xref: /freebsd/sys/geom/raid/tr_raid1e.c (revision 9768746ba83efa02837c5b9c66348db6e900208f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 #define N	2
49 
50 SYSCTL_DECL(_kern_geom_raid_raid1e);
51 
52 #define RAID1E_REBUILD_SLAB	(1 << 20) /* One transation in a rebuild */
53 static int g_raid1e_rebuild_slab = RAID1E_REBUILD_SLAB;
54 SYSCTL_UINT(_kern_geom_raid_raid1e, OID_AUTO, rebuild_slab_size, CTLFLAG_RWTUN,
55     &g_raid1e_rebuild_slab, 0,
56     "Amount of the disk to rebuild each read/write cycle of the rebuild.");
57 
58 #define RAID1E_REBUILD_FAIR_IO 20 /* use 1/x of the available I/O */
59 static int g_raid1e_rebuild_fair_io = RAID1E_REBUILD_FAIR_IO;
60 SYSCTL_UINT(_kern_geom_raid_raid1e, OID_AUTO, rebuild_fair_io, CTLFLAG_RWTUN,
61     &g_raid1e_rebuild_fair_io, 0,
62     "Fraction of the I/O bandwidth to use when disk busy for rebuild.");
63 
64 #define RAID1E_REBUILD_CLUSTER_IDLE 100
65 static int g_raid1e_rebuild_cluster_idle = RAID1E_REBUILD_CLUSTER_IDLE;
66 SYSCTL_UINT(_kern_geom_raid_raid1e, OID_AUTO, rebuild_cluster_idle, CTLFLAG_RWTUN,
67     &g_raid1e_rebuild_cluster_idle, 0,
68     "Number of slabs to do each time we trigger a rebuild cycle");
69 
70 #define RAID1E_REBUILD_META_UPDATE 1024 /* update meta data every 1GB or so */
71 static int g_raid1e_rebuild_meta_update = RAID1E_REBUILD_META_UPDATE;
72 SYSCTL_UINT(_kern_geom_raid_raid1e, OID_AUTO, rebuild_meta_update, CTLFLAG_RWTUN,
73     &g_raid1e_rebuild_meta_update, 0,
74     "When to update the meta data.");
75 
76 static MALLOC_DEFINE(M_TR_RAID1E, "tr_raid1e_data", "GEOM_RAID RAID1E data");
77 
78 #define TR_RAID1E_NONE 0
79 #define TR_RAID1E_REBUILD 1
80 #define TR_RAID1E_RESYNC 2
81 
82 #define TR_RAID1E_F_DOING_SOME	0x1
83 #define TR_RAID1E_F_LOCKED	0x2
84 #define TR_RAID1E_F_ABORT	0x4
85 
86 struct g_raid_tr_raid1e_object {
87 	struct g_raid_tr_object	 trso_base;
88 	int			 trso_starting;
89 	int			 trso_stopping;
90 	int			 trso_type;
91 	int			 trso_recover_slabs; /* slabs before rest */
92 	int			 trso_fair_io;
93 	int			 trso_meta_update;
94 	int			 trso_flags;
95 	struct g_raid_subdisk	*trso_failed_sd; /* like per volume */
96 	void			*trso_buffer;	 /* Buffer space */
97 	off_t			 trso_lock_pos; /* Locked range start. */
98 	off_t			 trso_lock_len; /* Locked range length. */
99 	struct bio		 trso_bio;
100 };
101 
102 static g_raid_tr_taste_t g_raid_tr_taste_raid1e;
103 static g_raid_tr_event_t g_raid_tr_event_raid1e;
104 static g_raid_tr_start_t g_raid_tr_start_raid1e;
105 static g_raid_tr_stop_t g_raid_tr_stop_raid1e;
106 static g_raid_tr_iostart_t g_raid_tr_iostart_raid1e;
107 static g_raid_tr_iodone_t g_raid_tr_iodone_raid1e;
108 static g_raid_tr_kerneldump_t g_raid_tr_kerneldump_raid1e;
109 static g_raid_tr_locked_t g_raid_tr_locked_raid1e;
110 static g_raid_tr_idle_t g_raid_tr_idle_raid1e;
111 static g_raid_tr_free_t g_raid_tr_free_raid1e;
112 
113 static kobj_method_t g_raid_tr_raid1e_methods[] = {
114 	KOBJMETHOD(g_raid_tr_taste,	g_raid_tr_taste_raid1e),
115 	KOBJMETHOD(g_raid_tr_event,	g_raid_tr_event_raid1e),
116 	KOBJMETHOD(g_raid_tr_start,	g_raid_tr_start_raid1e),
117 	KOBJMETHOD(g_raid_tr_stop,	g_raid_tr_stop_raid1e),
118 	KOBJMETHOD(g_raid_tr_iostart,	g_raid_tr_iostart_raid1e),
119 	KOBJMETHOD(g_raid_tr_iodone,	g_raid_tr_iodone_raid1e),
120 	KOBJMETHOD(g_raid_tr_kerneldump, g_raid_tr_kerneldump_raid1e),
121 	KOBJMETHOD(g_raid_tr_locked,	g_raid_tr_locked_raid1e),
122 	KOBJMETHOD(g_raid_tr_idle,	g_raid_tr_idle_raid1e),
123 	KOBJMETHOD(g_raid_tr_free,	g_raid_tr_free_raid1e),
124 	{ 0, 0 }
125 };
126 
127 static struct g_raid_tr_class g_raid_tr_raid1e_class = {
128 	"RAID1E",
129 	g_raid_tr_raid1e_methods,
130 	sizeof(struct g_raid_tr_raid1e_object),
131 	.trc_enable = 1,
132 	.trc_priority = 200,
133 	.trc_accept_unmapped = 1
134 };
135 
136 static void g_raid_tr_raid1e_rebuild_abort(struct g_raid_tr_object *tr);
137 static void g_raid_tr_raid1e_maybe_rebuild(struct g_raid_tr_object *tr,
138     struct g_raid_subdisk *sd);
139 static int g_raid_tr_raid1e_select_read_disk(struct g_raid_volume *vol,
140     int no, off_t off, off_t len, u_int mask);
141 
142 static inline void
143 V2P(struct g_raid_volume *vol, off_t virt,
144     int *disk, off_t *offset, off_t *start)
145 {
146 	off_t nstrip;
147 	u_int strip_size;
148 
149 	strip_size = vol->v_strip_size;
150 	/* Strip number. */
151 	nstrip = virt / strip_size;
152 	/* Start position in strip. */
153 	*start = virt % strip_size;
154 	/* Disk number. */
155 	*disk = (nstrip * N) % vol->v_disks_count;
156 	/* Strip start position in disk. */
157 	*offset = ((nstrip * N) / vol->v_disks_count) * strip_size;
158 }
159 
160 static inline void
161 P2V(struct g_raid_volume *vol, int disk, off_t offset,
162     off_t *virt, int *copy)
163 {
164 	off_t nstrip, start;
165 	u_int strip_size;
166 
167 	strip_size = vol->v_strip_size;
168 	/* Start position in strip. */
169 	start = offset % strip_size;
170 	/* Physical strip number. */
171 	nstrip = (offset / strip_size) * vol->v_disks_count + disk;
172 	/* Number of physical strip (copy) inside virtual strip. */
173 	*copy = nstrip % N;
