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 static g_attrchanged_t vdev_geom_attrchanged;
  46 struct g_class zfs_vdev_class = {
  47         .name = "ZFS::VDEV",
  48         .version = G_VERSION,
  49         .attrchanged = vdev_geom_attrchanged,
  50 };
  51
  52 DECLARE_GEOM_CLASS(zfs_vdev_class, zfs_vdev);
  53
  54 SYSCTL_DECL(_vfs_zfs_vdev);
  55 /* Don't send BIO_FLUSH. */
  56 static int vdev_geom_bio_flush_disable;
  57 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_flush_disable, CTLFLAG_RWTUN,
  58     &vdev_geom_bio_flush_disable, 0, "Disable BIO_FLUSH");
  59 /* Don't send BIO_DELETE. */
  60 static int vdev_geom_bio_delete_disable;
  61 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_delete_disable, CTLFLAG_RWTUN,
  62     &vdev_geom_bio_delete_disable, 0, "Disable BIO_DELETE");
  63
  64 static void
  65 vdev_geom_set_rotation_rate(vdev_t *vd, struct g_consumer *cp)
  66 {
  67         int error;
  68         uint16_t rate;
  69
  70         error = g_getattr("GEOM::rotation_rate", cp, &rate);
  71         if (error == 0)
  72                 vd->vdev_rotation_rate = rate;
  73         else
  74                 vd->vdev_rotation_rate = VDEV_RATE_UNKNOWN;
  75 }
  76
  77 static void
  78 vdev_geom_attrchanged(struct g_consumer *cp, const char *attr)
  79 {
  80         vdev_t *vd;
  81         spa_t *spa;
  82         char *physpath;
  83         int error, physpath_len;
  84
  85         vd = cp->private;
  86         if (vd == NULL)
  87                 return;
  88
  89         if (strcmp(attr, "GEOM::rotation_rate") == 0) {
  90                 vdev_geom_set_rotation_rate(vd, cp);
  91                 return;
  92         }
  93
  94         if (strcmp(attr, "GEOM::physpath") != 0)
  95                 return;
  96
  97         if (g_access(cp, 1, 0, 0) != 0)
  98                 return;
  99
 100         /*
 101          * Record/Update physical path information for this device.
 102          */
 103         spa = vd->vdev_spa;
 104         physpath_len = MAXPATHLEN;
 105         physpath = g_malloc(physpath_len, M_WAITOK|M_ZERO);
 106         error = g_io_getattr("GEOM::physpath", cp, &physpath_len, physpath);
 107         g_access(cp, -1, 0, 0);
 108         if (error == 0) {
 109                 char *old_physpath;
 110
 111                 old_physpath = vd->vdev_physpath;
 112                 vd->vdev_physpath = spa_strdup(physpath);
 113                 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
 114
 115                 if (old_physpath != NULL) {
 116                         int held_lock;
 117
 118                         held_lock = spa_config_held(spa, SCL_STATE, RW_WRITER);
 119                         if (held_lock == 0) {
 120                                 g_topology_unlock();
 121                                 spa_config_enter(spa, SCL_STATE, FTAG,
 122                                     RW_WRITER);
 123                         }
 124
 125                         spa_strfree(old_physpath);
 126
 127                         if (held_lock == 0) {
 128                                 spa_config_exit(spa, SCL_STATE, FTAG);
 129                                 g_topology_lock();
 130                         }
 131                 }
 132         }
 133         g_free(physpath);
 134 }
 135
 136 static void
 137 vdev_geom_orphan(struct g_consumer *cp)
 138 {
 139         vdev_t *vd;
 140
 141         g_topology_assert();
 142
 143         vd = cp->private;
 144         if (vd == NULL) {
 145                 /* Vdev close in progress.  Ignore the event. */
 146                 return;
 147         }
 148
 149         /*
 150          * Orphan callbacks occur from the GEOM event thread.
