sys/geom/mirror/g_mirror.c

   1 /*-
   2  * Copyright (c) 2004-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include <sys/param.h>
  31 #include <sys/systm.h>
  32 #include <sys/kernel.h>
  33 #include <sys/module.h>
  34 #include <sys/limits.h>
  35 #include <sys/lock.h>
  36 #include <sys/mutex.h>
  37 #include <sys/bio.h>
  38 #include <sys/sbuf.h>
  39 #include <sys/sysctl.h>
  40 #include <sys/malloc.h>
  41 #include <sys/eventhandler.h>
  42 #include <vm/uma.h>
  43 #include <geom/geom.h>
  44 #include <sys/proc.h>
  45 #include <sys/kthread.h>
  46 #include <sys/sched.h>
  47 #include <geom/mirror/g_mirror.h>
  48
  49 FEATURE(geom_mirror, "GEOM mirroring support");
  50
  51 static MALLOC_DEFINE(M_MIRROR, "mirror_data", "GEOM_MIRROR Data");
  52
  53 SYSCTL_DECL(_kern_geom);
  54 static SYSCTL_NODE(_kern_geom, OID_AUTO, mirror, CTLFLAG_RW, 0,
  55     "GEOM_MIRROR stuff");
  56 u_int g_mirror_debug = 0;
  57 TUNABLE_INT("kern.geom.mirror.debug", &g_mirror_debug);
  58 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, debug, CTLFLAG_RW, &g_mirror_debug, 0,
  59     "Debug level");
  60 static u_int g_mirror_timeout = 4;
  61 TUNABLE_INT("kern.geom.mirror.timeout", &g_mirror_timeout);
  62 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, timeout, CTLFLAG_RW, &g_mirror_timeout,
  63     0, "Time to wait on all mirror components");
  64 static u_int g_mirror_idletime = 5;
  65 TUNABLE_INT("kern.geom.mirror.idletime", &g_mirror_idletime);
  66 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, idletime, CTLFLAG_RW,
  67     &g_mirror_idletime, 0, "Mark components as clean when idling");
  68 static u_int g_mirror_disconnect_on_failure = 1;
  69 TUNABLE_INT("kern.geom.mirror.disconnect_on_failure",
  70     &g_mirror_disconnect_on_failure);
  71 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, disconnect_on_failure, CTLFLAG_RW,
  72     &g_mirror_disconnect_on_failure, 0, "Disconnect component on I/O failure.");
  73 static u_int g_mirror_syncreqs = 2;
  74 TUNABLE_INT("kern.geom.mirror.sync_requests", &g_mirror_syncreqs);
  75 SYSCTL_UINT(_kern_geom_mirror, OID_AUTO, sync_requests, CTLFLAG_RDTUN,
  76     &g_mirror_syncreqs, 0, "Parallel synchronization I/O requests.");
  77
  78 #define MSLEEP(ident, mtx, priority, wmesg, timeout)    do {            \
  79         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, (ident));       \
  80         msleep((ident), (mtx), (priority), (wmesg), (timeout));         \
  81         G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, (ident));       \
  82 } while (0)
  83
  84 static eventhandler_tag g_mirror_post_sync = NULL;
  85 static int g_mirror_shutdown = 0;
  86
  87 static int g_mirror_destroy_geom(struct gctl_req *req, struct g_class *mp,
  88     struct g_geom *gp);
  89 static g_taste_t g_mirror_taste;
  90 static void g_mirror_init(struct g_class *mp);
  91 static void g_mirror_fini(struct g_class *mp);
  92
  93 struct g_class g_mirror_class = {
  94         .name = G_MIRROR_CLASS_NAME,
  95         .version = G_VERSION,
  96         .ctlreq = g_mirror_config,
  97         .taste = g_mirror_taste,
  98         .destroy_geom = g_mirror_destroy_geom,
  99         .init = g_mirror_init,
 100         .fini = g_mirror_fini
 101 };
 102
 103
 104 static void g_mirror_destroy_provider(struct g_mirror_softc *sc);
 105 static int g_mirror_update_disk(struct g_mirror_disk *disk, u_int state);
 106 static void g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force);
 107 static void g_mirror_dumpconf(struct sbuf *sb, const char *indent,
 108     struct g_geom *gp, struct g_consumer *cp, struct g_provider *pp);
 109 static void g_mirror_sync_stop(struct g_mirror_disk *disk, int type);
 110 static void g_mirror_register_request(struct bio *bp);
 111 static void g_mirror_sync_release(struct g_mirror_softc *sc);
 112
 113
 114 static const char *
 115 g_mirror_disk_state2str(int state)
 116 {
 117
 118         switch (state) {
 119         case G_MIRROR_DISK_STATE_NONE:
 120                 return ("NONE");
 121         case G_MIRROR_DISK_STATE_NEW:
 122                 return ("NEW");
 123         case G_MIRROR_DISK_STATE_ACTIVE:
 124                 return ("ACTIVE");
 125         case G_MIRROR_DISK_STATE_STALE:
 126                 return ("STALE");
 127         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
 128                 return ("SYNCHRONIZING");
 129         case G_MIRROR_DISK_STATE_DISCONNECTED:
 130                 return ("DISCONNECTED");
 131         case G_MIRROR_DISK_STATE_DESTROY:
 132                 return ("DESTROY");
 133         default:
 134                 return ("INVALID");
 135         }
 136 }
 137
 138 static const char *
 139 g_mirror_device_state2str(int state)
 140 {
 141
 142         switch (state) {
 143         case G_MIRROR_DEVICE_STATE_STARTING:
 144                 return ("STARTING");
 145         case G_MIRROR_DEVICE_STATE_RUNNING:
 146                 return ("RUNNING");
 147         default:
 148                 return ("INVALID");
 149         }
 150 }
 151
 152 static const char *
 153 g_mirror_get_diskname(struct g_mirror_disk *disk)
 154 {
 155
 156         if (disk->d_consumer == NULL || disk->d_consumer->provider == NULL)
 157                 return ("[unknown]");
 158         return (disk->d_name);
 159 }
 160
 161 /*
 162  * --- Events handling functions ---
 163  * Events in geom_mirror are used to maintain disks and device status
 164  * from one thread to simplify locking.
 165  */
 166 static void
 167 g_mirror_event_free(struct g_mirror_event *ep)
 168 {
 169
 170         free(ep, M_MIRROR);
 171 }
 172
 173 int
 174 g_mirror_event_send(void *arg, int state, int flags)
 175 {
 176         struct g_mirror_softc *sc;
 177         struct g_mirror_disk *disk;
 178         struct g_mirror_event *ep;
 179         int error;
 180
 181         ep = malloc(sizeof(*ep), M_MIRROR, M_WAITOK);
 182         G_MIRROR_DEBUG(4, "%s: Sending event %p.", __func__, ep);
 183         if ((flags & G_MIRROR_EVENT_DEVICE) != 0) {
 184                 disk = NULL;
 185                 sc = arg;
 186         } else {
 187                 disk = arg;
 188                 sc = disk->d_softc;
 189         }
 190         ep->e_disk = disk;
 191         ep->e_state = state;
 192         ep->e_flags = flags;
 193         ep->e_error = 0;
 194         mtx_lock(&sc->sc_events_mtx);
 195         TAILQ_INSERT_TAIL(&sc->sc_events, ep, e_next);
 196         mtx_unlock(&sc->sc_events_mtx);
 197         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
 198         mtx_lock(&sc->sc_queue_mtx);
 199         wakeup(sc);
 200         mtx_unlock(&sc->sc_queue_mtx);
 201         if ((flags & G_MIRROR_EVENT_DONTWAIT) != 0)
 202                 return (0);
 203         sx_assert(&sc->sc_lock, SX_XLOCKED);
 204         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, ep);
 205         sx_xunlock(&sc->sc_lock);
 206         while ((ep->e_flags & G_MIRROR_EVENT_DONE) == 0) {
 207                 mtx_lock(&sc->sc_events_mtx);
 208                 MSLEEP(ep, &sc->sc_events_mtx, PRIBIO | PDROP, "m:event",
 209                     hz * 5);
 210         }
 211         error = ep->e_error;
 212         g_mirror_event_free(ep);
 213         sx_xlock(&sc->sc_lock);
 214         return (error);
 215 }
 216
 217 static struct g_mirror_event *
 218 g_mirror_event_get(struct g_mirror_softc *sc)
 219 {
 220         struct g_mirror_event *ep;
 221
 222         mtx_lock(&sc->sc_events_mtx);
 223         ep = TAILQ_FIRST(&sc->sc_events);
 224         mtx_unlock(&sc->sc_events_mtx);
 225         return (ep);
 226 }
 227
 228 static void
 229 g_mirror_event_remove(struct g_mirror_softc *sc, struct g_mirror_event *ep)
 230 {
 231
 232         mtx_lock(&sc->sc_events_mtx);
 233         TAILQ_REMOVE(&sc->sc_events, ep, e_next);
 234         mtx_unlock(&sc->sc_events_mtx);
 235 }
 236
 237 static void
 238 g_mirror_event_cancel(struct g_mirror_disk *disk)
 239 {
 240         struct g_mirror_softc *sc;
 241         struct g_mirror_event *ep, *tmpep;
 242
 243         sc = disk->d_softc;
 244         sx_assert(&sc->sc_lock, SX_XLOCKED);
 245
 246         mtx_lock(&sc->sc_events_mtx);
 247         TAILQ_FOREACH_SAFE(ep, &sc->sc_events, e_next, tmpep) {
 248                 if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0)
 249                         continue;
 250                 if (ep->e_disk != disk)
 251                         continue;
 252                 TAILQ_REMOVE(&sc->sc_events, ep, e_next);
 253                 if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
 254                         g_mirror_event_free(ep);
 255                 else {
 256                         ep->e_error = ECANCELED;
 257                         wakeup(ep);
 258                 }
 259         }
 260         mtx_unlock(&sc->sc_events_mtx);
 261 }
 262
 263 /*
 264  * Return the number of disks in given state.
 265  * If state is equal to -1, count all connected disks.
 266  */
 267 u_int
 268 g_mirror_ndisks(struct g_mirror_softc *sc, int state)
 269 {
 270         struct g_mirror_disk *disk;
 271         u_int n = 0;
 272
 273         sx_assert(&sc->sc_lock, SX_LOCKED);
 274
 275         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 276                 if (state == -1 || disk->d_state == state)
 277                         n++;
 278         }
 279         return (n);
 280 }
 281
 282 /*
 283  * Find a disk in mirror by its disk ID.
 284  */
 285 static struct g_mirror_disk *
 286 g_mirror_id2disk(struct g_mirror_softc *sc, uint32_t id)
 287 {
 288         struct g_mirror_disk *disk;
 289
 290         sx_assert(&sc->sc_lock, SX_XLOCKED);
 291
 292         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 293                 if (disk->d_id == id)
 294                         return (disk);
 295         }
 296         return (NULL);
 297 }
 298
 299 static u_int
 300 g_mirror_nrequests(struct g_mirror_softc *sc, struct g_consumer *cp)
 301 {
 302         struct bio *bp;
 303         u_int nreqs = 0;
 304
 305         mtx_lock(&sc->sc_queue_mtx);
 306         TAILQ_FOREACH(bp, &sc->sc_queue.queue, bio_queue) {
 307                 if (bp->bio_from == cp)
 308                         nreqs++;
 309         }
 310         mtx_unlock(&sc->sc_queue_mtx);
 311         return (nreqs);
 312 }
 313
 314 static int
 315 g_mirror_is_busy(struct g_mirror_softc *sc, struct g_consumer *cp)
 316 {
 317
 318         if (cp->index > 0) {
 319                 G_MIRROR_DEBUG(2,
 320                     "I/O requests for %s exist, can't destroy it now.",
 321                     cp->provider->name);
 322                 return (1);
 323         }
 324         if (g_mirror_nrequests(sc, cp) > 0) {
 325                 G_MIRROR_DEBUG(2,
 326                     "I/O requests for %s in queue, can't destroy it now.",
 327                     cp->provider->name);
 328                 return (1);
 329         }
 330         return (0);
 331 }
 332
 333 static void
 334 g_mirror_destroy_consumer(void *arg, int flags __unused)
 335 {
 336         struct g_consumer *cp;
 337
 338         g_topology_assert();
 339
 340         cp = arg;
 341         G_MIRROR_DEBUG(1, "Consumer %s destroyed.", cp->provider->name);
 342         g_detach(cp);
 343         g_destroy_consumer(cp);
 344 }
 345
 346 static void
 347 g_mirror_kill_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
 348 {
 349         struct g_provider *pp;
 350         int retaste_wait;
 351
 352         g_topology_assert();
 353
 354         cp->private = NULL;
 355         if (g_mirror_is_busy(sc, cp))
 356                 return;
 357         pp = cp->provider;
 358         retaste_wait = 0;
 359         if (cp->acw == 1) {
 360                 if ((pp->geom->flags & G_GEOM_WITHER) == 0)
 361                         retaste_wait = 1;
 362         }
 363         G_MIRROR_DEBUG(2, "Access %s r%dw%de%d = %d", pp->name, -cp->acr,
 364             -cp->acw, -cp->ace, 0);
 365         if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
 366                 g_access(cp, -cp->acr, -cp->acw, -cp->ace);
 367         if (retaste_wait) {
 368                 /*
 369                  * After retaste event was send (inside g_access()), we can send
 370                  * event to detach and destroy consumer.
 371                  * A class, which has consumer to the given provider connected
 372                  * will not receive retaste event for the provider.
 373                  * This is the way how I ignore retaste events when I close
 374                  * consumers opened for write: I detach and destroy consumer
 375                  * after retaste event is sent.
 376                  */
 377                 g_post_event(g_mirror_destroy_consumer, cp, M_WAITOK, NULL);
 378                 return;
 379         }
 380         G_MIRROR_DEBUG(1, "Consumer %s destroyed.", pp->name);
 381         g_detach(cp);
 382         g_destroy_consumer(cp);
 383 }
 384
 385 static int
 386 g_mirror_connect_disk(struct g_mirror_disk *disk, struct g_provider *pp)
 387 {
 388         struct g_consumer *cp;
 389         int error;
 390
 391         g_topology_assert_not();
 392         KASSERT(disk->d_consumer == NULL,
 393             ("Disk already connected (device %s).", disk->d_softc->sc_name));
 394
 395         g_topology_lock();
 396         cp = g_new_consumer(disk->d_softc->sc_geom);
 397         error = g_attach(cp, pp);
 398         if (error != 0) {
 399                 g_destroy_consumer(cp);
 400                 g_topology_unlock();
 401                 return (error);
 402         }
 403         error = g_access(cp, 1, 1, 1);
 404         if (error != 0) {
 405                 g_detach(cp);
 406                 g_destroy_consumer(cp);
 407                 g_topology_unlock();
 408                 G_MIRROR_DEBUG(0, "Cannot open consumer %s (error=%d).",
 409                     pp->name, error);
 410                 return (error);
 411         }
 412         g_topology_unlock();
 413         disk->d_consumer = cp;
 414         disk->d_consumer->private = disk;
 415         disk->d_consumer->index = 0;
 416
 417         G_MIRROR_DEBUG(2, "Disk %s connected.", g_mirror_get_diskname(disk));
 418         return (0);
 419 }
 420
 421 static void
 422 g_mirror_disconnect_consumer(struct g_mirror_softc *sc, struct g_consumer *cp)
 423 {
 424
 425         g_topology_assert();
 426
 427         if (cp == NULL)
 428                 return;
 429         if (cp->provider != NULL)
 430                 g_mirror_kill_consumer(sc, cp);
 431         else
 432                 g_destroy_consumer(cp);
 433 }
 434
 435 /*
 436  * Initialize disk. This means allocate memory, create consumer, attach it
 437  * to the provider and open access (r1w1e1) to it.