174 	/* Offset in virtual space. */
175 	*virt = (nstrip / N) * strip_size + start;
176 }
177 
178 static int
179 g_raid_tr_taste_raid1e(struct g_raid_tr_object *tr, struct g_raid_volume *vol)
180 {
181 	struct g_raid_tr_raid1e_object *trs;
182 
183 	trs = (struct g_raid_tr_raid1e_object *)tr;
184 	if (tr->tro_volume->v_raid_level != G_RAID_VOLUME_RL_RAID1E ||
185 	    tr->tro_volume->v_raid_level_qualifier != G_RAID_VOLUME_RLQ_R1EA)
186 		return (G_RAID_TR_TASTE_FAIL);
187 	trs->trso_starting = 1;
188 	return (G_RAID_TR_TASTE_SUCCEED);
189 }
190 
191 static int
192 g_raid_tr_update_state_raid1e_even(struct g_raid_volume *vol)
193 {
194 	struct g_raid_softc *sc;
195 	struct g_raid_subdisk *sd, *bestsd, *worstsd;
196 	int i, j, state, sstate;
197 
198 	sc = vol->v_softc;
199 	state = G_RAID_VOLUME_S_OPTIMAL;
200 	for (i = 0; i < vol->v_disks_count / N; i++) {
201 		bestsd = &vol->v_subdisks[i * N];
202 		for (j = 1; j < N; j++) {
203 			sd = &vol->v_subdisks[i * N + j];
204 			if (sd->sd_state > bestsd->sd_state)
205 				bestsd = sd;
206 			else if (sd->sd_state == bestsd->sd_state &&
207 			    (sd->sd_state == G_RAID_SUBDISK_S_REBUILD ||
208 			     sd->sd_state == G_RAID_SUBDISK_S_RESYNC) &&
209 			    sd->sd_rebuild_pos > bestsd->sd_rebuild_pos)
210 				bestsd = sd;
211 		}
212 		if (bestsd->sd_state >= G_RAID_SUBDISK_S_UNINITIALIZED &&
213 		    bestsd->sd_state != G_RAID_SUBDISK_S_ACTIVE) {
214 			/* We found reasonable candidate. */
215 			G_RAID_DEBUG1(1, sc,
216 			    "Promote subdisk %s:%d from %s to ACTIVE.",
217 			    vol->v_name, bestsd->sd_pos,
218 			    g_raid_subdisk_state2str(bestsd->sd_state));
219 			g_raid_change_subdisk_state(bestsd,
220 			    G_RAID_SUBDISK_S_ACTIVE);
221 			g_raid_write_metadata(sc,
222 			    vol, bestsd, bestsd->sd_disk);
223 		}
224 		worstsd = &vol->v_subdisks[i * N];
225 		for (j = 1; j < N; j++) {
226 			sd = &vol->v_subdisks[i * N + j];
227 			if (sd->sd_state < worstsd->sd_state)
228 				worstsd = sd;
229 		}
230 		if (worstsd->sd_state == G_RAID_SUBDISK_S_ACTIVE)
231 			sstate = G_RAID_VOLUME_S_OPTIMAL;
232 		else if (worstsd->sd_state >= G_RAID_SUBDISK_S_STALE)
233 			sstate = G_RAID_VOLUME_S_SUBOPTIMAL;
234 		else if (bestsd->sd_state == G_RAID_SUBDISK_S_ACTIVE)
235 			sstate = G_RAID_VOLUME_S_DEGRADED;
236 		else
237 			sstate = G_RAID_VOLUME_S_BROKEN;
238 		if (sstate < state)
239 			state = sstate;
240 	}
241 	return (state);
242 }
243 
244 static int
245 g_raid_tr_update_state_raid1e_odd(struct g_raid_volume *vol)
246 {
247 	struct g_raid_softc *sc;
248 	struct g_raid_subdisk *sd, *bestsd, *worstsd;
249 	int i, j, state, sstate;
250 
251 	sc = vol->v_softc;
252 	if (g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE) ==
253 	    vol->v_disks_count)
254 		return (G_RAID_VOLUME_S_OPTIMAL);
255 	for (i = 0; i < vol->v_disks_count; i++) {
256 		sd = &vol->v_subdisks[i];
257 		if (sd->sd_state == G_RAID_SUBDISK_S_UNINITIALIZED) {
258 			/* We found reasonable candidate. */
259 			G_RAID_DEBUG1(1, sc,
260 			    "Promote subdisk %s:%d from %s to STALE.",
261 			    vol->v_name, sd->sd_pos,
262 			    g_raid_subdisk_state2str(sd->sd_state));
263 			g_raid_change_subdisk_state(sd,
264 			    G_RAID_SUBDISK_S_STALE);
265 			g_raid_write_metadata(sc, vol, sd, sd->sd_disk);
266 		}
267 	}
268 	state = G_RAID_VOLUME_S_OPTIMAL;
269 	for (i = 0; i < vol->v_disks_count; i++) {
270 		bestsd = &vol->v_subdisks[i];
271 		worstsd = &vol->v_subdisks[i];
272 		for (j = 1; j < N; j++) {
273 			sd = &vol->v_subdisks[(i + j) % vol->v_disks_count];
274 			if (sd->sd_state > bestsd->sd_state)
275 				bestsd = sd;
276 			else if (sd->sd_state == bestsd->sd_state &&
277 			    (sd->sd_state == G_RAID_SUBDISK_S_REBUILD ||
278 			     sd->sd_state == G_RAID_SUBDISK_S_RESYNC) &&
279 			    sd->sd_rebuild_pos > bestsd->sd_rebuild_pos)
280 				bestsd = sd;
281 			if (sd->sd_state < worstsd->sd_state)
282 				worstsd = sd;
283 		}
284 		if (worstsd->sd_state == G_RAID_SUBDISK_S_ACTIVE)
285 			sstate = G_RAID_VOLUME_S_OPTIMAL;
286 		else if (worstsd->sd_state >= G_RAID_SUBDISK_S_STALE)
287 			sstate = G_RAID_VOLUME_S_SUBOPTIMAL;
288 		else if (bestsd->sd_state >= G_RAID_SUBDISK_S_STALE)
289 			sstate = G_RAID_VOLUME_S_DEGRADED;
290 		else
291 			sstate = G_RAID_VOLUME_S_BROKEN;
292 		if (sstate < state)
293 			state = sstate;
294 	}
295 	return (state);
296 }
297 
298 static int
299 g_raid_tr_update_state_raid1e(struct g_raid_volume *vol,
300     struct g_raid_subdisk *sd)
301 {
302 	struct g_raid_tr_raid1e_object *trs;
303 	struct g_raid_softc *sc;
304 	u_int s;
305 
306 	sc = vol->v_softc;
307 	trs = (struct g_raid_tr_raid1e_object *)vol->v_tr;
308 	if (trs->trso_stopping &&
309 	    (trs->trso_flags & TR_RAID1E_F_DOING_SOME) == 0)
310 		s = G_RAID_VOLUME_S_STOPPED;
311 	else if (trs->trso_starting)
312 		s = G_RAID_VOLUME_S_STARTING;
313 	else {
314 		if ((vol->v_disks_count % N) == 0)
315 			s = g_raid_tr_update_state_raid1e_even(vol);
316 		else
317 			s = g_raid_tr_update_state_raid1e_odd(vol);
318 	}
319 	if (s != vol->v_state) {
320 		g_raid_event_send(vol, G_RAID_VOLUME_S_ALIVE(s) ?
