sys/geom/mirror/g_mirror.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/param.h>
  33 #include <sys/systm.h>
  34 #include <sys/bio.h>
  35 #include <sys/eventhandler.h>
  36 #include <sys/fail.h>
  37 #include <sys/kernel.h>
  38 #include <sys/kthread.h>
  39 #include <sys/limits.h>
  40 #include <sys/lock.h>
  41 #include <sys/malloc.h>
  42 #include <sys/mutex.h>
  43 #include <sys/proc.h>
  44 #include <sys/sbuf.h>
  45 #include <sys/sched.h>
  46 #include <sys/sx.h>
  47 #include <sys/sysctl.h>
  48
  49 #include <geom/geom.h>
  50 #include <geom/geom_dbg.h>
  51 #include <geom/mirror/g_mirror.h>
  52
  53 FEATURE(geom_mirror, "GEOM mirroring support");
  54
  55 static MALLOC_DEFINE(M_MIRROR, "mirror_data", "GEOM_MIRROR Data");
  56
  57 SYSCTL_DECL(_kern_geom);
  58 static SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
  59     "GEOM_MIRROR stuff");
  60 int g_mirror_debug = 0;
  61 SYSCTL_INT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RWTUN, &g_mirror_debug, 0,
  62     "Debug level");
  63 bool g_launch_mirror_before_timeout = true;
  64 SYSCTL_BOOL(_kern_geom_mirror, OID_AUTO, launch_mirror_before_timeout,
  65     CTLFLAG_RWTUN, &g_launch_mirror_before_timeout, 0,
  66     "If false, force gmirror to wait out the full kern.geom.mirror.timeout "
  67     "before launching mirrors");
  68 static u_int g_mirror_timeout = 4;
  69 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RWTUN, &g_mirror_timeout,
  70     0, "Time to wait on all mirror components");
  71 static u_int g_mirror_idletime = 5;
  72 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, idletime, CTLFLAG_RWTUN,
  73     &g_mirror_idletime, 0, "Mark components as clean when idling");
  74 static u_int g_mirror_disconnect_on_failure = 1;
  75 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, disconnect_on_failure, CTLFLAG_RWTUN,
  76     &g_mirror_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
  77 static u_int g_mirror_syncreqs = 2;
  78 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
  79     &g_mirror_syncreqs, 0, "Parallel synchronization I/O requests.");
  80 static u_int g_mirror_sync_period = 5;
  81 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_update_period, CTLFLAG_RWTUN,
  82     &g_mirror_sync_period, 0,
  83     "Metadata update period during synchronization, in seconds");
  84
  85 #define MSLEEP(ident, mtx, priority, wmesg, timeout)    do {            \
  86         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));       \
  87         msleep((ident), (mtx), (priority), (wmesg), (timeout));         \
  88         G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, (ident));       \
  89 } while (0)
  90
  91 static eventhandler_tag g_mirror_post_sync = NULL;
  92 static int g_mirror_shutdown = 0;
  93
  94 static g_ctl_destroy_geom_t g_mirror_destroy_geom;
  95 static g_taste_t g_mirror_taste;
  96 static g_init_t g_mirror_init;
  97 static g_fini_t g_mirror_fini;
  98 static g_provgone_t g_mirror_providergone;
  99 static g_resize_t g_mirror_resize;
 100
 101 struct g_class g_mirror_class = {
 102         .name = G_MIRROR_CLASS_NAME,
 103         .version = G_VERSION,
 104         .ctlreq = g_mirror_config,
 105         .taste = g_mirror_taste,
 106         .destroy_geom = g_mirror_destroy_geom,
 107         .init = g_mirror_init,
 108         .fini = g_mirror_fini,
 109         .providergone = g_mirror_providergone,
 110         .resize = g_mirror_resize
 111 };
 112
 113
 114 static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
 115 static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
 116 static void g_mirror_update_device(struct g_mirror_softc *sc, bool force);
 117 static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
 118     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
 119 static int g_mirror_refresh_device(struct g_mirror_softc *sc,
 120     const struct g_provider *pp, const struct g_mirror_metadata *md);
 121 static void g_mirror_sync_reinit(const struct g_mirror_disk *disk,
 122     struct bio *bp, off_t offset);
 123 static void g_mirror_sync_stop(struct g_mirror_disk *disk, int type);
 124 static void g_mirror_register_request(struct g_mirror_softc *sc,
 125     struct bio *bp);
 126 static void g_mirror_sync_release(struct g_mirror_softc *sc);
 127
 128
 129 static const char *
 130 g_mirror_disk_state2str(int state)
 131 {
 132
 133         switch (state) {
 134         case G_MIRROR_DISK_STATE_NONE:
 135                 return ("NONE");
 136         case G_MIRROR_DISK_STATE_NEW:
 137                 return ("NEW");
 138         case G_MIRROR_DISK_STATE_ACTIVE:
 139                 return ("ACTIVE");
 140         case G_MIRROR_DISK_STATE_STALE:
 141                 return ("STALE");
 142         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
 143                 return ("SYNCHRONIZING");
 144         case G_MIRROR_DISK_STATE_DISCONNECTED:
 145                 return ("DISCONNECTED");
 146         case G_MIRROR_DISK_STATE_DESTROY:
 147                 return ("DESTROY");
 148         default:
 149                 return ("INVALID");
 150         }
 151 }
 152
 153 static const char *
 154 g_mirror_device_state2str(int state)
 155 {
 156
 157         switch (state) {
 158         case G_MIRROR_DEVICE_STATE_STARTING:
 159                 return ("STARTING");
 160         case G_MIRROR_DEVICE_STATE_RUNNING:
 161                 return ("RUNNING");
 162         default:
 163                 return ("INVALID");
 164         }
 165 }
 166
 167 static const char *
 168 g_mirror_get_diskname(struct g_mirror_disk *disk)
 169 {
 170
 171         if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
 172                 return ("[unknown]");
 173         return (disk->d_name);
 174 }
 175
 176 /*
 177  * --- Events handling functions ---
 178  * Events in geom_mirror are used to maintain disks and device status
 179  * from one thread to simplify locking.
 180  */
 181 static void
 182 g_mirror_event_free(struct g_mirror_event *ep)
 183 {
 184
 185         free(ep, M_MIRROR);
 186 }
 187
 188 int
 189 g_mirror_event_send(void *arg, int state, int flags)
 190 {
 191         struct g_mirror_softc *sc;
 192         struct g_mirror_disk *disk;
 193         struct g_mirror_event *ep;
 194         int error;
 195
 196         ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
 197         G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
 198         if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
 199                 disk = NULL;
 200                 sc = arg;
 201         } else {
 202                 disk = arg;
 203                 sc = disk->d_softc;
 204         }
 205         ep->e_disk = disk;
 206         ep->e_state = state;
 207         ep->e_flags = flags;
 208         ep->e_error = 0;
 209         mtx_lock(&sc->sc_events_mtx);
 210         TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
 211         mtx_unlock(&sc->sc_events_mtx);
 212         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
 213         mtx_lock(&sc->sc_queue_mtx);
 214         wakeup(sc);
 215         mtx_unlock(&sc->sc_queue_mtx);
 216         if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
 217                 return (0);
 218         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
 219         sx_xunlock(&sc->sc_lock);
 220         while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
 221                 mtx_lock(&sc->sc_events_mtx);
 222                 MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "m:event",
 223                     hz * 5);
 224         }
 225         error = ep->e_error;
 226         g_mirror_event_free(ep);
 227         sx_xlock(&sc->sc_lock);
 228         return (error);
 229 }
 230
 231 static struct g_mirror_event *
 232 g_mirror_event_first(struct g_mirror_softc *sc)
 233 {
 234         struct g_mirror_event *ep;
 235
 236         mtx_lock(&sc->sc_events_mtx);
 237         ep = TAILQ_FIRST(&sc->sc_events);
 238         mtx_unlock(&sc->sc_events_mtx);
 239         return (ep);
 240 }
 241
 242 static void
 243 g_mirror_event_remove(struct g_mirror_softc *sc, struct g_mirror_event *ep)
 244 {
 245
 246         mtx_lock(&sc->sc_events_mtx);
 247         TAILQ_REMOVE(&sc->sc_events, ep, e_next);
 248         mtx_unlock(&sc->sc_events_mtx);
 249 }
 250
 251 static void
 252 g_mirror_event_cancel(struct g_mirror_disk *disk)
 253 {
 254         struct g_mirror_softc *sc;
 255         struct g_mirror_event *ep, *tmpep;
 256
 257         sc = disk->d_softc;
 258         sx_assert(&sc->sc_lock, SX_XLOCKED);
 259
 260         mtx_lock(&sc->sc_events_mtx);
 261         TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
 262                 if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
 263                         continue;
 264                 if (ep->e_disk != disk)
 265                         continue;
 266                 TAILQ_REMOVE(&sc->sc_events, ep, e_next);
 267                 if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
 268                         g_mirror_event_free(ep);
 269                 else {
 270                         ep->e_error = ECANCELED;
 271                         wakeup(ep);
 272                 }
 273         }
 274         mtx_unlock(&sc->sc_events_mtx);
 275 }
 276
 277 /*
 278  * Return the number of disks in given state.
 279  * If state is equal to -1, count all connected disks.
 280  */
 281 u_int
 282 g_mirror_ndisks(struct g_mirror_softc *sc, int state)
 283 {
 284         struct g_mirror_disk *disk;
 285         u_int n = 0;
 286
 287         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 288                 if (state == -1 || disk->d_state == state)
 289                         n++;
 290         }
 291         return (n);
 292 }
 293
 294 /*
 295  * Find a disk in mirror by its disk ID.
 296  */
 297 static struct g_mirror_disk *
 298 g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
 299 {
 300         struct g_mirror_disk *disk;
 301
 302         sx_assert(&sc->sc_lock, SX_XLOCKED);
 303
 304         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 305                 if (disk->d_id == id)
 306                         return (disk);
 307         }
 308         return (NULL);
 309 }
 310
 311 static u_int
 312 g_mirror_nrequests(struct g_mirror_softc *sc, struct g_consumer *cp)
 313 {
 314         struct bio *bp;
 315         u_int nreqs = 0;
 316
 317         mtx_lock(&sc->sc_queue_mtx);
 318         TAILQ_FOREACH(bp, &sc->sc_queue, bio_queue) {
 319                 if (bp->bio_from == cp)
 320                         nreqs++;
 321         }
 322         mtx_unlock(&sc->sc_queue_mtx);
 323         return (nreqs);
 324 }
 325
 326 static int
 327 g_mirror_is_busy(struct g_mirror_softc *sc, struct g_consumer *cp)
 328 {
 329
 330         if (cp->index > 0) {
 331                 G_MIRROR_DEBUG(2,
 332                     "I/O requests for %s exist, can't destroy it now.",
 333                     cp->provider->name);
 334                 return (1);
 335         }
 336         if (g_mirror_nrequests(sc, cp) > 0) {
 337                 G_MIRROR_DEBUG(2,
 338                     "I/O requests for %s in queue, can't destroy it now.",
 339                     cp->provider->name);
 340                 return (1);
 341         }
 342         return (0);
 343 }
 344
 345 static void
 346 g_mirror_destroy_consumer(void *arg, int flags __unused)
 347 {
 348         struct g_consumer *cp;
 349
 350         g_topology_assert();
 351
 352         cp = arg;
 353         G_MIRROR_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
 354         g_detach(cp);
 355         g_destroy_consumer(cp);
 356 }
 357
 358 static void
 359 g_mirror_kill_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
 360 {
 361         struct g_provider *pp;
 362         int retaste_wait;
 363
 364         g_topology_assert();
 365
 366         cp->private = NULL;
 367         if (g_mirror_is_busy(sc, cp))
 368                 return;
 369         pp = cp->provider;
 370         retaste_wait = 0;
 371         if (cp->acw == 1) {
 372                 if ((pp->geom->flags & G_GEOM_WITHER) == 0)
 373                         retaste_wait = 1;
 374         }
 375         G_MIRROR_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
 376             -cp->acw, -cp->ace, 0);
 377         if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
 378                 g_access(cp, -cp->acr, -cp->acw, -cp->ace);
 379         if (retaste_wait) {
 380                 /*
 381                  * After retaste event was send (inside g_access()), we can send
 382                  * event to detach and destroy consumer.
 383                  * A class, which has consumer to the given provider connected
 384                  * will not receive retaste event for the provider.
 385                  * This is the way how I ignore retaste events when I close
 386                  * consumers opened for write: I detach and destroy consumer
 387                  * after retaste event is sent.
 388                  */
 389                 g_post_event(g_mirror_destroy_consumer, cp, M_WAITOK, NULL);
 390                 return;
 391         }
 392         G_MIRROR_DEBUG(1, "Consumer %s destroyed.", pp->name);
 393         g_detach(cp);
 394         g_destroy_consumer(cp);
 395 }
 396
 397 static int
 398 g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
 399 {
 400         struct g_consumer *cp;
 401         int error;
 402
 403         g_topology_assert_not();
 404         KASSERT(disk->d_consumer == NULL,
 405             ("Disk already connected (device %s).", disk->d_softc->sc_name));
 406
 407         g_topology_lock();
 408         cp = g_new_consumer(disk->d_softc->sc_geom);
 409         cp->flags |= G_CF_DIRECT_RECEIVE;
 410         error = g_attach(cp, pp);
 411         if (error != 0) {
 412                 g_destroy_consumer(cp);
 413                 g_topology_unlock();
 414                 return (error);
 415         }
 416         error = g_access(cp, 1, 1, 1);
 417         if (error != 0) {
 418                 g_detach(cp);
 419                 g_destroy_consumer(cp);
 420                 g_topology_unlock();
 421                 G_MIRROR_DEBUG(0, "Cannot open consumer %s (error=%d).",
 422                     pp->name, error);
 423                 return (error);
 424         }
 425         g_topology_unlock();
 426         disk->d_consumer = cp;
 427         disk->d_consumer->private = disk;
 428         disk->d_consumer->index = 0;
 429
 430         G_MIRROR_DEBUG(2, "Disk %s connected.", g_mirror_get_diskname(disk));
 431         return (0);
 432 }
 433
 434 static void
 435 g_mirror_disconnect_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
 436 {
 437
 438         g_topology_assert();
 439
 440         if (cp == NULL)
 441                 return;
 442         if (cp->provider != NULL)
 443                 g_mirror_kill_consumer(sc, cp);
 444         else
 445                 g_destroy_consumer(cp);
 446 }
 447
 448 /*
 449  * Initialize disk. This means allocate memory, create consumer, attach it
 450  * to the provider and open access (r1w1e1) to it.
