sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_geom.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2006 Pawel Jakub Dawidek <pjd@FreeBSD.org>
  23  * All rights reserved.
  24  *
  25  * Portions Copyright (c) 2012 Martin Matuska <mm@FreeBSD.org>
  26  */
  27
  28 #include <sys/zfs_context.h>
  29 #include <sys/param.h>
  30 #include <sys/kernel.h>
  31 #include <sys/bio.h>
  32 #include <sys/disk.h>
  33 #include <sys/spa.h>
  34 #include <sys/spa_impl.h>
  35 #include <sys/vdev_impl.h>
  36 #include <sys/fs/zfs.h>
  37 #include <sys/zio.h>
  38 #include <geom/geom.h>
  39 #include <geom/geom_int.h>
  40
  41 /*
  42  * Virtual device vector for GEOM.
  43  */
  44
  45 static g_attrchanged_t vdev_geom_attrchanged;
  46 struct g_class zfs_vdev_class = {
  47         .name = "ZFS::VDEV",
  48         .version = G_VERSION,
  49         .attrchanged = vdev_geom_attrchanged,
  50 };
  51
  52 DECLARE_GEOM_CLASS(zfs_vdev_class, zfs_vdev);
  53
  54 SYSCTL_DECL(_vfs_zfs_vdev);
  55 /* Don't send BIO_FLUSH. */
  56 static int vdev_geom_bio_flush_disable;
  57 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_flush_disable, CTLFLAG_RWTUN,
  58     &vdev_geom_bio_flush_disable, 0, "Disable BIO_FLUSH");
  59 /* Don't send BIO_DELETE. */
  60 static int vdev_geom_bio_delete_disable;
  61 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_delete_disable, CTLFLAG_RWTUN,
  62     &vdev_geom_bio_delete_disable, 0, "Disable BIO_DELETE");
  63
  64 /* Declare local functions */
  65 static void vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read);
  66
  67 /*
  68  * Thread local storage used to indicate when a thread is probing geoms
  69  * for their guids.  If NULL, this thread is not tasting geoms.  If non NULL,
  70  * it is looking for a replacement for the vdev_t* that is its value.
  71  */
  72 uint_t zfs_geom_probe_vdev_key;
  73
  74 static void
  75 vdev_geom_set_rotation_rate(vdev_t *vd, struct g_consumer *cp)
  76 {
  77         int error;
  78         uint16_t rate;
  79
  80         error = g_getattr("GEOM::rotation_rate", cp, &rate);
  81         if (error == 0)
  82                 vd->vdev_rotation_rate = rate;
  83         else
  84                 vd->vdev_rotation_rate = VDEV_RATE_UNKNOWN;
  85 }
  86
  87 static void
  88 vdev_geom_set_physpath(struct g_consumer *cp, boolean_t do_null_update)
  89 {
  90         boolean_t needs_update = B_FALSE;
  91         vdev_t *vd;
  92         char *physpath;
  93         int error, physpath_len;
  94
  95         if (g_access(cp, 1, 0, 0) != 0)
  96                 return;
  97
  98         vd = cp->private;
  99         physpath_len = MAXPATHLEN;
 100         physpath = g_malloc(physpath_len, M_WAITOK|M_ZERO);
 101         error = g_io_getattr("GEOM::physpath", cp, &physpath_len, physpath);
 102         g_access(cp, -1, 0, 0);
 103         if (error == 0) {
 104                 char *old_physpath;
 105
 106                 /* g_topology lock ensures that vdev has not been closed */
 107                 g_topology_assert();
 108                 old_physpath = vd->vdev_physpath;
 109                 vd->vdev_physpath = spa_strdup(physpath);
 110
 111                 if (old_physpath != NULL) {
 112                         needs_update = (strcmp(old_physpath,
 113                                                 vd->vdev_physpath) != 0);
 114                         spa_strfree(old_physpath);
 115                 } else
 116                         needs_update = do_null_update;
 117         }
 118         g_free(physpath);
 119
 120         /*
 121          * If the physical path changed, update the config.
 122          * Only request an update for previously unset physpaths if
 123          * requested by the caller.
