sys/geom/raid/tr_raid1.c

   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_raid1);
  46
  47 #define RAID1_REBUILD_SLAB      (1 << 20) /* One transation in a rebuild */
  48 static int g_raid1_rebuild_slab = RAID1_REBUILD_SLAB;
  49 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_slab_size, CTLFLAG_RWTUN,
  50     &g_raid1_rebuild_slab, 0,
  51     "Amount of the disk to rebuild each read/write cycle of the rebuild.");
  52
  53 #define RAID1_REBUILD_FAIR_IO 20 /* use 1/x of the available I/O */
  54 static int g_raid1_rebuild_fair_io = RAID1_REBUILD_FAIR_IO;
  55 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_fair_io, CTLFLAG_RWTUN,
  56     &g_raid1_rebuild_fair_io, 0,
  57     "Fraction of the I/O bandwidth to use when disk busy for rebuild.");
  58
  59 #define RAID1_REBUILD_CLUSTER_IDLE 100
  60 static int g_raid1_rebuild_cluster_idle = RAID1_REBUILD_CLUSTER_IDLE;
  61 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_cluster_idle, CTLFLAG_RWTUN,
  62     &g_raid1_rebuild_cluster_idle, 0,
  63     "Number of slabs to do each time we trigger a rebuild cycle");
  64
  65 #define RAID1_REBUILD_META_UPDATE 1024 /* update meta data every 1GB or so */
  66 static int g_raid1_rebuild_meta_update = RAID1_REBUILD_META_UPDATE;
  67 SYSCTL_UINT(_kern_geom_raid_raid1, OID_AUTO, rebuild_meta_update, CTLFLAG_RWTUN,
  68     &g_raid1_rebuild_meta_update, 0,
  69     "When to update the meta data.");
  70
  71 static MALLOC_DEFINE(M_TR_RAID1, "tr_raid1_data", "GEOM_RAID RAID1 data");
  72
  73 #define TR_RAID1_NONE 0
  74 #define TR_RAID1_REBUILD 1
  75 #define TR_RAID1_RESYNC 2
  76
  77 #define TR_RAID1_F_DOING_SOME   0x1
  78 #define TR_RAID1_F_LOCKED       0x2
  79 #define TR_RAID1_F_ABORT        0x4
  80
  81 struct g_raid_tr_raid1_object {
  82         struct g_raid_tr_object  trso_base;
  83         int                      trso_starting;
  84         int                      trso_stopping;
  85         int                      trso_type;
  86         int                      trso_recover_slabs; /* slabs before rest */
  87         int                      trso_fair_io;
  88         int                      trso_meta_update;
  89         int                      trso_flags;
  90         struct g_raid_subdisk   *trso_failed_sd; /* like per volume */
  91         void                    *trso_buffer;    /* Buffer space */
  92         struct bio               trso_bio;
  93 };
  94
  95 static g_raid_tr_taste_t g_raid_tr_taste_raid1;
  96 static g_raid_tr_event_t g_raid_tr_event_raid1;
  97 static g_raid_tr_start_t g_raid_tr_start_raid1;
  98 static g_raid_tr_stop_t g_raid_tr_stop_raid1;
  99 static g_raid_tr_iostart_t g_raid_tr_iostart_raid1;
 100 static g_raid_tr_iodone_t g_raid_tr_iodone_raid1;
 101 static g_raid_tr_kerneldump_t g_raid_tr_kerneldump_raid1;
 102 static g_raid_tr_locked_t g_raid_tr_locked_raid1;
 103 static g_raid_tr_idle_t g_raid_tr_idle_raid1;
 104 static g_raid_tr_free_t g_raid_tr_free_raid1;
 105
 106 static kobj_method_t g_raid_tr_raid1_methods[] = {
 107         KOBJMETHOD(g_raid_tr_taste,     g_raid_tr_taste_raid1),
 108         KOBJMETHOD(g_raid_tr_event,     g_raid_tr_event_raid1),
 109         KOBJMETHOD(g_raid_tr_start,     g_raid_tr_start_raid1),
 110         KOBJMETHOD(g_raid_tr_stop,      g_raid_tr_stop_raid1),
 111         KOBJMETHOD(g_raid_tr_iostart,   g_raid_tr_iostart_raid1),
 112         KOBJMETHOD(g_raid_tr_iodone,    g_raid_tr_iodone_raid1),
 113         KOBJMETHOD(g_raid_tr_kerneldump, g_raid_tr_kerneldump_raid1),
 114         KOBJMETHOD(g_raid_tr_locked,    g_raid_tr_locked_raid1),
 115         KOBJMETHOD(g_raid_tr_idle,      g_raid_tr_idle_raid1),
 116         KOBJMETHOD(g_raid_tr_free,      g_raid_tr_free_raid1),
 117         { 0, 0 }
 118 };
 119
 120 static struct g_raid_tr_class g_raid_tr_raid1_class = {
 121         "RAID1",
 122         g_raid_tr_raid1_methods,
 123         sizeof(struct g_raid_tr_raid1_object),
 124         .trc_enable = 1,
 125         .trc_priority = 100,
 126         .trc_accept_unmapped = 1
 127 };
 128
 129 static void g_raid_tr_raid1_rebuild_abort(struct g_raid_tr_object *tr);
 130 static void g_raid_tr_raid1_maybe_rebuild(struct g_raid_tr_object *tr,
 131     struct g_raid_subdisk *sd);
 132
 133 static int
 134 g_raid_tr_taste_raid1(struct g_raid_tr_object *tr, struct g_raid_volume *vol)
 135 {
 136         struct g_raid_tr_raid1_object *trs;
 137
 138         trs = (struct g_raid_tr_raid1_object *)tr;
 139         if (tr->tro_volume->v_raid_level != G_RAID_VOLUME_RL_RAID1 ||
 140             (tr->tro_volume->v_raid_level_qualifier != G_RAID_VOLUME_RLQ_R1SM &&
