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         if (vd == NULL)
  73                 return;
  74
  75         /*
  76          * Orphan callbacks occur from the GEOM event thread.
  77          * Concurrent with this call, new I/O requests may be
  78          * working their way through GEOM about to find out
  79          * (only once executed by the g_down thread) that we've
  80          * been orphaned from our disk provider.  These I/Os
  81          * must be retired before we can detach our consumer.
  82          * This is most easily achieved by acquiring the
  83          * SPA ZIO configuration lock as a writer, but doing
  84          * so with the GEOM topology lock held would cause
  85          * a lock order reversal.  Instead, rely on the SPA's
  86          * async removal support to invoke a close on this
  87          * vdev once it is safe to do so.
  88          */
  89         zfs_post_remove(vd->vdev_spa, vd);
  90         vd->vdev_remove_wanted = B_TRUE;
  91         spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
  92 }
  93
  94 static struct g_consumer *
  95 vdev_geom_attach(struct g_provider *pp)
  96 {
  97         struct g_geom *gp;
  98         struct g_consumer *cp;
  99
 100         g_topology_assert();
 101
 102         ZFS_LOG(1, "Attaching to %s.", pp->name);
 103         /* Do we have geom already? No? Create one. */
 104         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 105                 if (gp->flags & G_GEOM_WITHER)
 106                         continue;
 107                 if (strcmp(gp->name, "zfs::vdev") != 0)
 108                         continue;
 109                 break;
 110         }
 111         if (gp == NULL) {
 112                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 113                 gp->orphan = vdev_geom_orphan;
 114                 cp = g_new_consumer(gp);
 115                 if (g_attach(cp, pp) != 0) {
 116                         g_wither_geom(gp, ENXIO);
 117                         return (NULL);
 118                 }
 119                 if (g_access(cp, 1, 0, 1) != 0) {
 120                         g_wither_geom(gp, ENXIO);
 121                         return (NULL);
 122                 }
 123                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 124         } else {
 125                 /* Check if we are already connected to this provider. */
 126                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 127                         if (cp->provider == pp) {
 128                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 129                                 break;
 130                         }
 131                 }
 132                 if (cp == NULL) {
 133                         cp = g_new_consumer(gp);
 134                         if (g_attach(cp, pp) != 0) {
 135                                 g_destroy_consumer(cp);
 136                                 return (NULL);
 137                         }
 138                         if (g_access(cp, 1, 0, 1) != 0) {
 139                                 g_detach(cp);
 140                                 g_destroy_consumer(cp);
 141                                 return (NULL);
 142                         }
 143                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 144                 } else {
 145                         if (g_access(cp, 1, 0, 1) != 0)
 146                                 return (NULL);
 147                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 148                 }
 149         }
 150         return (cp);
 151 }
 152
 153 static void
 154 vdev_geom_detach(void *arg, int flag __unused)
 155 {
 156         struct g_geom *gp;
 157         struct g_consumer *cp;
 158
 159         g_topology_assert();
 160         cp = arg;
 161         gp = cp->geom;
 162
 163         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 164         g_access(cp, -1, 0, -1);
 165         /* Destroy consumer on last close. */
 166         if (cp->acr == 0 && cp->ace == 0) {
 167                 ZFS_LOG(1, "Destroyed consumer to %s.", cp->provider->name);
 168                 if (cp->acw > 0)
 169                         g_access(cp, 0, -cp->acw, 0);
 170                 g_detach(cp);
 171                 g_destroy_consumer(cp);
 172         }
 173         /* Destroy geom if there are no consumers left. */
 174         if (LIST_EMPTY(&gp->consumer)) {
 175                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 176                 g_wither_geom(gp, ENXIO);
 177         }
 178 }
 179
 180 static uint64_t
 181 nvlist_get_guid(nvlist_t *list)
 182 {
 183         uint64_t value;
 184
 185         value = 0;
 186         nvlist_lookup_uint64(list, ZPOOL_CONFIG_GUID, &value);
 187         return (value);
 188 }
 189
 190 static int
 191 vdev_geom_io(struct g_consumer *cp, int cmd, void *data, off_t offset, off_t size)
