sys/geom/mirror/g_mirror.c

   1 /*-
   2  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@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/systm.h>
  32 #include <sys/fail.h>
  33 #include <sys/kernel.h>
  34 #include <sys/module.h>
  35 #include <sys/limits.h>
  36 #include <sys/lock.h>
  37 #include <sys/mutex.h>
  38 #include <sys/bio.h>
  39 #include <sys/sbuf.h>
  40 #include <sys/sysctl.h>
  41 #include <sys/malloc.h>
  42 #include <sys/eventhandler.h>
  43 #include <vm/uma.h>
  44 #include <geom/geom.h>
  45 #include <sys/proc.h>
  46 #include <sys/kthread.h>
  47 #include <sys/sched.h>
  48 #include <geom/mirror/g_mirror.h>
  49
  50 FEATURE(geom_mirror, "GEOM mirroring support");
  51
  52 static MALLOC_DEFINE(M_MIRROR, "mirror_data", "GEOM_MIRROR Data");
  53
  54 SYSCTL_DECL(_kern_geom);
  55 static SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW, 0,
  56     "GEOM_MIRROR stuff");
  57 u_int g_mirror_debug = 0;
  58 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RWTUN, &g_mirror_debug, 0,
  59     "Debug level");
  60 static u_int g_mirror_timeout = 4;
  61 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RWTUN, &g_mirror_timeout,
  62     0, "Time to wait on all mirror components");
  63 static u_int g_mirror_idletime = 5;
  64 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, idletime, CTLFLAG_RWTUN,
  65     &g_mirror_idletime, 0, "Mark components as clean when idling");
  66 static u_int g_mirror_disconnect_on_failure = 1;
  67 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, disconnect_on_failure, CTLFLAG_RWTUN,
  68     &g_mirror_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
  69 static u_int g_mirror_syncreqs = 2;
  70 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
  71     &g_mirror_syncreqs, 0, "Parallel synchronization I/O requests.");
  72
  73 #define MSLEEP(ident, mtx, priority, wmesg, timeout)    do {            \
  74         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));       \
  75         msleep((ident), (mtx), (priority), (wmesg), (timeout));         \
  76         G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, (ident));       \
  77 } while (0)
  78
  79 static eventhandler_tag g_mirror_post_sync = NULL;
  80 static int g_mirror_shutdown = 0;
  81
  82 static int g_mirror_destroy_geom(struct gctl_req *req, struct g_class *mp,
  83     struct g_geom *gp);
  84 static g_taste_t g_mirror_taste;
  85 static g_resize_t g_mirror_resize;
  86 static void g_mirror_init(struct g_class *mp);
  87 static void g_mirror_fini(struct g_class *mp);
  88
  89 struct g_class g_mirror_class = {
  90         .name = G_MIRROR_CLASS_NAME,
  91         .version = G_VERSION,
  92         .ctlreq = g_mirror_config,
  93         .taste = g_mirror_taste,
  94         .destroy_geom = g_mirror_destroy_geom,
  95         .init = g_mirror_init,
  96         .fini = g_mirror_fini,
  97         .resize = g_mirror_resize
  98 };
  99
 100
 101 static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
 102 static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
 103 static void g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force);
 104 static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
 105     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
 106 static void g_mirror_sync_stop(struct g_mirror_disk *disk, int type);
 107 static void g_mirror_register_request(struct bio *bp);
 108 static void g_mirror_sync_release(struct g_mirror_softc *sc);
 109
 110
 111 static const char *
 112 g_mirror_disk_state2str(int state)
 113 {
 114
 115         switch (state) {
 116         case G_MIRROR_DISK_STATE_NONE:
 117                 return ("NONE");
 118         case G_MIRROR_DISK_STATE_NEW:
 119                 return ("NEW");
 120         case G_MIRROR_DISK_STATE_ACTIVE:
 121                 return ("ACTIVE");
 122         case G_MIRROR_DISK_STATE_STALE:
 123                 return ("STALE");
 124         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
 125                 return ("SYNCHRONIZING");
 126         case G_MIRROR_DISK_STATE_DISCONNECTED:
 127                 return ("DISCONNECTED");
 128         case G_MIRROR_DISK_STATE_DESTROY:
 129                 return ("DESTROY");
 130         default:
 131                 return ("INVALID");
 132         }
 133 }
 134
 135 static const char *
 136 g_mirror_device_state2str(int state)
 137 {
 138
 139         switch (state) {
 140         case G_MIRROR_DEVICE_STATE_STARTING:
 141                 return ("STARTING");
 142         case G_MIRROR_DEVICE_STATE_RUNNING:
 143                 return ("RUNNING");
 144         default:
 145                 return ("INVALID");
 146         }
 147 }
 148
 149 static const char *
 150 g_mirror_get_diskname(struct g_mirror_disk *disk)
 151 {
 152
 153         if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
 154                 return ("[unknown]");
 155         return (disk->d_name);
 156 }
 157
 158 /*
 159  * --- Events handling functions ---
 160  * Events in geom_mirror are used to maintain disks and device status
 161  * from one thread to simplify locking.
 162  */
 163 static void
 164 g_mirror_event_free(struct g_mirror_event *ep)
 165 {
 166
 167         free(ep, M_MIRROR);
 168 }
 169
 170 int
 171 g_mirror_event_send(void *arg, int state, int flags)
 172 {
 173         struct g_mirror_softc *sc;
 174         struct g_mirror_disk *disk;
 175         struct g_mirror_event *ep;
 176         int error;
 177
 178         ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
 179         G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
 180         if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
 181                 disk = NULL;
 182                 sc = arg;
 183         } else {
 184                 disk = arg;
 185                 sc = disk->d_softc;
 186         }
 187         ep->e_disk = disk;
 188         ep->e_state = state;
 189         ep->e_flags = flags;
 190         ep->e_error = 0;
 191         mtx_lock(&sc->sc_events_mtx);
 192         TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
 193         mtx_unlock(&sc->sc_events_mtx);
 194         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
 195         mtx_lock(&sc->sc_queue_mtx);
 196         wakeup(sc);
 197         mtx_unlock(&sc->sc_queue_mtx);
 198         if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
 199                 return (0);
 200         sx_assert(&sc->sc_lock, SX_XLOCKED);
 201         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
 202         sx_xunlock(&sc->sc_lock);
 203         while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
 204                 mtx_lock(&sc->sc_events_mtx);
 205                 MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "m:event",
 206                     hz * 5);
 207         }
 208         error = ep->e_error;
 209         g_mirror_event_free(ep);
 210         sx_xlock(&sc->sc_lock);
 211         return (error);
 212 }
 213
 214 static struct g_mirror_event *
 215 g_mirror_event_get(struct g_mirror_softc *sc)
 216 {
 217         struct g_mirror_event *ep;
 218
 219         mtx_lock(&sc->sc_events_mtx);
 220         ep = TAILQ_FIRST(&sc->sc_events);
 221         mtx_unlock(&sc->sc_events_mtx);
 222         return (ep);
 223 }
 224
 225 static void
 226 g_mirror_event_remove(struct g_mirror_softc *sc, struct g_mirror_event *ep)
 227 {
 228
 229         mtx_lock(&sc->sc_events_mtx);
 230         TAILQ_REMOVE(&sc->sc_events, ep, e_next);
 231         mtx_unlock(&sc->sc_events_mtx);
 232 }
 233
 234 static void
 235 g_mirror_event_cancel(struct g_mirror_disk *disk)
 236 {
 237         struct g_mirror_softc *sc;
 238         struct g_mirror_event *ep, *tmpep;
 239
 240         sc = disk->d_softc;
 241         sx_assert(&sc->sc_lock, SX_XLOCKED);
 242
 243         mtx_lock(&sc->sc_events_mtx);
 244         TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
 245                 if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
 246                         continue;
 247                 if (ep->e_disk != disk)
 248                         continue;
 249                 TAILQ_REMOVE(&sc->sc_events, ep, e_next);
 250                 if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
 251                         g_mirror_event_free(ep);
 252                 else {
 253                         ep->e_error = ECANCELED;
 254                         wakeup(ep);
 255                 }
 256         }
 257         mtx_unlock(&sc->sc_events_mtx);
 258 }
 259
 260 /*
 261  * Return the number of disks in given state.
 262  * If state is equal to -1, count all connected disks.
 263  */
 264 u_int
 265 g_mirror_ndisks(struct g_mirror_softc *sc, int state)
 266 {
 267         struct g_mirror_disk *disk;
 268         u_int n = 0;
 269
 270         sx_assert(&sc->sc_lock, SX_LOCKED);
 271
 272         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 273                 if (state == -1 || disk->d_state == state)
 274                         n++;
 275         }
 276         return (n);
 277 }
 278
 279 /*
 280  * Find a disk in mirror by its disk ID.
 281  */
 282 static struct g_mirror_disk *
 283 g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
 284 {
 285         struct g_mirror_disk *disk;
 286
 287         sx_assert(&sc->sc_lock, SX_XLOCKED);
 288
 289         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 290                 if (disk->d_id == id)
 291                         return (disk);
 292         }
 293         return (NULL);
 294 }
 295
 296 static u_int
 297 g_mirror_nrequests(struct g_mirror_softc *sc, struct g_consumer *cp)
 298 {
 299         struct bio *bp;
 300         u_int nreqs = 0;
 301
 302         mtx_lock(&sc->sc_queue_mtx);
 303         TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
 304                 if (bp->bio_from == cp)
 305                         nreqs++;
 306         }
 307         mtx_unlock(&sc->sc_queue_mtx);
 308         return (nreqs);
 309 }
 310
 311 static int
 312 g_mirror_is_busy(struct g_mirror_softc *sc, struct g_consumer *cp)
 313 {
 314
 315         if (cp->index > 0) {
 316                 G_MIRROR_DEBUG(2,
 317                     "I/O requests for %s exist, can't destroy it now.",
 318                     cp->provider->name);
 319                 return (1);
 320         }
 321         if (g_mirror_nrequests(sc, cp) > 0) {
 322                 G_MIRROR_DEBUG(2,
 323                     "I/O requests for %s in queue, can't destroy it now.",
 324                     cp->provider->name);
 325                 return (1);
 326         }
 327         return (0);
 328 }
 329
 330 static void
 331 g_mirror_destroy_consumer(void *arg, int flags __unused)
 332 {
 333         struct g_consumer *cp;
 334
 335         g_topology_assert();
 336
 337         cp = arg;
 338         G_MIRROR_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
 339         g_detach(cp);
 340         g_destroy_consumer(cp);
 341 }
 342
 343 static void
 344 g_mirror_kill_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
 345 {
 346         struct g_provider *pp;
 347         int retaste_wait;
 348
 349         g_topology_assert();
 350
 351         cp->private = NULL;
 352         if (g_mirror_is_busy(sc, cp))
 353                 return;
 354         pp = cp->provider;
 355         retaste_wait = 0;
 356         if (cp->acw == 1) {
 357                 if ((pp->geom->flags & G_GEOM_WITHER) == 0)
 358                         retaste_wait = 1;
 359         }
 360         G_MIRROR_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
 361             -cp->acw, -cp->ace, 0);
 362         if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
 363                 g_access(cp, -cp->acr, -cp->acw, -cp->ace);
 364         if (retaste_wait) {
 365                 /*
 366                  * After retaste event was send (inside g_access()), we can send
 367                  * event to detach and destroy consumer.
 368                  * A class, which has consumer to the given provider connected
 369                  * will not receive retaste event for the provider.
 370                  * This is the way how I ignore retaste events when I close
 371                  * consumers opened for write: I detach and destroy consumer
 372                  * after retaste event is sent.
 373                  */
 374                 g_post_event(g_mirror_destroy_consumer, cp, M_WAITOK, NULL);
 375                 return;
 376         }
 377         G_MIRROR_DEBUG(1, "Consumer %s destroyed.", pp->name);
 378         g_detach(cp);
 379         g_destroy_consumer(cp);
 380 }
 381
 382 static int
 383 g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
 384 {
 385         struct g_consumer *cp;
 386         int error;
 387
 388         g_topology_assert_not();
 389         KASSERT(disk->d_consumer == NULL,
 390             ("Disk already connected (device %s).", disk->d_softc->sc_name));
 391
 392         g_topology_lock();
 393         cp = g_new_consumer(disk->d_softc->sc_geom);
 394         cp->flags |= G_CF_DIRECT_RECEIVE;
 395         error = g_attach(cp, pp);
 396         if (error != 0) {
 397                 g_destroy_consumer(cp);
 398                 g_topology_unlock();
 399                 return (error);
 400         }
 401         error = g_access(cp, 1, 1, 1);
 402         if (error != 0) {
 403                 g_detach(cp);
 404                 g_destroy_consumer(cp);
 405                 g_topology_unlock();
 406                 G_MIRROR_DEBUG(0, "Cannot open consumer %s (error=%d).",
 407                     pp->name, error);
 408                 return (error);
 409         }
 410         g_topology_unlock();
 411         disk->d_consumer = cp;
 412         disk->d_consumer->private = disk;
 413         disk->d_consumer->index = 0;
 414
 415         G_MIRROR_DEBUG(2, "Disk %s connected.", g_mirror_get_diskname(disk));
 416         return (0);
 417 }
 418
 419 static void
 420 g_mirror_disconnect_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
 421 {
 422
 423         g_topology_assert();
 424
 425         if (cp == NULL)
 426                 return;
 427         if (cp->provider != NULL)
 428                 g_mirror_kill_consumer(sc, cp);
 429         else
 430                 g_destroy_consumer(cp);
 431 }
 432
 433 /*
 434  * Initialize disk. This means allocate memory, create consumer, attach it
 435  * to the provider and open access (r1w1e1) to it.
