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                                 if (bp->bio_error == ENXIO &&
 979                                     bp->bio_cmd == BIO_READ)
 980                                         sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
 981                                 else if (bp->bio_error == ENXIO)
 982                                         sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID_NOW;
 983                                 else
 984                                         sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
 985                                 g_mirror_event_send(disk,
 986                                     G_MIRROR_DISK_STATE_DISCONNECTED,
 987                                     G_MIRROR_EVENT_DONTWAIT);
 988                         }
 989                 }
 990                 switch (pbp->bio_cmd) {
 991                 case BIO_DELETE:
 992                 case BIO_WRITE:
 993                         pbp->bio_inbed--;
 994                         pbp->bio_children--;
 995                         break;
 996                 }
 997         }
 998         g_destroy_bio(bp);
 999
1000         switch (pbp->bio_cmd) {
1001         case BIO_READ:
1002                 if (pbp->bio_inbed < pbp->bio_children)
1003                         break;
1004                 if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1)
1005                         g_io_deliver(pbp, pbp->bio_error);
1006                 else {
1007                         pbp->bio_error = 0;
1008                         mtx_lock(&sc->sc_queue_mtx);
1009                         bioq_insert_tail(&sc->sc_queue, pbp);
1010                         mtx_unlock(&sc->sc_queue_mtx);
1011                         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1012                         wakeup(sc);
1013                 }
1014                 break;
1015         case BIO_DELETE:
1016         case BIO_WRITE:
1017                 if (pbp->bio_children == 0) {
1018                         /*
1019                          * All requests failed.
1020                          */
1021                 } else if (pbp->bio_inbed < pbp->bio_children) {
1022                         /* Do nothing. */
1023                         break;
1024                 } else if (pbp->bio_children == pbp->bio_inbed) {
1025                         /* Some requests succeeded. */
1026                         pbp->bio_error = 0;
1027                         pbp->bio_completed = pbp->bio_length;
1028                 }
1029                 bioq_remove(&sc->sc_inflight, pbp);
1030                 /* Release delayed sync requests if possible. */
1031                 g_mirror_sync_release(sc);
1032                 g_io_deliver(pbp, pbp->bio_error);
1033                 break;
1034         default:
1035                 KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
1036                 break;
1037         }
1038 }
1039
1040 static void
1041 g_mirror_sync_done(struct bio *bp)
1042 {
1043         struct g_mirror_softc *sc;
1044
1045         G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
1046         sc = bp->bio_from->geom->softc;
1047         bp->bio_cflags = G_MIRROR_BIO_FLAG_SYNC;
1048         mtx_lock(&sc->sc_queue_mtx);
1049         bioq_insert_tail(&sc->sc_queue, bp);
1050         mtx_unlock(&sc->sc_queue_mtx);
1051         wakeup(sc);
1052 }
1053
1054 static void
1055 g_mirror_candelete(struct bio *bp)
1056 {
1057         struct g_mirror_softc *sc;
1058         struct g_mirror_disk *disk;
1059         int *val;
1060
1061         sc = bp->bio_to->private;
1062         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1063                 if (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE)
1064                         break;
1065         }
1066         val = (int *)bp->bio_data;
1067         *val = (disk != NULL);
1068         g_io_deliver(bp, 0);
1069 }
1070
1071 static void
1072 g_mirror_kernel_dump(struct bio *bp)
1073 {
1074         struct g_mirror_softc *sc;
1075         struct g_mirror_disk *disk;
1076         struct bio *cbp;
1077         struct g_kerneldump *gkd;
1078
1079         /*
1080          * We configure dumping to the first component, because this component
1081          * will be used for reading with 'prefer' balance algorithm.
1082          * If the component with the higest priority is currently disconnected
1083          * we will not be able to read the dump after the reboot if it will be
1084          * connected and synchronized later. Can we do something better?
1085          */
1086         sc = bp->bio_to->private;
1087         disk = LIST_FIRST(&sc->sc_disks);
1088
1089         gkd = (struct g_kerneldump *)bp->bio_data;
1090         if (gkd->length > bp->bio_to->mediasize)
1091                 gkd->length = bp->bio_to->mediasize;
1092         cbp = g_clone_bio(bp);
1093         if (cbp == NULL) {
1094                 g_io_deliver(bp, ENOMEM);
1095                 return;
1096         }
1097         cbp->bio_done = g_std_done;
1098         g_io_request(cbp, disk->d_consumer);
1099         G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1100             g_mirror_get_diskname(disk));
1101 }
1102
1103 static void
1104 g_mirror_flush(struct g_mirror_softc *sc, struct bio *bp)
1105 {
1106         struct bio_queue_head queue;
1107         struct g_mirror_disk *disk;
1108         struct g_consumer *cp;
1109         struct bio *cbp;
1110
1111         bioq_init(&queue);
1112         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1113                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1114                         continue;
1115                 cbp = g_clone_bio(bp);
1116                 if (cbp == NULL) {
1117                         while ((cbp = bioq_takefirst(&queue)) != NULL)
1118                                 g_destroy_bio(cbp);
1119                         if (bp->bio_error == 0)
1120                                 bp->bio_error = ENOMEM;
1121                         g_io_deliver(bp, bp->bio_error);
1122                         return;
1123                 }
1124                 bioq_insert_tail(&queue, cbp);
1125                 cbp->bio_done = g_mirror_flush_done;
1126                 cbp->bio_caller1 = disk;
1127                 cbp->bio_to = disk->d_consumer->provider;
1128         }
1129         while ((cbp = bioq_takefirst(&queue)) != NULL) {
1130                 G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1131                 disk = cbp->bio_caller1;
1132                 cbp->bio_caller1 = NULL;
1133                 cp = disk->d_consumer;
1134                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1135                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1136                     cp->acr, cp->acw, cp->ace));
1137                 g_io_request(cbp, disk->d_consumer);
1138         }
1139 }
1140
1141 static void
1142 g_mirror_start(struct bio *bp)
1143 {
1144         struct g_mirror_softc *sc;
1145
1146         sc = bp->bio_to->private;
1147         /*
1148          * If sc == NULL or there are no valid disks, provider's error
1149          * should be set and g_mirror_start() should not be called at all.
1150          */
1151         KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1152             ("Provider's error should be set (error=%d)(mirror=%s).",
1153             bp->bio_to->error, bp->bio_to->name));
1154         G_MIRROR_LOGREQ(3, bp, "Request received.");
1155
1156         switch (bp->bio_cmd) {
1157         case BIO_READ:
1158         case BIO_WRITE:
1159         case BIO_DELETE:
1160                 break;
1161         case BIO_FLUSH:
1162                 g_mirror_flush(sc, bp);
1163                 return;
1164         case BIO_GETATTR:
1165                 if (!strcmp(bp->bio_attribute, "GEOM::candelete")) {
1166                         g_mirror_candelete(bp);
1167                         return;
1168                 } else if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1169                         g_mirror_kernel_dump(bp);
1170                         return;
1171                 }
1172                 /* FALLTHROUGH */
1173         default:
1174                 g_io_deliver(bp, EOPNOTSUPP);
1175                 return;
1176         }
1177         mtx_lock(&sc->sc_queue_mtx);
1178         bioq_insert_tail(&sc->sc_queue, bp);
1179         mtx_unlock(&sc->sc_queue_mtx);
1180         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1181         wakeup(sc);
1182 }
1183
1184 /*
1185  * Return TRUE if the given request is colliding with a in-progress
1186  * synchronization request.
1187  */
1188 static int
1189 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1190 {
1191         struct g_mirror_disk *disk;
1192         struct bio *sbp;
1193         off_t rstart, rend, sstart, send;
1194         u_int i;
1195
1196         if (sc->sc_sync.ds_ndisks == 0)
1197                 return (0);
1198         rstart = bp->bio_offset;
1199         rend = bp->bio_offset + bp->bio_length;
1200         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1201                 if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1202                         continue;
1203                 for (i = 0; i < g_mirror_syncreqs; i++) {
1204                         sbp = disk->d_sync.ds_bios[i];
1205                         if (sbp == NULL)
1206                                 continue;
1207                         sstart = sbp->bio_offset;
1208                         send = sbp->bio_offset + sbp->bio_length;
1209                         if (rend > sstart && rstart < send)
1210                                 return (1);
1211                 }
1212         }
1213         return (0);
1214 }
1215
1216 /*
1217  * Return TRUE if the given sync request is colliding with a in-progress regular
1218  * request.
1219  */
1220 static int
1221 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1222 {
1223         off_t rstart, rend, sstart, send;
1224         struct bio *bp;
1225
1226         if (sc->sc_sync.ds_ndisks == 0)
1227                 return (0);
1228         sstart = sbp->bio_offset;
1229         send = sbp->bio_offset + sbp->bio_length;
1230         TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1231                 rstart = bp->bio_offset;
1232                 rend = bp->bio_offset + bp->bio_length;
1233                 if (rend > sstart && rstart < send)
1234                         return (1);
1235         }
1236         return (0);
1237 }
1238
1239 /*
1240  * Puts request onto delayed queue.
1241  */
1242 static void
1243 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1244 {
1245
1246         G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1247         bioq_insert_head(&sc->sc_regular_delayed, bp);
1248 }
1249
1250 /*
1251  * Puts synchronization request onto delayed queue.
1252  */
1253 static void
1254 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1255 {
1256
1257         G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1258         bioq_insert_tail(&sc->sc_sync_delayed, bp);
1259 }
1260
1261 /*
1262  * Releases delayed regular requests which don't collide anymore with sync
1263  * requests.
1264  */
1265 static void
1266 g_mirror_regular_release(struct g_mirror_softc *sc)
1267 {
1268         struct bio *bp, *bp2;
1269
1270         TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1271                 if (g_mirror_sync_collision(sc, bp))
1272                         continue;
1273                 bioq_remove(&sc->sc_regular_delayed, bp);
1274                 G_MIRROR_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1275                 mtx_lock(&sc->sc_queue_mtx);
1276                 bioq_insert_head(&sc->sc_queue, bp);
1277 #if 0
1278                 /*
1279                  * wakeup() is not needed, because this function is called from
1280                  * the worker thread.
