sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_geom.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
  23  * All rights reserved.
  24  *
  25  * Portions Copyright (c) 2012 Martin Matuska <mm@FreeBSD.org>
  26  */
  27
  28 #include <sys/zfs_context.h>
  29 #include <sys/param.h>
  30 #include <sys/kernel.h>
  31 #include <sys/bio.h>
  32 #include <sys/disk.h>
  33 #include <sys/spa.h>
  34 #include <sys/spa_impl.h>
  35 #include <sys/vdev_impl.h>
  36 #include <sys/fs/zfs.h>
  37 #include <sys/zio.h>
  38 #include <geom/geom.h>
  39 #include <geom/geom_int.h>
  40
  41 /*
  42  * Virtual device vector for GEOM.
  43  */
  44
  45 struct g_class zfs_vdev_class = {
  46         .name = "ZFS::VDEV",
  47         .version = G_VERSION,
  48 };
  49
  50 DECLARE_GEOM_CLASS(zfs_vdev_class, zfs_vdev);
  51
  52 SYSCTL_DECL(_vfs_zfs_vdev);
  53 /* Don't send BIO_FLUSH. */
  54 static int vdev_geom_bio_flush_disable = 0;
  55 TUNABLE_INT("vfs.zfs.vdev.bio_flush_disable", &vdev_geom_bio_flush_disable);
  56 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_flush_disable, CTLFLAG_RW,
  57     &vdev_geom_bio_flush_disable, 0, "Disable BIO_FLUSH");
  58 /* Don't send BIO_DELETE. */
  59 static int vdev_geom_bio_delete_disable = 0;
  60 TUNABLE_INT("vfs.zfs.vdev.bio_delete_disable", &vdev_geom_bio_delete_disable);
  61 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_delete_disable, CTLFLAG_RW,
  62     &vdev_geom_bio_delete_disable, 0, "Disable BIO_DELETE");
  63
  64 static void
  65 vdev_geom_orphan(struct g_consumer *cp)
  66 {
  67         vdev_t *vd;
  68
  69         g_topology_assert();
  70
  71         vd = cp->private;
  72
  73         /*
  74          * Orphan callbacks occur from the GEOM event thread.
  75          * Concurrent with this call, new I/O requests may be
  76          * working their way through GEOM about to find out
  77          * (only once executed by the g_down thread) that we've
  78          * been orphaned from our disk provider.  These I/Os
  79          * must be retired before we can detach our consumer.
  80          * This is most easily achieved by acquiring the
  81          * SPA ZIO configuration lock as a writer, but doing
  82          * so with the GEOM topology lock held would cause
  83          * a lock order reversal.  Instead, rely on the SPA's
  84          * async removal support to invoke a close on this
  85          * vdev once it is safe to do so.
