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 /* Declare local functions */
  67 static void vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read);
  68
  69 /*
  70  * Thread local storage used to indicate when a thread is probing geoms
  71  * for their guids.  If NULL, this thread is not tasting geoms.  If non NULL,
  72  * it is looking for a replacement for the vdev_t* that is its value.
  73  */
  74 uint_t zfs_geom_probe_vdev_key;
  75
  76 static void
  77 vdev_geom_set_rotation_rate(vdev_t *vd, struct g_consumer *cp)
  78 {
  79         int error;
  80         uint16_t rate;
  81
  82         error = g_getattr("GEOM::rotation_rate", cp, &rate);
  83         if (error == 0)
  84                 vd->vdev_rotation_rate = rate;
  85         else
  86                 vd->vdev_rotation_rate = VDEV_RATE_UNKNOWN;
  87 }
  88
  89 static void
  90 vdev_geom_set_physpath(struct g_consumer *cp, boolean_t do_null_update)
  91 {
  92         boolean_t needs_update = B_FALSE;
  93         vdev_t *vd;
  94         char *physpath;
  95         int error, physpath_len;
  96
  97         if (g_access(cp, 1, 0, 0) != 0)
  98                 return;
  99
 100         vd = cp->private;
 101         physpath_len = MAXPATHLEN;
 102         physpath = g_malloc(physpath_len, M_WAITOK|M_ZERO);
 103         error = g_io_getattr("GEOM::physpath", cp, &physpath_len, physpath);
 104         g_access(cp, -1, 0, 0);
 105         if (error == 0) {
 106                 char *old_physpath;
 107
 108                 /* g_topology lock ensures that vdev has not been closed */
 109                 g_topology_assert();
 110                 old_physpath = vd->vdev_physpath;
 111                 vd->vdev_physpath = spa_strdup(physpath);
 112
 113                 if (old_physpath != NULL) {
 114                         needs_update = (strcmp(old_physpath,
 115                                                 vd->vdev_physpath) != 0);
 116                         spa_strfree(old_physpath);
 117                 } else
 118                         needs_update = do_null_update;
 119         }
 120         g_free(physpath);
 121
 122         /*
 123          * If the physical path changed, update the config.
 124          * Only request an update for previously unset physpaths if
 125          * requested by the caller.
 126          */
 127         if (needs_update)
 128                 spa_async_request(vd->vdev_spa, SPA_ASYNC_CONFIG_UPDATE);
 129
 130 }
 131
 132 static void
 133 vdev_geom_attrchanged(struct g_consumer *cp, const char *attr)
 134 {
 135         vdev_t *vd;
 136         char *old_physpath;
 137         int error;
 138
 139         vd = cp->private;
 140         if (vd == NULL)
 141                 return;
 142
 143         if (strcmp(attr, "GEOM::rotation_rate") == 0) {
 144                 vdev_geom_set_rotation_rate(vd, cp);
 145                 return;
 146         }
 147
 148         if (strcmp(attr, "GEOM::physpath") == 0) {
 149                 vdev_geom_set_physpath(cp, /*do_null_update*/B_TRUE);
 150                 return;
 151         }
 152 }
 153
 154 static void
 155 vdev_geom_orphan(struct g_consumer *cp)
 156 {
 157         vdev_t *vd;
 158
 159         g_topology_assert();
 160
 161         vd = cp->private;
 162         if (vd == NULL) {
 163                 /* Vdev close in progress.  Ignore the event. */
 164                 return;
 165         }
 166
 167         /*
 168          * Orphan callbacks occur from the GEOM event thread.
 169          * Concurrent with this call, new I/O requests may be
 170          * working their way through GEOM about to find out
 171          * (only once executed by the g_down thread) that we've
 172          * been orphaned from our disk provider.  These I/Os
 173          * must be retired before we can detach our consumer.
 174          * This is most easily achieved by acquiring the
 175          * SPA ZIO configuration lock as a writer, but doing
 176          * so with the GEOM topology lock held would cause
 177          * a lock order reversal.  Instead, rely on the SPA's
 178          * async removal support to invoke a close on this
 179          * vdev once it is safe to do so.