 451  */
 452 static struct g_mirror_disk *
 453 g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
 454     struct g_mirror_metadata *md, int *errorp)
 455 {
 456         struct g_mirror_disk *disk;
 457         int i, error;
 458
 459         disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
 460         if (disk == NULL) {
 461                 error = ENOMEM;
 462                 goto fail;
 463         }
 464         disk->d_softc = sc;
 465         error = g_mirror_connect_disk(disk, pp);
 466         if (error != 0)
 467                 goto fail;
 468         disk->d_id = md->md_did;
 469         disk->d_state = G_MIRROR_DISK_STATE_NONE;
 470         disk->d_priority = md->md_priority;
 471         disk->d_flags = md->md_dflags;
 472         error = g_getattr("GEOM::candelete", disk->d_consumer, &i);
 473         if (error == 0 && i != 0)
 474                 disk->d_flags |= G_MIRROR_DISK_FLAG_CANDELETE;
 475         if (md->md_provider[0] != '\0')
 476                 disk->d_flags |= G_MIRROR_DISK_FLAG_HARDCODED;
 477         disk->d_sync.ds_consumer = NULL;
 478         disk->d_sync.ds_offset = md->md_sync_offset;
 479         disk->d_sync.ds_offset_done = md->md_sync_offset;
 480         disk->d_sync.ds_update_ts = time_uptime;
 481         disk->d_genid = md->md_genid;
 482         disk->d_sync.ds_syncid = md->md_syncid;
 483         disk->d_init_ndisks = md->md_all;
 484         disk->d_init_slice = md->md_slice;
 485         disk->d_init_balance = md->md_balance;
 486         disk->d_init_mediasize = md->md_mediasize;
 487         if (errorp != NULL)
 488                 *errorp = 0;
 489         return (disk);
 490 fail:
 491         if (errorp != NULL)
 492                 *errorp = error;
 493         if (disk != NULL)
 494                 free(disk, M_MIRROR);
 495         return (NULL);
 496 }
 497
 498 static void
 499 g_mirror_destroy_disk(struct g_mirror_disk *disk)
 500 {
 501         struct g_mirror_softc *sc;
 502
 503         g_topology_assert_not();
 504         sc = disk->d_softc;
 505         sx_assert(&sc->sc_lock, SX_XLOCKED);
 506
 507         g_topology_lock();
 508         LIST_REMOVE(disk, d_next);
 509         g_topology_unlock();
 510         g_mirror_event_cancel(disk);
 511         if (sc->sc_hint == disk)
 512                 sc->sc_hint = NULL;
 513         switch (disk->d_state) {
 514         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
 515                 g_mirror_sync_stop(disk, 1);
 516                 /* FALLTHROUGH */
 517         case G_MIRROR_DISK_STATE_NEW:
 518         case G_MIRROR_DISK_STATE_STALE:
 519         case G_MIRROR_DISK_STATE_ACTIVE:
 520                 g_topology_lock();
 521                 g_mirror_disconnect_consumer(sc, disk->d_consumer);
 522                 g_topology_unlock();
 523                 free(disk, M_MIRROR);
 524                 break;
 525         default:
 526                 KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
 527                     g_mirror_get_diskname(disk),
 528                     g_mirror_disk_state2str(disk->d_state)));
 529         }
 530 }
 531
 532 static void
 533 g_mirror_free_device(struct g_mirror_softc *sc)
 534 {
 535
 536         g_topology_assert();
 537
 538         mtx_destroy(&sc->sc_queue_mtx);
 539         mtx_destroy(&sc->sc_events_mtx);
 540         mtx_destroy(&sc->sc_done_mtx);
 541         sx_destroy(&sc->sc_lock);
 542         free(sc, M_MIRROR);
 543 }
 544
 545 static void
 546 g_mirror_providergone(struct g_provider *pp)
 547 {
 548         struct g_mirror_softc *sc = pp->private;
 549
 550         if ((--sc->sc_refcnt) == 0)
 551                 g_mirror_free_device(sc);
 552 }
 553
 554 static void
 555 g_mirror_destroy_device(struct g_mirror_softc *sc)
 556 {
 557         struct g_mirror_disk *disk;
 558         struct g_mirror_event *ep;
 559         struct g_geom *gp;
 560         struct g_consumer *cp, *tmpcp;
 561
 562         g_topology_assert_not();
 563         sx_assert(&sc->sc_lock, SX_XLOCKED);
 564
 565         gp = sc->sc_geom;
 566         if (sc->sc_provider != NULL)
 567                 g_mirror_destroy_provider(sc);
 568         for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
 569             disk = LIST_FIRST(&sc->sc_disks)) {
 570                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
 571                 g_mirror_update_metadata(disk);
 572                 g_mirror_destroy_disk(disk);
 573         }
 574         while ((ep = g_mirror_event_first(sc)) != NULL) {
 575                 g_mirror_event_remove(sc, ep);
 576                 if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
 577                         g_mirror_event_free(ep);
 578                 else {
 579                         ep->e_error = ECANCELED;
 580                         ep->e_flags |= G_MIRROR_EVENT_DONE;
 581                         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
 582                         mtx_lock(&sc->sc_events_mtx);
 583                         wakeup(ep);
 584                         mtx_unlock(&sc->sc_events_mtx);
 585                 }
 586         }
 587         callout_drain(&sc->sc_callout);
 588
 589         g_topology_lock();
 590         LIST_FOREACH_SAFE(cp, &sc->sc_sync.ds_geom->consumer, consumer, tmpcp) {
 591                 g_mirror_disconnect_consumer(sc, cp);
 592         }
 593         g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
 594         G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
 595         g_wither_geom(gp, ENXIO);
 596         sx_xunlock(&sc->sc_lock);
 597         if ((--sc->sc_refcnt) == 0)
 598                 g_mirror_free_device(sc);
 599         g_topology_unlock();
 600 }
 601
 602 static void
 603 g_mirror_orphan(struct g_consumer *cp)
 604 {
 605         struct g_mirror_disk *disk;
 606
 607         g_topology_assert();
 608
 609         disk = cp->private;
 610         if (disk == NULL)
 611                 return;
 612         disk->d_softc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
 613         g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
 614             G_MIRROR_EVENT_DONTWAIT);
 615 }
 616
 617 /*
 618  * Function should return the next active disk on the list.
 619  * It is possible that it will be the same disk as given.
 620  * If there are no active disks on list, NULL is returned.
 621  */
 622 static __inline struct g_mirror_disk *
 623 g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
 624 {
 625         struct g_mirror_disk *dp;
 626
 627         for (dp = LIST_NEXT(disk, d_next); dp != disk;
 628             dp = LIST_NEXT(dp, d_next)) {
 629                 if (dp == NULL)
 630                         dp = LIST_FIRST(&sc->sc_disks);
 631                 if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
 632                         break;
 633         }
 634         if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
 635                 return (NULL);
 636         return (dp);
 637 }
 638
 639 static struct g_mirror_disk *
 640 g_mirror_get_disk(struct g_mirror_softc *sc)
 641 {
 642         struct g_mirror_disk *disk;
 643
 644         if (sc->sc_hint == NULL) {
 645                 sc->sc_hint = LIST_FIRST(&sc->sc_disks);
 646                 if (sc->sc_hint == NULL)
 647                         return (NULL);
 648         }
 649         disk = sc->sc_hint;
 650         if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
 651                 disk = g_mirror_find_next(sc, disk);
 652                 if (disk == NULL)
 653                         return (NULL);
 654         }
 655         sc->sc_hint = g_mirror_find_next(sc, disk);
 656         return (disk);
 657 }
 658
 659 static int
 660 g_mirror_write_metadata(struct g_mirror_disk *disk,
 661     struct g_mirror_metadata *md)
 662 {
 663         struct g_mirror_softc *sc;
 664         struct g_consumer *cp;
 665         off_t offset, length;
 666         u_char *sector;
 667         int error = 0;
 668
 669         g_topology_assert_not();
 670         sc = disk->d_softc;
 671         sx_assert(&sc->sc_lock, SX_LOCKED);
 672
 673         cp = disk->d_consumer;
 674         KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
 675         KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
 676         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
 677             ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
 678             cp->acw, cp->ace));
 679         length = cp->provider->sectorsize;
 680         offset = cp->provider->mediasize - length;
 681         sector = malloc((size_t)length, M_MIRROR, M_WAITOK | M_ZERO);
 682         if (md != NULL &&
 683             (sc->sc_flags & G_MIRROR_DEVICE_FLAG_WIPE) == 0) {
 684                 /*
 685                  * Handle the case, when the size of parent provider reduced.
 686                  */
 687                 if (offset < md->md_mediasize)
 688                         error = ENOSPC;
 689                 else
 690                         mirror_metadata_encode(md, sector);
 691         }
 692         KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_metadata_write, error);
 693         if (error == 0)
 694                 error = g_write_data(cp, offset, sector, length);
 695         free(sector, M_MIRROR);
 696         if (error != 0) {
 697                 if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
 698                         disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
 699                         G_MIRROR_DEBUG(0, "Cannot write metadata on %s "
 700                             "(device=%s, error=%d).",
 701                             g_mirror_get_diskname(disk), sc->sc_name, error);
 702                 } else {
 703                         G_MIRROR_DEBUG(1, "Cannot write metadata on %s "
 704                             "(device=%s, error=%d).",
 705                             g_mirror_get_diskname(disk), sc->sc_name, error);
 706                 }
 707                 if (g_mirror_disconnect_on_failure &&
 708                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
 709                         sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
 710                         g_mirror_event_send(disk,
 711                             G_MIRROR_DISK_STATE_DISCONNECTED,
 712                             G_MIRROR_EVENT_DONTWAIT);
 713                 }
 714         }
 715         return (error);
 716 }
 717
 718 static int
 719 g_mirror_clear_metadata(struct g_mirror_disk *disk)
 720 {
 721         int error;
 722
 723         g_topology_assert_not();
 724         sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
 725
 726         if (disk->d_softc->sc_type != G_MIRROR_TYPE_AUTOMATIC)
 727                 return (0);
 728         error = g_mirror_write_metadata(disk, NULL);
 729         if (error == 0) {
 730                 G_MIRROR_DEBUG(2, "Metadata on %s cleared.",
 731                     g_mirror_get_diskname(disk));
 732         } else {
 733                 G_MIRROR_DEBUG(0,
 734                     "Cannot clear metadata on disk %s (error=%d).",
 735                     g_mirror_get_diskname(disk), error);
 736         }
 737         return (error);
 738 }
 739
 740 void
 741 g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
 742     struct g_mirror_metadata *md)
 743 {
 744
 745         strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
 746         md->md_version = G_MIRROR_VERSION;
 747         strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
 748         md->md_mid = sc->sc_id;
 749         md->md_all = sc->sc_ndisks;
 750         md->md_slice = sc->sc_slice;
 751         md->md_balance = sc->sc_balance;
 752         md->md_genid = sc->sc_genid;
 753         md->md_mediasize = sc->sc_mediasize;
 754         md->md_sectorsize = sc->sc_sectorsize;
 755         md->md_mflags = (sc->sc_flags & G_MIRROR_DEVICE_FLAG_MASK);
 756         bzero(md->md_provider, sizeof(md->md_provider));
 757         if (disk == NULL) {
 758                 md->md_did = arc4random();
 759                 md->md_priority = 0;
 760                 md->md_syncid = 0;
 761                 md->md_dflags = 0;
 762                 md->md_sync_offset = 0;
 763                 md->md_provsize = 0;
 764         } else {
 765                 md->md_did = disk->d_id;
 766                 md->md_priority = disk->d_priority;
 767                 md->md_syncid = disk->d_sync.ds_syncid;
 768                 md->md_dflags = (disk->d_flags & G_MIRROR_DISK_FLAG_MASK);
 769                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
 770                         md->md_sync_offset = disk->d_sync.ds_offset_done;
 771                 else
 772                         md->md_sync_offset = 0;
 773                 if ((disk->d_flags & G_MIRROR_DISK_FLAG_HARDCODED) != 0) {
 774                         strlcpy(md->md_provider,
 775                             disk->d_consumer->provider->name,
 776                             sizeof(md->md_provider));
 777                 }
 778                 md->md_provsize = disk->d_consumer->provider->mediasize;
 779         }
 780 }
 781
 782 void
 783 g_mirror_update_metadata(struct g_mirror_disk *disk)
 784 {
 785         struct g_mirror_softc *sc;
 786         struct g_mirror_metadata md;
 787         int error;
 788
 789         g_topology_assert_not();
 790         sc = disk->d_softc;
 791         sx_assert(&sc->sc_lock, SX_LOCKED);
 792
 793         if (sc->sc_type != G_MIRROR_TYPE_AUTOMATIC)
 794                 return;
 795         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WIPE) == 0)
 796                 g_mirror_fill_metadata(sc, disk, &md);
 797         error = g_mirror_write_metadata(disk, &md);
 798         if (error == 0) {
 799                 G_MIRROR_DEBUG(2, "Metadata on %s updated.",
 800                     g_mirror_get_diskname(disk));
 801         } else {
 802                 G_MIRROR_DEBUG(0,
 803                     "Cannot update metadata on disk %s (error=%d).",
 804                     g_mirror_get_diskname(disk), error);
 805         }
 806 }
 807
 808 static void
 809 g_mirror_bump_syncid(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_syncid++;
 820         G_MIRROR_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
 821             sc->sc_syncid);
 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_sync.ds_syncid = sc->sc_syncid;
 826                         g_mirror_update_metadata(disk);
 827                 }
 828         }
 829 }
 830
 831 static void
 832 g_mirror_bump_genid(struct g_mirror_softc *sc)
 833 {
 834         struct g_mirror_disk *disk;
 835
 836         g_topology_assert_not();
 837         sx_assert(&sc->sc_lock, SX_XLOCKED);
 838         KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
 839             ("%s called with no active disks (device=%s).", __func__,
 840             sc->sc_name));
 841
 842         sc->sc_genid++;
 843         G_MIRROR_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
 844             sc->sc_genid);
 845         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 846                 if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
 847                     disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
 848                         disk->d_genid = sc->sc_genid;
 849                         g_mirror_update_metadata(disk);
 850                 }
 851         }
 852 }
 853
 854 static int
 855 g_mirror_idle(struct g_mirror_softc *sc, int acw)
 856 {
 857         struct g_mirror_disk *disk;
 858         int timeout;
 859
 860         g_topology_assert_not();
 861         sx_assert(&sc->sc_lock, SX_XLOCKED);
 862
 863         if (sc->sc_provider == NULL)
 864                 return (0);
 865         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
 866                 return (0);
 867         if (sc->sc_idle)
 868                 return (0);
 869         if (sc->sc_writes > 0)
 870                 return (0);
 871         if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
 872                 timeout = g_mirror_idletime - (time_uptime - sc->sc_last_write);
 873                 if (!g_mirror_shutdown && timeout > 0)
 874                         return (timeout);
 875         }
 876         sc->sc_idle = 1;
 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(2, "Disk %s (device %s) marked as clean.",
 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         return (0);
 886 }
 887
 888 static void
 889 g_mirror_unidle(struct g_mirror_softc *sc)
 890 {
 891         struct g_mirror_disk *disk;
 892
 893         g_topology_assert_not();
 894         sx_assert(&sc->sc_lock, SX_XLOCKED);
 895
 896         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
 897                 return;
 898         sc->sc_idle = 0;
 899         sc->sc_last_write = time_uptime;
 900         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 901                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
 902                         continue;
 903                 G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as dirty.",
 904                     g_mirror_get_diskname(disk), sc->sc_name);
 905                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
 906                 g_mirror_update_metadata(disk);
 907         }
 908 }
 909
 910 static void
 911 g_mirror_done(struct bio *bp)
 912 {