 192 {
 193         struct bio *bp;
 194         u_char *p;
 195         off_t off, maxio;
 196         int error;
 197
 198         ASSERT((offset % cp->provider->sectorsize) == 0);
 199         ASSERT((size % cp->provider->sectorsize) == 0);
 200
 201         bp = g_alloc_bio();
 202         off = offset;
 203         offset += size;
 204         p = data;
 205         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 206         error = 0;
 207
 208         for (; off < offset; off += maxio, p += maxio, size -= maxio) {
 209                 bzero(bp, sizeof(*bp));
 210                 bp->bio_cmd = cmd;
 211                 bp->bio_done = NULL;
 212                 bp->bio_offset = off;
 213                 bp->bio_length = MIN(size, maxio);
 214                 bp->bio_data = p;
 215                 g_io_request(bp, cp);
 216                 error = biowait(bp, "vdev_geom_io");
 217                 if (error != 0)
 218                         break;
 219         }
 220
 221         g_destroy_bio(bp);
 222         return (error);
 223 }
 224
 225 static void
 226 vdev_geom_taste_orphan(struct g_consumer *cp)
 227 {
 228
 229         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
 230             cp->provider->name));
 231 }
 232
 233 static int
 234 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 235 {
 236         struct g_provider *pp;
 237         vdev_label_t *label;
 238         char *p, *buf;
 239         size_t buflen;
 240         uint64_t psize;
 241         off_t offset, size;
 242         uint64_t guid, state, txg;
 243         int error, l, len;
 244
 245         g_topology_assert_not();
 246
 247         pp = cp->provider;
 248         ZFS_LOG(1, "Reading config from %s...", pp->name);
 249
 250         psize = pp->mediasize;
 251         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 252
 253         size = sizeof(*label) + pp->sectorsize -
 254             ((sizeof(*label) - 1) % pp->sectorsize) - 1;
 255
 256         guid = 0;
 257         label = kmem_alloc(size, KM_SLEEP);
 258         buflen = sizeof(label->vl_vdev_phys.vp_nvlist);
 259
 260         *config = NULL;
 261         for (l = 0; l < VDEV_LABELS; l++) {
 262
 263                 offset = vdev_label_offset(psize, l, 0);
 264                 if ((offset % pp->sectorsize) != 0)
 265                         continue;
 266
 267                 if (vdev_geom_io(cp, BIO_READ, label, offset, size) != 0)
 268                         continue;
 269                 buf = label->vl_vdev_phys.vp_nvlist;
 270
 271                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 272                         continue;
 273
 274                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 275                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 276                         nvlist_free(*config);
 277                         *config = NULL;
 278                         continue;
 279                 }
 280
 281                 if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
 282                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 283                     &txg) != 0 || txg == 0)) {
 284                         nvlist_free(*config);
 285                         *config = NULL;
 286                         continue;
 287                 }
 288
 289                 break;
 290         }
 291
 292         kmem_free(label, size);
 293         return (*config == NULL ? ENOENT : 0);
 294 }
 295
 296 static void
 297 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 298 {
 299         nvlist_t **new_configs;
 300         uint64_t i;
 301
 302         if (id < *count)
 303                 return;
 304         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 305             KM_SLEEP);
 306         for (i = 0; i < *count; i++)
 307                 new_configs[i] = (*configs)[i];
 308         if (*configs != NULL)
 309                 kmem_free(*configs, *count * sizeof(void *));
 310         *configs = new_configs;
 311         *count = id + 1;
 312 }
 313
 314 static void
 315 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 316     const char *name, uint64_t* known_pool_guid)
 317 {
 318         nvlist_t *vdev_tree;
 319         uint64_t pool_guid;
 320         uint64_t vdev_guid, known_guid;
 321         uint64_t id, txg, known_txg;
 322         char *pname;
 323         int i;
 324
 325         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 326             strcmp(pname, name) != 0)
 327                 goto ignore;
 328
 329         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 330                 goto ignore;
 331
 332         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 333                 goto ignore;
 334
 335         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 336                 goto ignore;
 337
 338         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 339                 goto ignore;