 141              tr->tro_volume->v_raid_level_qualifier != G_RAID_VOLUME_RLQ_R1MM))
 142                 return (G_RAID_TR_TASTE_FAIL);
 143         trs->trso_starting = 1;
 144         return (G_RAID_TR_TASTE_SUCCEED);
 145 }
 146
 147 static int
 148 g_raid_tr_update_state_raid1(struct g_raid_volume *vol,
 149     struct g_raid_subdisk *sd)
 150 {
 151         struct g_raid_tr_raid1_object *trs;
 152         struct g_raid_softc *sc;
 153         struct g_raid_subdisk *tsd, *bestsd;
 154         u_int s;
 155         int i, na, ns;
 156
 157         sc = vol->v_softc;
 158         trs = (struct g_raid_tr_raid1_object *)vol->v_tr;
 159         if (trs->trso_stopping &&
 160             (trs->trso_flags & TR_RAID1_F_DOING_SOME) == 0)
 161                 s = G_RAID_VOLUME_S_STOPPED;
 162         else if (trs->trso_starting)
 163                 s = G_RAID_VOLUME_S_STARTING;
 164         else {
 165                 /* Make sure we have at least one ACTIVE disk. */
 166                 na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
 167                 if (na == 0) {
 168                         /*
 169                          * Critical situation! We have no any active disk!
 170                          * Choose the best disk we have to make it active.
 171                          */
 172                         bestsd = &vol->v_subdisks[0];
 173                         for (i = 1; i < vol->v_disks_count; i++) {
 174                                 tsd = &vol->v_subdisks[i];
 175                                 if (tsd->sd_state > bestsd->sd_state)
 176                                         bestsd = tsd;
 177                                 else if (tsd->sd_state == bestsd->sd_state &&
 178                                     (tsd->sd_state == G_RAID_SUBDISK_S_REBUILD ||
 179                                      tsd->sd_state == G_RAID_SUBDISK_S_RESYNC) &&
 180                                     tsd->sd_rebuild_pos > bestsd->sd_rebuild_pos)
 181                                         bestsd = tsd;
 182                         }
 183                         if (bestsd->sd_state >= G_RAID_SUBDISK_S_UNINITIALIZED) {
 184                                 /* We found reasonable candidate. */
 185                                 G_RAID_DEBUG1(1, sc,
 186                                     "Promote subdisk %s:%d from %s to ACTIVE.",
 187                                     vol->v_name, bestsd->sd_pos,
 188                                     g_raid_subdisk_state2str(bestsd->sd_state));
 189                                 g_raid_change_subdisk_state(bestsd,
 190                                     G_RAID_SUBDISK_S_ACTIVE);
 191                                 g_raid_write_metadata(sc,
 192                                     vol, bestsd, bestsd->sd_disk);
 193                         }
 194                 }
 195                 na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
 196                 ns = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
 197                      g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
 198                 if (na == vol->v_disks_count)
 199                         s = G_RAID_VOLUME_S_OPTIMAL;
 200                 else if (na + ns == vol->v_disks_count)
 201                         s = G_RAID_VOLUME_S_SUBOPTIMAL;
 202                 else if (na > 0)
 203                         s = G_RAID_VOLUME_S_DEGRADED;
 204                 else
 205                         s = G_RAID_VOLUME_S_BROKEN;
 206                 g_raid_tr_raid1_maybe_rebuild(vol->v_tr, sd);
 207         }
 208         if (s != vol->v_state) {
 209                 g_raid_event_send(vol, G_RAID_VOLUME_S_ALIVE(s) ?
 210                     G_RAID_VOLUME_E_UP : G_RAID_VOLUME_E_DOWN,
 211                     G_RAID_EVENT_VOLUME);
 212                 g_raid_change_volume_state(vol, s);
 213                 if (!trs->trso_starting && !trs->trso_stopping)
 214                         g_raid_write_metadata(sc, vol, NULL, NULL);
 215         }
 216         return (0);
 217 }
 218
 219 static void
 220 g_raid_tr_raid1_fail_disk(struct g_raid_softc *sc, struct g_raid_subdisk *sd,
 221     struct g_raid_disk *disk)
 222 {
 223         /*
 224          * We don't fail the last disk in the pack, since it still has decent
 225          * data on it and that's better than failing the disk if it is the root
 226          * file system.
 227          *
 228          * XXX should this be controlled via a tunable?  It makes sense for
 229          * the volume that has / on it.  I can't think of a case where we'd
 230          * want the volume to go away on this kind of event.