 436  */
 437 static struct g_mirror_disk *
 438 g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
 439     struct g_mirror_metadata *md, int *errorp)
 440 {
 441         struct g_mirror_disk *disk;
 442         int i, error;
 443
 444         disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
 445         if (disk == NULL) {
 446                 error = ENOMEM;
 447                 goto fail;
 448         }
 449         disk->d_softc = sc;
 450         error = g_mirror_connect_disk(disk, pp);
 451         if (error != 0)
 452                 goto fail;
 453         disk->d_id = md->md_did;
 454         disk->d_state = G_MIRROR_DISK_STATE_NONE;
 455         disk->d_priority = md->md_priority;
 456         disk->d_flags = md->md_dflags;
 457         error = g_getattr("GEOM::candelete", disk->d_consumer, &i);
 458         if (error == 0 && i != 0)
 459                 disk->d_flags |= G_MIRROR_DISK_FLAG_CANDELETE;
 460         if (md->md_provider[0] != '\0')
 461                 disk->d_flags |= G_MIRROR_DISK_FLAG_HARDCODED;
 462         disk->d_sync.ds_consumer = NULL;
 463         disk->d_sync.ds_offset = md->md_sync_offset;
 464         disk->d_sync.ds_offset_done = md->md_sync_offset;
 465         disk->d_genid = md->md_genid;
 466         disk->d_sync.ds_syncid = md->md_syncid;
 467         if (errorp != NULL)
 468                 *errorp = 0;
 469         return (disk);
 470 fail:
 471         if (errorp != NULL)
 472                 *errorp = error;
 473         if (disk != NULL)
 474                 free(disk, M_MIRROR);
 475         return (NULL);
 476 }
 477
 478 static void
 479 g_mirror_destroy_disk(struct g_mirror_disk *disk)
 480 {
 481         struct g_mirror_softc *sc;
 482
 483         g_topology_assert_not();
 484         sc = disk->d_softc;
 485         sx_assert(&sc->sc_lock, SX_XLOCKED);
 486
 487         LIST_REMOVE(disk, d_next);
 488         g_mirror_event_cancel(disk);
 489         if (sc->sc_hint == disk)
 490                 sc->sc_hint = NULL;
 491         switch (disk->d_state) {
 492         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
 493                 g_mirror_sync_stop(disk, 1);
 494                 /* FALLTHROUGH */
 495         case G_MIRROR_DISK_STATE_NEW:
 496         case G_MIRROR_DISK_STATE_STALE:
 497         case G_MIRROR_DISK_STATE_ACTIVE:
 498                 g_topology_lock();
 499                 g_mirror_disconnect_consumer(sc, disk->d_consumer);
 500                 g_topology_unlock();
 501                 free(disk, M_MIRROR);
 502                 break;
 503         default:
 504                 KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
 505                     g_mirror_get_diskname(disk),
 506                     g_mirror_disk_state2str(disk->d_state)));
 507         }
 508 }
 509
 510 static void
 511 g_mirror_destroy_device(struct g_mirror_softc *sc)
 512 {
 513         struct g_mirror_disk *disk;
 514         struct g_mirror_event *ep;
 515         struct g_geom *gp;
 516         struct g_consumer *cp, *tmpcp;
 517
 518         g_topology_assert_not();
 519         sx_assert(&sc->sc_lock, SX_XLOCKED);
 520
 521         gp = sc->sc_geom;
 522         if (sc->sc_provider != NULL)
 523                 g_mirror_destroy_provider(sc);
 524         for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
 525             disk = LIST_FIRST(&sc->sc_disks)) {
 526                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
 527                 g_mirror_update_metadata(disk);
 528                 g_mirror_destroy_disk(disk);
 529         }
 530         while ((ep = g_mirror_event_get(sc)) != NULL) {
 531                 g_mirror_event_remove(sc, ep);
 532                 if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
 533                         g_mirror_event_free(ep);
 534                 else {
 535                         ep->e_error = ECANCELED;
 536                         ep->e_flags |= G_MIRROR_EVENT_DONE;
 537                         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
 538                         mtx_lock(&sc->sc_events_mtx);
 539                         wakeup(ep);
 540                         mtx_unlock(&sc->sc_events_mtx);
 541                 }
 542         }
 543         callout_drain(&sc->sc_callout);
 544
 545         g_topology_lock();
 546         LIST_FOREACH_SAFE(cp, &sc->sc_sync.ds_geom->consumer, consumer, tmpcp) {
 547                 g_mirror_disconnect_consumer(sc, cp);
 548         }
 549         g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
 550         G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
 551         g_wither_geom(gp, ENXIO);
 552         g_topology_unlock();
 553         mtx_destroy(&sc->sc_queue_mtx);
 554         mtx_destroy(&sc->sc_events_mtx);
 555         mtx_destroy(&sc->sc_done_mtx);
 556         sx_xunlock(&sc->sc_lock);
 557         sx_destroy(&sc->sc_lock);
 558 }
 559
 560 static void
 561 g_mirror_orphan(struct g_consumer *cp)
 562 {
 563         struct g_mirror_disk *disk;
 564
 565         g_topology_assert();
 566
 567         disk = cp->private;
 568         if (disk == NULL)
 569                 return;
 570         disk->d_softc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
 571         g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
 572             G_MIRROR_EVENT_DONTWAIT);
 573 }
 574
 575 /*
 576  * Function should return the next active disk on the list.
 577  * It is possible that it will be the same disk as given.
 578  * If there are no active disks on list, NULL is returned.
 579  */
 580 static __inline struct g_mirror_disk *
 581 g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
 582 {
 583         struct g_mirror_disk *dp;
 584
 585         for (dp = LIST_NEXT(disk, d_next); dp != disk;
 586             dp = LIST_NEXT(dp, d_next)) {
 587                 if (dp == NULL)
 588                         dp = LIST_FIRST(&sc->sc_disks);
 589                 if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
 590                         break;
 591         }
 592         if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
 593                 return (NULL);
 594         return (dp);
 595 }
 596
 597 static struct g_mirror_disk *
 598 g_mirror_get_disk(struct g_mirror_softc *sc)
 599 {
 600         struct g_mirror_disk *disk;
 601
 602         if (sc->sc_hint == NULL) {
 603                 sc->sc_hint = LIST_FIRST(&sc->sc_disks);
 604                 if (sc->sc_hint == NULL)
 605                         return (NULL);
 606         }
 607         disk = sc->sc_hint;
 608         if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
 609                 disk = g_mirror_find_next(sc, disk);
 610                 if (disk == NULL)
 611                         return (NULL);
 612         }
 613         sc->sc_hint = g_mirror_find_next(sc, disk);
 614         return (disk);
 615 }
 616
 617 static int
 618 g_mirror_write_metadata(struct g_mirror_disk *disk,
 619     struct g_mirror_metadata *md)
 620 {
 621         struct g_mirror_softc *sc;
 622         struct g_consumer *cp;
 623         off_t offset, length;
 624         u_char *sector;
 625         int error = 0;
 626
 627         g_topology_assert_not();
 628         sc = disk->d_softc;
 629         sx_assert(&sc->sc_lock, SX_LOCKED);
 630
 631         cp = disk->d_consumer;
 632         KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
 633         KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
 634         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
 635             ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
 636             cp->acw, cp->ace));
 637         length = cp->provider->sectorsize;
 638         offset = cp->provider->mediasize - length;
 639         sector = malloc((size_t)length, M_MIRROR, M_WAITOK | M_ZERO);
 640         if (md != NULL &&
 641             (sc->sc_flags & G_MIRROR_DEVICE_FLAG_WIPE) == 0) {
 642                 /*
 643                  * Handle the case, when the size of parent provider reduced.
 644                  */
 645                 if (offset < md->md_mediasize)
 646                         error = ENOSPC;
 647                 else
 648                         mirror_metadata_encode(md, sector);
 649         }
 650         KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_metadata_write, error);
 651         if (error == 0)
 652                 error = g_write_data(cp, offset, sector, length);
 653         free(sector, M_MIRROR);
 654         if (error != 0) {
 655                 if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
 656                         disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
 657                         G_MIRROR_DEBUG(0, "Cannot write metadata on %s "
 658                             "(device=%s, error=%d).",
 659                             g_mirror_get_diskname(disk), sc->sc_name, error);
 660                 } else {
 661                         G_MIRROR_DEBUG(1, "Cannot write metadata on %s "
 662                             "(device=%s, error=%d).",
 663                             g_mirror_get_diskname(disk), sc->sc_name, error);
 664                 }
 665                 if (g_mirror_disconnect_on_failure &&
 666                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
 667                         sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
 668                         g_mirror_event_send(disk,
 669                             G_MIRROR_DISK_STATE_DISCONNECTED,
 670                             G_MIRROR_EVENT_DONTWAIT);
 671                 }
 672         }
 673         return (error);
 674 }
 675
 676 static int
 677 g_mirror_clear_metadata(struct g_mirror_disk *disk)
 678 {
 679         int error;
 680
 681         g_topology_assert_not();
 682         sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
 683
 684         error = g_mirror_write_metadata(disk, NULL);
 685         if (error == 0) {
 686                 G_MIRROR_DEBUG(2, "Metadata on %s cleared.",
 687                     g_mirror_get_diskname(disk));
 688         } else {
 689                 G_MIRROR_DEBUG(0,
 690                     "Cannot clear metadata on disk %s (error=%d).",
 691                     g_mirror_get_diskname(disk), error);
 692         }
 693         return (error);
 694 }
 695
 696 void
 697 g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
 698     struct g_mirror_metadata *md)
 699 {
 700
 701         strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
 702         md->md_version = G_MIRROR_VERSION;
 703         strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
 704         md->md_mid = sc->sc_id;
 705         md->md_all = sc->sc_ndisks;
 706         md->md_slice = sc->sc_slice;
 707         md->md_balance = sc->sc_balance;
 708         md->md_genid = sc->sc_genid;
 709         md->md_mediasize = sc->sc_mediasize;
 710         md->md_sectorsize = sc->sc_sectorsize;
 711         md->md_mflags = (sc->sc_flags & G_MIRROR_DEVICE_FLAG_MASK);
 712         bzero(md->md_provider, sizeof(md->md_provider));
 713         if (disk == NULL) {
 714                 md->md_did = arc4random();
 715                 md->md_priority = 0;
 716                 md->md_syncid = 0;
 717                 md->md_dflags = 0;
 718                 md->md_sync_offset = 0;
 719                 md->md_provsize = 0;
 720         } else {
 721                 md->md_did = disk->d_id;
 722                 md->md_priority = disk->d_priority;
 723                 md->md_syncid = disk->d_sync.ds_syncid;
 724                 md->md_dflags = (disk->d_flags & G_MIRROR_DISK_FLAG_MASK);
 725                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
 726                         md->md_sync_offset = disk->d_sync.ds_offset_done;
 727                 else
 728                         md->md_sync_offset = 0;
 729                 if ((disk->d_flags & G_MIRROR_DISK_FLAG_HARDCODED) != 0) {
 730                         strlcpy(md->md_provider,
 731                             disk->d_consumer->provider->name,
 732                             sizeof(md->md_provider));
 733                 }
 734                 md->md_provsize = disk->d_consumer->provider->mediasize;
 735         }
 736 }
 737
 738 void
 739 g_mirror_update_metadata(struct g_mirror_disk *disk)
 740 {
 741         struct g_mirror_softc *sc;
 742         struct g_mirror_metadata md;
 743         int error;
 744
 745         g_topology_assert_not();
 746         sc = disk->d_softc;
 747         sx_assert(&sc->sc_lock, SX_LOCKED);
 748
 749         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WIPE) == 0)
 750                 g_mirror_fill_metadata(sc, disk, &md);
 751         error = g_mirror_write_metadata(disk, &md);
 752         if (error == 0) {
 753                 G_MIRROR_DEBUG(2, "Metadata on %s updated.",
 754                     g_mirror_get_diskname(disk));
 755         } else {
 756                 G_MIRROR_DEBUG(0,
 757                     "Cannot update metadata on disk %s (error=%d).",
 758                     g_mirror_get_diskname(disk), error);
 759         }
 760 }
 761
 762 static void
 763 g_mirror_bump_syncid(struct g_mirror_softc *sc)
 764 {
 765         struct g_mirror_disk *disk;
 766
 767         g_topology_assert_not();
 768         sx_assert(&sc->sc_lock, SX_XLOCKED);
 769         KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
 770             ("%s called with no active disks (device=%s).", __func__,
 771             sc->sc_name));
 772
 773         sc->sc_syncid++;
 774         G_MIRROR_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
 775             sc->sc_syncid);
 776         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 777                 if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
 778                     disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
 779                         disk->d_sync.ds_syncid = sc->sc_syncid;
 780                         g_mirror_update_metadata(disk);
 781                 }
 782         }
 783 }
 784
 785 static void
 786 g_mirror_bump_genid(struct g_mirror_softc *sc)
 787 {
 788         struct g_mirror_disk *disk;
 789
 790         g_topology_assert_not();
 791         sx_assert(&sc->sc_lock, SX_XLOCKED);
 792         KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
 793             ("%s called with no active disks (device=%s).", __func__,
 794             sc->sc_name));
 795
 796         sc->sc_genid++;
 797         G_MIRROR_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
 798             sc->sc_genid);
 799         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 800                 if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
 801                     disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
 802                         disk->d_genid = sc->sc_genid;
 803                         g_mirror_update_metadata(disk);
 804                 }
 805         }
 806 }
 807
 808 static int
 809 g_mirror_idle(struct g_mirror_softc *sc, int acw)
 810 {
 811         struct g_mirror_disk *disk;
 812         int timeout;
 813
 814         g_topology_assert_not();
 815         sx_assert(&sc->sc_lock, SX_XLOCKED);
 816
 817         if (sc->sc_provider == NULL)
 818                 return (0);
 819         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
 820                 return (0);
 821         if (sc->sc_idle)
 822                 return (0);
 823         if (sc->sc_writes > 0)
 824                 return (0);
 825         if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
 826                 timeout = g_mirror_idletime - (time_uptime - sc->sc_last_write);
 827                 if (!g_mirror_shutdown && timeout > 0)
 828                         return (timeout);
 829         }
 830         sc->sc_idle = 1;
 831         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 832                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
 833                         continue;
 834                 G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as clean.",
 835                     g_mirror_get_diskname(disk), sc->sc_name);
 836                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
 837                 g_mirror_update_metadata(disk);
 838         }
 839         return (0);
 840 }
 841
 842 static void
 843 g_mirror_unidle(struct g_mirror_softc *sc)
 844 {
 845         struct g_mirror_disk *disk;
 846
 847         g_topology_assert_not();
 848         sx_assert(&sc->sc_lock, SX_XLOCKED);