 913         struct g_mirror_softc *sc;
 914
 915         sc = bp->bio_from->geom->softc;
 916         bp->bio_cflags = G_MIRROR_BIO_FLAG_REGULAR;
 917         mtx_lock(&sc->sc_queue_mtx);
 918         TAILQ_INSERT_TAIL(&sc->sc_queue, bp, bio_queue);
 919         mtx_unlock(&sc->sc_queue_mtx);
 920         wakeup(sc);
 921 }
 922
 923 static void
 924 g_mirror_regular_request_error(struct g_mirror_softc *sc,
 925     struct g_mirror_disk *disk, struct bio *bp)
 926 {
 927
 928         if ((bp->bio_cmd == BIO_FLUSH || bp->bio_cmd == BIO_SPEEDUP) &&
 929             bp->bio_error == EOPNOTSUPP)
 930                 return;
 931
 932         if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
 933                 disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
 934                 G_MIRROR_LOGREQ(0, bp, "Request failed (error=%d).",
 935                     bp->bio_error);
 936         } else {
 937                 G_MIRROR_LOGREQ(1, bp, "Request failed (error=%d).",
 938                     bp->bio_error);
 939         }
 940         if (g_mirror_disconnect_on_failure &&
 941             g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
 942                 if (bp->bio_error == ENXIO &&
 943                     bp->bio_cmd == BIO_READ)
 944                         sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
 945                 else if (bp->bio_error == ENXIO)
 946                         sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID_NOW;
 947                 else
 948                         sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
 949                 g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
 950                     G_MIRROR_EVENT_DONTWAIT);
 951         }
 952 }
 953
 954 static void
 955 g_mirror_regular_request(struct g_mirror_softc *sc, struct bio *bp)
 956 {
 957         struct g_mirror_disk *disk;
 958         struct bio *pbp;
 959
 960         g_topology_assert_not();
 961         KASSERT(sc->sc_provider == bp->bio_parent->bio_to,
 962             ("regular request %p with unexpected origin", bp));
 963
 964         pbp = bp->bio_parent;
 965         bp->bio_from->index--;
 966         if (bp->bio_cmd == BIO_WRITE || bp->bio_cmd == BIO_DELETE)
 967                 sc->sc_writes--;
 968         disk = bp->bio_from->private;
 969         if (disk == NULL) {
 970                 g_topology_lock();
 971                 g_mirror_kill_consumer(sc, bp->bio_from);
 972                 g_topology_unlock();
 973         }
 974
 975         switch (bp->bio_cmd) {
 976         case BIO_READ:
 977                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_read,
 978                     bp->bio_error);
 979                 break;
 980         case BIO_WRITE:
 981                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_write,
 982                     bp->bio_error);
 983                 break;
 984         case BIO_DELETE:
 985                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_delete,
 986                     bp->bio_error);
 987                 break;
 988         case BIO_FLUSH:
 989                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_flush,
 990                     bp->bio_error);
 991                 break;
 992         case BIO_SPEEDUP:
 993                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_regular_request_speedup,
 994                     bp->bio_error);
 995                 break;
 996         }
 997
 998         pbp->bio_inbed++;
 999         KASSERT(pbp->bio_inbed <= pbp->bio_children,
1000             ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
1001             pbp->bio_children));
1002         if (bp->bio_error == 0 && pbp->bio_error == 0) {
1003                 G_MIRROR_LOGREQ(3, bp, "Request delivered.");
1004                 g_destroy_bio(bp);
1005                 if (pbp->bio_children == pbp->bio_inbed) {
1006                         G_MIRROR_LOGREQ(3, pbp, "Request delivered.");
1007                         pbp->bio_completed = pbp->bio_length;
1008                         if (pbp->bio_cmd == BIO_WRITE ||
1009                             pbp->bio_cmd == BIO_DELETE) {
1010                                 TAILQ_REMOVE(&sc->sc_inflight, pbp, bio_queue);
1011                                 /* Release delayed sync requests if possible. */
1012                                 g_mirror_sync_release(sc);
1013                         }
1014                         g_io_deliver(pbp, pbp->bio_error);
1015                 }
1016                 return;
1017         } else if (bp->bio_error != 0) {
1018                 if (pbp->bio_error == 0)
1019                         pbp->bio_error = bp->bio_error;
1020                 if (disk != NULL)
1021                         g_mirror_regular_request_error(sc, disk, bp);
1022                 switch (pbp->bio_cmd) {
1023                 case BIO_DELETE:
1024                 case BIO_WRITE:
1025                 case BIO_FLUSH:
1026                 case BIO_SPEEDUP:
1027                         pbp->bio_inbed--;
1028                         pbp->bio_children--;
1029                         break;
1030                 }
1031         }
1032         g_destroy_bio(bp);
1033
1034         switch (pbp->bio_cmd) {
1035         case BIO_READ:
1036                 if (pbp->bio_inbed < pbp->bio_children)
1037                         break;
1038                 if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1)
1039                         g_io_deliver(pbp, pbp->bio_error);
1040                 else {
1041                         pbp->bio_error = 0;
1042                         mtx_lock(&sc->sc_queue_mtx);
1043                         TAILQ_INSERT_TAIL(&sc->sc_queue, pbp, bio_queue);
1044                         mtx_unlock(&sc->sc_queue_mtx);
1045                         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1046                         wakeup(sc);
1047                 }
1048                 break;
1049         case BIO_DELETE:
1050         case BIO_WRITE:
1051         case BIO_FLUSH:
1052         case BIO_SPEEDUP:
1053                 if (pbp->bio_children == 0) {
1054                         /*
1055                          * All requests failed.
1056                          */
1057                 } else if (pbp->bio_inbed < pbp->bio_children) {
1058                         /* Do nothing. */
1059                         break;
1060                 } else if (pbp->bio_children == pbp->bio_inbed) {
1061                         /* Some requests succeeded. */
1062                         pbp->bio_error = 0;
1063                         pbp->bio_completed = pbp->bio_length;
1064                 }
1065                 if (pbp->bio_cmd == BIO_WRITE || pbp->bio_cmd == BIO_DELETE) {
1066                         TAILQ_REMOVE(&sc->sc_inflight, pbp, bio_queue);
1067                         /* Release delayed sync requests if possible. */
1068                         g_mirror_sync_release(sc);
1069                 }
1070                 g_io_deliver(pbp, pbp->bio_error);
1071                 break;
1072         default:
1073                 KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
1074                 break;
1075         }
1076 }
1077
1078 static void
1079 g_mirror_sync_done(struct bio *bp)
1080 {
1081         struct g_mirror_softc *sc;
1082
1083         G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
1084         sc = bp->bio_from->geom->softc;
1085         bp->bio_cflags = G_MIRROR_BIO_FLAG_SYNC;
1086         mtx_lock(&sc->sc_queue_mtx);
1087         TAILQ_INSERT_TAIL(&sc->sc_queue, bp, bio_queue);
1088         mtx_unlock(&sc->sc_queue_mtx);
1089         wakeup(sc);
1090 }
1091
1092 static void
1093 g_mirror_candelete(struct bio *bp)
1094 {
1095         struct g_mirror_softc *sc;
1096         struct g_mirror_disk *disk;
1097         int val;
1098
1099         sc = bp->bio_to->private;
1100         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1101                 if (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE)
1102                         break;
1103         }
1104         val = disk != NULL;
1105         g_handleattr(bp, "GEOM::candelete", &val, sizeof(val));
1106 }
1107
1108 static void
1109 g_mirror_kernel_dump(struct bio *bp)
1110 {
1111         struct g_mirror_softc *sc;
1112         struct g_mirror_disk *disk;
1113         struct bio *cbp;
1114         struct g_kerneldump *gkd;
1115
1116         /*
1117          * We configure dumping to the first component, because this component
1118          * will be used for reading with 'prefer' balance algorithm.
1119          * If the component with the highest priority is currently disconnected
1120          * we will not be able to read the dump after the reboot if it will be
1121          * connected and synchronized later. Can we do something better?
1122          */
1123         sc = bp->bio_to->private;
1124         disk = LIST_FIRST(&sc->sc_disks);
1125
1126         gkd = (struct g_kerneldump *)bp->bio_data;
1127         if (gkd->length > bp->bio_to->mediasize)
1128                 gkd->length = bp->bio_to->mediasize;
1129         cbp = g_clone_bio(bp);
1130         if (cbp == NULL) {
1131                 g_io_deliver(bp, ENOMEM);
1132                 return;
1133         }
1134         cbp->bio_done = g_std_done;
1135         g_io_request(cbp, disk->d_consumer);
1136         G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1137             g_mirror_get_diskname(disk));
1138 }
1139
1140 static void
1141 g_mirror_start(struct bio *bp)
1142 {
1143         struct g_mirror_softc *sc;
1144
1145         sc = bp->bio_to->private;
1146         /*
1147          * If sc == NULL or there are no valid disks, provider's error
1148          * should be set and g_mirror_start() should not be called at all.
1149          */
1150         KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1151             ("Provider's error should be set (error=%d)(mirror=%s).",
1152             bp->bio_to->error, bp->bio_to->name));
1153         G_MIRROR_LOGREQ(3, bp, "Request received.");
1154
1155         switch (bp->bio_cmd) {
1156         case BIO_READ:
1157         case BIO_WRITE:
1158         case BIO_DELETE:
1159         case BIO_SPEEDUP:
1160         case BIO_FLUSH:
1161                 break;
1162         case BIO_GETATTR:
1163                 if (!strcmp(bp->bio_attribute, "GEOM::candelete")) {
1164                         g_mirror_candelete(bp);
1165                         return;
1166                 } else if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1167                         g_mirror_kernel_dump(bp);
1168                         return;
1169                 }
1170                 /* FALLTHROUGH */
1171         default:
1172                 g_io_deliver(bp, EOPNOTSUPP);
1173                 return;
1174         }
1175         mtx_lock(&sc->sc_queue_mtx);
1176         if (bp->bio_to->error != 0) {
1177                 mtx_unlock(&sc->sc_queue_mtx);
1178                 g_io_deliver(bp, bp->bio_to->error);
1179                 return;
1180         }
1181         TAILQ_INSERT_TAIL(&sc->sc_queue, bp, bio_queue);
1182         mtx_unlock(&sc->sc_queue_mtx);
1183         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1184         wakeup(sc);
1185 }
1186
1187 /*
1188  * Return TRUE if the given request is colliding with a in-progress
1189  * synchronization request.
1190  */
1191 static bool
1192 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1193 {
1194         struct g_mirror_disk *disk;
1195         struct bio *sbp;
1196         off_t rstart, rend, sstart, send;
1197         u_int i;
1198
1199         if (sc->sc_sync.ds_ndisks == 0)
1200                 return (false);
1201         rstart = bp->bio_offset;
1202         rend = bp->bio_offset + bp->bio_length;
1203         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1204                 if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1205                         continue;
1206                 for (i = 0; i < g_mirror_syncreqs; i++) {
1207                         sbp = disk->d_sync.ds_bios[i];
1208                         if (sbp == NULL)
1209                                 continue;
1210                         sstart = sbp->bio_offset;
1211                         send = sbp->bio_offset + sbp->bio_length;
1212                         if (rend > sstart && rstart < send)
1213                                 return (true);
1214                 }
1215         }
1216         return (false);
1217 }
1218
1219 /*
1220  * Return TRUE if the given sync request is colliding with a in-progress regular
1221  * request.
1222  */
1223 static bool
1224 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1225 {
1226         off_t rstart, rend, sstart, send;
1227         struct bio *bp;
1228
1229         if (sc->sc_sync.ds_ndisks == 0)
1230                 return (false);
1231         sstart = sbp->bio_offset;
1232         send = sbp->bio_offset + sbp->bio_length;
1233         TAILQ_FOREACH(bp, &sc->sc_inflight, bio_queue) {
1234                 rstart = bp->bio_offset;
1235                 rend = bp->bio_offset + bp->bio_length;
1236                 if (rend > sstart && rstart < send)
1237                         return (true);
1238         }
1239         return (false);
1240 }
1241
1242 /*
1243  * Puts regular request onto delayed queue.
1244  */
1245 static void
1246 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1247 {
1248
1249         G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1250         TAILQ_INSERT_TAIL(&sc->sc_regular_delayed, bp, bio_queue);
1251 }
1252
1253 /*
1254  * Puts synchronization request onto delayed queue.
1255  */
1256 static void
1257 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1258 {
1259
1260         G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1261         TAILQ_INSERT_TAIL(&sc->sc_sync_delayed, bp, bio_queue);
1262 }
1263
1264 /*
1265  * Requeue delayed regular requests.
1266  */
1267 static void
1268 g_mirror_regular_release(struct g_mirror_softc *sc)
1269 {
1270         struct bio *bp;
1271
1272         if ((bp = TAILQ_FIRST(&sc->sc_regular_delayed)) == NULL)
1273                 return;
1274         if (g_mirror_sync_collision(sc, bp))
1275                 return;
1276
1277         G_MIRROR_DEBUG(2, "Requeuing regular requests after collision.");
1278         mtx_lock(&sc->sc_queue_mtx);
1279         TAILQ_CONCAT(&sc->sc_regular_delayed, &sc->sc_queue, bio_queue);
1280         TAILQ_SWAP(&sc->sc_regular_delayed, &sc->sc_queue, bio, bio_queue);
1281         mtx_unlock(&sc->sc_queue_mtx);
1282 }
1283
1284 /*
1285  * Releases delayed sync requests which don't collide anymore with regular
1286  * requests.
1287  */
1288 static void
1289 g_mirror_sync_release(struct g_mirror_softc *sc)
1290 {
1291         struct bio *bp, *bp2;
1292
1293         TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed, bio_queue, bp2) {
1294                 if (g_mirror_regular_collision(sc, bp))
1295                         continue;
1296                 TAILQ_REMOVE(&sc->sc_sync_delayed, bp, bio_queue);
1297                 G_MIRROR_LOGREQ(2, bp,
1298                     "Releasing delayed synchronization request.");
1299                 g_io_request(bp, bp->bio_from);
1300         }
1301 }
1302
1303 /*
1304  * Free a synchronization request and clear its slot in the array.
1305  */
1306 static void
1307 g_mirror_sync_request_free(struct g_mirror_disk *disk, struct bio *bp)
1308 {
1309         int idx;
1310
1311         if (disk != NULL && disk->d_sync.ds_bios != NULL) {
1312                 idx = (int)(uintptr_t)bp->bio_caller1;
1313                 KASSERT(disk->d_sync.ds_bios[idx] == bp,
1314                     ("unexpected sync BIO at %p:%d", disk, idx));
1315                 disk->d_sync.ds_bios[idx] = NULL;
1316         }
1317         free(bp->bio_data, M_MIRROR);
1318         g_destroy_bio(bp);
1319 }
1320
1321 /*
1322  * Handle synchronization requests.
1323  * Every synchronization request is a two-step process: first, a read request is
1324  * sent to the mirror provider via the sync consumer. If that request completes
1325  * successfully, it is converted to a write and sent to the disk being
1326  * synchronized. If the write also completes successfully, the synchronization
1327  * offset is advanced and a new read request is submitted.
1328  */
1329 static void
1330 g_mirror_sync_request(struct g_mirror_softc *sc, struct bio *bp)
1331 {
1332         struct g_mirror_disk *disk;
1333         struct g_mirror_disk_sync *sync;
1334
1335         KASSERT((bp->bio_cmd == BIO_READ &&
1336             bp->bio_from->geom == sc->sc_sync.ds_geom) ||
1337             (bp->bio_cmd == BIO_WRITE && bp->bio_from->geom == sc->sc_geom),
1338             ("Sync BIO %p with unexpected origin", bp));
1339
1340         bp->bio_from->index--;
1341         disk = bp->bio_from->private;
1342         if (disk == NULL) {
1343                 sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1344                 g_topology_lock();
1345                 g_mirror_kill_consumer(sc, bp->bio_from);
1346                 g_topology_unlock();
1347                 g_mirror_sync_request_free(NULL, bp);
1348                 sx_xlock(&sc->sc_lock);
1349                 return;
1350         }
1351
1352         sync = &disk->d_sync;
1353
1354         /*
1355          * Synchronization request.
