this module.
635 	 */
636 	if (trs->trso_failed_sd != NULL &&
637 	    !(bp->bio_cflags & G_RAID_BIO_FLAG_SPECIAL)) {
638 		/* Make this new or running now round short. */
639 		trs->trso_recover_slabs = 0;
640 		if (--trs->trso_fair_io <= 0) {
641 			trs->trso_fair_io = g_raid1_rebuild_fair_io;
642 			g_raid_tr_raid1_rebuild_some(tr);
643 		}
644 	}
645 	switch (bp->bio_cmd) {
646 	case BIO_READ:
647 		g_raid_tr_iostart_raid1_read(tr, bp);
648 		break;
649 	case BIO_WRITE:
650 		g_raid_tr_iostart_raid1_write(tr, bp);
651 		break;
652 	case BIO_DELETE:
653 		g_raid_iodone(bp, EIO);
654 		break;
655 	case BIO_FLUSH:
656 		g_raid_tr_flush_common(tr, bp);
657 		break;
658 	default:
659 		KASSERT(1 == 0, ("Invalid command here: %u (volume=%s)",
660 		    bp->bio_cmd, vol->v_name));
661 		break;
662 	}
663 }
664 
665 static void
666 g_raid_tr_iodone_raid1(struct g_raid_tr_object *tr,
667     struct g_raid_subdisk *sd, struct bio *bp)
668 {
669 	struct bio *cbp;
670 	struct g_raid_subdisk *nsd;
671 	struct g_raid_volume *vol;
672 	struct bio *pbp;
673 	struct g_raid_tr_raid1_object *trs;
674 	uintptr_t *mask;
675 	int error, do_write;
676 
677 	trs = (struct g_raid_tr_raid1_object *)tr;
678 	vol = tr->tro_volume;
679 	if (bp->bio_cflags & G_RAID_BIO_FLAG_SYNC) {
680 		/*
681 		 * This operation is part of a rebuild or resync operation.
682 		 * See what work just got done, then schedule the next bit of
683 		 * work, if any.  Rebuild/resync is done a little bit at a
684 		 * time.  Either when a timeout happens, or after we get a
685 		 * bunch of I/Os to the disk (to make sure an active system
686 		 * will complete in a sane amount of time).
687 		 *
688 		 * We are setup to do differing amounts of work for each of
689 		 * these cases.  so long as the slabs is smallish (less than
690 		 * 50 or so, I'd guess, but that's just a WAG), we shouldn't
691 		 * have any bio starvation issues.  For active disks, we do
692 		 * 5MB of data, for inactive ones, we do 50MB.
693 		 */
694 		if (trs->trso_type == TR_RAID1_REBUILD) {
695 			if (bp->bio_cmd == BIO_READ) {
696 
697 				/* Immediately abort rebuild, if requested. */
698 				if (trs->trso_flags & TR_RAID1_F_ABORT) {
699 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
700 					g_raid_tr_raid1_rebuild_abort(tr);
701 					return;
702 				}
703 
704 				/* On read error, skip and cross fingers. */
705 				if (bp->bio_error != 0) {
706 					G_RAID_LOGREQ(0, bp,
707 					    "Read error during rebuild (%d), "
708 					    "possible data loss!",
709 					    bp->bio_error);
710 					goto rebuild_round_done;
711 				}
712 
713 				/*
714 				 * The read operation finished, queue the
715 				 * write and get out.
716 				 */
717 				G_RAID_LOGREQ(4, bp, "rebuild read done. %d",
718 				    bp->bio_error);
719 				bp->bio_cmd = BIO_WRITE;
720 				bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
721 				G_RAID_LOGREQ(4, bp, "Queueing rebuild write.");
722 				g_raid_subdisk_iostart(trs->trso_failed_sd, bp);
723 			} else {
724 				/*
725 				 * The write operation just finished.  Do
726 				 * another.  We keep cloning the master bio
727 				 * since it has the right buffers allocated to
728 				 * it.
