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 struct consumer_vdev_elem {
  53         SLIST_ENTRY(consumer_vdev_elem) elems;
  54         vdev_t                          *vd;
  55 };
  56
  57 SLIST_HEAD(consumer_priv_t, consumer_vdev_elem);
  58 _Static_assert(sizeof(((struct g_consumer*)NULL)->private)
  59     == sizeof(struct consumer_priv_t*),
  60     "consumer_priv_t* can't be stored in g_consumer.private");
  61
  62 DECLARE_GEOM_CLASS(zfs_vdev_class, zfs_vdev);
  63
  64 SYSCTL_DECL(_vfs_zfs_vdev);
  65 /* Don't send BIO_FLUSH. */
  66 static int vdev_geom_bio_flush_disable;
  67 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_flush_disable, CTLFLAG_RWTUN,
  68     &vdev_geom_bio_flush_disable, 0, "Disable BIO_FLUSH");
  69 /* Don't send BIO_DELETE. */
  70 static int vdev_geom_bio_delete_disable;
  71 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_delete_disable, CTLFLAG_RWTUN,
  72     &vdev_geom_bio_delete_disable, 0, "Disable BIO_DELETE");
  73
  74 /* Declare local functions */
  75 static void vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read);
  76
  77 /*
  78  * Thread local storage used to indicate when a thread is probing geoms
  79  * for their guids.  If NULL, this thread is not tasting geoms.  If non NULL,
  80  * it is looking for a replacement for the vdev_t* that is its value.
  81  */
  82 uint_t zfs_geom_probe_vdev_key;
  83
  84 static void
  85 vdev_geom_set_rotation_rate(vdev_t *vd, struct g_consumer *cp)
  86 {
  87         int error;
  88         uint16_t rate;
  89
  90         error = g_getattr("GEOM::rotation_rate", cp, &rate);
  91         if (error == 0 && rate == 1)
  92                 vd->vdev_nonrot = B_TRUE;
  93         else
  94                 vd->vdev_nonrot = B_FALSE;
  95 }
  96
  97 static void
  98 vdev_geom_set_physpath(vdev_t *vd, struct g_consumer *cp,
  99                        boolean_t do_null_update)
 100 {
 101         boolean_t needs_update = B_FALSE;
 102         char *physpath;
 103         int error, physpath_len;
 104
 105         physpath_len = MAXPATHLEN;
 106         physpath = g_malloc(physpath_len, M_WAITOK|M_ZERO);
 107         error = g_io_getattr("GEOM::physpath", cp, &physpath_len, physpath);
 108         if (error == 0) {
 109                 char *old_physpath;
 110
 111                 /* g_topology lock ensures that vdev has not been closed */
 112                 g_topology_assert();
 113                 old_physpath = vd->vdev_physpath;
 114                 vd->vdev_physpath = spa_strdup(physpath);
 115
 116                 if (old_physpath != NULL) {
 117                         needs_update = (strcmp(old_physpath,
 118                                                 vd->vdev_physpath) != 0);
 119                         spa_strfree(old_physpath);
 120                 } else
 121                         needs_update = do_null_update;
 122         }
 123         g_free(physpath);
 124
 125         /*
 126          * If the physical path changed, update the config.
 127          * Only request an update for previously unset physpaths if
 128          * requested by the caller.
