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