 180          */
 181         vd->vdev_remove_wanted = B_TRUE;
 182         spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
 183 }
 184
 185 static struct g_consumer *
 186 vdev_geom_attach(struct g_provider *pp, vdev_t *vd)
 187 {
 188         struct g_geom *gp;
 189         struct g_consumer *cp;
 190         int error;
 191
 192         g_topology_assert();
 193
 194         ZFS_LOG(1, "Attaching to %s.", pp->name);
 195
 196         if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize)) {
 197                 ZFS_LOG(1, "Failing attach of %s. Incompatible sectorsize %d\n",
 198                     pp->name, pp->sectorsize);
 199                 return (NULL);
 200         } else if (pp->mediasize < SPA_MINDEVSIZE) {
 201                 ZFS_LOG(1, "Failing attach of %s. Incompatible mediasize %ju\n",
 202                     pp->name, pp->mediasize);
 203                 return (NULL);
 204         }
 205
 206         /* Do we have geom already? No? Create one. */
 207         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 208                 if (gp->flags & G_GEOM_WITHER)
 209                         continue;
 210                 if (strcmp(gp->name, "zfs::vdev") != 0)
 211                         continue;
 212                 break;
 213         }
 214         if (gp == NULL) {
 215                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 216                 gp->orphan = vdev_geom_orphan;
 217                 gp->attrchanged = vdev_geom_attrchanged;
 218                 cp = g_new_consumer(gp);
 219                 error = g_attach(cp, pp);
 220                 if (error != 0) {
 221                         ZFS_LOG(1, "%s(%d): g_attach failed: %d\n", __func__,
 222                             __LINE__, error);
 223                         vdev_geom_detach(cp, B_FALSE);
 224                         return (NULL);
 225                 }
 226                 error = g_access(cp, 1, 0, 1);
 227                 if (error != 0) {
 228                         ZFS_LOG(1, "%s(%d): g_access failed: %d\n", __func__,
 229                                __LINE__, error);
 230                         vdev_geom_detach(cp, B_FALSE);
 231                         return (NULL);
 232                 }
 233                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 234         } else {
 235                 /* Check if we are already connected to this provider. */
 236                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 237                         if (cp->provider == pp) {
 238                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 239                                 break;
 240                         }
 241                 }
 242                 if (cp == NULL) {
 243                         cp = g_new_consumer(gp);
 244                         error = g_attach(cp, pp);
 245                         if (error != 0) {
 246                                 ZFS_LOG(1, "%s(%d): g_attach failed: %d\n",
 247                                     __func__, __LINE__, error);
 248                                 vdev_geom_detach(cp, B_FALSE);
 249                                 return (NULL);
 250                         }
 251                         error = g_access(cp, 1, 0, 1);
 252                         if (error != 0) {
 253                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 254                                     __func__, __LINE__, error);
 255                                 vdev_geom_detach(cp, B_FALSE);
 256                                 return (NULL);
 257                         }
 258                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 259                 } else {
 260                         error = g_access(cp, 1, 0, 1);
 261                         if (error != 0) {
 262                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 263                                     __func__, __LINE__, error);
 264                                 return (NULL);
 265                         }
 266                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 267                 }
 268         }
 269
 270         /*
 271          * BUG: cp may already belong to a vdev.  This could happen if:
 272          * 1) That vdev is a shared spare, or
 273          * 2) We are trying to reopen a missing vdev and we are scanning by
 274          *    guid.  In that case, we'll ultimately fail to open this consumer,
 275          *    but not until after setting the private field.