 124          */
 125         if (needs_update)
 126                 spa_async_request(vd->vdev_spa, SPA_ASYNC_CONFIG_UPDATE);
 127
 128 }
 129
 130 static void
 131 vdev_geom_attrchanged(struct g_consumer *cp, const char *attr)
 132 {
 133         vdev_t *vd;
 134         char *old_physpath;
 135         int error;
 136
 137         vd = cp->private;
 138         if (vd == NULL)
 139                 return;
 140
 141         if (strcmp(attr, "GEOM::rotation_rate") == 0) {
 142                 vdev_geom_set_rotation_rate(vd, cp);
 143                 return;
 144         }
 145
 146         if (strcmp(attr, "GEOM::physpath") == 0) {
 147                 vdev_geom_set_physpath(cp, /*do_null_update*/B_TRUE);
 148                 return;
 149         }
 150 }
 151
 152 static void
 153 vdev_geom_orphan(struct g_consumer *cp)
 154 {
 155         vdev_t *vd;
 156
 157         g_topology_assert();
 158
 159         vd = cp->private;
 160         if (vd == NULL) {
 161                 /* Vdev close in progress.  Ignore the event. */
 162                 return;
 163         }
 164
 165         /*
 166          * Orphan callbacks occur from the GEOM event thread.
 167          * Concurrent with this call, new I/O requests may be
 168          * working their way through GEOM about to find out
 169          * (only once executed by the g_down thread) that we've
 170          * been orphaned from our disk provider.  These I/Os
 171          * must be retired before we can detach our consumer.
 172          * This is most easily achieved by acquiring the
 173          * SPA ZIO configuration lock as a writer, but doing
 174          * so with the GEOM topology lock held would cause
 175          * a lock order reversal.  Instead, rely on the SPA's
 176          * async removal support to invoke a close on this
 177          * vdev once it is safe to do so.
 178          */
 179         vd->vdev_remove_wanted = B_TRUE;
 180         spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
 181 }
 182
 183 static struct g_consumer *
 184 vdev_geom_attach(struct g_provider *pp, vdev_t *vd)
 185 {
 186         struct g_geom *gp;
 187         struct g_consumer *cp;
 188         int error;
 189
 190         g_topology_assert();
 191
 192         ZFS_LOG(1, "Attaching to %s.", pp->name);
 193
 194         if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize)) {
 195                 ZFS_LOG(1, "Failing attach of %s. Incompatible sectorsize %d\n",
 196                     pp->name, pp->sectorsize);
 197                 return (NULL);
 198         } else if (pp->mediasize < SPA_MINDEVSIZE) {
 199                 ZFS_LOG(1, "Failing attach of %s. Incompatible mediasize %ju\n",
 200                     pp->name, pp->mediasize);
 201                 return (NULL);
 202         }
 203
 204         /* Do we have geom already? No? Create one. */
 205         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 206                 if (gp->flags & G_GEOM_WITHER)
 207                         continue;
 208                 if (strcmp(gp->name, "zfs::vdev") != 0)
 209                         continue;
 210                 break;
 211         }
 212         if (gp == NULL) {
 213                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 214                 gp->orphan = vdev_geom_orphan;
 215                 gp->attrchanged = vdev_geom_attrchanged;
 216                 cp = g_new_consumer(gp);
 217                 error = g_attach(cp, pp);
 218                 if (error != 0) {
 219                         ZFS_LOG(1, "%s(%d): g_attach failed: %d\n", __func__,
 220                             __LINE__, error);
 221                         vdev_geom_detach(cp, B_FALSE);
 222                         return (NULL);
 223                 }
 224                 error = g_access(cp, 1, 0, 1);
 225                 if (error != 0) {
 226                         ZFS_LOG(1, "%s(%d): g_access failed: %d", __func__,
 227                                __LINE__, error);
 228                         vdev_geom_detach(cp, B_FALSE);
 229                         return (NULL);
 230                 }
 231                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 232         } else {
 233                 /* Check if we are already connected to this provider. */
 234                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 235                         if (cp->provider == pp) {
 236                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 237                                 break;
 238                         }
 239                 }
 240                 if (cp == NULL) {
 241                         cp = g_new_consumer(gp);
 242                         error = g_attach(cp, pp);
 243                         if (error != 0) {
 244                                 ZFS_LOG(1, "%s(%d): g_attach failed: %d\n",
 245                                     __func__, __LINE__, error);
 246                                 vdev_geom_detach(cp, B_FALSE);
 247                                 return (NULL);
 248                         }
 249                         error = g_access(cp, 1, 0, 1);
 250                         if (error != 0) {
 251                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 252                                     __func__, __LINE__, error);
 253                                 vdev_geom_detach(cp, B_FALSE);
 254                                 return (NULL);
 255                         }
 256                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 257                 } else {
 258                         error = g_access(cp, 1, 0, 1);
 259                         if (error != 0) {
 260                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 261                                     __func__, __LINE__, error);
 262                                 return (NULL);
 263                         }
 264                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 265                 }
 266         }
 267
 268         /*
 269          * BUG: cp may already belong to a vdev.  This could happen if:
 270          * 1) That vdev is a shared spare, or
 271          * 2) We are trying to reopen a missing vdev and we are scanning by
 272          *    guid.  In that case, we'll ultimately fail to open this consumer,
 273          *    but not until after setting the private field.
