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_DESTROYED ||
 274                     state > POOL_STATE_L2CACHE) {
 275                         nvlist_free(*config);
 276                         *config = NULL;
 277                         continue;
 278                 }
 279
 280                 if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
 281                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 282                     &txg) != 0 || txg == 0)) {
 283                         nvlist_free(*config);
 284                         *config = NULL;
 285                         continue;
 286                 }
 287
 288                 break;
 289         }
 290
 291         kmem_free(label, size);
 292         return (*config == NULL ? ENOENT : 0);
 293 }
 294
 295 static void
 296 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 297 {
 298         nvlist_t **new_configs;
 299         uint64_t i;
 300
 301         if (id < *count)
 302                 return;
 303         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 304             KM_SLEEP);
 305         for (i = 0; i < *count; i++)
 306                 new_configs[i] = (*configs)[i];
 307         if (*configs != NULL)
 308                 kmem_free(*configs, *count * sizeof(void *));
 309         *configs = new_configs;
 310         *count = id + 1;
 311 }
 312
 313 static void
 314 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 315     const char *name, uint64_t* known_pool_guid)
 316 {
 317         nvlist_t *vdev_tree;
 318         uint64_t pool_guid;
 319         uint64_t vdev_guid, known_guid;
 320         uint64_t id, txg, known_txg;
 321         char *pname;
 322         int i;
 323
 324         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 325             strcmp(pname, name) != 0)
 326                 goto ignore;
 327
 328         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 329                 goto ignore;
 330
 331         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 332                 goto ignore;
 333
 334         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 335                 goto ignore;
 336
 337         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 338                 goto ignore;
 339
 340         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 341
 342         if (*known_pool_guid != 0) {
 343                 if (pool_guid != *known_pool_guid)
 344                         goto ignore;
 345         } else
 346                 *known_pool_guid = pool_guid;
 347
 348         resize_configs(configs, count, id);
 349
 350         if ((*configs)[id] != NULL) {
 351                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 352                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 353                 if (txg <= known_txg)
 354                         goto ignore;
 355                 nvlist_free((*configs)[id]);
 356         }
 357
 358         (*configs)[id] = cfg;
 359         return;
 360
 361 ignore:
 362         nvlist_free(cfg);
 363 }
 364
 365 static int
 366 vdev_geom_attach_taster(struct g_consumer *cp, struct g_provider *pp)
 367 {
 368         int error;
 369
 370         if (pp->flags & G_PF_WITHER)
 371                 return (EINVAL);
 372         if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize))
 373                 return (EINVAL);
 374         g_attach(cp, pp);
 375         error = g_access(cp, 1, 0, 0);
 376         if (error != 0)
 377                 g_detach(cp);
 378         return (error);
 379 }
 380
 381 static void
 382 vdev_geom_detach_taster(struct g_consumer *cp)
 383 {
 384         g_access(cp, -1, 0, 0);
 385         g_detach(cp);
 386 }
 387
 388 int
 389 vdev_geom_read_pool_label(const char *name,
 390     nvlist_t ***configs, uint64_t *count)
 391 {
 392         struct g_class *mp;
 393         struct g_geom *gp, *zgp;
 394         struct g_provider *pp;
 395         struct g_consumer *zcp;
 396         nvlist_t *vdev_cfg;
 397         uint64_t pool_guid;
 398         int error;
 399
 400         DROP_GIANT();
 401         g_topology_lock();
 402
 403         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 404         /* This orphan function should be never called. */
 405         zgp->orphan = vdev_geom_taste_orphan;
 406         zcp = g_new_consumer(zgp);
 407
 408         *configs = NULL;
 409         *count = 0;
 410         pool_guid = 0;
 411         LIST_FOREACH(mp, &g_classes, class) {
 412                 if (mp == &zfs_vdev_class)
 413                         continue;
 414                 LIST_FOREACH(gp, &mp->geom, geom) {
 415                         if (gp->flags & G_GEOM_WITHER)
 416                                 continue;
 417                         LIST_FOREACH(pp, &gp->provider, provider) {
 418                                 if (pp->flags & G_PF_WITHER)
 419                                         continue;
 420                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 421                                         continue;
 422                                 g_topology_unlock();
 423                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 424                                 g_topology_lock();
 425                                 vdev_geom_detach_taster(zcp);
 426                                 if (error)
 427                                         continue;
 428                                 ZFS_LOG(1, "successfully read vdev config");
 429
 430                                 process_vdev_config(configs, count,
 431                                     vdev_cfg, name, &pool_guid);
