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 /* Declare local functions */
  65 static void vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read);
  66
  67 /*
  68  * Thread local storage used to indicate when a thread is probing geoms
  69  * for their guids.  If NULL, this thread is not tasting geoms.  If non NULL,
  70  * it is looking for a replacement for the vdev_t* that is its value.
  71  */
  72 uint_t zfs_geom_probe_vdev_key;
  73
  74 static void
  75 vdev_geom_set_rotation_rate(vdev_t *vd, struct g_consumer *cp)
  76 {
  77         int error;
  78         uint16_t rate;
  79
  80         error = g_getattr("GEOM::rotation_rate", cp, &rate);
  81         if (error == 0)
  82                 vd->vdev_rotation_rate = rate;
  83         else
  84                 vd->vdev_rotation_rate = VDEV_RATE_UNKNOWN;
  85 }
  86
  87 static void
  88 vdev_geom_attrchanged(struct g_consumer *cp, const char *attr)
  89 {
  90         vdev_t *vd;
  91         spa_t *spa;
  92         char *physpath;
  93         int error, physpath_len;
  94
  95         vd = cp->private;
  96         if (vd == NULL)
  97                 return;
  98
  99         if (strcmp(attr, "GEOM::rotation_rate") == 0) {
 100                 vdev_geom_set_rotation_rate(vd, cp);
 101                 return;
 102         }
 103
 104         if (strcmp(attr, "GEOM::physpath") != 0)
 105                 return;
 106
 107         if (g_access(cp, 1, 0, 0) != 0)
 108                 return;
 109
 110         /*
 111          * Record/Update physical path information for this device.
 112          */
 113         spa = vd->vdev_spa;
 114         physpath_len = MAXPATHLEN;
 115         physpath = g_malloc(physpath_len, M_WAITOK|M_ZERO);
 116         error = g_io_getattr("GEOM::physpath", cp, &physpath_len, physpath);
 117         g_access(cp, -1, 0, 0);
 118         if (error == 0) {
 119                 char *old_physpath;
 120
 121                 /* g_topology lock ensures that vdev has not been closed */
 122                 g_topology_assert();
 123                 old_physpath = vd->vdev_physpath;
 124                 vd->vdev_physpath = spa_strdup(physpath);
 125                 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
 126
 127                 if (old_physpath != NULL)
 128                         spa_strfree(old_physpath);
 129         }
 130         g_free(physpath);
 131 }
 132
 133 static void
 134 vdev_geom_orphan(struct g_consumer *cp)
 135 {
 136         vdev_t *vd;
 137
 138         g_topology_assert();
 139
 140         vd = cp->private;
 141         if (vd == NULL) {
 142                 /* Vdev close in progress.  Ignore the event. */
 143                 return;
 144         }
 145
 146         /*
 147          * Orphan callbacks occur from the GEOM event thread.
 148          * Concurrent with this call, new I/O requests may be
 149          * working their way through GEOM about to find out
 150          * (only once executed by the g_down thread) that we've
 151          * been orphaned from our disk provider.  These I/Os
 152          * must be retired before we can detach our consumer.
 153          * This is most easily achieved by acquiring the
 154          * SPA ZIO configuration lock as a writer, but doing
 155          * so with the GEOM topology lock held would cause
 156          * a lock order reversal.  Instead, rely on the SPA's
 157          * async removal support to invoke a close on this
 158          * vdev once it is safe to do so.
