xref: /freebsd/sys/geom/raid/tr_raid1.c (revision 6486b015fc84e96725fef22b0e3363351399ae83)
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_R1SM &&
149 	     tr->tro_volume->v_raid_level_qualifier != G_RAID_VOLUME_RLQ_R1MM))
150 		return (G_RAID_TR_TASTE_FAIL);
151 	trs->trso_starting = 1;
152 	return (G_RAID_TR_TASTE_SUCCEED);
153 }
154 
155 static int
156 g_raid_tr_update_state_raid1(struct g_raid_volume *vol,
157     struct g_raid_subdisk *sd)
158 {
159 	struct g_raid_tr_raid1_object *trs;
160 	struct g_raid_softc *sc;
161 	struct g_raid_subdisk *tsd, *bestsd;
162 	u_int s;
163 	int i, na, ns;
164 
165 	sc = vol->v_softc;
166 	trs = (struct g_raid_tr_raid1_object *)vol->v_tr;
167 	if (trs->trso_stopping &&
168 	    (trs->trso_flags & TR_RAID1_F_DOING_SOME) == 0)
169 		s = G_RAID_VOLUME_S_STOPPED;
170 	else if (trs->trso_starting)
171 		s = G_RAID_VOLUME_S_STARTING;
172 	else {
173 		/* Make sure we have at least one ACTIVE disk. */
174 		na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
175 		if (na == 0) {
176 			/*
177 			 * Critical situation! We have no any active disk!
178 			 * Choose the best disk we have to make it active.
179 			 */
180 			bestsd = &vol->v_subdisks[0];
181 			for (i = 1; i < vol->v_disks_count; i++) {
182 				tsd = &vol->v_subdisks[i];
183 				if (tsd->sd_state > bestsd->sd_state)
184 					bestsd = tsd;
185 				else if (tsd->sd_state == bestsd->sd_state &&
186 				    (tsd->sd_state == G_RAID_SUBDISK_S_REBUILD ||
187 				     tsd->sd_state == G_RAID_SUBDISK_S_RESYNC) &&
188 				    tsd->sd_rebuild_pos > bestsd->sd_rebuild_pos)
189 					bestsd = tsd;
190 			}
191 			if (bestsd->sd_state >= G_RAID_SUBDISK_S_UNINITIALIZED) {
192 				/* We found reasonable candidate. */
193 				G_RAID_DEBUG1(1, sc,
194 				    "Promote subdisk %s:%d from %s to ACTIVE.",
195 				    vol->v_name, bestsd->sd_pos,
196 				    g_raid_subdisk_state2str(bestsd->sd_state));
197 				g_raid_change_subdisk_state(bestsd,
198 				    G_RAID_SUBDISK_S_ACTIVE);
199 				g_raid_write_metadata(sc,
200 				    vol, bestsd, bestsd->sd_disk);
201 			}
202 		}
203 		na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
204 		ns = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
205 		     g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
206 		if (na == vol->v_disks_count)
207 			s = G_RAID_VOLUME_S_OPTIMAL;
208 		else if (na + ns == vol->v_disks_count)
209 			s = G_RAID_VOLUME_S_SUBOPTIMAL;
210 		else if (na > 0)
211 			s = G_RAID_VOLUME_S_DEGRADED;
212 		else
213 			s = G_RAID_VOLUME_S_BROKEN;
214 		g_raid_tr_raid1_maybe_rebuild(vol->v_tr, sd);
215 	}
216 	if (s != vol->v_state) {
217 		g_raid_event_send(vol, G_RAID_VOLUME_S_ALIVE(s) ?
218 		    G_RAID_VOLUME_E_UP : G_RAID_VOLUME_E_DOWN,
219 		    G_RAID_EVENT_VOLUME);
220 		g_raid_change_volume_state(vol, s);
221 		if (!trs->trso_starting && !trs->trso_stopping)
222 			g_raid_write_metadata(sc, vol, NULL, NULL);
223 	}
224 	return (0);
225 }
226 
227 static void
228 g_raid_tr_raid1_fail_disk(struct g_raid_softc *sc, struct g_raid_subdisk *sd,
229     struct g_raid_disk *disk)
230 {
231 	/*
232 	 * We don't fail the last disk in the pack, since it still has decent
233 	 * data on it and that's better than failing the disk if it is the root
234 	 * file system.
235 	 *
236 	 * XXX should this be controlled via a tunable?  It makes sense for
237 	 * the volume that has / on it.  I can't think of a case where we'd
238 	 * want the volume to go away on this kind of event.