 231          */
 232         if (g_raid_nsubdisks(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == 1 &&
 233             g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE) == sd)
 234                 return;
 235         g_raid_fail_disk(sc, sd, disk);
 236 }
 237
 238 static void
 239 g_raid_tr_raid1_rebuild_some(struct g_raid_tr_object *tr)
 240 {
 241         struct g_raid_tr_raid1_object *trs;
 242         struct g_raid_subdisk *sd, *good_sd;
 243         struct bio *bp;
 244
 245         trs = (struct g_raid_tr_raid1_object *)tr;
 246         if (trs->trso_flags & TR_RAID1_F_DOING_SOME)
 247                 return;
 248         sd = trs->trso_failed_sd;
 249         good_sd = g_raid_get_subdisk(sd->sd_volume, G_RAID_SUBDISK_S_ACTIVE);
 250         if (good_sd == NULL) {
 251                 g_raid_tr_raid1_rebuild_abort(tr);
 252                 return;
 253         }
 254         bp = &trs->trso_bio;
 255         memset(bp, 0, sizeof(*bp));
 256         bp->bio_offset = sd->sd_rebuild_pos;
 257         bp->bio_length = MIN(g_raid1_rebuild_slab,
 258             sd->sd_size - sd->sd_rebuild_pos);
 259         bp->bio_data = trs->trso_buffer;
 260         bp->bio_cmd = BIO_READ;
 261         bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
 262         bp->bio_caller1 = good_sd;
 263         trs->trso_flags |= TR_RAID1_F_DOING_SOME;
 264         trs->trso_flags |= TR_RAID1_F_LOCKED;
 265         g_raid_lock_range(sd->sd_volume,        /* Lock callback starts I/O */
 266            bp->bio_offset, bp->bio_length, NULL, bp);
 267 }
 268
 269 static void
 270 g_raid_tr_raid1_rebuild_done(struct g_raid_tr_raid1_object *trs)
 271 {
 272         struct g_raid_volume *vol;
 273         struct g_raid_subdisk *sd;
 274
 275         vol = trs->trso_base.tro_volume;
 276         sd = trs->trso_failed_sd;
 277         g_raid_write_metadata(vol->v_softc, vol, sd, sd->sd_disk);
 278         free(trs->trso_buffer, M_TR_RAID1);
 279         trs->trso_buffer = NULL;
 280         trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
 281         trs->trso_type = TR_RAID1_NONE;
 282         trs->trso_recover_slabs = 0;
 283         trs->trso_failed_sd = NULL;
 284         g_raid_tr_update_state_raid1(vol, NULL);
 285 }
 286
 287 static void
 288 g_raid_tr_raid1_rebuild_finish(struct g_raid_tr_object *tr)
 289 {
 290         struct g_raid_tr_raid1_object *trs;
 291         struct g_raid_subdisk *sd;
 292
 293         trs = (struct g_raid_tr_raid1_object *)tr;
 294         sd = trs->trso_failed_sd;
 295         G_RAID_DEBUG1(0, tr->tro_volume->v_softc,
 296             "Subdisk %s:%d-%s rebuild completed.",
 297             sd->sd_volume->v_name, sd->sd_pos,
 298             sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
 299         g_raid_change_subdisk_state(sd, G_RAID_SUBDISK_S_ACTIVE);
 300         sd->sd_rebuild_pos = 0;
 301         g_raid_tr_raid1_rebuild_done(trs);
 302 }
 303
 304 static void
 305 g_raid_tr_raid1_rebuild_abort(struct g_raid_tr_object *tr)
 306 {
 307         struct g_raid_tr_raid1_object *trs;
 308         struct g_raid_subdisk *sd;
 309         struct g_raid_volume *vol;
 310         off_t len;
 311
 312         vol = tr->tro_volume;
 313         trs = (struct g_raid_tr_raid1_object *)tr;
 314         sd = trs->trso_failed_sd;
 315         if (trs->trso_flags & TR_RAID1_F_DOING_SOME) {
 316                 G_RAID_DEBUG1(1, vol->v_softc,
 317                     "Subdisk %s:%d-%s rebuild is aborting.",
 318                     sd->sd_volume->v_name, sd->sd_pos,
 319                     sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
 320                 trs->trso_flags |= TR_RAID1_F_ABORT;
 321         } else {
 322                 G_RAID_DEBUG1(0, vol->v_softc,
 323                     "Subdisk %s:%d-%s rebuild aborted.",
 324                     sd->sd_volume->v_name, sd->sd_pos,
 325                     sd->sd_disk ? g_raid_get_diskname(sd->sd_disk) : "[none]");
 326                 trs->trso_flags &= ~TR_RAID1_F_ABORT;
 327                 if (trs->trso_flags & TR_RAID1_F_LOCKED) {
 328                         trs->trso_flags &= ~TR_RAID1_F_LOCKED;
 329                         len = MIN(g_raid1_rebuild_slab,
 330                             sd->sd_size - sd->sd_rebuild_pos);
 331                         g_raid_unlock_range(tr->tro_volume,
 332                             sd->sd_rebuild_pos, len);
 333                 }
 334                 g_raid_tr_raid1_rebuild_done(trs);
 335         }
 336 }
 337
 338 static void
 339 g_raid_tr_raid1_rebuild_start(struct g_raid_tr_object *tr)
 340 {
 341         struct g_raid_volume *vol;
 342         struct g_raid_tr_raid1_object *trs;
 343         struct g_raid_subdisk *sd, *fsd;
 344
 345         vol = tr->tro_volume;
 346         trs = (struct g_raid_tr_raid1_object *)tr;
 347         if (trs->trso_failed_sd) {
 348                 G_RAID_DEBUG1(1, vol->v_softc,
 349                     "Already rebuild in start rebuild. pos %jd\n",
 350                     (intmax_t)trs->trso_failed_sd->sd_rebuild_pos);
 351                 return;
 352         }
 353         sd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_ACTIVE);
 354         if (sd == NULL) {
 355                 G_RAID_DEBUG1(1, vol->v_softc,
 356                     "No active disk to rebuild.  night night.");