 129          */
 130         if (needs_update)
 131                 spa_async_request(vd->vdev_spa, SPA_ASYNC_CONFIG_UPDATE);
 132
 133 }
 134
 135 static void
 136 vdev_geom_attrchanged(struct g_consumer *cp, const char *attr)
 137 {
 138         char *old_physpath;
 139         struct consumer_priv_t *priv;
 140         struct consumer_vdev_elem *elem;
 141         int error;
 142
 143         priv = (struct consumer_priv_t*)&cp->private;
 144         if (SLIST_EMPTY(priv))
 145                 return;
 146
 147         SLIST_FOREACH(elem, priv, elems) {
 148                 vdev_t *vd = elem->vd;
 149                 if (strcmp(attr, "GEOM::rotation_rate") == 0) {
 150                         vdev_geom_set_rotation_rate(vd, cp);
 151                         return;
 152                 }
 153                 if (strcmp(attr, "GEOM::physpath") == 0) {
 154                         vdev_geom_set_physpath(vd, cp, /*null_update*/B_TRUE);
 155                         return;
 156                 }
 157         }
 158 }
 159
 160 static void
 161 vdev_geom_resize(struct g_consumer *cp)
 162 {
 163         struct consumer_priv_t *priv;
 164         struct consumer_vdev_elem *elem;
 165         spa_t *spa;
 166         vdev_t *vd;
 167
 168         priv = (struct consumer_priv_t *)&cp->private;
 169         if (SLIST_EMPTY(priv))
 170                 return;
 171
 172         SLIST_FOREACH(elem, priv, elems) {
 173                 vd = elem->vd;
 174                 if (vd->vdev_state != VDEV_STATE_HEALTHY)
 175                         continue;
 176                 spa = vd->vdev_spa;
 177                 if (!spa->spa_autoexpand)
 178                         continue;
 179                 vdev_online(spa, vd->vdev_guid, ZFS_ONLINE_EXPAND, NULL);
 180         }
 181 }
 182
 183 static void
 184 vdev_geom_orphan(struct g_consumer *cp)
 185 {
 186         struct consumer_priv_t *priv;
 187         struct consumer_vdev_elem *elem;
 188
 189         g_topology_assert();
 190
 191         priv = (struct consumer_priv_t*)&cp->private;
 192         if (SLIST_EMPTY(priv))
 193                 /* Vdev close in progress.  Ignore the event. */
 194                 return;
 195
 196         /*
 197          * Orphan callbacks occur from the GEOM event thread.
 198          * Concurrent with this call, new I/O requests may be
 199          * working their way through GEOM about to find out
 200          * (only once executed by the g_down thread) that we've
 201          * been orphaned from our disk provider.  These I/Os
 202          * must be retired before we can detach our consumer.
 203          * This is most easily achieved by acquiring the
 204          * SPA ZIO configuration lock as a writer, but doing
 205          * so with the GEOM topology lock held would cause
 206          * a lock order reversal.  Instead, rely on the SPA's
 207          * async removal support to invoke a close on this
 208          * vdev once it is safe to do so.
 209          */
 210         SLIST_FOREACH(elem, priv, elems) {
 211                 vdev_t *vd = elem->vd;
 212
 213                 vd->vdev_remove_wanted = B_TRUE;
 214                 spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
 215         }
 216 }
 217
 218 static struct g_consumer *
 219 vdev_geom_attach(struct g_provider *pp, vdev_t *vd, boolean_t sanity)
 220 {
 221         struct g_geom *gp;
 222         struct g_consumer *cp;
 223         int error;
 224
 225         g_topology_assert();
 226
 227         ZFS_LOG(1, "Attaching to %s.", pp->name);
 228
 229         if (sanity) {
 230                 if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize)) {
 231                         ZFS_LOG(1, "Failing attach of %s. "
 232                                    "Incompatible sectorsize %d\n",
 233                             pp->name, pp->sectorsize);
 234                         return (NULL);
 235                 } else if (pp->mediasize < SPA_MINDEVSIZE) {
 236                         ZFS_LOG(1, "Failing attach of %s. "
 237                                    "Incompatible mediasize %ju\n",
 238                             pp->name, pp->mediasize);
 239                         return (NULL);
 240                 }
 241         }
 242
 243         /* Do we have geom already? No? Create one. */
 244         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 245                 if (gp->flags & G_GEOM_WITHER)
 246                         continue;
 247                 if (strcmp(gp->name, "zfs::vdev") != 0)
 248                         continue;
 249                 break;
 250         }
 251         if (gp == NULL) {
 252                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 253                 gp->orphan = vdev_geom_orphan;
 254                 gp->attrchanged = vdev_geom_attrchanged;
 255                 gp->resize = vdev_geom_resize;
 256                 cp = g_new_consumer(gp);
 257                 error = g_attach(cp, pp);
 258                 if (error != 0) {