239 	 */
240 	if (g_raid_nsubdisks(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == 1 &&
241 	    g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == sd)
242 		return;
243 	g_raid_fail_disk(sc, sd, disk);
244 }
245 
246 static void
247 g_raid_tr_raid1_rebuild_some(struct g_raid_tr_object *tr)
248 {
249 	struct g_raid_tr_raid1_object *trs;
250 	struct g_raid_subdisk *sd, *good_sd;
251 	struct bio *bp;
252 
253 	trs = (struct g_raid_tr_raid1_object *)tr;
254 	if (trs->trso_flags & TR_RAID1_F_DOING_SOME)
255 		return;
256 	sd = trs->trso_failed_sd;
257 	good_sd = g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE);
258 	if (good_sd == NULL) {
259 		g_raid_tr_raid1_rebuild_abort(tr);
260 		return;
261 	}
262 	bp = &trs->trso_bio;
263 	memset(bp, 0, sizeof(*bp));
264 	bp->bio_offset = sd->sd_rebuild_pos;
265 	bp->bio_length = MIN(g_raid1_rebuild_slab,
266 	    sd->sd_size - sd->sd_rebuild_pos);
267 	bp->bio_data = trs->trso_buffer;
268 	bp->bio_cmd = BIO_READ;
269 	bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
270 	bp->bio_caller1 = good_sd;
271 	trs->trso_flags |= TR_RAID1_F_DOING_SOME;
272 	trs->trso_flags |= TR_RAID1_F_LOCKED;
273 	g_raid_lock_range(sd->sd_volume,	/* Lock callback starts I/O */
274 	   bp->bio_offset, bp->bio_length, NULL, bp);
275 }
276 
277 static void
278 g_raid_tr_raid1_rebuild_done(struct g_raid_tr_raid1_object *trs)
279 {
280 	struct g_raid_volume *vol;
281 	struct g_raid_subdisk *sd;
282 
283 	vol = trs->trso_base.tro_volume;
284 	sd = trs->trso_failed_sd;
285 	g_raid_write_metadata(vol->v_softc, vol, sd, sd->sd_disk);
286 	free(trs->trso_buffer, M_TR_RAID1);
287 	trs->trso_buffer = NULL;
288 	trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
289 	trs->trso_type = TR_RAID1_NONE;
290 	trs->trso_recover_slabs = 0;
291 	trs->trso_failed_sd = NULL;
292 	g_raid_tr_update_state_raid1(vol, NULL);
293 }
294 
295 static void
296 g_raid_tr_raid1_rebuild_finish(struct g_raid_tr_object *tr)
297 {
298 	struct g_raid_tr_raid1_object *trs;
299 	struct g_raid_subdisk *sd;
300 
301 	trs = (struct g_raid_tr_raid1_object *)tr;
302 	sd = trs->trso_failed_sd;
303 	G_RAID_DEBUG1(0, tr->tro_volume->v_softc,
304 	    "Subdisk %s:%d-%s rebuild completed.",
305 	    sd->sd_volume->v_name, sd->sd_pos,
306 	    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
307 	g_raid_change_subdisk_state(sd, G_RAID_SUBDISK_S_ACTIVE);
308 	sd->sd_rebuild_pos = 0;
309 	g_raid_tr_raid1_rebuild_done(trs);
310 }
311 
312 static void
313 g_raid_tr_raid1_rebuild_abort(struct g_raid_tr_object *tr)
314 {
315 	struct g_raid_tr_raid1_object *trs;
316 	struct g_raid_subdisk *sd;
317 	struct g_raid_volume *vol;
318 	off_t len;
319 
320 	vol = tr->tro_volume;
321 	trs = (struct g_raid_tr_raid1_object *)tr;
322 	sd = trs->trso_failed_sd;
323 	if (trs->trso_flags & TR_RAID1_F_DOING_SOME) {
324 		G_RAID_DEBUG1(1, vol->v_softc,
325 		    "Subdisk %s:%d-%s rebuild is aborting.",
326 		    sd->sd_volume->v_name, sd->sd_pos,
327 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
328 		trs->trso_flags |= TR_RAID1_F_ABORT;
329 	} else {
330 		G_RAID_DEBUG1(0, vol->v_softc,
331 		    "Subdisk %s:%d-%s rebuild aborted.",