 276          * The solution is to:
 277          * 1) Don't set the private field until after the open succeeds, and
 278          * 2) Set it to a linked list of vdevs, not just a single vdev
 279          */
 280         cp->private = vd;
 281         if (vd != NULL) {
 282                 vd->vdev_tsd = cp;
 283                 vdev_geom_set_physpath(cp, /*do_null_update*/B_FALSE);
 284         }
 285
 286         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
 287         return (cp);
 288 }
 289
 290 static void
 291 vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read)
 292 {
 293         struct g_geom *gp;
 294         vdev_t *vd;
 295
 296         g_topology_assert();
 297
 298         ZFS_LOG(1, "Detaching consumer. Provider %s.",
 299             cp->provider && cp->provider->name ? cp->provider->name : "NULL");
 300
 301         vd = cp->private;
 302         cp->private = NULL;
 303
 304         gp = cp->geom;
 305         if (open_for_read)
 306                 g_access(cp, -1, 0, -1);
 307         /* Destroy consumer on last close. */
 308         if (cp->acr == 0 && cp->ace == 0) {
 309                 if (cp->acw > 0)
 310                         g_access(cp, 0, -cp->acw, 0);
 311                 if (cp->provider != NULL) {
 312                         ZFS_LOG(1, "Destroying consumer to %s.",
 313                             cp->provider->name ? cp->provider->name : "NULL");
 314                         g_detach(cp);
 315                 }
 316                 g_destroy_consumer(cp);
 317         }
 318         /* Destroy geom if there are no consumers left. */
 319         if (LIST_EMPTY(&gp->consumer)) {
 320                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 321                 g_wither_geom(gp, ENXIO);
 322         }
 323 }
 324
 325 static void
 326 vdev_geom_close_locked(vdev_t *vd)
 327 {
 328         struct g_consumer *cp;
 329
 330         g_topology_assert();
 331
 332         cp = vd->vdev_tsd;
 333         vd->vdev_tsd = NULL;
 334         vd->vdev_delayed_close = B_FALSE;
 335         if (cp == NULL)
 336                 return;
 337
 338         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 339
 340         vdev_geom_detach(cp, B_TRUE);
 341 }
 342
 343 static void
 344 nvlist_get_guids(nvlist_t *list, uint64_t *pguid, uint64_t *vguid)
 345 {
 346
 347         (void) nvlist_lookup_uint64(list, ZPOOL_CONFIG_GUID, vguid);
 348         (void) nvlist_lookup_uint64(list, ZPOOL_CONFIG_POOL_GUID, pguid);
 349 }
 350
 351 /*
 352  * Issue one or more bios to the vdev in parallel
 353  * cmds, datas, offsets, errors, and sizes are arrays of length ncmds.  Each IO
 354  * operation is described by parallel entries from each array.  There may be
 355  * more bios actually issued than entries in the array
 356  */
 357 static void
 358 vdev_geom_io(struct g_consumer *cp, int *cmds, void **datas, off_t *offsets,
 359     off_t *sizes, int *errors, int ncmds)
 360 {
 361         struct bio **bios;
 362         u_char *p;
 363         off_t off, maxio, s, end;
 364         int i, n_bios, j;
 365         size_t bios_size;
 366
 367         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 368         n_bios = 0;
 369
 370         /* How many bios are required for all commands ? */
 371         for (i = 0; i < ncmds; i++)
 372                 n_bios += (sizes[i] + maxio - 1) / maxio;
 373
 374         /* Allocate memory for the bios */
 375         bios_size = n_bios * sizeof(struct bio*);
 376         bios = kmem_zalloc(bios_size, KM_SLEEP);
 377
 378         /* Prepare and issue all of the bios */
 379         for (i = j = 0; i < ncmds; i++) {
 380                 off = offsets[i];
 381                 p = datas[i];
 382                 s = sizes[i];
 383                 end = off + s;
 384                 ASSERT((off % cp->provider->sectorsize) == 0);
 385                 ASSERT((s % cp->provider->sectorsize) == 0);
 386
 387                 for (; off < end; off += maxio, p += maxio, s -= maxio, j++) {
 388                         bios[j] = g_alloc_bio();
 389                         bios[j]->bio_cmd = cmds[i];
 390                         bios[j]->bio_done = NULL;
 391                         bios[j]->bio_offset = off;
 392                         bios[j]->bio_length = MIN(s, maxio);
 393                         bios[j]->bio_data = p;
 394                         g_io_request(bios[j], cp);
 395                 }
 396         }
 397         ASSERT(j == n_bios);