 159          */
 160         vd->vdev_remove_wanted = B_TRUE;
 161         spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
 162 }
 163
 164 static struct g_consumer *
 165 vdev_geom_attach(struct g_provider *pp, vdev_t *vd)
 166 {
 167         struct g_geom *gp;
 168         struct g_consumer *cp;
 169         int error;
 170
 171         g_topology_assert();
 172
 173         ZFS_LOG(1, "Attaching to %s.", pp->name);
 174
 175         if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize)) {
 176                 ZFS_LOG(1, "Failing attach of %s. Incompatible sectorsize %d\n",
 177                     pp->name, pp->sectorsize);
 178                 return (NULL);
 179         } else if (pp->mediasize < SPA_MINDEVSIZE) {
 180                 ZFS_LOG(1, "Failing attach of %s. Incompatible mediasize %ju\n",
 181                     pp->name, pp->mediasize);
 182                 return (NULL);
 183         }
 184
 185         /* Do we have geom already? No? Create one. */
 186         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 187                 if (gp->flags & G_GEOM_WITHER)
 188                         continue;
 189                 if (strcmp(gp->name, "zfs::vdev") != 0)
 190                         continue;
 191                 break;
 192         }
 193         if (gp == NULL) {
 194                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 195                 gp->orphan = vdev_geom_orphan;
 196                 gp->attrchanged = vdev_geom_attrchanged;
 197                 cp = g_new_consumer(gp);
 198                 error = g_attach(cp, pp);
 199                 if (error != 0) {
 200                         ZFS_LOG(1, "%s(%d): g_attach failed: %d\n", __func__,
 201                             __LINE__, error);
 202                         vdev_geom_detach(cp, B_FALSE);
 203                         return (NULL);
 204                 }
 205                 error = g_access(cp, 1, 0, 1);
 206                 if (error != 0) {
 207                         ZFS_LOG(1, "%s(%d): g_access failed: %d\n", __func__,
 208                                __LINE__, error);
 209                         vdev_geom_detach(cp, B_FALSE);
 210                         return (NULL);
 211                 }
 212                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 213         } else {
 214                 /* Check if we are already connected to this provider. */
 215                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 216                         if (cp->provider == pp) {
 217                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 218                                 break;
 219                         }
 220                 }
 221                 if (cp == NULL) {
 222                         cp = g_new_consumer(gp);
 223                         error = g_attach(cp, pp);
 224                         if (error != 0) {
 225                                 ZFS_LOG(1, "%s(%d): g_attach failed: %d\n",
 226                                     __func__, __LINE__, error);
 227                                 vdev_geom_detach(cp, B_FALSE);
 228                                 return (NULL);
 229                         }
 230                         error = g_access(cp, 1, 0, 1);
 231                         if (error != 0) {
 232                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 233                                     __func__, __LINE__, error);
 234                                 vdev_geom_detach(cp, B_FALSE);
 235                                 return (NULL);
 236                         }
 237                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 238                 } else {
 239                         error = g_access(cp, 1, 0, 1);
 240                         if (error != 0) {
 241                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 242                                     __func__, __LINE__, error);
 243                                 return (NULL);
 244                         }
 245                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 246                 }
 247         }
 248
 249         /*
 250          * BUG: cp may already belong to a vdev.  This could happen if:
 251          * 1) That vdev is a shared spare, or
 252          * 2) We are trying to reopen a missing vdev and we are scanning by
 253          *    guid.  In that case, we'll ultimately fail to open this consumer,