 849
 850         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
 851                 return;
 852         sc->sc_idle = 0;
 853         sc->sc_last_write = time_uptime;
 854         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 855                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
 856                         continue;
 857                 G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as dirty.",
 858                     g_mirror_get_diskname(disk), sc->sc_name);
 859                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
 860                 g_mirror_update_metadata(disk);
 861         }
 862 }
 863
 864 static void
 865 g_mirror_flush_done(struct bio *bp)
 866 {
 867         struct g_mirror_softc *sc;
 868         struct bio *pbp;
 869
 870         pbp = bp->bio_parent;
 871         sc = pbp->bio_to->geom->softc;
 872         mtx_lock(&sc->sc_done_mtx);
 873         if (pbp->bio_error == 0)
 874                 pbp->bio_error = bp->bio_error;
 875         pbp->bio_completed += bp->bio_completed;
 876         pbp->bio_inbed++;
 877         if (pbp->bio_children == pbp->bio_inbed) {
 878                 mtx_unlock(&sc->sc_done_mtx);
 879                 g_io_deliver(pbp, pbp->bio_error);
 880         } else
 881                 mtx_unlock(&sc->sc_done_mtx);
 882         g_destroy_bio(bp);
 883 }
 884
 885 static void
 886 g_mirror_done(struct bio *bp)
 887 {
 888         struct g_mirror_softc *sc;
 889
 890         sc = bp->bio_from->geom->softc;
 891         bp->bio_cflags = G_MIRROR_BIO_FLAG_REGULAR;
 892         mtx_lock(&sc->sc_queue_mtx);
 893         bioq_insert_tail(&sc->sc_queue, bp);
 894         mtx_unlock(&sc->sc_queue_mtx);
 895         wakeup(sc);
 896 }
 897
 898 static void
 899 g_mirror_regular_request(struct bio *bp)
 900 {
 901         struct g_mirror_softc *sc;
 902         struct g_mirror_disk *disk;
 903         struct bio *pbp;
 904
 905         g_topology_assert_not();
 906
 907         pbp = bp->bio_parent;
 908         sc = pbp->bio_to->geom->softc;
 909         bp->bio_from->index--;
 910         if (bp->bio_cmd == BIO_WRITE)
 911                 sc->sc_writes--;
 912         disk = bp->bio_from->private;
 913         if (disk == NULL) {
 914                 g_topology_lock();
 915                 g_mirror_kill_consumer(sc, bp->bio_from);
 916                 g_topology_unlock();
 917         }
 918
 919         if (bp->bio_cmd == BIO_READ)
 920                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_read,
 921                     bp->bio_error);
 922         else if (bp->bio_cmd == BIO_WRITE)
 923                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_write,
 924                     bp->bio_error);
 925
 926         pbp->bio_inbed++;
 927         KASSERT(pbp->bio_inbed <= pbp->bio_children,
 928             ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
 929             pbp->bio_children));
 930         if (bp->bio_error == 0 && pbp->bio_error == 0) {
 931                 G_MIRROR_LOGREQ(3, bp, "Request delivered.");
 932                 g_destroy_bio(bp);
 933                 if (pbp->bio_children == pbp->bio_inbed) {
 934                         G_MIRROR_LOGREQ(3, pbp, "Request delivered.");
 935                         pbp->bio_completed = pbp->bio_length;
 936                         if (pbp->bio_cmd == BIO_WRITE ||
 937                             pbp->bio_cmd == BIO_DELETE) {
 938                                 bioq_remove(&sc->sc_inflight, pbp);
 939                                 /* Release delayed sync requests if possible. */
 940                                 g_mirror_sync_release(sc);
 941                         }
 942                         g_io_deliver(pbp, pbp->bio_error);
 943                 }
 944                 return;
 945         } else if (bp->bio_error != 0) {
 946                 if (pbp->bio_error == 0)
 947                         pbp->bio_error = bp->bio_error;
 948                 if (disk != NULL) {
 949                         if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
 950                                 disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
 951                                 G_MIRROR_LOGREQ(0, bp,
 952                                     "Request failed (error=%d).",
 953                                     bp->bio_error);
 954                         } else {
 955                                 G_MIRROR_LOGREQ(1, bp,
 956                                     "Request failed (error=%d).",
 957                                     bp->bio_error);
 958                         }
 959                         if (g_mirror_disconnect_on_failure &&
 960                             g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1)
 961                         {
 962                                 sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
 963                                 g_mirror_event_send(disk,
 964                                     G_MIRROR_DISK_STATE_DISCONNECTED,
 965                                     G_MIRROR_EVENT_DONTWAIT);
 966                         }
 967                 }
 968                 switch (pbp->bio_cmd) {
 969                 case BIO_DELETE:
 970                 case BIO_WRITE:
 971                         pbp->bio_inbed--;
 972                         pbp->bio_children--;
 973                         break;
 974                 }
 975         }
 976         g_destroy_bio(bp);
 977
 978         switch (pbp->bio_cmd) {
 979         case BIO_READ:
 980                 if (pbp->bio_inbed < pbp->bio_children)
 981                         break;
 982                 if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1)
 983                         g_io_deliver(pbp, pbp->bio_error);
 984                 else {
 985                         pbp->bio_error = 0;
 986                         mtx_lock(&sc->sc_queue_mtx);
 987                         bioq_insert_tail(&sc->sc_queue, pbp);
 988                         mtx_unlock(&sc->sc_queue_mtx);
 989                         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
 990                         wakeup(sc);
 991                 }
 992                 break;
 993         case BIO_DELETE:
 994         case BIO_WRITE:
 995                 if (pbp->bio_children == 0) {
 996                         /*
 997                          * All requests failed.
 998                          */
 999                 } else if (pbp->bio_inbed < pbp->bio_children) {
1000                         /* Do nothing. */
1001                         break;
1002                 } else if (pbp->bio_children == pbp->bio_inbed) {
1003                         /* Some requests succeeded. */
1004                         pbp->bio_error = 0;
1005                         pbp->bio_completed = pbp->bio_length;
1006                 }
1007                 bioq_remove(&sc->sc_inflight, pbp);
1008                 /* Release delayed sync requests if possible. */
1009                 g_mirror_sync_release(sc);
1010                 g_io_deliver(pbp, pbp->bio_error);
1011                 break;
1012         default:
1013                 KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
1014                 break;
1015         }
1016 }
1017
1018 static void
1019 g_mirror_sync_done(struct bio *bp)
1020 {
1021         struct g_mirror_softc *sc;
1022
1023         G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
1024         sc = bp->bio_from->geom->softc;
1025         bp->bio_cflags = G_MIRROR_BIO_FLAG_SYNC;
1026         mtx_lock(&sc->sc_queue_mtx);
1027         bioq_insert_tail(&sc->sc_queue, bp);
1028         mtx_unlock(&sc->sc_queue_mtx);
1029         wakeup(sc);
1030 }
1031
1032 static void
1033 g_mirror_candelete(struct bio *bp)
1034 {
1035         struct g_mirror_softc *sc;
1036         struct g_mirror_disk *disk;
1037         int *val;
1038
1039         sc = bp->bio_to->geom->softc;
1040         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1041                 if (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE)
1042                         break;
1043         }
1044         val = (int *)bp->bio_data;
1045         *val = (disk != NULL);
1046         g_io_deliver(bp, 0);
1047 }
1048
1049 static void
1050 g_mirror_kernel_dump(struct bio *bp)
1051 {
1052         struct g_mirror_softc *sc;
1053         struct g_mirror_disk *disk;
1054         struct bio *cbp;
1055         struct g_kerneldump *gkd;
1056
1057         /*
1058          * We configure dumping to the first component, because this component
1059          * will be used for reading with 'prefer' balance algorithm.
1060          * If the component with the highest priority is currently disconnected
1061          * we will not be able to read the dump after the reboot if it will be
1062          * connected and synchronized later. Can we do something better?
1063          */
1064         sc = bp->bio_to->geom->softc;
1065         disk = LIST_FIRST(&sc->sc_disks);
1066
1067         gkd = (struct g_kerneldump *)bp->bio_data;
1068         if (gkd->length > bp->bio_to->mediasize)
1069                 gkd->length = bp->bio_to->mediasize;
1070         cbp = g_clone_bio(bp);
1071         if (cbp == NULL) {
1072                 g_io_deliver(bp, ENOMEM);
1073                 return;
1074         }
1075         cbp->bio_done = g_std_done;
1076         g_io_request(cbp, disk->d_consumer);
1077         G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1078             g_mirror_get_diskname(disk));
1079 }
1080
1081 static void
1082 g_mirror_flush(struct g_mirror_softc *sc, struct bio *bp)
1083 {
1084         struct bio_queue_head queue;
1085         struct g_mirror_disk *disk;
1086         struct g_consumer *cp;
1087         struct bio *cbp;
1088
1089         bioq_init(&queue);
1090         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1091                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1092                         continue;
1093                 cbp = g_clone_bio(bp);
1094                 if (cbp == NULL) {
1095                         while ((cbp = bioq_takefirst(&queue)) != NULL)
1096                                 g_destroy_bio(cbp);
1097                         if (bp->bio_error == 0)
1098                                 bp->bio_error = ENOMEM;
1099                         g_io_deliver(bp, bp->bio_error);
1100                         return;
1101                 }
1102                 bioq_insert_tail(&queue, cbp);
1103                 cbp->bio_done = g_mirror_flush_done;
1104                 cbp->bio_caller1 = disk;
1105                 cbp->bio_to = disk->d_consumer->provider;
1106         }
1107         while ((cbp = bioq_takefirst(&queue)) != NULL) {
1108                 G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1109                 disk = cbp->bio_caller1;
1110                 cbp->bio_caller1 = NULL;
1111                 cp = disk->d_consumer;
1112                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1113                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1114                     cp->acr, cp->acw, cp->ace));
1115                 g_io_request(cbp, disk->d_consumer);
1116         }
1117 }
1118
1119 static void
1120 g_mirror_start(struct bio *bp)
1121 {
1122         struct g_mirror_softc *sc;
1123
1124         sc = bp->bio_to->geom->softc;
1125         /*
1126          * If sc == NULL or there are no valid disks, provider's error
1127          * should be set and g_mirror_start() should not be called at all.
1128          */
1129         KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1130             ("Provider's error should be set (error=%d)(mirror=%s).",
1131             bp->bio_to->error, bp->bio_to->name));
1132         G_MIRROR_LOGREQ(3, bp, "Request received.");
1133
1134         switch (bp->bio_cmd) {
1135         case BIO_READ:
1136         case BIO_WRITE:
1137         case BIO_DELETE:
1138                 break;
1139         case BIO_FLUSH:
1140                 g_mirror_flush(sc, bp);
1141                 return;
1142         case BIO_GETATTR:
1143                 if (!strcmp(bp->bio_attribute, "GEOM::candelete")) {
1144                         g_mirror_candelete(bp);
1145                         return;
1146                 } else if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1147                         g_mirror_kernel_dump(bp);
1148                         return;
1149                 }
1150                 /* FALLTHROUGH */
1151         default:
1152                 g_io_deliver(bp, EOPNOTSUPP);
1153                 return;
1154         }
1155         mtx_lock(&sc->sc_queue_mtx);
1156         bioq_insert_tail(&sc->sc_queue, bp);
1157         mtx_unlock(&sc->sc_queue_mtx);
1158         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1159         wakeup(sc);
1160 }
1161
1162 /*
1163  * Return TRUE if the given request is colliding with a in-progress
1164  * synchronization request.
1165  */
1166 static int
1167 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1168 {
1169         struct g_mirror_disk *disk;
1170         struct bio *sbp;
1171         off_t rstart, rend, sstart, send;
1172         u_int i;
1173
1174         if (sc->sc_sync.ds_ndisks == 0)
1175                 return (0);
1176         rstart = bp->bio_offset;
1177         rend = bp->bio_offset + bp->bio_length;
1178         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1179                 if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1180                         continue;
1181                 for (i = 0; i < g_mirror_syncreqs; i++) {
1182                         sbp = disk->d_sync.ds_bios[i];
1183                         if (sbp == NULL)
1184                                 continue;
1185                         sstart = sbp->bio_offset;
1186                         send = sbp->bio_offset + sbp->bio_length;
1187                         if (rend > sstart && rstart < send)
1188                                 return (1);
1189                 }
1190         }
1191         return (0);
1192 }
1193
1194 /*
1195  * Return TRUE if the given sync request is colliding with a in-progress regular
1196  * request.
1197  */
1198 static int
1199 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1200 {
1201         off_t rstart, rend, sstart, send;
1202         struct bio *bp;
1203
1204         if (sc->sc_sync.ds_ndisks == 0)
1205                 return (0);
1206         sstart = sbp->bio_offset;
1207         send = sbp->bio_offset + sbp->bio_length;
1208         TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1209                 rstart = bp->bio_offset;
1210                 rend = bp->bio_offset + bp->bio_length;
1211                 if (rend > sstart && rstart < send)
1212                         return (1);
1213         }
1214         return (0);
1215 }
1216
1217 /*
1218  * Puts request onto delayed queue.
1219  */
1220 static void
1221 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1222 {
1223
1224         G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1225         bioq_insert_head(&sc->sc_regular_delayed, bp);
1226 }
1227
1228 /*
1229  * Puts synchronization request onto delayed queue.
1230  */
1231 static void
1232 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1233 {
1234
1235         G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1236         bioq_insert_tail(&sc->sc_sync_delayed, bp);
1237 }
1238
1239 /*
1240  * Releases delayed regular requests which don't collide anymore with sync
1241  * requests.
1242  */
1243 static void
1244 g_mirror_regular_release(struct g_mirror_softc *sc)
1245 {
1246         struct bio *bp, *bp2;
1247
1248         TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1249                 if (g_mirror_sync_collision(sc, bp))
1250                         continue;
1251                 bioq_remove(&sc->sc_regular_delayed, bp);
1252                 G_MIRROR_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1253                 mtx_lock(&sc->sc_queue_mtx);
1254                 bioq_insert_head(&sc->sc_queue, bp);
1255 #if 0
1256                 /*
1257                  * wakeup() is not needed, because this function is called from
1258                  * the worker thread.
1259                  */
1260                 wakeup(&sc->sc_queue);
1261 #endif
1262                 mtx_unlock(&sc->sc_queue_mtx);
1263         }
1264 }
1265
1266 /*
1267  * Releases delayed sync requests which don't collide anymore with regular
1268  * requests.