 432                         }
 433                 }
 434         }
 435
 436         g_destroy_consumer(zcp);
 437         g_destroy_geom(zgp);
 438         g_topology_unlock();
 439         PICKUP_GIANT();
 440
 441         return (*count > 0 ? 0 : ENOENT);
 442 }
 443
 444 static uint64_t
 445 vdev_geom_read_guid(struct g_consumer *cp)
 446 {
 447         nvlist_t *config;
 448         uint64_t guid;
 449
 450         g_topology_assert_not();
 451
 452         guid = 0;
 453         if (vdev_geom_read_config(cp, &config) == 0) {
 454                 guid = nvlist_get_guid(config);
 455                 nvlist_free(config);
 456         }
 457         return (guid);
 458 }
 459
 460 static struct g_consumer *
 461 vdev_geom_attach_by_guid(uint64_t guid)
 462 {
 463         struct g_class *mp;
 464         struct g_geom *gp, *zgp;
 465         struct g_provider *pp;
 466         struct g_consumer *cp, *zcp;
 467         uint64_t pguid;
 468
 469         g_topology_assert();
 470
 471         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 472         /* This orphan function should be never called. */
 473         zgp->orphan = vdev_geom_taste_orphan;
 474         zcp = g_new_consumer(zgp);
 475
 476         cp = NULL;
 477         LIST_FOREACH(mp, &g_classes, class) {
 478                 if (mp == &zfs_vdev_class)
 479                         continue;
 480                 LIST_FOREACH(gp, &mp->geom, geom) {
 481                         if (gp->flags & G_GEOM_WITHER)
 482                                 continue;
 483                         LIST_FOREACH(pp, &gp->provider, provider) {
 484                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 485                                         continue;
 486                                 g_topology_unlock();
 487                                 pguid = vdev_geom_read_guid(zcp);
 488                                 g_topology_lock();
 489                                 vdev_geom_detach_taster(zcp);
 490                                 if (pguid != guid)
 491                                         continue;
 492                                 cp = vdev_geom_attach(pp);
 493                                 if (cp == NULL) {
 494                                         printf("ZFS WARNING: Unable to attach to %s.\n",
 495                                             pp->name);
 496                                         continue;
 497                                 }
 498                                 break;
 499                         }
 500                         if (cp != NULL)
 501                                 break;
 502                 }
 503                 if (cp != NULL)
 504                         break;
 505         }
 506 end:
 507         g_destroy_consumer(zcp);
 508         g_destroy_geom(zgp);
 509         return (cp);
 510 }
 511
 512 static struct g_consumer *
 513 vdev_geom_open_by_guid(vdev_t *vd)
 514 {
 515         struct g_consumer *cp;
 516         char *buf;
 517         size_t len;
 518
 519         g_topology_assert();
 520
 521         ZFS_LOG(1, "Searching by guid [%ju].", (uintmax_t)vd->vdev_guid);
 522         cp = vdev_geom_attach_by_guid(vd->vdev_guid);
 523         if (cp != NULL) {
 524                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 525                 buf = kmem_alloc(len, KM_SLEEP);
 526
 527                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 528                 spa_strfree(vd->vdev_path);
 529                 vd->vdev_path = buf;
 530
 531                 ZFS_LOG(1, "Attach by guid [%ju] succeeded, provider %s.",
 532                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 533         } else {
 534                 ZFS_LOG(1, "Search by guid [%ju] failed.",
 535                     (uintmax_t)vd->vdev_guid);
 536         }
 537
 538         return (cp);
 539 }
 540
 541 static struct g_consumer *
 542 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 543 {
 544         struct g_provider *pp;
 545         struct g_consumer *cp;
 546         uint64_t guid;
 547
 548         g_topology_assert();
 549
 550         cp = NULL;
 551         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 552         if (pp != NULL) {
 553                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 554                 cp = vdev_geom_attach(pp);
 555                 if (cp != NULL && check_guid && ISP2(pp->sectorsize) &&
 556                     pp->sectorsize <= VDEV_PAD_SIZE) {
 557                         g_topology_unlock();
 558                         guid = vdev_geom_read_guid(cp);
 559                         g_topology_lock();
 560                         if (guid != vd->vdev_guid) {
 561                                 vdev_geom_detach(cp, 0);
 562                                 cp = NULL;
 563                                 ZFS_LOG(1, "guid mismatch for provider %s: "
 564                                     "%ju != %ju.", vd->vdev_path,
 565                                     (uintmax_t)vd->vdev_guid, (uintmax_t)guid);
 566                         } else {
 567                                 ZFS_LOG(1, "guid match for provider %s.",
 568                                     vd->vdev_path);
 569                         }
 570                 }
 571         }
 572
 573         return (cp);
 574 }
 575
 576 static int
 577 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 578     uint64_t *ashift)
 579 {
 580         struct g_provider *pp;
 581         struct g_consumer *cp;
 582         size_t bufsize;
 583         int error;
 584
 585         /*
 586          * We must have a pathname, and it must be absolute.