 259                         ZFS_LOG(1, "%s(%d): g_attach failed: %d\n", __func__,
 260                             __LINE__, error);
 261                         vdev_geom_detach(cp, B_FALSE);
 262                         return (NULL);
 263                 }
 264                 error = g_access(cp, 1, 0, 1);
 265                 if (error != 0) {
 266                         ZFS_LOG(1, "%s(%d): g_access failed: %d\n", __func__,
 267                                __LINE__, error);
 268                         vdev_geom_detach(cp, B_FALSE);
 269                         return (NULL);
 270                 }
 271                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 272         } else {
 273                 /* Check if we are already connected to this provider. */
 274                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 275                         if (cp->provider == pp) {
 276                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 277                                 break;
 278                         }
 279                 }
 280                 if (cp == NULL) {
 281                         cp = g_new_consumer(gp);
 282                         error = g_attach(cp, pp);
 283                         if (error != 0) {
 284                                 ZFS_LOG(1, "%s(%d): g_attach failed: %d\n",
 285                                     __func__, __LINE__, error);
 286                                 vdev_geom_detach(cp, B_FALSE);
 287                                 return (NULL);
 288                         }
 289                         error = g_access(cp, 1, 0, 1);
 290                         if (error != 0) {
 291                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 292                                     __func__, __LINE__, error);
 293                                 vdev_geom_detach(cp, B_FALSE);
 294                                 return (NULL);
 295                         }
 296                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 297                 } else {
 298                         error = g_access(cp, 1, 0, 1);
 299                         if (error != 0) {
 300                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 301                                     __func__, __LINE__, error);
 302                                 return (NULL);
 303                         }
 304                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 305                 }
 306         }
 307
 308         if (vd != NULL)
 309                 vd->vdev_tsd = cp;
 310
 311         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
 312         return (cp);
 313 }
 314
 315 static void
 316 vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read)
 317 {
 318         struct g_geom *gp;
 319
 320         g_topology_assert();
 321
 322         ZFS_LOG(1, "Detaching from %s.",
 323             cp->provider && cp->provider->name ? cp->provider->name : "NULL");
 324
 325         gp = cp->geom;
 326         if (open_for_read)
 327                 g_access(cp, -1, 0, -1);
 328         /* Destroy consumer on last close. */
 329         if (cp->acr == 0 && cp->ace == 0) {
 330                 if (cp->acw > 0)
 331                         g_access(cp, 0, -cp->acw, 0);
 332                 if (cp->provider != NULL) {
 333                         ZFS_LOG(1, "Destroying consumer for %s.",
 334                             cp->provider->name ? cp->provider->name : "NULL");
 335                         g_detach(cp);
 336                 }
 337                 g_destroy_consumer(cp);
 338         }
 339         /* Destroy geom if there are no consumers left. */
 340         if (LIST_EMPTY(&gp->consumer)) {
 341                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 342                 g_wither_geom(gp, ENXIO);
 343         }
 344 }
 345
 346 static void
 347 vdev_geom_close_locked(vdev_t *vd)
 348 {
 349         struct g_consumer *cp;
 350         struct consumer_priv_t *priv;
 351         struct consumer_vdev_elem *elem, *elem_temp;
 352
 353         g_topology_assert();
 354
 355         cp = vd->vdev_tsd;
 356         vd->vdev_delayed_close = B_FALSE;
 357         if (cp == NULL)
 358                 return;
 359
 360         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 361         KASSERT(cp->private != NULL, ("%s: cp->private is NULL", __func__));
 362         priv = (struct consumer_priv_t*)&cp->private;
 363         vd->vdev_tsd = NULL;
 364         SLIST_FOREACH_SAFE(elem, priv, elems, elem_temp) {
 365                 if (elem->vd == vd) {
 366                         SLIST_REMOVE(priv, elem, consumer_vdev_elem, elems);
 367                         g_free(elem);
 368                 }
 369         }
 370
 371         vdev_geom_detach(cp, B_TRUE);
 372 }
 373
 374 /*
 375  * Issue one or more bios to the vdev in parallel
 376  * cmds, datas, offsets, errors, and sizes are arrays of length ncmds.  Each IO
 377  * operation is described by parallel entries from each array.  There may be