 340
 341         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 342
 343         if (*known_pool_guid != 0) {
 344                 if (pool_guid != *known_pool_guid)
 345                         goto ignore;
 346         } else
 347                 *known_pool_guid = pool_guid;
 348
 349         resize_configs(configs, count, id);
 350
 351         if ((*configs)[id] != NULL) {
 352                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 353                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 354                 if (txg <= known_txg)
 355                         goto ignore;
 356                 nvlist_free((*configs)[id]);
 357         }
 358
 359         (*configs)[id] = cfg;
 360         return;
 361
 362 ignore:
 363         nvlist_free(cfg);
 364 }
 365
 366 static int
 367 vdev_geom_attach_taster(struct g_consumer *cp, struct g_provider *pp)
 368 {
 369         int error;
 370
 371         if (pp->flags & G_PF_WITHER)
 372                 return (EINVAL);
 373         g_attach(cp, pp);
 374         error = g_access(cp, 1, 0, 0);
 375         if (error == 0) {
 376                 if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize))
 377                         error = EINVAL;
 378                 else if (pp->mediasize < SPA_MINDEVSIZE)
 379                         error = EINVAL;
 380                 if (error != 0)
 381                         g_access(cp, -1, 0, 0);
 382         }
 383         if (error != 0)
 384                 g_detach(cp);
 385         return (error);
 386 }
 387
 388 static void
 389 vdev_geom_detach_taster(struct g_consumer *cp)
 390 {
 391         g_access(cp, -1, 0, 0);
 392         g_detach(cp);
 393 }
 394
 395 int
 396 vdev_geom_read_pool_label(const char *name,
 397     nvlist_t ***configs, uint64_t *count)
 398 {
 399         struct g_class *mp;
 400         struct g_geom *gp, *zgp;
 401         struct g_provider *pp;
 402         struct g_consumer *zcp;
 403         nvlist_t *vdev_cfg;
 404         uint64_t pool_guid;
 405         int error;
 406
 407         DROP_GIANT();
 408         g_topology_lock();
 409
 410         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 411         /* This orphan function should be never called. */
 412         zgp->orphan = vdev_geom_taste_orphan;
 413         zcp = g_new_consumer(zgp);
 414
 415         *configs = NULL;
 416         *count = 0;
 417         pool_guid = 0;
 418         LIST_FOREACH(mp, &g_classes, class) {
 419                 if (mp == &zfs_vdev_class)
 420                         continue;
 421                 LIST_FOREACH(gp, &mp->geom, geom) {
 422                         if (gp->flags & G_GEOM_WITHER)
 423                                 continue;
 424                         LIST_FOREACH(pp, &gp->provider, provider) {
 425                                 if (pp->flags & G_PF_WITHER)
 426                                         continue;
 427                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 428                                         continue;
 429                                 g_topology_unlock();
 430                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 431                                 g_topology_lock();
 432                                 vdev_geom_detach_taster(zcp);
 433                                 if (error)
 434                                         continue;
 435                                 ZFS_LOG(1, "successfully read vdev config");
 436
 437                                 process_vdev_config(configs, count,
 438                                     vdev_cfg, name, &pool_guid);
 439                         }
 440                 }
 441         }
 442
 443         g_destroy_consumer(zcp);
 444         g_destroy_geom(zgp);
 445         g_topology_unlock();
 446         PICKUP_GIANT();
 447
 448         return (*count > 0 ? 0 : ENOENT);
 449 }
 450
 451 static uint64_t
 452 vdev_geom_read_guid(struct g_consumer *cp)
 453 {
 454         nvlist_t *config;
 455         uint64_t guid;
 456
 457         g_topology_assert_not();
 458
 459         guid = 0;
 460         if (vdev_geom_read_config(cp, &config) == 0) {
 461                 guid = nvlist_get_guid(config);
 462                 nvlist_free(config);
 463         }
 464         return (guid);
 465 }
 466
 467 static struct g_consumer *
 468 vdev_geom_attach_by_guid(uint64_t guid)
 469 {
 470         struct g_class *mp;
 471         struct g_geom *gp, *zgp;
 472         struct g_provider *pp;
 473         struct g_consumer *cp, *zcp;
 474         uint64_t pguid;
 475
 476         g_topology_assert();
 477
 478         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 479         /* This orphan function should be never called. */
 480         zgp->orphan = vdev_geom_taste_orphan;
 481         zcp = g_new_consumer(zgp);
 482
 483         cp = NULL;
 484         LIST_FOREACH(mp, &g_classes, class) {