  86          */
  87         zfs_post_remove(vd->vdev_spa, vd);
  88         vd->vdev_remove_wanted = B_TRUE;
  89         spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
  90 }
  91
  92 static struct g_consumer *
  93 vdev_geom_attach(struct g_provider *pp)
  94 {
  95         struct g_geom *gp;
  96         struct g_consumer *cp;
  97
  98         g_topology_assert();
  99
 100         ZFS_LOG(1, "Attaching to %s.", pp->name);
 101         /* Do we have geom already? No? Create one. */
 102         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 103                 if (gp->flags & G_GEOM_WITHER)
 104                         continue;
 105                 if (strcmp(gp->name, "zfs::vdev") != 0)
 106                         continue;
 107                 break;
 108         }
 109         if (gp == NULL) {
 110                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 111                 gp->orphan = vdev_geom_orphan;
 112                 cp = g_new_consumer(gp);
 113                 if (g_attach(cp, pp) != 0) {
 114                         g_wither_geom(gp, ENXIO);
 115                         return (NULL);
 116                 }
 117                 if (g_access(cp, 1, 0, 1) != 0) {
 118                         g_wither_geom(gp, ENXIO);
 119                         return (NULL);
 120                 }
 121                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 122         } else {
 123                 /* Check if we are already connected to this provider. */
 124                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 125                         if (cp->provider == pp) {
 126                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 127                                 break;
 128                         }
 129                 }
 130                 if (cp == NULL) {
 131                         cp = g_new_consumer(gp);
 132                         if (g_attach(cp, pp) != 0) {
 133                                 g_destroy_consumer(cp);
 134                                 return (NULL);
 135                         }
 136                         if (g_access(cp, 1, 0, 1) != 0) {
 137                                 g_detach(cp);
 138                                 g_destroy_consumer(cp);
 139                                 return (NULL);
 140                         }
 141                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 142                 } else {
 143                         if (g_access(cp, 1, 0, 1) != 0)
 144                                 return (NULL);
 145                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 146                 }
 147         }
 148         return (cp);
 149 }
 150
 151 static void
 152 vdev_geom_detach(void *arg, int flag __unused)
 153 {
 154         struct g_geom *gp;
 155         struct g_consumer *cp;
 156
 157         g_topology_assert();
 158         cp = arg;
 159         gp = cp->geom;
 160
 161         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 162         g_access(cp, -1, 0, -1);
 163         /* Destroy consumer on last close. */
 164         if (cp->acr == 0 && cp->ace == 0) {
 165                 ZFS_LOG(1, "Destroyed consumer to %s.", cp->provider->name);
 166                 if (cp->acw > 0)
 167                         g_access(cp, 0, -cp->acw, 0);
 168                 g_detach(cp);
 169                 g_destroy_consumer(cp);
 170         }
 171         /* Destroy geom if there are no consumers left. */
 172         if (LIST_EMPTY(&gp->consumer)) {
 173                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 174                 g_wither_geom(gp, ENXIO);
 175         }
 176 }
 177
 178 static uint64_t
 179 nvlist_get_guid(nvlist_t *list)
 180 {
 181         uint64_t value;
 182
 183         value = 0;
 184         nvlist_lookup_uint64(list, ZPOOL_CONFIG_GUID, &value);
 185         return (value);
 186 }
 187
 188 static int
 189 vdev_geom_io(struct g_consumer *cp, int cmd, void *data, off_t offset, off_t size)
 190 {
 191         struct bio *bp;
 192         u_char *p;
 193         off_t off, maxio;
 194         int error;
 195
 196         ASSERT((offset % cp->provider->sectorsize) == 0);
 197         ASSERT((size % cp->provider->sectorsize) == 0);
 198
 199         bp = g_alloc_bio();
 200         off = offset;
 201         offset += size;
 202         p = data;
 203         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 204         error = 0;
 205
 206         for (; off < offset; off += maxio, p += maxio, size -= maxio) {
 207                 bzero(bp, sizeof(*bp));
 208                 bp->bio_cmd = cmd;
 209                 bp->bio_done = NULL;
 210                 bp->bio_offset = off;
 211                 bp->bio_length = MIN(size, maxio);