1281                  */
1282                 wakeup(&sc->sc_queue);
1283 #endif
1284                 mtx_unlock(&sc->sc_queue_mtx);
1285         }
1286 }
1287
1288 /*
1289  * Releases delayed sync requests which don't collide anymore with regular
1290  * requests.
1291  */
1292 static void
1293 g_mirror_sync_release(struct g_mirror_softc *sc)
1294 {
1295         struct bio *bp, *bp2;
1296
1297         TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1298                 if (g_mirror_regular_collision(sc, bp))
1299                         continue;
1300                 bioq_remove(&sc->sc_sync_delayed, bp);
1301                 G_MIRROR_LOGREQ(2, bp,
1302                     "Releasing delayed synchronization request.");
1303                 g_io_request(bp, bp->bio_from);
1304         }
1305 }
1306
1307 /*
1308  * Handle synchronization requests.
1309  * Every synchronization request is two-steps process: first, READ request is
1310  * send to active provider and then WRITE request (with read data) to the provider
1311  * beeing synchronized. When WRITE is finished, new synchronization request is
1312  * send.
1313  */
1314 static void
1315 g_mirror_sync_request(struct bio *bp)
1316 {
1317         struct g_mirror_softc *sc;
1318         struct g_mirror_disk *disk;
1319
1320         bp->bio_from->index--;
1321         sc = bp->bio_from->geom->softc;
1322         disk = bp->bio_from->private;
1323         if (disk == NULL) {
1324                 sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1325                 g_topology_lock();
1326                 g_mirror_kill_consumer(sc, bp->bio_from);
1327                 g_topology_unlock();
1328                 free(bp->bio_data, M_MIRROR);
1329                 g_destroy_bio(bp);
1330                 sx_xlock(&sc->sc_lock);
1331                 return;
1332         }
1333
1334         /*
1335          * Synchronization request.
1336          */
1337         switch (bp->bio_cmd) {
1338         case BIO_READ:
1339             {
1340                 struct g_consumer *cp;
1341
1342                 if (bp->bio_error != 0) {
1343                         G_MIRROR_LOGREQ(0, bp,
1344                             "Synchronization request failed (error=%d).",
1345                             bp->bio_error);
1346                         g_destroy_bio(bp);
1347                         return;
1348                 }
1349                 G_MIRROR_LOGREQ(3, bp,
1350                     "Synchronization request half-finished.");
1351                 bp->bio_cmd = BIO_WRITE;
1352                 bp->bio_cflags = 0;
1353                 cp = disk->d_consumer;
1354                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1355                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1356                     cp->acr, cp->acw, cp->ace));
1357                 cp->index++;
1358                 g_io_request(bp, cp);
1359                 return;
1360             }
1361         case BIO_WRITE:
1362             {
1363                 struct g_mirror_disk_sync *sync;
1364                 off_t offset;
1365                 void *data;
1366                 int i;
1367
1368                 if (bp->bio_error != 0) {
1369                         G_MIRROR_LOGREQ(0, bp,
1370                             "Synchronization request failed (error=%d).",
1371                             bp->bio_error);
1372                         g_destroy_bio(bp);
1373                         sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1374                         g_mirror_event_send(disk,
1375                             G_MIRROR_DISK_STATE_DISCONNECTED,
1376                             G_MIRROR_EVENT_DONTWAIT);
1377                         return;
1378                 }
1379                 G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1380                 sync = &disk->d_sync;
1381                 if (sync->ds_offset >= sc->sc_mediasize ||
1382                     sync->ds_consumer == NULL ||
1383                     (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1384                         /* Don't send more synchronization requests. */
1385                         sync->ds_inflight--;
1386                         if (sync->ds_bios != NULL) {
1387                                 i = (int)(uintptr_t)bp->bio_caller1;
1388                                 sync->ds_bios[i] = NULL;
1389                         }
1390                         free(bp->bio_data, M_MIRROR);
1391                         g_destroy_bio(bp);
1392                         if (sync->ds_inflight > 0)
1393                                 return;
1394                         if (sync->ds_consumer == NULL ||
1395                             (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1396                                 return;
1397                         }
1398                         /* Disk up-to-date, activate it. */
1399                         g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1400                             G_MIRROR_EVENT_DONTWAIT);
1401                         return;
1402                 }
1403
1404                 /* Send next synchronization request. */
1405                 data = bp->bio_data;
1406                 bzero(bp, sizeof(*bp));
1407                 bp->bio_cmd = BIO_READ;
1408                 bp->bio_offset = sync->ds_offset;
1409                 bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1410                 sync->ds_offset += bp->bio_length;
1411                 bp->bio_done = g_mirror_sync_done;
1412                 bp->bio_data = data;
1413                 bp->bio_from = sync->ds_consumer;
1414                 bp->bio_to = sc->sc_provider;
1415                 G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1416                 sync->ds_consumer->index++;
1417                 /*
1418                  * Delay the request if it is colliding with a regular request.
1419                  */
1420                 if (g_mirror_regular_collision(sc, bp))
1421                         g_mirror_sync_delay(sc, bp);
1422                 else
1423                         g_io_request(bp, sync->ds_consumer);
1424
1425                 /* Release delayed requests if possible. */
1426                 g_mirror_regular_release(sc);
1427
1428                 /* Find the smallest offset */
1429                 offset = sc->sc_mediasize;
1430                 for (i = 0; i < g_mirror_syncreqs; i++) {
1431                         bp = sync->ds_bios[i];
1432                         if (bp->bio_offset < offset)
1433                                 offset = bp->bio_offset;
1434                 }
1435                 if (sync->ds_offset_done + (MAXPHYS * 100) < offset) {
1436                         /* Update offset_done on every 100 blocks. */
1437                         sync->ds_offset_done = offset;
1438                         g_mirror_update_metadata(disk);
1439                 }
1440                 return;
1441             }
1442         default:
1443                 KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1444                     bp->bio_cmd, sc->sc_name));
1445                 break;
1446         }
1447 }
1448
1449 static void
1450 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1451 {
1452         struct g_mirror_disk *disk;
1453         struct g_consumer *cp;
1454         struct bio *cbp;
1455
1456         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1457                 if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1458                         break;
1459         }
1460         if (disk == NULL) {
1461                 if (bp->bio_error == 0)
1462                         bp->bio_error = ENXIO;
1463                 g_io_deliver(bp, bp->bio_error);
1464                 return;
1465         }
1466         cbp = g_clone_bio(bp);
1467         if (cbp == NULL) {
1468                 if (bp->bio_error == 0)
1469                         bp->bio_error = ENOMEM;
1470                 g_io_deliver(bp, bp->bio_error);
1471                 return;
1472         }
1473         /*
1474          * Fill in the component buf structure.
1475          */
1476         cp = disk->d_consumer;
1477         cbp->bio_done = g_mirror_done;
1478         cbp->bio_to = cp->provider;
1479         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1480         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1481             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1482             cp->acw, cp->ace));
1483         cp->index++;
1484         g_io_request(cbp, cp);
1485 }
1486
1487 static void
1488 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1489 {
1490         struct g_mirror_disk *disk;
1491         struct g_consumer *cp;
1492         struct bio *cbp;
1493
1494         disk = g_mirror_get_disk(sc);
1495         if (disk == NULL) {
1496                 if (bp->bio_error == 0)
1497                         bp->bio_error = ENXIO;
1498                 g_io_deliver(bp, bp->bio_error);
1499                 return;
1500         }
1501         cbp = g_clone_bio(bp);
1502         if (cbp == NULL) {
1503                 if (bp->bio_error == 0)
1504                         bp->bio_error = ENOMEM;
1505                 g_io_deliver(bp, bp->bio_error);
1506                 return;
1507         }
1508         /*
1509          * Fill in the component buf structure.
1510          */
1511         cp = disk->d_consumer;
1512         cbp->bio_done = g_mirror_done;
1513         cbp->bio_to = cp->provider;
1514         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1515         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1516             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1517             cp->acw, cp->ace));
1518         cp->index++;
1519         g_io_request(cbp, cp);
1520 }
1521
1522 #define TRACK_SIZE  (1 * 1024 * 1024)
1523 #define LOAD_SCALE      256
1524 #define ABS(x)          (((x) >= 0) ? (x) : (-(x)))
1525
1526 static void
1527 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1528 {
1529         struct g_mirror_disk *disk, *dp;
1530         struct g_consumer *cp;
1531         struct bio *cbp;
1532         int prio, best;
1533
1534         /* Find a disk with the smallest load. */
1535         disk = NULL;
1536         best = INT_MAX;
1537         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1538                 if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1539                         continue;
1540                 prio = dp->load;
1541                 /* If disk head is precisely in position - highly prefer it. */
1542                 if (dp->d_last_offset == bp->bio_offset)
1543                         prio -= 2 * LOAD_SCALE;
1544                 else
1545                 /* If disk head is close to position - prefer it. */
1546                 if (ABS(dp->d_last_offset - bp->bio_offset) < TRACK_SIZE)
1547                         prio -= 1 * LOAD_SCALE;
1548                 if (prio <= best) {
1549                         disk = dp;
1550                         best = prio;
1551                 }
1552         }
1553         KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1554         cbp = g_clone_bio(bp);
1555         if (cbp == NULL) {
1556                 if (bp->bio_error == 0)
1557                         bp->bio_error = ENOMEM;
1558                 g_io_deliver(bp, bp->bio_error);
1559                 return;
1560         }
1561         /*
1562          * Fill in the component buf structure.
1563          */
1564         cp = disk->d_consumer;
1565         cbp->bio_done = g_mirror_done;
1566         cbp->bio_to = cp->provider;
1567         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1568         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1569             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1570             cp->acw, cp->ace));
1571         cp->index++;
1572         /* Remember last head position */
1573         disk->d_last_offset = bp->bio_offset + bp->bio_length;
1574         /* Update loads. */
1575         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1576                 dp->load = (dp->d_consumer->index * LOAD_SCALE +
1577                     dp->load * 7) / 8;
1578         }
1579         g_io_request(cbp, cp);
1580 }
1581
1582 static void
1583 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1584 {
1585         struct bio_queue_head queue;
1586         struct g_mirror_disk *disk;
1587         struct g_consumer *cp;
1588         struct bio *cbp;
1589         off_t left, mod, offset, slice;
1590         u_char *data;
1591         u_int ndisks;
1592
1593         if (bp->bio_length <= sc->sc_slice) {
1594                 g_mirror_request_round_robin(sc, bp);
1595                 return;
1596         }
1597         ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1598         slice = bp->bio_length / ndisks;
1599         mod = slice % sc->sc_provider->sectorsize;
1600         if (mod != 0)
1601                 slice += sc->sc_provider->sectorsize - mod;
1602         /*
1603          * Allocate all bios before sending any request, so we can
1604          * return ENOMEM in nice and clean way.