321 		    G_RAID_VOLUME_E_UP : G_RAID_VOLUME_E_DOWN,
322 		    G_RAID_EVENT_VOLUME);
323 		g_raid_change_volume_state(vol, s);
324 		if (!trs->trso_starting && !trs->trso_stopping)
325 			g_raid_write_metadata(sc, vol, NULL, NULL);
326 	}
327 	if (!trs->trso_starting && !trs->trso_stopping)
328 		g_raid_tr_raid1e_maybe_rebuild(vol->v_tr, sd);
329 	return (0);
330 }
331 
332 static void
333 g_raid_tr_raid1e_fail_disk(struct g_raid_softc *sc, struct g_raid_subdisk *sd,
334     struct g_raid_disk *disk)
335 {
336 	struct g_raid_volume *vol;
337 
338 	vol = sd->sd_volume;
339 	/*
340 	 * We don't fail the last disk in the pack, since it still has decent
341 	 * data on it and that's better than failing the disk if it is the root
342 	 * file system.
343 	 *
344 	 * XXX should this be controlled via a tunable?  It makes sense for
345 	 * the volume that has / on it.  I can't think of a case where we'd
346 	 * want the volume to go away on this kind of event.
347 	 */
348 	if ((g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE) +
349 	     g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC) +
350 	     g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
351 	     g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_UNINITIALIZED) <
352 	     vol->v_disks_count) &&
353 	    (sd->sd_state >= G_RAID_SUBDISK_S_UNINITIALIZED))
354 		return;
355 	g_raid_fail_disk(sc, sd, disk);
356 }
357 
358 static void
359 g_raid_tr_raid1e_rebuild_done(struct g_raid_tr_raid1e_object *trs)
360 {
361 	struct g_raid_volume *vol;
362 	struct g_raid_subdisk *sd;
363 
364 	vol = trs->trso_base.tro_volume;
365 	sd = trs->trso_failed_sd;
366 	g_raid_write_metadata(vol->v_softc, vol, sd, sd->sd_disk);
367 	free(trs->trso_buffer, M_TR_RAID1E);
368 	trs->trso_buffer = NULL;
369 	trs->trso_flags &= ~TR_RAID1E_F_DOING_SOME;
370 	trs->trso_type = TR_RAID1E_NONE;
371 	trs->trso_recover_slabs = 0;
372 	trs->trso_failed_sd = NULL;
373 	g_raid_tr_update_state_raid1e(vol, NULL);
374 }
375 
376 static void
377 g_raid_tr_raid1e_rebuild_finish(struct g_raid_tr_object *tr)
378 {
379 	struct g_raid_tr_raid1e_object *trs;
380 	struct g_raid_subdisk *sd;
381 
382 	trs = (struct g_raid_tr_raid1e_object *)tr;
383 	sd = trs->trso_failed_sd;
384 	G_RAID_DEBUG1(0, tr->tro_volume->v_softc,
385 	    "Subdisk %s:%d-%s rebuild completed.",
386 	    sd->sd_volume->v_name, sd->sd_pos,
387 	    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
388 	g_raid_change_subdisk_state(sd, G_RAID_SUBDISK_S_ACTIVE);
389 	sd->sd_rebuild_pos = 0;
390 	g_raid_tr_raid1e_rebuild_done(trs);
391 }
392 
393 static void
394 g_raid_tr_raid1e_rebuild_abort(struct g_raid_tr_object *tr)
395 {
396 	struct g_raid_tr_raid1e_object *trs;
397 	struct g_raid_subdisk *sd;
398 	struct g_raid_volume *vol;
399 
400 	vol = tr->tro_volume;
401 	trs = (struct g_raid_tr_raid1e_object *)tr;
402 	sd = trs->trso_failed_sd;
403 	if (trs->trso_flags & TR_RAID1E_F_DOING_SOME) {
404 		G_RAID_DEBUG1(1, vol->v_softc,
405 		    "Subdisk %s:%d-%s rebuild is aborting.",
406 		    sd->sd_volume->v_name, sd->sd_pos,
407 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
408 		trs->trso_flags |= TR_RAID1E_F_ABORT;
409 	} else {
410 		G_RAID_DEBUG1(0, vol->v_softc,
411 		    "Subdisk %s:%d-%s rebuild aborted.",
412 		    sd->sd_volume->v_name, sd->sd_pos,
413 		    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
414 		trs->trso_flags &= ~TR_RAID1E_F_ABORT;
415 		if (trs->trso_flags & TR_RAID1E_F_LOCKED) {
416 			trs->trso_flags &= ~TR_RAID1E_F_LOCKED;
417 			g_raid_unlock_range(tr->tro_volume,
418 			    trs->trso_lock_pos, trs->trso_lock_len);
419 		}
420 		g_raid_tr_raid1e_rebuild_done(trs);
421 	}
422 }
423 
424 static void
425 g_raid_tr_raid1e_rebuild_some(struct g_raid_tr_object *tr)
426 {
427 	struct g_raid_tr_raid1e_object *trs;
428 	struct g_raid_softc *sc;
429 	struct g_raid_volume *vol;
430 	struct g_raid_subdisk *sd;
431 	struct bio *bp;
432 	off_t len, virtual, vend, offset, start;
433 	int disk, copy, best;
434 
435 	trs = (struct g_raid_tr_raid1e_object *)tr;
436 	if (trs->trso_flags & TR_RAID1E_F_DOING_SOME)
437 		return;
438 	vol = tr->tro_volume;
439 	sc = vol->v_softc;
440 	sd = trs->trso_failed_sd;
441 
442 	while (1) {
443 		if (sd->sd_rebuild_pos >= sd->sd_size) {
444 			g_raid_tr_raid1e_rebuild_finish(tr);
445 			return;
446 		}
447 		/* Get virtual offset from physical rebuild position. */
448 		P2V(vol, sd->sd_pos, sd->sd_rebuild_pos, &virtual, &copy);
449 		/* Get physical offset back to get first stripe position. */
450 		V2P(vol, virtual, &disk, &offset, &start);
451 		/* Calculate contignous data length. */
452 		len = MIN(g_raid1e_rebuild_slab,
453 		    sd->sd_size - sd->sd_rebuild_pos);
454 		if ((vol->v_disks_count % N) != 0)
455 			len = MIN(len, vol->v_strip_size - start);
456 		/* Find disk with most accurate data. */
457 		best = g_raid_tr_raid1e_select_read_disk(vol, disk,
458 		    offset + start, len, 0);
459 		if (best < 0) {
460 			/* There is no any valid disk. */
461 			g_raid_tr_raid1e_rebuild_abort(tr);
462 			return;
463 		} else if (best != copy) {
464 			/* Some other disk has better data. */
465 			break;
466 		}
467 		/* We have the most accurate data. Skip the range. */
468 		G_RAID_DEBUG1(3, sc, "Skipping rebuild for range %ju - %ju",
469 		    sd->sd_rebuild_pos, sd->sd_rebuild_pos + len);
470 		sd->sd_rebuild_pos += len;
471 	}
472 
473 	bp = &trs->trso_bio;
474 	memset(bp, 0, sizeof(*bp));
475 	bp->bio_offset = offset + start +
476 	    ((disk + best >= vol->v_disks_count) ? vol->v_strip_size : 0);
477 	bp->bio_length = len;
478 	bp->bio_data = trs->trso_buffer;
479 	bp->bio_cmd = BIO_READ;
480 	bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
481 	bp->bio_caller1 = &vol->v_subdisks[(disk + best) % vol->v_disks_count];
482 	G_RAID_LOGREQ(3, bp, "Queueing rebuild read");
483 	/*
484 	 * If we are crossing stripe boundary, correct affected virtual
485 	 * range we should lock.