 438  */
 439 static struct g_mirror_disk *
 440 g_mirror_init_disk(struct g_mirror_softc *sc, struct g_provider *pp,
 441     struct g_mirror_metadata *md, int *errorp)
 442 {
 443         struct g_mirror_disk *disk;
 444         int i, error;
 445
 446         disk = malloc(sizeof(*disk), M_MIRROR, M_NOWAIT | M_ZERO);
 447         if (disk == NULL) {
 448                 error = ENOMEM;
 449                 goto fail;
 450         }
 451         disk->d_softc = sc;
 452         error = g_mirror_connect_disk(disk, pp);
 453         if (error != 0)
 454                 goto fail;
 455         disk->d_id = md->md_did;
 456         disk->d_state = G_MIRROR_DISK_STATE_NONE;
 457         disk->d_priority = md->md_priority;
 458         disk->d_flags = md->md_dflags;
 459         error = g_getattr("GEOM::candelete", disk->d_consumer, &i);
 460         if (error == 0 && i != 0)
 461                 disk->d_flags |= G_MIRROR_DISK_FLAG_CANDELETE;
 462         if (md->md_provider[0] != '\0')
 463                 disk->d_flags |= G_MIRROR_DISK_FLAG_HARDCODED;
 464         disk->d_sync.ds_consumer = NULL;
 465         disk->d_sync.ds_offset = md->md_sync_offset;
 466         disk->d_sync.ds_offset_done = md->md_sync_offset;
 467         disk->d_genid = md->md_genid;
 468         disk->d_sync.ds_syncid = md->md_syncid;
 469         if (errorp != NULL)
 470                 *errorp = 0;
 471         return (disk);
 472 fail:
 473         if (errorp != NULL)
 474                 *errorp = error;
 475         if (disk != NULL)
 476                 free(disk, M_MIRROR);
 477         return (NULL);
 478 }
 479
 480 static void
 481 g_mirror_destroy_disk(struct g_mirror_disk *disk)
 482 {
 483         struct g_mirror_softc *sc;
 484
 485         g_topology_assert_not();
 486         sc = disk->d_softc;
 487         sx_assert(&sc->sc_lock, SX_XLOCKED);
 488
 489         LIST_REMOVE(disk, d_next);
 490         g_mirror_event_cancel(disk);
 491         if (sc->sc_hint == disk)
 492                 sc->sc_hint = NULL;
 493         switch (disk->d_state) {
 494         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
 495                 g_mirror_sync_stop(disk, 1);
 496                 /* FALLTHROUGH */
 497         case G_MIRROR_DISK_STATE_NEW:
 498         case G_MIRROR_DISK_STATE_STALE:
 499         case G_MIRROR_DISK_STATE_ACTIVE:
 500                 g_topology_lock();
 501                 g_mirror_disconnect_consumer(sc, disk->d_consumer);
 502                 g_topology_unlock();
 503                 free(disk, M_MIRROR);
 504                 break;
 505         default:
 506                 KASSERT(0 == 1, ("Wrong disk state (%s, %s).",
 507                     g_mirror_get_diskname(disk),
 508                     g_mirror_disk_state2str(disk->d_state)));
 509         }
 510 }
 511
 512 static void
 513 g_mirror_destroy_device(struct g_mirror_softc *sc)
 514 {
 515         struct g_mirror_disk *disk;
 516         struct g_mirror_event *ep;
 517         struct g_geom *gp;
 518         struct g_consumer *cp, *tmpcp;
 519
 520         g_topology_assert_not();
 521         sx_assert(&sc->sc_lock, SX_XLOCKED);
 522
 523         gp = sc->sc_geom;
 524         if (sc->sc_provider != NULL)
 525                 g_mirror_destroy_provider(sc);
 526         for (disk = LIST_FIRST(&sc->sc_disks); disk != NULL;
 527             disk = LIST_FIRST(&sc->sc_disks)) {
 528                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
 529                 g_mirror_update_metadata(disk);
 530                 g_mirror_destroy_disk(disk);
 531         }
 532         while ((ep = g_mirror_event_get(sc)) != NULL) {
 533                 g_mirror_event_remove(sc, ep);
 534                 if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0)
 535                         g_mirror_event_free(ep);
 536                 else {
 537                         ep->e_error = ECANCELED;
 538                         ep->e_flags |= G_MIRROR_EVENT_DONE;
 539                         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, ep);
 540                         mtx_lock(&sc->sc_events_mtx);
 541                         wakeup(ep);
 542                         mtx_unlock(&sc->sc_events_mtx);
 543                 }
 544         }
 545         callout_drain(&sc->sc_callout);
 546
 547         g_topology_lock();
 548         LIST_FOREACH_SAFE(cp, &sc->sc_sync.ds_geom->consumer, consumer, tmpcp) {
 549                 g_mirror_disconnect_consumer(sc, cp);
 550         }
 551         g_wither_geom(sc->sc_sync.ds_geom, ENXIO);
 552         G_MIRROR_DEBUG(0, "Device %s destroyed.", gp->name);
 553         g_wither_geom(gp, ENXIO);
 554         g_topology_unlock();
 555         mtx_destroy(&sc->sc_queue_mtx);
 556         mtx_destroy(&sc->sc_events_mtx);
 557         sx_xunlock(&sc->sc_lock);
 558         sx_destroy(&sc->sc_lock);
 559 }
 560
 561 static void
 562 g_mirror_orphan(struct g_consumer *cp)
 563 {
 564         struct g_mirror_disk *disk;
 565
 566         g_topology_assert();
 567
 568         disk = cp->private;
 569         if (disk == NULL)
 570                 return;
 571         disk->d_softc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
 572         g_mirror_event_send(disk, G_MIRROR_DISK_STATE_DISCONNECTED,
 573             G_MIRROR_EVENT_DONTWAIT);
 574 }
 575
 576 /*
 577  * Function should return the next active disk on the list.
 578  * It is possible that it will be the same disk as given.
 579  * If there are no active disks on list, NULL is returned.
 580  */
 581 static __inline struct g_mirror_disk *
 582 g_mirror_find_next(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
 583 {
 584         struct g_mirror_disk *dp;
 585
 586         for (dp = LIST_NEXT(disk, d_next); dp != disk;
 587             dp = LIST_NEXT(dp, d_next)) {
 588                 if (dp == NULL)
 589                         dp = LIST_FIRST(&sc->sc_disks);
 590                 if (dp->d_state == G_MIRROR_DISK_STATE_ACTIVE)
 591                         break;
 592         }
 593         if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
 594                 return (NULL);
 595         return (dp);
 596 }
 597
 598 static struct g_mirror_disk *
 599 g_mirror_get_disk(struct g_mirror_softc *sc)
 600 {
 601         struct g_mirror_disk *disk;
 602
 603         if (sc->sc_hint == NULL) {
 604                 sc->sc_hint = LIST_FIRST(&sc->sc_disks);
 605                 if (sc->sc_hint == NULL)
 606                         return (NULL);
 607         }
 608         disk = sc->sc_hint;
 609         if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE) {
 610                 disk = g_mirror_find_next(sc, disk);
 611                 if (disk == NULL)
 612                         return (NULL);
 613         }
 614         sc->sc_hint = g_mirror_find_next(sc, disk);
 615         return (disk);
 616 }
 617
 618 static int
 619 g_mirror_write_metadata(struct g_mirror_disk *disk,
 620     struct g_mirror_metadata *md)
 621 {
 622         struct g_mirror_softc *sc;
 623         struct g_consumer *cp;
 624         off_t offset, length;
 625         u_char *sector;
 626         int error = 0;
 627
 628         g_topology_assert_not();
 629         sc = disk->d_softc;
 630         sx_assert(&sc->sc_lock, SX_LOCKED);
 631
 632         cp = disk->d_consumer;
 633         KASSERT(cp != NULL, ("NULL consumer (%s).", sc->sc_name));
 634         KASSERT(cp->provider != NULL, ("NULL provider (%s).", sc->sc_name));
 635         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
 636             ("Consumer %s closed? (r%dw%de%d).", cp->provider->name, cp->acr,
 637             cp->acw, cp->ace));
 638         length = cp->provider->sectorsize;
 639         offset = cp->provider->mediasize - length;
 640         sector = malloc((size_t)length, M_MIRROR, M_WAITOK | M_ZERO);
 641         if (md != NULL)
 642                 mirror_metadata_encode(md, sector);
 643         error = g_write_data(cp, offset, sector, length);
 644         free(sector, M_MIRROR);
 645         if (error != 0) {
 646                 if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
 647                         disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
 648                         G_MIRROR_DEBUG(0, "Cannot write metadata on %s "
 649                             "(device=%s, error=%d).",
 650                             g_mirror_get_diskname(disk), sc->sc_name, error);
 651                 } else {
 652                         G_MIRROR_DEBUG(1, "Cannot write metadata on %s "
 653                             "(device=%s, error=%d).",
 654                             g_mirror_get_diskname(disk), sc->sc_name, error);
 655                 }
 656                 if (g_mirror_disconnect_on_failure &&
 657                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1) {
 658                         sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
 659                         g_mirror_event_send(disk,
 660                             G_MIRROR_DISK_STATE_DISCONNECTED,
 661                             G_MIRROR_EVENT_DONTWAIT);
 662                 }
 663         }
 664         return (error);
 665 }
 666
 667 static int
 668 g_mirror_clear_metadata(struct g_mirror_disk *disk)
 669 {
 670         int error;
 671
 672         g_topology_assert_not();
 673         sx_assert(&disk->d_softc->sc_lock, SX_LOCKED);
 674
 675         error = g_mirror_write_metadata(disk, NULL);
 676         if (error == 0) {
 677                 G_MIRROR_DEBUG(2, "Metadata on %s cleared.",
 678                     g_mirror_get_diskname(disk));
 679         } else {
 680                 G_MIRROR_DEBUG(0,
 681                     "Cannot clear metadata on disk %s (error=%d).",
 682                     g_mirror_get_diskname(disk), error);
 683         }
 684         return (error);
 685 }
 686
 687 void
 688 g_mirror_fill_metadata(struct g_mirror_softc *sc, struct g_mirror_disk *disk,
 689     struct g_mirror_metadata *md)
 690 {
 691
 692         strlcpy(md->md_magic, G_MIRROR_MAGIC, sizeof(md->md_magic));
 693         md->md_version = G_MIRROR_VERSION;
 694         strlcpy(md->md_name, sc->sc_name, sizeof(md->md_name));
 695         md->md_mid = sc->sc_id;
 696         md->md_all = sc->sc_ndisks;
 697         md->md_slice = sc->sc_slice;
 698         md->md_balance = sc->sc_balance;
 699         md->md_genid = sc->sc_genid;
 700         md->md_mediasize = sc->sc_mediasize;
 701         md->md_sectorsize = sc->sc_sectorsize;
 702         md->md_mflags = (sc->sc_flags & G_MIRROR_DEVICE_FLAG_MASK);
 703         bzero(md->md_provider, sizeof(md->md_provider));
 704         if (disk == NULL) {
 705                 md->md_did = arc4random();
 706                 md->md_priority = 0;
 707                 md->md_syncid = 0;
 708                 md->md_dflags = 0;
 709                 md->md_sync_offset = 0;
 710                 md->md_provsize = 0;
 711         } else {
 712                 md->md_did = disk->d_id;
 713                 md->md_priority = disk->d_priority;
 714                 md->md_syncid = disk->d_sync.ds_syncid;
 715                 md->md_dflags = (disk->d_flags & G_MIRROR_DISK_FLAG_MASK);
 716                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
 717                         md->md_sync_offset = disk->d_sync.ds_offset_done;
 718                 else
 719                         md->md_sync_offset = 0;
 720                 if ((disk->d_flags & G_MIRROR_DISK_FLAG_HARDCODED) != 0) {
 721                         strlcpy(md->md_provider,
 722                             disk->d_consumer->provider->name,
 723                             sizeof(md->md_provider));
 724                 }
 725                 md->md_provsize = disk->d_consumer->provider->mediasize;
 726         }
 727 }
 728
 729 void
 730 g_mirror_update_metadata(struct g_mirror_disk *disk)
 731 {
 732         struct g_mirror_softc *sc;
 733         struct g_mirror_metadata md;
 734         int error;
 735
 736         g_topology_assert_not();
 737         sc = disk->d_softc;
 738         sx_assert(&sc->sc_lock, SX_LOCKED);
 739
 740         g_mirror_fill_metadata(sc, disk, &md);
 741         error = g_mirror_write_metadata(disk, &md);
 742         if (error == 0) {
 743                 G_MIRROR_DEBUG(2, "Metadata on %s updated.",
 744                     g_mirror_get_diskname(disk));
 745         } else {
 746                 G_MIRROR_DEBUG(0,
 747                     "Cannot update metadata on disk %s (error=%d).",
 748                     g_mirror_get_diskname(disk), error);
 749         }
 750 }
 751
 752 static void
 753 g_mirror_bump_syncid(struct g_mirror_softc *sc)
 754 {
 755         struct g_mirror_disk *disk;
 756
 757         g_topology_assert_not();
 758         sx_assert(&sc->sc_lock, SX_XLOCKED);
 759         KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
 760             ("%s called with no active disks (device=%s).", __func__,
 761             sc->sc_name));
 762
 763         sc->sc_syncid++;
 764         G_MIRROR_DEBUG(1, "Device %s: syncid bumped to %u.", sc->sc_name,
 765             sc->sc_syncid);
 766         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 767                 if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
 768                     disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
 769                         disk->d_sync.ds_syncid = sc->sc_syncid;
 770                         g_mirror_update_metadata(disk);
 771                 }
 772         }
 773 }
 774
 775 static void
 776 g_mirror_bump_genid(struct g_mirror_softc *sc)
 777 {
 778         struct g_mirror_disk *disk;
 779
 780         g_topology_assert_not();
 781         sx_assert(&sc->sc_lock, SX_XLOCKED);
 782         KASSERT(g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 0,
 783             ("%s called with no active disks (device=%s).", __func__,
 784             sc->sc_name));
 785
 786         sc->sc_genid++;
 787         G_MIRROR_DEBUG(1, "Device %s: genid bumped to %u.", sc->sc_name,
 788             sc->sc_genid);
 789         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 790                 if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
 791                     disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
 792                         disk->d_genid = sc->sc_genid;
 793                         g_mirror_update_metadata(disk);
 794                 }
 795         }
 796 }
 797
 798 static int
 799 g_mirror_idle(struct g_mirror_softc *sc, int acw)
 800 {
 801         struct g_mirror_disk *disk;
 802         int timeout;
 803
 804         g_topology_assert_not();
 805         sx_assert(&sc->sc_lock, SX_XLOCKED);
 806
 807         if (sc->sc_provider == NULL)
 808                 return (0);
 809         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
 810                 return (0);
 811         if (sc->sc_idle)
 812                 return (0);
 813         if (sc->sc_writes > 0)
 814                 return (0);
 815         if (acw > 0 || (acw == -1 && sc->sc_provider->acw > 0)) {
 816                 timeout = g_mirror_idletime - (time_uptime - sc->sc_last_write);
 817                 if (!g_mirror_shutdown && timeout > 0)
 818                         return (timeout);
 819         }
 820         sc->sc_idle = 1;
 821         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 822                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
 823                         continue;
 824                 G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
 825                     g_mirror_get_diskname(disk), sc->sc_name);
 826                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
 827                 g_mirror_update_metadata(disk);
 828         }
 829         return (0);
 830 }
 831
 832 static void
 833 g_mirror_unidle(struct g_mirror_softc *sc)
 834 {
 835         struct g_mirror_disk *disk;
 836
 837         g_topology_assert_not();
 838         sx_assert(&sc->sc_lock, SX_XLOCKED);
 839
 840         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
 841                 return;
 842         sc->sc_idle = 0;
 843         sc->sc_last_write = time_uptime;
 844         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
 845                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
 846                         continue;
 847                 G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
 848                     g_mirror_get_diskname(disk), sc->sc_name);
 849                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
 850                 g_mirror_update_metadata(disk);
 851         }
 852 }
 853
 854 static void
 855 g_mirror_done(struct bio *bp)
 856 {
 857         struct g_mirror_softc *sc;
 858
 859         sc = bp->bio_from->geom->softc;
 860         bp->bio_cflags = G_MIRROR_BIO_FLAG_REGULAR;
 861         mtx_lock(&sc->sc_queue_mtx);
 862         bioq_disksort(&sc->sc_queue, bp);
 863         mtx_unlock(&sc->sc_queue_mtx);
 864         wakeup(sc);
 865 }
 866
 867 static void
 868 g_mirror_regular_request(struct bio *bp)
 869 {
 870         struct g_mirror_softc *sc;
 871         struct g_mirror_disk *disk;
 872         struct bio *pbp;
 873
 874         g_topology_assert_not();
 875
 876         pbp = bp->bio_parent;
 877         sc = pbp->bio_to->geom->softc;
 878         bp->bio_from->index--;
 879         if (bp->bio_cmd == BIO_WRITE)
 880                 sc->sc_writes--;
 881         disk = bp->bio_from->private;
 882         if (disk == NULL) {
 883                 g_topology_lock();
 884                 g_mirror_kill_consumer(sc, bp->bio_from);
 885                 g_topology_unlock();
 886         }
 887
 888         pbp->bio_inbed++;
 889         KASSERT(pbp->bio_inbed <= pbp->bio_children,
 890             ("bio_inbed (%u) is bigger than bio_children (%u).", pbp->bio_inbed,
 891             pbp->bio_children));
 892         if (bp->bio_error == 0 && pbp->bio_error == 0) {
 893                 G_MIRROR_LOGREQ(3, bp, "Request delivered.");
 894                 g_destroy_bio(bp);
 895                 if (pbp->bio_children == pbp->bio_inbed) {
 896                         G_MIRROR_LOGREQ(3, pbp, "Request delivered.");
 897                         pbp->bio_completed = pbp->bio_length;
 898                         if (pbp->bio_cmd == BIO_WRITE ||
 899                             pbp->bio_cmd == BIO_DELETE) {
 900                                 bioq_remove(&sc->sc_inflight, pbp);
 901                                 /* Release delayed sync requests if possible. */
 902                                 g_mirror_sync_release(sc);
 903                         }
 904                         g_io_deliver(pbp, pbp->bio_error);
 905                 }
 906                 return;
 907         } else if (bp->bio_error != 0) {
 908                 if (pbp->bio_error == 0)
 909                         pbp->bio_error = bp->bio_error;
 910                 if (disk != NULL) {
 911                         if ((disk->d_flags & G_MIRROR_DISK_FLAG_BROKEN) == 0) {
 912                                 disk->d_flags |= G_MIRROR_DISK_FLAG_BROKEN;
 913                                 G_MIRROR_LOGREQ(0, bp,
 914                                     "Request failed (error=%d).",
 915                                     bp->bio_error);
 916                         } else {
 917                                 G_MIRROR_LOGREQ(1, bp,
 918                                     "Request failed (error=%d).",
 919                                     bp->bio_error);
 920                         }
 921                         if (g_mirror_disconnect_on_failure &&
 922                             g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) > 1)
 923                         {
 924                                 sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
 925                                 g_mirror_event_send(disk,
 926                                     G_MIRROR_DISK_STATE_DISCONNECTED,
 927                                     G_MIRROR_EVENT_DONTWAIT);
 928                         }
 929                 }
 930                 switch (pbp->bio_cmd) {
 931                 case BIO_DELETE:
 932                 case BIO_WRITE:
 933                         pbp->bio_inbed--;
 934                         pbp->bio_children--;
 935                         break;
 936                 }
 937         }
 938         g_destroy_bio(bp);
 939
 940         switch (pbp->bio_cmd) {
 941         case BIO_READ:
 942                 if (pbp->bio_inbed < pbp->bio_children)
 943                         break;
 944                 if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 1)
 945                         g_io_deliver(pbp, pbp->bio_error);
 946                 else {
 947                         pbp->bio_error = 0;
 948                         mtx_lock(&sc->sc_queue_mtx);
 949                         bioq_disksort(&sc->sc_queue, pbp);
 950                         mtx_unlock(&sc->sc_queue_mtx);
 951                         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
 952                         wakeup(sc);
 953                 }
 954                 break;
 955         case BIO_DELETE:
 956         case BIO_WRITE:
 957                 if (pbp->bio_children == 0) {
 958                         /*
 959                          * All requests failed.