1605          */
1606         left = bp->bio_length;
1607         offset = bp->bio_offset;
1608         data = bp->bio_data;
1609         bioq_init(&queue);
1610         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1611                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1612                         continue;
1613                 cbp = g_clone_bio(bp);
1614                 if (cbp == NULL) {
1615                         while ((cbp = bioq_takefirst(&queue)) != NULL)
1616                                 g_destroy_bio(cbp);
1617                         if (bp->bio_error == 0)
1618                                 bp->bio_error = ENOMEM;
1619                         g_io_deliver(bp, bp->bio_error);
1620                         return;
1621                 }
1622                 bioq_insert_tail(&queue, cbp);
1623                 cbp->bio_done = g_mirror_done;
1624                 cbp->bio_caller1 = disk;
1625                 cbp->bio_to = disk->d_consumer->provider;
1626                 cbp->bio_offset = offset;
1627                 cbp->bio_data = data;
1628                 cbp->bio_length = MIN(left, slice);
1629                 left -= cbp->bio_length;
1630                 if (left == 0)
1631                         break;
1632                 offset += cbp->bio_length;
1633                 data += cbp->bio_length;
1634         }
1635         while ((cbp = bioq_takefirst(&queue)) != NULL) {
1636                 G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1637                 disk = cbp->bio_caller1;
1638                 cbp->bio_caller1 = NULL;
1639                 cp = disk->d_consumer;
1640                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1641                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1642                     cp->acr, cp->acw, cp->ace));
1643                 disk->d_consumer->index++;
1644                 g_io_request(cbp, disk->d_consumer);
1645         }
1646 }
1647
1648 static void
1649 g_mirror_register_request(struct bio *bp)
1650 {
1651         struct g_mirror_softc *sc;
1652
1653         sc = bp->bio_to->private;
1654         switch (bp->bio_cmd) {
1655         case BIO_READ:
1656                 switch (sc->sc_balance) {
1657                 case G_MIRROR_BALANCE_LOAD:
1658                         g_mirror_request_load(sc, bp);
1659                         break;
1660                 case G_MIRROR_BALANCE_PREFER:
1661                         g_mirror_request_prefer(sc, bp);
1662                         break;
1663                 case G_MIRROR_BALANCE_ROUND_ROBIN:
1664                         g_mirror_request_round_robin(sc, bp);
1665                         break;
1666                 case G_MIRROR_BALANCE_SPLIT:
1667                         g_mirror_request_split(sc, bp);
1668                         break;
1669                 }
1670                 return;
1671         case BIO_WRITE:
1672         case BIO_DELETE:
1673             {
1674                 struct g_mirror_disk *disk;
1675                 struct g_mirror_disk_sync *sync;
1676                 struct bio_queue_head queue;
1677                 struct g_consumer *cp;
1678                 struct bio *cbp;
1679
1680                 /*
1681                  * Delay the request if it is colliding with a synchronization
1682                  * request.
1683                  */
1684                 if (g_mirror_sync_collision(sc, bp)) {
1685                         g_mirror_regular_delay(sc, bp);
1686                         return;
1687                 }
1688
1689                 if (sc->sc_idle)
1690                         g_mirror_unidle(sc);
1691                 else
1692                         sc->sc_last_write = time_uptime;
1693
1694                 /*
1695                  * Allocate all bios before sending any request, so we can
1696                  * return ENOMEM in nice and clean way.
1697                  */
1698                 bioq_init(&queue);
1699                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1700                         sync = &disk->d_sync;
1701                         switch (disk->d_state) {
1702                         case G_MIRROR_DISK_STATE_ACTIVE:
1703                                 break;
1704                         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1705                                 if (bp->bio_offset >= sync->ds_offset)
1706                                         continue;
1707                                 break;
1708                         default:
1709                                 continue;
1710                         }
1711                         if (bp->bio_cmd == BIO_DELETE &&
1712                             (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE) == 0)
1713                                 continue;
1714                         cbp = g_clone_bio(bp);
1715                         if (cbp == NULL) {
1716                                 while ((cbp = bioq_takefirst(&queue)) != NULL)
1717                                         g_destroy_bio(cbp);
1718                                 if (bp->bio_error == 0)
1719                                         bp->bio_error = ENOMEM;
1720                                 g_io_deliver(bp, bp->bio_error);
1721                                 return;
1722                         }
1723                         bioq_insert_tail(&queue, cbp);
1724                         cbp->bio_done = g_mirror_done;
1725                         cp = disk->d_consumer;
1726                         cbp->bio_caller1 = cp;
1727                         cbp->bio_to = cp->provider;
1728                         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1729                             ("Consumer %s not opened (r%dw%de%d).",
1730                             cp->provider->name, cp->acr, cp->acw, cp->ace));
1731                 }
1732                 if (bioq_first(&queue) == NULL) {
1733                         g_io_deliver(bp, EOPNOTSUPP);
1734                         return;
1735                 }
1736                 while ((cbp = bioq_takefirst(&queue)) != NULL) {
1737                         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1738                         cp = cbp->bio_caller1;
1739                         cbp->bio_caller1 = NULL;
1740                         cp->index++;
1741                         sc->sc_writes++;
1742                         g_io_request(cbp, cp);
1743                 }
1744                 /*
1745                  * Put request onto inflight queue, so we can check if new
1746                  * synchronization requests don't collide with it.
1747                  */
1748                 bioq_insert_tail(&sc->sc_inflight, bp);
1749                 /*
1750                  * Bump syncid on first write.
1751                  */
1752                 if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1753                         sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1754                         g_mirror_bump_syncid(sc);
1755                 }
1756                 return;
1757             }
1758         default:
1759                 KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1760                     bp->bio_cmd, sc->sc_name));
1761                 break;
1762         }
1763 }
1764
1765 static int
1766 g_mirror_can_destroy(struct g_mirror_softc *sc)
1767 {
1768         struct g_geom *gp;
1769         struct g_consumer *cp;
1770
1771         g_topology_assert();
1772         gp = sc->sc_geom;
1773         if (gp->softc == NULL)
1774                 return (1);
1775         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_TASTING) != 0)
1776                 return (0);
1777         LIST_FOREACH(cp, &gp->consumer, consumer) {
1778                 if (g_mirror_is_busy(sc, cp))
1779                         return (0);
1780         }
1781         gp = sc->sc_sync.ds_geom;
1782         LIST_FOREACH(cp, &gp->consumer, consumer) {
1783                 if (g_mirror_is_busy(sc, cp))
1784                         return (0);
1785         }
1786         G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1787             sc->sc_name);
1788         return (1);
1789 }
1790
1791 static int
1792 g_mirror_try_destroy(struct g_mirror_softc *sc)
1793 {
1794
1795         if (sc->sc_rootmount != NULL) {
1796                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1797                     sc->sc_rootmount);
1798                 root_mount_rel(sc->sc_rootmount);
1799                 sc->sc_rootmount = NULL;
1800         }
1801         g_topology_lock();
1802         if (!g_mirror_can_destroy(sc)) {
1803                 g_topology_unlock();
1804                 return (0);
1805         }
1806         sc->sc_geom->softc = NULL;
1807         sc->sc_sync.ds_geom->softc = NULL;
1808         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WAIT) != 0) {
1809                 g_topology_unlock();
1810                 G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1811                     &sc->sc_worker);
1812                 /* Unlock sc_lock here, as it can be destroyed after wakeup. */
1813                 sx_xunlock(&sc->sc_lock);
1814                 wakeup(&sc->sc_worker);
1815                 sc->sc_worker = NULL;
1816         } else {
1817                 g_topology_unlock();
1818                 g_mirror_destroy_device(sc);
1819         }
1820         return (1);
1821 }
1822
1823 /*
1824  * Worker thread.
1825  */
1826 static void
1827 g_mirror_worker(void *arg)
1828 {
1829         struct g_mirror_softc *sc;
1830         struct g_mirror_event *ep;
1831         struct bio *bp;
1832         int timeout;
1833
1834         sc = arg;
1835         thread_lock(curthread);
1836         sched_prio(curthread, PRIBIO);
1837         thread_unlock(curthread);
1838
1839         sx_xlock(&sc->sc_lock);
1840         for (;;) {
1841                 G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1842                 /*
1843                  * First take a look at events.
1844                  * This is important to handle events before any I/O requests.
1845                  */
1846                 ep = g_mirror_event_get(sc);
1847                 if (ep != NULL) {
1848                         g_mirror_event_remove(sc, ep);
1849                         if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1850                                 /* Update only device status. */
1851                                 G_MIRROR_DEBUG(3,
1852                                     "Running event for device %s.",
1853                                     sc->sc_name);
1854                                 ep->e_error = 0;
1855                                 g_mirror_update_device(sc, 1);
1856                         } else {
1857                                 /* Update disk status. */
1858                                 G_MIRROR_DEBUG(3, "Running event for disk %s.",
1859                                      g_mirror_get_diskname(ep->e_disk));
1860                                 ep->e_error = g_mirror_update_disk(ep->e_disk,
1861                                     ep->e_state);
1862                                 if (ep->e_error == 0)
1863                                         g_mirror_update_device(sc, 0);
1864                         }
1865                         if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1866                                 KASSERT(ep->e_error == 0,
1867                                     ("Error cannot be handled."));
1868                                 g_mirror_event_free(ep);
1869                         } else {
1870                                 ep->e_flags |= G_MIRROR_EVENT_DONE;
1871                                 G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1872                                     ep);
1873                                 mtx_lock(&sc->sc_events_mtx);
1874                                 wakeup(ep);
1875                                 mtx_unlock(&sc->sc_events_mtx);
1876                         }
1877                         if ((sc->sc_flags &
1878                             G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1879                                 if (g_mirror_try_destroy(sc)) {
1880                                         curthread->td_pflags &= ~TDP_GEOM;
1881                                         G_MIRROR_DEBUG(1, "Thread exiting.");
1882                                         kproc_exit(0);
1883                                 }
1884                         }
1885                         G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1886                         continue;
1887                 }
1888                 /*
1889                  * Check if we can mark array as CLEAN and if we can't take
1890                  * how much seconds should we wait.