 212                 bp->bio_data = p;
 213                 g_io_request(bp, cp);
 214                 error = biowait(bp, "vdev_geom_io");
 215                 if (error != 0)
 216                         break;
 217         }
 218
 219         g_destroy_bio(bp);
 220         return (error);
 221 }
 222
 223 static void
 224 vdev_geom_taste_orphan(struct g_consumer *cp)
 225 {
 226
 227         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
 228             cp->provider->name));
 229 }
 230
 231 static int
 232 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 233 {
 234         struct g_provider *pp;
 235         vdev_label_t *label;
 236         char *p, *buf;
 237         size_t buflen;
 238         uint64_t psize;
 239         off_t offset, size;
 240         uint64_t guid, state, txg;
 241         int error, l, len;
 242
 243         g_topology_assert_not();
 244
 245         pp = cp->provider;
 246         ZFS_LOG(1, "Reading config from %s...", pp->name);
 247
 248         psize = pp->mediasize;
 249         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 250
 251         size = sizeof(*label) + pp->sectorsize -
 252             ((sizeof(*label) - 1) % pp->sectorsize) - 1;
 253
 254         guid = 0;
 255         label = kmem_alloc(size, KM_SLEEP);
 256         buflen = sizeof(label->vl_vdev_phys.vp_nvlist);
 257
 258         *config = NULL;
 259         for (l = 0; l < VDEV_LABELS; l++) {
 260
 261                 offset = vdev_label_offset(psize, l, 0);
 262                 if ((offset % pp->sectorsize) != 0)
 263                         continue;
 264
 265                 if (vdev_geom_io(cp, BIO_READ, label, offset, size) != 0)
 266                         continue;
 267                 buf = label->vl_vdev_phys.vp_nvlist;
 268
 269                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 270                         continue;
 271
 272                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 273                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 274                         nvlist_free(*config);
 275                         *config = NULL;
 276                         continue;
 277                 }
 278
 279                 if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
 280                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 281                     &txg) != 0 || txg == 0)) {
 282                         nvlist_free(*config);
 283                         *config = NULL;
 284                         continue;
 285                 }
 286
 287                 break;
 288         }
 289
 290         kmem_free(label, size);
 291         return (*config == NULL ? ENOENT : 0);
 292 }
 293
 294 static void
 295 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 296 {
 297         nvlist_t **new_configs;
 298         uint64_t i;
 299
 300         if (id < *count)
 301                 return;
 302         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 303             KM_SLEEP);
 304         for (i = 0; i < *count; i++)
 305                 new_configs[i] = (*configs)[i];
 306         if (*configs != NULL)
 307                 kmem_free(*configs, *count * sizeof(void *));
 308         *configs = new_configs;
 309         *count = id + 1;
 310 }
 311
 312 static void
 313 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 314     const char *name, uint64_t* known_pool_guid)
 315 {
 316         nvlist_t *vdev_tree;
 317         uint64_t pool_guid;
 318         uint64_t vdev_guid, known_guid;
 319         uint64_t id, txg, known_txg;
 320         char *pname;
 321         int i;
 322
 323         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 324             strcmp(pname, name) != 0)
 325                 goto ignore;
 326
 327         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 328                 goto ignore;
 329
 330         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 331                 goto ignore;
 332
 333         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 334                 goto ignore;
 335
 336         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 337                 goto ignore;
 338
 339         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 340
 341         if (*known_pool_guid != 0) {
 342                 if (pool_guid != *known_pool_guid)
 343                         goto ignore;
 344         } else
 345                 *known_pool_guid = pool_guid;
 346
 347         resize_configs(configs, count, id);
 348
 349         if ((*configs)[id] != NULL) {
 350                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 351                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 352                 if (txg <= known_txg)
 353                         goto ignore;