332 		    sd->sd_volume->v_name, sd->sd_pos,
333 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
334 		trs->trso_flags &= ~TR_RAID1_F_ABORT;
335 		if (trs->trso_flags & TR_RAID1_F_LOCKED) {
336 			trs->trso_flags &= ~TR_RAID1_F_LOCKED;
337 			len = MIN(g_raid1_rebuild_slab,
338 			    sd->sd_size - sd->sd_rebuild_pos);
339 			g_raid_unlock_range(tr->tro_volume,
340 			    sd->sd_rebuild_pos, len);
341 		}
342 		g_raid_tr_raid1_rebuild_done(trs);
343 	}
344 }
345 
346 static void
347 g_raid_tr_raid1_rebuild_start(struct g_raid_tr_object *tr)
348 {
349 	struct g_raid_volume *vol;
350 	struct g_raid_tr_raid1_object *trs;
351 	struct g_raid_subdisk *sd, *fsd;
352 
353 	vol = tr->tro_volume;
354 	trs = (struct g_raid_tr_raid1_object *)tr;
355 	if (trs->trso_failed_sd) {
356 		G_RAID_DEBUG1(1, vol->v_softc,
357 		    "Already rebuild in start rebuild. pos %jd\n",
358 		    (intmax_t)trs->trso_failed_sd->sd_rebuild_pos);
359 		return;
360 	}
361 	sd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_ACTIVE);
362 	if (sd == NULL) {
363 		G_RAID_DEBUG1(1, vol->v_softc,
364 		    "No active disk to rebuild.  night night.");
365 		return;
366 	}
367 	fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_RESYNC);
368 	if (fsd == NULL)
369 		fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_REBUILD);
370 	if (fsd == NULL) {
371 		fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_STALE);
372 		if (fsd != NULL) {
373 			fsd->sd_rebuild_pos = 0;
374 			g_raid_change_subdisk_state(fsd,
375 			    G_RAID_SUBDISK_S_RESYNC);
376 			g_raid_write_metadata(vol->v_softc, vol, fsd, NULL);
377 		} else {
378 			fsd = g_raid_get_subdisk(vol,
379 			    G_RAID_SUBDISK_S_UNINITIALIZED);
380 			if (fsd == NULL)
381 				fsd = g_raid_get_subdisk(vol,
382 				    G_RAID_SUBDISK_S_NEW);
383 			if (fsd != NULL) {
384 				fsd->sd_rebuild_pos = 0;
385 				g_raid_change_subdisk_state(fsd,
386 				    G_RAID_SUBDISK_S_REBUILD);
387 				g_raid_write_metadata(vol->v_softc,
388 				    vol, fsd, NULL);
389 			}
390 		}
391 	}
392 	if (fsd == NULL) {
393 		G_RAID_DEBUG1(1, vol->v_softc,
394 		    "No failed disk to rebuild.  night night.");
395 		return;
396 	}
397 	trs->trso_failed_sd = fsd;
398 	G_RAID_DEBUG1(0, vol->v_softc,
399 	    "Subdisk %s:%d-%s rebuild start at %jd.",
400 	    fsd->sd_volume->v_name, fsd->sd_pos,
401 	    fsd->sd_disk ? g_raid_get_diskname(fsd->sd_disk) : "[none]",
402 	    trs->trso_failed_sd->sd_rebuild_pos);
403 	trs->trso_type = TR_RAID1_REBUILD;
404 	trs->trso_buffer = malloc(g_raid1_rebuild_slab, M_TR_RAID1, M_WAITOK);
405 	trs->trso_meta_update = g_raid1_rebuild_meta_update;
406 	g_raid_tr_raid1_rebuild_some(tr);
407 }
408 
409 
410 static void
411 g_raid_tr_raid1_maybe_rebuild(struct g_raid_tr_object *tr,
412     struct g_raid_subdisk *sd)
413 {
414 	struct g_raid_volume *vol;
415 	struct g_raid_tr_raid1_object *trs;
416 	int na, nr;
417 
418 	/*
419 	 * If we're stopping, don't do anything.  If we don't have at least one
420 	 * good disk and one bad disk, we don't do anything.  And if there's a
421 	 * 'good disk' stored in the trs, then we're in progress and we punt.
422 	 * If we make it past all these checks, we need to rebuild.