 378  * more bios actually issued than entries in the array
 379  */
 380 static void
 381 vdev_geom_io(struct g_consumer *cp, int *cmds, void **datas, off_t *offsets,
 382     off_t *sizes, int *errors, int ncmds)
 383 {
 384         struct bio **bios;
 385         u_char *p;
 386         off_t off, maxio, s, end;
 387         int i, n_bios, j;
 388         size_t bios_size;
 389
 390         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 391         n_bios = 0;
 392
 393         /* How many bios are required for all commands ? */
 394         for (i = 0; i < ncmds; i++)
 395                 n_bios += (sizes[i] + maxio - 1) / maxio;
 396
 397         /* Allocate memory for the bios */
 398         bios_size = n_bios * sizeof(struct bio*);
 399         bios = kmem_zalloc(bios_size, KM_SLEEP);
 400
 401         /* Prepare and issue all of the bios */
 402         for (i = j = 0; i < ncmds; i++) {
 403                 off = offsets[i];
 404                 p = datas[i];
 405                 s = sizes[i];
 406                 end = off + s;
 407                 ASSERT((off % cp->provider->sectorsize) == 0);
 408                 ASSERT((s % cp->provider->sectorsize) == 0);
 409
 410                 for (; off < end; off += maxio, p += maxio, s -= maxio, j++) {
 411                         bios[j] = g_alloc_bio();
 412                         bios[j]->bio_cmd = cmds[i];
 413                         bios[j]->bio_done = NULL;
 414                         bios[j]->bio_offset = off;
 415                         bios[j]->bio_length = MIN(s, maxio);
 416                         bios[j]->bio_data = p;
 417                         g_io_request(bios[j], cp);
 418                 }
 419         }
 420         ASSERT(j == n_bios);
 421
 422         /* Wait for all of the bios to complete, and clean them up */
 423         for (i = j = 0; i < ncmds; i++) {
 424                 off = offsets[i];
 425                 s = sizes[i];
 426                 end = off + s;
 427
 428                 for (; off < end; off += maxio, s -= maxio, j++) {
 429                         errors[i] = biowait(bios[j], "vdev_geom_io") || errors[i];
 430                         g_destroy_bio(bios[j]);
 431                 }
 432         }
 433         kmem_free(bios, bios_size);
 434 }
 435
 436 /*
 437  * Read the vdev config from a device.  Return the number of valid labels that
 438  * were found.  The vdev config will be returned in config if and only if at
 439  * least one valid label was found.
 440  */
 441 static int
 442 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **configp)
 443 {
 444         struct g_provider *pp;
 445         nvlist_t *config;
 446         vdev_phys_t *vdev_lists[VDEV_LABELS];
 447         char *buf;
 448         size_t buflen;
 449         uint64_t psize, state, txg;
 450         off_t offsets[VDEV_LABELS];
 451         off_t size;
 452         off_t sizes[VDEV_LABELS];
 453         int cmds[VDEV_LABELS];
 454         int errors[VDEV_LABELS];
 455         int l, nlabels;
 456
 457         g_topology_assert_not();
 458
 459         pp = cp->provider;
 460         ZFS_LOG(1, "Reading config from %s...", pp->name);
 461
 462         psize = pp->mediasize;
 463         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 464
 465         size = sizeof(*vdev_lists[0]) + pp->sectorsize -
 466             ((sizeof(*vdev_lists[0]) - 1) % pp->sectorsize) - 1;
 467
 468         buflen = sizeof(vdev_lists[0]->vp_nvlist);
 469
 470         /* Create all of the IO requests */
 471         for (l = 0; l < VDEV_LABELS; l++) {
 472                 cmds[l] = BIO_READ;
 473                 vdev_lists[l] = kmem_alloc(size, KM_SLEEP);
 474                 offsets[l] = vdev_label_offset(psize, l, 0) + VDEV_SKIP_SIZE;
 475                 sizes[l] = size;
 476                 errors[l] = 0;
 477                 ASSERT(offsets[l] % pp->sectorsize == 0);
 478         }
 479
 480         /* Issue the IO requests */
 481         vdev_geom_io(cp, cmds, (void**)vdev_lists, offsets, sizes, errors,
 482             VDEV_LABELS);
 483
 484         /* Parse the labels */
 485         config = *configp = NULL;
 486         nlabels = 0;
 487         for (l = 0; l < VDEV_LABELS; l++) {
 488                 if (errors[l] != 0)
 489                         continue;
 490
 491                 buf = vdev_lists[l]->vp_nvlist;
 492
 493                 if (nvlist_unpack(buf, buflen, &config, 0) != 0)
 494                         continue;
 495
 496                 if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE,
 497                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 498                         nvlist_free(config);
 499                         continue;
 500                 }
 501
 502                 if (state != POOL_STATE_SPARE &&
 503                     state != POOL_STATE_L2CACHE &&
 504                     (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG,
 505                     &txg) != 0 || txg == 0)) {
 506                         nvlist_free(config);
 507                         continue;