 354                 nvlist_free((*configs)[id]);
 355         }
 356
 357         (*configs)[id] = cfg;
 358         return;
 359
 360 ignore:
 361         nvlist_free(cfg);
 362 }
 363
 364 static int
 365 vdev_geom_attach_taster(struct g_consumer *cp, struct g_provider *pp)
 366 {
 367         int error;
 368
 369         if (pp->flags & G_PF_WITHER)
 370                 return (EINVAL);
 371         if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize))
 372                 return (EINVAL);
 373         g_attach(cp, pp);
 374         error = g_access(cp, 1, 0, 0);
 375         if (error != 0)
 376                 g_detach(cp);
 377         return (error);
 378 }
 379
 380 static void
 381 vdev_geom_detach_taster(struct g_consumer *cp)
 382 {
 383         g_access(cp, -1, 0, 0);
 384         g_detach(cp);
 385 }
 386
 387 int
 388 vdev_geom_read_pool_label(const char *name,
 389     nvlist_t ***configs, uint64_t *count)
 390 {
 391         struct g_class *mp;
 392         struct g_geom *gp, *zgp;
 393         struct g_provider *pp;
 394         struct g_consumer *zcp;
 395         nvlist_t *vdev_cfg;
 396         uint64_t pool_guid;
 397         int error;
 398
 399         DROP_GIANT();
 400         g_topology_lock();
 401
 402         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 403         /* This orphan function should be never called. */
 404         zgp->orphan = vdev_geom_taste_orphan;
 405         zcp = g_new_consumer(zgp);
 406
 407         *configs = NULL;
 408         *count = 0;
 409         pool_guid = 0;
 410         LIST_FOREACH(mp, &g_classes, class) {
 411                 if (mp == &zfs_vdev_class)
 412                         continue;
 413                 LIST_FOREACH(gp, &mp->geom, geom) {
 414                         if (gp->flags & G_GEOM_WITHER)
 415                                 continue;
 416                         LIST_FOREACH(pp, &gp->provider, provider) {
 417                                 if (pp->flags & G_PF_WITHER)
 418                                         continue;
 419                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 420                                         continue;
 421                                 g_topology_unlock();
 422                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 423                                 g_topology_lock();
 424                                 vdev_geom_detach_taster(zcp);
 425                                 if (error)
 426                                         continue;
 427                                 ZFS_LOG(1, "successfully read vdev config");
 428
 429                                 process_vdev_config(configs, count,
 430                                     vdev_cfg, name, &pool_guid);
 431                         }
 432                 }
 433         }
 434
 435         g_destroy_consumer(zcp);
 436         g_destroy_geom(zgp);
 437         g_topology_unlock();
 438         PICKUP_GIANT();
 439
 440         return (*count > 0 ? 0 : ENOENT);
 441 }
 442
 443 static uint64_t
 444 vdev_geom_read_guid(struct g_consumer *cp)
 445 {
 446         nvlist_t *config;
 447         uint64_t guid;
 448
 449         g_topology_assert_not();
 450
 451         guid = 0;
 452         if (vdev_geom_read_config(cp, &config) == 0) {
 453                 guid = nvlist_get_guid(config);
 454                 nvlist_free(config);
 455         }
 456         return (guid);
 457 }
 458
 459 static struct g_consumer *
 460 vdev_geom_attach_by_guid(uint64_t guid)
 461 {
 462         struct g_class *mp;
 463         struct g_geom *gp, *zgp;
 464         struct g_provider *pp;
 465         struct g_consumer *cp, *zcp;
 466         uint64_t pguid;
 467
 468         g_topology_assert();
 469
 470         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 471         /* This orphan function should be never called. */
 472         zgp->orphan = vdev_geom_taste_orphan;
 473         zcp = g_new_consumer(zgp);
 474
 475         cp = NULL;
 476         LIST_FOREACH(mp, &g_classes, class) {
 477                 if (mp == &zfs_vdev_class)
 478                         continue;
 479                 LIST_FOREACH(gp, &mp->geom, geom) {
 480                         if (gp->flags & G_GEOM_WITHER)
 481                                 continue;
 482                         LIST_FOREACH(pp, &gp->provider, provider) {
 483                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 484                                         continue;
 485                                 g_topology_unlock();
 486                                 pguid = vdev_geom_read_guid(zcp);
 487                                 g_topology_lock();
 488                                 vdev_geom_detach_taster(zcp);