 151          * Concurrent with this call, new I/O requests may be
 152          * working their way through GEOM about to find out
 153          * (only once executed by the g_down thread) that we've
 154          * been orphaned from our disk provider.  These I/Os
 155          * must be retired before we can detach our consumer.
 156          * This is most easily achieved by acquiring the
 157          * SPA ZIO configuration lock as a writer, but doing
 158          * so with the GEOM topology lock held would cause
 159          * a lock order reversal.  Instead, rely on the SPA's
 160          * async removal support to invoke a close on this
 161          * vdev once it is safe to do so.
 162          */
 163         zfs_post_remove(vd->vdev_spa, vd);
 164         vd->vdev_remove_wanted = B_TRUE;
 165         spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
 166 }
 167
 168 static struct g_consumer *
 169 vdev_geom_attach(struct g_provider *pp, vdev_t *vd)
 170 {
 171         struct g_geom *gp;
 172         struct g_consumer *cp;
 173
 174         g_topology_assert();
 175
 176         ZFS_LOG(1, "Attaching to %s.", pp->name);
 177         /* Do we have geom already? No? Create one. */
 178         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 179                 if (gp->flags & G_GEOM_WITHER)
 180                         continue;
 181                 if (strcmp(gp->name, "zfs::vdev") != 0)
 182                         continue;
 183                 break;
 184         }
 185         if (gp == NULL) {
 186                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 187                 gp->orphan = vdev_geom_orphan;
 188                 gp->attrchanged = vdev_geom_attrchanged;
 189                 cp = g_new_consumer(gp);
 190                 if (g_attach(cp, pp) != 0) {
 191                         g_wither_geom(gp, ENXIO);
 192                         return (NULL);
 193                 }
 194                 if (g_access(cp, 1, 0, 1) != 0) {
 195                         g_wither_geom(gp, ENXIO);
 196                         return (NULL);
 197                 }
 198                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 199         } else {
 200                 /* Check if we are already connected to this provider. */
 201                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 202                         if (cp->provider == pp) {
 203                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 204                                 break;
 205                         }
 206                 }
 207                 if (cp == NULL) {
 208                         cp = g_new_consumer(gp);
 209                         if (g_attach(cp, pp) != 0) {
 210                                 g_destroy_consumer(cp);
 211                                 return (NULL);
 212                         }
 213                         if (g_access(cp, 1, 0, 1) != 0) {
 214                                 g_detach(cp);
 215                                 g_destroy_consumer(cp);
 216                                 return (NULL);
 217                         }
 218                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 219                 } else {
 220                         if (g_access(cp, 1, 0, 1) != 0)
 221                                 return (NULL);
 222                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 223                 }
 224         }
 225
 226         /*
 227          * BUG: cp may already belong to a vdev.  This could happen if:
 228          * 1) That vdev is a shared spare, or
 229          * 2) We are trying to reopen a missing vdev and we are scanning by
 230          *    guid.  In that case, we'll ultimately fail to open this consumer,
 231          *    but not until after setting the private field.