 508                 }
 509
 510                 if (*configp != NULL)
 511                         nvlist_free(*configp);
 512                 *configp = config;
 513
 514                 nlabels++;
 515         }
 516
 517         /* Free the label storage */
 518         for (l = 0; l < VDEV_LABELS; l++)
 519                 kmem_free(vdev_lists[l], size);
 520
 521         return (nlabels);
 522 }
 523
 524 static void
 525 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 526 {
 527         nvlist_t **new_configs;
 528         uint64_t i;
 529
 530         if (id < *count)
 531                 return;
 532         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 533             KM_SLEEP);
 534         for (i = 0; i < *count; i++)
 535                 new_configs[i] = (*configs)[i];
 536         if (*configs != NULL)
 537                 kmem_free(*configs, *count * sizeof(void *));
 538         *configs = new_configs;
 539         *count = id + 1;
 540 }
 541
 542 static void
 543 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 544     const char *name, uint64_t* known_pool_guid)
 545 {
 546         nvlist_t *vdev_tree;
 547         uint64_t pool_guid;
 548         uint64_t vdev_guid, known_guid;
 549         uint64_t id, txg, known_txg;
 550         char *pname;
 551         int i;
 552
 553         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 554             strcmp(pname, name) != 0)
 555                 goto ignore;
 556
 557         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 558                 goto ignore;
 559
 560         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 561                 goto ignore;
 562
 563         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 564                 goto ignore;
 565
 566         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 567                 goto ignore;
 568
 569         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 570
 571         if (*known_pool_guid != 0) {
 572                 if (pool_guid != *known_pool_guid)
 573                         goto ignore;
 574         } else
 575                 *known_pool_guid = pool_guid;
 576
 577         resize_configs(configs, count, id);
 578
 579         if ((*configs)[id] != NULL) {
 580                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 581                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 582                 if (txg <= known_txg)
 583                         goto ignore;
 584                 nvlist_free((*configs)[id]);
 585         }
 586
 587         (*configs)[id] = cfg;
 588         return;
 589
 590 ignore:
 591         nvlist_free(cfg);
 592 }
 593
 594 int
 595 vdev_geom_read_pool_label(const char *name,
 596     nvlist_t ***configs, uint64_t *count)
 597 {
 598         struct g_class *mp;
 599         struct g_geom *gp;
 600         struct g_provider *pp;
 601         struct g_consumer *zcp;
 602         nvlist_t *vdev_cfg;
 603         uint64_t pool_guid;
 604         int error, nlabels;
 605
 606         DROP_GIANT();
 607         g_topology_lock();
 608
 609         *configs = NULL;
 610         *count = 0;
 611         pool_guid = 0;
 612         LIST_FOREACH(mp, &g_classes, class) {
 613                 if (mp == &zfs_vdev_class)
 614                         continue;
 615                 LIST_FOREACH(gp, &mp->geom, geom) {
 616                         if (gp->flags & G_GEOM_WITHER)
 617                                 continue;
 618                         LIST_FOREACH(pp, &gp->provider, provider) {
 619                                 if (pp->flags & G_PF_WITHER)
 620                                         continue;
 621                                 zcp = vdev_geom_attach(pp, NULL, B_TRUE);
 622                                 if (zcp == NULL)
 623                                         continue;
 624                                 g_topology_unlock();
 625                                 nlabels = vdev_geom_read_config(zcp, &vdev_cfg);
 626                                 g_topology_lock();
 627                                 vdev_geom_detach(zcp, B_TRUE);
 628                                 if (nlabels == 0)
 629                                         continue;
 630                                 ZFS_LOG(1, "successfully read vdev config");
 631
 632                                 process_vdev_config(configs, count,
 633                                     vdev_cfg, name, &pool_guid);
 634                         }
 635                 }
 636         }
 637         g_topology_unlock();
 638         PICKUP_GIANT();
 639
 640         return (*count > 0 ? 0 : ENOENT);
 641 }
 642
 643 enum match {
 644         NO_MATCH = 0,           /* No matching labels found */
 645         TOPGUID_MATCH = 1,      /* Labels match top guid, not vdev guid*/
 646         ZERO_MATCH = 1,         /* Should never be returned */
 647         ONE_MATCH = 2,          /* 1 label matching the vdev_guid */
 648         TWO_MATCH = 3,          /* 2 label matching the vdev_guid */