 485                 if (mp == &zfs_vdev_class)
 486                         continue;
 487                 LIST_FOREACH(gp, &mp->geom, geom) {
 488                         if (gp->flags & G_GEOM_WITHER)
 489                                 continue;
 490                         LIST_FOREACH(pp, &gp->provider, provider) {
 491                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 492                                         continue;
 493                                 g_topology_unlock();
 494                                 pguid = vdev_geom_read_guid(zcp);
 495                                 g_topology_lock();
 496                                 vdev_geom_detach_taster(zcp);
 497                                 if (pguid != guid)
 498                                         continue;
 499                                 cp = vdev_geom_attach(pp);
 500                                 if (cp == NULL) {
 501                                         printf("ZFS WARNING: Unable to attach to %s.\n",
 502                                             pp->name);
 503                                         continue;
 504                                 }
 505                                 break;
 506                         }
 507                         if (cp != NULL)
 508                                 break;
 509                 }
 510                 if (cp != NULL)
 511                         break;
 512         }
 513 end:
 514         g_destroy_consumer(zcp);
 515         g_destroy_geom(zgp);
 516         return (cp);
 517 }
 518
 519 static struct g_consumer *
 520 vdev_geom_open_by_guid(vdev_t *vd)
 521 {
 522         struct g_consumer *cp;
 523         char *buf;
 524         size_t len;
 525
 526         g_topology_assert();
 527
 528         ZFS_LOG(1, "Searching by guid [%ju].", (uintmax_t)vd->vdev_guid);
 529         cp = vdev_geom_attach_by_guid(vd->vdev_guid);
 530         if (cp != NULL) {
 531                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 532                 buf = kmem_alloc(len, KM_SLEEP);
 533
 534                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 535                 spa_strfree(vd->vdev_path);
 536                 vd->vdev_path = buf;
 537
 538                 ZFS_LOG(1, "Attach by guid [%ju] succeeded, provider %s.",
 539                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 540         } else {
 541                 ZFS_LOG(1, "Search by guid [%ju] failed.",
 542                     (uintmax_t)vd->vdev_guid);
 543         }
 544
 545         return (cp);
 546 }
 547
 548 static struct g_consumer *
 549 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 550 {
 551         struct g_provider *pp;
 552         struct g_consumer *cp;
 553         uint64_t guid;
 554
 555         g_topology_assert();
 556
 557         cp = NULL;
 558         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 559         if (pp != NULL) {
 560                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 561                 cp = vdev_geom_attach(pp);
 562                 if (cp != NULL && check_guid && ISP2(pp->sectorsize) &&
 563                     pp->sectorsize <= VDEV_PAD_SIZE) {
 564                         g_topology_unlock();
 565                         guid = vdev_geom_read_guid(cp);
 566                         g_topology_lock();
 567                         if (guid != vd->vdev_guid) {
 568                                 vdev_geom_detach(cp, 0);
 569                                 cp = NULL;
 570                                 ZFS_LOG(1, "guid mismatch for provider %s: "
 571                                     "%ju != %ju.", vd->vdev_path,
 572                                     (uintmax_t)vd->vdev_guid, (uintmax_t)guid);
 573                         } else {
 574                                 ZFS_LOG(1, "guid match for provider %s.",
 575                                     vd->vdev_path);
 576                         }
 577                 }
 578         }
 579
 580         return (cp);
 581 }
 582
 583 static int
 584 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 585     uint64_t *logical_ashift, uint64_t *physical_ashift)
 586 {
 587         struct g_provider *pp;
 588         struct g_consumer *cp;
 589         size_t bufsize;
 590         int error;
 591
 592         /*
 593          * We must have a pathname, and it must be absolute.
 594          */
 595         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 596                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 597                 return (EINVAL);
 598         }
 599
 600         vd->vdev_tsd = NULL;
 601
 602         DROP_GIANT();
 603         g_topology_lock();
 604         error = 0;
 605
 606         /*
 607          * If we're creating or splitting a pool, just find the GEOM provider
 608          * by its name and ignore GUID mismatches.