 232          * The solution is to:
 233          * 1) Don't set the private field until after the open succeeds, and
 234          * 2) Set it to a linked list of vdevs, not just a single vdev
 235          */
 236         cp->private = vd;
 237         vd->vdev_tsd = cp;
 238
 239         /* Fetch initial physical path information for this device. */
 240         vdev_geom_attrchanged(cp, "GEOM::physpath");
 241
 242         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
 243         return (cp);
 244 }
 245
 246 static void
 247 vdev_geom_close_locked(vdev_t *vd)
 248 {
 249         struct g_geom *gp;
 250         struct g_consumer *cp;
 251
 252         g_topology_assert();
 253
 254         cp = vd->vdev_tsd;
 255         if (cp == NULL)
 256                 return;
 257
 258         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 259         KASSERT(vd->vdev_tsd == cp, ("%s: vdev_tsd is not cp", __func__));
 260         vd->vdev_tsd = NULL;
 261         vd->vdev_delayed_close = B_FALSE;
 262         cp->private = NULL;
 263
 264         gp = cp->geom;
 265         g_access(cp, -1, 0, -1);
 266         /* Destroy consumer on last close. */
 267         if (cp->acr == 0 && cp->ace == 0) {
 268                 if (cp->acw > 0)
 269                         g_access(cp, 0, -cp->acw, 0);
 270                 if (cp->provider != NULL) {
 271                         ZFS_LOG(1, "Destroyed consumer to %s.",
 272                             cp->provider->name);
 273                         g_detach(cp);
 274                 }
 275                 g_destroy_consumer(cp);
 276         }
 277         /* Destroy geom if there are no consumers left. */
 278         if (LIST_EMPTY(&gp->consumer)) {
 279                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 280                 g_wither_geom(gp, ENXIO);
 281         }
 282 }
 283
 284 static void
 285 nvlist_get_guids(nvlist_t *list, uint64_t *pguid, uint64_t *vguid)
 286 {
 287
 288         nvlist_lookup_uint64(list, ZPOOL_CONFIG_GUID, vguid);
 289         nvlist_lookup_uint64(list, ZPOOL_CONFIG_POOL_GUID, pguid);
 290 }
 291
 292 static int
 293 vdev_geom_io(struct g_consumer *cp, int cmd, void *data, off_t offset, off_t size)
 294 {
 295         struct bio *bp;
 296         u_char *p;
 297         off_t off, maxio;
 298         int error;
 299
 300         ASSERT((offset % cp->provider->sectorsize) == 0);
 301         ASSERT((size % cp->provider->sectorsize) == 0);
 302
 303         bp = g_alloc_bio();
 304         off = offset;
 305         offset += size;
 306         p = data;
 307         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 308         error = 0;
 309
 310         for (; off < offset; off += maxio, p += maxio, size -= maxio) {
 311                 bzero(bp, sizeof(*bp));
 312                 bp->bio_cmd = cmd;
 313                 bp->bio_done = NULL;
 314                 bp->bio_offset = off;
 315                 bp->bio_length = MIN(size, maxio);
 316                 bp->bio_data = p;
 317                 g_io_request(bp, cp);
 318                 error = biowait(bp, "vdev_geom_io");
 319                 if (error != 0)
 320                         break;
 321         }
 322
 323         g_destroy_bio(bp);
 324         return (error);
 325 }
 326
 327 static void
 328 vdev_geom_taste_orphan(struct g_consumer *cp)
 329 {
 330
 331         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
 332             cp->provider->name));
 333 }
 334
 335 static int
 336 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 337 {
 338         struct g_provider *pp;
 339         vdev_label_t *label;
 340         char *p, *buf;
 341         size_t buflen;
 342         uint64_t psize;
 343         off_t offset, size;
 344         uint64_t state, txg;
 345         int error, l, len;
 346
 347         g_topology_assert_not();
 348
 349         pp = cp->provider;
 350         ZFS_LOG(1, "Reading config from %s...", pp->name);
 351
 352         psize = pp->mediasize;
 353         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 354
 355         size = sizeof(*label) + pp->sectorsize -
 356             ((sizeof(*label) - 1) % pp->sectorsize) - 1;
 357
 358         label = kmem_alloc(size, KM_SLEEP);
 359         buflen = sizeof(label->vl_vdev_phys.vp_nvlist);
 360
 361         *config = NULL;
 362         for (l = 0; l < VDEV_LABELS; l++) {
 363
