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         /* Fetch initial physical path information for this device. */
 245         vdev_geom_attrchanged(cp, "GEOM::physpath");
 246
 247         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
 248         return (cp);
 249 }
 250
 251 static void
 252 vdev_geom_close_locked(vdev_t *vd)
 253 {
 254         struct g_geom *gp;
 255         struct g_consumer *cp;
 256
 257         g_topology_assert();
 258
 259         cp = vd->vdev_tsd;
 260         if (cp == NULL)
 261                 return;
 262
 263         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 264         KASSERT(vd->vdev_tsd == cp, ("%s: vdev_tsd is not cp", __func__));
 265         vd->vdev_tsd = NULL;
 266         vd->vdev_delayed_close = B_FALSE;
 267         cp->private = NULL;
 268
 269         gp = cp->geom;
 270         g_access(cp, -1, 0, -1);
 271         /* Destroy consumer on last close. */
 272         if (cp->acr == 0 && cp->ace == 0) {
 273                 if (cp->acw > 0)
 274                         g_access(cp, 0, -cp->acw, 0);
 275                 if (cp->provider != NULL) {
 276                         ZFS_LOG(1, "Destroyed consumer to %s.",
 277                             cp->provider->name);
 278                         g_detach(cp);
 279                 }
 280                 g_destroy_consumer(cp);
 281         }
 282         /* Destroy geom if there are no consumers left. */
 283         if (LIST_EMPTY(&gp->consumer)) {
 284                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 285                 g_wither_geom(gp, ENXIO);
 286         }
 287 }
 288
 289 static void
 290 nvlist_get_guids(nvlist_t *list, uint64_t *pguid, uint64_t *vguid)
 291 {
 292
 293         (void) nvlist_lookup_uint64(list, ZPOOL_CONFIG_GUID, vguid);
 294         (void) nvlist_lookup_uint64(list, ZPOOL_CONFIG_POOL_GUID, pguid);
 295 }
 296
 297 static int
 298 vdev_geom_io(struct g_consumer *cp, int cmd, void *data, off_t offset, off_t size)
 299 {
 300         struct bio *bp;
 301         u_char *p;
 302         off_t off, maxio;
 303         int error;
 304
 305         ASSERT((offset % cp->provider->sectorsize) == 0);
 306         ASSERT((size % cp->provider->sectorsize) == 0);
 307
 308         bp = g_alloc_bio();
 309         off = offset;
 310         offset += size;
 311         p = data;
 312         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 313         error = 0;
 314
 315         for (; off < offset; off += maxio, p += maxio, size -= maxio) {
 316                 bzero(bp, sizeof(*bp));
 317                 bp->bio_cmd = cmd;
 318                 bp->bio_done = NULL;
 319                 bp->bio_offset = off;
 320                 bp->bio_length = MIN(size, maxio);
 321                 bp->bio_data = p;
 322                 g_io_request(bp, cp);
 323                 error = biowait(bp, "vdev_geom_io");
 324                 if (error != 0)
 325                         break;
 326         }
 327
 328         g_destroy_bio(bp);
 329         return (error);
 330 }
 331
 332 static void
 333 vdev_geom_taste_orphan(struct g_consumer *cp)
 334 {
 335
 336         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
 337             cp->provider->name));
 338 }
 339
 340 static int
 341 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 342 {
 343         struct g_provider *pp;
 344         vdev_label_t *label;
 345         char *p, *buf;
 346         size_t buflen;
 347         uint64_t psize;
 348         off_t offset, size;
 349         uint64_t state, txg;
 350         int error, l, len;
 351
 352         g_topology_assert_not();
 353
 354         pp = cp->provider;
 355         ZFS_LOG(1, "Reading config from %s...", pp->name);
 356
 357         psize = pp->mediasize;
 358         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 359
 360         size = sizeof(*label) + pp->sectorsize -
 361             ((sizeof(*label) - 1) % pp->sectorsize) - 1;
 362
 363         label = kmem_alloc(size, KM_SLEEP);
 364         buflen = sizeof(label->vl_vdev_phys.vp_nvlist);
 365
 366         *config = NULL;
 367         for (l = 0; l < VDEV_LABELS; l++) {
 368
 369                 offset = vdev_label_offset(psize, l, 0);
 370                 if ((offset % pp->sectorsize) != 0)
 371                         continue;
 372