423 	 */
424 	vol = tr->tro_volume;
425 	trs = (struct g_raid_tr_raid1_object *)tr;
426 	if (trs->trso_stopping)
427 		return;
428 	na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
429 	nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_REBUILD) +
430 	    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
431 	switch(trs->trso_type) {
432 	case TR_RAID1_NONE:
433 		if (na == 0)
434 			return;
435 		if (nr == 0) {
436 			nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_NEW) +
437 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
438 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_UNINITIALIZED);
439 			if (nr == 0)
440 				return;
441 		}
442 		g_raid_tr_raid1_rebuild_start(tr);
443 		break;
444 	case TR_RAID1_REBUILD:
445 		if (na == 0 || nr == 0 || trs->trso_failed_sd == sd)
446 			g_raid_tr_raid1_rebuild_abort(tr);
447 		break;
448 	case TR_RAID1_RESYNC:
449 		break;
450 	}
451 }
452 
453 static int
454 g_raid_tr_event_raid1(struct g_raid_tr_object *tr,
455     struct g_raid_subdisk *sd, u_int event)
456 {
457 
458 	g_raid_tr_update_state_raid1(tr->tro_volume, sd);
459 	return (0);
460 }
461 
462 static int
463 g_raid_tr_start_raid1(struct g_raid_tr_object *tr)
464 {
465 	struct g_raid_tr_raid1_object *trs;
466 	struct g_raid_volume *vol;
467 
468 	trs = (struct g_raid_tr_raid1_object *)tr;
469 	vol = tr->tro_volume;
470 	trs->trso_starting = 0;
471 	g_raid_tr_update_state_raid1(vol, NULL);
472 	return (0);
473 }
474 
475 static int
476 g_raid_tr_stop_raid1(struct g_raid_tr_object *tr)
477 {
478 	struct g_raid_tr_raid1_object *trs;
479 	struct g_raid_volume *vol;
480 
481 	trs = (struct g_raid_tr_raid1_object *)tr;
482 	vol = tr->tro_volume;
483 	trs->trso_starting = 0;
484 	trs->trso_stopping = 1;
485 	g_raid_tr_update_state_raid1(vol, NULL);
486 	return (0);
487 }
488 
489 /*
490  * Select the disk to read from.  Take into account: subdisk state, running
491  * error recovery, average disk load, head position and possible cache hits.
492  */
493 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
494 static struct g_raid_subdisk *
495 g_raid_tr_raid1_select_read_disk(struct g_raid_volume *vol, struct bio *bp,
496     u_int mask)
497 {
498 	struct g_raid_subdisk *sd, *best;
499 	int i, prio, bestprio;
500 
501 	best = NULL;
502 	bestprio = INT_MAX;
503 	for (i = 0; i < vol->v_disks_count; i++) {
504 		sd = &vol->v_subdisks[i];
505 		if (sd->sd_state != G_RAID_SUBDISK_S_ACTIVE &&
506 		    ((sd->sd_state != G_RAID_SUBDISK_S_REBUILD &&
507 		      sd->sd_state != G_RAID_SUBDISK_S_RESYNC) ||
508 		     bp->bio_offset + bp->bio_length > sd->sd_rebuild_pos))
509 			continue;
510 		if ((mask & (1 << i)) != 0)
511 			continue;
512 		prio = G_RAID_SUBDISK_LOAD(sd);
513 		prio += min(sd->sd_recovery, 255) << 22;
514 		prio += (G_RAID_SUBDISK_S_ACTIVE - sd->sd_state) << 16;
515 		/* If disk head is precisely in position - highly prefer it. */
516 		if (G_RAID_SUBDISK_POS(sd) == bp->bio_offset)
517 			prio -= 2 * G_RAID_SUBDISK_LOAD_SCALE;
518 		else
519 		/* If disk head is close to position - prefer it. */
520 		if (ABS(G_RAID_SUBDISK_POS(sd) - bp->bio_offset) <
521 		    G_RAID_SUBDISK_TRACK_SIZE)
522 			prio -= 1 * G_RAID_SUBDISK_LOAD_SCALE;
523 		if (prio < bestprio) {
524 			best = sd;
525 			bestprio = prio;
526 		}
527 	}
528 	return (best);
529 }
530 
531 static void
532 g_raid_tr_iostart_raid1_read(struct g_raid_tr_object *tr, struct bio *bp)
533 {
534 	struct g_raid_subdisk *sd;
535 	struct bio *cbp;
536 
537 	sd = g_raid_tr_raid1_select_read_disk(tr->tro_volume, bp, 0);
538 	KASSERT(sd != NULL, ("No active disks in volume %s.",
539 		tr->tro_volume->v_name));
540 
541 	cbp = g_clone_bio(bp);
542 	if (cbp == NULL) {
543 		g_raid_iodone(bp, ENOMEM);
544 		return;
545 	}
546 
547 	g_raid_subdisk_iostart(sd, cbp);
548 }
549 
550 static void
551 g_raid_tr_iostart_raid1_write(struct g_raid_tr_object *tr, struct bio *bp)
552 {
553 	struct g_raid_volume *vol;
554 	struct g_raid_subdisk *sd;
555 	struct bio_queue_head queue;
556 	struct bio *cbp;
557 	int i;
558 
559 	vol = tr->tro_volume;
560 
561 	/*
562 	 * Allocate all bios before sending any request, so we can return
563 	 * ENOMEM in nice and clean way.