 364                 offset = vdev_label_offset(psize, l, 0);
 365                 if ((offset % pp->sectorsize) != 0)
 366                         continue;
 367
 368                 if (vdev_geom_io(cp, BIO_READ, label, offset, size) != 0)
 369                         continue;
 370                 buf = label->vl_vdev_phys.vp_nvlist;
 371
 372                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 373                         continue;
 374
 375                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 376                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 377                         nvlist_free(*config);
 378                         *config = NULL;
 379                         continue;
 380                 }
 381
 382                 if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
 383                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 384                     &txg) != 0 || txg == 0)) {
 385                         nvlist_free(*config);
 386                         *config = NULL;
 387                         continue;
 388                 }
 389
 390                 break;
 391         }
 392
 393         kmem_free(label, size);
 394         return (*config == NULL ? ENOENT : 0);
 395 }
 396
 397 static void
 398 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 399 {
 400         nvlist_t **new_configs;
 401         uint64_t i;
 402
 403         if (id < *count)
 404                 return;
 405         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 406             KM_SLEEP);
 407         for (i = 0; i < *count; i++)
 408                 new_configs[i] = (*configs)[i];
 409         if (*configs != NULL)
 410                 kmem_free(*configs, *count * sizeof(void *));
 411         *configs = new_configs;
 412         *count = id + 1;
 413 }
 414
 415 static void
 416 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 417     const char *name, uint64_t* known_pool_guid)
 418 {
 419         nvlist_t *vdev_tree;
 420         uint64_t pool_guid;
 421         uint64_t vdev_guid, known_guid;
 422         uint64_t id, txg, known_txg;
 423         char *pname;
 424         int i;
 425
 426         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 427             strcmp(pname, name) != 0)
 428                 goto ignore;
 429
 430         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 431                 goto ignore;
 432
 433         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 434                 goto ignore;
 435
 436         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 437                 goto ignore;
 438
 439         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 440                 goto ignore;
 441
 442         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 443
 444         if (*known_pool_guid != 0) {
 445                 if (pool_guid != *known_pool_guid)
 446                         goto ignore;
 447         } else
 448                 *known_pool_guid = pool_guid;
 449
 450         resize_configs(configs, count, id);
 451
 452         if ((*configs)[id] != NULL) {
 453                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 454                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 455                 if (txg <= known_txg)
 456                         goto ignore;
 457                 nvlist_free((*configs)[id]);
 458         }
 459
 460         (*configs)[id] = cfg;
 461         return;
 462
 463 ignore:
 464         nvlist_free(cfg);
 465 }
 466
 467 static int
 468 vdev_geom_attach_taster(struct g_consumer *cp, struct g_provider *pp)
 469 {
 470         int error;
 471
 472         if (pp->flags & G_PF_WITHER)
 473                 return (EINVAL);
 474         g_attach(cp, pp);
 475         error = g_access(cp, 1, 0, 0);
 476         if (error == 0) {
 477                 if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize))
 478                         error = EINVAL;
 479                 else if (pp->mediasize < SPA_MINDEVSIZE)
 480                         error = EINVAL;
 481                 if (error != 0)
 482                         g_access(cp, -1, 0, 0);
 483         }
 484         if (error != 0)
 485                 g_detach(cp);
 486         return (error);
 487 }
 488
 489 static void
 490 vdev_geom_detach_taster(struct g_consumer *cp)
 491 {
 492         g_access(cp, -1, 0, 0);