 960                          */
 961                 } else if (pbp->bio_inbed < pbp->bio_children) {
 962                         /* Do nothing. */
 963                         break;
 964                 } else if (pbp->bio_children == pbp->bio_inbed) {
 965                         /* Some requests succeeded. */
 966                         pbp->bio_error = 0;
 967                         pbp->bio_completed = pbp->bio_length;
 968                 }
 969                 bioq_remove(&sc->sc_inflight, pbp);
 970                 /* Release delayed sync requests if possible. */
 971                 g_mirror_sync_release(sc);
 972                 g_io_deliver(pbp, pbp->bio_error);
 973                 break;
 974         default:
 975                 KASSERT(1 == 0, ("Invalid request: %u.", pbp->bio_cmd));
 976                 break;
 977         }
 978 }
 979
 980 static void
 981 g_mirror_sync_done(struct bio *bp)
 982 {
 983         struct g_mirror_softc *sc;
 984
 985         G_MIRROR_LOGREQ(3, bp, "Synchronization request delivered.");
 986         sc = bp->bio_from->geom->softc;
 987         bp->bio_cflags = G_MIRROR_BIO_FLAG_SYNC;
 988         mtx_lock(&sc->sc_queue_mtx);
 989         bioq_disksort(&sc->sc_queue, bp);
 990         mtx_unlock(&sc->sc_queue_mtx);
 991         wakeup(sc);
 992 }
 993
 994 static void
 995 g_mirror_kernel_dump(struct bio *bp)
 996 {
 997         struct g_mirror_softc *sc;
 998         struct g_mirror_disk *disk;
 999         struct bio *cbp;
1000         struct g_kerneldump *gkd;
1001
1002         /*
1003          * We configure dumping to the first component, because this component
1004          * will be used for reading with 'prefer' balance algorithm.
1005          * If the component with the higest priority is currently disconnected
1006          * we will not be able to read the dump after the reboot if it will be
1007          * connected and synchronized later. Can we do something better?
1008          */
1009         sc = bp->bio_to->geom->softc;
1010         disk = LIST_FIRST(&sc->sc_disks);
1011
1012         gkd = (struct g_kerneldump *)bp->bio_data;
1013         if (gkd->length > bp->bio_to->mediasize)
1014                 gkd->length = bp->bio_to->mediasize;
1015         cbp = g_clone_bio(bp);
1016         if (cbp == NULL) {
1017                 g_io_deliver(bp, ENOMEM);
1018                 return;
1019         }
1020         cbp->bio_done = g_std_done;
1021         g_io_request(cbp, disk->d_consumer);
1022         G_MIRROR_DEBUG(1, "Kernel dump will go to %s.",
1023             g_mirror_get_diskname(disk));
1024 }
1025
1026 static void
1027 g_mirror_flush(struct g_mirror_softc *sc, struct bio *bp)
1028 {
1029         struct bio_queue_head queue;
1030         struct g_mirror_disk *disk;
1031         struct g_consumer *cp;
1032         struct bio *cbp;
1033
1034         bioq_init(&queue);
1035         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1036                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1037                         continue;
1038                 cbp = g_clone_bio(bp);
1039                 if (cbp == NULL) {
1040                         for (cbp = bioq_first(&queue); cbp != NULL;
1041                             cbp = bioq_first(&queue)) {
1042                                 bioq_remove(&queue, cbp);
1043                                 g_destroy_bio(cbp);
1044                         }
1045                         if (bp->bio_error == 0)
1046                                 bp->bio_error = ENOMEM;
1047                         g_io_deliver(bp, bp->bio_error);
1048                         return;
1049                 }
1050                 bioq_insert_tail(&queue, cbp);
1051                 cbp->bio_done = g_std_done;
1052                 cbp->bio_caller1 = disk;
1053                 cbp->bio_to = disk->d_consumer->provider;
1054         }
1055         for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1056                 bioq_remove(&queue, cbp);
1057                 G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1058                 disk = cbp->bio_caller1;
1059                 cbp->bio_caller1 = NULL;
1060                 cp = disk->d_consumer;
1061                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1062                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1063                     cp->acr, cp->acw, cp->ace));
1064                 g_io_request(cbp, disk->d_consumer);
1065         }
1066 }
1067
1068 static void
1069 g_mirror_start(struct bio *bp)
1070 {
1071         struct g_mirror_softc *sc;
1072
1073         sc = bp->bio_to->geom->softc;
1074         /*
1075          * If sc == NULL or there are no valid disks, provider's error
1076          * should be set and g_mirror_start() should not be called at all.
1077          */
1078         KASSERT(sc != NULL && sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1079             ("Provider's error should be set (error=%d)(mirror=%s).",
1080             bp->bio_to->error, bp->bio_to->name));
1081         G_MIRROR_LOGREQ(3, bp, "Request received.");
1082
1083         switch (bp->bio_cmd) {
1084         case BIO_READ:
1085         case BIO_WRITE:
1086         case BIO_DELETE:
1087                 break;
1088         case BIO_FLUSH:
1089                 g_mirror_flush(sc, bp);
1090                 return;
1091         case BIO_GETATTR:
1092                 if (g_handleattr_int(bp, "GEOM::candelete", 1))
1093                         return;
1094                 else if (strcmp("GEOM::kerneldump", bp->bio_attribute) == 0) {
1095                         g_mirror_kernel_dump(bp);
1096                         return;
1097                 }
1098                 /* FALLTHROUGH */
1099         default:
1100                 g_io_deliver(bp, EOPNOTSUPP);
1101                 return;
1102         }
1103         mtx_lock(&sc->sc_queue_mtx);
1104         bioq_disksort(&sc->sc_queue, bp);
1105         mtx_unlock(&sc->sc_queue_mtx);
1106         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
1107         wakeup(sc);
1108 }
1109
1110 /*
1111  * Return TRUE if the given request is colliding with a in-progress
1112  * synchronization request.
1113  */
1114 static int
1115 g_mirror_sync_collision(struct g_mirror_softc *sc, struct bio *bp)
1116 {
1117         struct g_mirror_disk *disk;
1118         struct bio *sbp;
1119         off_t rstart, rend, sstart, send;
1120         int i;
1121
1122         if (sc->sc_sync.ds_ndisks == 0)
1123                 return (0);
1124         rstart = bp->bio_offset;
1125         rend = bp->bio_offset + bp->bio_length;
1126         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1127                 if (disk->d_state != G_MIRROR_DISK_STATE_SYNCHRONIZING)
1128                         continue;
1129                 for (i = 0; i < g_mirror_syncreqs; i++) {
1130                         sbp = disk->d_sync.ds_bios[i];
1131                         if (sbp == NULL)
1132                                 continue;
1133                         sstart = sbp->bio_offset;
1134                         send = sbp->bio_offset + sbp->bio_length;
1135                         if (rend > sstart && rstart < send)
1136                                 return (1);
1137                 }
1138         }
1139         return (0);
1140 }
1141
1142 /*
1143  * Return TRUE if the given sync request is colliding with a in-progress regular
1144  * request.
1145  */
1146 static int
1147 g_mirror_regular_collision(struct g_mirror_softc *sc, struct bio *sbp)
1148 {
1149         off_t rstart, rend, sstart, send;
1150         struct bio *bp;
1151
1152         if (sc->sc_sync.ds_ndisks == 0)
1153                 return (0);
1154         sstart = sbp->bio_offset;
1155         send = sbp->bio_offset + sbp->bio_length;
1156         TAILQ_FOREACH(bp, &sc->sc_inflight.queue, bio_queue) {
1157                 rstart = bp->bio_offset;
1158                 rend = bp->bio_offset + bp->bio_length;
1159                 if (rend > sstart && rstart < send)
1160                         return (1);
1161         }
1162         return (0);
1163 }
1164
1165 /*
1166  * Puts request onto delayed queue.
1167  */
1168 static void
1169 g_mirror_regular_delay(struct g_mirror_softc *sc, struct bio *bp)
1170 {
1171
1172         G_MIRROR_LOGREQ(2, bp, "Delaying request.");
1173         bioq_insert_head(&sc->sc_regular_delayed, bp);
1174 }
1175
1176 /*
1177  * Puts synchronization request onto delayed queue.
1178  */
1179 static void
1180 g_mirror_sync_delay(struct g_mirror_softc *sc, struct bio *bp)
1181 {
1182
1183         G_MIRROR_LOGREQ(2, bp, "Delaying synchronization request.");
1184         bioq_insert_tail(&sc->sc_sync_delayed, bp);
1185 }
1186
1187 /*
1188  * Releases delayed regular requests which don't collide anymore with sync
1189  * requests.
1190  */
1191 static void
1192 g_mirror_regular_release(struct g_mirror_softc *sc)
1193 {
1194         struct bio *bp, *bp2;
1195
1196         TAILQ_FOREACH_SAFE(bp, &sc->sc_regular_delayed.queue, bio_queue, bp2) {
1197                 if (g_mirror_sync_collision(sc, bp))
1198                         continue;
1199                 bioq_remove(&sc->sc_regular_delayed, bp);
1200                 G_MIRROR_LOGREQ(2, bp, "Releasing delayed request (%p).", bp);
1201                 mtx_lock(&sc->sc_queue_mtx);
1202                 bioq_insert_head(&sc->sc_queue, bp);
1203 #if 0
1204                 /*
1205                  * wakeup() is not needed, because this function is called from
1206                  * the worker thread.
1207                  */
1208                 wakeup(&sc->sc_queue);
1209 #endif
1210                 mtx_unlock(&sc->sc_queue_mtx);
1211         }
1212 }
1213
1214 /*
1215  * Releases delayed sync requests which don't collide anymore with regular
1216  * requests.
1217  */
1218 static void
1219 g_mirror_sync_release(struct g_mirror_softc *sc)
1220 {
1221         struct bio *bp, *bp2;
1222
1223         TAILQ_FOREACH_SAFE(bp, &sc->sc_sync_delayed.queue, bio_queue, bp2) {
1224                 if (g_mirror_regular_collision(sc, bp))
1225                         continue;
1226                 bioq_remove(&sc->sc_sync_delayed, bp);
1227                 G_MIRROR_LOGREQ(2, bp,
1228                     "Releasing delayed synchronization request.");
1229                 g_io_request(bp, bp->bio_from);
1230         }
1231 }
1232
1233 /*
1234  * Handle synchronization requests.
1235  * Every synchronization request is two-steps process: first, READ request is
1236  * send to active provider and then WRITE request (with read data) to the provider
1237  * beeing synchronized. When WRITE is finished, new synchronization request is
1238  * send.
1239  */
1240 static void
1241 g_mirror_sync_request(struct bio *bp)
1242 {
1243         struct g_mirror_softc *sc;
1244         struct g_mirror_disk *disk;
1245
1246         bp->bio_from->index--;
1247         sc = bp->bio_from->geom->softc;
1248         disk = bp->bio_from->private;
1249         if (disk == NULL) {
1250                 sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
1251                 g_topology_lock();
1252                 g_mirror_kill_consumer(sc, bp->bio_from);
1253                 g_topology_unlock();
1254                 free(bp->bio_data, M_MIRROR);
1255                 g_destroy_bio(bp);
1256                 sx_xlock(&sc->sc_lock);
1257                 return;
1258         }
1259
1260         /*
1261          * Synchronization request.