1269  */
1270 static void
1271 g_mirror_sync_release(struct g_mirror_softc *sc)
1272 {
1273         struct bio *bp, *bp2;
1274
1275         TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1276                 if (g_mirror_regular_collision(sc, bp))
1277                         continue;
1278                 bioq_remove(&sc->sc_sync_delayed, bp);
1279                 G_MIRROR_LOGREQ(2, bp,
1280                     "Releasing delayed synchronization request.");
1281                 g_io_request(bp, bp->bio_from);
1282         }
1283 }
1284
1285 /*
1286  * Handle synchronization requests.
1287  * Every synchronization request is two-steps process: first, READ request is
1288  * send to active provider and then WRITE request (with read data) to the provider
1289  * being synchronized. When WRITE is finished, new synchronization request is
1290  * send.
1291  */
1292 static void
1293 g_mirror_sync_request(struct bio *bp)
1294 {
1295         struct g_mirror_softc *sc;
1296         struct g_mirror_disk *disk;
1297
1298         bp->bio_from->index--;
1299         sc = bp->bio_from->geom->softc;
1300         disk = bp->bio_from->private;
1301         if (disk == NULL) {
1302                 sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1303                 g_topology_lock();
1304                 g_mirror_kill_consumer(sc, bp->bio_from);
1305                 g_topology_unlock();
1306                 free(bp->bio_data, M_MIRROR);
1307                 g_destroy_bio(bp);
1308                 sx_xlock(&sc->sc_lock);
1309                 return;
1310         }
1311
1312         /*
1313          * Synchronization request.
1314          */
1315         switch (bp->bio_cmd) {
1316         case BIO_READ:
1317             {
1318                 struct g_consumer *cp;
1319
1320                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_sync_request_read,
1321                     bp->bio_error);
1322
1323                 if (bp->bio_error != 0) {
1324                         G_MIRROR_LOGREQ(0, bp,
1325                             "Synchronization request failed (error=%d).",
1326                             bp->bio_error);
1327                         g_destroy_bio(bp);
1328                         return;
1329                 }
1330                 G_MIRROR_LOGREQ(3, bp,
1331                     "Synchronization request half-finished.");
1332                 bp->bio_cmd = BIO_WRITE;
1333                 bp->bio_cflags = 0;
1334                 cp = disk->d_consumer;
1335                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1336                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1337                     cp->acr, cp->acw, cp->ace));
1338                 cp->index++;
1339                 g_io_request(bp, cp);
1340                 return;
1341             }
1342         case BIO_WRITE:
1343             {
1344                 struct g_mirror_disk_sync *sync;
1345                 off_t offset;
1346                 void *data;
1347                 int i;
1348
1349                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_sync_request_write,
1350                     bp->bio_error);
1351
1352                 if (bp->bio_error != 0) {
1353                         G_MIRROR_LOGREQ(0, bp,
1354                             "Synchronization request failed (error=%d).",
1355                             bp->bio_error);
1356                         g_destroy_bio(bp);
1357                         sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1358                         g_mirror_event_send(disk,
1359                             G_MIRROR_DISK_STATE_DISCONNECTED,
1360                             G_MIRROR_EVENT_DONTWAIT);
1361                         return;
1362                 }
1363                 G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1364                 sync = &disk->d_sync;
1365                 if (sync->ds_offset >= sc->sc_mediasize ||
1366                     sync->ds_consumer == NULL ||
1367                     (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1368                         /* Don't send more synchronization requests. */
1369                         sync->ds_inflight--;
1370                         if (sync->ds_bios != NULL) {
1371                                 i = (int)(uintptr_t)bp->bio_caller1;
1372                                 sync->ds_bios[i] = NULL;
1373                         }
1374                         free(bp->bio_data, M_MIRROR);
1375                         g_destroy_bio(bp);
1376                         if (sync->ds_inflight > 0)
1377                                 return;
1378                         if (sync->ds_consumer == NULL ||
1379                             (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1380                                 return;
1381                         }
1382                         /* Disk up-to-date, activate it. */
1383                         g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1384                             G_MIRROR_EVENT_DONTWAIT);
1385                         return;
1386                 }
1387
1388                 /* Send next synchronization request. */
1389                 data = bp->bio_data;
1390                 g_reset_bio(bp);
1391                 bp->bio_cmd = BIO_READ;
1392                 bp->bio_offset = sync->ds_offset;
1393                 bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1394                 sync->ds_offset += bp->bio_length;
1395                 bp->bio_done = g_mirror_sync_done;
1396                 bp->bio_data = data;
1397                 bp->bio_from = sync->ds_consumer;
1398                 bp->bio_to = sc->sc_provider;
1399                 G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1400                 sync->ds_consumer->index++;
1401                 /*
1402                  * Delay the request if it is colliding with a regular request.
1403                  */
1404                 if (g_mirror_regular_collision(sc, bp))
1405                         g_mirror_sync_delay(sc, bp);
1406                 else
1407                         g_io_request(bp, sync->ds_consumer);
1408
1409                 /* Release delayed requests if possible. */
1410                 g_mirror_regular_release(sc);
1411
1412                 /* Find the smallest offset */
1413                 offset = sc->sc_mediasize;
1414                 for (i = 0; i < g_mirror_syncreqs; i++) {
1415                         bp = sync->ds_bios[i];
1416                         if (bp->bio_offset < offset)
1417                                 offset = bp->bio_offset;
1418                 }
1419                 if (sync->ds_offset_done + (MAXPHYS * 100) < offset) {
1420                         /* Update offset_done on every 100 blocks. */
1421                         sync->ds_offset_done = offset;
1422                         g_mirror_update_metadata(disk);
1423                 }
1424                 return;
1425             }
1426         default:
1427                 KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1428                     bp->bio_cmd, sc->sc_name));
1429                 break;
1430         }
1431 }
1432
1433 static void
1434 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1435 {
1436         struct g_mirror_disk *disk;
1437         struct g_consumer *cp;
1438         struct bio *cbp;
1439
1440         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1441                 if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1442                         break;
1443         }
1444         if (disk == NULL) {
1445                 if (bp->bio_error == 0)
1446                         bp->bio_error = ENXIO;
1447                 g_io_deliver(bp, bp->bio_error);
1448                 return;
1449         }
1450         cbp = g_clone_bio(bp);
1451         if (cbp == NULL) {
1452                 if (bp->bio_error == 0)
1453                         bp->bio_error = ENOMEM;
1454                 g_io_deliver(bp, bp->bio_error);
1455                 return;
1456         }
1457         /*
1458          * Fill in the component buf structure.
1459          */
1460         cp = disk->d_consumer;
1461         cbp->bio_done = g_mirror_done;
1462         cbp->bio_to = cp->provider;
1463         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1464         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1465             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1466             cp->acw, cp->ace));
1467         cp->index++;
1468         g_io_request(cbp, cp);
1469 }
1470
1471 static void
1472 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1473 {
1474         struct g_mirror_disk *disk;
1475         struct g_consumer *cp;
1476         struct bio *cbp;
1477
1478         disk = g_mirror_get_disk(sc);
1479         if (disk == NULL) {
1480                 if (bp->bio_error == 0)
1481                         bp->bio_error = ENXIO;
1482                 g_io_deliver(bp, bp->bio_error);
1483                 return;
1484         }
1485         cbp = g_clone_bio(bp);
1486         if (cbp == NULL) {
1487                 if (bp->bio_error == 0)
1488                         bp->bio_error = ENOMEM;
1489                 g_io_deliver(bp, bp->bio_error);
1490                 return;
1491         }
1492         /*
1493          * Fill in the component buf structure.
1494          */
1495         cp = disk->d_consumer;
1496         cbp->bio_done = g_mirror_done;
1497         cbp->bio_to = cp->provider;
1498         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1499         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1500             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1501             cp->acw, cp->ace));
1502         cp->index++;
1503         g_io_request(cbp, cp);
1504 }
1505
1506 #define TRACK_SIZE  (1 * 1024 * 1024)
1507 #define LOAD_SCALE      256
1508 #define ABS(x)          (((x) >= 0) ? (x) : (-(x)))
1509
1510 static void
1511 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1512 {
1513         struct g_mirror_disk *disk, *dp;
1514         struct g_consumer *cp;
1515         struct bio *cbp;
1516         int prio, best;
1517
1518         /* Find a disk with the smallest load. */
1519         disk = NULL;
1520         best = INT_MAX;
1521         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1522                 if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1523                         continue;
1524                 prio = dp->load;
1525                 /* If disk head is precisely in position - highly prefer it. */
1526                 if (dp->d_last_offset == bp->bio_offset)
1527                         prio -= 2 * LOAD_SCALE;
1528                 else
1529                 /* If disk head is close to position - prefer it. */
1530                 if (ABS(dp->d_last_offset - bp->bio_offset) < TRACK_SIZE)
1531                         prio -= 1 * LOAD_SCALE;
1532                 if (prio <= best) {
1533                         disk = dp;
1534                         best = prio;
1535                 }
1536         }
1537         KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1538         cbp = g_clone_bio(bp);
1539         if (cbp == NULL) {
1540                 if (bp->bio_error == 0)
1541                         bp->bio_error = ENOMEM;
1542                 g_io_deliver(bp, bp->bio_error);
1543                 return;
1544         }
1545         /*
1546          * Fill in the component buf structure.
1547          */
1548         cp = disk->d_consumer;
1549         cbp->bio_done = g_mirror_done;
1550         cbp->bio_to = cp->provider;
1551         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1552         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1553             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1554             cp->acw, cp->ace));
1555         cp->index++;
1556         /* Remember last head position */
1557         disk->d_last_offset = bp->bio_offset + bp->bio_length;
1558         /* Update loads. */
1559         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1560                 dp->load = (dp->d_consumer->index * LOAD_SCALE +
1561                     dp->load * 7) / 8;
1562         }
1563         g_io_request(cbp, cp);
1564 }
1565
1566 static void
1567 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1568 {
1569         struct bio_queue_head queue;
1570         struct g_mirror_disk *disk;
1571         struct g_consumer *cp;
1572         struct bio *cbp;
1573         off_t left, mod, offset, slice;
1574         u_char *data;
1575         u_int ndisks;
1576
1577         if (bp->bio_length <= sc->sc_slice) {
1578                 g_mirror_request_round_robin(sc, bp);
1579                 return;
1580         }
1581         ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1582         slice = bp->bio_length / ndisks;
1583         mod = slice % sc->sc_provider->sectorsize;
1584         if (mod != 0)
1585                 slice += sc->sc_provider->sectorsize - mod;
1586         /*
1587          * Allocate all bios before sending any request, so we can
1588          * return ENOMEM in nice and clean way.
1589          */
1590         left = bp->bio_length;
1591         offset = bp->bio_offset;
1592         data = bp->bio_data;
1593         bioq_init(&queue);
1594         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1595                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1596                         continue;
1597                 cbp = g_clone_bio(bp);
1598                 if (cbp == NULL) {
1599                         while ((cbp = bioq_takefirst(&queue)) != NULL)
1600                                 g_destroy_bio(cbp);
1601                         if (bp->bio_error == 0)
1602                                 bp->bio_error = ENOMEM;
1603                         g_io_deliver(bp, bp->bio_error);
1604                         return;
1605                 }
1606                 bioq_insert_tail(&queue, cbp);
1607                 cbp->bio_done = g_mirror_done;
1608                 cbp->bio_caller1 = disk;
1609                 cbp->bio_to = disk->d_consumer->provider;
1610                 cbp->bio_offset = offset;
1611                 cbp->bio_data = data;
1612                 cbp->bio_length = MIN(left, slice);
1613                 left -= cbp->bio_length;
1614                 if (left == 0)
1615                         break;
1616                 offset += cbp->bio_length;
1617                 data += cbp->bio_length;
1618         }
1619         while ((cbp = bioq_takefirst(&queue)) != NULL) {
1620                 G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1621                 disk = cbp->bio_caller1;
1622                 cbp->bio_caller1 = NULL;
1623                 cp = disk->d_consumer;
1624                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1625                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1626                     cp->acr, cp->acw, cp->ace));
1627                 disk->d_consumer->index++;
1628                 g_io_request(cbp, disk->d_consumer);
1629         }
1630 }
1631
1632 static void
1633 g_mirror_register_request(struct bio *bp)
1634 {
1635         struct g_mirror_softc *sc;
1636
1637         sc = bp->bio_to->geom->softc;
1638         switch (bp->bio_cmd) {
1639         case BIO_READ:
1640                 switch (sc->sc_balance) {
1641                 case G_MIRROR_BALANCE_LOAD:
1642                         g_mirror_request_load(sc, bp);
1643                         break;
1644                 case G_MIRROR_BALANCE_PREFER:
1645                         g_mirror_request_prefer(sc, bp);
1646                         break;
1647                 case G_MIRROR_BALANCE_ROUND_ROBIN:
1648                         g_mirror_request_round_robin(sc, bp);
1649                         break;
1650                 case G_MIRROR_BALANCE_SPLIT:
1651                         g_mirror_request_split(sc, bp);
1652                         break;
1653                 }
1654                 return;
1655         case BIO_WRITE:
1656         case BIO_DELETE:
1657             {
1658                 struct g_mirror_disk *disk;
1659                 struct g_mirror_disk_sync *sync;
1660                 struct bio_queue_head queue;
1661                 struct g_consumer *cp;
1662                 struct bio *cbp;
1663
1664                 /*
1665                  * Delay the request if it is colliding with a synchronization
1666                  * request.
1667                  */
1668                 if (g_mirror_sync_collision(sc, bp)) {
1669                         g_mirror_regular_delay(sc, bp);
1670                         return;
1671                 }
1672
1673                 if (sc->sc_idle)
1674                         g_mirror_unidle(sc);
1675                 else
1676                         sc->sc_last_write = time_uptime;
1677
1678                 /*
1679                  * Allocate all bios before sending any request, so we can
1680                  * return ENOMEM in nice and clean way.