 493         g_detach(cp);
 494 }
 495
 496 int
 497 vdev_geom_read_pool_label(const char *name,
 498     nvlist_t ***configs, uint64_t *count)
 499 {
 500         struct g_class *mp;
 501         struct g_geom *gp, *zgp;
 502         struct g_provider *pp;
 503         struct g_consumer *zcp;
 504         nvlist_t *vdev_cfg;
 505         uint64_t pool_guid;
 506         int error;
 507
 508         DROP_GIANT();
 509         g_topology_lock();
 510
 511         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 512         /* This orphan function should be never called. */
 513         zgp->orphan = vdev_geom_taste_orphan;
 514         zcp = g_new_consumer(zgp);
 515
 516         *configs = NULL;
 517         *count = 0;
 518         pool_guid = 0;
 519         LIST_FOREACH(mp, &g_classes, class) {
 520                 if (mp == &zfs_vdev_class)
 521                         continue;
 522                 LIST_FOREACH(gp, &mp->geom, geom) {
 523                         if (gp->flags & G_GEOM_WITHER)
 524                                 continue;
 525                         LIST_FOREACH(pp, &gp->provider, provider) {
 526                                 if (pp->flags & G_PF_WITHER)
 527                                         continue;
 528                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 529                                         continue;
 530                                 g_topology_unlock();
 531                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 532                                 g_topology_lock();
 533                                 vdev_geom_detach_taster(zcp);
 534                                 if (error)
 535                                         continue;
 536                                 ZFS_LOG(1, "successfully read vdev config");
 537
 538                                 process_vdev_config(configs, count,
 539                                     vdev_cfg, name, &pool_guid);
 540                         }
 541                 }
 542         }
 543
 544         g_destroy_consumer(zcp);
 545         g_destroy_geom(zgp);
 546         g_topology_unlock();
 547         PICKUP_GIANT();
 548
 549         return (*count > 0 ? 0 : ENOENT);
 550 }
 551
 552 static void
 553 vdev_geom_read_guids(struct g_consumer *cp, uint64_t *pguid, uint64_t *vguid)
 554 {
 555         nvlist_t *config;
 556
 557         g_topology_assert_not();
 558
 559         *pguid = 0;
 560         *vguid = 0;
 561         if (vdev_geom_read_config(cp, &config) == 0) {
 562                 nvlist_get_guids(config, pguid, vguid);
 563                 nvlist_free(config);
 564         }
 565 }
 566
 567 static struct g_consumer *
 568 vdev_geom_attach_by_guids(vdev_t *vd)
 569 {
 570         struct g_class *mp;
 571         struct g_geom *gp, *zgp;
 572         struct g_provider *pp;
 573         struct g_consumer *cp, *zcp;
 574         uint64_t pguid, vguid;
 575
 576         g_topology_assert();
 577
 578         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 579         /* This orphan function should be never called. */
 580         zgp->orphan = vdev_geom_taste_orphan;
 581         zcp = g_new_consumer(zgp);
 582
 583         cp = NULL;
 584         LIST_FOREACH(mp, &g_classes, class) {
 585                 if (mp == &zfs_vdev_class)
 586                         continue;
 587                 LIST_FOREACH(gp, &mp->geom, geom) {
 588                         if (gp->flags & G_GEOM_WITHER)
 589                                 continue;
 590                         LIST_FOREACH(pp, &gp->provider, provider) {
 591                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 592                                         continue;
 593                                 g_topology_unlock();
 594                                 vdev_geom_read_guids(zcp, &pguid, &vguid);
 595                                 g_topology_lock();
 596                                 vdev_geom_detach_taster(zcp);
 597                                 /*
 598                                  * Check that the label's vdev guid matches the
 599                                  * desired guid.  If the label has a pool guid,
 600                                  * check that it matches too. (Inactive spares
 601                                  * and L2ARCs do not have any pool guid in the
 602                                  * label.)