486 	 */
487 	if (start + len > vol->v_strip_size) {
488 		P2V(vol, sd->sd_pos, sd->sd_rebuild_pos + len, &vend, &copy);
489 		len = vend - virtual;
490 	}
491 	trs->trso_flags |= TR_RAID1E_F_DOING_SOME;
492 	trs->trso_flags |= TR_RAID1E_F_LOCKED;
493 	trs->trso_lock_pos = virtual;
494 	trs->trso_lock_len = len;
495 	/* Lock callback starts I/O */
496 	g_raid_lock_range(sd->sd_volume, virtual, len, NULL, bp);
497 }
498 
499 static void
500 g_raid_tr_raid1e_rebuild_start(struct g_raid_tr_object *tr)
501 {
502 	struct g_raid_volume *vol;
503 	struct g_raid_tr_raid1e_object *trs;
504 	struct g_raid_subdisk *sd;
505 
506 	vol = tr->tro_volume;
507 	trs = (struct g_raid_tr_raid1e_object *)tr;
508 	if (trs->trso_failed_sd) {
509 		G_RAID_DEBUG1(1, vol->v_softc,
510 		    "Already rebuild in start rebuild. pos %jd\n",
511 		    (intmax_t)trs->trso_failed_sd->sd_rebuild_pos);
512 		return;
513 	}
514 	sd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_RESYNC);
515 	if (sd == NULL)
516 		sd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_REBUILD);
517 	if (sd == NULL) {
518 		sd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_STALE);
519 		if (sd != NULL) {
520 			sd->sd_rebuild_pos = 0;
521 			g_raid_change_subdisk_state(sd,
522 			    G_RAID_SUBDISK_S_RESYNC);
523 			g_raid_write_metadata(vol->v_softc, vol, sd, NULL);
524 		} else {
525 			sd = g_raid_get_subdisk(vol,
526 			    G_RAID_SUBDISK_S_UNINITIALIZED);
527 			if (sd == NULL)
528 				sd = g_raid_get_subdisk(vol,
529 				    G_RAID_SUBDISK_S_NEW);
530 			if (sd != NULL) {
531 				sd->sd_rebuild_pos = 0;
532 				g_raid_change_subdisk_state(sd,
533 				    G_RAID_SUBDISK_S_REBUILD);
534 				g_raid_write_metadata(vol->v_softc,
535 				    vol, sd, NULL);
536 			}
537 		}
538 	}
539 	if (sd == NULL) {
540 		G_RAID_DEBUG1(1, vol->v_softc,
541 		    "No failed disk to rebuild.  night night.");
542 		return;
543 	}
544 	trs->trso_failed_sd = sd;
545 	G_RAID_DEBUG1(0, vol->v_softc,
546 	    "Subdisk %s:%d-%s rebuild start at %jd.",
547 	    sd->sd_volume->v_name, sd->sd_pos,
548 	    sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]",
549 	    trs->trso_failed_sd->sd_rebuild_pos);
550 	trs->trso_type = TR_RAID1E_REBUILD;
551 	trs->trso_buffer = malloc(g_raid1e_rebuild_slab, M_TR_RAID1E, M_WAITOK);
552 	trs->trso_meta_update = g_raid1e_rebuild_meta_update;
553 	g_raid_tr_raid1e_rebuild_some(tr);
554 }
555 
556 static void
557 g_raid_tr_raid1e_maybe_rebuild(struct g_raid_tr_object *tr,
558     struct g_raid_subdisk *sd)
559 {
560 	struct g_raid_volume *vol;
561 	struct g_raid_tr_raid1e_object *trs;
562 	int nr;
563 
564 	vol = tr->tro_volume;
565 	trs = (struct g_raid_tr_raid1e_object *)tr;
566 	if (trs->trso_stopping)
567 		return;
568 	nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_REBUILD) +
569 	    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
570 	switch(trs->trso_type) {
571 	case TR_RAID1E_NONE:
572 		if (vol->v_state < G_RAID_VOLUME_S_DEGRADED)
573 			return;
574 		if (nr == 0) {
575 			nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_NEW) +
576 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
577 			    g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_UNINITIALIZED);
578 			if (nr == 0)
579 				return;
580 		}
581 		g_raid_tr_raid1e_rebuild_start(tr);
582 		break;
583 	case TR_RAID1E_REBUILD:
584 		if (vol->v_state < G_RAID_VOLUME_S_DEGRADED || nr == 0 ||
585 		    trs->trso_failed_sd == sd)
586 			g_raid_tr_raid1e_rebuild_abort(tr);
587 		break;
588 	case TR_RAID1E_RESYNC:
589 		break;
590 	}
591 }
592 
593 static int
594 g_raid_tr_event_raid1e(struct g_raid_tr_object *tr,
595     struct g_raid_subdisk *sd, u_int event)
596 {
597 
598 	g_raid_tr_update_state_raid1e(tr->tro_volume, sd);
599 	return (0);
600 }
601 
602 static int
603 g_raid_tr_start_raid1e(struct g_raid_tr_object *tr)
604 {
605 	struct g_raid_tr_raid1e_object *trs;
606 	struct g_raid_volume *vol;
607 
608 	trs = (struct g_raid_tr_raid1e_object *)tr;
609 	vol = tr->tro_volume;
610 	trs->trso_starting = 0;
611 	g_raid_tr_update_state_raid1e(vol, NULL);
612 	return (0);
613 }
614 
615 static int
616 g_raid_tr_stop_raid1e(struct g_raid_tr_object *tr)
617 {
618 	struct g_raid_tr_raid1e_object *trs;
619 	struct g_raid_volume *vol;
620 
621 	trs = (struct g_raid_tr_raid1e_object *)tr;
622 	vol = tr->tro_volume;
623 	trs->trso_starting = 0;
624 	trs->trso_stopping = 1;
625 	g_raid_tr_update_state_raid1e(vol, NULL);
626 	return (0);
627 }
628 
629 /*
630  * Select the disk to read from.  Take into account: subdisk state, running
631  * error recovery, average disk load, head position and possible cache hits.