 398
 399         /* Wait for all of the bios to complete, and clean them up */
 400         for (i = j = 0; i < ncmds; i++) {
 401                 off = offsets[i];
 402                 s = sizes[i];
 403                 end = off + s;
 404
 405                 for (; off < end; off += maxio, s -= maxio, j++) {
 406                         errors[i] = biowait(bios[j], "vdev_geom_io") || errors[i];
 407                         g_destroy_bio(bios[j]);
 408                 }
 409         }
 410         kmem_free(bios, bios_size);
 411 }
 412
 413 static int
 414 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 415 {
 416         struct g_provider *pp;
 417         vdev_phys_t *vdev_lists[VDEV_LABELS];
 418         char *p, *buf;
 419         size_t buflen;
 420         uint64_t psize, state, txg;
 421         off_t offsets[VDEV_LABELS];
 422         off_t size;
 423         off_t sizes[VDEV_LABELS];
 424         int cmds[VDEV_LABELS];
 425         int errors[VDEV_LABELS];
 426         int l, len;
 427
 428         g_topology_assert_not();
 429
 430         pp = cp->provider;
 431         ZFS_LOG(1, "Reading config from %s...", pp->name);
 432
 433         psize = pp->mediasize;
 434         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 435
 436         size = sizeof(*vdev_lists[0]) + pp->sectorsize -
 437             ((sizeof(*vdev_lists[0]) - 1) % pp->sectorsize) - 1;
 438
 439         buflen = sizeof(vdev_lists[0]->vp_nvlist);
 440
 441         *config = NULL;
 442         /* Create all of the IO requests */
 443         for (l = 0; l < VDEV_LABELS; l++) {
 444                 cmds[l] = BIO_READ;
 445                 vdev_lists[l] = kmem_alloc(size, KM_SLEEP);
 446                 offsets[l] = vdev_label_offset(psize, l, 0) + VDEV_SKIP_SIZE;
 447                 sizes[l] = size;
 448                 errors[l] = 0;
 449                 ASSERT(offsets[l] % pp->sectorsize == 0);
 450         }
 451
 452         /* Issue the IO requests */
 453         vdev_geom_io(cp, cmds, (void**)vdev_lists, offsets, sizes, errors,
 454             VDEV_LABELS);
 455
 456         /* Parse the labels */
 457         for (l = 0; l < VDEV_LABELS; l++) {
 458                 if (errors[l] != 0)
 459                         continue;
 460
 461                 buf = vdev_lists[l]->vp_nvlist;
 462
 463                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 464                         continue;
 465
 466                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 467                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 468                         nvlist_free(*config);
 469                         *config = NULL;
 470                         continue;
 471                 }
 472
 473                 if (state != POOL_STATE_SPARE &&
 474                     state != POOL_STATE_L2CACHE &&
 475                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 476                     &txg) != 0 || txg == 0)) {
 477                         nvlist_free(*config);
 478                         *config = NULL;
 479                         continue;
 480                 }
 481
 482                 break;
 483         }
 484
 485         /* Free the label storage */
 486         for (l = 0; l < VDEV_LABELS; l++)
 487                 kmem_free(vdev_lists[l], size);
 488
 489         return (*config == NULL ? ENOENT : 0);
 490 }
 491
 492 static void
 493 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 494 {
 495         nvlist_t **new_configs;
 496         uint64_t i;
 497
 498         if (id < *count)
 499                 return;
 500         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 501             KM_SLEEP);
 502         for (i = 0; i < *count; i++)
 503                 new_configs[i] = (*configs)[i];
 504         if (*configs != NULL)
 505                 kmem_free(*configs, *count * sizeof(void *));
 506         *configs = new_configs;
 507         *count = id + 1;
 508 }
 509
 510 static void
 511 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 512     const char *name, uint64_t* known_pool_guid)
 513 {
 514         nvlist_t *vdev_tree;
 515         uint64_t pool_guid;
 516         uint64_t vdev_guid, known_guid;
 517         uint64_t id, txg, known_txg;
 518         char *pname;
 519         int i;
 520
 521         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 522             strcmp(pname, name) != 0)
 523                 goto ignore;
 524