564 	 */
565 	bioq_init(&queue);
566 	for (i = 0; i < vol->v_disks_count; i++) {
567 		sd = &vol->v_subdisks[i];
568 		switch (sd->sd_state) {
569 		case G_RAID_SUBDISK_S_ACTIVE:
570 			break;
571 		case G_RAID_SUBDISK_S_REBUILD:
572 			/*
573 			 * When rebuilding, only part of this subdisk is
574 			 * writable, the rest will be written as part of the
575 			 * that process.
576 			 */
577 			if (bp->bio_offset >= sd->sd_rebuild_pos)
578 				continue;
579 			break;
580 		case G_RAID_SUBDISK_S_STALE:
581 		case G_RAID_SUBDISK_S_RESYNC:
582 			/*
583 			 * Resyncing still writes on the theory that the
584 			 * resync'd disk is very close and writing it will
585 			 * keep it that way better if we keep up while
586 			 * resyncing.
587 			 */
588 			break;
589 		default:
590 			continue;
591 		}
592 		cbp = g_clone_bio(bp);
593 		if (cbp == NULL)
594 			goto failure;
595 		cbp->bio_caller1 = sd;
596 		bioq_insert_tail(&queue, cbp);
597 	}
598 	for (cbp = bioq_first(&queue); cbp != NULL;
599 	    cbp = bioq_first(&queue)) {
600 		bioq_remove(&queue, cbp);
601 		sd = cbp->bio_caller1;
602 		cbp->bio_caller1 = NULL;
603 		g_raid_subdisk_iostart(sd, cbp);
604 	}
605 	return;
606 failure:
607 	for (cbp = bioq_first(&queue); cbp != NULL;
608 	    cbp = bioq_first(&queue)) {
609 		bioq_remove(&queue, cbp);
610 		g_destroy_bio(cbp);
611 	}
612 	if (bp->bio_error == 0)
613 		bp->bio_error = ENOMEM;
614 	g_raid_iodone(bp, bp->bio_error);
615 }
616 
617 static void
618 g_raid_tr_iostart_raid1(struct g_raid_tr_object *tr, struct bio *bp)
619 {
620 	struct g_raid_volume *vol;
621 	struct g_raid_tr_raid1_object *trs;
622 
623 	vol = tr->tro_volume;
624 	trs = (struct g_raid_tr_raid1_object *)tr;
625 	if (vol->v_state != G_RAID_VOLUME_S_OPTIMAL &&
626 	    vol->v_state != G_RAID_VOLUME_S_SUBOPTIMAL &&
627 	    vol->v_state != G_RAID_VOLUME_S_DEGRADED) {
628 		g_raid_iodone(bp, EIO);
629 		return;
630 	}
631 	/*
632 	 * If we're rebuilding, squeeze in rebuild activity every so often,
633 	 * even when the disk is busy.  Be sure to only count real I/O
634 	 * to the disk.  All 'SPECIAL' I/O is traffic generated to the disk
635 	 * by this module.
636 	 */
637 	if (trs->trso_failed_sd != NULL &&
638 	    !(bp->bio_cflags & G_RAID_BIO_FLAG_SPECIAL)) {
639 		/* Make this new or running now round short. */
640 		trs->trso_recover_slabs = 0;
641 		if (--trs->trso_fair_io <= 0) {
642 			trs->trso_fair_io = g_raid1_rebuild_fair_io;
643 			g_raid_tr_raid1_rebuild_some(tr);
644 		}
645 	}
646 	switch (bp->bio_cmd) {
647 	case BIO_READ:
648 		g_raid_tr_iostart_raid1_read(tr, bp);
649 		break;
650 	case BIO_WRITE:
651 		g_raid_tr_iostart_raid1_write(tr, bp);
652 		break;
653 	case BIO_DELETE:
654 		g_raid_iodone(bp, EIO);
655 		break;
656 	case BIO_FLUSH:
657 		g_raid_tr_flush_common(tr, bp);
658 		break;
659 	default:
660 		KASSERT(1 == 0, ("Invalid command here: %u (volume=%s)",
661 		    bp->bio_cmd, vol->v_name));
662 		break;
663 	}
664 }
665 
666 static void
667 g_raid_tr_iodone_raid1(struct g_raid_tr_object *tr,
668     struct g_raid_subdisk *sd, struct bio *bp)
669 {
670 	struct bio *cbp;
671 	struct g_raid_subdisk *nsd;
672 	struct g_raid_volume *vol;
673 	struct bio *pbp;
674 	struct g_raid_tr_raid1_object *trs;
675 	uintptr_t *mask;
676 	int error, do_write;
677 
678 	trs = (struct g_raid_tr_raid1_object *)tr;
679 	vol = tr->tro_volume;
680 	if (bp->bio_cflags & G_RAID_BIO_FLAG_SYNC) {
681 		/*
682 		 * This operation is part of a rebuild or resync operation.