632  */
633 #define ABS(x)		(((x) >= 0) ? (x) : (-(x)))
634 static int
635 g_raid_tr_raid1e_select_read_disk(struct g_raid_volume *vol,
636     int no, off_t off, off_t len, u_int mask)
637 {
638 	struct g_raid_subdisk *sd;
639 	off_t offset;
640 	int i, best, prio, bestprio;
641 
642 	best = -1;
643 	bestprio = INT_MAX;
644 	for (i = 0; i < N; i++) {
645 		sd = &vol->v_subdisks[(no + i) % vol->v_disks_count];
646 		offset = off;
647 		if (no + i >= vol->v_disks_count)
648 			offset += vol->v_strip_size;
649 
650 		prio = G_RAID_SUBDISK_LOAD(sd);
651 		if ((mask & (1 << sd->sd_pos)) != 0)
652 			continue;
653 		switch (sd->sd_state) {
654 		case G_RAID_SUBDISK_S_ACTIVE:
655 			break;
656 		case G_RAID_SUBDISK_S_RESYNC:
657 			if (offset + off < sd->sd_rebuild_pos)
658 				break;
659 			/* FALLTHROUGH */
660 		case G_RAID_SUBDISK_S_STALE:
661 			prio += i << 24;
662 			break;
663 		case G_RAID_SUBDISK_S_REBUILD:
664 			if (offset + off < sd->sd_rebuild_pos)
665 				break;
666 			/* FALLTHROUGH */
667 		default:
668 			continue;
669 		}
670 		prio += min(sd->sd_recovery, 255) << 16;
671 		/* If disk head is precisely in position - highly prefer it. */
672 		if (G_RAID_SUBDISK_POS(sd) == offset)
673 			prio -= 2 * G_RAID_SUBDISK_LOAD_SCALE;
674 		else
675 		/* If disk head is close to position - prefer it. */
676 		if (ABS(G_RAID_SUBDISK_POS(sd) - offset) <
677 		    G_RAID_SUBDISK_TRACK_SIZE)
678 			prio -= 1 * G_RAID_SUBDISK_LOAD_SCALE;
679 		if (prio < bestprio) {
680 			bestprio = prio;
681 			best = i;
682 		}
683 	}
684 	return (best);
685 }
686 
687 static void
688 g_raid_tr_iostart_raid1e_read(struct g_raid_tr_object *tr, struct bio *bp)
689 {
690 	struct g_raid_volume *vol;
691 	struct g_raid_subdisk *sd;
692 	struct bio_queue_head queue;
693 	struct bio *cbp;
694 	char *addr;
695 	off_t offset, start, length, remain;
696 	u_int no, strip_size;
697 	int best;
698 
699 	vol = tr->tro_volume;
700 	if ((bp->bio_flags & BIO_UNMAPPED) != 0)
701 		addr = NULL;
702 	else
703 		addr = bp->bio_data;
704 	strip_size = vol->v_strip_size;
705 	V2P(vol, bp->bio_offset, &no, &offset, &start);
706 	remain = bp->bio_length;
707 	bioq_init(&queue);
708 	while (remain > 0) {
709 		length = MIN(strip_size - start, remain);
710 		best = g_raid_tr_raid1e_select_read_disk(vol,
711 		    no, offset, length, 0);
712 		KASSERT(best >= 0, ("No readable disk in volume %s!",
713 		    vol->v_name));
714 		no += best;
715 		if (no >= vol->v_disks_count) {
716 			no -= vol->v_disks_count;
717 			offset += strip_size;
718 		}
719 		cbp = g_clone_bio(bp);
720 		if (cbp == NULL)
721 			goto failure;
722 		cbp->bio_offset = offset + start;
723 		cbp->bio_length = length;
724 		if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
725 			cbp->bio_ma_offset += (uintptr_t)addr;
726 			cbp->bio_ma += cbp->bio_ma_offset / PAGE_SIZE;
727 			cbp->bio_ma_offset %= PAGE_SIZE;
728 			cbp->bio_ma_n = round_page(cbp->bio_ma_offset +
729 			    cbp->bio_length) / PAGE_SIZE;
730 		} else
731 			cbp->bio_data = addr;
732 		cbp->bio_caller1 = &vol->v_subdisks[no];
733 		bioq_insert_tail(&queue, cbp);
734 		no += N - best;
735 		if (no >= vol->v_disks_count) {
736 			no -= vol->v_disks_count;
737 			offset += strip_size;
738 		}
739 		remain -= length;
740 		addr += length;
741 		start = 0;
742 	}
743 	while ((cbp = bioq_takefirst(&queue)) != NULL) {
744 		sd = cbp->bio_caller1;
745 		cbp->bio_caller1 = NULL;
746 		g_raid_subdisk_iostart(sd, cbp);
747 	}
748 	return;
749 failure:
750 	while ((cbp = bioq_takefirst(&queue)) != NULL)
751 		g_destroy_bio(cbp);
752 	if (bp->bio_error == 0)
753 		bp->bio_error = ENOMEM;
754 	g_raid_iodone(bp, bp->bio_error);
755 }
756 
757 static void
758 g_raid_tr_iostart_raid1e_write(struct g_raid_tr_object *tr, struct bio *bp)
759 {
760 	struct g_raid_volume *vol;
761 	struct g_raid_subdisk *sd;
762 	struct bio_queue_head queue;
763 	struct bio *cbp;
764 	char *addr;
765 	off_t offset, start, length, remain;
766 	u_int no, strip_size;
767 	int i;
768 
769 	vol = tr->tro_volume;
770 	if ((bp->bio_flags & BIO_UNMAPPED) != 0)
771 		addr = NULL;
772 	else
773 		addr = bp->bio_data;
774 	strip_size = vol->v_strip_size;
775 	V2P(vol, bp->bio_offset, &no, &offset, &start);
776 	remain = bp->bio_length;
777 	bioq_init(&queue);
778 	while (remain > 0) {
779 		length = MIN(strip_size - start, remain);
780 		for (i = 0; i < N; i++) {
781 			sd = &vol->v_subdisks[no];
782 			switch (sd->sd_state) {
783 			case G_RAID_SUBDISK_S_ACTIVE:
784 			case G_RAID_SUBDISK_S_STALE:
785 			case G_RAID_SUBDISK_S_RESYNC:
786 				break;
787 			case G_RAID_SUBDISK_S_REBUILD:
788 				if (offset + start >= sd->sd_rebuild_pos)
789 					goto nextdisk;
790 				break;
791 			default:
792 				goto nextdisk;
793 			}
794 			cbp = g_clone_bio(bp);
795 			if (cbp == NULL)
796 				goto failure;
797 			cbp->bio_offset = offset + start;
798 			cbp->bio_length = length;
799 			if ((bp->bio_flags & BIO_UNMAPPED) != 0 &&
800 			    bp->bio_cmd != BIO_DELETE) {
801 				cbp->bio_ma_offset += (uintptr_t)addr;
802 				cbp->bio_ma += cbp->bio_ma_offset / PAGE_SIZE;
803 				cbp->bio_ma_offset %= PAGE_SIZE;
804 				cbp->bio_ma_n = round_page(cbp->bio_ma_offset +
805 				    cbp->bio_length) / PAGE_SIZE;
806 			} else
807 				cbp->bio_data = addr;
808 			cbp->bio_caller1 = sd;
809 			bioq_insert_tail(&queue, cbp);
810 nextdisk:
811 			if (++no >= vol->v_disks_count) {
812 				no = 0;
813 				offset += strip_size;
814 			}
815 		}
816 		remain -= length;
817 		if (bp->bio_cmd != BIO_DELETE)
818 			addr += length;
819 		start = 0;
820 	}
821 	while ((cbp = bioq_takefirst(&queue)) != NULL) {
822 		sd = cbp->bio_caller1;
823 		cbp->bio_caller1 = NULL;
824 		g_raid_subdisk_iostart(sd, cbp);
825 	}
826 	return;
827 failure:
828 	while ((cbp = bioq_takefirst(&queue)) != NULL)
829 		g_destroy_bio(cbp);
830 	if (bp->bio_error == 0)
831 		bp->bio_error = ENOMEM;
832 	g_raid_iodone(bp, bp->bio_error);
833 }
834 
835 static void
836 g_raid_tr_iostart_raid1e(struct g_raid_tr_object *tr, struct bio *bp)
837 {
838 	struct g_raid_volume *vol;
839 	struct g_raid_tr_raid1e_object *trs;
840 
841 	vol = tr->tro_volume;
842 	trs = (struct g_raid_tr_raid1e_object *)tr;
843 	if (vol->v_state != G_RAID_VOLUME_S_OPTIMAL &&
844 	    vol->v_state != G_RAID_VOLUME_S_SUBOPTIMAL &&
845 	    vol->v_state != G_RAID_VOLUME_S_DEGRADED) {
846 		g_raid_iodone(bp, EIO);
847 		return;
848 	}
849 	/*
850 	 * If we're rebuilding, squeeze in rebuild activity every so often,
851 	 * even when the disk is busy.  Be sure to only count real I/O
852 	 * to the disk.  All 'SPECIAL' I/O is traffic generated to the disk
853 	 * by this module.
854 	 */
855 	if (trs->trso_failed_sd != NULL &&
856 	    !(bp->bio_cflags & G_RAID_BIO_FLAG_SPECIAL)) {
857 		/* Make this new or running now round short. */
858 		trs->trso_recover_slabs = 0;
859 		if (--trs->trso_fair_io <= 0) {
860 			trs->trso_fair_io = g_raid1e_rebuild_fair_io;
861 			g_raid_tr_raid1e_rebuild_some(tr);
862 		}
863 	}
864 	switch (bp->bio_cmd) {
865 	case BIO_READ:
866 		g_raid_tr_iostart_raid1e_read(tr, bp);
867 		break;
868 	case BIO_WRITE:
869 	case BIO_DELETE:
870 		g_raid_tr_iostart_raid1e_write(tr, bp);
871 		break;
872 	case BIO_SPEEDUP:
873 	case BIO_FLUSH:
874 		g_raid_tr_flush_common(tr, bp);
875 		break;
876 	default:
877 		KASSERT(1 == 0, ("Invalid command here: %u (volume=%s)",
878 		    bp->bio_cmd, vol->v_name));
879 		break;
880 	}
881 }
882 
883 static void
884 g_raid_tr_iodone_raid1e(struct g_raid_tr_object *tr,
885     struct g_raid_subdisk *sd, struct bio *bp)
886 {
887 	struct bio *cbp;
888 	struct g_raid_subdisk *nsd;
889 	struct g_raid_volume *vol;
890 	struct bio *pbp;
891 	struct g_raid_tr_raid1e_object *trs;
892 	off_t virtual, offset, start;
893 	uintptr_t mask;
894 	int error, do_write, copy, disk, best;
895 
896 	trs = (struct g_raid_tr_raid1e_object *)tr;
897 	vol = tr->tro_volume;
898 	if (bp->bio_cflags & G_RAID_BIO_FLAG_SYNC) {
899 		if (trs->trso_type == TR_RAID1E_REBUILD) {
900 			nsd = trs->trso_failed_sd;
901 			if (bp->bio_cmd == BIO_READ) {
902 				/* Immediately abort rebuild, if requested. */
903 				if (trs->trso_flags & TR_RAID1E_F_ABORT) {
904 					trs->trso_flags &= ~TR_RAID1E_F_DOING_SOME;
905 					g_raid_tr_raid1e_rebuild_abort(tr);
906 					return;
907 				}
908 
909 				/* On read error, skip and cross fingers. */
910 				if (bp->bio_error != 0) {
911 					G_RAID_LOGREQ(0, bp,
912 					    "Read error during rebuild (%d), "
913 					    "possible data loss!",
914 					    bp->bio_error);
915 					goto rebuild_round_done;
916 				}
917 
918 				/*
919 				 * The read operation finished, queue the
920 				 * write and get out.
921 				 */
922 				G_RAID_LOGREQ(3, bp, "Rebuild read done: %d",
923 				    bp->bio_error);
924 				bp->bio_cmd = BIO_WRITE;
925 				bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
926 				bp->bio_offset = nsd->sd_rebuild_pos;
927 				G_RAID_LOGREQ(3, bp, "Queueing rebuild write.");
928 				g_raid_subdisk_iostart(nsd, bp);
929 			} else {
930 				/*
931 				 * The write operation just finished.  Do
932 				 * another.  We keep cloning the master bio
933 				 * since it has the right buffers allocated to
934 				 * it.