1891                  */
1892                 timeout = g_mirror_idle(sc, -1);
1893                 /*
1894                  * Now I/O requests.
1895                  */
1896                 /* Get first request from the queue. */
1897                 mtx_lock(&sc->sc_queue_mtx);
1898                 bp = bioq_takefirst(&sc->sc_queue);
1899                 if (bp == NULL) {
1900                         if ((sc->sc_flags &
1901                             G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1902                                 mtx_unlock(&sc->sc_queue_mtx);
1903                                 if (g_mirror_try_destroy(sc)) {
1904                                         curthread->td_pflags &= ~TDP_GEOM;
1905                                         G_MIRROR_DEBUG(1, "Thread exiting.");
1906                                         kproc_exit(0);
1907                                 }
1908                                 mtx_lock(&sc->sc_queue_mtx);
1909                         }
1910                         sx_xunlock(&sc->sc_lock);
1911                         /*
1912                          * XXX: We can miss an event here, because an event
1913                          *      can be added without sx-device-lock and without
1914                          *      mtx-queue-lock. Maybe I should just stop using
1915                          *      dedicated mutex for events synchronization and
1916                          *      stick with the queue lock?
1917                          *      The event will hang here until next I/O request
1918                          *      or next event is received.
1919                          */
1920                         MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
1921                             timeout * hz);
1922                         sx_xlock(&sc->sc_lock);
1923                         G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
1924                         continue;
1925                 }
1926                 mtx_unlock(&sc->sc_queue_mtx);
1927
1928                 if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
1929                     (bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0) {
1930                         g_mirror_sync_request(bp);      /* READ */
1931                 } else if (bp->bio_to != sc->sc_provider) {
1932                         if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
1933                                 g_mirror_regular_request(bp);
1934                         else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
1935                                 g_mirror_sync_request(bp);      /* WRITE */
1936                         else {
1937                                 KASSERT(0,
1938                                     ("Invalid request cflags=0x%hhx to=%s.",
1939                                     bp->bio_cflags, bp->bio_to->name));
1940                         }
1941                 } else {
1942                         g_mirror_register_request(bp);
1943                 }
1944                 G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
1945         }
1946 }
1947
1948 static void
1949 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
1950 {
1951
1952         sx_assert(&sc->sc_lock, SX_LOCKED);
1953
1954         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
1955                 return;
1956         if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
1957                 G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
1958                     g_mirror_get_diskname(disk), sc->sc_name);
1959                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1960         } else if (sc->sc_idle &&
1961             (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
1962                 G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
1963                     g_mirror_get_diskname(disk), sc->sc_name);
1964                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
1965         }
1966 }
1967
1968 static void
1969 g_mirror_sync_start(struct g_mirror_disk *disk)
1970 {
1971         struct g_mirror_softc *sc;
1972         struct g_consumer *cp;
1973         struct bio *bp;
1974         int error, i;
1975
1976         g_topology_assert_not();
1977         sc = disk->d_softc;
1978         sx_assert(&sc->sc_lock, SX_LOCKED);
1979
1980         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1981             ("Disk %s is not marked for synchronization.",
1982             g_mirror_get_diskname(disk)));
1983         KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1984             ("Device not in RUNNING state (%s, %u).", sc->sc_name,
1985             sc->sc_state));
1986
1987         sx_xunlock(&sc->sc_lock);
1988         g_topology_lock();
1989         cp = g_new_consumer(sc->sc_sync.ds_geom);
1990         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
1991         error = g_attach(cp, sc->sc_provider);
1992         KASSERT(error == 0,
1993             ("Cannot attach to %s (error=%d).", sc->sc_name, error));
1994         error = g_access(cp, 1, 0, 0);
1995         KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
1996         g_topology_unlock();
1997         sx_xlock(&sc->sc_lock);
1998
1999         G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
2000             g_mirror_get_diskname(disk));
2001         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) == 0)
2002                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
2003         KASSERT(disk->d_sync.ds_consumer == NULL,
2004             ("Sync consumer already exists (device=%s, disk=%s).",
2005             sc->sc_name, g_mirror_get_diskname(disk)));
2006
2007         disk->d_sync.ds_consumer = cp;
2008         disk->d_sync.ds_consumer->private = disk;
2009         disk->d_sync.ds_consumer->index = 0;
2010
2011         /*
2012          * Allocate memory for synchronization bios and initialize them.
2013          */
2014         disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
2015             M_MIRROR, M_WAITOK);
2016         for (i = 0; i < g_mirror_syncreqs; i++) {
2017                 bp = g_alloc_bio();
2018                 disk->d_sync.ds_bios[i] = bp;
2019                 bp->bio_parent = NULL;
2020                 bp->bio_cmd = BIO_READ;
2021                 bp->bio_data = malloc(MAXPHYS, M_MIRROR, M_WAITOK);
2022                 bp->bio_cflags = 0;
2023                 bp->bio_offset = disk->d_sync.ds_offset;
2024                 bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
2025                 disk->d_sync.ds_offset += bp->bio_length;
2026                 bp->bio_done = g_mirror_sync_done;
2027                 bp->bio_from = disk->d_sync.ds_consumer;
2028                 bp->bio_to = sc->sc_provider;
2029                 bp->bio_caller1 = (void *)(uintptr_t)i;
2030         }
2031
2032         /* Increase the number of disks in SYNCHRONIZING state. */
2033         sc->sc_sync.ds_ndisks++;
2034         /* Set the number of in-flight synchronization requests. */
2035         disk->d_sync.ds_inflight = g_mirror_syncreqs;
2036
2037         /*
2038          * Fire off first synchronization requests.
2039          */
2040         for (i = 0; i < g_mirror_syncreqs; i++) {
2041                 bp = disk->d_sync.ds_bios[i];
2042                 G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
2043                 disk->d_sync.ds_consumer->index++;
2044                 /*
2045                  * Delay the request if it is colliding with a regular request.
2046                  */
2047                 if (g_mirror_regular_collision(sc, bp))
2048                         g_mirror_sync_delay(sc, bp);
2049                 else
2050                         g_io_request(bp, disk->d_sync.ds_consumer);
2051         }
2052 }
2053
2054 /*
2055  * Stop synchronization process.
2056  * type: 0 - synchronization finished
2057  *       1 - synchronization stopped
2058  */
2059 static void
2060 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
2061 {
2062         struct g_mirror_softc *sc;
2063         struct g_consumer *cp;
2064
2065         g_topology_assert_not();
2066         sc = disk->d_softc;
2067         sx_assert(&sc->sc_lock, SX_LOCKED);
2068
2069         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2070             ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2071             g_mirror_disk_state2str(disk->d_state)));
2072         if (disk->d_sync.ds_consumer == NULL)
2073                 return;
2074
2075         if (type == 0) {
2076                 G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2077                     sc->sc_name, g_mirror_get_diskname(disk));
2078         } else /* if (type == 1) */ {
2079                 G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2080                     sc->sc_name, g_mirror_get_diskname(disk));
2081         }
2082         free(disk->d_sync.ds_bios, M_MIRROR);
2083         disk->d_sync.ds_bios = NULL;
2084         cp = disk->d_sync.ds_consumer;
2085         disk->d_sync.ds_consumer = NULL;
2086         disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2087         sc->sc_sync.ds_ndisks--;
2088         sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2089         g_topology_lock();
2090         g_mirror_kill_consumer(sc, cp);
2091         g_topology_unlock();
2092         sx_xlock(&sc->sc_lock);
2093 }
2094
2095 static void
2096 g_mirror_launch_provider(struct g_mirror_softc *sc)
2097 {
2098         struct g_mirror_disk *disk;
2099         struct g_provider *pp, *dp;
2100
2101         sx_assert(&sc->sc_lock, SX_LOCKED);
2102
2103         g_topology_lock();
2104         pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
2105         pp->flags |= G_PF_DIRECT_RECEIVE;
2106         pp->mediasize = sc->sc_mediasize;
2107         pp->sectorsize = sc->sc_sectorsize;
2108         pp->stripesize = 0;
2109         pp->stripeoffset = 0;
2110
2111         /* Splitting of unmapped BIO's could work but isn't implemented now */
2112         if (sc->sc_balance != G_MIRROR_BALANCE_SPLIT)
2113                 pp->flags |= G_PF_ACCEPT_UNMAPPED;
2114
2115         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2116                 if (disk->d_consumer && disk->d_consumer->provider) {
2117                         dp = disk->d_consumer->provider;
2118                         if (dp->stripesize > pp->stripesize) {
2119                                 pp->stripesize = dp->stripesize;
2120                                 pp->stripeoffset = dp->stripeoffset;
2121                         }
2122                         /* A provider underneath us doesn't support unmapped */
2123                         if ((dp->flags & G_PF_ACCEPT_UNMAPPED) == 0) {
2124                                 G_MIRROR_DEBUG(0, "Cancelling unmapped "
2125                                     "because of %s.", dp->name);
2126                                 pp->flags &= ~G_PF_ACCEPT_UNMAPPED;
2127                         }
2128                 }
2129         }
2130         pp->private = sc;
2131         sc->sc_refcnt++;
2132         sc->sc_provider = pp;
2133         g_error_provider(pp, 0);
2134         g_topology_unlock();
2135         G_MIRROR_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2136             g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE), sc->sc_ndisks);
2137         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2138                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2139                         g_mirror_sync_start(disk);
2140         }
2141 }
2142
2143 static void
2144 g_mirror_destroy_provider(struct g_mirror_softc *sc)
2145 {
2146         struct g_mirror_disk *disk;
2147         struct bio *bp;
2148
2149         g_topology_assert_not();
2150         KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2151             sc->sc_name));
2152
2153         g_topology_lock();
2154         g_error_provider(sc->sc_provider, ENXIO);
2155         mtx_lock(&sc->sc_queue_mtx);
2156         while ((bp = bioq_takefirst(&sc->sc_queue)) != NULL) {
2157                 /*
2158                  * Abort any pending I/O that wasn't generated by us.