729 				 */
730 				G_RAID_LOGREQ(4, bp,
731 				    "rebuild write done. Error %d",
732 				    bp->bio_error);
733 				nsd = trs->trso_failed_sd;
734 				if (bp->bio_error != 0 ||
735 				    trs->trso_flags & TR_RAID1_F_ABORT) {
736 					if ((trs->trso_flags &
737 					    TR_RAID1_F_ABORT) == 0) {
738 						g_raid_tr_raid1_fail_disk(sd->sd_softc,
739 						    nsd, nsd->sd_disk);
740 					}
741 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
742 					g_raid_tr_raid1_rebuild_abort(tr);
743 					return;
744 				}
745 rebuild_round_done:
746 				nsd = trs->trso_failed_sd;
747 				trs->trso_flags &= ~TR_RAID1_F_LOCKED;
748 				g_raid_unlock_range(sd->sd_volume,
749 				    bp->bio_offset, bp->bio_length);
750 				nsd->sd_rebuild_pos += bp->bio_length;
751 				if (nsd->sd_rebuild_pos >= nsd->sd_size) {
752 					g_raid_tr_raid1_rebuild_finish(tr);
753 					return;
754 				}
755 
756 				/* Abort rebuild if we are stopping */
757 				if (trs->trso_stopping) {
758 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
759 					g_raid_tr_raid1_rebuild_abort(tr);
760 					return;
761 				}
762 
763 				if (--trs->trso_meta_update <= 0) {
764 					g_raid_write_metadata(vol->v_softc,
765 					    vol, nsd, nsd->sd_disk);
766 					trs->trso_meta_update =
767 					    g_raid1_rebuild_meta_update;
768 				}
769 				trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
770 				if (--trs->trso_recover_slabs <= 0)
771 					return;
772 				g_raid_tr_raid1_rebuild_some(tr);
773 			}
774 		} else if (trs->trso_type == TR_RAID1_RESYNC) {
775 			/*
776 			 * read good sd, read bad sd in parallel.  when both
777 			 * done, compare the buffers.  write good to the bad
778 			 * if different.  do the next bit of work.
779 			 */
780 			panic("Somehow, we think we're doing a resync");
781 		}
782 		return;
783 	}
784 	pbp = bp->bio_parent;
785 	pbp->bio_inbed++;
786 	if (bp->bio_cmd == BIO_READ && bp->bio_error != 0) {
787 		/*
788 		 * Read failed on first drive.  Retry the read error on
789 		 * another disk drive, if available, before erroring out the
790 		 * read.
791 		 */
792 		sd->sd_disk->d_read_errs++;
793 		G_RAID_LOGREQ(0, bp,
794 		    "Read error (%d), %d read errors total",
795 		    bp->bio_error, sd->sd_disk->d_read_errs);
796 
797 		/*
798 		 * If there are too many read errors, we move to degraded.
799 		 * XXX Do we want to FAIL the drive (eg, make the user redo
800 		 * everything to get it back in sync), or just degrade the
801 		 * drive, which kicks off a resync?
802 		 */
803 		do_write = 1;
804 		if (sd->sd_disk->d_read_errs > g_raid_read_err_thresh) {
805 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
806 			if (pbp->bio_children == 1)
807 				do_write = 0;
808 		}
809 
810 		/*
811 		 * Find the other disk, and try to do the I/O to it.
812 		 */
813 		mask = (uintptr_t *)(&pbp->bio_driver2);
814 		if (pbp->bio_children == 1) {
815 			/* Save original subdisk. */
816 			pbp->bio_driver1 = do_write ? sd : NULL;
817 			*mask = 0;
818 		}
819 		*mask |= 1 << sd->sd_pos;
820 		nsd = g_raid_tr_raid1_select_read_disk(vol, pbp, *mask);
821 		if (nsd != NULL && (cbp = g_clone_bio(pbp)) != NULL) {
822 			g_destroy_bio(bp);
823 			G_RAID_LOGREQ(2, cbp, "Retrying read from %d",
824 			    nsd->sd_pos);
825 			if (pbp->bio_children == 2 && do_write) {
826 				sd->sd_recovery++;
827 				cbp->bio_caller1 = nsd;
828 				pbp->bio_pflags = G_RAID_BIO_FLAG_LOCKED;
829 				/* Lock callback starts I/O */
830 				g_raid_lock_range(sd->sd_volume,
831 				    cbp->bio_offset, cbp->bio_length, pbp, cbp);
832 			} else {
833 				g_raid_subdisk_iostart(nsd, cbp);
834 			}
835 			return;
836 		}
837 		/*
838 		 * We can't retry.  Return the original error by falling
839 		 * through.  This will happen when there's only one good disk.
840 		 * We don't need to fail the raid, since its actual state is
841 		 * based on the state of the subdisks.
842 		 */
843 		G_RAID_LOGREQ(2, bp, "Couldn't retry read, failing it");
844 	}
845 	if (bp->bio_cmd == BIO_READ &&
846 	    bp->bio_error == 0 &&
847 	    pbp->bio_children > 1 &&
848 	    pbp->bio_driver1 != NULL) {
849 		/*
850 		 * If it was a read, and bio_children is >1, then we just
851 		 * recovered the data from the second drive.  We should try to
852 		 * write that data to the first drive if sector remapping is
853 		 * enabled.  A write should put the data in a new place on the
854 		 * disk, remapping the bad sector.  Do we need to do that by
855 		 * queueing a request to the main worker thread?  It doesn't
856 		 * affect the return code of this current read, and can be
857 		 * done at our liesure.  However, to make the code simpler, it
858 		 * is done syncrhonously.