 254          *    but not until after setting the private field.
 255          * The solution is to:
 256          * 1) Don't set the private field until after the open succeeds, and
 257          * 2) Set it to a linked list of vdevs, not just a single vdev
 258          */
 259         cp->private = vd;
 260         if (vd != NULL)
 261                 vd->vdev_tsd = cp;
 262
 263         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
 264         return (cp);
 265 }
 266
 267 static void
 268 vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read)
 269 {
 270         struct g_geom *gp;
 271         vdev_t *vd;
 272
 273         g_topology_assert();
 274
 275         ZFS_LOG(1, "Detaching consumer. Provider %s.",
 276             cp->provider && cp->provider->name ? cp->provider->name : "NULL");
 277
 278         vd = cp->private;
 279         cp->private = NULL;
 280
 281         gp = cp->geom;
 282         if (open_for_read)
 283                 g_access(cp, -1, 0, -1);
 284         /* Destroy consumer on last close. */
 285         if (cp->acr == 0 && cp->ace == 0) {
 286                 if (cp->acw > 0)
 287                         g_access(cp, 0, -cp->acw, 0);
 288                 if (cp->provider != NULL) {
 289                         ZFS_LOG(1, "Destroying consumer to %s.",
 290                             cp->provider->name ? cp->provider->name : "NULL");
 291                         g_detach(cp);
 292                 }
 293                 g_destroy_consumer(cp);
 294         }
 295         /* Destroy geom if there are no consumers left. */
 296         if (LIST_EMPTY(&gp->consumer)) {
 297                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 298                 g_wither_geom(gp, ENXIO);
 299         }
 300 }
 301
 302 static void
 303 vdev_geom_close_locked(vdev_t *vd)
 304 {
 305         struct g_consumer *cp;
 306
 307         g_topology_assert();
 308
 309         cp = vd->vdev_tsd;
 310         vd->vdev_tsd = NULL;
 311         vd->vdev_delayed_close = B_FALSE;
 312         if (cp == NULL)
 313                 return;
 314
 315         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 316
 317         vdev_geom_detach(cp, B_TRUE);
 318 }
 319
 320 static void
 321 nvlist_get_guids(nvlist_t *list, uint64_t *pguid, uint64_t *vguid)
 322 {
 323
 324         (void) nvlist_lookup_uint64(list, ZPOOL_CONFIG_GUID, vguid);
 325         (void) nvlist_lookup_uint64(list, ZPOOL_CONFIG_POOL_GUID, pguid);
 326 }
 327
 328 /*
 329  * Issue one or more bios to the vdev in parallel
 330  * cmds, datas, offsets, errors, and sizes are arrays of length ncmds.  Each IO
 331  * operation is described by parallel entries from each array.  There may be
 332  * more bios actually issued than entries in the array
 333  */
 334 static void
 335 vdev_geom_io(struct g_consumer *cp, int *cmds, void **datas, off_t *offsets,
 336     off_t *sizes, int *errors, int ncmds)
 337 {
 338         struct bio **bios;
 339         u_char *p;
 340         off_t off, maxio, s, end;
 341         int i, n_bios, j;
 342         size_t bios_size;
 343
 344         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 345         n_bios = 0;
 346
 347         /* How many bios are required for all commands ? */
 348         for (i = 0; i < ncmds; i++)
 349                 n_bios += (sizes[i] + maxio - 1) / maxio;
 350
 351         /* Allocate memory for the bios */
 352         bios_size = n_bios * sizeof(struct bio*);
 353         bios = kmem_zalloc(bios_size, KM_SLEEP);
 354
 355         /* Prepare and issue all of the bios */
 356         for (i = j = 0; i < ncmds; i++) {
 357                 off = offsets[i];
 358                 p = datas[i];
 359                 s = sizes[i];
 360                 end = off + s;
 361                 ASSERT((off % cp->provider->sectorsize) == 0);
 362                 ASSERT((s % cp->provider->sectorsize) == 0);
 363
 364                 for (; off < end; off += maxio, p += maxio, s -= maxio, j++) {
 365                         bios[j] = g_alloc_bio();
 366                         bios[j]->bio_cmd = cmds[i];
 367                         bios[j]->bio_done = NULL;
 368                         bios[j]->bio_offset = off;
 369                         bios[j]->bio_length = MIN(s, maxio);
 370                         bios[j]->bio_data = p;
 371                         g_io_request(bios[j], cp);
 372                 }
 373         }
 374         ASSERT(j == n_bios);
 375