1262          */
1263         switch (bp->bio_cmd) {
1264         case BIO_READ:
1265             {
1266                 struct g_consumer *cp;
1267
1268                 if (bp->bio_error != 0) {
1269                         G_MIRROR_LOGREQ(0, bp,
1270                             "Synchronization request failed (error=%d).",
1271                             bp->bio_error);
1272                         g_destroy_bio(bp);
1273                         return;
1274                 }
1275                 G_MIRROR_LOGREQ(3, bp,
1276                     "Synchronization request half-finished.");
1277                 bp->bio_cmd = BIO_WRITE;
1278                 bp->bio_cflags = 0;
1279                 cp = disk->d_consumer;
1280                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1281                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1282                     cp->acr, cp->acw, cp->ace));
1283                 cp->index++;
1284                 g_io_request(bp, cp);
1285                 return;
1286             }
1287         case BIO_WRITE:
1288             {
1289                 struct g_mirror_disk_sync *sync;
1290                 off_t offset;
1291                 void *data;
1292                 int i;
1293
1294                 if (bp->bio_error != 0) {
1295                         G_MIRROR_LOGREQ(0, bp,
1296                             "Synchronization request failed (error=%d).",
1297                             bp->bio_error);
1298                         g_destroy_bio(bp);
1299                         sc->sc_bump_id |= G_MIRROR_BUMP_GENID;
1300                         g_mirror_event_send(disk,
1301                             G_MIRROR_DISK_STATE_DISCONNECTED,
1302                             G_MIRROR_EVENT_DONTWAIT);
1303                         return;
1304                 }
1305                 G_MIRROR_LOGREQ(3, bp, "Synchronization request finished.");
1306                 sync = &disk->d_sync;
1307                 if (sync->ds_offset == sc->sc_mediasize ||
1308                     sync->ds_consumer == NULL ||
1309                     (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1310                         /* Don't send more synchronization requests. */
1311                         sync->ds_inflight--;
1312                         if (sync->ds_bios != NULL) {
1313                                 i = (int)(uintptr_t)bp->bio_caller1;
1314                                 sync->ds_bios[i] = NULL;
1315                         }
1316                         free(bp->bio_data, M_MIRROR);
1317                         g_destroy_bio(bp);
1318                         if (sync->ds_inflight > 0)
1319                                 return;
1320                         if (sync->ds_consumer == NULL ||
1321                             (sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1322                                 return;
1323                         }
1324                         /* Disk up-to-date, activate it. */
1325                         g_mirror_event_send(disk, G_MIRROR_DISK_STATE_ACTIVE,
1326                             G_MIRROR_EVENT_DONTWAIT);
1327                         return;
1328                 }
1329
1330                 /* Send next synchronization request. */
1331                 data = bp->bio_data;
1332                 bzero(bp, sizeof(*bp));
1333                 bp->bio_cmd = BIO_READ;
1334                 bp->bio_offset = sync->ds_offset;
1335                 bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1336                 sync->ds_offset += bp->bio_length;
1337                 bp->bio_done = g_mirror_sync_done;
1338                 bp->bio_data = data;
1339                 bp->bio_from = sync->ds_consumer;
1340                 bp->bio_to = sc->sc_provider;
1341                 G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1342                 sync->ds_consumer->index++;
1343                 /*
1344                  * Delay the request if it is colliding with a regular request.
1345                  */
1346                 if (g_mirror_regular_collision(sc, bp))
1347                         g_mirror_sync_delay(sc, bp);
1348                 else
1349                         g_io_request(bp, sync->ds_consumer);
1350
1351                 /* Release delayed requests if possible. */
1352                 g_mirror_regular_release(sc);
1353
1354                 /* Find the smallest offset */
1355                 offset = sc->sc_mediasize;
1356                 for (i = 0; i < g_mirror_syncreqs; i++) {
1357                         bp = sync->ds_bios[i];
1358                         if (bp->bio_offset < offset)
1359                                 offset = bp->bio_offset;
1360                 }
1361                 if (sync->ds_offset_done + (MAXPHYS * 100) < offset) {
1362                         /* Update offset_done on every 100 blocks. */
1363                         sync->ds_offset_done = offset;
1364                         g_mirror_update_metadata(disk);
1365                 }
1366                 return;
1367             }
1368         default:
1369                 KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1370                     bp->bio_cmd, sc->sc_name));
1371                 break;
1372         }
1373 }
1374
1375 static void
1376 g_mirror_request_prefer(struct g_mirror_softc *sc, struct bio *bp)
1377 {
1378         struct g_mirror_disk *disk;
1379         struct g_consumer *cp;
1380         struct bio *cbp;
1381
1382         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1383                 if (disk->d_state == G_MIRROR_DISK_STATE_ACTIVE)
1384                         break;
1385         }
1386         if (disk == NULL) {
1387                 if (bp->bio_error == 0)
1388                         bp->bio_error = ENXIO;
1389                 g_io_deliver(bp, bp->bio_error);
1390                 return;
1391         }
1392         cbp = g_clone_bio(bp);
1393         if (cbp == NULL) {
1394                 if (bp->bio_error == 0)
1395                         bp->bio_error = ENOMEM;
1396                 g_io_deliver(bp, bp->bio_error);
1397                 return;
1398         }
1399         /*
1400          * Fill in the component buf structure.
1401          */
1402         cp = disk->d_consumer;
1403         cbp->bio_done = g_mirror_done;
1404         cbp->bio_to = cp->provider;
1405         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1406         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1407             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1408             cp->acw, cp->ace));
1409         cp->index++;
1410         g_io_request(cbp, cp);
1411 }
1412
1413 static void
1414 g_mirror_request_round_robin(struct g_mirror_softc *sc, struct bio *bp)
1415 {
1416         struct g_mirror_disk *disk;
1417         struct g_consumer *cp;
1418         struct bio *cbp;
1419
1420         disk = g_mirror_get_disk(sc);
1421         if (disk == NULL) {
1422                 if (bp->bio_error == 0)
1423                         bp->bio_error = ENXIO;
1424                 g_io_deliver(bp, bp->bio_error);
1425                 return;
1426         }
1427         cbp = g_clone_bio(bp);
1428         if (cbp == NULL) {
1429                 if (bp->bio_error == 0)
1430                         bp->bio_error = ENOMEM;
1431                 g_io_deliver(bp, bp->bio_error);
1432                 return;
1433         }
1434         /*
1435          * Fill in the component buf structure.
1436          */
1437         cp = disk->d_consumer;
1438         cbp->bio_done = g_mirror_done;
1439         cbp->bio_to = cp->provider;
1440         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1441         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1442             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1443             cp->acw, cp->ace));
1444         cp->index++;
1445         g_io_request(cbp, cp);
1446 }
1447
1448 #define TRACK_SIZE  (1 * 1024 * 1024)
1449 #define LOAD_SCALE      256
1450 #define ABS(x)          (((x) >= 0) ? (x) : (-(x)))
1451
1452 static void
1453 g_mirror_request_load(struct g_mirror_softc *sc, struct bio *bp)
1454 {
1455         struct g_mirror_disk *disk, *dp;
1456         struct g_consumer *cp;
1457         struct bio *cbp;
1458         int prio, best;
1459
1460         /* Find a disk with the smallest load. */
1461         disk = NULL;
1462         best = INT_MAX;
1463         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1464                 if (dp->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1465                         continue;
1466                 prio = dp->load;
1467                 /* If disk head is precisely in position - highly prefer it. */
1468                 if (dp->d_last_offset == bp->bio_offset)
1469                         prio -= 2 * LOAD_SCALE;
1470                 else
1471                 /* If disk head is close to position - prefer it. */
1472                 if (ABS(dp->d_last_offset - bp->bio_offset) < TRACK_SIZE)
1473                         prio -= 1 * LOAD_SCALE;
1474                 if (prio <= best) {
1475                         disk = dp;
1476                         best = prio;
1477                 }
1478         }
1479         KASSERT(disk != NULL, ("NULL disk for %s.", sc->sc_name));
1480         cbp = g_clone_bio(bp);
1481         if (cbp == NULL) {
1482                 if (bp->bio_error == 0)
1483                         bp->bio_error = ENOMEM;
1484                 g_io_deliver(bp, bp->bio_error);
1485                 return;
1486         }
1487         /*
1488          * Fill in the component buf structure.
1489          */
1490         cp = disk->d_consumer;
1491         cbp->bio_done = g_mirror_done;
1492         cbp->bio_to = cp->provider;
1493         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1494         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1495             ("Consumer %s not opened (r%dw%de%d).", cp->provider->name, cp->acr,
1496             cp->acw, cp->ace));
1497         cp->index++;
1498         /* Remember last head position */
1499         disk->d_last_offset = bp->bio_offset + bp->bio_length;
1500         /* Update loads. */
1501         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
1502                 dp->load = (dp->d_consumer->index * LOAD_SCALE +
1503                     dp->load * 7) / 8;
1504         }
1505         g_io_request(cbp, cp);
1506 }
1507
1508 static void
1509 g_mirror_request_split(struct g_mirror_softc *sc, struct bio *bp)
1510 {
1511         struct bio_queue_head queue;
1512         struct g_mirror_disk *disk;
1513         struct g_consumer *cp;
1514         struct bio *cbp;
1515         off_t left, mod, offset, slice;
1516         u_char *data;
1517         u_int ndisks;
1518
1519         if (bp->bio_length <= sc->sc_slice) {
1520                 g_mirror_request_round_robin(sc, bp);
1521                 return;
1522         }
1523         ndisks = g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE);
1524         slice = bp->bio_length / ndisks;
1525         mod = slice % sc->sc_provider->sectorsize;
1526         if (mod != 0)
1527                 slice += sc->sc_provider->sectorsize - mod;
1528         /*
1529          * Allocate all bios before sending any request, so we can
1530          * return ENOMEM in nice and clean way.
1531          */
1532         left = bp->bio_length;
1533         offset = bp->bio_offset;
1534         data = bp->bio_data;
1535         bioq_init(&queue);
1536         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1537                 if (disk->d_state != G_MIRROR_DISK_STATE_ACTIVE)
1538                         continue;
1539                 cbp = g_clone_bio(bp);
1540                 if (cbp == NULL) {
1541                         for (cbp = bioq_first(&queue); cbp != NULL;
1542                             cbp = bioq_first(&queue)) {
1543                                 bioq_remove(&queue, cbp);
1544                                 g_destroy_bio(cbp);
1545                         }
1546                         if (bp->bio_error == 0)
1547                                 bp->bio_error = ENOMEM;
1548                         g_io_deliver(bp, bp->bio_error);
1549                         return;
1550                 }
1551                 bioq_insert_tail(&queue, cbp);
1552                 cbp->bio_done = g_mirror_done;
1553                 cbp->bio_caller1 = disk;
1554                 cbp->bio_to = disk->d_consumer->provider;
1555                 cbp->bio_offset = offset;
1556                 cbp->bio_data = data;
1557                 cbp->bio_length = MIN(left, slice);
1558                 left -= cbp->bio_length;
1559                 if (left == 0)
1560                         break;
1561                 offset += cbp->bio_length;
1562                 data += cbp->bio_length;
1563         }
1564         for (cbp = bioq_first(&queue); cbp != NULL; cbp = bioq_first(&queue)) {
1565                 bioq_remove(&queue, cbp);
1566                 G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1567                 disk = cbp->bio_caller1;
1568                 cbp->bio_caller1 = NULL;
1569                 cp = disk->d_consumer;
1570                 KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1571                     ("Consumer %s not opened (r%dw%de%d).", cp->provider->name,
1572                     cp->acr, cp->acw, cp->ace));
1573                 disk->d_consumer->index++;
1574                 g_io_request(cbp, disk->d_consumer);
1575         }
1576 }
1577
1578 static void
1579 g_mirror_register_request(struct bio *bp)
1580 {
1581         struct g_mirror_softc *sc;
1582
1583         sc = bp->bio_to->geom->softc;
1584         switch (bp->bio_cmd) {
1585         case BIO_READ:
1586                 switch (sc->sc_balance) {
1587                 case G_MIRROR_BALANCE_LOAD:
1588                         g_mirror_request_load(sc, bp);
1589                         break;
1590                 case G_MIRROR_BALANCE_PREFER:
1591                         g_mirror_request_prefer(sc, bp);
1592                         break;
1593                 case G_MIRROR_BALANCE_ROUND_ROBIN:
1594                         g_mirror_request_round_robin(sc, bp);
1595                         break;
1596                 case G_MIRROR_BALANCE_SPLIT:
1597                         g_mirror_request_split(sc, bp);
1598                         break;
1599                 }
1600                 return;
1601         case BIO_WRITE:
1602         case BIO_DELETE:
1603             {
1604                 struct g_mirror_disk *disk;
1605                 struct g_mirror_disk_sync *sync;
1606                 struct bio_queue_head queue;
1607                 struct g_consumer *cp;
1608                 struct bio *cbp;
1609
1610                 /*
1611                  * Delay the request if it is colliding with a synchronization
1612                  * request.
1613                  */
1614                 if (g_mirror_sync_collision(sc, bp)) {
1615                         g_mirror_regular_delay(sc, bp);
1616                         return;
1617                 }
1618
1619                 if (sc->sc_idle)
1620                         g_mirror_unidle(sc);
1621                 else
1622                         sc->sc_last_write = time_uptime;
1623
1624                 /*
1625                  * Allocate all bios before sending any request, so we can
1626                  * return ENOMEM in nice and clean way.
1627                  */
1628                 bioq_init(&queue);
1629                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
1630                         sync = &disk->d_sync;
1631                         switch (disk->d_state) {
1632                         case G_MIRROR_DISK_STATE_ACTIVE:
1633                                 break;
1634                         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
1635                                 if (bp->bio_offset >= sync->ds_offset)
1636                                         continue;
1637                                 break;
1638                         default:
1639                                 continue;
1640                         }
1641                         if (bp->bio_cmd == BIO_DELETE &&
1642                             (disk->d_flags & G_MIRROR_DISK_FLAG_CANDELETE) == 0)
1643                                 continue;
1644                         cbp = g_clone_bio(bp);
1645                         if (cbp == NULL) {
1646                                 for (cbp = bioq_first(&queue); cbp != NULL;
1647                                     cbp = bioq_first(&queue)) {
1648                                         bioq_remove(&queue, cbp);
1649                                         g_destroy_bio(cbp);
1650                                 }
1651                                 if (bp->bio_error == 0)
1652                                         bp->bio_error = ENOMEM;
1653                                 g_io_deliver(bp, bp->bio_error);
1654                                 return;
1655                         }
1656                         bioq_insert_tail(&queue, cbp);
1657                         cbp->bio_done = g_mirror_done;
1658                         cp = disk->d_consumer;
1659                         cbp->bio_caller1 = cp;
1660                         cbp->bio_to = cp->provider;
1661                         KASSERT(cp->acr >= 1 && cp->acw >= 1 && cp->ace >= 1,
1662                             ("Consumer %s not opened (r%dw%de%d).",
1663                             cp->provider->name, cp->acr, cp->acw, cp->ace));
1664                 }
1665                 for (cbp = bioq_first(&queue); cbp != NULL;
1666                     cbp = bioq_first(&queue)) {
1667                         bioq_remove(&queue, cbp);
1668                         G_MIRROR_LOGREQ(3, cbp, "Sending request.");
1669                         cp = cbp->bio_caller1;
1670                         cbp->bio_caller1 = NULL;
1671                         cp->index++;
1672                         sc->sc_writes++;
1673                         g_io_request(cbp, cp);
1674                 }
1675                 /*
1676                  * Put request onto inflight queue, so we can check if new
1677                  * synchronization requests don't collide with it.
1678                  */
1679                 bioq_insert_tail(&sc->sc_inflight, bp);
1680                 /*
1681                  * Bump syncid on first write.
1682                  */
1683                 if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0) {
1684                         sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
1685                         g_mirror_bump_syncid(sc);
1686                 }
1687                 return;
1688             }
1689         default:
1690                 KASSERT(1 == 0, ("Invalid command here: %u (device=%s)",
1691                     bp->bio_cmd, sc->sc_name));
1692                 break;
1693         }
1694 }
1695
1696 static int
1697 g_mirror_can_destroy(struct g_mirror_softc *sc)
1698 {
1699         struct g_geom *gp;
1700         struct g_consumer *cp;
1701
1702         g_topology_assert();
1703         gp = sc->sc_geom;
1704         if (gp->softc == NULL)
1705                 return (1);
1706         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_TASTING) != 0)
1707                 return (0);
1708         LIST_FOREACH(cp, &gp->consumer, consumer) {
1709                 if (g_mirror_is_busy(sc, cp))
1710                         return (0);
1711         }
1712         gp = sc->sc_sync.ds_geom;
1713         LIST_FOREACH(cp, &gp->consumer, consumer) {
1714                 if (g_mirror_is_busy(sc, cp))
1715                         return (0);
1716         }
1717         G_MIRROR_DEBUG(2, "No I/O requests for %s, it can be destroyed.",
1718             sc->sc_name);
1719         return (1);
1720 }
1721
1722 static int
1723 g_mirror_try_destroy(struct g_mirror_softc *sc)
1724 {
1725
1726         if (sc->sc_rootmount != NULL) {
1727                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
1728                     sc->sc_rootmount);
1729                 root_mount_rel(sc->sc_rootmount);
1730                 sc->sc_rootmount = NULL;
1731         }
1732         g_topology_lock();
1733         if (!g_mirror_can_destroy(sc)) {
1734                 g_topology_unlock();
1735                 return (0);
1736         }
1737         sc->sc_geom->softc = NULL;
1738         sc->sc_sync.ds_geom->softc = NULL;
1739         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_WAIT) != 0) {
1740                 g_topology_unlock();
1741                 G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1742                     &sc->sc_worker);
1743                 /* Unlock sc_lock here, as it can be destroyed after wakeup. */
1744                 sx_xunlock(&sc->sc_lock);
1745                 wakeup(&sc->sc_worker);
1746                 sc->sc_worker = NULL;
1747         } else {
1748                 g_topology_unlock();
1749                 g_mirror_destroy_device(sc);
1750                 free(sc, M_MIRROR);
1751         }
1752         return (1);
1753 }
1754
1755 /*
1756  * Worker thread.