859 		 */
860 		G_RAID_LOGREQ(3, bp, "Recovered data from other drive");
861 		cbp = g_clone_bio(pbp);
862 		if (cbp != NULL) {
863 			g_destroy_bio(bp);
864 			cbp->bio_cmd = BIO_WRITE;
865 			cbp->bio_cflags = G_RAID_BIO_FLAG_REMAP;
866 			G_RAID_LOGREQ(2, cbp,
867 			    "Attempting bad sector remap on failing drive.");
868 			g_raid_subdisk_iostart(pbp->bio_driver1, cbp);
869 			return;
870 		}
871 	}
872 	if (pbp->bio_pflags & G_RAID_BIO_FLAG_LOCKED) {
873 		/*
874 		 * We're done with a recovery, mark the range as unlocked.
875 		 * For any write errors, we agressively fail the disk since
876 		 * there was both a READ and a WRITE error at this location.
877 		 * Both types of errors generally indicates the drive is on
878 		 * the verge of total failure anyway.  Better to stop trusting
879 		 * it now.  However, we need to reset error to 0 in that case
880 		 * because we're not failing the original I/O which succeeded.
881 		 */
882 		if (bp->bio_cmd == BIO_WRITE && bp->bio_error) {
883 			G_RAID_LOGREQ(0, bp, "Remap write failed: "
884 			    "failing subdisk.");
885 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
886 			bp->bio_error = 0;
887 		}
888 		if (pbp->bio_driver1 != NULL) {
889 			((struct g_raid_subdisk *)pbp->bio_driver1)
890 			    ->sd_recovery--;
891 		}
892 		G_RAID_LOGREQ(2, bp, "REMAP done %d.", bp->bio_error);
893 		g_raid_unlock_range(sd->sd_volume, bp->bio_offset,
894 		    bp->bio_length);
895 	}
896 	error = bp->bio_error;
897 	g_destroy_bio(bp);
898 	if (pbp->bio_children == pbp->bio_inbed) {
899 		pbp->bio_completed = pbp->bio_length;
900 		g_raid_iodone(pbp, error);
901 	}
902 }
903 
904 static int
905 g_raid_tr_kerneldump_raid1(struct g_raid_tr_object *tr,
906     void *virtual, vm_offset_t physical, off_t offset, size_t length)
907 {
908 	struct g_raid_volume *vol;
909 	struct g_raid_subdisk *sd;
910 	int error, i, ok;
911 
912 	vol = tr->tro_volume;
913 	error = 0;
914 	ok = 0;
915 	for (i = 0; i < vol->v_disks_count; i++) {
916 		sd = &vol->v_subdisks[i];
917 		switch (sd->sd_state) {
918 		case G_RAID_SUBDISK_S_ACTIVE:
919 			break;
920 		case G_RAID_SUBDISK_S_REBUILD:
921 			/*
922 			 * When rebuilding, only part of this subdisk is
923 			 * writable, the rest will be written as part of the
924 			 * that process.
925 			 */
926 			if (offset >= sd->sd_rebuild_pos)
927 				continue;
928 			break;
929 		case G_RAID_SUBDISK_S_STALE:
930 		case G_RAID_SUBDISK_S_RESYNC:
931 			/*
932 			 * Resyncing still writes on the theory that the
933 			 * resync'd disk is very close and writing it will
934 			 * keep it that way better if we keep up while
935 			 * resyncing.
936 			 */
937 			break;
938 		default:
939 			continue;
940 		}
941 		error = g_raid_subdisk_kerneldump(sd,
942 		    virtual, physical, offset, length);
943 		if (error == 0)
944 			ok++;
945 	}
946 	return (ok > 0 ? 0 : error);
947 }
948 
949 static int
950 g_raid_tr_locked_raid1(struct g_raid_tr_object *tr, void *argp)
951 {
952 	struct bio *bp;
953 	struct g_raid_subdisk *sd;
954 
955 	bp = (struct bio *)argp;
956 	sd = (struct g_raid_subdisk *)bp->bio_caller1;
957 	g_raid_subdisk_iostart(sd, bp);
958 
959 	return (0);
960 }
961 
962 static int
963 g_raid_tr_idle_raid1(struct g_raid_tr_object *tr)
964 {
965 	struct g_raid_tr_raid1_object *trs;
966 
967 	trs = (struct g_raid_tr_raid1_object *)tr;
968 	trs->trso_fair_io = g_raid1_rebuild_fair_io;
969 	trs->trso_recover_slabs = g_raid1_rebuild_cluster_idle;
970 	if (trs->trso_type == TR_RAID1_REBUILD)
971 		g_raid_tr_raid1_rebuild_some(tr);
972 	return (0);
973 }
974 
975 static int
976 g_raid_tr_free_raid1(struct g_raid_tr_object *tr)
977 {
978 	struct g_raid_tr_raid1_object *trs;
979 
980 	trs = (struct g_raid_tr_raid1_object *)tr;
981 
982 	if (trs->trso_buffer != NULL) {
983 		free(trs->trso_buffer, M_TR_RAID1);
984 		trs->trso_buffer = NULL;
985 	}
986 	return (0);
987 }
988 
989 G_RAID_TR_DECLARE(g_raid_tr_raid1);
990