 609          */
 610         if (vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 611             vd->vdev_spa->spa_splitting_newspa == B_TRUE)
 612                 cp = vdev_geom_open_by_path(vd, 0);
 613         else {
 614                 cp = vdev_geom_open_by_path(vd, 1);
 615                 if (cp == NULL) {
 616                         /*
 617                          * The device at vd->vdev_path doesn't have the
 618                          * expected guid. The disks might have merely
 619                          * moved around so try all other GEOM providers
 620                          * to find one with the right guid.
 621                          */
 622                         cp = vdev_geom_open_by_guid(vd);
 623                 }
 624         }
 625
 626         if (cp == NULL) {
 627                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 628                 error = ENOENT;
 629         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 630             !ISP2(cp->provider->sectorsize)) {
 631                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 632                     vd->vdev_path);
 633                 vdev_geom_detach(cp, 0);
 634                 error = EINVAL;
 635                 cp = NULL;
 636         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 637                 int i;
 638
 639                 for (i = 0; i < 5; i++) {
 640                         error = g_access(cp, 0, 1, 0);
 641                         if (error == 0)
 642                                 break;
 643                         g_topology_unlock();
 644                         tsleep(vd, 0, "vdev", hz / 2);
 645                         g_topology_lock();
 646                 }
 647                 if (error != 0) {
 648                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 649                             vd->vdev_path, error);
 650                         vdev_geom_detach(cp, 0);
 651                         cp = NULL;
 652                 }
 653         }
 654         g_topology_unlock();
 655         PICKUP_GIANT();
 656         if (cp == NULL) {
 657                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 658                 return (error);
 659         }
 660
 661         cp->private = vd;
 662         vd->vdev_tsd = cp;
 663         pp = cp->provider;
 664
 665         /*
 666          * Determine the actual size of the device.
 667          */
 668         *max_psize = *psize = pp->mediasize;
 669
 670         /*
 671          * Determine the device's minimum transfer size and preferred
 672          * transfer size.
 673          */
 674         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 675         *physical_ashift = 0;
 676         if (pp->stripesize)
 677                 *physical_ashift = highbit(pp->stripesize) - 1;
 678
 679         /*
 680          * Clear the nowritecache settings, so that on a vdev_reopen()
 681          * we will try again.
 682          */
 683         vd->vdev_nowritecache = B_FALSE;
 684
 685         if (vd->vdev_physpath != NULL)
 686                 spa_strfree(vd->vdev_physpath);
 687         bufsize = sizeof("/dev/") + strlen(pp->name);
 688         vd->vdev_physpath = kmem_alloc(bufsize, KM_SLEEP);
 689         snprintf(vd->vdev_physpath, bufsize, "/dev/%s", pp->name);
 690
 691         return (0);
 692 }
 693
 694 static void
 695 vdev_geom_close(vdev_t *vd)
 696 {
 697         struct g_consumer *cp;
 698
 699         cp = vd->vdev_tsd;
 700         if (cp == NULL)
 701                 return;
 702         vd->vdev_tsd = NULL;
 703         vd->vdev_delayed_close = B_FALSE;
 704         cp->private = NULL;     /* XXX locking */
 705         g_post_event(vdev_geom_detach, cp, M_WAITOK, NULL);
 706 }
 707
 708 static void
 709 vdev_geom_io_intr(struct bio *bp)
 710 {
 711         vdev_t *vd;
 712         zio_t *zio;
 713
 714         zio = bp->bio_caller1;
 715         vd = zio->io_vd;
 716         zio->io_error = bp->bio_error;
 717         if (zio->io_error == 0 && bp->bio_resid != 0)
 718                 zio->io_error = EIO;
 719         if (bp->bio_cmd == BIO_FLUSH && bp->bio_error == ENOTSUP) {
 720                 /*
 721                  * If we get ENOTSUP, we know that no future
 722                  * attempts will ever succeed.  In this case we
 723                  * set a persistent bit so that we don't bother
 724                  * with the ioctl in the future.
 725                  */
 726                 vd->vdev_nowritecache = B_TRUE;
 727         }
 728         if (bp->bio_cmd == BIO_DELETE && bp->bio_error == ENOTSUP) {
 729                 /*
 730                  * If we get ENOTSUP, we know that no future
 731                  * attempts will ever succeed.  In this case we
 732                  * set a persistent bit so that we don't bother
 733                  * with the ioctl in the future.