 357                 return;
 358         }
 359         fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_RESYNC);
 360         if (fsd == NULL)
 361                 fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_REBUILD);
 362         if (fsd == NULL) {
 363                 fsd = g_raid_get_subdisk(vol, G_RAID_SUBDISK_S_STALE);
 364                 if (fsd != NULL) {
 365                         fsd->sd_rebuild_pos = 0;
 366                         g_raid_change_subdisk_state(fsd,
 367                             G_RAID_SUBDISK_S_RESYNC);
 368                         g_raid_write_metadata(vol->v_softc, vol, fsd, NULL);
 369                 } else {
 370                         fsd = g_raid_get_subdisk(vol,
 371                             G_RAID_SUBDISK_S_UNINITIALIZED);
 372                         if (fsd == NULL)
 373                                 fsd = g_raid_get_subdisk(vol,
 374                                     G_RAID_SUBDISK_S_NEW);
 375                         if (fsd != NULL) {
 376                                 fsd->sd_rebuild_pos = 0;
 377                                 g_raid_change_subdisk_state(fsd,
 378                                     G_RAID_SUBDISK_S_REBUILD);
 379                                 g_raid_write_metadata(vol->v_softc,
 380                                     vol, fsd, NULL);
 381                         }
 382                 }
 383         }
 384         if (fsd == NULL) {
 385                 G_RAID_DEBUG1(1, vol->v_softc,
 386                     "No failed disk to rebuild.  night night.");
 387                 return;
 388         }
 389         trs->trso_failed_sd = fsd;
 390         G_RAID_DEBUG1(0, vol->v_softc,
 391             "Subdisk %s:%d-%s rebuild start at %jd.",
 392             fsd->sd_volume->v_name, fsd->sd_pos,
 393             fsd->sd_disk ? g_raid_get_diskname(fsd->sd_disk) : "[none]",
 394             trs->trso_failed_sd->sd_rebuild_pos);
 395         trs->trso_type = TR_RAID1_REBUILD;
 396         trs->trso_buffer = malloc(g_raid1_rebuild_slab, M_TR_RAID1, M_WAITOK);
 397         trs->trso_meta_update = g_raid1_rebuild_meta_update;
 398         g_raid_tr_raid1_rebuild_some(tr);
 399 }
 400
 401
 402 static void
 403 g_raid_tr_raid1_maybe_rebuild(struct g_raid_tr_object *tr,
 404     struct g_raid_subdisk *sd)
 405 {
 406         struct g_raid_volume *vol;
 407         struct g_raid_tr_raid1_object *trs;
 408         int na, nr;
 409
 410         /*
 411          * If we're stopping, don't do anything.  If we don't have at least one
 412          * good disk and one bad disk, we don't do anything.  And if there's a
 413          * 'good disk' stored in the trs, then we're in progress and we punt.
 414          * If we make it past all these checks, we need to rebuild.
 415          */
 416         vol = tr->tro_volume;
 417         trs = (struct g_raid_tr_raid1_object *)tr;
 418         if (trs->trso_stopping)
 419                 return;
 420         na = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_ACTIVE);
 421         nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_REBUILD) +
 422             g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_RESYNC);
 423         switch(trs->trso_type) {
 424         case TR_RAID1_NONE:
 425                 if (na == 0)
 426                         return;
 427                 if (nr == 0) {
 428                         nr = g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_NEW) +
 429                             g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_STALE) +
 430                             g_raid_nsubdisks(vol, G_RAID_SUBDISK_S_UNINITIALIZED);
 431                         if (nr == 0)
 432                                 return;
 433                 }
 434                 g_raid_tr_raid1_rebuild_start(tr);
 435                 break;
 436         case TR_RAID1_REBUILD:
 437                 if (na == 0 || nr == 0 || trs->trso_failed_sd == sd)
 438                         g_raid_tr_raid1_rebuild_abort(tr);
 439                 break;
 440         case TR_RAID1_RESYNC:
 441                 break;
 442         }
 443 }
 444
 445 static int
 446 g_raid_tr_event_raid1(struct g_raid_tr_object *tr,
 447     struct g_raid_subdisk *sd, u_int event)
 448 {
 449
 450         g_raid_tr_update_state_raid1(tr->tro_volume, sd);
 451         return (0);
 452 }
 453
 454 static int
 455 g_raid_tr_start_raid1(struct g_raid_tr_object *tr)
 456 {
 457         struct g_raid_tr_raid1_object *trs;
 458         struct g_raid_volume *vol;
 459
 460         trs = (struct g_raid_tr_raid1_object *)tr;
 461         vol = tr->tro_volume;
 462         trs->trso_starting = 0;
 463         g_raid_tr_update_state_raid1(vol, NULL);
 464         return (0);
 465 }
 466
 467 static int
 468 g_raid_tr_stop_raid1(struct g_raid_tr_object *tr)
 469 {
 470         struct g_raid_tr_raid1_object *trs;
 471         struct g_raid_volume *vol;
 472
 473         trs = (struct g_raid_tr_raid1_object *)tr;
 474         vol = tr->tro_volume;
 475         trs->trso_starting = 0;
 476         trs->trso_stopping = 1;
 477         g_raid_tr_update_state_raid1(vol, NULL);
 478         return (0);
 479 }
 480
 481 /*
 482  * Select the disk to read from.  Take into account: subdisk state, running
 483  * error recovery, average disk load, head position and possible cache hits.