2159                  * Synchronization requests and requests destined for individual
2160                  * mirror components can be destroyed immediately.
2161                  */
2162                 if (bp->bio_to == sc->sc_provider &&
2163                     bp->bio_from->geom != sc->sc_sync.ds_geom) {
2164                         g_io_deliver(bp, ENXIO);
2165                 } else {
2166                         if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
2167                                 free(bp->bio_data, M_MIRROR);
2168                         g_destroy_bio(bp);
2169                 }
2170         }
2171         mtx_unlock(&sc->sc_queue_mtx);
2172         g_wither_provider(sc->sc_provider, ENXIO);
2173         sc->sc_provider = NULL;
2174         G_MIRROR_DEBUG(0, "Device %s: provider destroyed.", sc->sc_name);
2175         g_topology_unlock();
2176         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2177                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2178                         g_mirror_sync_stop(disk, 1);
2179         }
2180 }
2181
2182 static void
2183 g_mirror_go(void *arg)
2184 {
2185         struct g_mirror_softc *sc;
2186
2187         sc = arg;
2188         G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2189         g_mirror_event_send(sc, 0,
2190             G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2191 }
2192
2193 static u_int
2194 g_mirror_determine_state(struct g_mirror_disk *disk)
2195 {
2196         struct g_mirror_softc *sc;
2197         u_int state;
2198
2199         sc = disk->d_softc;
2200         if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2201                 if ((disk->d_flags &
2202                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2203                         /* Disk does not need synchronization. */
2204                         state = G_MIRROR_DISK_STATE_ACTIVE;
2205                 } else {
2206                         if ((sc->sc_flags &
2207                              G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2208                             (disk->d_flags &
2209                              G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2210                                 /*
2211                                  * We can start synchronization from
2212                                  * the stored offset.
2213                                  */
2214                                 state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2215                         } else {
2216                                 state = G_MIRROR_DISK_STATE_STALE;
2217                         }
2218                 }
2219         } else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2220                 /*
2221                  * Reset all synchronization data for this disk,
2222                  * because if it even was synchronized, it was
2223                  * synchronized to disks with different syncid.
2224                  */
2225                 disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2226                 disk->d_sync.ds_offset = 0;
2227                 disk->d_sync.ds_offset_done = 0;
2228                 disk->d_sync.ds_syncid = sc->sc_syncid;
2229                 if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2230                     (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2231                         state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2232                 } else {
2233                         state = G_MIRROR_DISK_STATE_STALE;
2234                 }
2235         } else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2236                 /*
2237                  * Not good, NOT GOOD!
2238                  * It means that mirror was started on stale disks
2239                  * and more fresh disk just arrive.
2240                  * If there were writes, mirror is broken, sorry.
2241                  * I think the best choice here is don't touch
2242                  * this disk and inform the user loudly.
2243                  */
2244                 G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2245                     "disk (%s) arrives!! It will not be connected to the "
2246                     "running device.", sc->sc_name,
2247                     g_mirror_get_diskname(disk));
2248                 g_mirror_destroy_disk(disk);
2249                 state = G_MIRROR_DISK_STATE_NONE;
2250                 /* Return immediately, because disk was destroyed. */
2251                 return (state);
2252         }
2253         G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2254             g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2255         return (state);
2256 }
2257
2258 /*
2259  * Update device state.
2260  */
2261 static void
2262 g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force)
2263 {
2264         struct g_mirror_disk *disk;
2265         u_int state;
2266
2267         sx_assert(&sc->sc_lock, SX_XLOCKED);
2268
2269         switch (sc->sc_state) {
2270         case G_MIRROR_DEVICE_STATE_STARTING:
2271             {
2272                 struct g_mirror_disk *pdisk, *tdisk;
2273                 u_int dirty, ndisks, genid, syncid;
2274
2275                 KASSERT(sc->sc_provider == NULL,
2276                     ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2277                 /*
2278                  * Are we ready? We are, if all disks are connected or
2279                  * if we have any disks and 'force' is true.
2280                  */
2281                 ndisks = g_mirror_ndisks(sc, -1);
2282                 if (sc->sc_ndisks == ndisks || (force && ndisks > 0)) {
2283                         ;
2284                 } else if (ndisks == 0) {
2285                         /*
2286                          * Disks went down in starting phase, so destroy
2287                          * device.
2288                          */
2289                         callout_drain(&sc->sc_callout);
2290                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2291                         G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2292                             sc->sc_rootmount);
2293                         root_mount_rel(sc->sc_rootmount);
2294                         sc->sc_rootmount = NULL;
2295                         return;
2296                 } else {
2297                         return;
2298                 }
2299
2300                 /*
2301                  * Activate all disks with the biggest syncid.
2302                  */
2303                 if (force) {
2304                         /*
2305                          * If 'force' is true, we have been called due to
2306                          * timeout, so don't bother canceling timeout.
2307                          */
2308                         ndisks = 0;
2309                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2310                                 if ((disk->d_flags &
2311                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2312                                         ndisks++;
2313                                 }
2314                         }
2315                         if (ndisks == 0) {
2316                                 /* No valid disks found, destroy device. */
2317                                 sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2318                                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2319                                     __LINE__, sc->sc_rootmount);
2320                                 root_mount_rel(sc->sc_rootmount);
2321                                 sc->sc_rootmount = NULL;
2322                                 return;
2323                         }
2324                 } else {
2325                         /* Cancel timeout. */
2326                         callout_drain(&sc->sc_callout);
2327                 }
2328
2329                 /*
2330                  * Find the biggest genid.
2331                  */
2332                 genid = 0;
2333                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2334                         if (disk->d_genid > genid)
2335                                 genid = disk->d_genid;
2336                 }
2337                 sc->sc_genid = genid;
2338                 /*
2339                  * Remove all disks without the biggest genid.
2340                  */
2341                 LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2342                         if (disk->d_genid < genid) {
2343                                 G_MIRROR_DEBUG(0,
2344                                     "Component %s (device %s) broken, skipping.",
2345                                     g_mirror_get_diskname(disk), sc->sc_name);
2346                                 g_mirror_destroy_disk(disk);
2347                         }
2348                 }
2349
2350                 /*
2351                  * Find the biggest syncid.
2352                  */
2353                 syncid = 0;
2354                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2355                         if (disk->d_sync.ds_syncid > syncid)
2356                                 syncid = disk->d_sync.ds_syncid;
2357                 }
2358
2359                 /*
2360                  * Here we need to look for dirty disks and if all disks
2361                  * with the biggest syncid are dirty, we have to choose
2362                  * one with the biggest priority and rebuild the rest.
2363                  */
2364                 /*
2365                  * Find the number of dirty disks with the biggest syncid.
2366                  * Find the number of disks with the biggest syncid.
2367                  * While here, find a disk with the biggest priority.
2368                  */
2369                 dirty = ndisks = 0;
2370                 pdisk = NULL;
2371                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2372                         if (disk->d_sync.ds_syncid != syncid)
2373                                 continue;
2374                         if ((disk->d_flags &
2375                             G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2376                                 continue;
2377                         }
2378                         ndisks++;
2379                         if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2380                                 dirty++;
2381                                 if (pdisk == NULL ||
2382                                     pdisk->d_priority < disk->d_priority) {
2383                                         pdisk = disk;
2384                                 }
2385                         }
2386                 }
2387                 if (dirty == 0) {
2388                         /* No dirty disks at all, great. */
2389                 } else if (dirty == ndisks) {
2390                         /*
2391                          * Force synchronization for all dirty disks except one
2392                          * with the biggest priority.
2393                          */
2394                         KASSERT(pdisk != NULL, ("pdisk == NULL"));
2395                         G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2396                             "master disk for synchronization.",
2397                             g_mirror_get_diskname(pdisk), sc->sc_name);
2398                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2399                                 if (disk->d_sync.ds_syncid != syncid)
2400                                         continue;
2401                                 if ((disk->d_flags &
2402                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2403                                         continue;
2404                                 }
2405                                 KASSERT((disk->d_flags &
2406                                     G_MIRROR_DISK_FLAG_DIRTY) != 0,
2407                                     ("Disk %s isn't marked as dirty.",
2408                                     g_mirror_get_diskname(disk)));
2409                                 /* Skip the disk with the biggest priority. */
2410                                 if (disk == pdisk)
2411                                         continue;
2412                                 disk->d_sync.ds_syncid = 0;
2413                         }
2414                 } else if (dirty < ndisks) {
2415                         /*
2416                          * Force synchronization for all dirty disks.
2417                          * We have some non-dirty disks.
2418                          */
2419                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2420                                 if (disk->d_sync.ds_syncid != syncid)
2421                                         continue;
2422                                 if ((disk->d_flags &
2423                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2424                                         continue;
2425                                 }
2426                                 if ((disk->d_flags &
2427                                     G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2428                                         continue;
2429                                 }
2430                                 disk->d_sync.ds_syncid = 0;
2431                         }
2432                 }
2433
2434                 /* Reset hint. */
2435                 sc->sc_hint = NULL;
2436                 sc->sc_syncid = syncid;
2437                 if (force) {
2438                         /* Remember to bump syncid on first write. */
2439                         sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2440                 }
2441                 state = G_MIRROR_DEVICE_STATE_RUNNING;
2442                 G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2443                     sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2444                     g_mirror_device_state2str(state));
2445                 sc->sc_state = state;
2446                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2447                         state = g_mirror_determine_state(disk);
2448                         g_mirror_event_send(disk, state,
2449                             G_MIRROR_EVENT_DONTWAIT);
2450                         if (state == G_MIRROR_DISK_STATE_STALE)
2451                                 sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2452                 }
2453                 break;
2454             }
2455         case G_MIRROR_DEVICE_STATE_RUNNING:
2456                 if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2457                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2458                         /*
2459                          * No active disks or no disks at all,
2460                          * so destroy device.
2461                          */
2462                         if (sc->sc_provider != NULL)
2463                                 g_mirror_destroy_provider(sc);
2464                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2465                         break;
2466                 } else if (g_mirror_ndisks(sc,
2467                     G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2468                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2469                         /*
2470                          * We have active disks, launch provider if it doesn't
2471                          * exist.