1356          */
1357         switch (bp->bio_cmd) {
1358         case BIO_READ: {
1359                 struct g_consumer *cp;
1360
1361                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_sync_request_read,
1362                     bp->bio_error);
1363
1364                 if (bp->bio_error != 0) {
1365                         G_MIRROR_LOGREQ(0, bp,
1366                             "Synchronization request failed (error=%d).",
1367                             bp->bio_error);
1368
1369                         /*
1370                          * The read error will trigger a syncid bump, so there's
1371                          * no need to do that here.
1372                          *
1373                          * The read error handling for regular requests will
1374                          * retry the read from all active mirrors before passing
1375                          * the error back up, so there's no need to retry here.
1376                          */
1377                         g_mirror_sync_request_free(disk, bp);
1378                         g_mirror_event_send(disk,
1379                             G_MIRROR_DISK_STATE_DISCONNECTED,
1380                             G_MIRROR_EVENT_DONTWAIT);
1381                         return;
1382                 }
1383                 G_MIRROR_LOGREQ(3, bp,
1384                     "Synchronization request half-finished.");
1385                 bp->bio_cmd = BIO_WRITE;
1386                 bp->bio_cflags = 0;
1387                 cp = disk->d_consumer;
1388                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1389                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1390                     cp->acr, cp->acw, cp->ace));
1391                 cp->index++;
1392                 g_io_request(bp, cp);
1393                 return;
1394         }
1395         case BIO_WRITE: {
1396                 off_t offset;
1397                 int i;
1398
1399                 KFAIL_POINT_ERROR(DEBUG_FP, g_mirror_sync_request_write,
1400                     bp->bio_error);
1401
1402                 if (bp->bio_error != 0) {
1403                         G_MIRROR_LOGREQ(0, bp,
1404                             "Synchronization request failed (error=%d).",
1405                             bp->bio_error);
1406                         g_mirror_sync_request_free(disk, bp);
1407                         sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1408                         g_mirror_event_send(disk,
1409                             G_MIRROR_DISK_STATE_DISCONNECTED,
1410                             G_MIRROR_EVENT_DONTWAIT);
1411                         return;
1412                 }
1413                 G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1414                 if (sync->ds_offset >= sc->sc_mediasize ||
1415                     sync->ds_consumer == NULL ||
1416                     (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1417                         /* Don't send more synchronization requests. */
1418                         sync->ds_inflight--;
1419                         g_mirror_sync_request_free(disk, bp);
1420                         if (sync->ds_inflight > 0)
1421                                 return;
1422                         if (sync->ds_consumer == NULL ||
1423                             (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1424                                 return;
1425                         }
1426                         /* Disk up-to-date, activate it. */
1427                         g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1428                             G_MIRROR_EVENT_DONTWAIT);
1429                         return;
1430                 }
1431
1432                 /* Send next synchronization request. */
1433                 g_mirror_sync_reinit(disk, bp, sync->ds_offset);
1434                 sync->ds_offset += bp->bio_length;
1435
1436                 G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1437                 sync->ds_consumer->index++;
1438
1439                 /*
1440                  * Delay the request if it is colliding with a regular request.
1441                  */
1442                 if (g_mirror_regular_collision(sc, bp))
1443                         g_mirror_sync_delay(sc, bp);
1444                 else
1445                         g_io_request(bp, sync->ds_consumer);
1446
1447                 /* Requeue delayed requests if possible. */
1448                 g_mirror_regular_release(sc);
1449
1450                 /* Find the smallest offset */
1451                 offset = sc->sc_mediasize;
1452                 for (i = 0; i < g_mirror_syncreqs; i++) {
1453                         bp = sync->ds_bios[i];
1454                         if (bp != NULL && bp->bio_offset < offset)
1455                                 offset = bp->bio_offset;
1456                 }
1457                 if (g_mirror_sync_period > 0 &&
1458                     time_uptime - sync->ds_update_ts > g_mirror_sync_period) {
1459                         sync->ds_offset_done = offset;
1460                         g_mirror_update_metadata(disk);
1461                         sync->ds_update_ts = time_uptime;
1462                 }
1463                 return;
1464         }
1465         default:
1466                 panic("Invalid I/O request %p", bp);
1467         }
1468 }
1469
1470 static void
1471 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1472 {
1473         struct g_mirror_disk *disk;
1474         struct g_consumer *cp;
1475         struct bio *cbp;
1476
1477         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1478                 if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1479                         break;
1480         }
1481         if (disk == NULL) {
1482                 if (bp->bio_error == 0)
1483                         bp->bio_error = ENXIO;
1484                 g_io_deliver(bp, bp->bio_error);
1485                 return;
1486         }
1487         cbp = g_clone_bio(bp);
1488         if (cbp == NULL) {
1489                 if (bp->bio_error == 0)
1490                         bp->bio_error = ENOMEM;
1491                 g_io_deliver(bp, bp->bio_error);
1492                 return;
1493         }
1494         /*
1495          * Fill in the component buf structure.
1496          */
1497         cp = disk->d_consumer;
1498         cbp->bio_done = g_mirror_done;
1499         cbp->bio_to = cp->provider;
1500         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1501         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1502             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1503             cp->acw, cp->ace));
1504         cp->index++;
1505         g_io_request(cbp, cp);
1506 }
1507
1508 static void
1509 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1510 {
1511         struct g_mirror_disk *disk;
1512         struct g_consumer *cp;
1513         struct bio *cbp;
1514
1515         disk = g_mirror_get_disk(sc);
1516         if (disk == NULL) {
1517                 if (bp->bio_error == 0)
1518                         bp->bio_error = ENXIO;
1519                 g_io_deliver(bp, bp->bio_error);
1520                 return;
1521         }
1522         cbp = g_clone_bio(bp);
1523         if (cbp == NULL) {
1524                 if (bp->bio_error == 0)
1525                         bp->bio_error = ENOMEM;
1526                 g_io_deliver(bp, bp->bio_error);
1527                 return;
1528         }
1529         /*
1530          * Fill in the component buf structure.
1531          */
1532         cp = disk->d_consumer;
1533         cbp->bio_done = g_mirror_done;
1534         cbp->bio_to = cp->provider;
1535         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1536         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1537             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1538             cp->acw, cp->ace));
1539         cp->index++;
1540         g_io_request(cbp, cp);
1541 }
1542
1543 #define TRACK_SIZE  (1 * 1024 * 1024)
1544 #define LOAD_SCALE      256
1545 #define ABS(x)          (((x) >= 0) ? (x) : (-(x)))
1546
1547 static void
1548 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1549 {
1550         struct g_mirror_disk *disk, *dp;
1551         struct g_consumer *cp;
1552         struct bio *cbp;
1553         int prio, best;
1554
1555         /* Find a disk with the smallest load. */
1556         disk = NULL;
1557         best = INT_MAX;
1558         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1559                 if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1560                         continue;
1561                 prio = dp->load;
1562                 /* If disk head is precisely in position - highly prefer it. */
1563                 if (dp->d_last_offset == bp->bio_offset)
1564                         prio -= 2 * LOAD_SCALE;
1565                 else
1566                 /* If disk head is close to position - prefer it. */
1567                 if (ABS(dp->d_last_offset - bp->bio_offset) < TRACK_SIZE)
1568                         prio -= 1 * LOAD_SCALE;
1569                 if (prio <= best) {
1570                         disk = dp;
1571                         best = prio;
1572                 }
1573         }
1574         KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1575         cbp = g_clone_bio(bp);
1576         if (cbp == NULL) {
1577                 if (bp->bio_error == 0)
1578                         bp->bio_error = ENOMEM;
1579                 g_io_deliver(bp, bp->bio_error);
1580                 return;
1581         }
1582         /*
1583          * Fill in the component buf structure.
1584          */
1585         cp = disk->d_consumer;
1586         cbp->bio_done = g_mirror_done;
1587         cbp->bio_to = cp->provider;
1588         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1589         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1590             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1591             cp->acw, cp->ace));
1592         cp->index++;
1593         /* Remember last head position */
1594         disk->d_last_offset = bp->bio_offset + bp->bio_length;
1595         /* Update loads. */
1596         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1597                 dp->load = (dp->d_consumer->index * LOAD_SCALE +
1598                     dp->load * 7) / 8;
1599         }
1600         g_io_request(cbp, cp);
1601 }
1602
1603 static void
1604 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1605 {
1606         struct bio_queue queue;
1607         struct g_mirror_disk *disk;
1608         struct g_consumer *cp;
1609         struct bio *cbp;
1610         off_t left, mod, offset, slice;
1611         u_char *data;
1612         u_int ndisks;
1613
1614         if (bp->bio_length <= sc->sc_slice) {
1615                 g_mirror_request_round_robin(sc, bp);
1616                 return;
1617         }
1618         ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1619         slice = bp->bio_length / ndisks;
1620         mod = slice % sc->sc_provider->sectorsize;
1621         if (mod != 0)
1622                 slice += sc->sc_provider->sectorsize - mod;
1623         /*
1624          * Allocate all bios before sending any request, so we can
1625          * return ENOMEM in nice and clean way.
1626          */
1627         left = bp->bio_length;
1628         offset = bp->bio_offset;
1629         data = bp->bio_data;
1630         TAILQ_INIT(&queue);
1631         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1632                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1633                         continue;
1634                 cbp = g_clone_bio(bp);
1635                 if (cbp == NULL) {
1636                         while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1637                                 TAILQ_REMOVE(&queue, cbp, bio_queue);
1638                                 g_destroy_bio(cbp);
1639                         }
1640                         if (bp->bio_error == 0)
1641                                 bp->bio_error = ENOMEM;
1642                         g_io_deliver(bp, bp->bio_error);
1643                         return;
1644                 }
1645                 TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
1646                 cbp->bio_done = g_mirror_done;
1647                 cbp->bio_caller1 = disk;
1648                 cbp->bio_to = disk->d_consumer->provider;
1649                 cbp->bio_offset = offset;
1650                 cbp->bio_data = data;
1651                 cbp->bio_length = MIN(left, slice);
1652                 left -= cbp->bio_length;
1653                 if (left == 0)
1654                         break;
1655                 offset += cbp->bio_length;
1656                 data += cbp->bio_length;
1657         }
1658         while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1659                 TAILQ_REMOVE(&queue, cbp, bio_queue);
1660                 G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1661                 disk = cbp->bio_caller1;
1662                 cbp->bio_caller1 = NULL;
1663                 cp = disk->d_consumer;
1664                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1665                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1666                     cp->acr, cp->acw, cp->ace));
1667                 disk->d_consumer->index++;
1668                 g_io_request(cbp, disk->d_consumer);
1669         }
1670 }
1671
1672 static void
1673 g_mirror_register_request(struct g_mirror_softc *sc, struct bio *bp)
1674 {
1675         struct bio_queue queue;
1676         struct bio *cbp;
1677         struct g_consumer *cp;
1678         struct g_mirror_disk *disk;
1679
1680         sx_assert(&sc->sc_lock, SA_XLOCKED);
1681
1682         /*
1683          * To avoid ordering issues, if a write is deferred because of a
1684          * collision with a sync request, all I/O is deferred until that
1685          * write is initiated.
1686          */
1687         if (bp->bio_from->geom != sc->sc_sync.ds_geom &&
1688             !TAILQ_EMPTY(&sc->sc_regular_delayed)) {
1689                 g_mirror_regular_delay(sc, bp);
1690                 return;
1691         }
1692
1693         switch (bp->bio_cmd) {
1694         case BIO_READ:
1695                 switch (sc->sc_balance) {
1696                 case G_MIRROR_BALANCE_LOAD:
1697                         g_mirror_request_load(sc, bp);
1698                         break;
1699                 case G_MIRROR_BALANCE_PREFER:
1700                         g_mirror_request_prefer(sc, bp);
1701                         break;
1702                 case G_MIRROR_BALANCE_ROUND_ROBIN:
1703                         g_mirror_request_round_robin(sc, bp);
1704                         break;
1705                 case G_MIRROR_BALANCE_SPLIT:
1706                         g_mirror_request_split(sc, bp);
1707                         break;
1708                 }
1709                 return;
1710         case BIO_WRITE:
1711         case BIO_DELETE:
1712                 /*
1713                  * Delay the request if it is colliding with a synchronization
1714                  * request.
1715                  */
1716                 if (g_mirror_sync_collision(sc, bp)) {
1717                         g_mirror_regular_delay(sc, bp);
1718                         return;
1719                 }
1720
1721                 if (sc->sc_idle)
1722                         g_mirror_unidle(sc);
1723                 else
1724                         sc->sc_last_write = time_uptime;
1725
1726                 /*
1727                  * Bump syncid on first write.
1728                  */
1729                 if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1730                         sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1731                         g_mirror_bump_syncid(sc);
1732                 }
1733
1734                 /*
1735                  * Allocate all bios before sending any request, so we can
1736                  * return ENOMEM in nice and clean way.
1737                  */
1738                 TAILQ_INIT(&queue);
1739                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1740                         switch (disk->d_state) {
1741                         case G_MIRROR_DISK_STATE_ACTIVE:
1742                                 break;
1743                         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1744                                 if (bp->bio_offset >= disk->d_sync.ds_offset)
1745                                         continue;
1746                                 break;
1747                         default:
1748                                 continue;
1749                         }
1750                         if (bp->bio_cmd == BIO_DELETE &&
1751                             (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE) == 0)
1752                                 continue;
1753                         cbp = g_clone_bio(bp);
1754                         if (cbp == NULL) {
1755                                 while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1756                                         TAILQ_REMOVE(&queue, cbp, bio_queue);
1757                                         g_destroy_bio(cbp);
1758                                 }
1759                                 if (bp->bio_error == 0)
1760                                         bp->bio_error = ENOMEM;
1761                                 g_io_deliver(bp, bp->bio_error);
1762                                 return;
1763                         }
1764                         TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
1765                         cbp->bio_done = g_mirror_done;
1766                         cp = disk->d_consumer;
1767                         cbp->bio_caller1 = cp;
1768                         cbp->bio_to = cp->provider;
1769                         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1770                             ("Consumer %s not opened (r%dw%de%d).",
1771                             cp->provider->name, cp->acr, cp->acw, cp->ace));
1772                 }
1773                 if (TAILQ_EMPTY(&queue)) {
1774                         KASSERT(bp->bio_cmd == BIO_DELETE,
1775                             ("No consumers for regular request %p", bp));
1776                         g_io_deliver(bp, EOPNOTSUPP);
1777                         return;
1778                 }
1779                 while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1780                         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1781                         TAILQ_REMOVE(&queue, cbp, bio_queue);
1782                         cp = cbp->bio_caller1;
1783                         cbp->bio_caller1 = NULL;
1784                         cp->index++;
1785                         sc->sc_writes++;
1786                         g_io_request(cbp, cp);
1787                 }
1788                 /*
1789                  * Put request onto inflight queue, so we can check if new
1790                  * synchronization requests don't collide with it.