1757  */
1758 static void
1759 g_mirror_worker(void *arg)
1760 {
1761         struct g_mirror_softc *sc;
1762         struct g_mirror_event *ep;
1763         struct bio *bp;
1764         int timeout;
1765
1766         sc = arg;
1767         thread_lock(curthread);
1768         sched_prio(curthread, PRIBIO);
1769         thread_unlock(curthread);
1770
1771         sx_xlock(&sc->sc_lock);
1772         for (;;) {
1773                 G_MIRROR_DEBUG(5, "%s: Let's see...", __func__);
1774                 /*
1775                  * First take a look at events.
1776                  * This is important to handle events before any I/O requests.
1777                  */
1778                 ep = g_mirror_event_get(sc);
1779                 if (ep != NULL) {
1780                         g_mirror_event_remove(sc, ep);
1781                         if ((ep->e_flags & G_MIRROR_EVENT_DEVICE) != 0) {
1782                                 /* Update only device status. */
1783                                 G_MIRROR_DEBUG(3,
1784                                     "Running event for device %s.",
1785                                     sc->sc_name);
1786                                 ep->e_error = 0;
1787                                 g_mirror_update_device(sc, 1);
1788                         } else {
1789                                 /* Update disk status. */
1790                                 G_MIRROR_DEBUG(3, "Running event for disk %s.",
1791                                      g_mirror_get_diskname(ep->e_disk));
1792                                 ep->e_error = g_mirror_update_disk(ep->e_disk,
1793                                     ep->e_state);
1794                                 if (ep->e_error == 0)
1795                                         g_mirror_update_device(sc, 0);
1796                         }
1797                         if ((ep->e_flags & G_MIRROR_EVENT_DONTWAIT) != 0) {
1798                                 KASSERT(ep->e_error == 0,
1799                                     ("Error cannot be handled."));
1800                                 g_mirror_event_free(ep);
1801                         } else {
1802                                 ep->e_flags |= G_MIRROR_EVENT_DONE;
1803                                 G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__,
1804                                     ep);
1805                                 mtx_lock(&sc->sc_events_mtx);
1806                                 wakeup(ep);
1807                                 mtx_unlock(&sc->sc_events_mtx);
1808                         }
1809                         if ((sc->sc_flags &
1810                             G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1811                                 if (g_mirror_try_destroy(sc)) {
1812                                         curthread->td_pflags &= ~TDP_GEOM;
1813                                         G_MIRROR_DEBUG(1, "Thread exiting.");
1814                                         kproc_exit(0);
1815                                 }
1816                         }
1817                         G_MIRROR_DEBUG(5, "%s: I'm here 1.", __func__);
1818                         continue;
1819                 }
1820                 /*
1821                  * Check if we can mark array as CLEAN and if we can't take
1822                  * how much seconds should we wait.
1823                  */
1824                 timeout = g_mirror_idle(sc, -1);
1825                 /*
1826                  * Now I/O requests.
1827                  */
1828                 /* Get first request from the queue. */
1829                 mtx_lock(&sc->sc_queue_mtx);
1830                 bp = bioq_first(&sc->sc_queue);
1831                 if (bp == NULL) {
1832                         if ((sc->sc_flags &
1833                             G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
1834                                 mtx_unlock(&sc->sc_queue_mtx);
1835                                 if (g_mirror_try_destroy(sc)) {
1836                                         curthread->td_pflags &= ~TDP_GEOM;
1837                                         G_MIRROR_DEBUG(1, "Thread exiting.");
1838                                         kproc_exit(0);
1839                                 }
1840                                 mtx_lock(&sc->sc_queue_mtx);
1841                         }
1842                         sx_xunlock(&sc->sc_lock);
1843                         /*
1844                          * XXX: We can miss an event here, because an event
1845                          *      can be added without sx-device-lock and without
1846                          *      mtx-queue-lock. Maybe I should just stop using
1847                          *      dedicated mutex for events synchronization and
1848                          *      stick with the queue lock?
1849                          *      The event will hang here until next I/O request
1850                          *      or next event is received.
1851                          */
1852                         MSLEEP(sc, &sc->sc_queue_mtx, PRIBIO | PDROP, "m:w1",
1853                             timeout * hz);
1854                         sx_xlock(&sc->sc_lock);
1855                         G_MIRROR_DEBUG(5, "%s: I'm here 4.", __func__);
1856                         continue;
1857                 }
1858                 bioq_remove(&sc->sc_queue, bp);
1859                 mtx_unlock(&sc->sc_queue_mtx);
1860
1861                 if (bp->bio_from->geom == sc->sc_sync.ds_geom &&
1862                     (bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0) {
1863                         g_mirror_sync_request(bp);      /* READ */
1864                 } else if (bp->bio_to != sc->sc_provider) {
1865                         if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_REGULAR) != 0)
1866                                 g_mirror_regular_request(bp);
1867                         else if ((bp->bio_cflags & G_MIRROR_BIO_FLAG_SYNC) != 0)
1868                                 g_mirror_sync_request(bp);      /* WRITE */
1869                         else {
1870                                 KASSERT(0,
1871                                     ("Invalid request cflags=0x%hhx to=%s.",
1872                                     bp->bio_cflags, bp->bio_to->name));
1873                         }
1874                 } else {
1875                         g_mirror_register_request(bp);
1876                 }
1877                 G_MIRROR_DEBUG(5, "%s: I'm here 9.", __func__);
1878         }
1879 }
1880
1881 static void
1882 g_mirror_update_idle(struct g_mirror_softc *sc, struct g_mirror_disk *disk)
1883 {
1884
1885         sx_assert(&sc->sc_lock, SX_LOCKED);
1886
1887         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) != 0)
1888                 return;
1889         if (!sc->sc_idle && (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) == 0) {
1890                 G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as dirty.",
1891                     g_mirror_get_diskname(disk), sc->sc_name);
1892                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1893         } else if (sc->sc_idle &&
1894             (disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
1895                 G_MIRROR_DEBUG(1, "Disk %s (device %s) marked as clean.",
1896                     g_mirror_get_diskname(disk), sc->sc_name);
1897                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
1898         }
1899 }
1900
1901 static void
1902 g_mirror_sync_start(struct g_mirror_disk *disk)
1903 {
1904         struct g_mirror_softc *sc;
1905         struct g_consumer *cp;
1906         struct bio *bp;
1907         int error, i;
1908
1909         g_topology_assert_not();
1910         sc = disk->d_softc;
1911         sx_assert(&sc->sc_lock, SX_LOCKED);
1912
1913         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
1914             ("Disk %s is not marked for synchronization.",
1915             g_mirror_get_diskname(disk)));
1916         KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
1917             ("Device not in RUNNING state (%s, %u).", sc->sc_name,
1918             sc->sc_state));
1919
1920         sx_xunlock(&sc->sc_lock);
1921         g_topology_lock();
1922         cp = g_new_consumer(sc->sc_sync.ds_geom);
1923         error = g_attach(cp, sc->sc_provider);
1924         KASSERT(error == 0,
1925             ("Cannot attach to %s (error=%d).", sc->sc_name, error));
1926         error = g_access(cp, 1, 0, 0);
1927         KASSERT(error == 0, ("Cannot open %s (error=%d).", sc->sc_name, error));
1928         g_topology_unlock();
1929         sx_xlock(&sc->sc_lock);
1930
1931         G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s.", sc->sc_name,
1932             g_mirror_get_diskname(disk));
1933         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOFAILSYNC) == 0)
1934                 disk->d_flags |= G_MIRROR_DISK_FLAG_DIRTY;
1935         KASSERT(disk->d_sync.ds_consumer == NULL,
1936             ("Sync consumer already exists (device=%s, disk=%s).",
1937             sc->sc_name, g_mirror_get_diskname(disk)));
1938
1939         disk->d_sync.ds_consumer = cp;
1940         disk->d_sync.ds_consumer->private = disk;
1941         disk->d_sync.ds_consumer->index = 0;
1942
1943         /*
1944          * Allocate memory for synchronization bios and initialize them.
1945          */
1946         disk->d_sync.ds_bios = malloc(sizeof(struct bio *) * g_mirror_syncreqs,
1947             M_MIRROR, M_WAITOK);
1948         for (i = 0; i < g_mirror_syncreqs; i++) {
1949                 bp = g_alloc_bio();
1950                 disk->d_sync.ds_bios[i] = bp;
1951                 bp->bio_parent = NULL;
1952                 bp->bio_cmd = BIO_READ;
1953                 bp->bio_data = malloc(MAXPHYS, M_MIRROR, M_WAITOK);
1954                 bp->bio_cflags = 0;
1955                 bp->bio_offset = disk->d_sync.ds_offset;
1956                 bp->bio_length = MIN(MAXPHYS, sc->sc_mediasize - bp->bio_offset);
1957                 disk->d_sync.ds_offset += bp->bio_length;
1958                 bp->bio_done = g_mirror_sync_done;
1959                 bp->bio_from = disk->d_sync.ds_consumer;
1960                 bp->bio_to = sc->sc_provider;
1961                 bp->bio_caller1 = (void *)(uintptr_t)i;
1962         }
1963
1964         /* Increase the number of disks in SYNCHRONIZING state. */
1965         sc->sc_sync.ds_ndisks++;
1966         /* Set the number of in-flight synchronization requests. */
1967         disk->d_sync.ds_inflight = g_mirror_syncreqs;
1968
1969         /*
1970          * Fire off first synchronization requests.
1971          */
1972         for (i = 0; i < g_mirror_syncreqs; i++) {
1973                 bp = disk->d_sync.ds_bios[i];
1974                 G_MIRROR_LOGREQ(3, bp, "Sending synchronization request.");
1975                 disk->d_sync.ds_consumer->index++;
1976                 /*
1977                  * Delay the request if it is colliding with a regular request.
1978                  */
1979                 if (g_mirror_regular_collision(sc, bp))
1980                         g_mirror_sync_delay(sc, bp);
1981                 else
1982                         g_io_request(bp, disk->d_sync.ds_consumer);
1983         }
1984 }
1985
1986 /*
1987  * Stop synchronization process.
1988  * type: 0 - synchronization finished
1989  *       1 - synchronization stopped
1990  */
1991 static void
1992 g_mirror_sync_stop(struct g_mirror_disk *disk, int type)
1993 {
1994         struct g_mirror_softc *sc;
1995         struct g_consumer *cp;
1996
1997         g_topology_assert_not();
1998         sc = disk->d_softc;
1999         sx_assert(&sc->sc_lock, SX_LOCKED);
2000
2001         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2002             ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2003             g_mirror_disk_state2str(disk->d_state)));
2004         if (disk->d_sync.ds_consumer == NULL)
2005                 return;
2006
2007         if (type == 0) {
2008                 G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s finished.",
2009                     sc->sc_name, g_mirror_get_diskname(disk));
2010         } else /* if (type == 1) */ {
2011                 G_MIRROR_DEBUG(0, "Device %s: rebuilding provider %s stopped.",
2012                     sc->sc_name, g_mirror_get_diskname(disk));
2013         }
2014         free(disk->d_sync.ds_bios, M_MIRROR);
2015         disk->d_sync.ds_bios = NULL;
2016         cp = disk->d_sync.ds_consumer;
2017         disk->d_sync.ds_consumer = NULL;
2018         disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2019         sc->sc_sync.ds_ndisks--;
2020         sx_xunlock(&sc->sc_lock); /* Avoid recursion on sc_lock. */
2021         g_topology_lock();
2022         g_mirror_kill_consumer(sc, cp);
2023         g_topology_unlock();
2024         sx_xlock(&sc->sc_lock);
2025 }
2026
2027 static void
2028 g_mirror_launch_provider(struct g_mirror_softc *sc)
2029 {
2030         struct g_mirror_disk *disk;
2031         struct g_provider *pp, *dp;
2032
2033         sx_assert(&sc->sc_lock, SX_LOCKED);
2034
2035         g_topology_lock();
2036         pp = g_new_providerf(sc->sc_geom, "mirror/%s", sc->sc_name);
2037         pp->mediasize = sc->sc_mediasize;
2038         pp->sectorsize = sc->sc_sectorsize;
2039         pp->stripesize = 0;
2040         pp->stripeoffset = 0;
2041
2042         /* Splitting of unmapped BIO's could work but isn't implemented now */
2043         if (sc->sc_balance != G_MIRROR_BALANCE_SPLIT)
2044                 pp->flags |= G_PF_ACCEPT_UNMAPPED;
2045
2046         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2047                 if (disk->d_consumer && disk->d_consumer->provider) {
2048                         dp = disk->d_consumer->provider;
2049                         if (dp->stripesize > pp->stripesize) {
2050                                 pp->stripesize = dp->stripesize;
2051                                 pp->stripeoffset = dp->stripeoffset;
2052                         }
2053                         /* A provider underneath us doesn't support unmapped */
2054                         if ((dp->flags & G_PF_ACCEPT_UNMAPPED) == 0) {
2055                                 G_MIRROR_DEBUG(0, "Cancelling unmapped "
2056                                     "because of %s.", dp->name);
2057                                 pp->flags &= ~G_PF_ACCEPT_UNMAPPED;
2058                         }
2059                 }
2060         }
2061         sc->sc_provider = pp;
2062         g_error_provider(pp, 0);
2063         g_topology_unlock();
2064         G_MIRROR_DEBUG(0, "Device %s launched (%u/%u).", pp->name,
2065             g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE), sc->sc_ndisks);
2066         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2067                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2068                         g_mirror_sync_start(disk);
2069         }
2070 }
2071
2072 static void
2073 g_mirror_destroy_provider(struct g_mirror_softc *sc)
2074 {
2075         struct g_mirror_disk *disk;
2076         struct bio *bp;
2077
2078         g_topology_assert_not();
2079         KASSERT(sc->sc_provider != NULL, ("NULL provider (device=%s).",
2080             sc->sc_name));
2081
2082         g_topology_lock();
2083         g_error_provider(sc->sc_provider, ENXIO);
2084         mtx_lock(&sc->sc_queue_mtx);
2085         while ((bp = bioq_first(&sc->sc_queue)) != NULL) {
2086                 bioq_remove(&sc->sc_queue, bp);
2087                 g_io_deliver(bp, ENXIO);
2088         }
2089         mtx_unlock(&sc->sc_queue_mtx);
2090         G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.", sc->sc_name,
2091             sc->sc_provider->name);
2092         sc->sc_provider->flags |= G_PF_WITHER;
2093         g_orphan_provider(sc->sc_provider, ENXIO);
2094         g_topology_unlock();
2095         sc->sc_provider = NULL;
2096         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2097                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING)
2098                         g_mirror_sync_stop(disk, 1);
2099         }
2100 }
2101
2102 static void
2103 g_mirror_go(void *arg)
2104 {
2105         struct g_mirror_softc *sc;
2106
2107         sc = arg;
2108         G_MIRROR_DEBUG(0, "Force device %s start due to timeout.", sc->sc_name);
2109         g_mirror_event_send(sc, 0,
2110             G_MIRROR_EVENT_DONTWAIT | G_MIRROR_EVENT_DEVICE);
2111 }
2112
2113 static u_int
2114 g_mirror_determine_state(struct g_mirror_disk *disk)
2115 {
2116         struct g_mirror_softc *sc;
2117         u_int state;
2118
2119         sc = disk->d_softc;
2120         if (sc->sc_syncid == disk->d_sync.ds_syncid) {
2121                 if ((disk->d_flags &
2122                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2123                         /* Disk does not need synchronization. */
2124                         state = G_MIRROR_DISK_STATE_ACTIVE;
2125                 } else {
2126                         if ((sc->sc_flags &
2127                              G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2128                             (disk->d_flags &
2129                              G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2130                                 /*
2131                                  * We can start synchronization from
2132                                  * the stored offset.