 587          */
 588         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 589                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 590                 return (EINVAL);
 591         }
 592
 593         vd->vdev_tsd = NULL;
 594
 595         DROP_GIANT();
 596         g_topology_lock();
 597         error = 0;
 598
 599         /*
 600          * If we're creating or splitting a pool, just find the GEOM provider
 601          * by its name and ignore GUID mismatches.
 602          */
 603         if (vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 604             vd->vdev_spa->spa_splitting_newspa == B_TRUE)
 605                 cp = vdev_geom_open_by_path(vd, 0);
 606         else {
 607                 cp = vdev_geom_open_by_path(vd, 1);
 608                 if (cp == NULL) {
 609                         /*
 610                          * The device at vd->vdev_path doesn't have the
 611                          * expected guid. The disks might have merely
 612                          * moved around so try all other GEOM providers
 613                          * to find one with the right guid.
 614                          */
 615                         cp = vdev_geom_open_by_guid(vd);
 616                 }
 617         }
 618
 619         if (cp == NULL) {
 620                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 621                 error = ENOENT;
 622         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 623             !ISP2(cp->provider->sectorsize)) {
 624                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 625                     vd->vdev_path);
 626                 vdev_geom_detach(cp, 0);
 627                 error = EINVAL;
 628                 cp = NULL;
 629         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 630                 int i;
 631
 632                 for (i = 0; i < 5; i++) {
 633                         error = g_access(cp, 0, 1, 0);
 634                         if (error == 0)
 635                                 break;
 636                         g_topology_unlock();
 637                         tsleep(vd, 0, "vdev", hz / 2);
 638                         g_topology_lock();
 639                 }
 640                 if (error != 0) {
 641                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 642                             vd->vdev_path, error);
 643                         vdev_geom_detach(cp, 0);
 644                         cp = NULL;
 645                 }
 646         }
 647         g_topology_unlock();
 648         PICKUP_GIANT();
 649         if (cp == NULL) {
 650                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 651                 return (error);
 652         }
 653
 654         cp->private = vd;
 655         vd->vdev_tsd = cp;
 656         pp = cp->provider;
 657
 658         /*
 659          * Determine the actual size of the device.
 660          */
 661         *max_psize = *psize = pp->mediasize;
 662
 663         /*
 664          * Determine the device's minimum transfer size.
 665          */
 666         *ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 667
 668         /*
 669          * Clear the nowritecache settings, so that on a vdev_reopen()
 670          * we will try again.
 671          */
 672         vd->vdev_nowritecache = B_FALSE;
 673
 674         if (vd->vdev_physpath != NULL)
 675                 spa_strfree(vd->vdev_physpath);
 676         bufsize = sizeof("/dev/") + strlen(pp->name);
 677         vd->vdev_physpath = kmem_alloc(bufsize, KM_SLEEP);
 678         snprintf(vd->vdev_physpath, bufsize, "/dev/%s", pp->name);
 679
 680         return (0);
 681 }
 682
 683 static void
 684 vdev_geom_close(vdev_t *vd)
 685 {
 686         struct g_consumer *cp;
 687
 688         cp = vd->vdev_tsd;
 689         if (cp == NULL)
 690                 return;
 691         vd->vdev_tsd = NULL;
 692         vd->vdev_delayed_close = B_FALSE;
 693         g_post_event(vdev_geom_detach, cp, M_WAITOK, NULL);
 694 }
 695
 696 static void
 697 vdev_geom_io_intr(struct bio *bp)
 698 {
 699         vdev_t *vd;
 700         zio_t *zio;
 701
 702         zio = bp->bio_caller1;
 703         vd = zio->io_vd;
 704         zio->io_error = bp->bio_error;
 705         if (zio->io_error == 0 && bp->bio_resid != 0)
 706                 zio->io_error = EIO;
 707         if (bp->bio_cmd == BIO_FLUSH && bp->bio_error == ENOTSUP) {
 708                 /*
 709                  * If we get ENOTSUP, we know that no future
 710                  * attempts will ever succeed.  In this case we
 711                  * set a persistent bit so that we don't bother
 712                  * with the ioctl in the future.
 713                  */
 714                 vd->vdev_nowritecache = B_TRUE;
 715         }
 716         if (bp->bio_cmd == BIO_DELETE && bp->bio_error == ENOTSUP) {
 717                 /*
 718                  * If we get ENOTSUP, we know that no future
 719                  * attempts will ever succeed.  In this case we
 720                  * set a persistent bit so that we don't bother
 721                  * with the ioctl in the future.