 525         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 526                 goto ignore;
 527
 528         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 529                 goto ignore;
 530
 531         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 532                 goto ignore;
 533
 534         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 535                 goto ignore;
 536
 537         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 538
 539         if (*known_pool_guid != 0) {
 540                 if (pool_guid != *known_pool_guid)
 541                         goto ignore;
 542         } else
 543                 *known_pool_guid = pool_guid;
 544
 545         resize_configs(configs, count, id);
 546
 547         if ((*configs)[id] != NULL) {
 548                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 549                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 550                 if (txg <= known_txg)
 551                         goto ignore;
 552                 nvlist_free((*configs)[id]);
 553         }
 554
 555         (*configs)[id] = cfg;
 556         return;
 557
 558 ignore:
 559         nvlist_free(cfg);
 560 }
 561
 562 int
 563 vdev_geom_read_pool_label(const char *name,
 564     nvlist_t ***configs, uint64_t *count)
 565 {
 566         struct g_class *mp;
 567         struct g_geom *gp;
 568         struct g_provider *pp;
 569         struct g_consumer *zcp;
 570         nvlist_t *vdev_cfg;
 571         uint64_t pool_guid;
 572         int error;
 573
 574         DROP_GIANT();
 575         g_topology_lock();
 576
 577         *configs = NULL;
 578         *count = 0;
 579         pool_guid = 0;
 580         LIST_FOREACH(mp, &g_classes, class) {
 581                 if (mp == &zfs_vdev_class)
 582                         continue;
 583                 LIST_FOREACH(gp, &mp->geom, geom) {
 584                         if (gp->flags & G_GEOM_WITHER)
 585                                 continue;
 586                         LIST_FOREACH(pp, &gp->provider, provider) {
 587                                 if (pp->flags & G_PF_WITHER)
 588                                         continue;
 589                                 zcp = vdev_geom_attach(pp, NULL);
 590                                 if (zcp == NULL)
 591                                         continue;
 592                                 g_topology_unlock();
 593                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 594                                 g_topology_lock();
 595                                 vdev_geom_detach(zcp, B_TRUE);
 596                                 if (error)
 597                                         continue;
 598                                 ZFS_LOG(1, "successfully read vdev config");
 599
 600                                 process_vdev_config(configs, count,
 601                                     vdev_cfg, name, &pool_guid);
 602                         }
 603                 }
 604         }
 605         g_topology_unlock();
 606         PICKUP_GIANT();
 607
 608         return (*count > 0 ? 0 : ENOENT);
 609 }
 610
 611 static void
 612 vdev_geom_read_guids(struct g_consumer *cp, uint64_t *pguid, uint64_t *vguid)
 613 {
 614         nvlist_t *config;
 615
 616         g_topology_assert_not();
 617
 618         *pguid = 0;
 619         *vguid = 0;
 620         if (vdev_geom_read_config(cp, &config) == 0) {
 621                 nvlist_get_guids(config, pguid, vguid);
 622                 nvlist_free(config);
 623         }
 624 }
 625
 626 static boolean_t
 627 vdev_attach_ok(vdev_t *vd, struct g_provider *pp)
 628 {
 629         uint64_t pool_guid;
 630         uint64_t vdev_guid;
 631         struct g_consumer *zcp;
 632         boolean_t pool_ok;
 633         boolean_t vdev_ok;
 634
 635         zcp = vdev_geom_attach(pp, NULL);
 636         if (zcp == NULL) {
 637                 ZFS_LOG(1, "Unable to attach tasting instance to %s.",
 638                     pp->name);
 639                 return (B_FALSE);
 640         }
 641         g_topology_unlock();
 642         vdev_geom_read_guids(zcp, &pool_guid, &vdev_guid);
 643         g_topology_lock();
 644         vdev_geom_detach(zcp, B_TRUE);
 645
 646         /*
 647          * Check that the label's vdev guid matches the desired guid.  If the
 648          * label has a pool guid, check that it matches too. (Inactive spares
 649          * and L2ARCs do not have any pool guid in the label.)