683 		 * See what work just got done, then schedule the next bit of
684 		 * work, if any.  Rebuild/resync is done a little bit at a
685 		 * time.  Either when a timeout happens, or after we get a
686 		 * bunch of I/Os to the disk (to make sure an active system
687 		 * will complete in a sane amount of time).
688 		 *
689 		 * We are setup to do differing amounts of work for each of
690 		 * these cases.  so long as the slabs is smallish (less than
691 		 * 50 or so, I'd guess, but that's just a WAG), we shouldn't
692 		 * have any bio starvation issues.  For active disks, we do
693 		 * 5MB of data, for inactive ones, we do 50MB.
694 		 */
695 		if (trs->trso_type == TR_RAID1_REBUILD) {
696 			if (bp->bio_cmd == BIO_READ) {
697 
698 				/* Immediately abort rebuild, if requested. */
699 				if (trs->trso_flags & TR_RAID1_F_ABORT) {
700 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
701 					g_raid_tr_raid1_rebuild_abort(tr);
702 					return;
703 				}
704 
705 				/* On read error, skip and cross fingers. */
706 				if (bp->bio_error != 0) {
707 					G_RAID_LOGREQ(0, bp,
708 					    "Read error during rebuild (%d), "
709 					    "possible data loss!",
710 					    bp->bio_error);
711 					goto rebuild_round_done;
712 				}
713 
714 				/*
715 				 * The read operation finished, queue the
716 				 * write and get out.
717 				 */
718 				G_RAID_LOGREQ(4, bp, "rebuild read done. %d",
719 				    bp->bio_error);
720 				bp->bio_cmd = BIO_WRITE;
721 				bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
722 				G_RAID_LOGREQ(4, bp, "Queueing rebuild write.");
723 				g_raid_subdisk_iostart(trs->trso_failed_sd, bp);
724 			} else {
725 				/*
726 				 * The write operation just finished.  Do
727 				 * another.  We keep cloning the master bio
728 				 * since it has the right buffers allocated to
729 				 * it.
730 				 */
731 				G_RAID_LOGREQ(4, bp,
732 				    "rebuild write done. Error %d",
733 				    bp->bio_error);
734 				nsd = trs->trso_failed_sd;
735 				if (bp->bio_error != 0 ||
736 				    trs->trso_flags & TR_RAID1_F_ABORT) {
737 					if ((trs->trso_flags &
738 					    TR_RAID1_F_ABORT) == 0) {
739 						g_raid_tr_raid1_fail_disk(sd->sd_softc,
740 						    nsd, nsd->sd_disk);
741 					}
742 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
743 					g_raid_tr_raid1_rebuild_abort(tr);
744 					return;
745 				}
746 rebuild_round_done:
747 				nsd = trs->trso_failed_sd;
748 				trs->trso_flags &= ~TR_RAID1_F_LOCKED;
749 				g_raid_unlock_range(sd->sd_volume,
750 				    bp->bio_offset, bp->bio_length);
751 				nsd->sd_rebuild_pos += bp->bio_length;
752 				if (nsd->sd_rebuild_pos >= nsd->sd_size) {
753 					g_raid_tr_raid1_rebuild_finish(tr);
754 					return;
755 				}
756 
757 				/* Abort rebuild if we are stopping */
758 				if (trs->trso_stopping) {
759 					trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
760 					g_raid_tr_raid1_rebuild_abort(tr);
761 					return;
762 				}
763 
764 				if (--trs->trso_meta_update <= 0) {
765 					g_raid_write_metadata(vol->v_softc,
766 					    vol, nsd, nsd->sd_disk);
767 					trs->trso_meta_update =
768 					    g_raid1_rebuild_meta_update;
769 				}
770 				trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
771 				if (--trs->trso_recover_slabs <= 0)
772 					return;
773 				g_raid_tr_raid1_rebuild_some(tr);
774 			}
775 		} else if (trs->trso_type == TR_RAID1_RESYNC) {
776 			/*
777 			 * read good sd, read bad sd in parallel.  when both
778 			 * done, compare the buffers.  write good to the bad
779 			 * if different.  do the next bit of work.