1791                  */
1792                 TAILQ_INSERT_TAIL(&sc->sc_inflight, bp, bio_queue);
1793                 return;
1794         case BIO_SPEEDUP:
1795         case BIO_FLUSH:
1796                 TAILQ_INIT(&queue);
1797                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1798                         if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1799                                 continue;
1800                         cbp = g_clone_bio(bp);
1801                         if (cbp == NULL) {
1802                                 while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1803                                         TAILQ_REMOVE(&queue, cbp, bio_queue);
1804                                         g_destroy_bio(cbp);
1805                                 }
1806                                 if (bp->bio_error == 0)
1807                                         bp->bio_error = ENOMEM;
1808                                 g_io_deliver(bp, bp->bio_error);
1809                                 return;
1810                         }
1811                         TAILQ_INSERT_TAIL(&queue, cbp, bio_queue);
1812                         cbp->bio_done = g_mirror_done;
1813                         cbp->bio_caller1 = disk;
1814                         cbp->bio_to = disk->d_consumer->provider;
1815                 }
1816                 KASSERT(!TAILQ_EMPTY(&queue),
1817                     ("No consumers for regular request %p", bp));
1818                 while ((cbp = TAILQ_FIRST(&queue)) != NULL) {
1819                         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1820                         TAILQ_REMOVE(&queue, cbp, bio_queue);
1821                         disk = cbp->bio_caller1;
1822                         cbp->bio_caller1 = NULL;
1823                         cp = disk->d_consumer;
1824                         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1825                             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1826                             cp->acr, cp->acw, cp->ace));
1827                         cp->index++;
1828                         g_io_request(cbp, cp);
1829                 }
1830                 break;
1831         default:
1832                 KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1833                     bp->bio_cmd, sc->sc_name));
1834                 break;
1835         }
1836 }
1837
1838 static int
1839 g_mirror_can_destroy(struct g_mirror_softc *sc)
1840 {
1841         struct g_geom *gp;
1842         struct g_consumer *cp;
1843
1844         g_topology_assert();
1845         gp = sc->sc_geom;
1846         if (gp->softc == NULL)
1847                 return (1);
1848         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_TASTING) != 0)
1849                 return (0);
1850         LIST_FOREACH(cp, &gp->consumer, consumer) {
1851                 if (g_mirror_is_busy(sc, cp))
1852                         return (0);
1853         }
1854         gp = sc->sc_sync.ds_geom;
1855         LIST_FOREACH(cp, &gp->consumer, consumer) {
1856                 if (g_mirror_is_busy(sc, cp))
1857                         return (0);
1858         }
1859         G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1860             sc->sc_name);
1861         return (1);
1862 }
1863
1864 static int
1865 g_mirror_try_destroy(struct g_mirror_softc *sc)
1866 {
1867
1868         if (sc->sc_rootmount != NULL) {
1869                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1870                     sc->sc_rootmount);
1871                 root_mount_rel(sc->sc_rootmount);
1872                 sc->sc_rootmount = NULL;
1873         }
1874         g_topology_lock();
1875         if (!g_mirror_can_destroy(sc)) {
1876                 g_topology_unlock();
1877                 return (0);
1878         }
1879         sc->sc_geom->softc = NULL;
1880         sc->sc_sync.ds_geom->softc = NULL;
1881         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DRAIN) != 0) {
1882                 g_topology_unlock();
1883                 G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1884                     &sc->sc_worker);
1885                 /* Unlock sc_lock here, as it can be destroyed after wakeup. */
1886                 sx_xunlock(&sc->sc_lock);
1887                 wakeup(&sc->sc_worker);
1888                 sc->sc_worker = NULL;
1889         } else {
1890                 g_topology_unlock();
1891                 g_mirror_destroy_device(sc);
1892         }
1893         return (1);
1894 }
1895
1896 /*
1897  * Worker thread.
1898  */
1899 static void
1900 g_mirror_worker(void *arg)
1901 {
1902         struct g_mirror_softc *sc;
1903         struct g_mirror_event *ep;
1904         struct bio *bp;
1905         int timeout;
1906
1907         sc = arg;
1908         thread_lock(curthread);
1909         sched_prio(curthread, PRIBIO);
1910         thread_unlock(curthread);
1911
1912         sx_xlock(&sc->sc_lock);
1913         for (;;) {
1914                 G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1915                 /*
1916                  * First take a look at events.
1917                  * This is important to handle events before any I/O requests.
1918                  */
1919                 ep = g_mirror_event_first(sc);
1920                 if (ep != NULL) {
1921                         g_mirror_event_remove(sc, ep);
1922                         if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1923                                 /* Update only device status. */
1924                                 G_MIRROR_DEBUG(3,
1925                                     "Running event for device %s.",
1926                                     sc->sc_name);
1927                                 ep->e_error = 0;
1928                                 g_mirror_update_device(sc, true);
1929                         } else {
1930                                 /* Update disk status. */
1931                                 G_MIRROR_DEBUG(3, "Running event for disk %s.",
1932                                      g_mirror_get_diskname(ep->e_disk));
1933                                 ep->e_error = g_mirror_update_disk(ep->e_disk,
1934                                     ep->e_state);
1935                                 if (ep->e_error == 0)
1936                                         g_mirror_update_device(sc, false);
1937                         }
1938                         if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1939                                 KASSERT(ep->e_error == 0,
1940                                     ("Error cannot be handled."));
1941                                 g_mirror_event_free(ep);
1942                         } else {
1943                                 ep->e_flags |= G_MIRROR_EVENT_DONE;
1944                                 G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1945                                     ep);
1946                                 mtx_lock(&sc->sc_events_mtx);
1947                                 wakeup(ep);
1948                                 mtx_unlock(&sc->sc_events_mtx);
1949                         }
1950                         if ((sc->sc_flags &
1951                             G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1952                                 if (g_mirror_try_destroy(sc)) {
1953                                         curthread->td_pflags &= ~TDP_GEOM;
1954                                         G_MIRROR_DEBUG(1, "Thread exiting.");
1955                                         kproc_exit(0);
1956                                 }
1957                         }
1958                         G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1959                         continue;
1960                 }
1961
1962                 /*
1963                  * Check if we can mark array as CLEAN and if we can't take
1964                  * how much seconds should we wait.
1965                  */
1966                 timeout = g_mirror_idle(sc, -1);
1967
1968                 /*
1969                  * Handle I/O requests.
1970                  */
1971                 mtx_lock(&sc->sc_queue_mtx);
1972                 bp = TAILQ_FIRST(&sc->sc_queue);
1973                 if (bp != NULL)
1974                         TAILQ_REMOVE(&sc->sc_queue, bp, bio_queue);
1975                 else {
1976                         if ((sc->sc_flags &
1977                             G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1978                                 mtx_unlock(&sc->sc_queue_mtx);
1979                                 if (g_mirror_try_destroy(sc)) {
1980                                         curthread->td_pflags &= ~TDP_GEOM;
1981                                         G_MIRROR_DEBUG(1, "Thread exiting.");
1982                                         kproc_exit(0);
1983                                 }
1984                                 mtx_lock(&sc->sc_queue_mtx);
1985                                 if (!TAILQ_EMPTY(&sc->sc_queue)) {
1986                                         mtx_unlock(&sc->sc_queue_mtx);
1987                                         continue;
1988                                 }
1989                         }
1990                         if (g_mirror_event_first(sc) != NULL) {
1991                                 mtx_unlock(&sc->sc_queue_mtx);
1992                                 continue;
1993                         }
1994                         sx_xunlock(&sc->sc_lock);
1995                         MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
1996                             timeout * hz);
1997                         sx_xlock(&sc->sc_lock);
1998                         G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
1999                         continue;
2000                 }
2001                 mtx_unlock(&sc->sc_queue_mtx);
2002
2003                 if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
2004                     (bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0) {
2005                         /*
2006                          * Handle completion of the first half (the read) of a
2007                          * block synchronization operation.
2008                          */
2009                         g_mirror_sync_request(sc, bp);
2010                 } else if (bp->bio_to != sc->sc_provider) {
2011                         if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
2012                                 /*
2013                                  * Handle completion of a regular I/O request.
2014                                  */
2015                                 g_mirror_regular_request(sc, bp);
2016                         else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
2017                                 /*
2018                                  * Handle completion of the second half (the
2019                                  * write) of a block synchronization operation.
2020                                  */
2021                                 g_mirror_sync_request(sc, bp);
2022                         else {
2023                                 KASSERT(0,
2024                                     ("Invalid request cflags=0x%hx to=%s.",
2025                                     bp->bio_cflags, bp->bio_to->name));
2026                         }
2027                 } else {
2028                         /*
2029                          * Initiate an I/O request.
2030                          */
2031                         g_mirror_register_request(sc, bp);
2032                 }
2033                 G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
2034         }
2035 }
2036
2037 static void
2038 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
2039 {
2040
2041         sx_assert(&sc->sc_lock, SX_LOCKED);
2042
2043         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
2044                 return;
2045         if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2046                 G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as dirty.",
2047                     g_mirror_get_diskname(disk), sc->sc_name);
2048                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
2049         } else if (sc->sc_idle &&
2050             (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2051                 G_MIRROR_DEBUG(2, "Disk %s (device %s) marked as clean.",
2052                     g_mirror_get_diskname(disk), sc->sc_name);
2053                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2054         }
2055 }
2056
2057 static void
2058 g_mirror_sync_reinit(const struct g_mirror_disk *disk, struct bio *bp,
2059     off_t offset)
2060 {
2061         void *data;
2062         int idx;
2063
2064         data = bp->bio_data;
2065         idx = (int)(uintptr_t)bp->bio_caller1;
2066         g_reset_bio(bp);
2067
2068         bp->bio_cmd = BIO_READ;
2069         bp->bio_data = data;
2070         bp->bio_done = g_mirror_sync_done;
2071         bp->bio_from = disk->d_sync.ds_consumer;
2072         bp->bio_to = disk->d_softc->sc_provider;
2073         bp->bio_caller1 = (void *)(uintptr_t)idx;
2074         bp->bio_offset = offset;
2075         bp->bio_length = MIN(MAXPHYS,
2076             disk->d_softc->sc_mediasize - bp->bio_offset);
2077 }
2078
2079 static void
2080 g_mirror_sync_start(struct g_mirror_disk *disk)
2081 {
2082         struct g_mirror_softc *sc;
2083         struct g_mirror_disk_sync *sync;
2084         struct g_consumer *cp;
2085         struct bio *bp;
2086         int error, i;
2087
2088         g_topology_assert_not();
2089         sc = disk->d_softc;
2090         sync = &disk->d_sync;
2091         sx_assert(&sc->sc_lock, SX_LOCKED);
2092
2093         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2094             ("Disk %s is not marked for synchronization.",
2095             g_mirror_get_diskname(disk)));
2096         KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2097             ("Device not in RUNNING state (%s, %u).", sc->sc_name,
2098             sc->sc_state));
2099
2100         sx_xunlock(&sc->sc_lock);
2101         g_topology_lock();
2102         cp = g_new_consumer(sc->sc_sync.ds_geom);
2103         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2104         error = g_attach(cp, sc->sc_provider);
2105         KASSERT(error == 0,
2106             ("Cannot attach to %s (error=%d).", sc->sc_name, error));
2107         error = g_access(cp, 1, 0, 0);
2108         KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
2109         g_topology_unlock();
2110         sx_xlock(&sc->sc_lock);
2111
2112         G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
2113             g_mirror_get_diskname(disk));
2114         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) == 0)
2115                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
2116         KASSERT(sync->ds_consumer == NULL,
2117             ("Sync consumer already exists (device=%s, disk=%s).",
2118             sc->sc_name, g_mirror_get_diskname(disk)));
2119
2120         sync->ds_consumer = cp;
2121         sync->ds_consumer->private = disk;
2122         sync->ds_consumer->index = 0;
2123
2124         /*
2125          * Allocate memory for synchronization bios and initialize them.
2126          */
2127         sync->ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
2128             M_MIRROR, M_WAITOK);
2129         for (i = 0; i < g_mirror_syncreqs; i++) {
2130                 bp = g_alloc_bio();
2131                 sync->ds_bios[i] = bp;
2132
2133                 bp->bio_data = malloc(MAXPHYS, M_MIRROR, M_WAITOK);
2134                 bp->bio_caller1 = (void *)(uintptr_t)i;
2135                 g_mirror_sync_reinit(disk, bp, sync->ds_offset);
2136                 sync->ds_offset += bp->bio_length;
2137         }
2138
2139         /* Increase the number of disks in SYNCHRONIZING state. */
2140         sc->sc_sync.ds_ndisks++;
2141         /* Set the number of in-flight synchronization requests. */
2142         sync->ds_inflight = g_mirror_syncreqs;
2143
2144         /*
2145          * Fire off first synchronization requests.
2146          */
2147         for (i = 0; i < g_mirror_syncreqs; i++) {
2148                 bp = sync->ds_bios[i];
2149                 G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
2150                 sync->ds_consumer->index++;
2151                 /*
2152                  * Delay the request if it is colliding with a regular request.
2153                  */
2154                 if (g_mirror_regular_collision(sc, bp))
2155                         g_mirror_sync_delay(sc, bp);
2156                 else
2157                         g_io_request(bp, sync->ds_consumer);
2158         }
2159 }
2160
2161 /*
2162  * Stop synchronization process.
2163  * type: 0 - synchronization finished
2164  *       1 - synchronization stopped
2165  */
2166 static void
2167 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
2168 {
2169         struct g_mirror_softc *sc;
2170         struct g_consumer *cp;
2171
2172         g_topology_assert_not();
2173         sc = disk->d_softc;
2174         sx_assert(&sc->sc_lock, SX_LOCKED);
2175
2176         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2177             ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2178             g_mirror_disk_state2str(disk->d_state)));
2179         if (disk->d_sync.ds_consumer == NULL)
2180                 return;
2181
2182         if (type == 0) {
2183                 G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2184                     sc->sc_name, g_mirror_get_diskname(disk));
2185         } else /* if (type == 1) */ {
2186                 G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2187                     sc->sc_name, g_mirror_get_diskname(disk));
2188         }
2189         g_mirror_regular_release(sc);
2190         free(disk->d_sync.ds_bios, M_MIRROR);
2191         disk->d_sync.ds_bios = NULL;
2192         cp = disk->d_sync.ds_consumer;
2193         disk->d_sync.ds_consumer = NULL;
2194         disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2195         sc->sc_sync.ds_ndisks--;
2196         sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2197         g_topology_lock();
2198         g_mirror_kill_consumer(sc, cp);
2199         g_topology_unlock();
2200         sx_xlock(&sc->sc_lock);
2201 }
2202
2203 static void
2204 g_mirror_launch_provider(struct g_mirror_softc *sc)
2205 {
2206         struct g_mirror_disk *disk;
2207         struct g_provider *pp, *dp;
2208
2209         sx_assert(&sc->sc_lock, SX_LOCKED);
2210
2211         g_topology_lock();
2212         pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
2213         pp->flags |= G_PF_DIRECT_RECEIVE;
2214         pp->mediasize = sc->sc_mediasize;
2215         pp->sectorsize = sc->sc_sectorsize;
2216         pp->stripesize = 0;
2217         pp->stripeoffset = 0;
2218
2219         /* Splitting of unmapped BIO's could work but isn't implemented now */
2220         if (sc->sc_balance != G_MIRROR_BALANCE_SPLIT)
2221                 pp->flags |= G_PF_ACCEPT_UNMAPPED;
2222
2223         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2224                 if (disk->d_consumer && disk->d_consumer->provider) {
2225                         dp = disk->d_consumer->provider;
2226                         if (dp->stripesize > pp->stripesize) {
2227                                 pp->stripesize = dp->stripesize;
2228                                 pp->stripeoffset = dp->stripeoffset;
2229                         }
2230                         /* A provider underneath us doesn't support unmapped */
2231                         if ((dp->flags & G_PF_ACCEPT_UNMAPPED) == 0) {
2232                                 G_MIRROR_DEBUG(0, "Cancelling unmapped "
2233                                     "because of %s.", dp->name);
2234                                 pp->flags &= ~G_PF_ACCEPT_UNMAPPED;
2235                         }
2236                 }
2237         }
2238         pp->private = sc;
2239         sc->sc_refcnt++;
2240         sc->sc_provider = pp;
2241         g_error_provider(pp, 0);
2242         g_topology_unlock();
2243         G_MIRROR_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2244             g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE), sc->sc_ndisks);
2245         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2246                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2247                         g_mirror_sync_start(disk);
2248         }
2249 }
2250
2251 static void
2252 g_mirror_destroy_provider(struct g_mirror_softc *sc)
2253 {
2254         struct g_mirror_disk *disk;
2255         struct bio *bp;
2256
2257         g_topology_assert_not();
2258         KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2259             sc->sc_name));
2260
2261         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2262                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2263                         g_mirror_sync_stop(disk, 1);
2264         }
2265
2266         g_topology_lock();
2267         g_error_provider(sc->sc_provider, ENXIO);
2268         mtx_lock(&sc->sc_queue_mtx);
2269         while ((bp = TAILQ_FIRST(&sc->sc_queue)) != NULL) {
2270                 TAILQ_REMOVE(&sc->sc_queue, bp, bio_queue);
2271                 /*
2272                  * Abort any pending I/O that wasn't generated by us.