 484  */
 485 #define ABS(x)          (((x) >= 0) ? (x) : (-(x)))
 486 static struct g_raid_subdisk *
 487 g_raid_tr_raid1_select_read_disk(struct g_raid_volume *vol, struct bio *bp,
 488     u_int mask)
 489 {
 490         struct g_raid_subdisk *sd, *best;
 491         int i, prio, bestprio;
 492
 493         best = NULL;
 494         bestprio = INT_MAX;
 495         for (i = 0; i < vol->v_disks_count; i++) {
 496                 sd = &vol->v_subdisks[i];
 497                 if (sd->sd_state != G_RAID_SUBDISK_S_ACTIVE &&
 498                     ((sd->sd_state != G_RAID_SUBDISK_S_REBUILD &&
 499                       sd->sd_state != G_RAID_SUBDISK_S_RESYNC) ||
 500                      bp->bio_offset + bp->bio_length > sd->sd_rebuild_pos))
 501                         continue;
 502                 if ((mask & (1 << i)) != 0)
 503                         continue;
 504                 prio = G_RAID_SUBDISK_LOAD(sd);
 505                 prio += min(sd->sd_recovery, 255) << 22;
 506                 prio += (G_RAID_SUBDISK_S_ACTIVE - sd->sd_state) << 16;
 507                 /* If disk head is precisely in position - highly prefer it. */
 508                 if (G_RAID_SUBDISK_POS(sd) == bp->bio_offset)
 509                         prio -= 2 * G_RAID_SUBDISK_LOAD_SCALE;
 510                 else
 511                 /* If disk head is close to position - prefer it. */
 512                 if (ABS(G_RAID_SUBDISK_POS(sd) - bp->bio_offset) <
 513                     G_RAID_SUBDISK_TRACK_SIZE)
 514                         prio -= 1 * G_RAID_SUBDISK_LOAD_SCALE;
 515                 if (prio < bestprio) {
 516                         best = sd;
 517                         bestprio = prio;
 518                 }
 519         }
 520         return (best);
 521 }
 522
 523 static void
 524 g_raid_tr_iostart_raid1_read(struct g_raid_tr_object *tr, struct bio *bp)
 525 {
 526         struct g_raid_subdisk *sd;
 527         struct bio *cbp;
 528
 529         sd = g_raid_tr_raid1_select_read_disk(tr->tro_volume, bp, 0);
 530         KASSERT(sd != NULL, ("No active disks in volume %s.",
 531                 tr->tro_volume->v_name));
 532
 533         cbp = g_clone_bio(bp);
 534         if (cbp == NULL) {
 535                 g_raid_iodone(bp, ENOMEM);
 536                 return;
 537         }
 538
 539         g_raid_subdisk_iostart(sd, cbp);
 540 }
 541
 542 static void
 543 g_raid_tr_iostart_raid1_write(struct g_raid_tr_object *tr, struct bio *bp)
 544 {
 545         struct g_raid_volume *vol;
 546         struct g_raid_subdisk *sd;
 547         struct bio_queue_head queue;
 548         struct bio *cbp;
 549         int i;
 550
 551         vol = tr->tro_volume;
 552
 553         /*
 554          * Allocate all bios before sending any request, so we can return
 555          * ENOMEM in nice and clean way.
 556          */
 557         bioq_init(&queue);
 558         for (i = 0; i < vol->v_disks_count; i++) {
 559                 sd = &vol->v_subdisks[i];
 560                 switch (sd->sd_state) {
 561                 case G_RAID_SUBDISK_S_ACTIVE:
 562                         break;
 563                 case G_RAID_SUBDISK_S_REBUILD:
 564                         /*
 565                          * When rebuilding, only part of this subdisk is
 566                          * writable, the rest will be written as part of the
 567                          * that process.
 568                          */
 569                         if (bp->bio_offset >= sd->sd_rebuild_pos)
 570                                 continue;
 571                         break;
 572                 case G_RAID_SUBDISK_S_STALE:
 573                 case G_RAID_SUBDISK_S_RESYNC:
 574                         /*
 575                          * Resyncing still writes on the theory that the
 576                          * resync'd disk is very close and writing it will
 577                          * keep it that way better if we keep up while
 578                          * resyncing.
 579                          */
 580                         break;
 581                 default:
 582                         continue;
 583                 }
 584                 cbp = g_clone_bio(bp);
 585                 if (cbp == NULL)
 586                         goto failure;
 587                 cbp->bio_caller1 = sd;
 588                 bioq_insert_tail(&queue, cbp);
 589         }
 590         while ((cbp = bioq_takefirst(&queue)) != NULL) {
 591                 sd = cbp->bio_caller1;
 592                 cbp->bio_caller1 = NULL;
 593                 g_raid_subdisk_iostart(sd, cbp);
 594         }
 595         return;
 596 failure:
 597         while ((cbp = bioq_takefirst(&queue)) != NULL)
 598                 g_destroy_bio(cbp);
 599         if (bp->bio_error == 0)
 600                 bp->bio_error = ENOMEM;
 601         g_raid_iodone(bp, bp->bio_error);
 602 }
 603
 604 static void
 605 g_raid_tr_iostart_raid1(struct g_raid_tr_object *tr, struct bio *bp)
 606 {
 607         struct g_raid_volume *vol;
 608         struct g_raid_tr_raid1_object *trs;
 609
 610         vol = tr->tro_volume;
 611         trs = (struct g_raid_tr_raid1_object *)tr;
 612         if (vol->v_state != G_RAID_VOLUME_S_OPTIMAL &&
 613             vol->v_state != G_RAID_VOLUME_S_SUBOPTIMAL &&
 614             vol->v_state != G_RAID_VOLUME_S_DEGRADED) {
 615                 g_raid_iodone(bp, EIO);
 616                 return;
 617         }
 618         /*
 619          * If we're rebuilding, squeeze in rebuild activity every so often,
 620          * even when the disk is busy.  Be sure to only count real I/O
 621          * to the disk.  All 'SPECIAL' I/O is traffic generated to the disk
 622          * by this module.