2133                                  */
2134                                 state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2135                         } else {
2136                                 state = G_MIRROR_DISK_STATE_STALE;
2137                         }
2138                 }
2139         } else if (disk->d_sync.ds_syncid < sc->sc_syncid) {
2140                 /*
2141                  * Reset all synchronization data for this disk,
2142                  * because if it even was synchronized, it was
2143                  * synchronized to disks with different syncid.
2144                  */
2145                 disk->d_flags |= G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2146                 disk->d_sync.ds_offset = 0;
2147                 disk->d_sync.ds_offset_done = 0;
2148                 disk->d_sync.ds_syncid = sc->sc_syncid;
2149                 if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) == 0 ||
2150                     (disk->d_flags & G_MIRROR_DISK_FLAG_FORCE_SYNC) != 0) {
2151                         state = G_MIRROR_DISK_STATE_SYNCHRONIZING;
2152                 } else {
2153                         state = G_MIRROR_DISK_STATE_STALE;
2154                 }
2155         } else /* if (sc->sc_syncid < disk->d_sync.ds_syncid) */ {
2156                 /*
2157                  * Not good, NOT GOOD!
2158                  * It means that mirror was started on stale disks
2159                  * and more fresh disk just arrive.
2160                  * If there were writes, mirror is broken, sorry.
2161                  * I think the best choice here is don't touch
2162                  * this disk and inform the user loudly.
2163                  */
2164                 G_MIRROR_DEBUG(0, "Device %s was started before the freshest "
2165                     "disk (%s) arrives!! It will not be connected to the "
2166                     "running device.", sc->sc_name,
2167                     g_mirror_get_diskname(disk));
2168                 g_mirror_destroy_disk(disk);
2169                 state = G_MIRROR_DISK_STATE_NONE;
2170                 /* Return immediately, because disk was destroyed. */
2171                 return (state);
2172         }
2173         G_MIRROR_DEBUG(3, "State for %s disk: %s.",
2174             g_mirror_get_diskname(disk), g_mirror_disk_state2str(state));
2175         return (state);
2176 }
2177
2178 /*
2179  * Update device state.
2180  */
2181 static void
2182 g_mirror_update_device(struct g_mirror_softc *sc, boolean_t force)
2183 {
2184         struct g_mirror_disk *disk;
2185         u_int state;
2186
2187         sx_assert(&sc->sc_lock, SX_XLOCKED);
2188
2189         switch (sc->sc_state) {
2190         case G_MIRROR_DEVICE_STATE_STARTING:
2191             {
2192                 struct g_mirror_disk *pdisk, *tdisk;
2193                 u_int dirty, ndisks, genid, syncid;
2194
2195                 KASSERT(sc->sc_provider == NULL,
2196                     ("Non-NULL provider in STARTING state (%s).", sc->sc_name));
2197                 /*
2198                  * Are we ready? We are, if all disks are connected or
2199                  * if we have any disks and 'force' is true.
2200                  */
2201                 ndisks = g_mirror_ndisks(sc, -1);
2202                 if (sc->sc_ndisks == ndisks || (force && ndisks > 0)) {
2203                         ;
2204                 } else if (ndisks == 0) {
2205                         /*
2206                          * Disks went down in starting phase, so destroy
2207                          * device.
2208                          */
2209                         callout_drain(&sc->sc_callout);
2210                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2211                         G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p", __LINE__,
2212                             sc->sc_rootmount);
2213                         root_mount_rel(sc->sc_rootmount);
2214                         sc->sc_rootmount = NULL;
2215                         return;
2216                 } else {
2217                         return;
2218                 }
2219
2220                 /*
2221                  * Activate all disks with the biggest syncid.
2222                  */
2223                 if (force) {
2224                         /*
2225                          * If 'force' is true, we have been called due to
2226                          * timeout, so don't bother canceling timeout.
2227                          */
2228                         ndisks = 0;
2229                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2230                                 if ((disk->d_flags &
2231                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) == 0) {
2232                                         ndisks++;
2233                                 }
2234                         }
2235                         if (ndisks == 0) {
2236                                 /* No valid disks found, destroy device. */
2237                                 sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2238                                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2239                                     __LINE__, sc->sc_rootmount);
2240                                 root_mount_rel(sc->sc_rootmount);
2241                                 sc->sc_rootmount = NULL;
2242                                 return;
2243                         }
2244                 } else {
2245                         /* Cancel timeout. */
2246                         callout_drain(&sc->sc_callout);
2247                 }
2248
2249                 /*
2250                  * Find the biggest genid.
2251                  */
2252                 genid = 0;
2253                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2254                         if (disk->d_genid > genid)
2255                                 genid = disk->d_genid;
2256                 }
2257                 sc->sc_genid = genid;
2258                 /*
2259                  * Remove all disks without the biggest genid.
2260                  */
2261                 LIST_FOREACH_SAFE(disk, &sc->sc_disks, d_next, tdisk) {
2262                         if (disk->d_genid < genid) {
2263                                 G_MIRROR_DEBUG(0,
2264                                     "Component %s (device %s) broken, skipping.",
2265                                     g_mirror_get_diskname(disk), sc->sc_name);
2266                                 g_mirror_destroy_disk(disk);
2267                         }
2268                 }
2269
2270                 /*
2271                  * Find the biggest syncid.
2272                  */
2273                 syncid = 0;
2274                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2275                         if (disk->d_sync.ds_syncid > syncid)
2276                                 syncid = disk->d_sync.ds_syncid;
2277                 }
2278
2279                 /*
2280                  * Here we need to look for dirty disks and if all disks
2281                  * with the biggest syncid are dirty, we have to choose
2282                  * one with the biggest priority and rebuild the rest.
2283                  */
2284                 /*
2285                  * Find the number of dirty disks with the biggest syncid.
2286                  * Find the number of disks with the biggest syncid.
2287                  * While here, find a disk with the biggest priority.
2288                  */
2289                 dirty = ndisks = 0;
2290                 pdisk = NULL;
2291                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2292                         if (disk->d_sync.ds_syncid != syncid)
2293                                 continue;
2294                         if ((disk->d_flags &
2295                             G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2296                                 continue;
2297                         }
2298                         ndisks++;
2299                         if ((disk->d_flags & G_MIRROR_DISK_FLAG_DIRTY) != 0) {
2300                                 dirty++;
2301                                 if (pdisk == NULL ||
2302                                     pdisk->d_priority < disk->d_priority) {
2303                                         pdisk = disk;
2304                                 }
2305                         }
2306                 }
2307                 if (dirty == 0) {
2308                         /* No dirty disks at all, great. */
2309                 } else if (dirty == ndisks) {
2310                         /*
2311                          * Force synchronization for all dirty disks except one
2312                          * with the biggest priority.
2313                          */
2314                         KASSERT(pdisk != NULL, ("pdisk == NULL"));
2315                         G_MIRROR_DEBUG(1, "Using disk %s (device %s) as a "
2316                             "master disk for synchronization.",
2317                             g_mirror_get_diskname(pdisk), sc->sc_name);
2318                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2319                                 if (disk->d_sync.ds_syncid != syncid)
2320                                         continue;
2321                                 if ((disk->d_flags &
2322                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2323                                         continue;
2324                                 }
2325                                 KASSERT((disk->d_flags &
2326                                     G_MIRROR_DISK_FLAG_DIRTY) != 0,
2327                                     ("Disk %s isn't marked as dirty.",
2328                                     g_mirror_get_diskname(disk)));
2329                                 /* Skip the disk with the biggest priority. */
2330                                 if (disk == pdisk)
2331                                         continue;
2332                                 disk->d_sync.ds_syncid = 0;
2333                         }
2334                 } else if (dirty < ndisks) {
2335                         /*
2336                          * Force synchronization for all dirty disks.
2337                          * We have some non-dirty disks.
2338                          */
2339                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2340                                 if (disk->d_sync.ds_syncid != syncid)
2341                                         continue;
2342                                 if ((disk->d_flags &
2343                                     G_MIRROR_DISK_FLAG_SYNCHRONIZING) != 0) {
2344                                         continue;
2345                                 }
2346                                 if ((disk->d_flags &
2347                                     G_MIRROR_DISK_FLAG_DIRTY) == 0) {
2348                                         continue;
2349                                 }
2350                                 disk->d_sync.ds_syncid = 0;
2351                         }
2352                 }
2353
2354                 /* Reset hint. */
2355                 sc->sc_hint = NULL;
2356                 sc->sc_syncid = syncid;
2357                 if (force) {
2358                         /* Remember to bump syncid on first write. */
2359                         sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2360                 }
2361                 state = G_MIRROR_DEVICE_STATE_RUNNING;
2362                 G_MIRROR_DEBUG(1, "Device %s state changed from %s to %s.",
2363                     sc->sc_name, g_mirror_device_state2str(sc->sc_state),
2364                     g_mirror_device_state2str(state));
2365                 sc->sc_state = state;
2366                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2367                         state = g_mirror_determine_state(disk);
2368                         g_mirror_event_send(disk, state,
2369                             G_MIRROR_EVENT_DONTWAIT);
2370                         if (state == G_MIRROR_DISK_STATE_STALE)
2371                                 sc->sc_bump_id |= G_MIRROR_BUMP_SYNCID;
2372                 }
2373                 break;
2374             }
2375         case G_MIRROR_DEVICE_STATE_RUNNING:
2376                 if (g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE) == 0 &&
2377                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2378                         /*
2379                          * No active disks or no disks at all,
2380                          * so destroy device.
2381                          */
2382                         if (sc->sc_provider != NULL)
2383                                 g_mirror_destroy_provider(sc);
2384                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2385                         break;
2386                 } else if (g_mirror_ndisks(sc,
2387                     G_MIRROR_DISK_STATE_ACTIVE) > 0 &&
2388                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_NEW) == 0) {
2389                         /*
2390                          * We have active disks, launch provider if it doesn't
2391                          * exist.
2392                          */
2393                         if (sc->sc_provider == NULL)
2394                                 g_mirror_launch_provider(sc);
2395                         if (sc->sc_rootmount != NULL) {
2396                                 G_MIRROR_DEBUG(1, "root_mount_rel[%u] %p",
2397                                     __LINE__, sc->sc_rootmount);
2398                                 root_mount_rel(sc->sc_rootmount);
2399                                 sc->sc_rootmount = NULL;
2400                         }
2401                 }
2402                 /*
2403                  * Genid should be bumped immediately, so do it here.
2404                  */
2405                 if ((sc->sc_bump_id & G_MIRROR_BUMP_GENID) != 0) {
2406                         sc->sc_bump_id &= ~G_MIRROR_BUMP_GENID;
2407                         g_mirror_bump_genid(sc);
2408                 }
2409                 break;
2410         default:
2411                 KASSERT(1 == 0, ("Wrong device state (%s, %s).",
2412                     sc->sc_name, g_mirror_device_state2str(sc->sc_state)));
2413                 break;
2414         }
2415 }
2416
2417 /*
2418  * Update disk state and device state if needed.
2419  */
2420 #define DISK_STATE_CHANGED()    G_MIRROR_DEBUG(1,                       \
2421         "Disk %s state changed from %s to %s (device %s).",             \
2422         g_mirror_get_diskname(disk),                                    \
2423         g_mirror_disk_state2str(disk->d_state),                         \
2424         g_mirror_disk_state2str(state), sc->sc_name)
2425 static int
2426 g_mirror_update_disk(struct g_mirror_disk *disk, u_int state)
2427 {
2428         struct g_mirror_softc *sc;
2429
2430         sc = disk->d_softc;
2431         sx_assert(&sc->sc_lock, SX_XLOCKED);
2432
2433 again:
2434         G_MIRROR_DEBUG(3, "Changing disk %s state from %s to %s.",
2435             g_mirror_get_diskname(disk), g_mirror_disk_state2str(disk->d_state),
2436             g_mirror_disk_state2str(state));
2437         switch (state) {
2438         case G_MIRROR_DISK_STATE_NEW:
2439                 /*
2440                  * Possible scenarios:
2441                  * 1. New disk arrive.
2442                  */
2443                 /* Previous state should be NONE. */
2444                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NONE,
2445                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2446                     g_mirror_disk_state2str(disk->d_state)));
2447                 DISK_STATE_CHANGED();
2448
2449                 disk->d_state = state;
2450                 if (LIST_EMPTY(&sc->sc_disks))
2451                         LIST_INSERT_HEAD(&sc->sc_disks, disk, d_next);
2452                 else {
2453                         struct g_mirror_disk *dp;
2454
2455                         LIST_FOREACH(dp, &sc->sc_disks, d_next) {
2456                                 if (disk->d_priority >= dp->d_priority) {
2457                                         LIST_INSERT_BEFORE(dp, disk, d_next);
2458                                         dp = NULL;
2459                                         break;
2460                                 }
2461                                 if (LIST_NEXT(dp, d_next) == NULL)
2462                                         break;
2463                         }
2464                         if (dp != NULL)
2465                                 LIST_INSERT_AFTER(dp, disk, d_next);
2466                 }
2467                 G_MIRROR_DEBUG(1, "Device %s: provider %s detected.",
2468                     sc->sc_name, g_mirror_get_diskname(disk));
2469                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
2470                         break;
2471                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2472                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2473                     g_mirror_device_state2str(sc->sc_state),
2474                     g_mirror_get_diskname(disk),
2475                     g_mirror_disk_state2str(disk->d_state)));
2476                 state = g_mirror_determine_state(disk);
2477                 if (state != G_MIRROR_DISK_STATE_NONE)
2478                         goto again;
2479                 break;
2480         case G_MIRROR_DISK_STATE_ACTIVE:
2481                 /*
2482                  * Possible scenarios:
2483                  * 1. New disk does not need synchronization.
2484                  * 2. Synchronization process finished successfully.
2485                  */
2486                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2487                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2488                     g_mirror_device_state2str(sc->sc_state),
2489                     g_mirror_get_diskname(disk),
2490                     g_mirror_disk_state2str(disk->d_state)));
2491                 /* Previous state should be NEW or SYNCHRONIZING. */
2492                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW ||
2493                     disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2494                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2495                     g_mirror_disk_state2str(disk->d_state)));
2496                 DISK_STATE_CHANGED();
2497
2498                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2499                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_SYNCHRONIZING;
2500                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_FORCE_SYNC;
2501                         g_mirror_sync_stop(disk, 0);
2502                 }
2503                 disk->d_state = state;
2504                 disk->d_sync.ds_offset = 0;
2505                 disk->d_sync.ds_offset_done = 0;
2506                 g_mirror_update_idle(sc, disk);
2507                 g_mirror_update_metadata(disk);
2508                 G_MIRROR_DEBUG(1, "Device %s: provider %s activated.",
2509                     sc->sc_name, g_mirror_get_diskname(disk));
2510                 break;
2511         case G_MIRROR_DISK_STATE_STALE:
2512                 /*
2513                  * Possible scenarios:
2514                  * 1. Stale disk was connected.
2515                  */
2516                 /* Previous state should be NEW. */
2517                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2518                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2519                     g_mirror_disk_state2str(disk->d_state)));
2520                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2521                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2522                     g_mirror_device_state2str(sc->sc_state),
2523                     g_mirror_get_diskname(disk),
2524                     g_mirror_disk_state2str(disk->d_state)));
2525                 /*
2526                  * STALE state is only possible if device is marked
2527                  * NOAUTOSYNC.