 274          * The solution is to:
 275          * 1) Don't set the private field until after the open succeeds, and
 276          * 2) Set it to a linked list of vdevs, not just a single vdev
 277          */
 278         cp->private = vd;
 279         if (vd != NULL) {
 280                 vd->vdev_tsd = cp;
 281                 vdev_geom_set_physpath(cp, /*do_null_update*/B_FALSE);
 282         }
 283
 284         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
 285         return (cp);
 286 }
 287
 288 static void
 289 vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read)
 290 {
 291         struct g_geom *gp;
 292         vdev_t *vd;
 293
 294         g_topology_assert();
 295
 296         ZFS_LOG(1, "Detaching from %s.",
 297             cp->provider && cp->provider->name ? cp->provider->name : "NULL");
 298
 299         vd = cp->private;
 300         cp->private = NULL;
 301
 302         gp = cp->geom;
 303         if (open_for_read)
 304                 g_access(cp, -1, 0, -1);
 305         /* Destroy consumer on last close. */
 306         if (cp->acr == 0 && cp->ace == 0) {
 307                 if (cp->acw > 0)
 308                         g_access(cp, 0, -cp->acw, 0);
 309                 if (cp->provider != NULL) {
 310                         ZFS_LOG(1, "Destroying consumer for %s.",
 311                             cp->provider->name ? cp->provider->name : "NULL");
 312                         g_detach(cp);
 313                 }
 314                 g_destroy_consumer(cp);
 315         }
 316         /* Destroy geom if there are no consumers left. */
 317         if (LIST_EMPTY(&gp->consumer)) {
 318                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 319                 g_wither_geom(gp, ENXIO);
 320         }
 321 }
 322
 323 static void
 324 vdev_geom_close_locked(vdev_t *vd)
 325 {
 326         struct g_consumer *cp;
 327
 328         g_topology_assert();
 329
 330         cp = vd->vdev_tsd;
 331         vd->vdev_tsd = NULL;
 332         vd->vdev_delayed_close = B_FALSE;
 333         if (cp == NULL)
 334                 return;
 335
 336         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 337
 338         vdev_geom_detach(cp, B_TRUE);
 339 }
 340
 341 /*
 342  * Issue one or more bios to the vdev in parallel
 343  * cmds, datas, offsets, errors, and sizes are arrays of length ncmds.  Each IO
 344  * operation is described by parallel entries from each array.  There may be
 345  * more bios actually issued than entries in the array
 346  */
 347 static void
 348 vdev_geom_io(struct g_consumer *cp, int *cmds, void **datas, off_t *offsets,
 349     off_t *sizes, int *errors, int ncmds)
 350 {
 351         struct bio **bios;
 352         u_char *p;
 353         off_t off, maxio, s, end;
 354         int i, n_bios, j;
 355         size_t bios_size;
 356
 357         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 358         n_bios = 0;
 359
 360         /* How many bios are required for all commands ? */
 361         for (i = 0; i < ncmds; i++)
 362                 n_bios += (sizes[i] + maxio - 1) / maxio;
 363
 364         /* Allocate memory for the bios */
 365         bios_size = n_bios * sizeof(struct bio*);
 366         bios = kmem_zalloc(bios_size, KM_SLEEP);
 367
 368         /* Prepare and issue all of the bios */
 369         for (i = j = 0; i < ncmds; i++) {
 370                 off = offsets[i];
 371                 p = datas[i];
 372                 s = sizes[i];
 373                 end = off + s;
 374                 ASSERT((off % cp->provider->sectorsize) == 0);
 375                 ASSERT((s % cp->provider->sectorsize) == 0);
 376
 377                 for (; off < end; off += maxio, p += maxio, s -= maxio, j++) {
 378                         bios[j] = g_alloc_bio();
 379                         bios[j]->bio_cmd = cmds[i];
 380                         bios[j]->bio_done = NULL;
 381                         bios[j]->bio_offset = off;
 382                         bios[j]->bio_length = MIN(s, maxio);
 383                         bios[j]->bio_data = p;
 384                         g_io_request(bios[j], cp);
 385                 }
 386         }
 387         ASSERT(j == n_bios);
 388
 389         /* Wait for all of the bios to complete, and clean them up */
 390         for (i = j = 0; i < ncmds; i++) {
 391                 off = offsets[i];
 392                 s = sizes[i];
 393                 end = off + s;
 394
 395                 for (; off < end; off += maxio, s -= maxio, j++) {
 396                         errors[i] = biowait(bios[j], "vdev_geom_io") || errors[i];
 397                         g_destroy_bio(bios[j]);