 649         THREE_MATCH = 4,        /* 3 label matching the vdev_guid */
 650         FULL_MATCH = 5          /* all labels match the vdev_guid */
 651 };
 652
 653 static enum match
 654 vdev_attach_ok(vdev_t *vd, struct g_provider *pp)
 655 {
 656         nvlist_t *config;
 657         uint64_t pool_guid, top_guid, vdev_guid;
 658         struct g_consumer *cp;
 659         int nlabels;
 660
 661         cp = vdev_geom_attach(pp, NULL, B_TRUE);
 662         if (cp == NULL) {
 663                 ZFS_LOG(1, "Unable to attach tasting instance to %s.",
 664                     pp->name);
 665                 return (NO_MATCH);
 666         }
 667         g_topology_unlock();
 668         nlabels = vdev_geom_read_config(cp, &config);
 669         g_topology_lock();
 670         vdev_geom_detach(cp, B_TRUE);
 671         if (nlabels == 0) {
 672                 ZFS_LOG(1, "Unable to read config from %s.", pp->name);
 673                 return (NO_MATCH);
 674         }
 675
 676         pool_guid = 0;
 677         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid);
 678         top_guid = 0;
 679         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TOP_GUID, &top_guid);
 680         vdev_guid = 0;
 681         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_GUID, &vdev_guid);
 682         nvlist_free(config);
 683
 684         /*
 685          * Check that the label's pool guid matches the desired guid.
 686          * Inactive spares and L2ARCs do not have any pool guid in the label.
 687          */
 688         if (pool_guid != 0 && pool_guid != spa_guid(vd->vdev_spa)) {
 689                 ZFS_LOG(1, "pool guid mismatch for provider %s: %ju != %ju.",
 690                     pp->name,
 691                     (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)pool_guid);
 692                 return (NO_MATCH);
 693         }
 694
 695         /*
 696          * Check that the label's vdev guid matches the desired guid.
 697          * The second condition handles possible race on vdev detach, when
 698          * remaining vdev receives GUID of destroyed top level mirror vdev.
 699          */
 700         if (vdev_guid == vd->vdev_guid) {
 701                 ZFS_LOG(1, "guids match for provider %s.", pp->name);
 702                 return (ZERO_MATCH + nlabels);
 703         } else if (top_guid == vd->vdev_guid && vd == vd->vdev_top) {
 704                 ZFS_LOG(1, "top vdev guid match for provider %s.", pp->name);
 705                 return (TOPGUID_MATCH);
 706         }
 707         ZFS_LOG(1, "vdev guid mismatch for provider %s: %ju != %ju.",
 708             pp->name, (uintmax_t)vd->vdev_guid, (uintmax_t)vdev_guid);
 709         return (NO_MATCH);
 710 }
 711
 712 static struct g_consumer *
 713 vdev_geom_attach_by_guids(vdev_t *vd)
 714 {
 715         struct g_class *mp;
 716         struct g_geom *gp;
 717         struct g_provider *pp, *best_pp;
 718         struct g_consumer *cp;
 719         const char *vdpath;
 720         enum match match, best_match;
 721
 722         g_topology_assert();
 723
 724         vdpath = vd->vdev_path + sizeof("/dev/") - 1;
 725         cp = NULL;
 726         best_pp = NULL;
 727         best_match = NO_MATCH;
 728         LIST_FOREACH(mp, &g_classes, class) {
 729                 if (mp == &zfs_vdev_class)
 730                         continue;
 731                 LIST_FOREACH(gp, &mp->geom, geom) {
 732                         if (gp->flags & G_GEOM_WITHER)
 733                                 continue;
 734                         LIST_FOREACH(pp, &gp->provider, provider) {
 735                                 match = vdev_attach_ok(vd, pp);
 736                                 if (match > best_match) {
 737                                         best_match = match;
 738                                         best_pp = pp;
 739                                 } else if (match == best_match) {
 740                                         if (strcmp(pp->name, vdpath) == 0) {
 741                                                 best_pp = pp;
 742                                         }
 743                                 }
 744                                 if (match == FULL_MATCH)
 745                                         goto out;
 746                         }
 747                 }
 748         }
 749
 750 out:
 751         if (best_pp) {
 752                 cp = vdev_geom_attach(best_pp, vd, B_TRUE);
 753                 if (cp == NULL) {
 754                         printf("ZFS WARNING: Unable to attach to %s.\n",
 755                             best_pp->name);
 756                 }
 757         }
 758         return (cp);
 759 }
 760
 761 static struct g_consumer *
 762 vdev_geom_open_by_guids(vdev_t *vd)
 763 {
 764         struct g_consumer *cp;
 765         char *buf;
 766         size_t len;
 767
 768         g_topology_assert();
 769
 770         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 771                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 772         cp = vdev_geom_attach_by_guids(vd);