 373                 if (vdev_geom_io(cp, BIO_READ, label, offset, size) != 0)
 374                         continue;
 375                 buf = label->vl_vdev_phys.vp_nvlist;
 376
 377                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 378                         continue;
 379
 380                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 381                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 382                         nvlist_free(*config);
 383                         *config = NULL;
 384                         continue;
 385                 }
 386
 387                 if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
 388                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 389                     &txg) != 0 || txg == 0)) {
 390                         nvlist_free(*config);
 391                         *config = NULL;
 392                         continue;
 393                 }
 394
 395                 break;
 396         }
 397
 398         kmem_free(label, size);
 399         return (*config == NULL ? ENOENT : 0);
 400 }
 401
 402 static void
 403 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 404 {
 405         nvlist_t **new_configs;
 406         uint64_t i;
 407
 408         if (id < *count)
 409                 return;
 410         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 411             KM_SLEEP);
 412         for (i = 0; i < *count; i++)
 413                 new_configs[i] = (*configs)[i];
 414         if (*configs != NULL)
 415                 kmem_free(*configs, *count * sizeof(void *));
 416         *configs = new_configs;
 417         *count = id + 1;
 418 }
 419
 420 static void
 421 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 422     const char *name, uint64_t* known_pool_guid)
 423 {
 424         nvlist_t *vdev_tree;
 425         uint64_t pool_guid;
 426         uint64_t vdev_guid, known_guid;
 427         uint64_t id, txg, known_txg;
 428         char *pname;
 429         int i;
 430
 431         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 432             strcmp(pname, name) != 0)
 433                 goto ignore;
 434
 435         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 436                 goto ignore;
 437
 438         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 439                 goto ignore;
 440
 441         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 442                 goto ignore;
 443
 444         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 445                 goto ignore;
 446
 447         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 448
 449         if (*known_pool_guid != 0) {
 450                 if (pool_guid != *known_pool_guid)
 451                         goto ignore;
 452         } else
 453                 *known_pool_guid = pool_guid;
 454
 455         resize_configs(configs, count, id);
 456
 457         if ((*configs)[id] != NULL) {
 458                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 459                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 460                 if (txg <= known_txg)
 461                         goto ignore;
 462                 nvlist_free((*configs)[id]);
 463         }
 464
 465         (*configs)[id] = cfg;
 466         return;
 467
 468 ignore:
 469         nvlist_free(cfg);
 470 }
 471
 472 static int
 473 vdev_geom_attach_taster(struct g_consumer *cp, struct g_provider *pp)
 474 {
 475         int error;
 476
 477         if (pp->flags & G_PF_WITHER)
 478                 return (EINVAL);
 479         g_attach(cp, pp);
 480         error = g_access(cp, 1, 0, 0);
 481         if (error == 0) {
 482                 if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize))
 483                         error = EINVAL;
 484                 else if (pp->mediasize < SPA_MINDEVSIZE)
 485                         error = EINVAL;
 486                 if (error != 0)
 487                         g_access(cp, -1, 0, 0);
 488         }
 489         if (error != 0)
 490                 g_detach(cp);
 491         return (error);
 492 }
 493
 494 static void
 495 vdev_geom_detach_taster(struct g_consumer *cp)
 496 {
 497         g_access(cp, -1, 0, 0);
 498         g_detach(cp);
 499 }
 500
 501 int