 722                  */
 723                 vd->vdev_notrim = B_TRUE;
 724         }
 725         if (zio->io_error == EIO && !vd->vdev_remove_wanted) {
 726                 /*
 727                  * If provider's error is set we assume it is being
 728                  * removed.
 729                  */
 730                 if (bp->bio_to->error != 0) {
 731                         /*
 732                          * We post the resource as soon as possible, instead of
 733                          * when the async removal actually happens, because the
 734                          * DE is using this information to discard previous I/O
 735                          * errors.
 736                          */
 737                         /* XXX: zfs_post_remove() can sleep. */
 738                         zfs_post_remove(zio->io_spa, vd);
 739                         vd->vdev_remove_wanted = B_TRUE;
 740                         spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
 741                 } else if (!vd->vdev_delayed_close) {
 742                         vd->vdev_delayed_close = B_TRUE;
 743                 }
 744         }
 745         g_destroy_bio(bp);
 746         zio_interrupt(zio);
 747 }
 748
 749 static int
 750 vdev_geom_io_start(zio_t *zio)
 751 {
 752         vdev_t *vd;
 753         struct g_consumer *cp;
 754         struct bio *bp;
 755         int error;
 756
 757         vd = zio->io_vd;
 758
 759         if (zio->io_type == ZIO_TYPE_IOCTL) {
 760                 /* XXPOLICY */
 761                 if (!vdev_readable(vd)) {
 762                         zio->io_error = ENXIO;
 763                         return (ZIO_PIPELINE_CONTINUE);
 764                 }
 765
 766                 switch (zio->io_cmd) {
 767                 case DKIOCFLUSHWRITECACHE:
 768                         if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 769                                 break;
 770                         if (vd->vdev_nowritecache) {
 771                                 zio->io_error = ENOTSUP;
 772                                 break;
 773                         }
 774                         goto sendreq;
 775                 case DKIOCTRIM:
 776                         if (vdev_geom_bio_delete_disable)
 777                                 break;
 778                         if (vd->vdev_notrim) {
 779                                 zio->io_error = ENOTSUP;
 780                                 break;
 781                         }
 782                         goto sendreq;
 783                 default:
 784                         zio->io_error = ENOTSUP;
 785                 }
 786
 787                 return (ZIO_PIPELINE_CONTINUE);
 788         }
 789 sendreq:
 790         cp = vd->vdev_tsd;
 791         if (cp == NULL) {
 792                 zio->io_error = ENXIO;
 793                 return (ZIO_PIPELINE_CONTINUE);
 794         }
 795         bp = g_alloc_bio();
 796         bp->bio_caller1 = zio;
 797         switch (zio->io_type) {
 798         case ZIO_TYPE_READ:
 799         case ZIO_TYPE_WRITE:
 800                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 801                 bp->bio_data = zio->io_data;
 802                 bp->bio_offset = zio->io_offset;
 803                 bp->bio_length = zio->io_size;
 804                 break;
 805         case ZIO_TYPE_IOCTL:
 806                 switch (zio->io_cmd) {
 807                 case DKIOCFLUSHWRITECACHE:
 808                         bp->bio_cmd = BIO_FLUSH;
 809                         bp->bio_flags |= BIO_ORDERED;
 810                         bp->bio_data = NULL;
 811                         bp->bio_offset = cp->provider->mediasize;
 812                         bp->bio_length = 0;
 813                         break;
 814                 case DKIOCTRIM:
 815                         bp->bio_cmd = BIO_DELETE;
 816                         bp->bio_data = NULL;
 817                         bp->bio_offset = zio->io_offset;
 818                         bp->bio_length = zio->io_size;
 819                         break;
 820                 }
 821                 break;
 822         }
 823         bp->bio_done = vdev_geom_io_intr;
 824
 825         g_io_request(bp, cp);
 826
 827         return (ZIO_PIPELINE_STOP);
 828 }
 829
 830 static void
 831 vdev_geom_io_done(zio_t *zio)
 832 {
 833 }
 834
 835 static void
 836 vdev_geom_hold(vdev_t *vd)
 837 {
 838 }
 839
 840 static void
 841 vdev_geom_rele(vdev_t *vd)
 842 {
 843 }
 844
 845 vdev_ops_t vdev_geom_ops = {
 846         vdev_geom_open,
 847         vdev_geom_close,
 848         vdev_default_asize,
 849         vdev_geom_io_start,
 850         vdev_geom_io_done,
 851         NULL,
 852         vdev_geom_hold,
 853         vdev_geom_rele,
 854         VDEV_TYPE_DISK,         /* name of this vdev type */
 855         B_TRUE                  /* leaf vdev */
 856 };