 650          */
 651         if ((pool_guid == 0 || pool_guid == spa_guid(vd->vdev_spa)) &&
 652             vdev_guid == vd->vdev_guid) {
 653                 ZFS_LOG(1, "guids match for provider %s.", vd->vdev_path);
 654                 return (B_TRUE);
 655         } else {
 656                 ZFS_LOG(1, "guid mismatch for provider %s: "
 657                     "%ju:%ju != %ju:%ju.", vd->vdev_path,
 658                     (uintmax_t)spa_guid(vd->vdev_spa),
 659                     (uintmax_t)vd->vdev_guid,
 660                     (uintmax_t)pool_guid, (uintmax_t)vdev_guid);
 661                 return (B_FALSE);
 662         }
 663 }
 664
 665 static struct g_consumer *
 666 vdev_geom_attach_by_guids(vdev_t *vd)
 667 {
 668         struct g_class *mp;
 669         struct g_geom *gp;
 670         struct g_provider *pp;
 671         struct g_consumer *cp;
 672
 673         g_topology_assert();
 674
 675         cp = NULL;
 676         LIST_FOREACH(mp, &g_classes, class) {
 677                 if (mp == &zfs_vdev_class)
 678                         continue;
 679                 LIST_FOREACH(gp, &mp->geom, geom) {
 680                         if (gp->flags & G_GEOM_WITHER)
 681                                 continue;
 682                         LIST_FOREACH(pp, &gp->provider, provider) {
 683                                 if (!vdev_attach_ok(vd, pp))
 684                                         continue;
 685                                 cp = vdev_geom_attach(pp, vd);
 686                                 if (cp == NULL) {
 687                                         printf("ZFS WARNING: Unable to "
 688                                             "attach to %s.\n", pp->name);
 689                                         continue;
 690                                 }
 691                                 break;
 692                         }
 693                         if (cp != NULL)
 694                                 break;
 695                 }
 696                 if (cp != NULL)
 697                         break;
 698         }
 699 end:
 700         return (cp);
 701 }
 702
 703 static struct g_consumer *
 704 vdev_geom_open_by_guids(vdev_t *vd)
 705 {
 706         struct g_consumer *cp;
 707         char *buf;
 708         size_t len;
 709
 710         g_topology_assert();
 711
 712         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 713                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 714         cp = vdev_geom_attach_by_guids(vd);
 715         if (cp != NULL) {
 716                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 717                 buf = kmem_alloc(len, KM_SLEEP);
 718
 719                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 720                 spa_strfree(vd->vdev_path);
 721                 vd->vdev_path = buf;
 722
 723                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 724                     (uintmax_t)spa_guid(vd->vdev_spa),
 725                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 726         } else {
 727                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 728                     (uintmax_t)spa_guid(vd->vdev_spa),
 729                     (uintmax_t)vd->vdev_guid);
 730         }
 731
 732         return (cp);
 733 }
 734
 735 static struct g_consumer *
 736 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 737 {
 738         struct g_provider *pp;
 739         struct g_consumer *cp;
 740
 741         g_topology_assert();
 742
 743         cp = NULL;
 744         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 745         if (pp != NULL) {
 746                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 747                 if (!check_guid || vdev_attach_ok(vd, pp))
 748                         cp = vdev_geom_attach(pp, vd);
 749         }
 750
 751         return (cp);
 752 }
 753
 754 static int
 755 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 756     uint64_t *logical_ashift, uint64_t *physical_ashift)
 757 {
 758         struct g_provider *pp;
 759         struct g_consumer *cp;
 760         size_t bufsize;
 761         int error;
 762
 763         /* Set the TLS to indicate downstack that we should not access zvols*/
 764         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 765
 766         /*
 767          * We must have a pathname, and it must be absolute.
 768          */
 769         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 770                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 771                 return (EINVAL);
 772         }
 773
 774         vd->vdev_tsd = NULL;
 775
 776         DROP_GIANT();
 777         g_topology_lock();
 778         error = 0;
 779
 780         if (vd->vdev_spa->spa_splitting_newspa ||
 781             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 782              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 783              vd->vdev_spa->spa_load_state == SPA_LOAD_CREATE)) {
 784                 /*
 785                  * We are dealing with a vdev that hasn't been previously
 786                  * opened (since boot), and we are not loading an
 787                  * existing pool configuration.  This looks like a
 788                  * vdev add operation to a new or existing pool.
 789                  * Assume the user knows what he/she is doing and find
 790                  * GEOM provider by its name, ignoring GUID mismatches.
 791                  *
 792                  * XXPOLICY: It would be safer to only allow a device
 793                  *           that is unlabeled or labeled but missing
 794                  *           GUID information to be opened in this fashion,
 795                  *           unless we are doing a split, in which case we
 796                  *           should allow any guid.
 797                  */
 798                 cp = vdev_geom_open_by_path(vd, 0);
 799         } else {
 800                 /*
 801                  * Try using the recorded path for this device, but only
 802                  * accept it if its label data contains the expected GUIDs.
