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