 376         /* Wait for all of the bios to complete, and clean them up */
 377         for (i = j = 0; i < ncmds; i++) {
 378                 off = offsets[i];
 379                 s = sizes[i];
 380                 end = off + s;
 381
 382                 for (; off < end; off += maxio, s -= maxio, j++) {
 383                         errors[i] = biowait(bios[j], "vdev_geom_io") || errors[i];
 384                         g_destroy_bio(bios[j]);
 385                 }
 386         }
 387         kmem_free(bios, bios_size);
 388 }
 389
 390 static int
 391 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 392 {
 393         struct g_provider *pp;
 394         vdev_phys_t *vdev_lists[VDEV_LABELS];
 395         char *p, *buf;
 396         size_t buflen;
 397         uint64_t psize, state, txg;
 398         off_t offsets[VDEV_LABELS];
 399         off_t size;
 400         off_t sizes[VDEV_LABELS];
 401         int cmds[VDEV_LABELS];
 402         int errors[VDEV_LABELS];
 403         int l, len;
 404
 405         g_topology_assert_not();
 406
 407         pp = cp->provider;
 408         ZFS_LOG(1, "Reading config from %s...", pp->name);
 409
 410         psize = pp->mediasize;
 411         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 412
 413         size = sizeof(*vdev_lists[0]) + pp->sectorsize -
 414             ((sizeof(*vdev_lists[0]) - 1) % pp->sectorsize) - 1;
 415
 416         buflen = sizeof(vdev_lists[0]->vp_nvlist);
 417
 418         *config = NULL;
 419         /* Create all of the IO requests */
 420         for (l = 0; l < VDEV_LABELS; l++) {
 421                 cmds[l] = BIO_READ;
 422                 vdev_lists[l] = kmem_alloc(size, KM_SLEEP);
 423                 offsets[l] = vdev_label_offset(psize, l, 0) + VDEV_SKIP_SIZE;
 424                 sizes[l] = size;
 425                 errors[l] = 0;
 426                 ASSERT(offsets[l] % pp->sectorsize == 0);
 427         }
 428
 429         /* Issue the IO requests */
 430         vdev_geom_io(cp, cmds, (void**)vdev_lists, offsets, sizes, errors,
 431             VDEV_LABELS);
 432
 433         /* Parse the labels */
 434         for (l = 0; l < VDEV_LABELS; l++) {
 435                 if (errors[l] != 0)
 436                         continue;
 437
 438                 buf = vdev_lists[l]->vp_nvlist;
 439
 440                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 441                         continue;
 442
 443                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 444                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 445                         nvlist_free(*config);
 446                         *config = NULL;
 447                         continue;
 448                 }
 449
 450                 if (state != POOL_STATE_SPARE &&
 451                     state != POOL_STATE_L2CACHE &&
 452                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 453                     &txg) != 0 || txg == 0)) {
 454                         nvlist_free(*config);
 455                         *config = NULL;
 456                         continue;
 457                 }
 458
 459                 break;
 460         }
 461
 462         /* Free the label storage */
 463         for (l = 0; l < VDEV_LABELS; l++)
 464                 kmem_free(vdev_lists[l], size);
 465
 466         return (*config == NULL ? ENOENT : 0);
 467 }
 468
 469 static void
 470 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 471 {
 472         nvlist_t **new_configs;
 473         uint64_t i;
 474
 475         if (id < *count)
 476                 return;
 477         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 478             KM_SLEEP);
 479         for (i = 0; i < *count; i++)
 480                 new_configs[i] = (*configs)[i];
 481         if (*configs != NULL)
 482                 kmem_free(*configs, *count * sizeof(void *));
 483         *configs = new_configs;
 484         *count = id + 1;
 485 }
 486
 487 static void
 488 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 489     const char *name, uint64_t* known_pool_guid)
 490 {
 491         nvlist_t *vdev_tree;
 492         uint64_t pool_guid;
 493         uint64_t vdev_guid, known_guid;
 494         uint64_t id, txg, known_txg;
 495         char *pname;
 496         int i;
 497
 498         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 499             strcmp(pname, name) != 0)
 500                 goto ignore;
 501