2273                  * Synchronization requests and requests destined for individual
2274                  * mirror components can be destroyed immediately.
2275                  */
2276                 if (bp->bio_to == sc->sc_provider &&
2277                     bp->bio_from->geom != sc->sc_sync.ds_geom) {
2278                         g_io_deliver(bp, ENXIO);
2279                 } else {
2280                         if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
2281                                 free(bp->bio_data, M_MIRROR);
2282                         g_destroy_bio(bp);
2283                 }
2284         }
2285         mtx_unlock(&sc->sc_queue_mtx);
2286         g_wither_provider(sc->sc_provider, ENXIO);
2287         sc->sc_provider = NULL;
2288         G_MIRROR_DEBUG(0, "Device %s: provider destroyed.", sc->sc_name);
2289         g_topology_unlock();
2290 }
2291
2292 static void
2293 g_mirror_go(void *arg)
2294 {
2295         struct g_mirror_softc *sc;
2296
2297         sc = arg;
2298         G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2299         g_mirror_event_send(sc, 0,
2300             G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2301 }
2302
2303 static u_int
2304 g_mirror_determine_state(struct g_mirror_disk *disk)
2305 {
2306         struct g_mirror_softc *sc;
2307         u_int state;
2308
2309         sc = disk->d_softc;
2310         if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2311                 if ((disk->d_flags &
2312                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0 &&
2313                     (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 ||
2314                      (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0)) {
2315                         /* Disk does not need synchronization. */
2316                         state = G_MIRROR_DISK_STATE_ACTIVE;
2317                 } else {
2318                         if ((sc->sc_flags &
2319                              G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2320                             (disk->d_flags &
2321                              G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2322                                 /*
2323                                  * We can start synchronization from
2324                                  * the stored offset.
2325                                  */
2326                                 state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2327                         } else {
2328                                 state = G_MIRROR_DISK_STATE_STALE;
2329                         }
2330                 }
2331         } else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2332                 /*
2333                  * Reset all synchronization data for this disk,
2334                  * because if it even was synchronized, it was
2335                  * synchronized to disks with different syncid.
2336                  */
2337                 disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2338                 disk->d_sync.ds_offset = 0;
2339                 disk->d_sync.ds_offset_done = 0;
2340                 disk->d_sync.ds_syncid = sc->sc_syncid;
2341                 if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2342                     (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2343                         state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2344                 } else {
2345                         state = G_MIRROR_DISK_STATE_STALE;
2346                 }
2347         } else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2348                 /*
2349                  * Not good, NOT GOOD!
2350                  * It means that mirror was started on stale disks
2351                  * and more fresh disk just arrive.
2352                  * If there were writes, mirror is broken, sorry.
2353                  * I think the best choice here is don't touch
2354                  * this disk and inform the user loudly.
2355                  */
2356                 G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2357                     "disk (%s) arrives!! It will not be connected to the "
2358                     "running device.", sc->sc_name,
2359                     g_mirror_get_diskname(disk));
2360                 g_mirror_destroy_disk(disk);
2361                 state = G_MIRROR_DISK_STATE_NONE;
2362                 /* Return immediately, because disk was destroyed. */
2363                 return (state);
2364         }
2365         G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2366             g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2367         return (state);
2368 }
2369
2370 /*
2371  * Update device state.
2372  */
2373 static void
2374 g_mirror_update_device(struct g_mirror_softc *sc, bool force)
2375 {
2376         struct g_mirror_disk *disk;
2377         u_int state;
2378
2379         sx_assert(&sc->sc_lock, SX_XLOCKED);
2380
2381         switch (sc->sc_state) {
2382         case G_MIRROR_DEVICE_STATE_STARTING:
2383             {
2384                 struct g_mirror_disk *pdisk, *tdisk;
2385                 const char *mismatch;
2386                 uintmax_t found, newest;
2387                 u_int dirty, ndisks;
2388
2389                 /* Pre-flight checks */
2390                 LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2391                         /*
2392                          * Confirm we already detected the newest genid.
2393                          */
2394                         KASSERT(sc->sc_genid >= disk->d_genid,
2395                             ("%s: found newer genid %u (sc:%p had %u).", __func__,
2396                             disk->d_genid, sc, sc->sc_genid));
2397
2398                         /* Kick out any previously tasted stale components. */
2399                         if (disk->d_genid < sc->sc_genid) {
2400                                 G_MIRROR_DEBUG(0, "Stale 'genid' field on %s "
2401                                     "(device %s) (component=%u latest=%u), skipping.",
2402                                     g_mirror_get_diskname(disk), sc->sc_name,
2403                                     disk->d_genid, sc->sc_genid);
2404                                 g_mirror_destroy_disk(disk);
2405                                 sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2406                                 continue;
2407                         }
2408
2409                         /*
2410                          * Confirm we already detected the newest syncid.
2411                          */
2412                         KASSERT(sc->sc_syncid >= disk->d_sync.ds_syncid,
2413                             ("%s: found newer syncid %u (sc:%p had %u).",
2414                              __func__, disk->d_sync.ds_syncid, sc,
2415                              sc->sc_syncid));
2416
2417 #define DETECT_MISMATCH(field, name) \
2418                         if (mismatch == NULL &&                                 \
2419                             disk->d_init_ ## field != sc->sc_ ## field) {       \
2420                                 mismatch = name;                                \
2421                                 found = (intmax_t)disk->d_init_ ## field;       \
2422                                 newest = (intmax_t)sc->sc_ ## field;            \
2423                         }
2424                         mismatch = NULL;
2425                         DETECT_MISMATCH(ndisks, "md_all");
2426                         DETECT_MISMATCH(balance, "md_balance");
2427                         DETECT_MISMATCH(slice, "md_slice");
2428                         DETECT_MISMATCH(mediasize, "md_mediasize");
2429 #undef DETECT_MISMATCH
2430                         if (mismatch != NULL) {
2431                                 G_MIRROR_DEBUG(0, "Found a mismatching '%s' "
2432                                     "field on %s (device %s) (found=%ju "
2433                                     "newest=%ju).", mismatch,
2434                                     g_mirror_get_diskname(disk), sc->sc_name,
2435                                     found, newest);
2436                                 g_mirror_destroy_disk(disk);
2437                                 sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2438                                 continue;
2439                         }
2440                 }
2441
2442                 KASSERT(sc->sc_provider == NULL,
2443                     ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2444                 /*
2445                  * Are we ready? If the timeout (force is true) has expired, and
2446                  * any disks are present, then yes. If we're permitted to launch
2447                  * before the timeout has expired and the expected number of
2448                  * current-generation mirror disks have been tasted, then yes.
2449                  */
2450                 ndisks = g_mirror_ndisks(sc, -1);
2451                 if ((force && ndisks > 0) ||
2452                     (g_launch_mirror_before_timeout && ndisks == sc->sc_ndisks)) {
2453                         ;
2454                 } else if (ndisks == 0) {
2455                         /*
2456                          * Disks went down in starting phase, so destroy
2457                          * device.
2458                          */
2459                         callout_drain(&sc->sc_callout);
2460                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2461                         G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2462                             sc->sc_rootmount);
2463                         root_mount_rel(sc->sc_rootmount);
2464                         sc->sc_rootmount = NULL;
2465                         return;
2466                 } else {
2467                         return;
2468                 }
2469
2470                 /*
2471                  * Activate all disks with the biggest syncid.
2472                  */
2473                 if (force) {
2474                         /*
2475                          * If 'force' is true, we have been called due to
2476                          * timeout, so don't bother canceling timeout.
2477                          */
2478                         ndisks = 0;
2479                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2480                                 if ((disk->d_flags &
2481                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2482                                         ndisks++;
2483                                 }
2484                         }
2485                         if (ndisks == 0) {
2486                                 /* No valid disks found, destroy device. */
2487                                 sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2488                                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2489                                     __LINE__, sc->sc_rootmount);
2490                                 root_mount_rel(sc->sc_rootmount);
2491                                 sc->sc_rootmount = NULL;
2492                                 return;
2493                         }
2494                 } else {
2495                         /* Cancel timeout. */
2496                         callout_drain(&sc->sc_callout);
2497                 }
2498
2499                 /*
2500                  * Here we need to look for dirty disks and if all disks
2501                  * with the biggest syncid are dirty, we have to choose
2502                  * one with the biggest priority and rebuild the rest.
2503                  */
2504                 /*
2505                  * Find the number of dirty disks with the biggest syncid.
2506                  * Find the number of disks with the biggest syncid.
2507                  * While here, find a disk with the biggest priority.
2508                  */
2509                 dirty = ndisks = 0;
2510                 pdisk = NULL;
2511                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2512                         if (disk->d_sync.ds_syncid != sc->sc_syncid)
2513                                 continue;
2514                         if ((disk->d_flags &
2515                             G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2516                                 continue;
2517                         }
2518                         ndisks++;
2519                         if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2520                                 dirty++;
2521                                 if (pdisk == NULL ||
2522                                     pdisk->d_priority < disk->d_priority) {
2523                                         pdisk = disk;
2524                                 }
2525                         }
2526                 }
2527                 if (dirty == 0) {
2528                         /* No dirty disks at all, great. */
2529                 } else if (dirty == ndisks) {
2530                         /*
2531                          * Force synchronization for all dirty disks except one
2532                          * with the biggest priority.
2533                          */
2534                         KASSERT(pdisk != NULL, ("pdisk == NULL"));
2535                         G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2536                             "master disk for synchronization.",
2537                             g_mirror_get_diskname(pdisk), sc->sc_name);
2538                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2539                                 if (disk->d_sync.ds_syncid != sc->sc_syncid)
2540                                         continue;
2541                                 if ((disk->d_flags &
2542                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2543                                         continue;
2544                                 }
2545                                 KASSERT((disk->d_flags &
2546                                     G_MIRROR_DISK_FLAG_DIRTY) != 0,
2547                                     ("Disk %s isn't marked as dirty.",
2548                                     g_mirror_get_diskname(disk)));
2549                                 /* Skip the disk with the biggest priority. */
2550                                 if (disk == pdisk)
2551                                         continue;
2552                                 disk->d_sync.ds_syncid = 0;
2553                         }
2554                 } else if (dirty < ndisks) {
2555                         /*
2556                          * Force synchronization for all dirty disks.
2557                          * We have some non-dirty disks.
2558                          */
2559                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2560                                 if (disk->d_sync.ds_syncid != sc->sc_syncid)
2561                                         continue;
2562                                 if ((disk->d_flags &
2563                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2564                                         continue;
2565                                 }
2566                                 if ((disk->d_flags &
2567                                     G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2568                                         continue;
2569                                 }
2570                                 disk->d_sync.ds_syncid = 0;
2571                         }
2572                 }
2573
2574                 /* Reset hint. */
2575                 sc->sc_hint = NULL;
2576                 if (force) {
2577                         /* Remember to bump syncid on first write. */
2578                         sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2579                 }
2580                 state = G_MIRROR_DEVICE_STATE_RUNNING;
2581                 G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2582                     sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2583                     g_mirror_device_state2str(state));
2584                 sc->sc_state = state;
2585                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2586                         state = g_mirror_determine_state(disk);
2587                         g_mirror_event_send(disk, state,
2588                             G_MIRROR_EVENT_DONTWAIT);
2589                         if (state == G_MIRROR_DISK_STATE_STALE)
2590                                 sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2591                 }
2592                 break;
2593             }
2594         case G_MIRROR_DEVICE_STATE_RUNNING:
2595                 if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2596                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2597                         /*
2598                          * No usable disks, so destroy the device.
2599                          */
2600                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2601                         break;
2602                 } else if (g_mirror_ndisks(sc,
2603                     G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2604                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2605                         /*
2606                          * We have active disks, launch provider if it doesn't
2607                          * exist.
2608                          */
2609                         if (sc->sc_provider == NULL)
2610                                 g_mirror_launch_provider(sc);
2611                         if (sc->sc_rootmount != NULL) {
2612                                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2613                                     __LINE__, sc->sc_rootmount);
2614                                 root_mount_rel(sc->sc_rootmount);
2615                                 sc->sc_rootmount = NULL;
2616                         }
2617                 }
2618                 /*
2619                  * Genid should be bumped immediately, so do it here.
2620                  */
2621                 if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2622                         sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2623                         g_mirror_bump_genid(sc);
2624                 }
2625                 if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID_NOW) != 0) {
2626                         sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID_NOW;
2627                         g_mirror_bump_syncid(sc);
2628                 }
2629                 break;
2630         default:
2631                 KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2632                     sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2633                 break;
2634         }
2635 }
2636
2637 /*
2638  * Update disk state and device state if needed.
2639  */
2640 #define DISK_STATE_CHANGED()    G_MIRROR_DEBUG(1,                       \
2641         "Disk %s state changed from %s to %s (device %s).",             \
2642         g_mirror_get_diskname(disk),                                    \
2643         g_mirror_disk_state2str(disk->d_state),                         \
2644         g_mirror_disk_state2str(state), sc->sc_name)
2645 static int
2646 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2647 {
2648         struct g_mirror_softc *sc;
2649
2650         sc = disk->d_softc;
2651         sx_assert(&sc->sc_lock, SX_XLOCKED);
2652
2653 again:
2654         G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2655             g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2656             g_mirror_disk_state2str(state));
2657         switch (state) {
2658         case G_MIRROR_DISK_STATE_NEW:
2659                 /*
2660                  * Possible scenarios:
2661                  * 1. New disk arrive.
2662                  */
2663                 /* Previous state should be NONE. */
2664                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2665                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2666                     g_mirror_disk_state2str(disk->d_state)));
2667                 DISK_STATE_CHANGED();
2668
2669                 disk->d_state = state;
2670                 g_topology_lock();
2671                 if (LIST_EMPTY(&sc->sc_disks))
2672                         LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2673                 else {
2674                         struct g_mirror_disk *dp;
2675
2676                         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2677                                 if (disk->d_priority >= dp->d_priority) {
2678                                         LIST_INSERT_BEFORE(dp, disk, d_next);
2679                                         dp = NULL;
2680                                         break;
2681                                 }
2682                                 if (LIST_NEXT(dp, d_next) == NULL)
2683                                         break;
2684                         }
2685                         if (dp != NULL)
2686                                 LIST_INSERT_AFTER(dp, disk, d_next);
2687                 }
2688                 g_topology_unlock();
2689                 G_MIRROR_DEBUG(1, "Device %s: provider %s detected.",
2690                     sc->sc_name, g_mirror_get_diskname(disk));
2691                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2692                         break;
2693                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2694                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2695                     g_mirror_device_state2str(sc->sc_state),
2696                     g_mirror_get_diskname(disk),
2697                     g_mirror_disk_state2str(disk->d_state)));
2698                 state = g_mirror_determine_state(disk);
2699                 if (state != G_MIRROR_DISK_STATE_NONE)
2700                         goto again;
2701                 break;
2702         case G_MIRROR_DISK_STATE_ACTIVE:
2703                 /*
2704                  * Possible scenarios:
2705                  * 1. New disk does not need synchronization.