 502 vdev_geom_read_pool_label(const char *name,
 503     nvlist_t ***configs, uint64_t *count)
 504 {
 505         struct g_class *mp;
 506         struct g_geom *gp, *zgp;
 507         struct g_provider *pp;
 508         struct g_consumer *zcp;
 509         nvlist_t *vdev_cfg;
 510         uint64_t pool_guid;
 511         int error;
 512
 513         DROP_GIANT();
 514         g_topology_lock();
 515
 516         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 517         /* This orphan function should be never called. */
 518         zgp->orphan = vdev_geom_taste_orphan;
 519         zcp = g_new_consumer(zgp);
 520
 521         *configs = NULL;
 522         *count = 0;
 523         pool_guid = 0;
 524         LIST_FOREACH(mp, &g_classes, class) {
 525                 if (mp == &zfs_vdev_class)
 526                         continue;
 527                 LIST_FOREACH(gp, &mp->geom, geom) {
 528                         if (gp->flags & G_GEOM_WITHER)
 529                                 continue;
 530                         LIST_FOREACH(pp, &gp->provider, provider) {
 531                                 if (pp->flags & G_PF_WITHER)
 532                                         continue;
 533                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 534                                         continue;
 535                                 g_topology_unlock();
 536                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 537                                 g_topology_lock();
 538                                 vdev_geom_detach_taster(zcp);
 539                                 if (error)
 540                                         continue;
 541                                 ZFS_LOG(1, "successfully read vdev config");
 542
 543                                 process_vdev_config(configs, count,
 544                                     vdev_cfg, name, &pool_guid);
 545                         }
 546                 }
 547         }
 548
 549         g_destroy_consumer(zcp);
 550         g_destroy_geom(zgp);
 551         g_topology_unlock();
 552         PICKUP_GIANT();
 553
 554         return (*count > 0 ? 0 : ENOENT);
 555 }
 556
 557 static void
 558 vdev_geom_read_guids(struct g_consumer *cp, uint64_t *pguid, uint64_t *vguid)
 559 {
 560         nvlist_t *config;
 561
 562         g_topology_assert_not();
 563
 564         *pguid = 0;
 565         *vguid = 0;
 566         if (vdev_geom_read_config(cp, &config) == 0) {
 567                 nvlist_get_guids(config, pguid, vguid);
 568                 nvlist_free(config);
 569         }
 570 }
 571
 572 static struct g_consumer *
 573 vdev_geom_attach_by_guids(vdev_t *vd)
 574 {
 575         struct g_class *mp;
 576         struct g_geom *gp, *zgp;
 577         struct g_provider *pp;
 578         struct g_consumer *cp, *zcp;
 579         uint64_t pguid, vguid;
 580
 581         g_topology_assert();
 582
 583         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 584         /* This orphan function should be never called. */
 585         zgp->orphan = vdev_geom_taste_orphan;
 586         zcp = g_new_consumer(zgp);
 587
 588         cp = NULL;
 589         LIST_FOREACH(mp, &g_classes, class) {
 590                 if (mp == &zfs_vdev_class)
 591                         continue;
 592                 LIST_FOREACH(gp, &mp->geom, geom) {
 593                         if (gp->flags & G_GEOM_WITHER)
 594                                 continue;
 595                         LIST_FOREACH(pp, &gp->provider, provider) {
 596                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 597                                         continue;
 598                                 g_topology_unlock();
 599                                 vdev_geom_read_guids(zcp, &pguid, &vguid);
 600                                 g_topology_lock();
 601                                 vdev_geom_detach_taster(zcp);
 602                                 /*
 603                                  * Check that the label's vdev guid matches the
 604                                  * desired guid.  If the label has a pool guid,
 605                                  * check that it matches too. (Inactive spares
 606                                  * and L2ARCs do not have any pool guid in the
 607                                  * label.)