 489                                 if (pguid != guid)
 490                                         continue;
 491                                 cp = vdev_geom_attach(pp);
 492                                 if (cp == NULL) {
 493                                         printf("ZFS WARNING: Unable to attach to %s.\n",
 494                                             pp->name);
 495                                         continue;
 496                                 }
 497                                 break;
 498                         }
 499                         if (cp != NULL)
 500                                 break;
 501                 }
 502                 if (cp != NULL)
 503                         break;
 504         }
 505 end:
 506         g_destroy_consumer(zcp);
 507         g_destroy_geom(zgp);
 508         return (cp);
 509 }
 510
 511 static struct g_consumer *
 512 vdev_geom_open_by_guid(vdev_t *vd)
 513 {
 514         struct g_consumer *cp;
 515         char *buf;
 516         size_t len;
 517
 518         g_topology_assert();
 519
 520         ZFS_LOG(1, "Searching by guid [%ju].", (uintmax_t)vd->vdev_guid);
 521         cp = vdev_geom_attach_by_guid(vd->vdev_guid);
 522         if (cp != NULL) {
 523                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 524                 buf = kmem_alloc(len, KM_SLEEP);
 525
 526                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 527                 spa_strfree(vd->vdev_path);
 528                 vd->vdev_path = buf;
 529
 530                 ZFS_LOG(1, "Attach by guid [%ju] succeeded, provider %s.",
 531                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 532         } else {
 533                 ZFS_LOG(1, "Search by guid [%ju] failed.",
 534                     (uintmax_t)vd->vdev_guid);
 535         }
 536
 537         return (cp);
 538 }
 539
 540 static struct g_consumer *
 541 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 542 {
 543         struct g_provider *pp;
 544         struct g_consumer *cp;
 545         uint64_t guid;
 546
 547         g_topology_assert();
 548
 549         cp = NULL;
 550         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 551         if (pp != NULL) {
 552                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 553                 cp = vdev_geom_attach(pp);
 554                 if (cp != NULL && check_guid && ISP2(pp->sectorsize) &&
 555                     pp->sectorsize <= VDEV_PAD_SIZE) {
 556                         g_topology_unlock();
 557                         guid = vdev_geom_read_guid(cp);
 558                         g_topology_lock();
 559                         if (guid != vd->vdev_guid) {
 560                                 vdev_geom_detach(cp, 0);
 561                                 cp = NULL;
 562                                 ZFS_LOG(1, "guid mismatch for provider %s: "
 563                                     "%ju != %ju.", vd->vdev_path,
 564                                     (uintmax_t)vd->vdev_guid, (uintmax_t)guid);
 565                         } else {
 566                                 ZFS_LOG(1, "guid match for provider %s.",
 567                                     vd->vdev_path);
 568                         }
 569                 }
 570         }
 571
 572         return (cp);
 573 }
 574
 575 static int
 576 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 577     uint64_t *ashift)
 578 {
 579         struct g_provider *pp;
 580         struct g_consumer *cp;
 581         size_t bufsize;
 582         int error;
 583
 584         /*
 585          * We must have a pathname, and it must be absolute.
 586          */
 587         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 588                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 589                 return (EINVAL);
 590         }
 591
 592         vd->vdev_tsd = NULL;
 593
 594         DROP_GIANT();
 595         g_topology_lock();
 596         error = 0;
 597
 598         /*
 599          * If we're creating or splitting a pool, just find the GEOM provider
 600          * by its name and ignore GUID mismatches.
 601          */
 602         if (vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 603             vd->vdev_spa->spa_splitting_newspa == B_TRUE)
 604                 cp = vdev_geom_open_by_path(vd, 0);
 605         else {
 606                 cp = vdev_geom_open_by_path(vd, 1);
 607                 if (cp == NULL) {
 608                         /*
 609                          * The device at vd->vdev_path doesn't have the
 610                          * expected guid. The disks might have merely
 611                          * moved around so try all other GEOM providers
 612                          * to find one with the right guid.