 603                                 */
 604                                 if ((pguid != 0 &&
 605                                      pguid != spa_guid(vd->vdev_spa)) ||
 606                                     vguid != vd->vdev_guid)
 607                                         continue;
 608                                 cp = vdev_geom_attach(pp, vd);
 609                                 if (cp == NULL) {
 610                                         printf("ZFS WARNING: Unable to "
 611                                             "attach to %s.\n", pp->name);
 612                                         continue;
 613                                 }
 614                                 break;
 615                         }
 616                         if (cp != NULL)
 617                                 break;
 618                 }
 619                 if (cp != NULL)
 620                         break;
 621         }
 622 end:
 623         g_destroy_consumer(zcp);
 624         g_destroy_geom(zgp);
 625         return (cp);
 626 }
 627
 628 static struct g_consumer *
 629 vdev_geom_open_by_guids(vdev_t *vd)
 630 {
 631         struct g_consumer *cp;
 632         char *buf;
 633         size_t len;
 634
 635         g_topology_assert();
 636
 637         ZFS_LOG(1, "Searching by guid [%ju].", (uintmax_t)vd->vdev_guid);
 638         cp = vdev_geom_attach_by_guids(vd);
 639         if (cp != NULL) {
 640                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 641                 buf = kmem_alloc(len, KM_SLEEP);
 642
 643                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 644                 spa_strfree(vd->vdev_path);
 645                 vd->vdev_path = buf;
 646
 647                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 648                     (uintmax_t)spa_guid(vd->vdev_spa),
 649                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 650         } else {
 651                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 652                     (uintmax_t)spa_guid(vd->vdev_spa),
 653                     (uintmax_t)vd->vdev_guid);
 654         }
 655
 656         return (cp);
 657 }
 658
 659 static struct g_consumer *
 660 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 661 {
 662         struct g_provider *pp;
 663         struct g_consumer *cp;
 664         uint64_t pguid, vguid;
 665
 666         g_topology_assert();
 667
 668         cp = NULL;
 669         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 670         if (pp != NULL) {
 671                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 672                 cp = vdev_geom_attach(pp, vd);
 673                 if (cp != NULL && check_guid && ISP2(pp->sectorsize) &&
 674                     pp->sectorsize <= VDEV_PAD_SIZE) {
 675                         g_topology_unlock();
 676                         vdev_geom_read_guids(cp, &pguid, &vguid);
 677                         g_topology_lock();
 678                         if (pguid != spa_guid(vd->vdev_spa) ||
 679                             vguid != vd->vdev_guid) {
 680                                 vdev_geom_close_locked(vd);
 681                                 cp = NULL;
 682                                 ZFS_LOG(1, "guid mismatch for provider %s: "
 683                                     "%ju:%ju != %ju:%ju.", vd->vdev_path,
 684                                     (uintmax_t)spa_guid(vd->vdev_spa),
 685                                     (uintmax_t)vd->vdev_guid,
 686                                     (uintmax_t)pguid, (uintmax_t)vguid);
 687                         } else {
 688                                 ZFS_LOG(1, "guid match for provider %s.",
 689                                     vd->vdev_path);
 690                         }
 691                 }
 692         }
 693
 694         return (cp);
 695 }
 696
 697 static int
 698 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 699     uint64_t *logical_ashift, uint64_t *physical_ashift)
 700 {
 701         struct g_provider *pp;
 702         struct g_consumer *cp;
 703         size_t bufsize;
 704         int error;
 705
 706         /*
 707          * We must have a pathname, and it must be absolute.
 708          */
 709         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 710                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 711                 return (EINVAL);
 712         }
 713
 714         vd->vdev_tsd = NULL;
 715
 716         DROP_GIANT();
 717         g_topology_lock();
 718         error = 0;
 719
 720         if (vd->vdev_spa->spa_splitting_newspa ||
 721             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 722              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE)) {
 723                 /*
 724                  * We are dealing with a vdev that hasn't been previously
 725                  * opened (since boot), and we are not loading an
 726                  * existing pool configuration.  This looks like a
 727                  * vdev add operation to a new or existing pool.
 728                  * Assume the user knows what he/she is doing and find
 729                  * GEOM provider by its name, ignoring GUID mismatches.
 730                  *
 731                  * XXPOLICY: It would be safer to only allow a device
 732                  *           that is unlabeled or labeled but missing
 733                  *           GUID information to be opened in this fashion,
 734                  *           unless we are doing a split, in which case we
 735                  *           should allow any guid.
 736                  */
 737                 cp = vdev_geom_open_by_path(vd, 0);
 738         } else {
 739                 /*
 740                  * Try using the recorded path for this device, but only
 741                  * accept it if its label data contains the expected GUIDs.
 742                  */
 743                 cp = vdev_geom_open_by_path(vd, 1);
 744                 if (cp == NULL) {
 745                         /*
 746                          * The device at vd->vdev_path doesn't have the
 747                          * expected GUIDs. The disks might have merely
 748                          * moved around so try all other GEOM providers
 749                          * to find one with the right GUIDs.