 608                                 */
 609                                 if ((pguid != 0 &&
 610                                      pguid != spa_guid(vd->vdev_spa)) ||
 611                                     vguid != vd->vdev_guid)
 612                                         continue;
 613                                 cp = vdev_geom_attach(pp, vd);
 614                                 if (cp == NULL) {
 615                                         printf("ZFS WARNING: Unable to "
 616                                             "attach to %s.\n", pp->name);
 617                                         continue;
 618                                 }
 619                                 break;
 620                         }
 621                         if (cp != NULL)
 622                                 break;
 623                 }
 624                 if (cp != NULL)
 625                         break;
 626         }
 627 end:
 628         g_destroy_consumer(zcp);
 629         g_destroy_geom(zgp);
 630         return (cp);
 631 }
 632
 633 static struct g_consumer *
 634 vdev_geom_open_by_guids(vdev_t *vd)
 635 {
 636         struct g_consumer *cp;
 637         char *buf;
 638         size_t len;
 639
 640         g_topology_assert();
 641
 642         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 643                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 644         cp = vdev_geom_attach_by_guids(vd);
 645         if (cp != NULL) {
 646                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 647                 buf = kmem_alloc(len, KM_SLEEP);
 648
 649                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 650                 spa_strfree(vd->vdev_path);
 651                 vd->vdev_path = buf;
 652
 653                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 654                     (uintmax_t)spa_guid(vd->vdev_spa),
 655                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 656         } else {
 657                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 658                     (uintmax_t)spa_guid(vd->vdev_spa),
 659                     (uintmax_t)vd->vdev_guid);
 660         }
 661
 662         return (cp);
 663 }
 664
 665 static struct g_consumer *
 666 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 667 {
 668         struct g_provider *pp;
 669         struct g_consumer *cp;
 670         uint64_t pguid, vguid;
 671
 672         g_topology_assert();
 673
 674         cp = NULL;
 675         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 676         if (pp != NULL) {
 677                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 678                 cp = vdev_geom_attach(pp, vd);
 679                 if (cp != NULL && check_guid && ISP2(pp->sectorsize) &&
 680                     pp->sectorsize <= VDEV_PAD_SIZE) {
 681                         g_topology_unlock();
 682                         vdev_geom_read_guids(cp, &pguid, &vguid);
 683                         g_topology_lock();
 684                         /*
 685                          * Check that the label's vdev guid matches the
 686                          * desired guid.  If the label has a pool guid,
 687                          * check that it matches too. (Inactive spares
 688                          * and L2ARCs do not have any pool guid in the
 689                          * label.)
 690                          */
 691                         if ((pguid != 0 &&
 692                             pguid != spa_guid(vd->vdev_spa)) ||
 693                             vguid != vd->vdev_guid) {
 694                                 vdev_geom_close_locked(vd);
 695                                 cp = NULL;
 696                                 ZFS_LOG(1, "guid mismatch for provider %s: "
 697                                     "%ju:%ju != %ju:%ju.", vd->vdev_path,
 698                                     (uintmax_t)spa_guid(vd->vdev_spa),
 699                                     (uintmax_t)vd->vdev_guid,
 700                                     (uintmax_t)pguid, (uintmax_t)vguid);
 701                         } else {
 702                                 ZFS_LOG(1, "guid match for provider %s.",
 703                                     vd->vdev_path);
 704                         }
 705                 }
 706         }
 707
 708         return (cp);
 709 }
 710
 711 static int
 712 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 713     uint64_t *logical_ashift, uint64_t *physical_ashift)
 714 {
 715         struct g_provider *pp;
 716         struct g_consumer *cp;
 717         size_t bufsize;
 718         int error;
 719
 720         /*
 721          * We must have a pathname, and it must be absolute.
 722          */
 723         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 724                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 725                 return (EINVAL);
 726         }
 727
 728         vd->vdev_tsd = NULL;
 729
 730         DROP_GIANT();
 731         g_topology_lock();
 732         error = 0;
 733
 734         if (vd->vdev_spa->spa_splitting_newspa ||
 735             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 736              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE)) {
 737                 /*
 738                  * We are dealing with a vdev that hasn't been previously
 739                  * opened (since boot), and we are not loading an
 740                  * existing pool configuration.  This looks like a
 741                  * vdev add operation to a new or existing pool.
 742                  * Assume the user knows what he/she is doing and find
 743                  * GEOM provider by its name, ignoring GUID mismatches.
 744                  *
 745                  * XXPOLICY: It would be safer to only allow a device
 746                  *           that is unlabeled or labeled but missing
 747                  *           GUID information to be opened in this fashion,
 748                  *           unless we are doing a split, in which case we
 749                  *           should allow any guid.
 750                  */
 751                 cp = vdev_geom_open_by_path(vd, 0);
 752         } else {
 753                 /*
 754                  * Try using the recorded path for this device, but only
 755                  * accept it if its label data contains the expected GUIDs.