 398                 }
 399         }
 400         kmem_free(bios, bios_size);
 401 }
 402
 403 static int
 404 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 405 {
 406         struct g_provider *pp;
 407         vdev_phys_t *vdev_lists[VDEV_LABELS];
 408         char *p, *buf;
 409         size_t buflen;
 410         uint64_t psize, state, txg;
 411         off_t offsets[VDEV_LABELS];
 412         off_t size;
 413         off_t sizes[VDEV_LABELS];
 414         int cmds[VDEV_LABELS];
 415         int errors[VDEV_LABELS];
 416         int l, len;
 417
 418         g_topology_assert_not();
 419
 420         pp = cp->provider;
 421         ZFS_LOG(1, "Reading config from %s...", pp->name);
 422
 423         psize = pp->mediasize;
 424         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 425
 426         size = sizeof(*vdev_lists[0]) + pp->sectorsize -
 427             ((sizeof(*vdev_lists[0]) - 1) % pp->sectorsize) - 1;
 428
 429         buflen = sizeof(vdev_lists[0]->vp_nvlist);
 430
 431         *config = NULL;
 432         /* Create all of the IO requests */
 433         for (l = 0; l < VDEV_LABELS; l++) {
 434                 cmds[l] = BIO_READ;
 435                 vdev_lists[l] = kmem_alloc(size, KM_SLEEP);
 436                 offsets[l] = vdev_label_offset(psize, l, 0) + VDEV_SKIP_SIZE;
 437                 sizes[l] = size;
 438                 errors[l] = 0;
 439                 ASSERT(offsets[l] % pp->sectorsize == 0);
 440         }
 441
 442         /* Issue the IO requests */
 443         vdev_geom_io(cp, cmds, (void**)vdev_lists, offsets, sizes, errors,
 444             VDEV_LABELS);
 445
 446         /* Parse the labels */
 447         for (l = 0; l < VDEV_LABELS; l++) {
 448                 if (errors[l] != 0)
 449                         continue;
 450
 451                 buf = vdev_lists[l]->vp_nvlist;
 452
 453                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 454                         continue;
 455
 456                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 457                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 458                         nvlist_free(*config);
 459                         *config = NULL;
 460                         continue;
 461                 }
 462
 463                 if (state != POOL_STATE_SPARE &&
 464                     state != POOL_STATE_L2CACHE &&
 465                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 466                     &txg) != 0 || txg == 0)) {
 467                         nvlist_free(*config);
 468                         *config = NULL;
 469                         continue;
 470                 }
 471
 472                 break;
 473         }
 474
 475         /* Free the label storage */
 476         for (l = 0; l < VDEV_LABELS; l++)
 477                 kmem_free(vdev_lists[l], size);
 478
 479         return (*config == NULL ? ENOENT : 0);
 480 }
 481
 482 static void
 483 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 484 {
 485         nvlist_t **new_configs;
 486         uint64_t i;
 487
 488         if (id < *count)
 489                 return;
 490         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 491             KM_SLEEP);
 492         for (i = 0; i < *count; i++)
 493                 new_configs[i] = (*configs)[i];
 494         if (*configs != NULL)
 495                 kmem_free(*configs, *count * sizeof(void *));
 496         *configs = new_configs;
 497         *count = id + 1;
 498 }
 499
 500 static void
 501 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 502     const char *name, uint64_t* known_pool_guid)
 503 {
 504         nvlist_t *vdev_tree;
 505         uint64_t pool_guid;
 506         uint64_t vdev_guid, known_guid;
 507         uint64_t id, txg, known_txg;
 508         char *pname;
 509         int i;
 510
 511         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 512             strcmp(pname, name) != 0)
 513                 goto ignore;
 514
 515         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 516                 goto ignore;
 517
 518         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 519                 goto ignore;
 520
 521         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 522                 goto ignore;
 523
 524         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 525                 goto ignore;
 526
 527         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 528
 529         if (*known_pool_guid != 0) {
 530                 if (pool_guid != *known_pool_guid)
 531                         goto ignore;