935 				 */
936 				G_RAID_LOGREQ(3, bp, "Rebuild write done: %d",
937 				    bp->bio_error);
938 				if (bp->bio_error != 0 ||
939 				    trs->trso_flags & TR_RAID1E_F_ABORT) {
940 					if ((trs->trso_flags &
941 					    TR_RAID1E_F_ABORT) == 0) {
942 						g_raid_tr_raid1e_fail_disk(sd->sd_softc,
943 						    nsd, nsd->sd_disk);
944 					}
945 					trs->trso_flags &= ~TR_RAID1E_F_DOING_SOME;
946 					g_raid_tr_raid1e_rebuild_abort(tr);
947 					return;
948 				}
949 rebuild_round_done:
950 				trs->trso_flags &= ~TR_RAID1E_F_LOCKED;
951 				g_raid_unlock_range(tr->tro_volume,
952 				    trs->trso_lock_pos, trs->trso_lock_len);
953 				nsd->sd_rebuild_pos += bp->bio_length;
954 				if (nsd->sd_rebuild_pos >= nsd->sd_size) {
955 					g_raid_tr_raid1e_rebuild_finish(tr);
956 					return;
957 				}
958 
959 				/* Abort rebuild if we are stopping */
960 				if (trs->trso_stopping) {
961 					trs->trso_flags &= ~TR_RAID1E_F_DOING_SOME;
962 					g_raid_tr_raid1e_rebuild_abort(tr);
963 					return;
964 				}
965 
966 				if (--trs->trso_meta_update <= 0) {
967 					g_raid_write_metadata(vol->v_softc,
968 					    vol, nsd, nsd->sd_disk);
969 					trs->trso_meta_update =
970 					    g_raid1e_rebuild_meta_update;
971 					/* Compensate short rebuild I/Os. */
972 					if ((vol->v_disks_count % N) != 0 &&
973 					    vol->v_strip_size <
974 					     g_raid1e_rebuild_slab) {
975 						trs->trso_meta_update *=
976 						    g_raid1e_rebuild_slab;
977 						trs->trso_meta_update /=
978 						    vol->v_strip_size;
979 					}
980 				}
981 				trs->trso_flags &= ~TR_RAID1E_F_DOING_SOME;
982 				if (--trs->trso_recover_slabs <= 0)
983 					return;
984 				/* Run next rebuild iteration. */
985 				g_raid_tr_raid1e_rebuild_some(tr);
986 			}
987 		} else if (trs->trso_type == TR_RAID1E_RESYNC) {
988 			/*
989 			 * read good sd, read bad sd in parallel.  when both
990 			 * done, compare the buffers.  write good to the bad
991 			 * if different.  do the next bit of work.
992 			 */
993 			panic("Somehow, we think we're doing a resync");
994 		}
995 		return;
996 	}
997 	pbp = bp->bio_parent;
998 	pbp->bio_inbed++;
999 	mask = (intptr_t)bp->bio_caller2;
1000 	if (bp->bio_cmd == BIO_READ && bp->bio_error != 0) {
1001 		/*
1002 		 * Read failed on first drive.  Retry the read error on
1003 		 * another disk drive, if available, before erroring out the
1004 		 * read.
1005 		 */
1006 		sd->sd_disk->d_read_errs++;
1007 		G_RAID_LOGREQ(0, bp,
1008 		    "Read error (%d), %d read errors total",
1009 		    bp->bio_error, sd->sd_disk->d_read_errs);
1010 
1011 		/*
1012 		 * If there are too many read errors, we move to degraded.
1013 		 * XXX Do we want to FAIL the drive (eg, make the user redo
1014 		 * everything to get it back in sync), or just degrade the
1015 		 * drive, which kicks off a resync?
1016 		 */
1017 		do_write = 0;
1018 		if (sd->sd_disk->d_read_errs > g_raid_read_err_thresh)
1019 			g_raid_tr_raid1e_fail_disk(sd->sd_softc, sd, sd->sd_disk);
1020 		else if (mask == 0)
1021 			do_write = 1;
1022 
1023 		/* Restore what we were doing. */
1024 		P2V(vol, sd->sd_pos, bp->bio_offset, &virtual, &copy);
1025 		V2P(vol, virtual, &disk, &offset, &start);
1026 
1027 		/* Find the other disk, and try to do the I/O to it. */
1028 		mask |= 1 << copy;
1029 		best = g_raid_tr_raid1e_select_read_disk(vol,
1030 		    disk, offset, start, mask);
1031 		if (best >= 0 && (cbp = g_clone_bio(pbp)) != NULL) {
1032 			disk += best;
1033 			if (disk >= vol->v_disks_count) {
1034 				disk -= vol->v_disks_count;
1035 				offset += vol->v_strip_size;
1036 			}
1037 			cbp->bio_offset = offset + start;
1038 			cbp->bio_length = bp->bio_length;
1039 			cbp->bio_data = bp->bio_data;
1040 			cbp->bio_ma = bp->bio_ma;
1041 			cbp->bio_ma_offset = bp->bio_ma_offset;
1042 			cbp->bio_ma_n = bp->bio_ma_n;
1043 			g_destroy_bio(bp);
1044 			nsd = &vol->v_subdisks[disk];
1045 			G_RAID_LOGREQ(2, cbp, "Retrying read from %d",
1046 			    nsd->sd_pos);
1047 			if (do_write)
1048 				mask |= 1 << 31;
1049 			if ((mask & (1U << 31)) != 0)
1050 				sd->sd_recovery++;
1051 			cbp->bio_caller2 = (void *)mask;
1052 			if (do_write) {
1053 				cbp->bio_caller1 = nsd;
1054 				/* Lock callback starts I/O */
1055 				g_raid_lock_range(sd->sd_volume,
1056 				    virtual, cbp->bio_length, pbp, cbp);
1057 			} else {
1058 				g_raid_subdisk_iostart(nsd, cbp);
1059 			}
1060 			return;
1061 		}
1062 		/*
1063 		 * We can't retry.  Return the original error by falling
1064 		 * through.  This will happen when there's only one good disk.
1065 		 * We don't need to fail the raid, since its actual state is
1066 		 * based on the state of the subdisks.
1067 		 */
1068 		G_RAID_LOGREQ(2, bp, "Couldn't retry read, failing it");
1069 	}
1070 	if (bp->bio_cmd == BIO_READ &&
1071 	    bp->bio_error == 0 &&
1072 	    (mask & (1U << 31)) != 0) {
1073 		G_RAID_LOGREQ(3, bp, "Recovered data from other drive");
1074 
1075 		/* Restore what we were doing. */
1076 		P2V(vol, sd->sd_pos, bp->bio_offset, &virtual, &copy);
1077 		V2P(vol, virtual, &disk, &offset, &start);
1078 
1079 		/* Find best disk to write. */
1080 		best = g_raid_tr_raid1e_select_read_disk(vol,
1081 		    disk, offset, start, ~mask);
1082 		if (best >= 0 && (cbp = g_clone_bio(pbp)) != NULL) {
1083 			disk += best;
1084 			if (disk >= vol->v_disks_count) {
1085 				disk -= vol->v_disks_count;
1086 				offset += vol->v_strip_size;
1087 			}
1088 			cbp->bio_offset = offset + start;
1089 			cbp->bio_cmd = BIO_WRITE;
1090 			cbp->bio_cflags = G_RAID_BIO_FLAG_REMAP;
1091 			cbp->bio_caller2 = (void *)mask;
1092 			g_destroy_bio(bp);
1093 			G_RAID_LOGREQ(2, cbp,
1094 			    "Attempting bad sector remap on failing drive.");
1095 			g_raid_subdisk_iostart(&vol->v_subdisks[disk], cbp);
1096 			return;
1097 		}
1098 	}
1099 	if ((mask & (1U << 31)) != 0) {
1100 		/*
1101 		 * We're done with a recovery, mark the range as unlocked.