 623          */
 624         if (trs->trso_failed_sd != NULL &&
 625             !(bp->bio_cflags & G_RAID_BIO_FLAG_SPECIAL)) {
 626                 /* Make this new or running now round short. */
 627                 trs->trso_recover_slabs = 0;
 628                 if (--trs->trso_fair_io <= 0) {
 629                         trs->trso_fair_io = g_raid1_rebuild_fair_io;
 630                         g_raid_tr_raid1_rebuild_some(tr);
 631                 }
 632         }
 633         switch (bp->bio_cmd) {
 634         case BIO_READ:
 635                 g_raid_tr_iostart_raid1_read(tr, bp);
 636                 break;
 637         case BIO_WRITE:
 638         case BIO_DELETE:
 639                 g_raid_tr_iostart_raid1_write(tr, bp);
 640                 break;
 641         case BIO_FLUSH:
 642                 g_raid_tr_flush_common(tr, bp);
 643                 break;
 644         default:
 645                 KASSERT(1 == 0, ("Invalid command here: %u (volume=%s)",
 646                     bp->bio_cmd, vol->v_name));
 647                 break;
 648         }
 649 }
 650
 651 static void
 652 g_raid_tr_iodone_raid1(struct g_raid_tr_object *tr,
 653     struct g_raid_subdisk *sd, struct bio *bp)
 654 {
 655         struct bio *cbp;
 656         struct g_raid_subdisk *nsd;
 657         struct g_raid_volume *vol;
 658         struct bio *pbp;
 659         struct g_raid_tr_raid1_object *trs;
 660         uintptr_t *mask;
 661         int error, do_write;
 662
 663         trs = (struct g_raid_tr_raid1_object *)tr;
 664         vol = tr->tro_volume;
 665         if (bp->bio_cflags & G_RAID_BIO_FLAG_SYNC) {
 666                 /*
 667                  * This operation is part of a rebuild or resync operation.
 668                  * See what work just got done, then schedule the next bit of
 669                  * work, if any.  Rebuild/resync is done a little bit at a
 670                  * time.  Either when a timeout happens, or after we get a
 671                  * bunch of I/Os to the disk (to make sure an active system
 672                  * will complete in a sane amount of time).
 673                  *
 674                  * We are setup to do differing amounts of work for each of
 675                  * these cases.  so long as the slabs is smallish (less than
 676                  * 50 or so, I'd guess, but that's just a WAG), we shouldn't
 677                  * have any bio starvation issues.  For active disks, we do
 678                  * 5MB of data, for inactive ones, we do 50MB.
 679                  */
 680                 if (trs->trso_type == TR_RAID1_REBUILD) {
 681                         if (bp->bio_cmd == BIO_READ) {
 682
 683                                 /* Immediately abort rebuild, if requested. */
 684                                 if (trs->trso_flags & TR_RAID1_F_ABORT) {
 685                                         trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
 686                                         g_raid_tr_raid1_rebuild_abort(tr);
 687                                         return;
 688                                 }
 689
 690                                 /* On read error, skip and cross fingers. */
 691                                 if (bp->bio_error != 0) {
 692                                         G_RAID_LOGREQ(0, bp,
 693                                             "Read error during rebuild (%d), "
 694                                             "possible data loss!",
 695                                             bp->bio_error);
 696                                         goto rebuild_round_done;
 697                                 }
 698
 699                                 /*
 700                                  * The read operation finished, queue the
 701                                  * write and get out.
 702                                  */
 703                                 G_RAID_LOGREQ(4, bp, "rebuild read done. %d",
 704                                     bp->bio_error);
 705                                 bp->bio_cmd = BIO_WRITE;
 706                                 bp->bio_cflags = G_RAID_BIO_FLAG_SYNC;
 707                                 G_RAID_LOGREQ(4, bp, "Queueing rebuild write.");
 708                                 g_raid_subdisk_iostart(trs->trso_failed_sd, bp);
 709                         } else {
 710                                 /*
 711                                  * The write operation just finished.  Do
 712                                  * another.  We keep cloning the master bio
 713                                  * since it has the right buffers allocated to
 714                                  * it.