780 			 */
781 			panic("Somehow, we think we're doing a resync");
782 		}
783 		return;
784 	}
785 	pbp = bp->bio_parent;
786 	pbp->bio_inbed++;
787 	if (bp->bio_cmd == BIO_READ && bp->bio_error != 0) {
788 		/*
789 		 * Read failed on first drive.  Retry the read error on
790 		 * another disk drive, if available, before erroring out the
791 		 * read.
792 		 */
793 		sd->sd_disk->d_read_errs++;
794 		G_RAID_LOGREQ(0, bp,
795 		    "Read error (%d), %d read errors total",
796 		    bp->bio_error, sd->sd_disk->d_read_errs);
797 
798 		/*
799 		 * If there are too many read errors, we move to degraded.
800 		 * XXX Do we want to FAIL the drive (eg, make the user redo
801 		 * everything to get it back in sync), or just degrade the
802 		 * drive, which kicks off a resync?
803 		 */
804 		do_write = 1;
805 		if (sd->sd_disk->d_read_errs > g_raid_read_err_thresh) {
806 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
807 			if (pbp->bio_children == 1)
808 				do_write = 0;
809 		}
810 
811 		/*
812 		 * Find the other disk, and try to do the I/O to it.
813 		 */
814 		mask = (uintptr_t *)(&pbp->bio_driver2);
815 		if (pbp->bio_children == 1) {
816 			/* Save original subdisk. */
817 			pbp->bio_driver1 = do_write ? sd : NULL;
818 			*mask = 0;
819 		}
820 		*mask |= 1 << sd->sd_pos;
821 		nsd = g_raid_tr_raid1_select_read_disk(vol, pbp, *mask);
822 		if (nsd != NULL && (cbp = g_clone_bio(pbp)) != NULL) {
823 			g_destroy_bio(bp);
824 			G_RAID_LOGREQ(2, cbp, "Retrying read from %d",
825 			    nsd->sd_pos);
826 			if (pbp->bio_children == 2 && do_write) {
827 				sd->sd_recovery++;
828 				cbp->bio_caller1 = nsd;
829 				pbp->bio_pflags = G_RAID_BIO_FLAG_LOCKED;
830 				/* Lock callback starts I/O */
831 				g_raid_lock_range(sd->sd_volume,
832 				    cbp->bio_offset, cbp->bio_length, pbp, cbp);
833 			} else {
834 				g_raid_subdisk_iostart(nsd, cbp);
835 			}
836 			return;
837 		}
838 		/*
839 		 * We can't retry.  Return the original error by falling
840 		 * through.  This will happen when there's only one good disk.
841 		 * We don't need to fail the raid, since its actual state is
842 		 * based on the state of the subdisks.
843 		 */
844 		G_RAID_LOGREQ(2, bp, "Couldn't retry read, failing it");
845 	}
846 	if (bp->bio_cmd == BIO_READ &&
847 	    bp->bio_error == 0 &&
848 	    pbp->bio_children > 1 &&
849 	    pbp->bio_driver1 != NULL) {
850 		/*
851 		 * If it was a read, and bio_children is >1, then we just
852 		 * recovered the data from the second drive.  We should try to
853 		 * write that data to the first drive if sector remapping is
854 		 * enabled.  A write should put the data in a new place on the
855 		 * disk, remapping the bad sector.  Do we need to do that by
856 		 * queueing a request to the main worker thread?  It doesn't
857 		 * affect the return code of this current read, and can be
858 		 * done at our liesure.  However, to make the code simpler, it
859 		 * is done syncrhonously.
860 		 */
861 		G_RAID_LOGREQ(3, bp, "Recovered data from other drive");
862 		cbp = g_clone_bio(pbp);
863 		if (cbp != NULL) {
864 			g_destroy_bio(bp);
865 			cbp->bio_cmd = BIO_WRITE;
866 			cbp->bio_cflags = G_RAID_BIO_FLAG_REMAP;
867 			G_RAID_LOGREQ(2, cbp,
868 			    "Attempting bad sector remap on failing drive.");
869 			g_raid_subdisk_iostart(pbp->bio_driver1, cbp);
870 			return;
871 		}
872 	}
873 	if (pbp->bio_pflags & G_RAID_BIO_FLAG_LOCKED) {
874 		/*
875 		 * We're done with a recovery, mark the range as unlocked.