 532         } else
 533                 *known_pool_guid = pool_guid;
 534
 535         resize_configs(configs, count, id);
 536
 537         if ((*configs)[id] != NULL) {
 538                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 539                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 540                 if (txg <= known_txg)
 541                         goto ignore;
 542                 nvlist_free((*configs)[id]);
 543         }
 544
 545         (*configs)[id] = cfg;
 546         return;
 547
 548 ignore:
 549         nvlist_free(cfg);
 550 }
 551
 552 int
 553 vdev_geom_read_pool_label(const char *name,
 554     nvlist_t ***configs, uint64_t *count)
 555 {
 556         struct g_class *mp;
 557         struct g_geom *gp;
 558         struct g_provider *pp;
 559         struct g_consumer *zcp;
 560         nvlist_t *vdev_cfg;
 561         uint64_t pool_guid;
 562         int error;
 563
 564         DROP_GIANT();
 565         g_topology_lock();
 566
 567         *configs = NULL;
 568         *count = 0;
 569         pool_guid = 0;
 570         LIST_FOREACH(mp, &g_classes, class) {
 571                 if (mp == &zfs_vdev_class)
 572                         continue;
 573                 LIST_FOREACH(gp, &mp->geom, geom) {
 574                         if (gp->flags & G_GEOM_WITHER)
 575                                 continue;
 576                         LIST_FOREACH(pp, &gp->provider, provider) {
 577                                 if (pp->flags & G_PF_WITHER)
 578                                         continue;
 579                                 zcp = vdev_geom_attach(pp, NULL);
 580                                 if (zcp == NULL)
 581                                         continue;
 582                                 g_topology_unlock();
 583                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 584                                 g_topology_lock();
 585                                 vdev_geom_detach(zcp, B_TRUE);
 586                                 if (error)
 587                                         continue;
 588                                 ZFS_LOG(1, "successfully read vdev config");
 589
 590                                 process_vdev_config(configs, count,
 591                                     vdev_cfg, name, &pool_guid);
 592                         }
 593                 }
 594         }
 595         g_topology_unlock();
 596         PICKUP_GIANT();
 597
 598         return (*count > 0 ? 0 : ENOENT);
 599 }
 600
 601 enum match {
 602         NO_MATCH,
 603         TOP_MATCH,
 604         FULL_MATCH
 605 };
 606
 607 static enum match
 608 vdev_attach_ok(vdev_t *vd, struct g_provider *pp)
 609 {
 610         nvlist_t *config;
 611         uint64_t pool_guid, top_guid, vdev_guid;
 612         struct g_consumer *cp;
 613
 614         cp = vdev_geom_attach(pp, NULL);
 615         if (cp == NULL) {
 616                 ZFS_LOG(1, "Unable to attach tasting instance to %s.",
 617                     pp->name);
 618                 return (NO_MATCH);
 619         }
 620         g_topology_unlock();
 621         if (vdev_geom_read_config(cp, &config) != 0) {
 622                 g_topology_lock();
 623                 vdev_geom_detach(cp, B_TRUE);
 624                 ZFS_LOG(1, "Unable to read config from %s.", pp->name);
 625                 return (NO_MATCH);
 626         }
 627         g_topology_lock();
 628         vdev_geom_detach(cp, B_TRUE);
 629
 630         pool_guid = 0;
 631         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid);
 632         top_guid = 0;
 633         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TOP_GUID, &top_guid);
 634         vdev_guid = 0;
 635         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_GUID, &vdev_guid);
 636         nvlist_free(config);
 637
 638         /*
 639          * Check that the label's pool guid matches the desired guid.
 640          * Inactive spares and L2ARCs do not have any pool guid in the label.
 641          */
 642         if (pool_guid != 0 && pool_guid != spa_guid(vd->vdev_spa)) {
 643                 ZFS_LOG(1, "pool guid mismatch for provider %s: %ju != %ju.",
 644                     pp->name,
 645                     (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)pool_guid);
 646                 return (NO_MATCH);
 647         }
 648
 649         /*
 650          * Check that the label's vdev guid matches the desired guid.
 651          * The second condition handles possible race on vdev detach, when
 652          * remaining vdev receives GUID of destroyed top level mirror vdev.