1681                  */
1682                 bioq_init(&queue);
1683                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1684                         sync = &disk->d_sync;
1685                         switch (disk->d_state) {
1686                         case G_MIRROR_DISK_STATE_ACTIVE:
1687                                 break;
1688                         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1689                                 if (bp->bio_offset >= sync->ds_offset)
1690                                         continue;
1691                                 break;
1692                         default:
1693                                 continue;
1694                         }
1695                         if (bp->bio_cmd == BIO_DELETE &&
1696                             (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE) == 0)
1697                                 continue;
1698                         cbp = g_clone_bio(bp);
1699                         if (cbp == NULL) {
1700                                 while ((cbp = bioq_takefirst(&queue)) != NULL)
1701                                         g_destroy_bio(cbp);
1702                                 if (bp->bio_error == 0)
1703                                         bp->bio_error = ENOMEM;
1704                                 g_io_deliver(bp, bp->bio_error);
1705                                 return;
1706                         }
1707                         bioq_insert_tail(&queue, cbp);
1708                         cbp->bio_done = g_mirror_done;
1709                         cp = disk->d_consumer;
1710                         cbp->bio_caller1 = cp;
1711                         cbp->bio_to = cp->provider;
1712                         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1713                             ("Consumer %s not opened (r%dw%de%d).",
1714                             cp->provider->name, cp->acr, cp->acw, cp->ace));
1715                 }
1716                 if (bioq_first(&queue) == NULL) {
1717                         g_io_deliver(bp, EOPNOTSUPP);
1718                         return;
1719                 }
1720                 while ((cbp = bioq_takefirst(&queue)) != NULL) {
1721                         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1722                         cp = cbp->bio_caller1;
1723                         cbp->bio_caller1 = NULL;
1724                         cp->index++;
1725                         sc->sc_writes++;
1726                         g_io_request(cbp, cp);
1727                 }
1728                 /*
1729                  * Put request onto inflight queue, so we can check if new
1730                  * synchronization requests don't collide with it.
1731                  */
1732                 bioq_insert_tail(&sc->sc_inflight, bp);
1733                 /*
1734                  * Bump syncid on first write.
1735                  */
1736                 if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1737                         sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1738                         g_mirror_bump_syncid(sc);
1739                 }
1740                 return;
1741             }
1742         default:
1743                 KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1744                     bp->bio_cmd, sc->sc_name));
1745                 break;
1746         }
1747 }
1748
1749 static int
1750 g_mirror_can_destroy(struct g_mirror_softc *sc)
1751 {
1752         struct g_geom *gp;
1753         struct g_consumer *cp;
1754
1755         g_topology_assert();
1756         gp = sc->sc_geom;
1757         if (gp->softc == NULL)
1758                 return (1);
1759         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_TASTING) != 0)
1760                 return (0);
1761         LIST_FOREACH(cp, &gp->consumer, consumer) {
1762                 if (g_mirror_is_busy(sc, cp))
1763                         return (0);
1764         }
1765         gp = sc->sc_sync.ds_geom;
1766         LIST_FOREACH(cp, &gp->consumer, consumer) {
1767                 if (g_mirror_is_busy(sc, cp))
1768                         return (0);
1769         }
1770         G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1771             sc->sc_name);
1772         return (1);
1773 }
1774
1775 static int
1776 g_mirror_try_destroy(struct g_mirror_softc *sc)
1777 {
1778
1779         if (sc->sc_rootmount != NULL) {
1780                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1781                     sc->sc_rootmount);
1782                 root_mount_rel(sc->sc_rootmount);
1783                 sc->sc_rootmount = NULL;
1784         }
1785         g_topology_lock();
1786         if (!g_mirror_can_destroy(sc)) {
1787                 g_topology_unlock();
1788                 return (0);
1789         }
1790         sc->sc_geom->softc = NULL;
1791         sc->sc_sync.ds_geom->softc = NULL;
1792         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WAIT) != 0) {
1793                 g_topology_unlock();
1794                 G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1795                     &sc->sc_worker);
1796                 /* Unlock sc_lock here, as it can be destroyed after wakeup. */
1797                 sx_xunlock(&sc->sc_lock);
1798                 wakeup(&sc->sc_worker);
1799                 sc->sc_worker = NULL;
1800         } else {
1801                 g_topology_unlock();
1802                 g_mirror_destroy_device(sc);
1803                 free(sc, M_MIRROR);
1804         }
1805         return (1);
1806 }
1807
1808 /*
1809  * Worker thread.
1810  */
1811 static void
1812 g_mirror_worker(void *arg)
1813 {
1814         struct g_mirror_softc *sc;
1815         struct g_mirror_event *ep;
1816         struct bio *bp;
1817         int timeout;
1818
1819         sc = arg;
1820         thread_lock(curthread);
1821         sched_prio(curthread, PRIBIO);
1822         thread_unlock(curthread);
1823
1824         sx_xlock(&sc->sc_lock);
1825         for (;;) {
1826                 G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1827                 /*
1828                  * First take a look at events.
1829                  * This is important to handle events before any I/O requests.
1830                  */
1831                 ep = g_mirror_event_get(sc);
1832                 if (ep != NULL) {
1833                         g_mirror_event_remove(sc, ep);
1834                         if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1835                                 /* Update only device status. */
1836                                 G_MIRROR_DEBUG(3,
1837                                     "Running event for device %s.",
1838                                     sc->sc_name);
1839                                 ep->e_error = 0;
1840                                 g_mirror_update_device(sc, 1);
1841                         } else {
1842                                 /* Update disk status. */
1843                                 G_MIRROR_DEBUG(3, "Running event for disk %s.",
1844                                      g_mirror_get_diskname(ep->e_disk));
1845                                 ep->e_error = g_mirror_update_disk(ep->e_disk,
1846                                     ep->e_state);
1847                                 if (ep->e_error == 0)
1848                                         g_mirror_update_device(sc, 0);
1849                         }
1850                         if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1851                                 KASSERT(ep->e_error == 0,
1852                                     ("Error cannot be handled."));
1853                                 g_mirror_event_free(ep);
1854                         } else {
1855                                 ep->e_flags |= G_MIRROR_EVENT_DONE;
1856                                 G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1857                                     ep);
1858                                 mtx_lock(&sc->sc_events_mtx);
1859                                 wakeup(ep);
1860                                 mtx_unlock(&sc->sc_events_mtx);
1861                         }
1862                         if ((sc->sc_flags &
1863                             G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1864                                 if (g_mirror_try_destroy(sc)) {
1865                                         curthread->td_pflags &= ~TDP_GEOM;
1866                                         G_MIRROR_DEBUG(1, "Thread exiting.");
1867                                         kproc_exit(0);
1868                                 }
1869                         }
1870                         G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1871                         continue;
1872                 }
1873                 /*
1874                  * Check if we can mark array as CLEAN and if we can't take
1875                  * how much seconds should we wait.
1876                  */
1877                 timeout = g_mirror_idle(sc, -1);
1878                 /*
1879                  * Now I/O requests.
1880                  */
1881                 /* Get first request from the queue. */
1882                 mtx_lock(&sc->sc_queue_mtx);
1883                 bp = bioq_takefirst(&sc->sc_queue);
1884                 if (bp == NULL) {
1885                         if ((sc->sc_flags &
1886                             G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1887                                 mtx_unlock(&sc->sc_queue_mtx);
1888                                 if (g_mirror_try_destroy(sc)) {
1889                                         curthread->td_pflags &= ~TDP_GEOM;
1890                                         G_MIRROR_DEBUG(1, "Thread exiting.");
1891                                         kproc_exit(0);
1892                                 }
1893                                 mtx_lock(&sc->sc_queue_mtx);
1894                         }
1895                         sx_xunlock(&sc->sc_lock);
1896                         /*
1897                          * XXX: We can miss an event here, because an event
1898                          *      can be added without sx-device-lock and without
1899                          *      mtx-queue-lock. Maybe I should just stop using
1900                          *      dedicated mutex for events synchronization and
1901                          *      stick with the queue lock?
1902                          *      The event will hang here until next I/O request
1903                          *      or next event is received.
1904                          */
1905                         MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
1906                             timeout * hz);
1907                         sx_xlock(&sc->sc_lock);
1908                         G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
1909                         continue;
1910                 }
1911                 mtx_unlock(&sc->sc_queue_mtx);
1912
1913                 if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
1914                     (bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0) {
1915                         g_mirror_sync_request(bp);      /* READ */
1916                 } else if (bp->bio_to != sc->sc_provider) {
1917                         if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
1918                                 g_mirror_regular_request(bp);
1919                         else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
1920                                 g_mirror_sync_request(bp);      /* WRITE */
1921                         else {
1922                                 KASSERT(0,
1923                                     ("Invalid request cflags=0x%hx to=%s.",
1924                                     bp->bio_cflags, bp->bio_to->name));
1925                         }
1926                 } else {
1927                         g_mirror_register_request(bp);
1928                 }
1929                 G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
1930         }
1931 }
1932
1933 static void
1934 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
1935 {
1936
1937         sx_assert(&sc->sc_lock, SX_LOCKED);
1938
1939         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
1940                 return;
1941         if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
1942                 G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as dirty.",
1943                     g_mirror_get_diskname(disk), sc->sc_name);
1944                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1945         } else if (sc->sc_idle &&
1946             (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
1947                 G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as clean.",
1948                     g_mirror_get_diskname(disk), sc->sc_name);
1949                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
1950         }
1951 }
1952
1953 static void
1954 g_mirror_sync_start(struct g_mirror_disk *disk)
1955 {
1956         struct g_mirror_softc *sc;
1957         struct g_consumer *cp;
1958         struct bio *bp;
1959         int error, i;
1960
1961         g_topology_assert_not();
1962         sc = disk->d_softc;
1963         sx_assert(&sc->sc_lock, SX_LOCKED);
1964
1965         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1966             ("Disk %s is not marked for synchronization.",
1967             g_mirror_get_diskname(disk)));
1968         KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1969             ("Device not in RUNNING state (%s, %u).", sc->sc_name,
1970             sc->sc_state));
1971
1972         sx_xunlock(&sc->sc_lock);
1973         g_topology_lock();
1974         cp = g_new_consumer(sc->sc_sync.ds_geom);
1975         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
1976         error = g_attach(cp, sc->sc_provider);
1977         KASSERT(error == 0,
1978             ("Cannot attach to %s (error=%d).", sc->sc_name, error));
1979         error = g_access(cp, 1, 0, 0);
1980         KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
1981         g_topology_unlock();
1982         sx_xlock(&sc->sc_lock);
1983
1984         G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
1985             g_mirror_get_diskname(disk));
1986         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) == 0)
1987                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1988         KASSERT(disk->d_sync.ds_consumer == NULL,
1989             ("Sync consumer already exists (device=%s, disk=%s).",
1990             sc->sc_name, g_mirror_get_diskname(disk)));
1991
1992         disk->d_sync.ds_consumer = cp;
1993         disk->d_sync.ds_consumer->private = disk;
1994         disk->d_sync.ds_consumer->index = 0;
1995
1996         /*
1997          * Allocate memory for synchronization bios and initialize them.
1998          */
1999         disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
2000             M_MIRROR, M_WAITOK);
2001         for (i = 0; i < g_mirror_syncreqs; i++) {
2002                 bp = g_alloc_bio();
2003                 disk->d_sync.ds_bios[i] = bp;
2004                 bp->bio_parent = NULL;
2005                 bp->bio_cmd = BIO_READ;
2006                 bp->bio_data = malloc(MAXPHYS, M_MIRROR, M_WAITOK);
2007                 bp->bio_cflags = 0;
2008                 bp->bio_offset = disk->d_sync.ds_offset;
2009                 bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
2010                 disk->d_sync.ds_offset += bp->bio_length;
2011                 bp->bio_done = g_mirror_sync_done;
2012                 bp->bio_from = disk->d_sync.ds_consumer;
2013                 bp->bio_to = sc->sc_provider;
2014                 bp->bio_caller1 = (void *)(uintptr_t)i;
2015         }
2016
2017         /* Increase the number of disks in SYNCHRONIZING state. */
2018         sc->sc_sync.ds_ndisks++;
2019         /* Set the number of in-flight synchronization requests. */
2020         disk->d_sync.ds_inflight = g_mirror_syncreqs;
2021
2022         /*
2023          * Fire off first synchronization requests.
2024          */
2025         for (i = 0; i < g_mirror_syncreqs; i++) {
2026                 bp = disk->d_sync.ds_bios[i];
2027                 G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
2028                 disk->d_sync.ds_consumer->index++;
2029                 /*
2030                  * Delay the request if it is colliding with a regular request.
2031                  */
2032                 if (g_mirror_regular_collision(sc, bp))
2033                         g_mirror_sync_delay(sc, bp);
2034                 else
2035                         g_io_request(bp, disk->d_sync.ds_consumer);
2036         }
2037 }
2038
2039 /*
2040  * Stop synchronization process.
2041  * type: 0 - synchronization finished
2042  *       1 - synchronization stopped
2043  */
2044 static void
2045 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
2046 {
2047         struct g_mirror_softc *sc;
2048         struct g_consumer *cp;
2049
2050         g_topology_assert_not();
2051         sc = disk->d_softc;
2052         sx_assert(&sc->sc_lock, SX_LOCKED);
2053
2054         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2055             ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2056             g_mirror_disk_state2str(disk->d_state)));
2057         if (disk->d_sync.ds_consumer == NULL)
2058                 return;
2059
2060         if (type == 0) {
2061                 G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2062                     sc->sc_name, g_mirror_get_diskname(disk));
2063         } else /* if (type == 1) */ {
2064                 G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2065                     sc->sc_name, g_mirror_get_diskname(disk));
2066         }
2067         free(disk->d_sync.ds_bios, M_MIRROR);
2068         disk->d_sync.ds_bios = NULL;
2069         cp = disk->d_sync.ds_consumer;
2070         disk->d_sync.ds_consumer = NULL;
2071         disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2072         sc->sc_sync.ds_ndisks--;
2073         sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2074         g_topology_lock();
2075         g_mirror_kill_consumer(sc, cp);
2076         g_topology_unlock();
2077         sx_xlock(&sc->sc_lock);
2078 }
2079
2080 static void
2081 g_mirror_launch_provider(struct g_mirror_softc *sc)
2082 {
2083         struct g_mirror_disk *disk;
2084         struct g_provider *pp, *dp;
2085
2086         sx_assert(&sc->sc_lock, SX_LOCKED);
2087
2088         g_topology_lock();
2089         pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
2090         pp->flags |= G_PF_DIRECT_RECEIVE;
2091         pp->mediasize = sc->sc_mediasize;
2092         pp->sectorsize = sc->sc_sectorsize;
2093         pp->stripesize = 0;
2094         pp->stripeoffset = 0;
2095
2096         /* Splitting of unmapped BIO's could work but isn't implemented now */
2097         if (sc->sc_balance != G_MIRROR_BALANCE_SPLIT)
2098                 pp->flags |= G_PF_ACCEPT_UNMAPPED;
2099
2100         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2101                 if (disk->d_consumer && disk->d_consumer->provider) {
2102                         dp = disk->d_consumer->provider;
2103                         if (dp->stripesize > pp->stripesize) {
2104                                 pp->stripesize = dp->stripesize;
2105                                 pp->stripeoffset = dp->stripeoffset;
2106                         }
2107                         /* A provider underneath us doesn't support unmapped */
2108                         if ((dp->flags & G_PF_ACCEPT_UNMAPPED) == 0) {
2109                                 G_MIRROR_DEBUG(0, "Cancelling unmapped "
2110                                     "because of %s.", dp->name);
2111                                 pp->flags &= ~G_PF_ACCEPT_UNMAPPED;
2112                         }
2113                 }
2114         }
2115         sc->sc_provider = pp;
2116         g_error_provider(pp, 0);
2117         g_topology_unlock();
2118         G_MIRROR_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2119             g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE), sc->sc_ndisks);
2120         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2121                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2122                         g_mirror_sync_start(disk);
2123         }
2124 }
2125
2126 static void
2127 g_mirror_destroy_provider(struct g_mirror_softc *sc)
2128 {
2129         struct g_mirror_disk *disk;
2130         struct bio *bp;
2131
2132         g_topology_assert_not();
2133         KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2134             sc->sc_name));
2135
2136         g_topology_lock();
2137         g_error_provider(sc->sc_provider, ENXIO);
2138         mtx_lock(&sc->sc_queue_mtx);
2139         while ((bp = bioq_takefirst(&sc->sc_queue)) != NULL) {
2140                 /*
2141                  * Abort any pending I/O that wasn't generated by us.