 803                  */
 804                 cp = vdev_geom_open_by_path(vd, 1);
 805                 if (cp == NULL) {
 806                         /*
 807                          * The device at vd->vdev_path doesn't have the
 808                          * expected GUIDs. The disks might have merely
 809                          * moved around so try all other GEOM providers
 810                          * to find one with the right GUIDs.
 811                          */
 812                         cp = vdev_geom_open_by_guids(vd);
 813                 }
 814         }
 815
 816         /* Clear the TLS now that tasting is done */
 817         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 818
 819         if (cp == NULL) {
 820                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 821                 error = ENOENT;
 822         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 823             !ISP2(cp->provider->sectorsize)) {
 824                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 825                     vd->vdev_path);
 826
 827                 vdev_geom_close_locked(vd);
 828                 error = EINVAL;
 829                 cp = NULL;
 830         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 831                 int i;
 832
 833                 for (i = 0; i < 5; i++) {
 834                         error = g_access(cp, 0, 1, 0);
 835                         if (error == 0)
 836                                 break;
 837                         g_topology_unlock();
 838                         tsleep(vd, 0, "vdev", hz / 2);
 839                         g_topology_lock();
 840                 }
 841                 if (error != 0) {
 842                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 843                             vd->vdev_path, error);
 844                         vdev_geom_close_locked(vd);
 845                         cp = NULL;
 846                 }
 847         }
 848
 849         /* Fetch initial physical path information for this device. */
 850         if (cp != NULL)
 851                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 852
 853         g_topology_unlock();
 854         PICKUP_GIANT();
 855         if (cp == NULL) {
 856                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 857                 return (error);
 858         }
 859         pp = cp->provider;
 860
 861         /*
 862          * Determine the actual size of the device.
 863          */
 864         *max_psize = *psize = pp->mediasize;
 865
 866         /*
 867          * Determine the device's minimum transfer size and preferred
 868          * transfer size.
 869          */
 870         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 871         *physical_ashift = 0;
 872         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 873             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 874                 *physical_ashift = highbit(pp->stripesize) - 1;
 875
 876         /*
 877          * Clear the nowritecache settings, so that on a vdev_reopen()
 878          * we will try again.
 879          */
 880         vd->vdev_nowritecache = B_FALSE;
 881
 882         /*
 883          * Determine the device's rotation rate.
 884          */
 885         vdev_geom_set_rotation_rate(vd, cp);
 886
 887         return (0);
 888 }
 889
 890 static void
 891 vdev_geom_close(vdev_t *vd)
 892 {
 893
 894         DROP_GIANT();
 895         g_topology_lock();
 896         vdev_geom_close_locked(vd);
 897         g_topology_unlock();
 898         PICKUP_GIANT();
 899 }
 900
 901 static void
 902 vdev_geom_io_intr(struct bio *bp)
 903 {
 904         vdev_t *vd;
 905         zio_t *zio;
 906
 907         zio = bp->bio_caller1;
 908         vd = zio->io_vd;
 909         zio->io_error = bp->bio_error;
 910         if (zio->io_error == 0 && bp->bio_resid != 0)
 911                 zio->io_error = SET_ERROR(EIO);
 912
 913         switch(zio->io_error) {
 914         case ENOTSUP:
 915                 /*
 916                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 917                  * that future attempts will never succeed. In this case
 918                  * we set a persistent flag so that we don't bother with
 919                  * requests in the future.
 920                  */
 921                 switch(bp->bio_cmd) {
 922                 case BIO_FLUSH:
 923                         vd->vdev_nowritecache = B_TRUE;
 924                         break;
 925                 case BIO_DELETE:
 926                         vd->vdev_notrim = B_TRUE;
 927                         break;
 928                 }
 929                 break;
 930         case ENXIO:
 931                 if (!vd->vdev_remove_wanted) {
 932                         /*
 933                          * If provider's error is set we assume it is being
 934                          * removed.