876 		 * For any write errors, we agressively fail the disk since
877 		 * there was both a READ and a WRITE error at this location.
878 		 * Both types of errors generally indicates the drive is on
879 		 * the verge of total failure anyway.  Better to stop trusting
880 		 * it now.  However, we need to reset error to 0 in that case
881 		 * because we're not failing the original I/O which succeeded.
882 		 */
883 		if (bp->bio_cmd == BIO_WRITE && bp->bio_error) {
884 			G_RAID_LOGREQ(0, bp, "Remap write failed: "
885 			    "failing subdisk.");
886 			g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
887 			bp->bio_error = 0;
888 		}
889 		if (pbp->bio_driver1 != NULL) {
890 			((struct g_raid_subdisk *)pbp->bio_driver1)
891 			    ->sd_recovery--;
892 		}
893 		G_RAID_LOGREQ(2, bp, "REMAP done %d.", bp->bio_error);
894 		g_raid_unlock_range(sd->sd_volume, bp->bio_offset,
895 		    bp->bio_length);
896 	}
897 	error = bp->bio_error;
898 	g_destroy_bio(bp);
899 	if (pbp->bio_children == pbp->bio_inbed) {
900 		pbp->bio_completed = pbp->bio_length;
901 		g_raid_iodone(pbp, error);
902 	}
903 }
904 
905 static int
906 g_raid_tr_kerneldump_raid1(struct g_raid_tr_object *tr,
907     void *virtual, vm_offset_t physical, off_t offset, size_t length)
908 {
909 	struct g_raid_volume *vol;
910 	struct g_raid_subdisk *sd;
911 	int error, i, ok;
912 
913 	vol = tr->tro_volume;
914 	error = 0;
915 	ok = 0;
916 	for (i = 0; i < vol->v_disks_count; i++) {
917 		sd = &vol->v_subdisks[i];
918 		switch (sd->sd_state) {
919 		case G_RAID_SUBDISK_S_ACTIVE:
920 			break;
921 		case G_RAID_SUBDISK_S_REBUILD:
922 			/*
923 			 * When rebuilding, only part of this subdisk is
924 			 * writable, the rest will be written as part of the
925 			 * that process.
926 			 */
927 			if (offset >= sd->sd_rebuild_pos)
928 				continue;
929 			break;
930 		case G_RAID_SUBDISK_S_STALE:
931 		case G_RAID_SUBDISK_S_RESYNC:
932 			/*
933 			 * Resyncing still writes on the theory that the
934 			 * resync'd disk is very close and writing it will
935 			 * keep it that way better if we keep up while
936 			 * resyncing.
937 			 */
938 			break;
939 		default:
940 			continue;
941 		}
942 		error = g_raid_subdisk_kerneldump(sd,
943 		    virtual, physical, offset, length);
944 		if (error == 0)
945 			ok++;
946 	}
947 	return (ok > 0 ? 0 : error);
948 }
949 
950 static int
951 g_raid_tr_locked_raid1(struct g_raid_tr_object *tr, void *argp)
952 {
953 	struct bio *bp;
954 	struct g_raid_subdisk *sd;
955 
956 	bp = (struct bio *)argp;
957 	sd = (struct g_raid_subdisk *)bp->bio_caller1;
958 	g_raid_subdisk_iostart(sd, bp);
959 
960 	return (0);
961 }
962 
963 static int
964 g_raid_tr_idle_raid1(struct g_raid_tr_object *tr)
965 {
966 	struct g_raid_tr_raid1_object *trs;
967 
968 	trs = (struct g_raid_tr_raid1_object *)tr;
969 	trs->trso_fair_io = g_raid1_rebuild_fair_io;
970 	trs->trso_recover_slabs = g_raid1_rebuild_cluster_idle;
971 	if (trs->trso_type == TR_RAID1_REBUILD)
972 		g_raid_tr_raid1_rebuild_some(tr);
973 	return (0);
974 }
975 
976 static int
977 g_raid_tr_free_raid1(struct g_raid_tr_object *tr)
978 {
979 	struct g_raid_tr_raid1_object *trs;
980 
981 	trs = (struct g_raid_tr_raid1_object *)tr;
982 
983 	if (trs->trso_buffer != NULL) {
984 		free(trs->trso_buffer, M_TR_RAID1);
985 		trs->trso_buffer = NULL;
986 	}
987 	return (0);
988 }
989 
990 G_RAID_TR_DECLARE(g_raid_tr_raid1);
991