 750                          */
 751                         cp = vdev_geom_open_by_guids(vd);
 752                 }
 753         }
 754
 755         if (cp == NULL) {
 756                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 757                 error = ENOENT;
 758         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 759             !ISP2(cp->provider->sectorsize)) {
 760                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 761                     vd->vdev_path);
 762
 763                 vdev_geom_close_locked(vd);
 764                 error = EINVAL;
 765                 cp = NULL;
 766         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 767                 int i;
 768
 769                 for (i = 0; i < 5; i++) {
 770                         error = g_access(cp, 0, 1, 0);
 771                         if (error == 0)
 772                                 break;
 773                         g_topology_unlock();
 774                         tsleep(vd, 0, "vdev", hz / 2);
 775                         g_topology_lock();
 776                 }
 777                 if (error != 0) {
 778                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 779                             vd->vdev_path, error);
 780                         vdev_geom_close_locked(vd);
 781                         cp = NULL;
 782                 }
 783         }
 784
 785         g_topology_unlock();
 786         PICKUP_GIANT();
 787         if (cp == NULL) {
 788                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 789                 return (error);
 790         }
 791         pp = cp->provider;
 792
 793         /*
 794          * Determine the actual size of the device.
 795          */
 796         *max_psize = *psize = pp->mediasize;
 797
 798         /*
 799          * Determine the device's minimum transfer size and preferred
 800          * transfer size.
 801          */
 802         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 803         *physical_ashift = 0;
 804         if (pp->stripesize)
 805                 *physical_ashift = highbit(pp->stripesize) - 1;
 806
 807         /*
 808          * Clear the nowritecache settings, so that on a vdev_reopen()
 809          * we will try again.
 810          */
 811         vd->vdev_nowritecache = B_FALSE;
 812
 813         /*
 814          * Determine the device's rotation rate.
 815          */
 816         vdev_geom_set_rotation_rate(vd, cp);
 817
 818         return (0);
 819 }
 820
 821 static void
 822 vdev_geom_close(vdev_t *vd)
 823 {
 824
 825         DROP_GIANT();
 826         g_topology_lock();
 827         vdev_geom_close_locked(vd);
 828         g_topology_unlock();
 829         PICKUP_GIANT();
 830 }
 831
 832 static void
 833 vdev_geom_io_intr(struct bio *bp)
 834 {
 835         vdev_t *vd;
 836         zio_t *zio;
 837
 838         zio = bp->bio_caller1;
 839         vd = zio->io_vd;
 840         zio->io_error = bp->bio_error;
 841         if (zio->io_error == 0 && bp->bio_resid != 0)
 842                 zio->io_error = SET_ERROR(EIO);
 843
 844         switch(zio->io_error) {
 845         case ENOTSUP:
 846                 /*
 847                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 848                  * that future attempts will never succeed. In this case
 849                  * we set a persistent flag so that we don't bother with
 850                  * requests in the future.
 851                  */
 852                 switch(bp->bio_cmd) {
 853                 case BIO_FLUSH:
 854                         vd->vdev_nowritecache = B_TRUE;
 855                         break;
 856                 case BIO_DELETE:
 857                         vd->vdev_notrim = B_TRUE;
 858                         break;
 859                 }
 860                 break;
 861         case ENXIO:
 862                 if (!vd->vdev_remove_wanted) {
 863                         /*
 864                          * If provider's error is set we assume it is being
 865                          * removed.