2528                  */
2529                 KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_NOAUTOSYNC) != 0,
2530                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2531                     g_mirror_device_state2str(sc->sc_state),
2532                     g_mirror_get_diskname(disk),
2533                     g_mirror_disk_state2str(disk->d_state)));
2534                 DISK_STATE_CHANGED();
2535
2536                 disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2537                 disk->d_state = state;
2538                 g_mirror_update_metadata(disk);
2539                 G_MIRROR_DEBUG(0, "Device %s: provider %s is stale.",
2540                     sc->sc_name, g_mirror_get_diskname(disk));
2541                 break;
2542         case G_MIRROR_DISK_STATE_SYNCHRONIZING:
2543                 /*
2544                  * Possible scenarios:
2545                  * 1. Disk which needs synchronization was connected.
2546                  */
2547                 /* Previous state should be NEW. */
2548                 KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2549                     ("Wrong disk state (%s, %s).", g_mirror_get_diskname(disk),
2550                     g_mirror_disk_state2str(disk->d_state)));
2551                 KASSERT(sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING,
2552                     ("Wrong device state (%s, %s, %s, %s).", sc->sc_name,
2553                     g_mirror_device_state2str(sc->sc_state),
2554                     g_mirror_get_diskname(disk),
2555                     g_mirror_disk_state2str(disk->d_state)));
2556                 DISK_STATE_CHANGED();
2557
2558                 if (disk->d_state == G_MIRROR_DISK_STATE_NEW)
2559                         disk->d_flags &= ~G_MIRROR_DISK_FLAG_DIRTY;
2560                 disk->d_state = state;
2561                 if (sc->sc_provider != NULL) {
2562                         g_mirror_sync_start(disk);
2563                         g_mirror_update_metadata(disk);
2564                 }
2565                 break;
2566         case G_MIRROR_DISK_STATE_DISCONNECTED:
2567                 /*
2568                  * Possible scenarios:
2569                  * 1. Device wasn't running yet, but disk disappear.
2570                  * 2. Disk was active and disapppear.
2571                  * 3. Disk disappear during synchronization process.
2572                  */
2573                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING) {
2574                         /*
2575                          * Previous state should be ACTIVE, STALE or
2576                          * SYNCHRONIZING.
2577                          */
2578                         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_ACTIVE ||
2579                             disk->d_state == G_MIRROR_DISK_STATE_STALE ||
2580                             disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING,
2581                             ("Wrong disk state (%s, %s).",
2582                             g_mirror_get_diskname(disk),
2583                             g_mirror_disk_state2str(disk->d_state)));
2584                 } else if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING) {
2585                         /* Previous state should be NEW. */
2586                         KASSERT(disk->d_state == G_MIRROR_DISK_STATE_NEW,
2587                             ("Wrong disk state (%s, %s).",
2588                             g_mirror_get_diskname(disk),
2589                             g_mirror_disk_state2str(disk->d_state)));
2590                         /*
2591                          * Reset bumping syncid if disk disappeared in STARTING
2592                          * state.
2593                          */
2594                         if ((sc->sc_bump_id & G_MIRROR_BUMP_SYNCID) != 0)
2595                                 sc->sc_bump_id &= ~G_MIRROR_BUMP_SYNCID;
2596 #ifdef  INVARIANTS
2597                 } else {
2598                         KASSERT(1 == 0, ("Wrong device state (%s, %s, %s, %s).",
2599                             sc->sc_name,
2600                             g_mirror_device_state2str(sc->sc_state),
2601                             g_mirror_get_diskname(disk),
2602                             g_mirror_disk_state2str(disk->d_state)));
2603 #endif
2604                 }
2605                 DISK_STATE_CHANGED();
2606                 G_MIRROR_DEBUG(0, "Device %s: provider %s disconnected.",
2607                     sc->sc_name, g_mirror_get_diskname(disk));
2608
2609                 g_mirror_destroy_disk(disk);
2610                 break;
2611         case G_MIRROR_DISK_STATE_DESTROY:
2612             {
2613                 int error;
2614
2615                 error = g_mirror_clear_metadata(disk);
2616                 if (error != 0)
2617                         return (error);
2618                 DISK_STATE_CHANGED();
2619                 G_MIRROR_DEBUG(0, "Device %s: provider %s destroyed.",
2620                     sc->sc_name, g_mirror_get_diskname(disk));
2621
2622                 g_mirror_destroy_disk(disk);
2623                 sc->sc_ndisks--;
2624                 LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2625                         g_mirror_update_metadata(disk);
2626                 }
2627                 break;
2628             }
2629         default:
2630                 KASSERT(1 == 0, ("Unknown state (%u).", state));
2631                 break;
2632         }
2633         return (0);
2634 }
2635 #undef  DISK_STATE_CHANGED
2636
2637 int
2638 g_mirror_read_metadata(struct g_consumer *cp, struct g_mirror_metadata *md)
2639 {
2640         struct g_provider *pp;
2641         u_char *buf;
2642         int error;
2643
2644         g_topology_assert();
2645
2646         error = g_access(cp, 1, 0, 0);
2647         if (error != 0)
2648                 return (error);
2649         pp = cp->provider;
2650         g_topology_unlock();
2651         /* Metadata are stored on last sector. */
2652         buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
2653             &error);
2654         g_topology_lock();
2655         g_access(cp, -1, 0, 0);
2656         if (buf == NULL) {
2657                 G_MIRROR_DEBUG(1, "Cannot read metadata from %s (error=%d).",
2658                     cp->provider->name, error);
2659                 return (error);
2660         }
2661
2662         /* Decode metadata. */
2663         error = mirror_metadata_decode(buf, md);
2664         g_free(buf);
2665         if (strcmp(md->md_magic, G_MIRROR_MAGIC) != 0)
2666                 return (EINVAL);
2667         if (md->md_version > G_MIRROR_VERSION) {
2668                 G_MIRROR_DEBUG(0,
2669                     "Kernel module is too old to handle metadata from %s.",
2670                     cp->provider->name);
2671                 return (EINVAL);
2672         }
2673         if (error != 0) {
2674                 G_MIRROR_DEBUG(1, "MD5 metadata hash mismatch for provider %s.",
2675                     cp->provider->name);
2676                 return (error);
2677         }
2678
2679         return (0);
2680 }
2681
2682 static int
2683 g_mirror_check_metadata(struct g_mirror_softc *sc, struct g_provider *pp,
2684     struct g_mirror_metadata *md)
2685 {
2686
2687         if (g_mirror_id2disk(sc, md->md_did) != NULL) {
2688                 G_MIRROR_DEBUG(1, "Disk %s (id=%u) already exists, skipping.",
2689                     pp->name, md->md_did);
2690                 return (EEXIST);
2691         }
2692         if (md->md_all != sc->sc_ndisks) {
2693                 G_MIRROR_DEBUG(1,
2694                     "Invalid '%s' field on disk %s (device %s), skipping.",
2695                     "md_all", pp->name, sc->sc_name);
2696                 return (EINVAL);
2697         }
2698         if (md->md_slice != sc->sc_slice) {
2699                 G_MIRROR_DEBUG(1,
2700                     "Invalid '%s' field on disk %s (device %s), skipping.",
2701                     "md_slice", pp->name, sc->sc_name);
2702                 return (EINVAL);
2703         }
2704         if (md->md_balance != sc->sc_balance) {
2705                 G_MIRROR_DEBUG(1,
2706                     "Invalid '%s' field on disk %s (device %s), skipping.",
2707                     "md_balance", pp->name, sc->sc_name);
2708                 return (EINVAL);
2709         }
2710         if (md->md_mediasize != sc->sc_mediasize) {
2711                 G_MIRROR_DEBUG(1,
2712                     "Invalid '%s' field on disk %s (device %s), skipping.",
2713                     "md_mediasize", pp->name, sc->sc_name);
2714                 return (EINVAL);
2715         }
2716         if (sc->sc_mediasize > pp->mediasize) {
2717                 G_MIRROR_DEBUG(1,
2718                     "Invalid size of disk %s (device %s), skipping.", pp->name,
2719                     sc->sc_name);
2720                 return (EINVAL);
2721         }
2722         if (md->md_sectorsize != sc->sc_sectorsize) {
2723                 G_MIRROR_DEBUG(1,
2724                     "Invalid '%s' field on disk %s (device %s), skipping.",
2725                     "md_sectorsize", pp->name, sc->sc_name);
2726                 return (EINVAL);
2727         }
2728         if ((sc->sc_sectorsize % pp->sectorsize) != 0) {
2729                 G_MIRROR_DEBUG(1,
2730                     "Invalid sector size of disk %s (device %s), skipping.",
2731                     pp->name, sc->sc_name);
2732                 return (EINVAL);
2733         }
2734         if ((md->md_mflags & ~G_MIRROR_DEVICE_FLAG_MASK) != 0) {
2735                 G_MIRROR_DEBUG(1,
2736                     "Invalid device flags on disk %s (device %s), skipping.",
2737                     pp->name, sc->sc_name);
2738                 return (EINVAL);
2739         }
2740         if ((md->md_dflags & ~G_MIRROR_DISK_FLAG_MASK) != 0) {
2741                 G_MIRROR_DEBUG(1,
2742                     "Invalid disk flags on disk %s (device %s), skipping.",
2743                     pp->name, sc->sc_name);
2744                 return (EINVAL);
2745         }
2746         return (0);
2747 }
2748
2749 int
2750 g_mirror_add_disk(struct g_mirror_softc *sc, struct g_provider *pp,
2751     struct g_mirror_metadata *md)
2752 {
2753         struct g_mirror_disk *disk;
2754         int error;
2755
2756         g_topology_assert_not();
2757         G_MIRROR_DEBUG(2, "Adding disk %s.", pp->name);
2758
2759         error = g_mirror_check_metadata(sc, pp, md);
2760         if (error != 0)
2761                 return (error);
2762         if (sc->sc_state == G_MIRROR_DEVICE_STATE_RUNNING &&
2763             md->md_genid < sc->sc_genid) {
2764                 G_MIRROR_DEBUG(0, "Component %s (device %s) broken, skipping.",
2765                     pp->name, sc->sc_name);
2766                 return (EINVAL);
2767         }
2768         disk = g_mirror_init_disk(sc, pp, md, &error);
2769         if (disk == NULL)
2770                 return (error);
2771         error = g_mirror_event_send(disk, G_MIRROR_DISK_STATE_NEW,
2772             G_MIRROR_EVENT_WAIT);
2773         if (error != 0)
2774                 return (error);
2775         if (md->md_version < G_MIRROR_VERSION) {
2776                 G_MIRROR_DEBUG(0, "Upgrading metadata on %s (v%d->v%d).",
2777                     pp->name, md->md_version, G_MIRROR_VERSION);
2778                 g_mirror_update_metadata(disk);
2779         }
2780         return (0);
2781 }
2782
2783 static void
2784 g_mirror_destroy_delayed(void *arg, int flag)
2785 {
2786         struct g_mirror_softc *sc;
2787         int error;
2788
2789         if (flag == EV_CANCEL) {
2790                 G_MIRROR_DEBUG(1, "Destroying canceled.");
2791                 return;
2792         }
2793         sc = arg;
2794         g_topology_unlock();
2795         sx_xlock(&sc->sc_lock);
2796         KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) == 0,
2797             ("DESTROY flag set on %s.", sc->sc_name));
2798         KASSERT((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0,
2799             ("DESTROYING flag not set on %s.", sc->sc_name));
2800         G_MIRROR_DEBUG(1, "Destroying %s (delayed).", sc->sc_name);
2801         error = g_mirror_destroy(sc, G_MIRROR_DESTROY_SOFT);
2802         if (error != 0) {
2803                 G_MIRROR_DEBUG(0, "Cannot destroy %s.", sc->sc_name);
2804                 sx_xunlock(&sc->sc_lock);
2805         }
2806         g_topology_lock();
2807 }
2808
2809 static int
2810 g_mirror_access(struct g_provider *pp, int acr, int acw, int ace)
2811 {
2812         struct g_mirror_softc *sc;
2813         int dcr, dcw, dce, error = 0;
2814
2815         g_topology_assert();
2816         G_MIRROR_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, acr,
2817             acw, ace);
2818
2819         sc = pp->geom->softc;
2820         if (sc == NULL && acr <= 0 && acw <= 0 && ace <= 0)
2821                 return (0);
2822         KASSERT(sc != NULL, ("NULL softc (provider=%s).", pp->name));
2823
2824         dcr = pp->acr + acr;
2825         dcw = pp->acw + acw;
2826         dce = pp->ace + ace;
2827
2828         g_topology_unlock();
2829         sx_xlock(&sc->sc_lock);
2830         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0 ||
2831             LIST_EMPTY(&sc->sc_disks)) {
2832                 if (acr > 0 || acw > 0 || ace > 0)
2833                         error = ENXIO;
2834                 goto end;
2835         }
2836         if (dcw == 0)
2837                 g_mirror_idle(sc, dcw);
2838         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROYING) != 0) {
2839                 if (acr > 0 || acw > 0 || ace > 0) {
2840                         error = ENXIO;
2841                         goto end;
2842                 }
2843                 if (dcr == 0 && dcw == 0 && dce == 0) {
2844                         g_post_event(g_mirror_destroy_delayed, sc, M_WAITOK,
2845                             sc, NULL);
2846                 }
2847         }
2848 end:
2849         sx_xunlock(&sc->sc_lock);
2850         g_topology_lock();
2851         return (error);
2852 }
2853
2854 static struct g_geom *
2855 g_mirror_create(struct g_class *mp, const struct g_mirror_metadata *md)
2856 {
2857         struct g_mirror_softc *sc;
2858         struct g_geom *gp;
2859         int error, timeout;
2860
2861         g_topology_assert();
2862         G_MIRROR_DEBUG(1, "Creating device %s (id=%u).", md->md_name,
2863             md->md_mid);
2864
2865         /* One disk is minimum. */
2866         if (md->md_all < 1)
2867                 return (NULL);
2868         /*
2869          * Action geom.
2870          */
2871         gp = g_new_geomf(mp, "%s", md->md_name);
2872         sc = malloc(sizeof(*sc), M_MIRROR, M_WAITOK | M_ZERO);
2873         gp->start = g_mirror_start;
2874         gp->orphan = g_mirror_orphan;
2875         gp->access = g_mirror_access;
2876         gp->dumpconf = g_mirror_dumpconf;
2877
2878         sc->sc_id = md->md_mid;
2879         sc->sc_slice = md->md_slice;
2880         sc->sc_balance = md->md_balance;
2881         sc->sc_mediasize = md->md_mediasize;
2882         sc->sc_sectorsize = md->md_sectorsize;
2883         sc->sc_ndisks = md->md_all;
2884         sc->sc_flags = md->md_mflags;
2885         sc->sc_bump_id = 0;
2886         sc->sc_idle = 1;
2887         sc->sc_last_write = time_uptime;
2888         sc->sc_writes = 0;
2889         sx_init(&sc->sc_lock, "gmirror:lock");
2890         bioq_init(&sc->sc_queue);
2891         mtx_init(&sc->sc_queue_mtx, "gmirror:queue", NULL, MTX_DEF);
2892         bioq_init(&sc->sc_regular_delayed);
2893         bioq_init(&sc->sc_inflight);
2894         bioq_init(&sc->sc_sync_delayed);
2895         LIST_INIT(&sc->sc_disks);
2896         TAILQ_INIT(&sc->sc_events);
2897         mtx_init(&sc->sc_events_mtx, "gmirror:events", NULL, MTX_DEF);
2898         callout_init(&sc->sc_callout, CALLOUT_MPSAFE);
2899         sc->sc_state = G_MIRROR_DEVICE_STATE_STARTING;
2900         gp->softc = sc;
2901         sc->sc_geom = gp;
2902         sc->sc_provider = NULL;
2903         /*
2904          * Synchronization geom.