 613                          */
 614                         cp = vdev_geom_open_by_guid(vd);
 615                 }
 616         }
 617
 618         if (cp == NULL) {
 619                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 620                 error = ENOENT;
 621         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 622             !ISP2(cp->provider->sectorsize)) {
 623                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 624                     vd->vdev_path);
 625                 vdev_geom_detach(cp, 0);
 626                 error = EINVAL;
 627                 cp = NULL;
 628         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 629                 int i;
 630
 631                 for (i = 0; i < 5; i++) {
 632                         error = g_access(cp, 0, 1, 0);
 633                         if (error == 0)
 634                                 break;
 635                         g_topology_unlock();
 636                         tsleep(vd, 0, "vdev", hz / 2);
 637                         g_topology_lock();
 638                 }
 639                 if (error != 0) {
 640                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 641                             vd->vdev_path, error);
 642                         vdev_geom_detach(cp, 0);
 643                         cp = NULL;
 644                 }
 645         }
 646         g_topology_unlock();
 647         PICKUP_GIANT();
 648         if (cp == NULL) {
 649                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 650                 return (error);
 651         }
 652
 653         cp->private = vd;
 654         vd->vdev_tsd = cp;
 655         pp = cp->provider;
 656
 657         /*
 658          * Determine the actual size of the device.
 659          */
 660         *max_psize = *psize = pp->mediasize;
 661
 662         /*
 663          * Determine the device's minimum transfer size.
 664          */
 665         *ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 666
 667         /*
 668          * Clear the nowritecache settings, so that on a vdev_reopen()
 669          * we will try again.
 670          */
 671         vd->vdev_nowritecache = B_FALSE;
 672
 673         if (vd->vdev_physpath != NULL)
 674                 spa_strfree(vd->vdev_physpath);
 675         bufsize = sizeof("/dev/") + strlen(pp->name);
 676         vd->vdev_physpath = kmem_alloc(bufsize, KM_SLEEP);
 677         snprintf(vd->vdev_physpath, bufsize, "/dev/%s", pp->name);
 678
 679         return (0);
 680 }
 681
 682 static void
 683 vdev_geom_close(vdev_t *vd)
 684 {
 685         struct g_consumer *cp;
 686
 687         cp = vd->vdev_tsd;
 688         if (cp == NULL)
 689                 return;
 690         vd->vdev_tsd = NULL;
 691         vd->vdev_delayed_close = B_FALSE;
 692         g_post_event(vdev_geom_detach, cp, M_WAITOK, NULL);
 693 }
 694
 695 static void
 696 vdev_geom_io_intr(struct bio *bp)
 697 {
 698         vdev_t *vd;
 699         zio_t *zio;
 700
 701         zio = bp->bio_caller1;
 702         vd = zio->io_vd;
 703         zio->io_error = bp->bio_error;
 704         if (zio->io_error == 0 && bp->bio_resid != 0)
 705                 zio->io_error = EIO;
 706         if (bp->bio_cmd == BIO_FLUSH && bp->bio_error == ENOTSUP) {
 707                 /*
 708                  * If we get ENOTSUP, we know that no future
 709                  * attempts will ever succeed.  In this case we
 710                  * set a persistent bit so that we don't bother
 711                  * with the ioctl in the future.
 712                  */
 713                 vd->vdev_nowritecache = B_TRUE;
 714         }
 715         if (bp->bio_cmd == BIO_DELETE && bp->bio_error == ENOTSUP) {
 716                 /*
 717                  * If we get ENOTSUP, we know that no future
 718                  * attempts will ever succeed.  In this case we
 719                  * set a persistent bit so that we don't bother
 720                  * with the ioctl in the future.
 721                  */
 722                 vd->vdev_notrim = B_TRUE;
 723         }
 724         if (zio->io_error == EIO && !vd->vdev_remove_wanted) {
 725                 /*
 726                  * If provider's error is set we assume it is being
 727                  * removed.
 728                  */
 729                 if (bp->bio_to->error != 0) {
 730                         /*
 731                          * We post the resource as soon as possible, instead of
 732                          * when the async removal actually happens, because the
 733                          * DE is using this information to discard previous I/O
 734                          * errors.