 502         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 503                 goto ignore;
 504
 505         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 506                 goto ignore;
 507
 508         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 509                 goto ignore;
 510
 511         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 512                 goto ignore;
 513
 514         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 515
 516         if (*known_pool_guid != 0) {
 517                 if (pool_guid != *known_pool_guid)
 518                         goto ignore;
 519         } else
 520                 *known_pool_guid = pool_guid;
 521
 522         resize_configs(configs, count, id);
 523
 524         if ((*configs)[id] != NULL) {
 525                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 526                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 527                 if (txg <= known_txg)
 528                         goto ignore;
 529                 nvlist_free((*configs)[id]);
 530         }
 531
 532         (*configs)[id] = cfg;
 533         return;
 534
 535 ignore:
 536         nvlist_free(cfg);
 537 }
 538
 539 int
 540 vdev_geom_read_pool_label(const char *name,
 541     nvlist_t ***configs, uint64_t *count)
 542 {
 543         struct g_class *mp;
 544         struct g_geom *gp;
 545         struct g_provider *pp;
 546         struct g_consumer *zcp;
 547         nvlist_t *vdev_cfg;
 548         uint64_t pool_guid;
 549         int error;
 550
 551         DROP_GIANT();
 552         g_topology_lock();
 553
 554         *configs = NULL;
 555         *count = 0;
 556         pool_guid = 0;
 557         LIST_FOREACH(mp, &g_classes, class) {
 558                 if (mp == &zfs_vdev_class)
 559                         continue;
 560                 LIST_FOREACH(gp, &mp->geom, geom) {
 561                         if (gp->flags & G_GEOM_WITHER)
 562                                 continue;
 563                         LIST_FOREACH(pp, &gp->provider, provider) {
 564                                 if (pp->flags & G_PF_WITHER)
 565                                         continue;
 566                                 zcp = vdev_geom_attach(pp, NULL);
 567                                 if (zcp == NULL)
 568                                         continue;
 569                                 g_topology_unlock();
 570                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 571                                 g_topology_lock();
 572                                 vdev_geom_detach(zcp, B_TRUE);
 573                                 if (error)
 574                                         continue;
 575                                 ZFS_LOG(1, "successfully read vdev config");
 576
 577                                 process_vdev_config(configs, count,
 578                                     vdev_cfg, name, &pool_guid);
 579                         }
 580                 }
 581         }
 582         g_topology_unlock();
 583         PICKUP_GIANT();
 584
 585         return (*count > 0 ? 0 : ENOENT);
 586 }
 587
 588 static void
 589 vdev_geom_read_guids(struct g_consumer *cp, uint64_t *pguid, uint64_t *vguid)
 590 {
 591         nvlist_t *config;
 592
 593         g_topology_assert_not();
 594
 595         *pguid = 0;
 596         *vguid = 0;
 597         if (vdev_geom_read_config(cp, &config) == 0) {
 598                 nvlist_get_guids(config, pguid, vguid);
 599                 nvlist_free(config);
 600         }
 601 }
 602
 603 static boolean_t
 604 vdev_attach_ok(vdev_t *vd, struct g_provider *pp)
 605 {
 606         uint64_t pool_guid;
 607         uint64_t vdev_guid;
 608         struct g_consumer *zcp;
 609         boolean_t pool_ok;
 610         boolean_t vdev_ok;
 611
 612         zcp = vdev_geom_attach(pp, NULL);
 613         if (zcp == NULL) {
 614                 ZFS_LOG(1, "Unable to attach tasting instance to %s.",
 615                     pp->name);
 616                 return (B_FALSE);
 617         }
 618         g_topology_unlock();
 619         vdev_geom_read_guids(zcp, &pool_guid, &vdev_guid);
 620         g_topology_lock();
 621         vdev_geom_detach(zcp, B_TRUE);
 622
 623         /*
 624          * Check that the label's vdev guid matches the desired guid.  If the
 625          * label has a pool guid, check that it matches too. (Inactive spares
 626          * and L2ARCs do not have any pool guid in the label.)