2905          */
2906         gp = g_new_geomf(mp, "%s.sync", md->md_name);
2907         gp->softc = sc;
2908         gp->orphan = g_mirror_orphan;
2909         sc->sc_sync.ds_geom = gp;
2910         sc->sc_sync.ds_ndisks = 0;
2911         error = kproc_create(g_mirror_worker, sc, &sc->sc_worker, 0, 0,
2912             "g_mirror %s", md->md_name);
2913         if (error != 0) {
2914                 G_MIRROR_DEBUG(1, "Cannot create kernel thread for %s.",
2915                     sc->sc_name);
2916                 g_destroy_geom(sc->sc_sync.ds_geom);
2917                 mtx_destroy(&sc->sc_events_mtx);
2918                 mtx_destroy(&sc->sc_queue_mtx);
2919                 sx_destroy(&sc->sc_lock);
2920                 g_destroy_geom(sc->sc_geom);
2921                 free(sc, M_MIRROR);
2922                 return (NULL);
2923         }
2924
2925         G_MIRROR_DEBUG(1, "Device %s created (%u components, id=%u).",
2926             sc->sc_name, sc->sc_ndisks, sc->sc_id);
2927
2928         sc->sc_rootmount = root_mount_hold("GMIRROR");
2929         G_MIRROR_DEBUG(1, "root_mount_hold %p", sc->sc_rootmount);
2930         /*
2931          * Run timeout.
2932          */
2933         timeout = g_mirror_timeout * hz;
2934         callout_reset(&sc->sc_callout, timeout, g_mirror_go, sc);
2935         return (sc->sc_geom);
2936 }
2937
2938 int
2939 g_mirror_destroy(struct g_mirror_softc *sc, int how)
2940 {
2941         struct g_mirror_disk *disk;
2942         struct g_provider *pp;
2943
2944         g_topology_assert_not();
2945         if (sc == NULL)
2946                 return (ENXIO);
2947         sx_assert(&sc->sc_lock, SX_XLOCKED);
2948
2949         pp = sc->sc_provider;
2950         if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
2951                 switch (how) {
2952                 case G_MIRROR_DESTROY_SOFT:
2953                         G_MIRROR_DEBUG(1,
2954                             "Device %s is still open (r%dw%de%d).", pp->name,
2955                             pp->acr, pp->acw, pp->ace);
2956                         return (EBUSY);
2957                 case G_MIRROR_DESTROY_DELAYED:
2958                         G_MIRROR_DEBUG(1,
2959                             "Device %s will be destroyed on last close.",
2960                             pp->name);
2961                         LIST_FOREACH(disk, &sc->sc_disks, d_next) {
2962                                 if (disk->d_state ==
2963                                     G_MIRROR_DISK_STATE_SYNCHRONIZING) {
2964                                         g_mirror_sync_stop(disk, 1);
2965                                 }
2966                         }
2967                         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROYING;
2968                         return (EBUSY);
2969                 case G_MIRROR_DESTROY_HARD:
2970                         G_MIRROR_DEBUG(1, "Device %s is still open, so it "
2971                             "can't be definitely removed.", pp->name);
2972                 }
2973         }
2974
2975         g_topology_lock();
2976         if (sc->sc_geom->softc == NULL) {
2977                 g_topology_unlock();
2978                 return (0);
2979         }
2980         sc->sc_geom->softc = NULL;
2981         sc->sc_sync.ds_geom->softc = NULL;
2982         g_topology_unlock();
2983
2984         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_DESTROY;
2985         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_WAIT;
2986         G_MIRROR_DEBUG(4, "%s: Waking up %p.", __func__, sc);
2987         sx_xunlock(&sc->sc_lock);
2988         mtx_lock(&sc->sc_queue_mtx);
2989         wakeup(sc);
2990         mtx_unlock(&sc->sc_queue_mtx);
2991         G_MIRROR_DEBUG(4, "%s: Sleeping %p.", __func__, &sc->sc_worker);
2992         while (sc->sc_worker != NULL)
2993                 tsleep(&sc->sc_worker, PRIBIO, "m:destroy", hz / 5);
2994         G_MIRROR_DEBUG(4, "%s: Woken up %p.", __func__, &sc->sc_worker);
2995         sx_xlock(&sc->sc_lock);
2996         g_mirror_destroy_device(sc);
2997         free(sc, M_MIRROR);
2998         return (0);
2999 }
3000
3001 static void
3002 g_mirror_taste_orphan(struct g_consumer *cp)
3003 {
3004
3005         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
3006             cp->provider->name));
3007 }
3008
3009 static struct g_geom *
3010 g_mirror_taste(struct g_class *mp, struct g_provider *pp, int flags __unused)
3011 {
3012         struct g_mirror_metadata md;
3013         struct g_mirror_softc *sc;
3014         struct g_consumer *cp;
3015         struct g_geom *gp;
3016         int error;
3017
3018         g_topology_assert();
3019         g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
3020         G_MIRROR_DEBUG(2, "Tasting %s.", pp->name);
3021
3022         gp = g_new_geomf(mp, "mirror:taste");
3023         /*
3024          * This orphan function should be never called.
3025          */
3026         gp->orphan = g_mirror_taste_orphan;
3027         cp = g_new_consumer(gp);
3028         g_attach(cp, pp);
3029         error = g_mirror_read_metadata(cp, &md);
3030         g_detach(cp);
3031         g_destroy_consumer(cp);
3032         g_destroy_geom(gp);
3033         if (error != 0)
3034                 return (NULL);
3035         gp = NULL;
3036
3037         if (md.md_provider[0] != '\0' &&
3038             !g_compare_names(md.md_provider, pp->name))
3039                 return (NULL);
3040         if (md.md_provsize != 0 && md.md_provsize != pp->mediasize)
3041                 return (NULL);
3042         if ((md.md_dflags & G_MIRROR_DISK_FLAG_INACTIVE) != 0) {
3043                 G_MIRROR_DEBUG(0,
3044                     "Device %s: provider %s marked as inactive, skipping.",
3045                     md.md_name, pp->name);
3046                 return (NULL);
3047         }
3048         if (g_mirror_debug >= 2)
3049                 mirror_metadata_dump(&md);
3050
3051         /*
3052          * Let's check if device already exists.
3053          */
3054         sc = NULL;
3055         LIST_FOREACH(gp, &mp->geom, geom) {
3056                 sc = gp->softc;
3057                 if (sc == NULL)
3058                         continue;
3059                 if (sc->sc_sync.ds_geom == gp)
3060                         continue;
3061                 if (strcmp(md.md_name, sc->sc_name) != 0)
3062                         continue;
3063                 if (md.md_mid != sc->sc_id) {
3064                         G_MIRROR_DEBUG(0, "Device %s already configured.",
3065                             sc->sc_name);
3066                         return (NULL);
3067                 }
3068                 break;
3069         }
3070         if (gp == NULL) {
3071                 gp = g_mirror_create(mp, &md);
3072                 if (gp == NULL) {
3073                         G_MIRROR_DEBUG(0, "Cannot create device %s.",
3074                             md.md_name);
3075                         return (NULL);
3076                 }
3077                 sc = gp->softc;
3078         }
3079         G_MIRROR_DEBUG(1, "Adding disk %s to %s.", pp->name, gp->name);
3080         g_topology_unlock();
3081         sx_xlock(&sc->sc_lock);
3082         sc->sc_flags |= G_MIRROR_DEVICE_FLAG_TASTING;
3083         error = g_mirror_add_disk(sc, pp, &md);
3084         if (error != 0) {
3085                 G_MIRROR_DEBUG(0, "Cannot add disk %s to %s (error=%d).",
3086                     pp->name, gp->name, error);
3087                 if (LIST_EMPTY(&sc->sc_disks)) {
3088                         g_cancel_event(sc);
3089                         g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3090                         g_topology_lock();
3091                         return (NULL);
3092                 }
3093                 gp = NULL;
3094         }
3095         sc->sc_flags &= ~G_MIRROR_DEVICE_FLAG_TASTING;
3096         if ((sc->sc_flags & G_MIRROR_DEVICE_FLAG_DESTROY) != 0) {
3097                 g_mirror_destroy(sc, G_MIRROR_DESTROY_HARD);
3098                 g_topology_lock();
3099                 return (NULL);
3100         }
3101         sx_xunlock(&sc->sc_lock);
3102         g_topology_lock();
3103         return (gp);
3104 }
3105
3106 static int
3107 g_mirror_destroy_geom(struct gctl_req *req __unused,
3108     struct g_class *mp __unused, struct g_geom *gp)
3109 {
3110         struct g_mirror_softc *sc;
3111         int error;
3112
3113         g_topology_unlock();
3114         sc = gp->softc;
3115         sx_xlock(&sc->sc_lock);
3116         g_cancel_event(sc);
3117         error = g_mirror_destroy(gp->softc, G_MIRROR_DESTROY_SOFT);
3118         if (error != 0)
3119                 sx_xunlock(&sc->sc_lock);
3120         g_topology_lock();
3121         return (error);
3122 }
3123
3124 static void
3125 g_mirror_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
3126     struct g_consumer *cp, struct g_provider *pp)
3127 {
3128         struct g_mirror_softc *sc;
3129
3130         g_topology_assert();
3131
3132         sc = gp->softc;
3133         if (sc == NULL)
3134                 return;
3135         /* Skip synchronization geom. */
3136         if (gp == sc->sc_sync.ds_geom)
3137                 return;
3138         if (pp != NULL) {
3139                 /* Nothing here. */
3140         } else if (cp != NULL) {
3141                 struct g_mirror_disk *disk;
3142
3143                 disk = cp->private;
3144                 if (disk == NULL)
3145                         return;
3146                 g_topology_unlock();
3147                 sx_xlock(&sc->sc_lock);
3148                 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)disk->d_id);
3149                 if (disk->d_state == G_MIRROR_DISK_STATE_SYNCHRONIZING) {
3150                         sbuf_printf(sb, "%s<Synchronized>", indent);
3151                         if (disk->d_sync.ds_offset == 0)
3152                                 sbuf_printf(sb, "0%%");
3153                         else {
3154                                 sbuf_printf(sb, "%u%%",
3155                                     (u_int)((disk->d_sync.ds_offset * 100) /
3156                                     sc->sc_provider->mediasize));
3157                         }
3158                         sbuf_printf(sb, "</Synchronized>\n");
3159                         if (disk->d_sync.ds_offset > 0) {
3160                                 sbuf_printf(sb, "%s<BytesSynced>%jd"
3161                                     "</BytesSynced>\n", indent,
3162                                     (intmax_t)disk->d_sync.ds_offset);
3163                         }
3164                 }
3165                 sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent,
3166                     disk->d_sync.ds_syncid);
3167                 sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent,
3168                     disk->d_genid);
3169                 sbuf_printf(sb, "%s<Flags>", indent);
3170                 if (disk->d_flags == 0)
3171                         sbuf_printf(sb, "NONE");
3172                 else {
3173                         int first = 1;
3174
3175 #define ADD_FLAG(flag, name)    do {                                    \
3176         if ((disk->d_flags & (flag)) != 0) {                            \
3177                 if (!first)                                             \
3178                         sbuf_printf(sb, ", ");                          \
3179                 else                                                    \
3180                         first = 0;                                      \
3181                 sbuf_printf(sb, name);                                  \
3182         }                                                               \
3183 } while (0)
3184                         ADD_FLAG(G_MIRROR_DISK_FLAG_DIRTY, "DIRTY");
3185                         ADD_FLAG(G_MIRROR_DISK_FLAG_HARDCODED, "HARDCODED");
3186                         ADD_FLAG(G_MIRROR_DISK_FLAG_INACTIVE, "INACTIVE");
3187                         ADD_FLAG(G_MIRROR_DISK_FLAG_SYNCHRONIZING,
3188                             "SYNCHRONIZING");
3189                         ADD_FLAG(G_MIRROR_DISK_FLAG_FORCE_SYNC, "FORCE_SYNC");
3190                         ADD_FLAG(G_MIRROR_DISK_FLAG_BROKEN, "BROKEN");
3191 #undef  ADD_FLAG
3192                 }
3193                 sbuf_printf(sb, "</Flags>\n");
3194                 sbuf_printf(sb, "%s<Priority>%u</Priority>\n", indent,
3195                     disk->d_priority);
3196                 sbuf_printf(sb, "%s<State>%s</State>\n", indent,
3197                     g_mirror_disk_state2str(disk->d_state));
3198                 sx_xunlock(&sc->sc_lock);
3199                 g_topology_lock();
3200         } else {
3201                 g_topology_unlock();
3202                 sx_xlock(&sc->sc_lock);
3203                 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id);
3204                 sbuf_printf(sb, "%s<SyncID>%u</SyncID>\n", indent, sc->sc_syncid);
3205                 sbuf_printf(sb, "%s<GenID>%u</GenID>\n", indent, sc->sc_genid);
3206                 sbuf_printf(sb, "%s<Flags>", indent);
3207                 if (sc->sc_flags == 0)
3208                         sbuf_printf(sb, "NONE");
3209                 else {
3210                         int first = 1;
3211
3212 #define ADD_FLAG(flag, name)    do {                                    \
3213         if ((sc->sc_flags & (flag)) != 0) {                             \
3214                 if (!first)                                             \
3215                         sbuf_printf(sb, ", ");                          \
3216                 else                                                    \
3217                         first = 0;                                      \
3218                 sbuf_printf(sb, name);                                  \
3219         }                                                               \
3220 } while (0)
3221                         ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOFAILSYNC, "NOFAILSYNC");
3222                         ADD_FLAG(G_MIRROR_DEVICE_FLAG_NOAUTOSYNC, "NOAUTOSYNC");
3223 #undef  ADD_FLAG
3224                 }
3225                 sbuf_printf(sb, "</Flags>\n");
3226                 sbuf_printf(sb, "%s<Slice>%u</Slice>\n", indent,
3227                     (u_int)sc->sc_slice);
3228                 sbuf_printf(sb, "%s<Balance>%s</Balance>\n", indent,
3229                     balance_name(sc->sc_balance));
3230                 sbuf_printf(sb, "%s<Components>%u</Components>\n", indent,
3231                     sc->sc_ndisks);
3232                 sbuf_printf(sb, "%s<State>", indent);
3233                 if (sc->sc_state == G_MIRROR_DEVICE_STATE_STARTING)
3234                         sbuf_printf(sb, "%s", "STARTING");
3235                 else if (sc->sc_ndisks ==
3236                     g_mirror_ndisks(sc, G_MIRROR_DISK_STATE_ACTIVE))
3237                         sbuf_printf(sb, "%s", "COMPLETE");
3238                 else
3239                         sbuf_printf(sb, "%s", "DEGRADED");
3240                 sbuf_printf(sb, "</State>\n");
3241                 sx_xunlock(&sc->sc_lock);
3242                 g_topology_lock();
3243         }
3244 }
3245
3246 static void
3247 g_mirror_shutdown_post_sync(void *arg, int howto)
3248 {
3249         struct g_class *mp;
3250         struct g_geom *gp, *gp2;
3251         struct g_mirror_softc *sc;
3252         int error;
3253
3254         mp = arg;
3255         DROP_GIANT();
3256         g_topology_lock();
3257         g_mirror_shutdown = 1;
3258         LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) {
3259                 if ((sc = gp->softc) == NULL)
3260                         continue;
3261                 /* Skip synchronization geom. */
3262                 if (gp == sc->sc_sync.ds_geom)
3263                         continue;
3264                 g_topology_unlock();
3265                 sx_xlock(&sc->sc_lock);
3266                 g_mirror_idle(sc, -1);
3267                 g_cancel_event(sc);
3268                 error = g_mirror_destroy(sc, G_MIRROR_DESTROY_DELAYED);
3269                 if (error != 0)
3270                         sx_xunlock(&sc->sc_lock);
3271                 g_topology_lock();
3272         }
3273         g_topology_unlock();
3274         PICKUP_GIANT();
3275 }
3276
3277 static void
3278 g_mirror_init(struct g_class *mp)
3279 {
3280
3281         g_mirror_post_sync = EVENTHANDLER_REGISTER(shutdown_post_sync,
3282             g_mirror_shutdown_post_sync, mp, SHUTDOWN_PRI_FIRST);
3283         if (g_mirror_post_sync == NULL)
3284                 G_MIRROR_DEBUG(0, "Warning! Cannot register shutdown event.");
3285 }
3286
3287 static void
3288 g_mirror_fini(struct g_class *mp)
3289 {
3290
3291         if (g_mirror_post_sync != NULL)
3292                 EVENTHANDLER_DEREGISTER(shutdown_post_sync, g_mirror_post_sync);
3293 }
3294
3295 DECLARE_GEOM_CLASS(g_mirror_class, g_mirror);