 653          */
 654         if (vdev_guid == vd->vdev_guid) {
 655                 ZFS_LOG(1, "guids match for provider %s.", pp->name);
 656                 return (FULL_MATCH);
 657         } else if (top_guid == vd->vdev_guid && vd == vd->vdev_top) {
 658                 ZFS_LOG(1, "top vdev guid match for provider %s.", pp->name);
 659                 return (TOP_MATCH);
 660         }
 661         ZFS_LOG(1, "vdev guid mismatch for provider %s: %ju != %ju.",
 662             pp->name, (uintmax_t)vd->vdev_guid, (uintmax_t)vdev_guid);
 663         return (NO_MATCH);
 664 }
 665
 666 static struct g_consumer *
 667 vdev_geom_attach_by_guids(vdev_t *vd)
 668 {
 669         struct g_class *mp;
 670         struct g_geom *gp;
 671         struct g_provider *pp;
 672         struct g_consumer *cp;
 673         enum match m;
 674
 675         g_topology_assert();
 676
 677         cp = NULL;
 678         LIST_FOREACH(mp, &g_classes, class) {
 679                 if (mp == &zfs_vdev_class)
 680                         continue;
 681                 LIST_FOREACH(gp, &mp->geom, geom) {
 682                         if (gp->flags & G_GEOM_WITHER)
 683                                 continue;
 684                         LIST_FOREACH(pp, &gp->provider, provider) {
 685                                 m = vdev_attach_ok(vd, pp);
 686                                 if (m == NO_MATCH)
 687                                         continue;
 688                                 if (cp != NULL) {
 689                                         if (m == FULL_MATCH)
 690                                                 vdev_geom_detach(cp, B_TRUE);
 691                                         else
 692                                                 continue;
 693                                 }
 694                                 cp = vdev_geom_attach(pp, vd);
 695                                 if (cp == NULL) {
 696                                         printf("ZFS WARNING: Unable to "
 697                                             "attach to %s.\n", pp->name);
 698                                         continue;
 699                                 }
 700                                 if (m == FULL_MATCH)
 701                                         return (cp);
 702                         }
 703                 }
 704         }
 705         return (cp);
 706 }
 707
 708 static struct g_consumer *
 709 vdev_geom_open_by_guids(vdev_t *vd)
 710 {
 711         struct g_consumer *cp;
 712         char *buf;
 713         size_t len;
 714
 715         g_topology_assert();
 716
 717         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 718                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 719         cp = vdev_geom_attach_by_guids(vd);
 720         if (cp != NULL) {
 721                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 722                 buf = kmem_alloc(len, KM_SLEEP);
 723
 724                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 725                 spa_strfree(vd->vdev_path);
 726                 vd->vdev_path = buf;
 727
 728                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 729                     (uintmax_t)spa_guid(vd->vdev_spa),
 730                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 731         } else {
 732                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 733                     (uintmax_t)spa_guid(vd->vdev_spa),
 734                     (uintmax_t)vd->vdev_guid);
 735         }
 736
 737         return (cp);
 738 }
 739
 740 static struct g_consumer *
 741 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 742 {
 743         struct g_provider *pp;
 744         struct g_consumer *cp;
 745
 746         g_topology_assert();
 747
 748         cp = NULL;
 749         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 750         if (pp != NULL) {
 751                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 752                 if (!check_guid || vdev_attach_ok(vd, pp) == FULL_MATCH)
 753                         cp = vdev_geom_attach(pp, vd);
 754         }
 755
 756         return (cp);
 757 }
 758
 759 static int
 760 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 761     uint64_t *logical_ashift, uint64_t *physical_ashift)
 762 {
 763         struct g_provider *pp;
 764         struct g_consumer *cp;
 765         size_t bufsize;
 766         int error;
 767
 768         /* Set the TLS to indicate downstack that we should not access zvols*/
 769         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 770
 771         /*
 772          * We must have a pathname, and it must be absolute.
 773          */
 774         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 775                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 776                 return (EINVAL);
 777         }
 778
 779         /*
 780          * Reopen the device if it's not currently open. Otherwise,
 781          * just update the physical size of the device.
 782          */
 783         if ((cp = vd->vdev_tsd) != NULL) {
 784                 ASSERT(vd->vdev_reopening);
 785                 goto skip_open;
 786         }
 787
 788         DROP_GIANT();
 789         g_topology_lock();
 790         error = 0;
 791
 792         if (vd->vdev_spa->spa_splitting_newspa ||
 793             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 794              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 795              vd->vdev_spa->spa_load_state == SPA_LOAD_CREATE)) {
 796                 /*
 797                  * We are dealing with a vdev that hasn't been previously
 798                  * opened (since boot), and we are not loading an
 799                  * existing pool configuration.  This looks like a
 800                  * vdev add operation to a new or existing pool.
 801                  * Assume the user knows what he/she is doing and find
 802                  * GEOM provider by its name, ignoring GUID mismatches.