 627          */
 628         if ((pool_guid == 0 || pool_guid == spa_guid(vd->vdev_spa)) &&
 629             vdev_guid == vd->vdev_guid) {
 630                 ZFS_LOG(1, "guids match for provider %s.", vd->vdev_path);
 631                 return (B_TRUE);
 632         } else {
 633                 ZFS_LOG(1, "guid mismatch for provider %s: "
 634                     "%ju:%ju != %ju:%ju.", vd->vdev_path,
 635                     (uintmax_t)spa_guid(vd->vdev_spa),
 636                     (uintmax_t)vd->vdev_guid,
 637                     (uintmax_t)pool_guid, (uintmax_t)vdev_guid);
 638                 return (B_FALSE);
 639         }
 640 }
 641
 642 static struct g_consumer *
 643 vdev_geom_attach_by_guids(vdev_t *vd)
 644 {
 645         struct g_class *mp;
 646         struct g_geom *gp;
 647         struct g_provider *pp;
 648         struct g_consumer *cp;
 649
 650         g_topology_assert();
 651
 652         cp = NULL;
 653         LIST_FOREACH(mp, &g_classes, class) {
 654                 if (mp == &zfs_vdev_class)
 655                         continue;
 656                 LIST_FOREACH(gp, &mp->geom, geom) {
 657                         if (gp->flags & G_GEOM_WITHER)
 658                                 continue;
 659                         LIST_FOREACH(pp, &gp->provider, provider) {
 660                                 if (!vdev_attach_ok(vd, pp))
 661                                         continue;
 662                                 cp = vdev_geom_attach(pp, vd);
 663                                 if (cp == NULL) {
 664                                         printf("ZFS WARNING: Unable to "
 665                                             "attach to %s.\n", pp->name);
 666                                         continue;
 667                                 }
 668                                 break;
 669                         }
 670                         if (cp != NULL)
 671                                 break;
 672                 }
 673                 if (cp != NULL)
 674                         break;
 675         }
 676 end:
 677         return (cp);
 678 }
 679
 680 static struct g_consumer *
 681 vdev_geom_open_by_guids(vdev_t *vd)
 682 {
 683         struct g_consumer *cp;
 684         char *buf;
 685         size_t len;
 686
 687         g_topology_assert();
 688
 689         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 690                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 691         cp = vdev_geom_attach_by_guids(vd);
 692         if (cp != NULL) {
 693                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 694                 buf = kmem_alloc(len, KM_SLEEP);
 695
 696                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 697                 spa_strfree(vd->vdev_path);
 698                 vd->vdev_path = buf;
 699
 700                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 701                     (uintmax_t)spa_guid(vd->vdev_spa),
 702                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 703         } else {
 704                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 705                     (uintmax_t)spa_guid(vd->vdev_spa),
 706                     (uintmax_t)vd->vdev_guid);
 707         }
 708
 709         return (cp);
 710 }
 711
 712 static struct g_consumer *
 713 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 714 {
 715         struct g_provider *pp;
 716         struct g_consumer *cp;
 717
 718         g_topology_assert();
 719
 720         cp = NULL;
 721         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 722         if (pp != NULL) {
 723                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 724                 if (!check_guid || vdev_attach_ok(vd, pp))
 725                         cp = vdev_geom_attach(pp, vd);
 726         }
 727
 728         return (cp);
 729 }
 730
 731 static int
 732 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 733     uint64_t *logical_ashift, uint64_t *physical_ashift)
 734 {
 735         struct g_provider *pp;
 736         struct g_consumer *cp;
 737         size_t bufsize;
 738         int error;
 739
 740         /* Set the TLS to indicate downstack that we should not access zvols*/
 741         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 742
 743         /*
 744          * We must have a pathname, and it must be absolute.
 745          */
 746         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 747                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 748                 return (EINVAL);
 749         }
 750
 751         vd->vdev_tsd = NULL;
 752
 753         DROP_GIANT();
 754         g_topology_lock();
 755         error = 0;
 756
 757         if (vd->vdev_spa->spa_splitting_newspa ||
 758             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 759              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE)) {
 760                 /*
 761                  * We are dealing with a vdev that hasn't been previously
 762                  * opened (since boot), and we are not loading an
 763                  * existing pool configuration.  This looks like a
 764                  * vdev add operation to a new or existing pool.
 765                  * Assume the user knows what he/she is doing and find
 766                  * GEOM provider by its name, ignoring GUID mismatches.
 767                  *
 768                  * XXPOLICY: It would be safer to only allow a device
 769                  *           that is unlabeled or labeled but missing
 770                  *           GUID information to be opened in this fashion,
 771                  *           unless we are doing a split, in which case we
 772                  *           should allow any guid.
 773                  */
 774                 cp = vdev_geom_open_by_path(vd, 0);
 775         } else {
 776                 /*
 777                  * Try using the recorded path for this device, but only
 778                  * accept it if its label data contains the expected GUIDs.
 779                  */
 780                 cp = vdev_geom_open_by_path(vd, 1);
 781                 if (cp == NULL) {
 782                         /*
 783                          * The device at vd->vdev_path doesn't have the
 784                          * expected GUIDs. The disks might have merely
 785                          * moved around so try all other GEOM providers
 786                          * to find one with the right GUIDs.