 735                          */
 736                         /* XXX: zfs_post_remove() can sleep. */
 737                         zfs_post_remove(zio->io_spa, vd);
 738                         vd->vdev_remove_wanted = B_TRUE;
 739                         spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
 740                 } else if (!vd->vdev_delayed_close) {
 741                         vd->vdev_delayed_close = B_TRUE;
 742                 }
 743         }
 744         g_destroy_bio(bp);
 745         zio_interrupt(zio);
 746 }
 747
 748 static int
 749 vdev_geom_io_start(zio_t *zio)
 750 {
 751         vdev_t *vd;
 752         struct g_consumer *cp;
 753         struct bio *bp;
 754         int error;
 755
 756         vd = zio->io_vd;
 757
 758         if (zio->io_type == ZIO_TYPE_IOCTL) {
 759                 /* XXPOLICY */
 760                 if (!vdev_readable(vd)) {
 761                         zio->io_error = ENXIO;
 762                         return (ZIO_PIPELINE_CONTINUE);
 763                 }
 764
 765                 switch (zio->io_cmd) {
 766                 case DKIOCFLUSHWRITECACHE:
 767                         if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 768                                 break;
 769                         if (vd->vdev_nowritecache) {
 770                                 zio->io_error = ENOTSUP;
 771                                 break;
 772                         }
 773                         goto sendreq;
 774                 case DKIOCTRIM:
 775                         if (vdev_geom_bio_delete_disable)
 776                                 break;
 777                         if (vd->vdev_notrim) {
 778                                 zio->io_error = ENOTSUP;
 779                                 break;
 780                         }
 781                         goto sendreq;
 782                 default:
 783                         zio->io_error = ENOTSUP;
 784                 }
 785
 786                 return (ZIO_PIPELINE_CONTINUE);
 787         }
 788 sendreq:
 789         cp = vd->vdev_tsd;
 790         if (cp == NULL) {
 791                 zio->io_error = ENXIO;
 792                 return (ZIO_PIPELINE_CONTINUE);
 793         }
 794         bp = g_alloc_bio();
 795         bp->bio_caller1 = zio;
 796         switch (zio->io_type) {
 797         case ZIO_TYPE_READ:
 798         case ZIO_TYPE_WRITE:
 799                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 800                 bp->bio_data = zio->io_data;
 801                 bp->bio_offset = zio->io_offset;
 802                 bp->bio_length = zio->io_size;
 803                 break;
 804         case ZIO_TYPE_IOCTL:
 805                 switch (zio->io_cmd) {
 806                 case DKIOCFLUSHWRITECACHE:
 807                         bp->bio_cmd = BIO_FLUSH;
 808                         bp->bio_flags |= BIO_ORDERED;
 809                         bp->bio_data = NULL;
 810                         bp->bio_offset = cp->provider->mediasize;
 811                         bp->bio_length = 0;
 812                         break;
 813                 case DKIOCTRIM:
 814                         bp->bio_cmd = BIO_DELETE;
 815                         bp->bio_data = NULL;
 816                         bp->bio_offset = zio->io_offset;
 817                         bp->bio_length = zio->io_size;
 818                         break;
 819                 }
 820                 break;
 821         }
 822         bp->bio_done = vdev_geom_io_intr;
 823
 824         g_io_request(bp, cp);
 825
 826         return (ZIO_PIPELINE_STOP);
 827 }
 828
 829 static void
 830 vdev_geom_io_done(zio_t *zio)
 831 {
 832 }
 833
 834 static void
 835 vdev_geom_hold(vdev_t *vd)
 836 {
 837 }
 838
 839 static void
 840 vdev_geom_rele(vdev_t *vd)
 841 {
 842 }
 843
 844 vdev_ops_t vdev_geom_ops = {
 845         vdev_geom_open,
 846         vdev_geom_close,
 847         vdev_default_asize,
 848         vdev_geom_io_start,
 849         vdev_geom_io_done,
 850         NULL,
 851         vdev_geom_hold,
 852         vdev_geom_rele,
 853         VDEV_TYPE_DISK,         /* name of this vdev type */
 854         B_TRUE                  /* leaf vdev */
 855 };