2706                  * 2. Synchronization process finished successfully.
2707                  */
2708                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2709                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2710                     g_mirror_device_state2str(sc->sc_state),
2711                     g_mirror_get_diskname(disk),
2712                     g_mirror_disk_state2str(disk->d_state)));
2713                 /* Previous state should be NEW or SYNCHRONIZING. */
2714                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2715                     disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2716                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2717                     g_mirror_disk_state2str(disk->d_state)));
2718                 DISK_STATE_CHANGED();
2719
2720                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2721                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2722                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2723                         g_mirror_sync_stop(disk, 0);
2724                 }
2725                 disk->d_state = state;
2726                 disk->d_sync.ds_offset = 0;
2727                 disk->d_sync.ds_offset_done = 0;
2728                 g_mirror_update_idle(sc, disk);
2729                 g_mirror_update_metadata(disk);
2730                 G_MIRROR_DEBUG(1, "Device %s: provider %s activated.",
2731                     sc->sc_name, g_mirror_get_diskname(disk));
2732                 break;
2733         case G_MIRROR_DISK_STATE_STALE:
2734                 /*
2735                  * Possible scenarios:
2736                  * 1. Stale disk was connected.
2737                  */
2738                 /* Previous state should be NEW. */
2739                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2740                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2741                     g_mirror_disk_state2str(disk->d_state)));
2742                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2743                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2744                     g_mirror_device_state2str(sc->sc_state),
2745                     g_mirror_get_diskname(disk),
2746                     g_mirror_disk_state2str(disk->d_state)));
2747                 /*
2748                  * STALE state is only possible if device is marked
2749                  * NOAUTOSYNC.
2750                  */
2751                 KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2752                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2753                     g_mirror_device_state2str(sc->sc_state),
2754                     g_mirror_get_diskname(disk),
2755                     g_mirror_disk_state2str(disk->d_state)));
2756                 DISK_STATE_CHANGED();
2757
2758                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2759                 disk->d_state = state;
2760                 g_mirror_update_metadata(disk);
2761                 G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2762                     sc->sc_name, g_mirror_get_diskname(disk));
2763                 break;
2764         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2765                 /*
2766                  * Possible scenarios:
2767                  * 1. Disk which needs synchronization was connected.
2768                  */
2769                 /* Previous state should be NEW. */
2770                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2771                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2772                     g_mirror_disk_state2str(disk->d_state)));
2773                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2774                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2775                     g_mirror_device_state2str(sc->sc_state),
2776                     g_mirror_get_diskname(disk),
2777                     g_mirror_disk_state2str(disk->d_state)));
2778                 DISK_STATE_CHANGED();
2779
2780                 if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2781                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2782                 disk->d_state = state;
2783                 if (sc->sc_provider != NULL) {
2784                         g_mirror_sync_start(disk);
2785                         g_mirror_update_metadata(disk);
2786                 }
2787                 break;
2788         case G_MIRROR_DISK_STATE_DISCONNECTED:
2789                 /*
2790                  * Possible scenarios:
2791                  * 1. Device wasn't running yet, but disk disappear.
2792                  * 2. Disk was active and disapppear.
2793                  * 3. Disk disappear during synchronization process.
2794                  */
2795                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2796                         /*
2797                          * Previous state should be ACTIVE, STALE or
2798                          * SYNCHRONIZING.
2799                          */
2800                         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2801                             disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2802                             disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2803                             ("Wrong disk state (%s, %s).",
2804                             g_mirror_get_diskname(disk),
2805                             g_mirror_disk_state2str(disk->d_state)));
2806                 } else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2807                         /* Previous state should be NEW. */
2808                         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2809                             ("Wrong disk state (%s, %s).",
2810                             g_mirror_get_diskname(disk),
2811                             g_mirror_disk_state2str(disk->d_state)));
2812                         /*
2813                          * Reset bumping syncid if disk disappeared in STARTING
2814                          * state.
2815                          */
2816                         if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2817                                 sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2818 #ifdef  INVARIANTS
2819                 } else {
2820                         KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2821                             sc->sc_name,
2822                             g_mirror_device_state2str(sc->sc_state),
2823                             g_mirror_get_diskname(disk),
2824                             g_mirror_disk_state2str(disk->d_state)));
2825 #endif
2826                 }
2827                 DISK_STATE_CHANGED();
2828                 G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2829                     sc->sc_name, g_mirror_get_diskname(disk));
2830
2831                 g_mirror_destroy_disk(disk);
2832                 break;
2833         case G_MIRROR_DISK_STATE_DESTROY:
2834             {
2835                 int error;
2836
2837                 error = g_mirror_clear_metadata(disk);
2838                 if (error != 0) {
2839                         G_MIRROR_DEBUG(0,
2840                             "Device %s: failed to clear metadata on %s: %d.",
2841                             sc->sc_name, g_mirror_get_diskname(disk), error);
2842                         break;
2843                 }
2844                 DISK_STATE_CHANGED();
2845                 G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2846                     sc->sc_name, g_mirror_get_diskname(disk));
2847
2848                 g_mirror_destroy_disk(disk);
2849                 sc->sc_ndisks--;
2850                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2851                         g_mirror_update_metadata(disk);
2852                 }
2853                 break;
2854             }
2855         default:
2856                 KASSERT(1 == 0, ("Unknown state (%u).", state));
2857                 break;
2858         }
2859         return (0);
2860 }
2861 #undef  DISK_STATE_CHANGED
2862
2863 int
2864 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2865 {
2866         struct g_provider *pp;
2867         u_char *buf;
2868         int error;
2869
2870         g_topology_assert();
2871
2872         error = g_access(cp, 1, 0, 0);
2873         if (error != 0)
2874                 return (error);
2875         pp = cp->provider;
2876         g_topology_unlock();
2877         /* Metadata are stored on last sector. */
2878         buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2879             &error);
2880         g_topology_lock();
2881         g_access(cp, -1, 0, 0);
2882         if (buf == NULL) {
2883                 G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2884                     cp->provider->name, error);
2885                 return (error);
2886         }
2887
2888         /* Decode metadata. */
2889         error = mirror_metadata_decode(buf, md);
2890         g_free(buf);
2891         if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2892                 return (EINVAL);
2893         if (md->md_version > G_MIRROR_VERSION) {
2894                 G_MIRROR_DEBUG(0,
2895                     "Kernel module is too old to handle metadata from %s.",
2896                     cp->provider->name);
2897                 return (EINVAL);
2898         }
2899         if (error != 0) {
2900                 G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2901                     cp->provider->name);
2902                 return (error);
2903         }
2904
2905         return (0);
2906 }
2907
2908 static int
2909 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2910     struct g_mirror_metadata *md)
2911 {
2912
2913         G_MIRROR_DEBUG(2, "%s: md_did 0x%u disk %s device %s md_all 0x%x "
2914             "sc_ndisks 0x%x md_slice 0x%x sc_slice 0x%x md_balance 0x%x "
2915             "sc_balance 0x%x sc_mediasize 0x%jx pp_mediasize 0x%jx "
2916             "md_sectorsize 0x%x sc_sectorsize 0x%x md_mflags 0x%jx "
2917             "md_dflags 0x%jx md_syncid 0x%x md_genid 0x%x md_priority 0x%x "
2918             "sc_state 0x%x.",
2919             __func__, md->md_did, pp->name, sc->sc_name, md->md_all,
2920             sc->sc_ndisks, md->md_slice, sc->sc_slice, md->md_balance,
2921             sc->sc_balance, (uintmax_t)sc->sc_mediasize,
2922             (uintmax_t)pp->mediasize, md->md_sectorsize, sc->sc_sectorsize,
2923             (uintmax_t)md->md_mflags, (uintmax_t)md->md_dflags, md->md_syncid,
2924             md->md_genid, md->md_priority, sc->sc_state);
2925
2926         if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2927                 G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2928                     pp->name, md->md_did);
2929                 return (EEXIST);
2930         }
2931         if (sc->sc_mediasize > pp->mediasize) {
2932                 G_MIRROR_DEBUG(1,
2933                     "Invalid size of disk %s (device %s), skipping.", pp->name,
2934                     sc->sc_name);
2935                 return (EINVAL);
2936         }
2937         if (md->md_sectorsize != sc->sc_sectorsize) {
2938                 G_MIRROR_DEBUG(1,
2939                     "Invalid '%s' field on disk %s (device %s), skipping.",
2940                     "md_sectorsize", pp->name, sc->sc_name);
2941                 return (EINVAL);
2942         }
2943         if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2944                 G_MIRROR_DEBUG(1,
2945                     "Invalid sector size of disk %s (device %s), skipping.",
2946                     pp->name, sc->sc_name);
2947                 return (EINVAL);
2948         }
2949         if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
2950                 G_MIRROR_DEBUG(1,
2951                     "Invalid device flags on disk %s (device %s), skipping.",
2952                     pp->name, sc->sc_name);
2953                 return (EINVAL);
2954         }
2955         if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
2956                 G_MIRROR_DEBUG(1,
2957                     "Invalid disk flags on disk %s (device %s), skipping.",
2958                     pp->name, sc->sc_name);
2959                 return (EINVAL);
2960         }
2961         return (0);
2962 }
2963
2964 int
2965 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
2966     struct g_mirror_metadata *md)
2967 {
2968         struct g_mirror_disk *disk;
2969         int error;
2970
2971         g_topology_assert_not();
2972         G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
2973
2974         error = g_mirror_check_metadata(sc, pp, md);
2975         if (error != 0)
2976                 return (error);
2977
2978         if (md->md_genid < sc->sc_genid) {
2979                 G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
2980                     pp->name, sc->sc_name);
2981                 return (EINVAL);
2982         }
2983
2984         /*
2985          * If the component disk we're tasting has newer metadata than the
2986          * STARTING gmirror device, refresh the device from the component.
2987          */
2988         error = g_mirror_refresh_device(sc, pp, md);
2989         if (error != 0)
2990                 return (error);
2991
2992         disk = g_mirror_init_disk(sc, pp, md, &error);
2993         if (disk == NULL)
2994                 return (error);
2995         error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
2996             G_MIRROR_EVENT_WAIT);
2997         if (error != 0)
2998                 return (error);
2999         if (md->md_version < G_MIRROR_VERSION) {
3000                 G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
3001                     pp->name, md->md_version, G_MIRROR_VERSION);
3002                 g_mirror_update_metadata(disk);
3003         }
3004         return (0);
3005 }
3006
3007 static void
3008 g_mirror_destroy_delayed(void *arg, int flag)
3009 {
3010         struct g_mirror_softc *sc;
3011         int error;
3012
3013         if (flag == EV_CANCEL) {
3014                 G_MIRROR_DEBUG(1, "Destroying canceled.");
3015                 return;
3016         }
3017         sc = arg;
3018         g_topology_unlock();
3019         sx_xlock(&sc->sc_lock);
3020         KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
3021             ("DESTROY flag set on %s.", sc->sc_name));
3022         KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_CLOSEWAIT) != 0,
3023             ("CLOSEWAIT flag not set on %s.", sc->sc_name));
3024         G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
3025         error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
3026         if (error != 0) {
3027                 G_MIRROR_DEBUG(0, "Cannot destroy %s (error=%d).",
3028                     sc->sc_name, error);
3029                 sx_xunlock(&sc->sc_lock);
3030         }
3031         g_topology_lock();
3032 }
3033
3034 static int
3035 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
3036 {
3037         struct g_mirror_softc *sc;
3038         int error = 0;
3039
3040         g_topology_assert();
3041         G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
3042             acw, ace);
3043
3044         sc = pp->private;
3045         KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
3046
3047         g_topology_unlock();
3048         sx_xlock(&sc->sc_lock);
3049         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
3050             (sc->sc_flags & G_MIRROR_DEVICE_FLAG_CLOSEWAIT) != 0 ||
3051             LIST_EMPTY(&sc->sc_disks)) {
3052                 if (acr > 0 || acw > 0 || ace > 0)
3053                         error = ENXIO;
3054                 goto end;
3055         }
3056         sc->sc_provider_open += acr + acw + ace;
3057         if (pp->acw + acw == 0)
3058                 g_mirror_idle(sc, 0);
3059         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_CLOSEWAIT) != 0 &&
3060             sc->sc_provider_open == 0)
3061                 g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK, sc, NULL);
3062 end:
3063         sx_xunlock(&sc->sc_lock);
3064         g_topology_lock();
3065         return (error);
3066 }
3067
3068 static void
3069 g_mirror_reinit_from_metadata(struct g_mirror_softc *sc,
3070     const struct g_mirror_metadata *md)
3071 {
3072
3073         sc->sc_genid = md->md_genid;
3074         sc->sc_syncid = md->md_syncid;
3075
3076         sc->sc_slice = md->md_slice;
3077         sc->sc_balance = md->md_balance;
3078         sc->sc_mediasize = md->md_mediasize;
3079         sc->sc_ndisks = md->md_all;
3080         sc->sc_flags &= ~G_MIRROR_DEVICE_FLAG_MASK;
3081         sc->sc_flags |= (md->md_mflags & G_MIRROR_DEVICE_FLAG_MASK);
3082 }
3083
3084 struct g_geom *
3085 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md,
3086     u_int type)
3087 {
3088         struct g_mirror_softc *sc;
3089         struct g_geom *gp;
3090         int error, timeout;
3091
3092         g_topology_assert();
3093         G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
3094             md->md_mid);
3095
3096         /* One disk is minimum. */
3097         if (md->md_all < 1)
3098                 return (NULL);
3099         /*
3100          * Action geom.
3101          */
3102         gp = g_new_geomf(mp, "%s", md->md_name);
3103         sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
3104         gp->start = g_mirror_start;
3105         gp->orphan = g_mirror_orphan;
3106         gp->access = g_mirror_access;
3107         gp->dumpconf = g_mirror_dumpconf;
3108
3109         sc->sc_type = type;
3110         sc->sc_id = md->md_mid;
3111         g_mirror_reinit_from_metadata(sc, md);
3112         sc->sc_sectorsize = md->md_sectorsize;
3113         sc->sc_bump_id = 0;
3114         sc->sc_idle = 1;
3115         sc->sc_last_write = time_uptime;
3116         sc->sc_writes = 0;
3117         sc->sc_refcnt = 1;
3118         sx_init(&sc->sc_lock, "gmirror:lock");
3119         TAILQ_INIT(&sc->sc_queue);
3120         mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
3121         TAILQ_INIT(&sc->sc_regular_delayed);
3122         TAILQ_INIT(&sc->sc_inflight);
3123         TAILQ_INIT(&sc->sc_sync_delayed);
3124         LIST_INIT(&sc->sc_disks);
3125         TAILQ_INIT(&sc->sc_events);
3126         mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
3127         callout_init(&sc->sc_callout, 1);
3128         mtx_init(&sc->sc_done_mtx, "gmirror:done", NULL, MTX_DEF);
3129         sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
3130         gp->softc = sc;
3131         sc->sc_geom = gp;
3132         sc->sc_provider = NULL;
3133         sc->sc_provider_open = 0;
3134         /*
3135          * Synchronization geom.