 715                                  */
 716                                 G_RAID_LOGREQ(4, bp,
 717                                     "rebuild write done. Error %d",
 718                                     bp->bio_error);
 719                                 nsd = trs->trso_failed_sd;
 720                                 if (bp->bio_error != 0 ||
 721                                     trs->trso_flags & TR_RAID1_F_ABORT) {
 722                                         if ((trs->trso_flags &
 723                                             TR_RAID1_F_ABORT) == 0) {
 724                                                 g_raid_tr_raid1_fail_disk(sd->sd_softc,
 725                                                     nsd, nsd->sd_disk);
 726                                         }
 727                                         trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
 728                                         g_raid_tr_raid1_rebuild_abort(tr);
 729                                         return;
 730                                 }
 731 rebuild_round_done:
 732                                 nsd = trs->trso_failed_sd;
 733                                 trs->trso_flags &= ~TR_RAID1_F_LOCKED;
 734                                 g_raid_unlock_range(sd->sd_volume,
 735                                     bp->bio_offset, bp->bio_length);
 736                                 nsd->sd_rebuild_pos += bp->bio_length;
 737                                 if (nsd->sd_rebuild_pos >= nsd->sd_size) {
 738                                         g_raid_tr_raid1_rebuild_finish(tr);
 739                                         return;
 740                                 }
 741
 742                                 /* Abort rebuild if we are stopping */
 743                                 if (trs->trso_stopping) {
 744                                         trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
 745                                         g_raid_tr_raid1_rebuild_abort(tr);
 746                                         return;
 747                                 }
 748
 749                                 if (--trs->trso_meta_update <= 0) {
 750                                         g_raid_write_metadata(vol->v_softc,
 751                                             vol, nsd, nsd->sd_disk);
 752                                         trs->trso_meta_update =
 753                                             g_raid1_rebuild_meta_update;
 754                                 }
 755                                 trs->trso_flags &= ~TR_RAID1_F_DOING_SOME;
 756                                 if (--trs->trso_recover_slabs <= 0)
 757                                         return;
 758                                 g_raid_tr_raid1_rebuild_some(tr);
 759                         }
 760                 } else if (trs->trso_type == TR_RAID1_RESYNC) {
 761                         /*
 762                          * read good sd, read bad sd in parallel.  when both
 763                          * done, compare the buffers.  write good to the bad
 764                          * if different.  do the next bit of work.
 765                          */
 766                         panic("Somehow, we think we're doing a resync");
 767                 }
 768                 return;
 769         }
 770         pbp = bp->bio_parent;
 771         pbp->bio_inbed++;
 772         if (bp->bio_cmd == BIO_READ && bp->bio_error != 0) {
 773                 /*
 774                  * Read failed on first drive.  Retry the read error on
 775                  * another disk drive, if available, before erroring out the
 776                  * read.
 777                  */
 778                 sd->sd_disk->d_read_errs++;
 779                 G_RAID_LOGREQ(0, bp,
 780                     "Read error (%d), %d read errors total",
 781                     bp->bio_error, sd->sd_disk->d_read_errs);
 782
 783                 /*
 784                  * If there are too many read errors, we move to degraded.
 785                  * XXX Do we want to FAIL the drive (eg, make the user redo
 786                  * everything to get it back in sync), or just degrade the
 787                  * drive, which kicks off a resync?
 788                  */
 789                 do_write = 1;
 790                 if (sd->sd_disk->d_read_errs > g_raid_read_err_thresh) {
 791                         g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
 792                         if (pbp->bio_children == 1)
 793                                 do_write = 0;
 794                 }
 795
 796                 /*
 797                  * Find the other disk, and try to do the I/O to it.
 798                  */
 799                 mask = (uintptr_t *)(&pbp->bio_driver2);
 800                 if (pbp->bio_children == 1) {
 801                         /* Save original subdisk. */
 802                         pbp->bio_driver1 = do_write ? sd : NULL;
 803                         *mask = 0;
 804                 }
 805                 *mask |= 1 << sd->sd_pos;
 806                 nsd = g_raid_tr_raid1_select_read_disk(vol, pbp, *mask);
 807                 if (nsd != NULL && (cbp = g_clone_bio(pbp)) != NULL) {
 808                         g_destroy_bio(bp);
 809                         G_RAID_LOGREQ(2, cbp, "Retrying read from %d",
 810                             nsd->sd_pos);
 811                         if (pbp->bio_children == 2 && do_write) {
 812                                 sd->sd_recovery++;
 813                                 cbp->bio_caller1 = nsd;
 814                                 pbp->bio_pflags = G_RAID_BIO_FLAG_LOCKED;
 815                                 /* Lock callback starts I/O */
 816                                 g_raid_lock_range(sd->sd_volume,
 817                                     cbp->bio_offset, cbp->bio_length, pbp, cbp);
 818                         } else {
 819                                 g_raid_subdisk_iostart(nsd, cbp);
 820                         }
 821                         return;
 822                 }
 823                 /*
 824                  * We can't retry.  Return the original error by falling
 825                  * through.  This will happen when there's only one good disk.
 826                  * We don't need to fail the raid, since its actual state is
 827                  * based on the state of the subdisks.
 828                  */
 829                 G_RAID_LOGREQ(2, bp, "Couldn't retry read, failing it");
 830         }
 831         if (bp->bio_cmd == BIO_READ &&
 832             bp->bio_error == 0 &&
 833             pbp->bio_children > 1 &&
 834             pbp->bio_driver1 != NULL) {
 835                 /*
 836                  * If it was a read, and bio_children is >1, then we just
 837                  * recovered the data from the second drive.  We should try to
 838                  * write that data to the first drive if sector remapping is
 839                  * enabled.  A write should put the data in a new place on the
 840                  * disk, remapping the bad sector.  Do we need to do that by
 841                  * queueing a request to the main worker thread?  It doesn't
 842                  * affect the return code of this current read, and can be
 843                  * done at our liesure.  However, to make the code simpler, it
 844                  * is done syncrhonously.
 845                  */
 846                 G_RAID_LOGREQ(3, bp, "Recovered data from other drive");
 847                 cbp = g_clone_bio(pbp);
 848                 if (cbp != NULL) {
 849                         g_destroy_bio(bp);
 850                         cbp->bio_cmd = BIO_WRITE;
 851                         cbp->bio_cflags = G_RAID_BIO_FLAG_REMAP;
 852                         G_RAID_LOGREQ(2, cbp,
 853                             "Attempting bad sector remap on failing drive.");
 854                         g_raid_subdisk_iostart(pbp->bio_driver1, cbp);
 855                         return;
 856                 }
 857         }
 858         if (pbp->bio_pflags & G_RAID_BIO_FLAG_LOCKED) {
 859                 /*
 860                  * We're done with a recovery, mark the range as unlocked.
