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