 935                          */
 936                         if (bp->bio_to->error != 0) {
 937                                 vd->vdev_remove_wanted = B_TRUE;
 938                                 spa_async_request(zio->io_spa,
 939                                     SPA_ASYNC_REMOVE);
 940                         } else if (!vd->vdev_delayed_close) {
 941                                 vd->vdev_delayed_close = B_TRUE;
 942                         }
 943                 }
 944                 break;
 945         }
 946         g_destroy_bio(bp);
 947         zio_delay_interrupt(zio);
 948 }
 949
 950 static void
 951 vdev_geom_io_start(zio_t *zio)
 952 {
 953         vdev_t *vd;
 954         struct g_consumer *cp;
 955         struct bio *bp;
 956         int error;
 957
 958         vd = zio->io_vd;
 959
 960         switch (zio->io_type) {
 961         case ZIO_TYPE_IOCTL:
 962                 /* XXPOLICY */
 963                 if (!vdev_readable(vd)) {
 964                         zio->io_error = SET_ERROR(ENXIO);
 965                         zio_interrupt(zio);
 966                         return;
 967                 } else {
 968                         switch (zio->io_cmd) {
 969                         case DKIOCFLUSHWRITECACHE:
 970                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 971                                         break;
 972                                 if (vd->vdev_nowritecache) {
 973                                         zio->io_error = SET_ERROR(ENOTSUP);
 974                                         break;
 975                                 }
 976                                 goto sendreq;
 977                         default:
 978                                 zio->io_error = SET_ERROR(ENOTSUP);
 979                         }
 980                 }
 981
 982                 zio_execute(zio);
 983                 return;
 984         case ZIO_TYPE_FREE:
 985                 if (vd->vdev_notrim) {
 986                         zio->io_error = SET_ERROR(ENOTSUP);
 987                 } else if (!vdev_geom_bio_delete_disable) {
 988                         goto sendreq;
 989                 }
 990                 zio_execute(zio);
 991                 return;
 992         }
 993 sendreq:
 994         ASSERT(zio->io_type == ZIO_TYPE_READ ||
 995             zio->io_type == ZIO_TYPE_WRITE ||
 996             zio->io_type == ZIO_TYPE_FREE ||
 997             zio->io_type == ZIO_TYPE_IOCTL);
 998
 999         cp = vd->vdev_tsd;
1000         if (cp == NULL) {
1001                 zio->io_error = SET_ERROR(ENXIO);
1002                 zio_interrupt(zio);
1003                 return;
1004         }
1005         bp = g_alloc_bio();
1006         bp->bio_caller1 = zio;
1007         switch (zio->io_type) {
1008         case ZIO_TYPE_READ:
1009         case ZIO_TYPE_WRITE:
1010                 zio->io_target_timestamp = zio_handle_io_delay(zio);
1011                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
1012                 bp->bio_data = zio->io_data;
1013                 bp->bio_offset = zio->io_offset;
1014                 bp->bio_length = zio->io_size;
1015                 break;
1016         case ZIO_TYPE_FREE:
1017                 bp->bio_cmd = BIO_DELETE;
1018                 bp->bio_data = NULL;
1019                 bp->bio_offset = zio->io_offset;
1020                 bp->bio_length = zio->io_size;
1021                 break;
1022         case ZIO_TYPE_IOCTL:
1023                 bp->bio_cmd = BIO_FLUSH;
1024                 bp->bio_flags |= BIO_ORDERED;
1025                 bp->bio_data = NULL;
1026                 bp->bio_offset = cp->provider->mediasize;
1027                 bp->bio_length = 0;
1028                 break;
1029         }
1030         bp->bio_done = vdev_geom_io_intr;
1031
1032         g_io_request(bp, cp);
1033 }
1034
1035 static void
1036 vdev_geom_io_done(zio_t *zio)
1037 {
1038 }
1039
1040 static void
1041 vdev_geom_hold(vdev_t *vd)
1042 {
1043 }
1044
1045 static void
1046 vdev_geom_rele(vdev_t *vd)
1047 {
1048 }
1049
1050 vdev_ops_t vdev_geom_ops = {
1051         vdev_geom_open,
1052         vdev_geom_close,
1053         vdev_default_asize,
1054         vdev_geom_io_start,
1055         vdev_geom_io_done,
1056         NULL,
1057         vdev_geom_hold,
1058         vdev_geom_rele,
1059         VDEV_TYPE_DISK,         /* name of this vdev type */
1060         B_TRUE                  /* leaf vdev */
1061 };