 756                  */
 757                 cp = vdev_geom_open_by_path(vd, 1);
 758                 if (cp == NULL) {
 759                         /*
 760                          * The device at vd->vdev_path doesn't have the
 761                          * expected GUIDs. The disks might have merely
 762                          * moved around so try all other GEOM providers
 763                          * to find one with the right GUIDs.
 764                          */
 765                         cp = vdev_geom_open_by_guids(vd);
 766                 }
 767         }
 768
 769         if (cp == NULL) {
 770                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 771                 error = ENOENT;
 772         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 773             !ISP2(cp->provider->sectorsize)) {
 774                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 775                     vd->vdev_path);
 776
 777                 vdev_geom_close_locked(vd);
 778                 error = EINVAL;
 779                 cp = NULL;
 780         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 781                 int i;
 782
 783                 for (i = 0; i < 5; i++) {
 784                         error = g_access(cp, 0, 1, 0);
 785                         if (error == 0)
 786                                 break;
 787                         g_topology_unlock();
 788                         tsleep(vd, 0, "vdev", hz / 2);
 789                         g_topology_lock();
 790                 }
 791                 if (error != 0) {
 792                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 793                             vd->vdev_path, error);
 794                         vdev_geom_close_locked(vd);
 795                         cp = NULL;
 796                 }
 797         }
 798
 799         g_topology_unlock();
 800         PICKUP_GIANT();
 801         if (cp == NULL) {
 802                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 803                 return (error);
 804         }
 805         pp = cp->provider;
 806
 807         /*
 808          * Determine the actual size of the device.
 809          */
 810         *max_psize = *psize = pp->mediasize;
 811
 812         /*
 813          * Determine the device's minimum transfer size and preferred
 814          * transfer size.
 815          */
 816         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 817         *physical_ashift = 0;
 818         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 819             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 820                 *physical_ashift = highbit(pp->stripesize) - 1;
 821
 822         /*
 823          * Clear the nowritecache settings, so that on a vdev_reopen()
 824          * we will try again.
 825          */
 826         vd->vdev_nowritecache = B_FALSE;
 827
 828         /*
 829          * Determine the device's rotation rate.
 830          */
 831         vdev_geom_set_rotation_rate(vd, cp);
 832
 833         return (0);
 834 }
 835
 836 static void
 837 vdev_geom_close(vdev_t *vd)
 838 {
 839
 840         DROP_GIANT();
 841         g_topology_lock();
 842         vdev_geom_close_locked(vd);
 843         g_topology_unlock();
 844         PICKUP_GIANT();
 845 }
 846
 847 static void
 848 vdev_geom_io_intr(struct bio *bp)
 849 {
 850         vdev_t *vd;
 851         zio_t *zio;
 852
 853         zio = bp->bio_caller1;
 854         vd = zio->io_vd;
 855         zio->io_error = bp->bio_error;
 856         if (zio->io_error == 0 && bp->bio_resid != 0)
 857                 zio->io_error = SET_ERROR(EIO);
 858
 859         switch(zio->io_error) {
 860         case ENOTSUP:
 861                 /*
 862                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 863                  * that future attempts will never succeed. In this case
 864                  * we set a persistent flag so that we don't bother with
 865                  * requests in the future.
 866                  */
 867                 switch(bp->bio_cmd) {
 868                 case BIO_FLUSH:
 869                         vd->vdev_nowritecache = B_TRUE;
 870                         break;
 871                 case BIO_DELETE:
 872                         vd->vdev_notrim = B_TRUE;
 873                         break;
 874                 }
 875                 break;
 876         case ENXIO:
 877                 if (!vd->vdev_remove_wanted) {
 878                         /*
 879                          * If provider's error is set we assume it is being
 880                          * removed.