 803                  *
 804                  * XXPOLICY: It would be safer to only allow a device
 805                  *           that is unlabeled or labeled but missing
 806                  *           GUID information to be opened in this fashion,
 807                  *           unless we are doing a split, in which case we
 808                  *           should allow any guid.
 809                  */
 810                 cp = vdev_geom_open_by_path(vd, 0);
 811         } else {
 812                 /*
 813                  * Try using the recorded path for this device, but only
 814                  * accept it if its label data contains the expected GUIDs.
 815                  */
 816                 cp = vdev_geom_open_by_path(vd, 1);
 817                 if (cp == NULL) {
 818                         /*
 819                          * The device at vd->vdev_path doesn't have the
 820                          * expected GUIDs. The disks might have merely
 821                          * moved around so try all other GEOM providers
 822                          * to find one with the right GUIDs.
 823                          */
 824                         cp = vdev_geom_open_by_guids(vd);
 825                 }
 826         }
 827
 828         /* Clear the TLS now that tasting is done */
 829         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 830
 831         if (cp == NULL) {
 832                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 833                 error = ENOENT;
 834         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 835             !ISP2(cp->provider->sectorsize)) {
 836                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 837                     vd->vdev_path);
 838
 839                 vdev_geom_close_locked(vd);
 840                 error = EINVAL;
 841                 cp = NULL;
 842         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 843                 int i;
 844
 845                 for (i = 0; i < 5; i++) {
 846                         error = g_access(cp, 0, 1, 0);
 847                         if (error == 0)
 848                                 break;
 849                         g_topology_unlock();
 850                         tsleep(vd, 0, "vdev", hz / 2);
 851                         g_topology_lock();
 852                 }
 853                 if (error != 0) {
 854                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 855                             vd->vdev_path, error);
 856                         vdev_geom_close_locked(vd);
 857                         cp = NULL;
 858                 }
 859         }
 860
 861         /* Fetch initial physical path information for this device. */
 862         if (cp != NULL)
 863                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 864
 865         g_topology_unlock();
 866         PICKUP_GIANT();
 867         if (cp == NULL) {
 868                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 869                 return (error);
 870         }
 871 skip_open:
 872         pp = cp->provider;
 873
 874         /*
 875          * Determine the actual size of the device.
 876          */
 877         *max_psize = *psize = pp->mediasize;
 878
 879         /*
 880          * Determine the device's minimum transfer size and preferred
 881          * transfer size.
 882          */
 883         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 884         *physical_ashift = 0;
 885         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 886             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 887                 *physical_ashift = highbit(pp->stripesize) - 1;
 888
 889         /*
 890          * Clear the nowritecache settings, so that on a vdev_reopen()
 891          * we will try again.
 892          */
 893         vd->vdev_nowritecache = B_FALSE;
 894
 895         /*
 896          * Determine the device's rotation rate.
 897          */
 898         vdev_geom_set_rotation_rate(vd, cp);
 899
 900         return (0);
 901 }
 902
 903 static void
 904 vdev_geom_close(vdev_t *vd)
 905 {
 906
 907         if (vd->vdev_reopening)
 908                 return;
 909
 910         DROP_GIANT();
 911         g_topology_lock();
 912         vdev_geom_close_locked(vd);
 913         g_topology_unlock();
 914         PICKUP_GIANT();
 915 }
 916
 917 static void
 918 vdev_geom_io_intr(struct bio *bp)
 919 {
 920         vdev_t *vd;
 921         zio_t *zio;
 922
 923         zio = bp->bio_caller1;
 924         vd = zio->io_vd;
 925         zio->io_error = bp->bio_error;
 926         if (zio->io_error == 0 && bp->bio_resid != 0)
 927                 zio->io_error = SET_ERROR(EIO);
 928
 929         switch(zio->io_error) {
 930         case ENOTSUP:
 931                 /*
 932                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 933                  * that future attempts will never succeed. In this case
 934                  * we set a persistent flag so that we don't bother with
 935                  * requests in the future.
 936                  */
 937                 switch(bp->bio_cmd) {
 938                 case BIO_FLUSH:
 939                         vd->vdev_nowritecache = B_TRUE;
 940                         break;
 941                 case BIO_DELETE:
 942                         vd->vdev_notrim = B_TRUE;
 943                         break;
 944                 }
 945                 break;
 946         case ENXIO:
 947                 if (!vd->vdev_remove_wanted) {
 948                         /*
 949                          * If provider's error is set we assume it is being
 950                          * removed.