3136          */
3137         gp = g_new_geomf(mp, "%s.sync", md->md_name);
3138         gp->softc = sc;
3139         gp->orphan = g_mirror_orphan;
3140         sc->sc_sync.ds_geom = gp;
3141         sc->sc_sync.ds_ndisks = 0;
3142         error = kproc_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
3143             "g_mirror %s", md->md_name);
3144         if (error != 0) {
3145                 G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
3146                     sc->sc_name);
3147                 g_destroy_geom(sc->sc_sync.ds_geom);
3148                 g_destroy_geom(sc->sc_geom);
3149                 g_mirror_free_device(sc);
3150                 return (NULL);
3151         }
3152
3153         G_MIRROR_DEBUG(1, "Device %s created (%u components, id=%u).",
3154             sc->sc_name, sc->sc_ndisks, sc->sc_id);
3155
3156         sc->sc_rootmount = root_mount_hold("GMIRROR");
3157         G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
3158         /*
3159          * Run timeout.
3160          */
3161         timeout = g_mirror_timeout * hz;
3162         callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
3163         return (sc->sc_geom);
3164 }
3165
3166 int
3167 g_mirror_destroy(struct g_mirror_softc *sc, int how)
3168 {
3169         struct g_mirror_disk *disk;
3170
3171         g_topology_assert_not();
3172         sx_assert(&sc->sc_lock, SX_XLOCKED);
3173
3174         if (sc->sc_provider_open != 0) {
3175                 switch (how) {
3176                 case G_MIRROR_DESTROY_SOFT:
3177                         G_MIRROR_DEBUG(1,
3178                             "Device %s is still open (%d).", sc->sc_name,
3179                             sc->sc_provider_open);
3180                         return (EBUSY);
3181                 case G_MIRROR_DESTROY_DELAYED:
3182                         G_MIRROR_DEBUG(1,
3183                             "Device %s will be destroyed on last close.",
3184                             sc->sc_name);
3185                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
3186                                 if (disk->d_state ==
3187                                     G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3188                                         g_mirror_sync_stop(disk, 1);
3189                                 }
3190                         }
3191                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_CLOSEWAIT;
3192                         return (EBUSY);
3193                 case G_MIRROR_DESTROY_HARD:
3194                         G_MIRROR_DEBUG(1, "Device %s is still open, so it "
3195                             "can't be definitely removed.", sc->sc_name);
3196                 }
3197         }
3198
3199         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
3200                 sx_xunlock(&sc->sc_lock);
3201                 return (0);
3202         }
3203         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
3204         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DRAIN;
3205         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
3206         sx_xunlock(&sc->sc_lock);
3207         mtx_lock(&sc->sc_queue_mtx);
3208         wakeup(sc);
3209         mtx_unlock(&sc->sc_queue_mtx);
3210         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
3211         while (sc->sc_worker != NULL)
3212                 tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
3213         G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
3214         sx_xlock(&sc->sc_lock);
3215         g_mirror_destroy_device(sc);
3216         return (0);
3217 }
3218
3219 static void
3220 g_mirror_taste_orphan(struct g_consumer *cp)
3221 {
3222
3223         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3224             cp->provider->name));
3225 }
3226
3227 static struct g_geom *
3228 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3229 {
3230         struct g_mirror_metadata md;
3231         struct g_mirror_softc *sc;
3232         struct g_consumer *cp;
3233         struct g_geom *gp;
3234         int error;
3235
3236         g_topology_assert();
3237         g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3238         G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
3239
3240         gp = g_new_geomf(mp, "mirror:taste");
3241         /*
3242          * This orphan function should be never called.
3243          */
3244         gp->orphan = g_mirror_taste_orphan;
3245         cp = g_new_consumer(gp);
3246         g_attach(cp, pp);
3247         error = g_mirror_read_metadata(cp, &md);
3248         g_detach(cp);
3249         g_destroy_consumer(cp);
3250         g_destroy_geom(gp);
3251         if (error != 0)
3252                 return (NULL);
3253         gp = NULL;
3254
3255         if (md.md_provider[0] != '\0' &&
3256             !g_compare_names(md.md_provider, pp->name))
3257                 return (NULL);
3258         if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3259                 return (NULL);
3260         if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
3261                 G_MIRROR_DEBUG(0,
3262                     "Device %s: provider %s marked as inactive, skipping.",
3263                     md.md_name, pp->name);
3264                 return (NULL);
3265         }
3266         if (g_mirror_debug >= 2)
3267                 mirror_metadata_dump(&md);
3268
3269         /*
3270          * Let's check if device already exists.
3271          */
3272         sc = NULL;
3273         LIST_FOREACH(gp, &mp->geom, geom) {
3274                 sc = gp->softc;
3275                 if (sc == NULL)
3276                         continue;
3277                 if (sc->sc_type != G_MIRROR_TYPE_AUTOMATIC)
3278                         continue;
3279                 if (sc->sc_sync.ds_geom == gp)
3280                         continue;
3281                 if (strcmp(md.md_name, sc->sc_name) != 0)
3282                         continue;
3283                 if (md.md_mid != sc->sc_id) {
3284                         G_MIRROR_DEBUG(0, "Device %s already configured.",
3285                             sc->sc_name);
3286                         return (NULL);
3287                 }
3288                 break;
3289         }
3290         if (gp == NULL) {
3291                 gp = g_mirror_create(mp, &md, G_MIRROR_TYPE_AUTOMATIC);
3292                 if (gp == NULL) {
3293                         G_MIRROR_DEBUG(0, "Cannot create device %s.",
3294                             md.md_name);
3295                         return (NULL);
3296                 }
3297                 sc = gp->softc;
3298         }
3299         G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3300         g_topology_unlock();
3301         sx_xlock(&sc->sc_lock);
3302         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_TASTING;
3303         error = g_mirror_add_disk(sc, pp, &md);
3304         sc->sc_flags &= ~G_MIRROR_DEVICE_FLAG_TASTING;
3305         if (error != 0) {
3306                 G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3307                     pp->name, gp->name, error);
3308                 if (LIST_EMPTY(&sc->sc_disks)) {
3309                         g_cancel_event(sc);
3310                         g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3311                         g_topology_lock();
3312                         return (NULL);
3313                 }
3314                 gp = NULL;
3315         }
3316         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
3317                 g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3318                 g_topology_lock();
3319                 return (NULL);
3320         }
3321         sx_xunlock(&sc->sc_lock);
3322         g_topology_lock();
3323         return (gp);
3324 }
3325
3326 static void
3327 g_mirror_resize(struct g_consumer *cp)
3328 {
3329         struct g_mirror_disk *disk;
3330
3331         g_topology_assert();
3332         g_trace(G_T_TOPOLOGY, "%s(%s)", __func__, cp->provider->name);
3333
3334         disk = cp->private;
3335         if (disk == NULL)
3336                 return;
3337         g_topology_unlock();
3338         g_mirror_update_metadata(disk);
3339         g_topology_lock();
3340 }
3341
3342 static int
3343 g_mirror_destroy_geom(struct gctl_req *req __unused,
3344     struct g_class *mp __unused, struct g_geom *gp)
3345 {
3346         struct g_mirror_softc *sc;
3347         int error;
3348
3349         g_topology_unlock();
3350         sc = gp->softc;
3351         sx_xlock(&sc->sc_lock);
3352         g_cancel_event(sc);
3353         error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3354         if (error != 0)
3355                 sx_xunlock(&sc->sc_lock);
3356         g_topology_lock();
3357         return (error);
3358 }
3359
3360 static void
3361 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3362     struct g_consumer *cp, struct g_provider *pp)
3363 {
3364         struct g_mirror_softc *sc;
3365
3366         g_topology_assert();
3367
3368         sc = gp->softc;
3369         if (sc == NULL)
3370                 return;
3371         /* Skip synchronization geom. */
3372         if (gp == sc->sc_sync.ds_geom)
3373                 return;
3374         if (pp != NULL) {
3375                 /* Nothing here. */
3376         } else if (cp != NULL) {
3377                 struct g_mirror_disk *disk;
3378
3379                 disk = cp->private;
3380                 if (disk == NULL)
3381                         return;
3382                 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3383                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3384                         sbuf_printf(sb, "%s<Synchronized>", indent);
3385                         if (disk->d_sync.ds_offset == 0)
3386                                 sbuf_cat(sb, "0%");
3387                         else
3388                                 sbuf_printf(sb, "%u%%",
3389                                     (u_int)((disk->d_sync.ds_offset * 100) /
3390                                     sc->sc_mediasize));
3391                         sbuf_cat(sb, "</Synchronized>\n");
3392                         if (disk->d_sync.ds_offset > 0)
3393                                 sbuf_printf(sb, "%s<BytesSynced>%jd"
3394                                     "</BytesSynced>\n", indent,
3395                                     (intmax_t)disk->d_sync.ds_offset);
3396                 }
3397                 sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3398                     disk->d_sync.ds_syncid);
3399                 sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3400                     disk->d_genid);
3401                 sbuf_printf(sb, "%s<Flags>", indent);
3402                 if (disk->d_flags == 0)
3403                         sbuf_cat(sb, "NONE");
3404                 else {
3405                         int first = 1;
3406
3407 #define ADD_FLAG(flag, name)    do {                                    \
3408         if ((disk->d_flags & (flag)) != 0) {                            \
3409                 if (!first)                                             \
3410                         sbuf_cat(sb, ", ");                             \
3411                 else                                                    \
3412                         first = 0;                                      \
3413                 sbuf_cat(sb, name);                                     \
3414         }                                                               \
3415 } while (0)
3416                         ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3417                         ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3418                         ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3419                         ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3420                             "SYNCHRONIZING");
3421                         ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3422                         ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3423 #undef  ADD_FLAG
3424                 }
3425                 sbuf_cat(sb, "</Flags>\n");
3426                 sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3427                     disk->d_priority);
3428                 sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3429                     g_mirror_disk_state2str(disk->d_state));
3430         } else {
3431                 sbuf_printf(sb, "%s<Type>", indent);
3432                 switch (sc->sc_type) {
3433                 case G_MIRROR_TYPE_AUTOMATIC:
3434                         sbuf_cat(sb, "AUTOMATIC");
3435                         break;
3436                 case G_MIRROR_TYPE_MANUAL:
3437                         sbuf_cat(sb, "MANUAL");
3438                         break;
3439                 default:
3440                         sbuf_cat(sb, "UNKNOWN");
3441                         break;
3442                 }
3443                 sbuf_cat(sb, "</Type>\n");
3444                 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3445                 sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3446                 sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3447                 sbuf_printf(sb, "%s<Flags>", indent);
3448                 if (sc->sc_flags == 0)
3449                         sbuf_cat(sb, "NONE");
3450                 else {
3451                         int first = 1;
3452
3453 #define ADD_FLAG(flag, name)    do {                                    \
3454         if ((sc->sc_flags & (flag)) != 0) {                             \
3455                 if (!first)                                             \
3456                         sbuf_cat(sb, ", ");                             \
3457                 else                                                    \
3458                         first = 0;                                      \
3459                 sbuf_cat(sb, name);                                     \
3460         }                                                               \
3461 } while (0)
3462                         ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3463                         ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3464 #undef  ADD_FLAG
3465                 }
3466                 sbuf_cat(sb, "</Flags>\n");
3467                 sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3468                     (u_int)sc->sc_slice);
3469                 sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3470                     balance_name(sc->sc_balance));
3471                 sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3472                     sc->sc_ndisks);
3473                 sbuf_printf(sb, "%s<State>", indent);
3474                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3475                         sbuf_printf(sb, "%s", "STARTING");
3476                 else if (sc->sc_ndisks ==
3477                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3478                         sbuf_printf(sb, "%s", "COMPLETE");
3479                 else
3480                         sbuf_printf(sb, "%s", "DEGRADED");
3481                 sbuf_cat(sb, "</State>\n");
3482         }
3483 }
3484
3485 static void
3486 g_mirror_shutdown_post_sync(void *arg, int howto)
3487 {
3488         struct g_class *mp;
3489         struct g_geom *gp, *gp2;
3490         struct g_mirror_softc *sc;
3491         int error;
3492
3493         if (KERNEL_PANICKED())
3494                 return;
3495
3496         mp = arg;
3497         g_topology_lock();
3498         g_mirror_shutdown = 1;
3499         LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3500                 if ((sc = gp->softc) == NULL)
3501                         continue;
3502                 /* Skip synchronization geom. */
3503                 if (gp == sc->sc_sync.ds_geom)
3504                         continue;
3505                 g_topology_unlock();
3506                 sx_xlock(&sc->sc_lock);
3507                 g_mirror_idle(sc, -1);
3508                 g_cancel_event(sc);
3509                 error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3510                 if (error != 0)
3511                         sx_xunlock(&sc->sc_lock);
3512                 g_topology_lock();
3513         }
3514         g_topology_unlock();
3515 }
3516
3517 static void
3518 g_mirror_init(struct g_class *mp)
3519 {
3520
3521         g_mirror_post_sync = EVENTHANDLER_REGISTER(shutdown_post_sync,
3522             g_mirror_shutdown_post_sync, mp, SHUTDOWN_PRI_FIRST);
3523         if (g_mirror_post_sync == NULL)
3524                 G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3525 }
3526
3527 static void
3528 g_mirror_fini(struct g_class *mp)
3529 {
3530
3531         if (g_mirror_post_sync != NULL)
3532                 EVENTHANDLER_DEREGISTER(shutdown_post_sync, g_mirror_post_sync);
3533 }
3534
3535 /*
3536  * Refresh the mirror device's metadata when gmirror encounters a newer
3537  * generation as the individual components are being added to the mirror set.
3538  */
3539 static int
3540 g_mirror_refresh_device(struct g_mirror_softc *sc, const struct g_provider *pp,
3541     const struct g_mirror_metadata *md)
3542 {
3543
3544         g_topology_assert_not();
3545         sx_assert(&sc->sc_lock, SX_XLOCKED);
3546
3547         KASSERT(sc->sc_genid <= md->md_genid,
3548             ("%s: attempted to refresh from stale component %s (device %s) "
3549             "(%u < %u).", __func__, pp->name, sc->sc_name, md->md_genid,
3550             sc->sc_genid));
3551
3552         if (sc->sc_genid > md->md_genid || (sc->sc_genid == md->md_genid &&
3553             sc->sc_syncid >= md->md_syncid))
3554                 return (0);
3555
3556         G_MIRROR_DEBUG(0, "Found newer version for device %s (genid: curr=%u "
3557             "new=%u; syncid: curr=%u new=%u; ndisks: curr=%u new=%u; "
3558             "provider=%s).", sc->sc_name, sc->sc_genid, md->md_genid,
3559             sc->sc_syncid, md->md_syncid, sc->sc_ndisks, md->md_all, pp->name);
3560
3561         if (sc->sc_state != G_MIRROR_DEVICE_STATE_STARTING) {
3562                 /* Probable data corruption detected */
3563                 G_MIRROR_DEBUG(0, "Cannot refresh metadata in %s state "
3564                     "(device=%s genid=%u). A stale mirror device was launched.",
3565                     g_mirror_device_state2str(sc->sc_state), sc->sc_name,
3566                     sc->sc_genid);
3567                 return (EINVAL);
3568         }
3569
3570         /* Update softc */
3571         g_mirror_reinit_from_metadata(sc, md);
3572
3573         G_MIRROR_DEBUG(1, "Refresh device %s (id=%u, state=%s) from disk %s "
3574             "(genid=%u syncid=%u md_all=%u).", sc->sc_name, md->md_mid,
3575             g_mirror_device_state2str(sc->sc_state), pp->name, md->md_genid,
3576             md->md_syncid, (unsigned)md->md_all);
3577
3578         return (0);
3579 }
3580
3581 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);
3582 MODULE_VERSION(geom_mirror, 0);