 787                          */
 788                         cp = vdev_geom_open_by_guids(vd);
 789                 }
 790         }
 791
 792         /* Clear the TLS now that tasting is done */
 793         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 794
 795         if (cp == NULL) {
 796                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 797                 error = ENOENT;
 798         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 799             !ISP2(cp->provider->sectorsize)) {
 800                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 801                     vd->vdev_path);
 802
 803                 vdev_geom_close_locked(vd);
 804                 error = EINVAL;
 805                 cp = NULL;
 806         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 807                 int i;
 808
 809                 for (i = 0; i < 5; i++) {
 810                         error = g_access(cp, 0, 1, 0);
 811                         if (error == 0)
 812                                 break;
 813                         g_topology_unlock();
 814                         tsleep(vd, 0, "vdev", hz / 2);
 815                         g_topology_lock();
 816                 }
 817                 if (error != 0) {
 818                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 819                             vd->vdev_path, error);
 820                         vdev_geom_close_locked(vd);
 821                         cp = NULL;
 822                 }
 823         }
 824
 825         /* Fetch initial physical path information for this device. */
 826         if (cp != NULL)
 827                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 828
 829         g_topology_unlock();
 830         PICKUP_GIANT();
 831         if (cp == NULL) {
 832                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 833                 return (error);
 834         }
 835         pp = cp->provider;
 836
 837         /*
 838          * Determine the actual size of the device.
 839          */
 840         *max_psize = *psize = pp->mediasize;
 841
 842         /*
 843          * Determine the device's minimum transfer size and preferred
 844          * transfer size.
 845          */
 846         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 847         *physical_ashift = 0;
 848         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 849             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 850                 *physical_ashift = highbit(pp->stripesize) - 1;
 851
 852         /*
 853          * Clear the nowritecache settings, so that on a vdev_reopen()
 854          * we will try again.
 855          */
 856         vd->vdev_nowritecache = B_FALSE;
 857
 858         /*
 859          * Determine the device's rotation rate.
 860          */
 861         vdev_geom_set_rotation_rate(vd, cp);
 862
 863         return (0);
 864 }
 865
 866 static void
 867 vdev_geom_close(vdev_t *vd)
 868 {
 869
 870         DROP_GIANT();
 871         g_topology_lock();
 872         vdev_geom_close_locked(vd);
 873         g_topology_unlock();
 874         PICKUP_GIANT();
 875 }
 876
 877 static void
 878 vdev_geom_io_intr(struct bio *bp)
 879 {
 880         vdev_t *vd;
 881         zio_t *zio;
 882
 883         zio = bp->bio_caller1;
 884         vd = zio->io_vd;
 885         zio->io_error = bp->bio_error;
 886         if (zio->io_error == 0 && bp->bio_resid != 0)
 887                 zio->io_error = SET_ERROR(EIO);
 888
 889         switch(zio->io_error) {
 890         case ENOTSUP:
 891                 /*
 892                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 893                  * that future attempts will never succeed. In this case
 894                  * we set a persistent flag so that we don't bother with
 895                  * requests in the future.
 896                  */
 897                 switch(bp->bio_cmd) {
 898                 case BIO_FLUSH:
 899                         vd->vdev_nowritecache = B_TRUE;
 900                         break;
 901                 case BIO_DELETE:
 902                         vd->vdev_notrim = B_TRUE;
 903                         break;
 904                 }
 905                 break;
 906         case ENXIO:
 907                 if (!vd->vdev_remove_wanted) {
 908                         /*
 909                          * If provider's error is set we assume it is being
 910                          * removed.