2142                  * Synchronization requests and requests destined for individual
2143                  * mirror components can be destroyed immediately.
2144                  */
2145                 if (bp->bio_to == sc->sc_provider &&
2146                     bp->bio_from->geom != sc->sc_sync.ds_geom) {
2147                         g_io_deliver(bp, ENXIO);
2148                 } else {
2149                         if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
2150                                 free(bp->bio_data, M_MIRROR);
2151                         g_destroy_bio(bp);
2152                 }
2153         }
2154         mtx_unlock(&sc->sc_queue_mtx);
2155         G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2156             sc->sc_provider->name);
2157         sc->sc_provider->flags |= G_PF_WITHER;
2158         g_orphan_provider(sc->sc_provider, ENXIO);
2159         g_topology_unlock();
2160         sc->sc_provider = NULL;
2161         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2162                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2163                         g_mirror_sync_stop(disk, 1);
2164         }
2165 }
2166
2167 static void
2168 g_mirror_go(void *arg)
2169 {
2170         struct g_mirror_softc *sc;
2171
2172         sc = arg;
2173         G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2174         g_mirror_event_send(sc, 0,
2175             G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2176 }
2177
2178 static u_int
2179 g_mirror_determine_state(struct g_mirror_disk *disk)
2180 {
2181         struct g_mirror_softc *sc;
2182         u_int state;
2183
2184         sc = disk->d_softc;
2185         if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2186                 if ((disk->d_flags &
2187                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2188                         /* Disk does not need synchronization. */
2189                         state = G_MIRROR_DISK_STATE_ACTIVE;
2190                 } else {
2191                         if ((sc->sc_flags &
2192                              G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2193                             (disk->d_flags &
2194                              G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2195                                 /*
2196                                  * We can start synchronization from
2197                                  * the stored offset.
2198                                  */
2199                                 state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2200                         } else {
2201                                 state = G_MIRROR_DISK_STATE_STALE;
2202                         }
2203                 }
2204         } else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2205                 /*
2206                  * Reset all synchronization data for this disk,
2207                  * because if it even was synchronized, it was
2208                  * synchronized to disks with different syncid.
2209                  */
2210                 disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2211                 disk->d_sync.ds_offset = 0;
2212                 disk->d_sync.ds_offset_done = 0;
2213                 disk->d_sync.ds_syncid = sc->sc_syncid;
2214                 if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2215                     (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2216                         state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2217                 } else {
2218                         state = G_MIRROR_DISK_STATE_STALE;
2219                 }
2220         } else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2221                 /*
2222                  * Not good, NOT GOOD!
2223                  * It means that mirror was started on stale disks
2224                  * and more fresh disk just arrive.
2225                  * If there were writes, mirror is broken, sorry.
2226                  * I think the best choice here is don't touch
2227                  * this disk and inform the user loudly.
2228                  */
2229                 G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2230                     "disk (%s) arrives!! It will not be connected to the "
2231                     "running device.", sc->sc_name,
2232                     g_mirror_get_diskname(disk));
2233                 g_mirror_destroy_disk(disk);
2234                 state = G_MIRROR_DISK_STATE_NONE;
2235                 /* Return immediately, because disk was destroyed. */
2236                 return (state);
2237         }
2238         G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2239             g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2240         return (state);
2241 }
2242
2243 /*
2244  * Update device state.
2245  */
2246 static void
2247 g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force)
2248 {
2249         struct g_mirror_disk *disk;
2250         u_int state;
2251
2252         sx_assert(&sc->sc_lock, SX_XLOCKED);
2253
2254         switch (sc->sc_state) {
2255         case G_MIRROR_DEVICE_STATE_STARTING:
2256             {
2257                 struct g_mirror_disk *pdisk, *tdisk;
2258                 u_int dirty, ndisks, genid, syncid;
2259
2260                 KASSERT(sc->sc_provider == NULL,
2261                     ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2262                 /*
2263                  * Are we ready? We are, if all disks are connected or
2264                  * if we have any disks and 'force' is true.
2265                  */
2266                 ndisks = g_mirror_ndisks(sc, -1);
2267                 if (sc->sc_ndisks == ndisks || (force && ndisks > 0)) {
2268                         ;
2269                 } else if (ndisks == 0) {
2270                         /*
2271                          * Disks went down in starting phase, so destroy
2272                          * device.
2273                          */
2274                         callout_drain(&sc->sc_callout);
2275                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2276                         G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2277                             sc->sc_rootmount);
2278                         root_mount_rel(sc->sc_rootmount);
2279                         sc->sc_rootmount = NULL;
2280                         return;
2281                 } else {
2282                         return;
2283                 }
2284
2285                 /*
2286                  * Activate all disks with the biggest syncid.
2287                  */
2288                 if (force) {
2289                         /*
2290                          * If 'force' is true, we have been called due to
2291                          * timeout, so don't bother canceling timeout.
2292                          */
2293                         ndisks = 0;
2294                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2295                                 if ((disk->d_flags &
2296                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2297                                         ndisks++;
2298                                 }
2299                         }
2300                         if (ndisks == 0) {
2301                                 /* No valid disks found, destroy device. */
2302                                 sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2303                                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2304                                     __LINE__, sc->sc_rootmount);
2305                                 root_mount_rel(sc->sc_rootmount);
2306                                 sc->sc_rootmount = NULL;
2307                                 return;
2308                         }
2309                 } else {
2310                         /* Cancel timeout. */
2311                         callout_drain(&sc->sc_callout);
2312                 }
2313
2314                 /*
2315                  * Find the biggest genid.
2316                  */
2317                 genid = 0;
2318                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2319                         if (disk->d_genid > genid)
2320                                 genid = disk->d_genid;
2321                 }
2322                 sc->sc_genid = genid;
2323                 /*
2324                  * Remove all disks without the biggest genid.
2325                  */
2326                 LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2327                         if (disk->d_genid < genid) {
2328                                 G_MIRROR_DEBUG(0,
2329                                     "Component %s (device %s) broken, skipping.",
2330                                     g_mirror_get_diskname(disk), sc->sc_name);
2331                                 g_mirror_destroy_disk(disk);
2332                         }
2333                 }
2334
2335                 /*
2336                  * Find the biggest syncid.
2337                  */
2338                 syncid = 0;
2339                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2340                         if (disk->d_sync.ds_syncid > syncid)
2341                                 syncid = disk->d_sync.ds_syncid;
2342                 }
2343
2344                 /*
2345                  * Here we need to look for dirty disks and if all disks
2346                  * with the biggest syncid are dirty, we have to choose
2347                  * one with the biggest priority and rebuild the rest.
2348                  */
2349                 /*
2350                  * Find the number of dirty disks with the biggest syncid.
2351                  * Find the number of disks with the biggest syncid.
2352                  * While here, find a disk with the biggest priority.
2353                  */
2354                 dirty = ndisks = 0;
2355                 pdisk = NULL;
2356                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2357                         if (disk->d_sync.ds_syncid != syncid)
2358                                 continue;
2359                         if ((disk->d_flags &
2360                             G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2361                                 continue;
2362                         }
2363                         ndisks++;
2364                         if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2365                                 dirty++;
2366                                 if (pdisk == NULL ||
2367                                     pdisk->d_priority < disk->d_priority) {
2368                                         pdisk = disk;
2369                                 }
2370                         }
2371                 }
2372                 if (dirty == 0) {
2373                         /* No dirty disks at all, great. */
2374                 } else if (dirty == ndisks) {
2375                         /*
2376                          * Force synchronization for all dirty disks except one
2377                          * with the biggest priority.
2378                          */
2379                         KASSERT(pdisk != NULL, ("pdisk == NULL"));
2380                         G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2381                             "master disk for synchronization.",
2382                             g_mirror_get_diskname(pdisk), sc->sc_name);
2383                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2384                                 if (disk->d_sync.ds_syncid != syncid)
2385                                         continue;
2386                                 if ((disk->d_flags &
2387                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2388                                         continue;
2389                                 }
2390                                 KASSERT((disk->d_flags &
2391                                     G_MIRROR_DISK_FLAG_DIRTY) != 0,
2392                                     ("Disk %s isn't marked as dirty.",
2393                                     g_mirror_get_diskname(disk)));
2394                                 /* Skip the disk with the biggest priority. */
2395                                 if (disk == pdisk)
2396                                         continue;
2397                                 disk->d_sync.ds_syncid = 0;
2398                         }
2399                 } else if (dirty < ndisks) {
2400                         /*
2401                          * Force synchronization for all dirty disks.
2402                          * We have some non-dirty disks.
2403                          */
2404                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2405                                 if (disk->d_sync.ds_syncid != syncid)
2406                                         continue;
2407                                 if ((disk->d_flags &
2408                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2409                                         continue;
2410                                 }
2411                                 if ((disk->d_flags &
2412                                     G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2413                                         continue;
2414                                 }
2415                                 disk->d_sync.ds_syncid = 0;
2416                         }
2417                 }
2418
2419                 /* Reset hint. */
2420                 sc->sc_hint = NULL;
2421                 sc->sc_syncid = syncid;
2422                 if (force) {
2423                         /* Remember to bump syncid on first write. */
2424                         sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2425                 }
2426                 state = G_MIRROR_DEVICE_STATE_RUNNING;
2427                 G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2428                     sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2429                     g_mirror_device_state2str(state));
2430                 sc->sc_state = state;
2431                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2432                         state = g_mirror_determine_state(disk);
2433                         g_mirror_event_send(disk, state,
2434                             G_MIRROR_EVENT_DONTWAIT);
2435                         if (state == G_MIRROR_DISK_STATE_STALE)
2436                                 sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2437                 }
2438                 break;
2439             }
2440         case G_MIRROR_DEVICE_STATE_RUNNING:
2441                 if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2442                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2443                         /*
2444                          * No active disks or no disks at all,
2445                          * so destroy device.
2446                          */
2447                         if (sc->sc_provider != NULL)
2448                                 g_mirror_destroy_provider(sc);
2449                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2450                         break;
2451                 } else if (g_mirror_ndisks(sc,
2452                     G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2453                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2454                         /*
2455                          * We have active disks, launch provider if it doesn't
2456                          * exist.
2457                          */
2458                         if (sc->sc_provider == NULL)
2459                                 g_mirror_launch_provider(sc);
2460                         if (sc->sc_rootmount != NULL) {
2461                                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2462                                     __LINE__, sc->sc_rootmount);
2463                                 root_mount_rel(sc->sc_rootmount);
2464                                 sc->sc_rootmount = NULL;
2465                         }
2466                 }
2467                 /*
2468                  * Genid should be bumped immediately, so do it here.
2469                  */
2470                 if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2471                         sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2472                         g_mirror_bump_genid(sc);
2473                 }
2474                 break;
2475         default:
2476                 KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2477                     sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2478                 break;
2479         }
2480 }
2481
2482 /*
2483  * Update disk state and device state if needed.
2484  */
2485 #define DISK_STATE_CHANGED()    G_MIRROR_DEBUG(1,                       \
2486         "Disk %s state changed from %s to %s (device %s).",             \
2487         g_mirror_get_diskname(disk),                                    \
2488         g_mirror_disk_state2str(disk->d_state),                         \
2489         g_mirror_disk_state2str(state), sc->sc_name)
2490 static int
2491 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2492 {
2493         struct g_mirror_softc *sc;
2494
2495         sc = disk->d_softc;
2496         sx_assert(&sc->sc_lock, SX_XLOCKED);
2497
2498 again:
2499         G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2500             g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2501             g_mirror_disk_state2str(state));
2502         switch (state) {
2503         case G_MIRROR_DISK_STATE_NEW:
2504                 /*
2505                  * Possible scenarios:
2506                  * 1. New disk arrive.
2507                  */
2508                 /* Previous state should be NONE. */
2509                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2510                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2511                     g_mirror_disk_state2str(disk->d_state)));
2512                 DISK_STATE_CHANGED();
2513
2514                 disk->d_state = state;
2515                 if (LIST_EMPTY(&sc->sc_disks))
2516                         LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2517                 else {
2518                         struct g_mirror_disk *dp;
2519
2520                         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2521                                 if (disk->d_priority >= dp->d_priority) {
2522                                         LIST_INSERT_BEFORE(dp, disk, d_next);
2523                                         dp = NULL;
2524                                         break;
2525                                 }
2526                                 if (LIST_NEXT(dp, d_next) == NULL)
2527                                         break;
2528                         }
2529                         if (dp != NULL)
2530                                 LIST_INSERT_AFTER(dp, disk, d_next);
2531                 }
2532                 G_MIRROR_DEBUG(1, "Device %s: provider %s detected.",
2533                     sc->sc_name, g_mirror_get_diskname(disk));
2534                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2535                         break;
2536                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2537                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2538                     g_mirror_device_state2str(sc->sc_state),
2539                     g_mirror_get_diskname(disk),
2540                     g_mirror_disk_state2str(disk->d_state)));
2541                 state = g_mirror_determine_state(disk);
2542                 if (state != G_MIRROR_DISK_STATE_NONE)
2543                         goto again;
2544                 break;
2545         case G_MIRROR_DISK_STATE_ACTIVE:
2546                 /*
2547                  * Possible scenarios:
2548                  * 1. New disk does not need synchronization.