2472                          */
2473                         if (sc->sc_provider == NULL)
2474                                 g_mirror_launch_provider(sc);
2475                         if (sc->sc_rootmount != NULL) {
2476                                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2477                                     __LINE__, sc->sc_rootmount);
2478                                 root_mount_rel(sc->sc_rootmount);
2479                                 sc->sc_rootmount = NULL;
2480                         }
2481                 }
2482                 /*
2483                  * Genid should be bumped immediately, so do it here.
2484                  */
2485                 if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2486                         sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2487                         g_mirror_bump_genid(sc);
2488                 }
2489                 if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID_NOW) != 0) {
2490                         sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID_NOW;
2491                         g_mirror_bump_syncid(sc);
2492                 }
2493                 break;
2494         default:
2495                 KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2496                     sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2497                 break;
2498         }
2499 }
2500
2501 /*
2502  * Update disk state and device state if needed.
2503  */
2504 #define DISK_STATE_CHANGED()    G_MIRROR_DEBUG(1,                       \
2505         "Disk %s state changed from %s to %s (device %s).",             \
2506         g_mirror_get_diskname(disk),                                    \
2507         g_mirror_disk_state2str(disk->d_state),                         \
2508         g_mirror_disk_state2str(state), sc->sc_name)
2509 static int
2510 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2511 {
2512         struct g_mirror_softc *sc;
2513
2514         sc = disk->d_softc;
2515         sx_assert(&sc->sc_lock, SX_XLOCKED);
2516
2517 again:
2518         G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2519             g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2520             g_mirror_disk_state2str(state));
2521         switch (state) {
2522         case G_MIRROR_DISK_STATE_NEW:
2523                 /*
2524                  * Possible scenarios:
2525                  * 1. New disk arrive.
2526                  */
2527                 /* Previous state should be NONE. */
2528                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2529                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2530                     g_mirror_disk_state2str(disk->d_state)));
2531                 DISK_STATE_CHANGED();
2532
2533                 disk->d_state = state;
2534                 if (LIST_EMPTY(&sc->sc_disks))
2535                         LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2536                 else {
2537                         struct g_mirror_disk *dp;
2538
2539                         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2540                                 if (disk->d_priority >= dp->d_priority) {
2541                                         LIST_INSERT_BEFORE(dp, disk, d_next);
2542                                         dp = NULL;
2543                                         break;
2544                                 }
2545                                 if (LIST_NEXT(dp, d_next) == NULL)
2546                                         break;
2547                         }
2548                         if (dp != NULL)
2549                                 LIST_INSERT_AFTER(dp, disk, d_next);
2550                 }
2551                 G_MIRROR_DEBUG(1, "Device %s: provider %s detected.",
2552                     sc->sc_name, g_mirror_get_diskname(disk));
2553                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2554                         break;
2555                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2556                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2557                     g_mirror_device_state2str(sc->sc_state),
2558                     g_mirror_get_diskname(disk),
2559                     g_mirror_disk_state2str(disk->d_state)));
2560                 state = g_mirror_determine_state(disk);
2561                 if (state != G_MIRROR_DISK_STATE_NONE)
2562                         goto again;
2563                 break;
2564         case G_MIRROR_DISK_STATE_ACTIVE:
2565                 /*
2566                  * Possible scenarios:
2567                  * 1. New disk does not need synchronization.
2568                  * 2. Synchronization process finished successfully.
2569                  */
2570                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2571                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2572                     g_mirror_device_state2str(sc->sc_state),
2573                     g_mirror_get_diskname(disk),
2574                     g_mirror_disk_state2str(disk->d_state)));
2575                 /* Previous state should be NEW or SYNCHRONIZING. */
2576                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2577                     disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2578                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2579                     g_mirror_disk_state2str(disk->d_state)));
2580                 DISK_STATE_CHANGED();
2581
2582                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2583                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2584                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2585                         g_mirror_sync_stop(disk, 0);
2586                 }
2587                 disk->d_state = state;
2588                 disk->d_sync.ds_offset = 0;
2589                 disk->d_sync.ds_offset_done = 0;
2590                 g_mirror_update_idle(sc, disk);
2591                 g_mirror_update_metadata(disk);
2592                 G_MIRROR_DEBUG(1, "Device %s: provider %s activated.",
2593                     sc->sc_name, g_mirror_get_diskname(disk));
2594                 break;
2595         case G_MIRROR_DISK_STATE_STALE:
2596                 /*
2597                  * Possible scenarios:
2598                  * 1. Stale disk was connected.
2599                  */
2600                 /* Previous state should be NEW. */
2601                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2602                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2603                     g_mirror_disk_state2str(disk->d_state)));
2604                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2605                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2606                     g_mirror_device_state2str(sc->sc_state),
2607                     g_mirror_get_diskname(disk),
2608                     g_mirror_disk_state2str(disk->d_state)));
2609                 /*
2610                  * STALE state is only possible if device is marked
2611                  * NOAUTOSYNC.
2612                  */
2613                 KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2614                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2615                     g_mirror_device_state2str(sc->sc_state),
2616                     g_mirror_get_diskname(disk),
2617                     g_mirror_disk_state2str(disk->d_state)));
2618                 DISK_STATE_CHANGED();
2619
2620                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2621                 disk->d_state = state;
2622                 g_mirror_update_metadata(disk);
2623                 G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2624                     sc->sc_name, g_mirror_get_diskname(disk));
2625                 break;
2626         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2627                 /*
2628                  * Possible scenarios:
2629                  * 1. Disk which needs synchronization was connected.
2630                  */
2631                 /* Previous state should be NEW. */
2632                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2633                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2634                     g_mirror_disk_state2str(disk->d_state)));
2635                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2636                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2637                     g_mirror_device_state2str(sc->sc_state),
2638                     g_mirror_get_diskname(disk),
2639                     g_mirror_disk_state2str(disk->d_state)));
2640                 DISK_STATE_CHANGED();
2641
2642                 if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2643                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2644                 disk->d_state = state;
2645                 if (sc->sc_provider != NULL) {
2646                         g_mirror_sync_start(disk);
2647                         g_mirror_update_metadata(disk);
2648                 }
2649                 break;
2650         case G_MIRROR_DISK_STATE_DISCONNECTED:
2651                 /*
2652                  * Possible scenarios:
2653                  * 1. Device wasn't running yet, but disk disappear.
2654                  * 2. Disk was active and disapppear.
2655                  * 3. Disk disappear during synchronization process.
2656                  */
2657                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2658                         /*
2659                          * Previous state should be ACTIVE, STALE or
2660                          * SYNCHRONIZING.
2661                          */
2662                         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2663                             disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2664                             disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2665                             ("Wrong disk state (%s, %s).",
2666                             g_mirror_get_diskname(disk),
2667                             g_mirror_disk_state2str(disk->d_state)));
2668                 } else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2669                         /* Previous state should be NEW. */
2670                         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2671                             ("Wrong disk state (%s, %s).",
2672                             g_mirror_get_diskname(disk),
2673                             g_mirror_disk_state2str(disk->d_state)));
2674                         /*
2675                          * Reset bumping syncid if disk disappeared in STARTING
2676                          * state.
2677                          */
2678                         if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2679                                 sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2680 #ifdef  INVARIANTS
2681                 } else {
2682                         KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2683                             sc->sc_name,
2684                             g_mirror_device_state2str(sc->sc_state),
2685                             g_mirror_get_diskname(disk),
2686                             g_mirror_disk_state2str(disk->d_state)));
2687 #endif
2688                 }
2689                 DISK_STATE_CHANGED();
2690                 G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2691                     sc->sc_name, g_mirror_get_diskname(disk));
2692
2693                 g_mirror_destroy_disk(disk);
2694                 break;
2695         case G_MIRROR_DISK_STATE_DESTROY:
2696             {
2697                 int error;
2698
2699                 error = g_mirror_clear_metadata(disk);
2700                 if (error != 0) {
2701                         G_MIRROR_DEBUG(0,
2702                             "Device %s: failed to clear metadata on %s: %d.",
2703                             sc->sc_name, g_mirror_get_diskname(disk), error);
2704                         break;
2705                 }
2706                 DISK_STATE_CHANGED();
2707                 G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2708                     sc->sc_name, g_mirror_get_diskname(disk));
2709
2710                 g_mirror_destroy_disk(disk);
2711                 sc->sc_ndisks--;
2712                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2713                         g_mirror_update_metadata(disk);
2714                 }
2715                 break;
2716             }
2717         default:
2718                 KASSERT(1 == 0, ("Unknown state (%u).", state));
2719                 break;
2720         }
2721         return (0);
2722 }
2723 #undef  DISK_STATE_CHANGED
2724
2725 int
2726 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2727 {
2728         struct g_provider *pp;
2729         u_char *buf;
2730         int error;
2731
2732         g_topology_assert();
2733
2734         error = g_access(cp, 1, 0, 0);
2735         if (error != 0)
2736                 return (error);
2737         pp = cp->provider;
2738         g_topology_unlock();
2739         /* Metadata are stored on last sector. */
2740         buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2741             &error);
2742         g_topology_lock();
2743         g_access(cp, -1, 0, 0);
2744         if (buf == NULL) {
2745                 G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2746                     cp->provider->name, error);
2747                 return (error);
2748         }
2749
2750         /* Decode metadata. */
2751         error = mirror_metadata_decode(buf, md);
2752         g_free(buf);