 951                          */
 952                         if (bp->bio_to->error != 0) {
 953                                 vd->vdev_remove_wanted = B_TRUE;
 954                                 spa_async_request(zio->io_spa,
 955                                     SPA_ASYNC_REMOVE);
 956                         } else if (!vd->vdev_delayed_close) {
 957                                 vd->vdev_delayed_close = B_TRUE;
 958                         }
 959                 }
 960                 break;
 961         }
 962         g_destroy_bio(bp);
 963         zio_delay_interrupt(zio);
 964 }
 965
 966 static void
 967 vdev_geom_io_start(zio_t *zio)
 968 {
 969         vdev_t *vd;
 970         struct g_consumer *cp;
 971         struct bio *bp;
 972         int error;
 973
 974         vd = zio->io_vd;
 975
 976         switch (zio->io_type) {
 977         case ZIO_TYPE_IOCTL:
 978                 /* XXPOLICY */
 979                 if (!vdev_readable(vd)) {
 980                         zio->io_error = SET_ERROR(ENXIO);
 981                         zio_interrupt(zio);
 982                         return;
 983                 } else {
 984                         switch (zio->io_cmd) {
 985                         case DKIOCFLUSHWRITECACHE:
 986                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 987                                         break;
 988                                 if (vd->vdev_nowritecache) {
 989                                         zio->io_error = SET_ERROR(ENOTSUP);
 990                                         break;
 991                                 }
 992                                 goto sendreq;
 993                         default:
 994                                 zio->io_error = SET_ERROR(ENOTSUP);
 995                         }
 996                 }
 997
 998                 zio_execute(zio);
 999                 return;
1000         case ZIO_TYPE_FREE:
1001                 if (vd->vdev_notrim) {
1002                         zio->io_error = SET_ERROR(ENOTSUP);
1003                 } else if (!vdev_geom_bio_delete_disable) {
1004                         goto sendreq;
1005                 }
1006                 zio_execute(zio);
1007                 return;
1008         }
1009 sendreq:
1010         ASSERT(zio->io_type == ZIO_TYPE_READ ||
1011             zio->io_type == ZIO_TYPE_WRITE ||
1012             zio->io_type == ZIO_TYPE_FREE ||
1013             zio->io_type == ZIO_TYPE_IOCTL);
1014
1015         cp = vd->vdev_tsd;
1016         if (cp == NULL) {
1017                 zio->io_error = SET_ERROR(ENXIO);
1018                 zio_interrupt(zio);
1019                 return;
1020         }
1021         bp = g_alloc_bio();
1022         bp->bio_caller1 = zio;
1023         switch (zio->io_type) {
1024         case ZIO_TYPE_READ:
1025         case ZIO_TYPE_WRITE:
1026                 zio->io_target_timestamp = zio_handle_io_delay(zio);
1027                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
1028                 bp->bio_data = zio->io_data;
1029                 bp->bio_offset = zio->io_offset;
1030                 bp->bio_length = zio->io_size;
1031                 break;
1032         case ZIO_TYPE_FREE:
1033                 bp->bio_cmd = BIO_DELETE;
1034                 bp->bio_data = NULL;
1035                 bp->bio_offset = zio->io_offset;
1036                 bp->bio_length = zio->io_size;
1037                 break;
1038         case ZIO_TYPE_IOCTL:
1039                 bp->bio_cmd = BIO_FLUSH;
1040                 bp->bio_flags |= BIO_ORDERED;
1041                 bp->bio_data = NULL;
1042                 bp->bio_offset = cp->provider->mediasize;
1043                 bp->bio_length = 0;
1044                 break;
1045         }
1046         bp->bio_done = vdev_geom_io_intr;
1047
1048         g_io_request(bp, cp);
1049 }
1050
1051 static void
1052 vdev_geom_io_done(zio_t *zio)
1053 {
1054 }
1055
1056 static void
1057 vdev_geom_hold(vdev_t *vd)
1058 {
1059 }
1060
1061 static void
1062 vdev_geom_rele(vdev_t *vd)
1063 {
1064 }
1065
1066 vdev_ops_t vdev_geom_ops = {
1067         vdev_geom_open,
1068         vdev_geom_close,
1069         vdev_default_asize,
1070         vdev_geom_io_start,
1071         vdev_geom_io_done,
1072         NULL,
1073         vdev_geom_hold,
1074         vdev_geom_rele,
1075         VDEV_TYPE_DISK,         /* name of this vdev type */
1076         B_TRUE                  /* leaf vdev */
1077 };