 911                          */
 912                         if (bp->bio_to->error != 0) {
 913                                 vd->vdev_remove_wanted = B_TRUE;
 914                                 spa_async_request(zio->io_spa,
 915                                     SPA_ASYNC_REMOVE);
 916                         } else if (!vd->vdev_delayed_close) {
 917                                 vd->vdev_delayed_close = B_TRUE;
 918                         }
 919                 }
 920                 break;
 921         }
 922         g_destroy_bio(bp);
 923         zio_delay_interrupt(zio);
 924 }
 925
 926 static void
 927 vdev_geom_io_start(zio_t *zio)
 928 {
 929         vdev_t *vd;
 930         struct g_consumer *cp;
 931         struct bio *bp;
 932         int error;
 933
 934         vd = zio->io_vd;
 935
 936         switch (zio->io_type) {
 937         case ZIO_TYPE_IOCTL:
 938                 /* XXPOLICY */
 939                 if (!vdev_readable(vd)) {
 940                         zio->io_error = SET_ERROR(ENXIO);
 941                         zio_interrupt(zio);
 942                         return;
 943                 } else {
 944                         switch (zio->io_cmd) {
 945                         case DKIOCFLUSHWRITECACHE:
 946                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 947                                         break;
 948                                 if (vd->vdev_nowritecache) {
 949                                         zio->io_error = SET_ERROR(ENOTSUP);
 950                                         break;
 951                                 }
 952                                 goto sendreq;
 953                         default:
 954                                 zio->io_error = SET_ERROR(ENOTSUP);
 955                         }
 956                 }
 957
 958                 zio_execute(zio);
 959                 return;
 960         case ZIO_TYPE_FREE:
 961                 if (vd->vdev_notrim) {
 962                         zio->io_error = SET_ERROR(ENOTSUP);
 963                 } else if (!vdev_geom_bio_delete_disable) {
 964                         goto sendreq;
 965                 }
 966                 zio_execute(zio);
 967                 return;
 968         }
 969 sendreq:
 970         ASSERT(zio->io_type == ZIO_TYPE_READ ||
 971             zio->io_type == ZIO_TYPE_WRITE ||
 972             zio->io_type == ZIO_TYPE_FREE ||
 973             zio->io_type == ZIO_TYPE_IOCTL);
 974
 975         cp = vd->vdev_tsd;
 976         if (cp == NULL) {
 977                 zio->io_error = SET_ERROR(ENXIO);
 978                 zio_interrupt(zio);
 979                 return;
 980         }
 981         bp = g_alloc_bio();
 982         bp->bio_caller1 = zio;
 983         switch (zio->io_type) {
 984         case ZIO_TYPE_READ:
 985         case ZIO_TYPE_WRITE:
 986                 zio->io_target_timestamp = zio_handle_io_delay(zio);
 987                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 988                 bp->bio_data = zio->io_data;
 989                 bp->bio_offset = zio->io_offset;
 990                 bp->bio_length = zio->io_size;
 991                 break;
 992         case ZIO_TYPE_FREE:
 993                 bp->bio_cmd = BIO_DELETE;
 994                 bp->bio_data = NULL;
 995                 bp->bio_offset = zio->io_offset;
 996                 bp->bio_length = zio->io_size;
 997                 break;
 998         case ZIO_TYPE_IOCTL:
 999                 bp->bio_cmd = BIO_FLUSH;
1000                 bp->bio_flags |= BIO_ORDERED;
1001                 bp->bio_data = NULL;
1002                 bp->bio_offset = cp->provider->mediasize;
1003                 bp->bio_length = 0;
1004                 break;
1005         }
1006         bp->bio_done = vdev_geom_io_intr;
1007
1008         g_io_request(bp, cp);
1009 }
1010
1011 static void
1012 vdev_geom_io_done(zio_t *zio)
1013 {
1014 }
1015
1016 static void
1017 vdev_geom_hold(vdev_t *vd)
1018 {
1019 }
1020
1021 static void
1022 vdev_geom_rele(vdev_t *vd)
1023 {
1024 }
1025
1026 vdev_ops_t vdev_geom_ops = {
1027         vdev_geom_open,
1028         vdev_geom_close,
1029         vdev_default_asize,
1030         vdev_geom_io_start,
1031         vdev_geom_io_done,
1032         NULL,
1033         vdev_geom_hold,
1034         vdev_geom_rele,
1035         VDEV_TYPE_DISK,         /* name of this vdev type */
1036         B_TRUE                  /* leaf vdev */
1037 };