2753         if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2754                 return (EINVAL);
2755         if (md->md_version > G_MIRROR_VERSION) {
2756                 G_MIRROR_DEBUG(0,
2757                     "Kernel module is too old to handle metadata from %s.",
2758                     cp->provider->name);
2759                 return (EINVAL);
2760         }
2761         if (error != 0) {
2762                 G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2763                     cp->provider->name);
2764                 return (error);
2765         }
2766
2767         return (0);
2768 }
2769
2770 static int
2771 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2772     struct g_mirror_metadata *md)
2773 {
2774
2775         if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2776                 G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2777                     pp->name, md->md_did);
2778                 return (EEXIST);
2779         }
2780         if (md->md_all != sc->sc_ndisks) {
2781                 G_MIRROR_DEBUG(1,
2782                     "Invalid '%s' field on disk %s (device %s), skipping.",
2783                     "md_all", pp->name, sc->sc_name);
2784                 return (EINVAL);
2785         }
2786         if (md->md_slice != sc->sc_slice) {
2787                 G_MIRROR_DEBUG(1,
2788                     "Invalid '%s' field on disk %s (device %s), skipping.",
2789                     "md_slice", pp->name, sc->sc_name);
2790                 return (EINVAL);
2791         }
2792         if (md->md_balance != sc->sc_balance) {
2793                 G_MIRROR_DEBUG(1,
2794                     "Invalid '%s' field on disk %s (device %s), skipping.",
2795                     "md_balance", pp->name, sc->sc_name);
2796                 return (EINVAL);
2797         }
2798 #if 0
2799         if (md->md_mediasize != sc->sc_mediasize) {
2800                 G_MIRROR_DEBUG(1,
2801                     "Invalid '%s' field on disk %s (device %s), skipping.",
2802                     "md_mediasize", pp->name, sc->sc_name);
2803                 return (EINVAL);
2804         }
2805 #endif
2806         if (sc->sc_mediasize > pp->mediasize) {
2807                 G_MIRROR_DEBUG(1,
2808                     "Invalid size of disk %s (device %s), skipping.", pp->name,
2809                     sc->sc_name);
2810                 return (EINVAL);
2811         }
2812         if (md->md_sectorsize != sc->sc_sectorsize) {
2813                 G_MIRROR_DEBUG(1,
2814                     "Invalid '%s' field on disk %s (device %s), skipping.",
2815                     "md_sectorsize", pp->name, sc->sc_name);
2816                 return (EINVAL);
2817         }
2818         if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2819                 G_MIRROR_DEBUG(1,
2820                     "Invalid sector size of disk %s (device %s), skipping.",
2821                     pp->name, sc->sc_name);
2822                 return (EINVAL);
2823         }
2824         if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
2825                 G_MIRROR_DEBUG(1,
2826                     "Invalid device flags on disk %s (device %s), skipping.",
2827                     pp->name, sc->sc_name);
2828                 return (EINVAL);
2829         }
2830         if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
2831                 G_MIRROR_DEBUG(1,
2832                     "Invalid disk flags on disk %s (device %s), skipping.",
2833                     pp->name, sc->sc_name);
2834                 return (EINVAL);
2835         }
2836         return (0);
2837 }
2838
2839 int
2840 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
2841     struct g_mirror_metadata *md)
2842 {
2843         struct g_mirror_disk *disk;
2844         int error;
2845
2846         g_topology_assert_not();
2847         G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
2848
2849         error = g_mirror_check_metadata(sc, pp, md);
2850         if (error != 0)
2851                 return (error);
2852         if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING &&
2853             md->md_genid < sc->sc_genid) {
2854                 G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
2855                     pp->name, sc->sc_name);
2856                 return (EINVAL);
2857         }
2858         disk = g_mirror_init_disk(sc, pp, md, &error);
2859         if (disk == NULL)
2860                 return (error);
2861         error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
2862             G_MIRROR_EVENT_WAIT);
2863         if (error != 0)
2864                 return (error);
2865         if (md->md_version < G_MIRROR_VERSION) {
2866                 G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
2867                     pp->name, md->md_version, G_MIRROR_VERSION);
2868                 g_mirror_update_metadata(disk);
2869         }
2870         return (0);
2871 }
2872
2873 static void
2874 g_mirror_destroy_delayed(void *arg, int flag)
2875 {
2876         struct g_mirror_softc *sc;
2877         int error;
2878
2879         if (flag == EV_CANCEL) {
2880                 G_MIRROR_DEBUG(1, "Destroying canceled.");
2881                 return;
2882         }
2883         sc = arg;
2884         g_topology_unlock();
2885         sx_xlock(&sc->sc_lock);
2886         KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
2887             ("DESTROY flag set on %s.", sc->sc_name));
2888         KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0,
2889             ("DESTROYING flag not set on %s.", sc->sc_name));
2890         G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
2891         error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
2892         if (error != 0) {
2893                 G_MIRROR_DEBUG(0, "Cannot destroy %s (error=%d).",
2894                     sc->sc_name, error);
2895                 sx_xunlock(&sc->sc_lock);
2896         }
2897         g_topology_lock();
2898 }
2899
2900 static int
2901 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
2902 {
2903         struct g_mirror_softc *sc;
2904         int error = 0;
2905
2906         g_topology_assert();
2907         G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
2908             acw, ace);
2909
2910         sc = pp->private;
2911         KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
2912
2913         g_topology_unlock();
2914         sx_xlock(&sc->sc_lock);
2915         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
2916             (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0 ||
2917             LIST_EMPTY(&sc->sc_disks)) {
2918                 if (acr > 0 || acw > 0 || ace > 0)
2919                         error = ENXIO;
2920                 goto end;
2921         }
2922         sc->sc_provider_open += acr + acw + ace;
2923         if (pp->acw + acw == 0)
2924                 g_mirror_idle(sc, 0);
2925         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0 &&
2926             sc->sc_provider_open == 0)
2927                 g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK, sc, NULL);
2928 end:
2929         sx_xunlock(&sc->sc_lock);
2930         g_topology_lock();
2931         return (error);
2932 }
2933
2934 static struct g_geom *
2935 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md)
2936 {
2937         struct g_mirror_softc *sc;
2938         struct g_geom *gp;
2939         int error, timeout;
2940
2941         g_topology_assert();
2942         G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
2943             md->md_mid);
2944
2945         /* One disk is minimum. */
2946         if (md->md_all < 1)
2947                 return (NULL);
2948         /*
2949          * Action geom.
2950          */
2951         gp = g_new_geomf(mp, "%s", md->md_name);
2952         sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
2953         gp->start = g_mirror_start;
2954         gp->orphan = g_mirror_orphan;
2955         gp->access = g_mirror_access;
2956         gp->dumpconf = g_mirror_dumpconf;
2957
2958         sc->sc_id = md->md_mid;
2959         sc->sc_slice = md->md_slice;
2960         sc->sc_balance = md->md_balance;
2961         sc->sc_mediasize = md->md_mediasize;
2962         sc->sc_sectorsize = md->md_sectorsize;
2963         sc->sc_ndisks = md->md_all;
2964         sc->sc_flags = md->md_mflags;
2965         sc->sc_bump_id = 0;
2966         sc->sc_idle = 1;
2967         sc->sc_last_write = time_uptime;
2968         sc->sc_writes = 0;
2969         sc->sc_refcnt = 1;
2970         sx_init(&sc->sc_lock, "gmirror:lock");
2971         bioq_init(&sc->sc_queue);
2972         mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
2973         bioq_init(&sc->sc_regular_delayed);
2974         bioq_init(&sc->sc_inflight);
2975         bioq_init(&sc->sc_sync_delayed);
2976         LIST_INIT(&sc->sc_disks);
2977         TAILQ_INIT(&sc->sc_events);
2978         mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
2979         callout_init(&sc->sc_callout, 1);
2980         mtx_init(&sc->sc_done_mtx, "gmirror:done", NULL, MTX_DEF);
2981         sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
2982         gp->softc = sc;
2983         sc->sc_geom = gp;
2984         sc->sc_provider = NULL;
2985         sc->sc_provider_open = 0;
2986         /*
2987          * Synchronization geom.
2988          */
2989         gp = g_new_geomf(mp, "%s.sync", md->md_name);
2990         gp->softc = sc;
2991         gp->orphan = g_mirror_orphan;
2992         sc->sc_sync.ds_geom = gp;
2993         sc->sc_sync.ds_ndisks = 0;
2994         error = kproc_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
2995             "g_mirror %s", md->md_name);
2996         if (error != 0) {
2997                 G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
2998                     sc->sc_name);
2999                 g_destroy_geom(sc->sc_sync.ds_geom);
3000                 g_destroy_geom(sc->sc_geom);
3001                 g_mirror_free_device(sc);
3002                 return (NULL);
3003         }
3004
3005         G_MIRROR_DEBUG(1, "Device %s created (%u components, id=%u).",
3006             sc->sc_name, sc->sc_ndisks, sc->sc_id);
3007
3008         sc->sc_rootmount = root_mount_hold("GMIRROR");
3009         G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3010         /*
3011          * Run timeout.
3012          */
3013         timeout = g_mirror_timeout * hz;
3014         callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
3015         return (sc->sc_geom);
3016 }
3017
3018 int
3019 g_mirror_destroy(struct g_mirror_softc *sc, int how)
3020 {
3021         struct g_mirror_disk *disk;
3022
3023         g_topology_assert_not();
3024         sx_assert(&sc->sc_lock, SX_XLOCKED);
3025
3026         if (sc->sc_provider_open != 0) {
3027                 switch (how) {
3028                 case G_MIRROR_DESTROY_SOFT:
3029                         G_MIRROR_DEBUG(1,
3030                             "Device %s is still open (%d).", sc->sc_name,
3031                             sc->sc_provider_open);
3032                         return (EBUSY);
3033                 case G_MIRROR_DESTROY_DELAYED:
3034                         G_MIRROR_DEBUG(1,
3035                             "Device %s will be destroyed on last close.",
3036                             sc->sc_name);
3037                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
3038                                 if (disk->d_state ==
3039                                     G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3040                                         g_mirror_sync_stop(disk, 1);
3041                                 }
3042                         }
3043                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROYING;
3044                         return (EBUSY);
3045                 case G_MIRROR_DESTROY_HARD:
3046                         G_MIRROR_DEBUG(1, "Device %s is still open, so it "
3047                             "can't be definitely removed.", sc->sc_name);
3048                 }
3049         }
3050
3051         g_topology_lock();
3052         if (sc->sc_geom->softc == NULL) {
3053                 g_topology_unlock();
3054                 return (0);
3055         }
3056         sc->sc_geom->softc = NULL;
3057         sc->sc_sync.ds_geom->softc = NULL;
3058         g_topology_unlock();
3059
3060         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
3061         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_WAIT;
3062         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3063         sx_xunlock(&sc->sc_lock);
3064         mtx_lock(&sc->sc_queue_mtx);
3065         wakeup(sc);
3066         mtx_unlock(&sc->sc_queue_mtx);
3067         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3068         while (sc->sc_worker != NULL)
3069                 tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
3070         G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3071         sx_xlock(&sc->sc_lock);
3072         g_mirror_destroy_device(sc);
3073         return (0);
3074 }
3075
3076 static void
3077 g_mirror_taste_orphan(struct g_consumer *cp)
3078 {
3079
3080         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3081             cp->provider->name));
3082 }
3083
3084 static struct g_geom *
3085 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3086 {
3087         struct g_mirror_metadata md;
3088         struct g_mirror_softc *sc;
3089         struct g_consumer *cp;
3090         struct g_geom *gp;
3091         int error;
3092
3093         g_topology_assert();
3094         g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3095         G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
3096
3097         gp = g_new_geomf(mp, "mirror:taste");
3098         /*
3099          * This orphan function should be never called.