 881                          */
 882                         if (bp->bio_to->error != 0) {
 883                                 vd->vdev_remove_wanted = B_TRUE;
 884                                 spa_async_request(zio->io_spa,
 885                                     SPA_ASYNC_REMOVE);
 886                         } else if (!vd->vdev_delayed_close) {
 887                                 vd->vdev_delayed_close = B_TRUE;
 888                         }
 889                 }
 890                 break;
 891         }
 892         g_destroy_bio(bp);
 893         zio_delay_interrupt(zio);
 894 }
 895
 896 static void
 897 vdev_geom_io_start(zio_t *zio)
 898 {
 899         vdev_t *vd;
 900         struct g_consumer *cp;
 901         struct bio *bp;
 902         int error;
 903
 904         vd = zio->io_vd;
 905
 906         switch (zio->io_type) {
 907         case ZIO_TYPE_IOCTL:
 908                 /* XXPOLICY */
 909                 if (!vdev_readable(vd)) {
 910                         zio->io_error = SET_ERROR(ENXIO);
 911                         zio_interrupt(zio);
 912                         return;
 913                 } else {
 914                         switch (zio->io_cmd) {
 915                         case DKIOCFLUSHWRITECACHE:
 916                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 917                                         break;
 918                                 if (vd->vdev_nowritecache) {
 919                                         zio->io_error = SET_ERROR(ENOTSUP);
 920                                         break;
 921                                 }
 922                                 goto sendreq;
 923                         default:
 924                                 zio->io_error = SET_ERROR(ENOTSUP);
 925                         }
 926                 }
 927
 928                 zio_execute(zio);
 929                 return;
 930         case ZIO_TYPE_FREE:
 931                 if (vd->vdev_notrim) {
 932                         zio->io_error = SET_ERROR(ENOTSUP);
 933                 } else if (!vdev_geom_bio_delete_disable) {
 934                         goto sendreq;
 935                 }
 936                 zio_execute(zio);
 937                 return;
 938         }
 939 sendreq:
 940         ASSERT(zio->io_type == ZIO_TYPE_READ ||
 941             zio->io_type == ZIO_TYPE_WRITE ||
 942             zio->io_type == ZIO_TYPE_FREE ||
 943             zio->io_type == ZIO_TYPE_IOCTL);
 944
 945         cp = vd->vdev_tsd;
 946         if (cp == NULL) {
 947                 zio->io_error = SET_ERROR(ENXIO);
 948                 zio_interrupt(zio);
 949                 return;
 950         }
 951         bp = g_alloc_bio();
 952         bp->bio_caller1 = zio;
 953         switch (zio->io_type) {
 954         case ZIO_TYPE_READ:
 955         case ZIO_TYPE_WRITE:
 956                 zio->io_target_timestamp = zio_handle_io_delay(zio);
 957                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 958                 bp->bio_data = zio->io_data;
 959                 bp->bio_offset = zio->io_offset;
 960                 bp->bio_length = zio->io_size;
 961                 break;
 962         case ZIO_TYPE_FREE:
 963                 bp->bio_cmd = BIO_DELETE;
 964                 bp->bio_data = NULL;
 965                 bp->bio_offset = zio->io_offset;
 966                 bp->bio_length = zio->io_size;
 967                 break;
 968         case ZIO_TYPE_IOCTL:
 969                 bp->bio_cmd = BIO_FLUSH;
 970                 bp->bio_flags |= BIO_ORDERED;
 971                 bp->bio_data = NULL;
 972                 bp->bio_offset = cp->provider->mediasize;
 973                 bp->bio_length = 0;
 974                 break;
 975         }
 976         bp->bio_done = vdev_geom_io_intr;
 977
 978         g_io_request(bp, cp);
 979 }
 980
 981 static void
 982 vdev_geom_io_done(zio_t *zio)
 983 {
 984 }
 985
 986 static void
 987 vdev_geom_hold(vdev_t *vd)
 988 {
 989 }
 990
 991 static void
 992 vdev_geom_rele(vdev_t *vd)
 993 {
 994 }
 995
 996 vdev_ops_t vdev_geom_ops = {
 997         vdev_geom_open,
 998         vdev_geom_close,
 999         vdev_default_asize,
1000         vdev_geom_io_start,
1001         vdev_geom_io_done,
1002         NULL,
1003         vdev_geom_hold,
1004         vdev_geom_rele,
1005         VDEV_TYPE_DISK,         /* name of this vdev type */
1006         B_TRUE                  /* leaf vdev */
1007 };