2549                  * 2. Synchronization process finished successfully.
2550                  */
2551                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2552                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2553                     g_mirror_device_state2str(sc->sc_state),
2554                     g_mirror_get_diskname(disk),
2555                     g_mirror_disk_state2str(disk->d_state)));
2556                 /* Previous state should be NEW or SYNCHRONIZING. */
2557                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2558                     disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2559                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2560                     g_mirror_disk_state2str(disk->d_state)));
2561                 DISK_STATE_CHANGED();
2562
2563                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2564                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2565                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2566                         g_mirror_sync_stop(disk, 0);
2567                 }
2568                 disk->d_state = state;
2569                 disk->d_sync.ds_offset = 0;
2570                 disk->d_sync.ds_offset_done = 0;
2571                 g_mirror_update_idle(sc, disk);
2572                 g_mirror_update_metadata(disk);
2573                 G_MIRROR_DEBUG(1, "Device %s: provider %s activated.",
2574                     sc->sc_name, g_mirror_get_diskname(disk));
2575                 break;
2576         case G_MIRROR_DISK_STATE_STALE:
2577                 /*
2578                  * Possible scenarios:
2579                  * 1. Stale disk was connected.
2580                  */
2581                 /* Previous state should be NEW. */
2582                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2583                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2584                     g_mirror_disk_state2str(disk->d_state)));
2585                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2586                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2587                     g_mirror_device_state2str(sc->sc_state),
2588                     g_mirror_get_diskname(disk),
2589                     g_mirror_disk_state2str(disk->d_state)));
2590                 /*
2591                  * STALE state is only possible if device is marked
2592                  * NOAUTOSYNC.
2593                  */
2594                 KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2595                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2596                     g_mirror_device_state2str(sc->sc_state),
2597                     g_mirror_get_diskname(disk),
2598                     g_mirror_disk_state2str(disk->d_state)));
2599                 DISK_STATE_CHANGED();
2600
2601                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2602                 disk->d_state = state;
2603                 g_mirror_update_metadata(disk);
2604                 G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2605                     sc->sc_name, g_mirror_get_diskname(disk));
2606                 break;
2607         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2608                 /*
2609                  * Possible scenarios:
2610                  * 1. Disk which needs synchronization was connected.
2611                  */
2612                 /* Previous state should be NEW. */
2613                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2614                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2615                     g_mirror_disk_state2str(disk->d_state)));
2616                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2617                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2618                     g_mirror_device_state2str(sc->sc_state),
2619                     g_mirror_get_diskname(disk),
2620                     g_mirror_disk_state2str(disk->d_state)));
2621                 DISK_STATE_CHANGED();
2622
2623                 if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2624                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2625                 disk->d_state = state;
2626                 if (sc->sc_provider != NULL) {
2627                         g_mirror_sync_start(disk);
2628                         g_mirror_update_metadata(disk);
2629                 }
2630                 break;
2631         case G_MIRROR_DISK_STATE_DISCONNECTED:
2632                 /*
2633                  * Possible scenarios:
2634                  * 1. Device wasn't running yet, but disk disappear.
2635                  * 2. Disk was active and disapppear.
2636                  * 3. Disk disappear during synchronization process.
2637                  */
2638                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2639                         /*
2640                          * Previous state should be ACTIVE, STALE or
2641                          * SYNCHRONIZING.
2642                          */
2643                         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2644                             disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2645                             disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2646                             ("Wrong disk state (%s, %s).",
2647                             g_mirror_get_diskname(disk),
2648                             g_mirror_disk_state2str(disk->d_state)));
2649                 } else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2650                         /* Previous state should be NEW. */
2651                         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2652                             ("Wrong disk state (%s, %s).",
2653                             g_mirror_get_diskname(disk),
2654                             g_mirror_disk_state2str(disk->d_state)));
2655                         /*
2656                          * Reset bumping syncid if disk disappeared in STARTING
2657                          * state.
2658                          */
2659                         if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2660                                 sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2661 #ifdef  INVARIANTS
2662                 } else {
2663                         KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2664                             sc->sc_name,
2665                             g_mirror_device_state2str(sc->sc_state),
2666                             g_mirror_get_diskname(disk),
2667                             g_mirror_disk_state2str(disk->d_state)));
2668 #endif
2669                 }
2670                 DISK_STATE_CHANGED();
2671                 G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2672                     sc->sc_name, g_mirror_get_diskname(disk));
2673
2674                 g_mirror_destroy_disk(disk);
2675                 break;
2676         case G_MIRROR_DISK_STATE_DESTROY:
2677             {
2678                 int error;
2679
2680                 error = g_mirror_clear_metadata(disk);
2681                 if (error != 0) {
2682                         G_MIRROR_DEBUG(0,
2683                             "Device %s: failed to clear metadata on %s: %d.",
2684                             sc->sc_name, g_mirror_get_diskname(disk), error);
2685                         break;
2686                 }
2687                 DISK_STATE_CHANGED();
2688                 G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2689                     sc->sc_name, g_mirror_get_diskname(disk));
2690
2691                 g_mirror_destroy_disk(disk);
2692                 sc->sc_ndisks--;
2693                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2694                         g_mirror_update_metadata(disk);
2695                 }
2696                 break;
2697             }
2698         default:
2699                 KASSERT(1 == 0, ("Unknown state (%u).", state));
2700                 break;
2701         }
2702         return (0);
2703 }
2704 #undef  DISK_STATE_CHANGED
2705
2706 int
2707 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2708 {
2709         struct g_provider *pp;
2710         u_char *buf;
2711         int error;
2712
2713         g_topology_assert();
2714
2715         error = g_access(cp, 1, 0, 0);
2716         if (error != 0)
2717                 return (error);
2718         pp = cp->provider;
2719         g_topology_unlock();
2720         /* Metadata are stored on last sector. */
2721         buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2722             &error);
2723         g_topology_lock();
2724         g_access(cp, -1, 0, 0);
2725         if (buf == NULL) {
2726                 G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2727                     cp->provider->name, error);
2728                 return (error);
2729         }
2730
2731         /* Decode metadata. */
2732         error = mirror_metadata_decode(buf, md);
2733         g_free(buf);
2734         if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2735                 return (EINVAL);
2736         if (md->md_version > G_MIRROR_VERSION) {
2737                 G_MIRROR_DEBUG(0,
2738                     "Kernel module is too old to handle metadata from %s.",
2739                     cp->provider->name);
2740                 return (EINVAL);
2741         }
2742         if (error != 0) {
2743                 G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2744                     cp->provider->name);
2745                 return (error);
2746         }
2747
2748         return (0);
2749 }
2750
2751 static int
2752 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2753     struct g_mirror_metadata *md)
2754 {
2755
2756         if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2757                 G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2758                     pp->name, md->md_did);
2759                 return (EEXIST);
2760         }
2761         if (md->md_all != sc->sc_ndisks) {
2762                 G_MIRROR_DEBUG(1,
2763                     "Invalid '%s' field on disk %s (device %s), skipping.",
2764                     "md_all", pp->name, sc->sc_name);
2765                 return (EINVAL);
2766         }
2767         if (md->md_slice != sc->sc_slice) {
2768                 G_MIRROR_DEBUG(1,
2769                     "Invalid '%s' field on disk %s (device %s), skipping.",
2770                     "md_slice", pp->name, sc->sc_name);
2771                 return (EINVAL);
2772         }
2773         if (md->md_balance != sc->sc_balance) {
2774                 G_MIRROR_DEBUG(1,
2775                     "Invalid '%s' field on disk %s (device %s), skipping.",
2776                     "md_balance", pp->name, sc->sc_name);
2777                 return (EINVAL);
2778         }
2779 #if 0
2780         if (md->md_mediasize != sc->sc_mediasize) {
2781                 G_MIRROR_DEBUG(1,
2782                     "Invalid '%s' field on disk %s (device %s), skipping.",
2783                     "md_mediasize", pp->name, sc->sc_name);
2784                 return (EINVAL);
2785         }
2786 #endif
2787         if (sc->sc_mediasize > pp->mediasize) {
2788                 G_MIRROR_DEBUG(1,
2789                     "Invalid size of disk %s (device %s), skipping.", pp->name,
2790                     sc->sc_name);
2791                 return (EINVAL);
2792         }
2793         if (md->md_sectorsize != sc->sc_sectorsize) {
2794                 G_MIRROR_DEBUG(1,
2795                     "Invalid '%s' field on disk %s (device %s), skipping.",
2796                     "md_sectorsize", pp->name, sc->sc_name);
2797                 return (EINVAL);
2798         }
2799         if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2800                 G_MIRROR_DEBUG(1,
2801                     "Invalid sector size of disk %s (device %s), skipping.",
2802                     pp->name, sc->sc_name);
2803                 return (EINVAL);
2804         }
2805         if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
2806                 G_MIRROR_DEBUG(1,
2807                     "Invalid device flags on disk %s (device %s), skipping.",
2808                     pp->name, sc->sc_name);
2809                 return (EINVAL);
2810         }
2811         if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
2812                 G_MIRROR_DEBUG(1,
2813                     "Invalid disk flags on disk %s (device %s), skipping.",
2814                     pp->name, sc->sc_name);
2815                 return (EINVAL);
2816         }
2817         return (0);
2818 }
2819
2820 int
2821 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
2822     struct g_mirror_metadata *md)
2823 {
2824         struct g_mirror_disk *disk;
2825         int error;
2826
2827         g_topology_assert_not();
2828         G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
2829
2830         error = g_mirror_check_metadata(sc, pp, md);
2831         if (error != 0)
2832                 return (error);
2833         if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING &&
2834             md->md_genid < sc->sc_genid) {
2835                 G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
2836                     pp->name, sc->sc_name);
2837                 return (EINVAL);
2838         }
2839         disk = g_mirror_init_disk(sc, pp, md, &error);
2840         if (disk == NULL)
2841                 return (error);
2842         error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
2843             G_MIRROR_EVENT_WAIT);
2844         if (error != 0)
2845                 return (error);
2846         if (md->md_version < G_MIRROR_VERSION) {
2847                 G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
2848                     pp->name, md->md_version, G_MIRROR_VERSION);
2849                 g_mirror_update_metadata(disk);
2850         }
2851         return (0);
2852 }
2853
2854 static void
2855 g_mirror_destroy_delayed(void *arg, int flag)
2856 {
2857         struct g_mirror_softc *sc;
2858         int error;
2859
2860         if (flag == EV_CANCEL) {
2861                 G_MIRROR_DEBUG(1, "Destroying canceled.");
2862                 return;
2863         }
2864         sc = arg;
2865         g_topology_unlock();
2866         sx_xlock(&sc->sc_lock);
2867         KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
2868             ("DESTROY flag set on %s.", sc->sc_name));
2869         KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0,
2870             ("DESTROYING flag not set on %s.", sc->sc_name));
2871         G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
2872         error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
2873         if (error != 0) {
2874                 G_MIRROR_DEBUG(0, "Cannot destroy %s (error=%d).",
2875                     sc->sc_name, error);
2876                 sx_xunlock(&sc->sc_lock);
2877         }
2878         g_topology_lock();
2879 }
2880
2881 static int
2882 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
2883 {
2884         struct g_mirror_softc *sc;
2885         int dcr, dcw, dce, error = 0;
2886
2887         g_topology_assert();
2888         G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
2889             acw, ace);
2890
2891         sc = pp->geom->softc;
2892         if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
2893                 return (0);
2894         KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
2895
2896         dcr = pp->acr + acr;
2897         dcw = pp->acw + acw;
2898         dce = pp->ace + ace;
2899
2900         g_topology_unlock();
2901         sx_xlock(&sc->sc_lock);
2902         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
2903             LIST_EMPTY(&sc->sc_disks)) {
2904                 if (acr > 0 || acw > 0 || ace > 0)
2905                         error = ENXIO;
2906                 goto end;
2907         }
2908         if (dcw == 0)
2909                 g_mirror_idle(sc, dcw);
2910         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0) {
2911                 if (acr > 0 || acw > 0 || ace > 0) {
2912                         error = ENXIO;
2913                         goto end;
2914                 }
2915                 if (dcr == 0 && dcw == 0 && dce == 0) {
2916                         g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK,
2917                             sc, NULL);
2918                 }
2919         }
2920 end:
2921         sx_xunlock(&sc->sc_lock);
2922         g_topology_lock();
2923         return (error);
2924 }
2925
2926 static struct g_geom *
2927 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md)
2928 {
2929         struct g_mirror_softc *sc;
2930         struct g_geom *gp;
2931         int error, timeout;
2932
2933         g_topology_assert();
2934         G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
2935             md->md_mid);
2936
2937         /* One disk is minimum. */
2938         if (md->md_all < 1)
2939                 return (NULL);
2940         /*
2941          * Action geom.
2942          */
2943         gp = g_new_geomf(mp, "%s", md->md_name);
2944         sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
2945         gp->start = g_mirror_start;
2946         gp->orphan = g_mirror_orphan;
2947         gp->access = g_mirror_access;
2948         gp->dumpconf = g_mirror_dumpconf;
2949
2950         sc->sc_id = md->md_mid;
2951         sc->sc_slice = md->md_slice;
2952         sc->sc_balance = md->md_balance;
2953         sc->sc_mediasize = md->md_mediasize;
2954         sc->sc_sectorsize = md->md_sectorsize;
2955         sc->sc_ndisks = md->md_all;
2956         sc->sc_flags = md->md_mflags;
2957         sc->sc_bump_id = 0;
2958         sc->sc_idle = 1;
2959         sc->sc_last_write = time_uptime;
2960         sc->sc_writes = 0;
2961         sx_init(&sc->sc_lock, "gmirror:lock");
2962         bioq_init(&sc->sc_queue);
2963         mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
2964         bioq_init(&sc->sc_regular_delayed);
2965         bioq_init(&sc->sc_inflight);
2966         bioq_init(&sc->sc_sync_delayed);
2967         LIST_INIT(&sc->sc_disks);
2968         TAILQ_INIT(&sc->sc_events);
2969         mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
2970         callout_init(&sc->sc_callout, 1);
2971         mtx_init(&sc->sc_done_mtx, "gmirror:done", NULL, MTX_DEF);
2972         sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
2973         gp->softc = sc;
2974         sc->sc_geom = gp;
2975         sc->sc_provider = NULL;
2976         /*
2977          * Synchronization geom.