1102 		 * For any write errors, we aggressively fail the disk since
1103 		 * there was both a READ and a WRITE error at this location.
1104 		 * Both types of errors generally indicates the drive is on
1105 		 * the verge of total failure anyway.  Better to stop trusting
1106 		 * it now.  However, we need to reset error to 0 in that case
1107 		 * because we're not failing the original I/O which succeeded.
1108 		 */
1109 
1110 		/* Restore what we were doing. */
1111 		P2V(vol, sd->sd_pos, bp->bio_offset, &virtual, &copy);
1112 		V2P(vol, virtual, &disk, &offset, &start);
1113 
1114 		for (copy = 0; copy < N; copy++) {
1115 			if ((mask & (1 << copy) ) != 0)
1116 				vol->v_subdisks[(disk + copy) %
1117 				    vol->v_disks_count].sd_recovery--;
1118 		}
1119 
1120 		if (bp->bio_cmd == BIO_WRITE && bp->bio_error) {
1121 			G_RAID_LOGREQ(0, bp, "Remap write failed: "
1122 			    "failing subdisk.");
1123 			g_raid_tr_raid1e_fail_disk(sd->sd_softc, sd, sd->sd_disk);
1124 			bp->bio_error = 0;
1125 		}
1126 		G_RAID_LOGREQ(2, bp, "REMAP done %d.", bp->bio_error);
1127 		g_raid_unlock_range(sd->sd_volume, virtual, bp->bio_length);
1128 	}
1129 	if (pbp->bio_cmd != BIO_READ) {
1130 		if (pbp->bio_inbed == 1 || pbp->bio_error != 0)
1131 			pbp->bio_error = bp->bio_error;
1132 		if (pbp->bio_cmd == BIO_WRITE && bp->bio_error != 0) {
1133 			G_RAID_LOGREQ(0, bp, "Write failed: failing subdisk.");
1134 			g_raid_tr_raid1e_fail_disk(sd->sd_softc, sd, sd->sd_disk);
1135 		}
1136 		error = pbp->bio_error;
1137 	} else
1138 		error = bp->bio_error;
1139 	g_destroy_bio(bp);
1140 	if (pbp->bio_children == pbp->bio_inbed) {
1141 		pbp->bio_completed = pbp->bio_length;
1142 		g_raid_iodone(pbp, error);
1143 	}
1144 }
1145 
1146 static int
1147 g_raid_tr_kerneldump_raid1e(struct g_raid_tr_object *tr, void *virtual,
1148     off_t boffset, size_t blength)
1149 {
1150 	struct g_raid_volume *vol;
1151 	struct g_raid_subdisk *sd;
1152 	struct bio_queue_head queue;
1153 	char *addr;
1154 	off_t offset, start, length, remain;
1155 	u_int no, strip_size;
1156 	int i, error;
1157 
1158 	vol = tr->tro_volume;
1159 	addr = virtual;
1160 	strip_size = vol->v_strip_size;
1161 	V2P(vol, boffset, &no, &offset, &start);
1162 	remain = blength;
1163 	bioq_init(&queue);
1164 	while (remain > 0) {
1165 		length = MIN(strip_size - start, remain);
1166 		for (i = 0; i < N; i++) {
1167 			sd = &vol->v_subdisks[no];
1168 			switch (sd->sd_state) {
1169 			case G_RAID_SUBDISK_S_ACTIVE:
1170 			case G_RAID_SUBDISK_S_STALE:
1171 			case G_RAID_SUBDISK_S_RESYNC:
1172 				break;
1173 			case G_RAID_SUBDISK_S_REBUILD:
1174 				if (offset + start >= sd->sd_rebuild_pos)
1175 					goto nextdisk;
1176 				break;
1177 			default:
1178 				goto nextdisk;
1179 			}
1180 			error = g_raid_subdisk_kerneldump(sd, addr,
1181 			    offset + start, length);
1182 			if (error != 0)
1183 				return (error);
1184 nextdisk:
1185 			if (++no >= vol->v_disks_count) {
1186 				no = 0;
1187 				offset += strip_size;
1188 			}
1189 		}
1190 		remain -= length;
1191 		addr += length;
1192 		start = 0;
1193 	}
1194 	return (0);
1195 }
1196 
1197 static int
1198 g_raid_tr_locked_raid1e(struct g_raid_tr_object *tr, void *argp)
1199 {
1200 	struct bio *bp;
1201 	struct g_raid_subdisk *sd;
1202 
1203 	bp = (struct bio *)argp;
1204 	sd = (struct g_raid_subdisk *)bp->bio_caller1;
1205 	g_raid_subdisk_iostart(sd, bp);
1206 
1207 	return (0);
1208 }
1209 
1210 static int
1211 g_raid_tr_idle_raid1e(struct g_raid_tr_object *tr)
1212 {
1213 	struct g_raid_tr_raid1e_object *trs;
1214 	struct g_raid_volume *vol;
1215 
1216 	vol = tr->tro_volume;
1217 	trs = (struct g_raid_tr_raid1e_object *)tr;
1218 	trs->trso_fair_io = g_raid1e_rebuild_fair_io;
1219 	trs->trso_recover_slabs = g_raid1e_rebuild_cluster_idle;
1220 	/* Compensate short rebuild I/Os. */
1221 	if ((vol->v_disks_count % N) != 0 &&
1222 	    vol->v_strip_size < g_raid1e_rebuild_slab) {
1223 		trs->trso_recover_slabs *= g_raid1e_rebuild_slab;
1224 		trs->trso_recover_slabs /= vol->v_strip_size;
1225 	}
1226 	if (trs->trso_type == TR_RAID1E_REBUILD)
1227 		g_raid_tr_raid1e_rebuild_some(tr);
1228 	return (0);
1229 }
1230 
1231 static int
1232 g_raid_tr_free_raid1e(struct g_raid_tr_object *tr)
1233 {
1234 	struct g_raid_tr_raid1e_object *trs;
1235 
1236 	trs = (struct g_raid_tr_raid1e_object *)tr;
1237 
1238 	if (trs->trso_buffer != NULL) {
1239 		free(trs->trso_buffer, M_TR_RAID1E);
1240 		trs->trso_buffer = NULL;
1241 	}
1242 	return (0);
1243 }
1244 
1245 G_RAID_TR_DECLARE(raid1e, "RAID1E");
1246