3100          */
3101         gp->orphan = g_mirror_taste_orphan;
3102         cp = g_new_consumer(gp);
3103         g_attach(cp, pp);
3104         error = g_mirror_read_metadata(cp, &md);
3105         g_detach(cp);
3106         g_destroy_consumer(cp);
3107         g_destroy_geom(gp);
3108         if (error != 0)
3109                 return (NULL);
3110         gp = NULL;
3111
3112         if (md.md_provider[0] != '\0' &&
3113             !g_compare_names(md.md_provider, pp->name))
3114                 return (NULL);
3115         if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3116                 return (NULL);
3117         if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
3118                 G_MIRROR_DEBUG(0,
3119                     "Device %s: provider %s marked as inactive, skipping.",
3120                     md.md_name, pp->name);
3121                 return (NULL);
3122         }
3123         if (g_mirror_debug >= 2)
3124                 mirror_metadata_dump(&md);
3125
3126         /*
3127          * Let's check if device already exists.
3128          */
3129         sc = NULL;
3130         LIST_FOREACH(gp, &mp->geom, geom) {
3131                 sc = gp->softc;
3132                 if (sc == NULL)
3133                         continue;
3134                 if (sc->sc_sync.ds_geom == gp)
3135                         continue;
3136                 if (strcmp(md.md_name, sc->sc_name) != 0)
3137                         continue;
3138                 if (md.md_mid != sc->sc_id) {
3139                         G_MIRROR_DEBUG(0, "Device %s already configured.",
3140                             sc->sc_name);
3141                         return (NULL);
3142                 }
3143                 break;
3144         }
3145         if (gp == NULL) {
3146                 gp = g_mirror_create(mp, &md);
3147                 if (gp == NULL) {
3148                         G_MIRROR_DEBUG(0, "Cannot create device %s.",
3149                             md.md_name);
3150                         return (NULL);
3151                 }
3152                 sc = gp->softc;
3153         }
3154         G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3155         g_topology_unlock();
3156         sx_xlock(&sc->sc_lock);
3157         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_TASTING;
3158         error = g_mirror_add_disk(sc, pp, &md);
3159         if (error != 0) {
3160                 G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3161                     pp->name, gp->name, error);
3162                 if (LIST_EMPTY(&sc->sc_disks)) {
3163                         g_cancel_event(sc);
3164                         g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3165                         g_topology_lock();
3166                         return (NULL);
3167                 }
3168                 gp = NULL;
3169         }
3170         sc->sc_flags &= ~G_MIRROR_DEVICE_FLAG_TASTING;
3171         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
3172                 g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3173                 g_topology_lock();
3174                 return (NULL);
3175         }
3176         sx_xunlock(&sc->sc_lock);
3177         g_topology_lock();
3178         return (gp);
3179 }
3180
3181 static void
3182 g_mirror_resize(struct g_consumer *cp)
3183 {
3184         struct g_mirror_disk *disk;
3185
3186         g_topology_assert();
3187         g_trace(G_T_TOPOLOGY, "%s(%s)", __func__, cp->provider->name);
3188
3189         disk = cp->private;
3190         if (disk == NULL)
3191                 return;
3192         g_topology_unlock();
3193         g_mirror_update_metadata(disk);
3194         g_topology_lock();
3195 }
3196
3197 static int
3198 g_mirror_destroy_geom(struct gctl_req *req __unused,
3199     struct g_class *mp __unused, struct g_geom *gp)
3200 {
3201         struct g_mirror_softc *sc;
3202         int error;
3203
3204         g_topology_unlock();
3205         sc = gp->softc;
3206         sx_xlock(&sc->sc_lock);
3207         g_cancel_event(sc);
3208         error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3209         if (error != 0)
3210                 sx_xunlock(&sc->sc_lock);
3211         g_topology_lock();
3212         return (error);
3213 }
3214
3215 static void
3216 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3217     struct g_consumer *cp, struct g_provider *pp)
3218 {
3219         struct g_mirror_softc *sc;
3220
3221         g_topology_assert();
3222
3223         sc = gp->softc;
3224         if (sc == NULL)
3225                 return;
3226         /* Skip synchronization geom. */
3227         if (gp == sc->sc_sync.ds_geom)
3228                 return;
3229         if (pp != NULL) {
3230                 /* Nothing here. */
3231         } else if (cp != NULL) {
3232                 struct g_mirror_disk *disk;
3233
3234                 disk = cp->private;
3235                 if (disk == NULL)
3236                         return;
3237                 g_topology_unlock();
3238                 sx_xlock(&sc->sc_lock);
3239                 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3240                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3241                         sbuf_printf(sb, "%s<Synchronized>", indent);
3242                         if (disk->d_sync.ds_offset == 0)
3243                                 sbuf_printf(sb, "0%%");
3244                         else {
3245                                 sbuf_printf(sb, "%u%%",
3246                                     (u_int)((disk->d_sync.ds_offset * 100) /
3247                                     sc->sc_provider->mediasize));
3248                         }
3249                         sbuf_printf(sb, "</Synchronized>\n");
3250                         if (disk->d_sync.ds_offset > 0) {
3251                                 sbuf_printf(sb, "%s<BytesSynced>%jd"
3252                                     "</BytesSynced>\n", indent,
3253                                     (intmax_t)disk->d_sync.ds_offset);
3254                         }
3255                 }
3256                 sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3257                     disk->d_sync.ds_syncid);
3258                 sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3259                     disk->d_genid);
3260                 sbuf_printf(sb, "%s<Flags>", indent);
3261                 if (disk->d_flags == 0)
3262                         sbuf_printf(sb, "NONE");
3263                 else {
3264                         int first = 1;
3265
3266 #define ADD_FLAG(flag, name)    do {                                    \
3267         if ((disk->d_flags & (flag)) != 0) {                            \
3268                 if (!first)                                             \
3269                         sbuf_printf(sb, ", ");                          \
3270                 else                                                    \
3271                         first = 0;                                      \
3272                 sbuf_printf(sb, name);                                  \
3273         }                                                               \
3274 } while (0)
3275                         ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3276                         ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3277                         ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3278                         ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3279                             "SYNCHRONIZING");
3280                         ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3281                         ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3282 #undef  ADD_FLAG
3283                 }
3284                 sbuf_printf(sb, "</Flags>\n");
3285                 sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3286                     disk->d_priority);
3287                 sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3288                     g_mirror_disk_state2str(disk->d_state));
3289                 sx_xunlock(&sc->sc_lock);
3290                 g_topology_lock();
3291         } else {
3292                 g_topology_unlock();
3293                 sx_xlock(&sc->sc_lock);
3294                 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3295                 sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3296                 sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3297                 sbuf_printf(sb, "%s<Flags>", indent);
3298                 if (sc->sc_flags == 0)
3299                         sbuf_printf(sb, "NONE");
3300                 else {
3301                         int first = 1;
3302
3303 #define ADD_FLAG(flag, name)    do {                                    \
3304         if ((sc->sc_flags & (flag)) != 0) {                             \
3305                 if (!first)                                             \
3306                         sbuf_printf(sb, ", ");                          \
3307                 else                                                    \
3308                         first = 0;                                      \
3309                 sbuf_printf(sb, name);                                  \
3310         }                                                               \
3311 } while (0)
3312                         ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3313                         ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3314 #undef  ADD_FLAG
3315                 }
3316                 sbuf_printf(sb, "</Flags>\n");
3317                 sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3318                     (u_int)sc->sc_slice);
3319                 sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3320                     balance_name(sc->sc_balance));
3321                 sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3322                     sc->sc_ndisks);
3323                 sbuf_printf(sb, "%s<State>", indent);
3324                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3325                         sbuf_printf(sb, "%s", "STARTING");
3326                 else if (sc->sc_ndisks ==
3327                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3328                         sbuf_printf(sb, "%s", "COMPLETE");
3329                 else
3330                         sbuf_printf(sb, "%s", "DEGRADED");
3331                 sbuf_printf(sb, "</State>\n");
3332                 sx_xunlock(&sc->sc_lock);
3333                 g_topology_lock();
3334         }
3335 }
3336
3337 static void
3338 g_mirror_shutdown_post_sync(void *arg, int howto)
3339 {
3340         struct g_class *mp;
3341         struct g_geom *gp, *gp2;
3342         struct g_mirror_softc *sc;
3343         int error;
3344
3345         mp = arg;
3346         DROP_GIANT();
3347         g_topology_lock();
3348         g_mirror_shutdown = 1;
3349         LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3350                 if ((sc = gp->softc) == NULL)
3351                         continue;
3352                 /* Skip synchronization geom. */
3353                 if (gp == sc->sc_sync.ds_geom)
3354                         continue;
3355                 g_topology_unlock();
3356                 sx_xlock(&sc->sc_lock);
3357                 g_mirror_idle(sc, -1);
3358                 g_cancel_event(sc);
3359                 error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3360                 if (error != 0)
3361                         sx_xunlock(&sc->sc_lock);
3362                 g_topology_lock();
3363         }
3364         g_topology_unlock();
3365         PICKUP_GIANT();
3366 }
3367
3368 static void
3369 g_mirror_init(struct g_class *mp)
3370 {
3371
3372         g_mirror_post_sync = EVENTHANDLER_REGISTER(shutdown_post_sync,
3373             g_mirror_shutdown_post_sync, mp, SHUTDOWN_PRI_FIRST);
3374         if (g_mirror_post_sync == NULL)
3375                 G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3376 }
3377
3378 static void
3379 g_mirror_fini(struct g_class *mp)
3380 {
3381
3382         if (g_mirror_post_sync != NULL)
3383                 EVENTHANDLER_DEREGISTER(shutdown_post_sync, g_mirror_post_sync);
3384 }
3385
3386 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);