2978          */
2979         gp = g_new_geomf(mp, "%s.sync", md->md_name);
2980         gp->softc = sc;
2981         gp->orphan = g_mirror_orphan;
2982         sc->sc_sync.ds_geom = gp;
2983         sc->sc_sync.ds_ndisks = 0;
2984         error = kproc_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
2985             "g_mirror %s", md->md_name);
2986         if (error != 0) {
2987                 G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
2988                     sc->sc_name);
2989                 g_destroy_geom(sc->sc_sync.ds_geom);
2990                 mtx_destroy(&sc->sc_done_mtx);
2991                 mtx_destroy(&sc->sc_events_mtx);
2992                 mtx_destroy(&sc->sc_queue_mtx);
2993                 sx_destroy(&sc->sc_lock);
2994                 g_destroy_geom(sc->sc_geom);
2995                 free(sc, M_MIRROR);
2996                 return (NULL);
2997         }
2998
2999         G_MIRROR_DEBUG(1, "Device %s created (%u components, id=%u).",
3000             sc->sc_name, sc->sc_ndisks, sc->sc_id);
3001
3002         sc->sc_rootmount = root_mount_hold("GMIRROR");
3003         G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3004         /*
3005          * Run timeout.
3006          */
3007         timeout = g_mirror_timeout * hz;
3008         callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
3009         return (sc->sc_geom);
3010 }
3011
3012 int
3013 g_mirror_destroy(struct g_mirror_softc *sc, int how)
3014 {
3015         struct g_mirror_disk *disk;
3016         struct g_provider *pp;
3017
3018         g_topology_assert_not();
3019         if (sc == NULL)
3020                 return (ENXIO);
3021         sx_assert(&sc->sc_lock, SX_XLOCKED);
3022
3023         pp = sc->sc_provider;
3024         if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0 ||
3025             SCHEDULER_STOPPED())) {
3026                 switch (how) {
3027                 case G_MIRROR_DESTROY_SOFT:
3028                         G_MIRROR_DEBUG(1,
3029                             "Device %s is still open (r%dw%de%d).", pp->name,
3030                             pp->acr, pp->acw, pp->ace);
3031                         return (EBUSY);
3032                 case G_MIRROR_DESTROY_DELAYED:
3033                         G_MIRROR_DEBUG(1,
3034                             "Device %s will be destroyed on last close.",
3035                             pp->name);
3036                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
3037                                 if (disk->d_state ==
3038                                     G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3039                                         g_mirror_sync_stop(disk, 1);
3040                                 }
3041                         }
3042                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROYING;
3043                         return (EBUSY);
3044                 case G_MIRROR_DESTROY_HARD:
3045                         G_MIRROR_DEBUG(1, "Device %s is still open, so it "
3046                             "can't be definitely removed.", pp->name);
3047                 }
3048         }
3049
3050         g_topology_lock();
3051         if (sc->sc_geom->softc == NULL) {
3052                 g_topology_unlock();
3053                 return (0);
3054         }
3055         sc->sc_geom->softc = NULL;
3056         sc->sc_sync.ds_geom->softc = NULL;
3057         g_topology_unlock();
3058
3059         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
3060         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_WAIT;
3061         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3062         sx_xunlock(&sc->sc_lock);
3063         mtx_lock(&sc->sc_queue_mtx);
3064         wakeup(sc);
3065         mtx_unlock(&sc->sc_queue_mtx);
3066         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3067         while (sc->sc_worker != NULL)
3068                 tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
3069         G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3070         sx_xlock(&sc->sc_lock);
3071         g_mirror_destroy_device(sc);
3072         free(sc, M_MIRROR);
3073         return (0);
3074 }
3075
3076 static void
3077 g_mirror_taste_orphan(struct g_consumer *cp)
3078 {
3079
3080         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3081             cp->provider->name));
3082 }
3083
3084 static struct g_geom *
3085 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3086 {
3087         struct g_mirror_metadata md;
3088         struct g_mirror_softc *sc;
3089         struct g_consumer *cp;
3090         struct g_geom *gp;
3091         int error;
3092
3093         g_topology_assert();
3094         g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3095         G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
3096
3097         gp = g_new_geomf(mp, "mirror:taste");
3098         /*
3099          * This orphan function should be never called.
3100          */
3101         gp->orphan = g_mirror_taste_orphan;
3102         cp = g_new_consumer(gp);
3103         g_attach(cp, pp);
3104         error = g_mirror_read_metadata(cp, &md);
3105         g_detach(cp);
3106         g_destroy_consumer(cp);
3107         g_destroy_geom(gp);
3108         if (error != 0)
3109                 return (NULL);
3110         gp = NULL;
3111
3112         if (md.md_provider[0] != '\0' &&
3113             !g_compare_names(md.md_provider, pp->name))
3114                 return (NULL);
3115         if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3116                 return (NULL);
3117         if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
3118                 G_MIRROR_DEBUG(0,
3119                     "Device %s: provider %s marked as inactive, skipping.",
3120                     md.md_name, pp->name);
3121                 return (NULL);
3122         }
3123         if (g_mirror_debug >= 2)
3124                 mirror_metadata_dump(&md);
3125
3126         /*
3127          * Let's check if device already exists.
3128          */
3129         sc = NULL;
3130         LIST_FOREACH(gp, &mp->geom, geom) {
3131                 sc = gp->softc;
3132                 if (sc == NULL)
3133                         continue;
3134                 if (sc->sc_sync.ds_geom == gp)
3135                         continue;
3136                 if (strcmp(md.md_name, sc->sc_name) != 0)
3137                         continue;
3138                 if (md.md_mid != sc->sc_id) {
3139                         G_MIRROR_DEBUG(0, "Device %s already configured.",
3140                             sc->sc_name);
3141                         return (NULL);
3142                 }
3143                 break;
3144         }
3145         if (gp == NULL) {
3146                 gp = g_mirror_create(mp, &md);
3147                 if (gp == NULL) {
3148                         G_MIRROR_DEBUG(0, "Cannot create device %s.",
3149                             md.md_name);
3150                         return (NULL);
3151                 }
3152                 sc = gp->softc;
3153         }
3154         G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3155         g_topology_unlock();
3156         sx_xlock(&sc->sc_lock);
3157         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_TASTING;
3158         error = g_mirror_add_disk(sc, pp, &md);
3159         if (error != 0) {
3160                 G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3161                     pp->name, gp->name, error);
3162                 if (LIST_EMPTY(&sc->sc_disks)) {
3163                         g_cancel_event(sc);
3164                         g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3165                         g_topology_lock();
3166                         return (NULL);
3167                 }
3168                 gp = NULL;
3169         }
3170         sc->sc_flags &= ~G_MIRROR_DEVICE_FLAG_TASTING;
3171         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
3172                 g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3173                 g_topology_lock();
3174                 return (NULL);
3175         }
3176         sx_xunlock(&sc->sc_lock);
3177         g_topology_lock();
3178         return (gp);
3179 }
3180
3181 static void
3182 g_mirror_resize(struct g_consumer *cp)
3183 {
3184         struct g_mirror_disk *disk;
3185
3186         g_topology_assert();
3187         g_trace(G_T_TOPOLOGY, "%s(%s)", __func__, cp->provider->name);
3188
3189         disk = cp->private;
3190         if (disk == NULL)
3191                 return;
3192         g_topology_unlock();
3193         g_mirror_update_metadata(disk);
3194         g_topology_lock();
3195 }
3196
3197 static int
3198 g_mirror_destroy_geom(struct gctl_req *req __unused,
3199     struct g_class *mp __unused, struct g_geom *gp)
3200 {
3201         struct g_mirror_softc *sc;
3202         int error;
3203
3204         g_topology_unlock();
3205         sc = gp->softc;
3206         sx_xlock(&sc->sc_lock);
3207         g_cancel_event(sc);
3208         error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3209         if (error != 0)
3210                 sx_xunlock(&sc->sc_lock);
3211         g_topology_lock();
3212         return (error);
3213 }
3214
3215 static void
3216 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3217     struct g_consumer *cp, struct g_provider *pp)
3218 {
3219         struct g_mirror_softc *sc;
3220
3221         g_topology_assert();
3222
3223         sc = gp->softc;
3224         if (sc == NULL)
3225                 return;
3226         /* Skip synchronization geom. */
3227         if (gp == sc->sc_sync.ds_geom)
3228                 return;
3229         if (pp != NULL) {
3230                 /* Nothing here. */
3231         } else if (cp != NULL) {
3232                 struct g_mirror_disk *disk;
3233
3234                 disk = cp->private;
3235                 if (disk == NULL)
3236                         return;
3237                 g_topology_unlock();
3238                 sx_xlock(&sc->sc_lock);
3239                 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3240                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3241                         sbuf_printf(sb, "%s<Synchronized>", indent);
3242                         if (disk->d_sync.ds_offset == 0)
3243                                 sbuf_printf(sb, "0%%");
3244                         else {
3245                                 sbuf_printf(sb, "%u%%",
3246                                     (u_int)((disk->d_sync.ds_offset * 100) /
3247                                     sc->sc_provider->mediasize));
3248                         }
3249                         sbuf_printf(sb, "</Synchronized>\n");
3250                         if (disk->d_sync.ds_offset > 0) {
3251                                 sbuf_printf(sb, "%s<BytesSynced>%jd"
3252                                     "</BytesSynced>\n", indent,
3253                                     (intmax_t)disk->d_sync.ds_offset);
3254                         }
3255                 }
3256                 sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3257                     disk->d_sync.ds_syncid);
3258                 sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3259                     disk->d_genid);
3260                 sbuf_printf(sb, "%s<Flags>", indent);
3261                 if (disk->d_flags == 0)
3262                         sbuf_printf(sb, "NONE");
3263                 else {
3264                         int first = 1;
3265
3266 #define ADD_FLAG(flag, name)    do {                                    \
3267         if ((disk->d_flags & (flag)) != 0) {                            \
3268                 if (!first)                                             \
3269                         sbuf_printf(sb, ", ");                          \
3270                 else                                                    \
3271                         first = 0;                                      \
3272                 sbuf_printf(sb, name);                                  \
3273         }                                                               \
3274 } while (0)
3275                         ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3276                         ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3277                         ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3278                         ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3279                             "SYNCHRONIZING");
3280                         ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3281                         ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3282 #undef  ADD_FLAG
3283                 }
3284                 sbuf_printf(sb, "</Flags>\n");
3285                 sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3286                     disk->d_priority);
3287                 sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3288                     g_mirror_disk_state2str(disk->d_state));
3289                 sx_xunlock(&sc->sc_lock);
3290                 g_topology_lock();
3291         } else {
3292                 g_topology_unlock();
3293                 sx_xlock(&sc->sc_lock);
3294                 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3295                 sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3296                 sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3297                 sbuf_printf(sb, "%s<Flags>", indent);
3298                 if (sc->sc_flags == 0)
3299                         sbuf_printf(sb, "NONE");
3300                 else {
3301                         int first = 1;
3302
3303 #define ADD_FLAG(flag, name)    do {                                    \
3304         if ((sc->sc_flags & (flag)) != 0) {                             \
3305                 if (!first)                                             \
3306                         sbuf_printf(sb, ", ");                          \
3307                 else                                                    \
3308                         first = 0;                                      \
3309                 sbuf_printf(sb, name);                                  \
3310         }                                                               \
3311 } while (0)
3312                         ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3313                         ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3314 #undef  ADD_FLAG
3315                 }
3316                 sbuf_printf(sb, "</Flags>\n");
3317                 sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3318                     (u_int)sc->sc_slice);
3319                 sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3320                     balance_name(sc->sc_balance));
3321                 sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3322                     sc->sc_ndisks);
3323                 sbuf_printf(sb, "%s<State>", indent);
3324                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3325                         sbuf_printf(sb, "%s", "STARTING");
3326                 else if (sc->sc_ndisks ==
3327                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3328                         sbuf_printf(sb, "%s", "COMPLETE");
3329                 else
3330                         sbuf_printf(sb, "%s", "DEGRADED");
3331                 sbuf_printf(sb, "</State>\n");
3332                 sx_xunlock(&sc->sc_lock);
3333                 g_topology_lock();
3334         }
3335 }
3336
3337 static void
3338 g_mirror_shutdown_post_sync(void *arg, int howto)
3339 {
3340         struct g_class *mp;
3341         struct g_geom *gp, *gp2;
3342         struct g_mirror_softc *sc;
3343         int error;
3344
3345         mp = arg;
3346         g_topology_lock();
3347         g_mirror_shutdown = 1;
3348         LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3349                 if ((sc = gp->softc) == NULL)
3350                         continue;
3351                 /* Skip synchronization geom. */
3352                 if (gp == sc->sc_sync.ds_geom)
3353                         continue;
3354                 g_topology_unlock();
3355                 sx_xlock(&sc->sc_lock);
3356                 g_mirror_idle(sc, -1);
3357                 g_cancel_event(sc);
3358                 error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3359                 if (error != 0)
3360                         sx_xunlock(&sc->sc_lock);
3361                 g_topology_lock();
3362         }
3363         g_topology_unlock();
3364 }
3365
3366 static void
3367 g_mirror_init(struct g_class *mp)
3368 {
3369
3370         g_mirror_post_sync = EVENTHANDLER_REGISTER(shutdown_post_sync,
3371             g_mirror_shutdown_post_sync, mp, SHUTDOWN_PRI_FIRST);
3372         if (g_mirror_post_sync == NULL)
3373                 G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3374 }
3375
3376 static void
3377 g_mirror_fini(struct g_class *mp)
3378 {
3379
3380         if (g_mirror_post_sync != NULL)
3381                 EVENTHANDLER_DEREGISTER(shutdown_post_sync, g_mirror_post_sync);
3382 }
3383
3384 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);