 861                  * For any write errors, we agressively fail the disk since
 862                  * there was both a READ and a WRITE error at this location.
 863                  * Both types of errors generally indicates the drive is on
 864                  * the verge of total failure anyway.  Better to stop trusting
 865                  * it now.  However, we need to reset error to 0 in that case
 866                  * because we're not failing the original I/O which succeeded.
 867                  */
 868                 if (bp->bio_cmd == BIO_WRITE && bp->bio_error) {
 869                         G_RAID_LOGREQ(0, bp, "Remap write failed: "
 870                             "failing subdisk.");
 871                         g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
 872                         bp->bio_error = 0;
 873                 }
 874                 if (pbp->bio_driver1 != NULL) {
 875                         ((struct g_raid_subdisk *)pbp->bio_driver1)
 876                             ->sd_recovery--;
 877                 }
 878                 G_RAID_LOGREQ(2, bp, "REMAP done %d.", bp->bio_error);
 879                 g_raid_unlock_range(sd->sd_volume, bp->bio_offset,
 880                     bp->bio_length);
 881         }
 882         if (pbp->bio_cmd != BIO_READ) {
 883                 if (pbp->bio_inbed == 1 || pbp->bio_error != 0)
 884                         pbp->bio_error = bp->bio_error;
 885                 if (pbp->bio_cmd == BIO_WRITE && bp->bio_error != 0) {
 886                         G_RAID_LOGREQ(0, bp, "Write failed: failing subdisk.");
 887                         g_raid_tr_raid1_fail_disk(sd->sd_softc, sd, sd->sd_disk);
 888                 }
 889                 error = pbp->bio_error;
 890         } else
 891                 error = bp->bio_error;
 892         g_destroy_bio(bp);
 893         if (pbp->bio_children == pbp->bio_inbed) {
 894                 pbp->bio_completed = pbp->bio_length;
 895                 g_raid_iodone(pbp, error);
 896         }
 897 }
 898
 899 static int
 900 g_raid_tr_kerneldump_raid1(struct g_raid_tr_object *tr,
 901     void *virtual, vm_offset_t physical, off_t offset, size_t length)
 902 {
 903         struct g_raid_volume *vol;
 904         struct g_raid_subdisk *sd;
 905         int error, i, ok;
 906
 907         vol = tr->tro_volume;
 908         error = 0;
 909         ok = 0;
 910         for (i = 0; i < vol->v_disks_count; i++) {
 911                 sd = &vol->v_subdisks[i];
 912                 switch (sd->sd_state) {
 913                 case G_RAID_SUBDISK_S_ACTIVE:
 914                         break;
 915                 case G_RAID_SUBDISK_S_REBUILD:
 916                         /*
 917                          * When rebuilding, only part of this subdisk is
 918                          * writable, the rest will be written as part of the
 919                          * that process.
 920                          */
 921                         if (offset >= sd->sd_rebuild_pos)
 922                                 continue;
 923                         break;
 924                 case G_RAID_SUBDISK_S_STALE:
 925                 case G_RAID_SUBDISK_S_RESYNC:
 926                         /*
 927                          * Resyncing still writes on the theory that the
 928                          * resync'd disk is very close and writing it will
 929                          * keep it that way better if we keep up while
 930                          * resyncing.
 931                          */
 932                         break;
 933                 default:
 934                         continue;
 935                 }
 936                 error = g_raid_subdisk_kerneldump(sd,
 937                     virtual, physical, offset, length);
 938                 if (error == 0)
 939                         ok++;
 940         }
 941         return (ok > 0 ? 0 : error);
 942 }
 943
 944 static int
 945 g_raid_tr_locked_raid1(struct g_raid_tr_object *tr, void *argp)
 946 {
 947         struct bio *bp;
 948         struct g_raid_subdisk *sd;
 949
 950         bp = (struct bio *)argp;
 951         sd = (struct g_raid_subdisk *)bp->bio_caller1;
 952         g_raid_subdisk_iostart(sd, bp);
 953
 954         return (0);
 955 }
 956
 957 static int
 958 g_raid_tr_idle_raid1(struct g_raid_tr_object *tr)
 959 {
 960         struct g_raid_tr_raid1_object *trs;
 961
 962         trs = (struct g_raid_tr_raid1_object *)tr;
 963         trs->trso_fair_io = g_raid1_rebuild_fair_io;
 964         trs->trso_recover_slabs = g_raid1_rebuild_cluster_idle;
 965         if (trs->trso_type == TR_RAID1_REBUILD)
 966                 g_raid_tr_raid1_rebuild_some(tr);
 967         return (0);
 968 }
 969
 970 static int
 971 g_raid_tr_free_raid1(struct g_raid_tr_object *tr)
 972 {
 973         struct g_raid_tr_raid1_object *trs;
 974
 975         trs = (struct g_raid_tr_raid1_object *)tr;
 976
 977         if (trs->trso_buffer != NULL) {
 978                 free(trs->trso_buffer, M_TR_RAID1);
 979                 trs->trso_buffer = NULL;
 980         }
 981         return (0);
 982 }
 983
 984 G_RAID_TR_DECLARE(raid1, "RAID1");