sys/geom/raid/tr_raid1.c

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