 866                          */
 867                         if (bp->bio_to->error != 0) {
 868                                 vd->vdev_remove_wanted = B_TRUE;
 869                                 spa_async_request(zio->io_spa,
 870                                     SPA_ASYNC_REMOVE);
 871                         } else if (!vd->vdev_delayed_close) {
 872                                 vd->vdev_delayed_close = B_TRUE;
 873                         }
 874                 }
 875                 break;
 876         }
 877         g_destroy_bio(bp);
 878         zio_interrupt(zio);
 879 }
 880
 881 static void
 882 vdev_geom_io_start(zio_t *zio)
 883 {
 884         vdev_t *vd;
 885         struct g_consumer *cp;
 886         struct bio *bp;
 887         int error;
 888
 889         vd = zio->io_vd;
 890
 891         switch (zio->io_type) {
 892         case ZIO_TYPE_IOCTL:
 893                 /* XXPOLICY */
 894                 if (!vdev_readable(vd)) {
 895                         zio->io_error = SET_ERROR(ENXIO);
 896                         zio_interrupt(zio);
 897                         return;
 898                 } else {
 899                         switch (zio->io_cmd) {
 900                         case DKIOCFLUSHWRITECACHE:
 901                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 902                                         break;
 903                                 if (vd->vdev_nowritecache) {
 904                                         zio->io_error = SET_ERROR(ENOTSUP);
 905                                         break;
 906                                 }
 907                                 goto sendreq;
 908                         default:
 909                                 zio->io_error = SET_ERROR(ENOTSUP);
 910                         }
 911                 }
 912
 913                 zio_execute(zio);
 914                 return;
 915         case ZIO_TYPE_FREE:
 916                 if (vd->vdev_notrim) {
 917                         zio->io_error = SET_ERROR(ENOTSUP);
 918                 } else if (!vdev_geom_bio_delete_disable) {
 919                         goto sendreq;
 920                 }
 921                 zio_execute(zio);
 922                 return;
 923         }
 924 sendreq:
 925         ASSERT(zio->io_type == ZIO_TYPE_READ ||
 926             zio->io_type == ZIO_TYPE_WRITE ||
 927             zio->io_type == ZIO_TYPE_FREE ||
 928             zio->io_type == ZIO_TYPE_IOCTL);
 929
 930         cp = vd->vdev_tsd;
 931         if (cp == NULL) {
 932                 zio->io_error = SET_ERROR(ENXIO);
 933                 zio_interrupt(zio);
 934                 return;
 935         }
 936         bp = g_alloc_bio();
 937         bp->bio_caller1 = zio;
 938         switch (zio->io_type) {
 939         case ZIO_TYPE_READ:
 940         case ZIO_TYPE_WRITE:
 941                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 942                 bp->bio_data = zio->io_data;
 943                 bp->bio_offset = zio->io_offset;
 944                 bp->bio_length = zio->io_size;
 945                 break;
 946         case ZIO_TYPE_FREE:
 947                 bp->bio_cmd = BIO_DELETE;
 948                 bp->bio_data = NULL;
 949                 bp->bio_offset = zio->io_offset;
 950                 bp->bio_length = zio->io_size;
 951                 break;
 952         case ZIO_TYPE_IOCTL:
 953                 bp->bio_cmd = BIO_FLUSH;
 954                 bp->bio_flags |= BIO_ORDERED;
 955                 bp->bio_data = NULL;
 956                 bp->bio_offset = cp->provider->mediasize;
 957                 bp->bio_length = 0;
 958                 break;
 959         }
 960         bp->bio_done = vdev_geom_io_intr;
 961
 962         g_io_request(bp, cp);
 963 }
 964
 965 static void
 966 vdev_geom_io_done(zio_t *zio)
 967 {
 968 }
 969
 970 static void
 971 vdev_geom_hold(vdev_t *vd)
 972 {
 973 }
 974
 975 static void
 976 vdev_geom_rele(vdev_t *vd)
 977 {
 978 }
 979
 980 vdev_ops_t vdev_geom_ops = {
 981         vdev_geom_open,
 982         vdev_geom_close,
 983         vdev_default_asize,
 984         vdev_geom_io_start,
 985         vdev_geom_io_done,
 986         NULL,
 987         vdev_geom_hold,
 988         vdev_geom_rele,
 989         VDEV_TYPE_DISK,         /* name of this vdev type */
 990         B_TRUE                  /* leaf vdev */
 991 };