 734                  */
 735                 vd->vdev_notrim = B_TRUE;
 736         }
 737         if (zio->io_error == ENXIO && !vd->vdev_remove_wanted) {
 738                 /*
 739                  * If provider's error is set we assume it is being
 740                  * removed.
 741                  */
 742                 if (bp->bio_to->error != 0) {
 743                         vd->vdev_remove_wanted = B_TRUE;
 744                         spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
 745                 } else if (!vd->vdev_delayed_close) {
 746                         vd->vdev_delayed_close = B_TRUE;
 747                 }
 748         }
 749         g_destroy_bio(bp);
 750         zio_interrupt(zio);
 751 }
 752
 753 static int
 754 vdev_geom_io_start(zio_t *zio)
 755 {
 756         vdev_t *vd;
 757         struct g_consumer *cp;
 758         struct bio *bp;
 759         int error;
 760
 761         vd = zio->io_vd;
 762
 763         if (zio->io_type == ZIO_TYPE_IOCTL) {
 764                 /* XXPOLICY */
 765                 if (!vdev_readable(vd)) {
 766                         zio->io_error = ENXIO;
 767                         return (ZIO_PIPELINE_CONTINUE);
 768                 }
 769
 770                 switch (zio->io_cmd) {
 771                 case DKIOCFLUSHWRITECACHE:
 772                         if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 773                                 break;
 774                         if (vd->vdev_nowritecache) {
 775                                 zio->io_error = ENOTSUP;
 776                                 break;
 777                         }
 778                         goto sendreq;
 779                 case DKIOCTRIM:
 780                         if (vdev_geom_bio_delete_disable)
 781                                 break;
 782                         if (vd->vdev_notrim) {
 783                                 zio->io_error = ENOTSUP;
 784                                 break;
 785                         }
 786                         goto sendreq;
 787                 default:
 788                         zio->io_error = ENOTSUP;
 789                 }
 790
 791                 return (ZIO_PIPELINE_CONTINUE);
 792         }
 793 sendreq:
 794         cp = vd->vdev_tsd;
 795         if (cp == NULL) {
 796                 zio->io_error = ENXIO;
 797                 return (ZIO_PIPELINE_CONTINUE);
 798         }
 799         bp = g_alloc_bio();
 800         bp->bio_caller1 = zio;
 801         switch (zio->io_type) {
 802         case ZIO_TYPE_READ:
 803         case ZIO_TYPE_WRITE:
 804                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 805                 bp->bio_data = zio->io_data;
 806                 bp->bio_offset = zio->io_offset;
 807                 bp->bio_length = zio->io_size;
 808                 break;
 809         case ZIO_TYPE_IOCTL:
 810                 switch (zio->io_cmd) {
 811                 case DKIOCFLUSHWRITECACHE:
 812                         bp->bio_cmd = BIO_FLUSH;
 813                         bp->bio_flags |= BIO_ORDERED;
 814                         bp->bio_data = NULL;
 815                         bp->bio_offset = cp->provider->mediasize;
 816                         bp->bio_length = 0;
 817                         break;
 818                 case DKIOCTRIM:
 819                         bp->bio_cmd = BIO_DELETE;
 820                         bp->bio_data = NULL;
 821                         bp->bio_offset = zio->io_offset;
 822                         bp->bio_length = zio->io_size;
 823                         break;
 824                 }
 825                 break;
 826         }
 827         bp->bio_done = vdev_geom_io_intr;
 828
 829         g_io_request(bp, cp);
 830
 831         return (ZIO_PIPELINE_STOP);
 832 }
 833
 834 static void
 835 vdev_geom_io_done(zio_t *zio)
 836 {
 837 }
 838
 839 static void
 840 vdev_geom_hold(vdev_t *vd)
 841 {
 842 }
 843
 844 static void
 845 vdev_geom_rele(vdev_t *vd)
 846 {
 847 }
 848
 849 vdev_ops_t vdev_geom_ops = {
 850         vdev_geom_open,
 851         vdev_geom_close,
 852         vdev_default_asize,
 853         vdev_geom_io_start,
 854         vdev_geom_io_done,
 855         NULL,
 856         vdev_geom_hold,
 857         vdev_geom_rele,
 858         VDEV_TYPE_DISK,         /* name of this vdev type */
 859         B_TRUE                  /* leaf vdev */
 860 };