 773         if (cp != NULL) {
 774                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 775                 buf = kmem_alloc(len, KM_SLEEP);
 776
 777                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 778                 spa_strfree(vd->vdev_path);
 779                 vd->vdev_path = buf;
 780
 781                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 782                     (uintmax_t)spa_guid(vd->vdev_spa),
 783                     (uintmax_t)vd->vdev_guid, cp->provider->name);
 784         } else {
 785                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 786                     (uintmax_t)spa_guid(vd->vdev_spa),
 787                     (uintmax_t)vd->vdev_guid);
 788         }
 789
 790         return (cp);
 791 }
 792
 793 static struct g_consumer *
 794 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 795 {
 796         struct g_provider *pp;
 797         struct g_consumer *cp;
 798
 799         g_topology_assert();
 800
 801         cp = NULL;
 802         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 803         if (pp != NULL) {
 804                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 805                 if (!check_guid || vdev_attach_ok(vd, pp) == FULL_MATCH)
 806                         cp = vdev_geom_attach(pp, vd, B_FALSE);
 807         }
 808
 809         return (cp);
 810 }
 811
 812 static int
 813 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 814     uint64_t *logical_ashift, uint64_t *physical_ashift)
 815 {
 816         struct g_provider *pp;
 817         struct g_consumer *cp;
 818         size_t bufsize;
 819         int error;
 820
 821         /* Set the TLS to indicate downstack that we should not access zvols*/
 822         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 823
 824         /*
 825          * We must have a pathname, and it must be absolute.
 826          */
 827         if (vd->vdev_path == NULL || strncmp(vd->vdev_path, "/dev/", 5) != 0) {
 828                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 829                 return (EINVAL);
 830         }
 831
 832         /*
 833          * Reopen the device if it's not currently open. Otherwise,
 834          * just update the physical size of the device.
 835          */
 836         if ((cp = vd->vdev_tsd) != NULL) {
 837                 ASSERT(vd->vdev_reopening);
 838                 goto skip_open;
 839         }
 840
 841         DROP_GIANT();
 842         g_topology_lock();
 843         error = 0;
 844
 845         if (vd->vdev_spa->spa_splitting_newspa ||
 846             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 847              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 848              vd->vdev_spa->spa_load_state == SPA_LOAD_CREATE)) {
 849                 /*
 850                  * We are dealing with a vdev that hasn't been previously
 851                  * opened (since boot), and we are not loading an
 852                  * existing pool configuration.  This looks like a
 853                  * vdev add operation to a new or existing pool.
 854                  * Assume the user knows what he/she is doing and find
 855                  * GEOM provider by its name, ignoring GUID mismatches.
 856                  *
 857                  * XXPOLICY: It would be safer to only allow a device
 858                  *           that is unlabeled or labeled but missing
 859                  *           GUID information to be opened in this fashion,
 860                  *           unless we are doing a split, in which case we
 861                  *           should allow any guid.
 862                  */
 863                 cp = vdev_geom_open_by_path(vd, 0);
 864         } else {
 865                 /*
 866                  * Try using the recorded path for this device, but only
 867                  * accept it if its label data contains the expected GUIDs.
 868                  */
 869                 cp = vdev_geom_open_by_path(vd, 1);
 870                 if (cp == NULL) {
 871                         /*
 872                          * The device at vd->vdev_path doesn't have the
 873                          * expected GUIDs. The disks might have merely
 874                          * moved around so try all other GEOM providers
 875                          * to find one with the right GUIDs.
 876                          */
 877                         cp = vdev_geom_open_by_guids(vd);
 878                 }
 879         }
 880
 881         /* Clear the TLS now that tasting is done */
 882         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 883
 884         if (cp == NULL) {
 885                 ZFS_LOG(1, "Vdev %s not found.", vd->vdev_path);
 886                 error = ENOENT;
 887         } else {
 888                 struct consumer_priv_t *priv;
 889                 struct consumer_vdev_elem *elem;
 890                 int spamode;
 891
 892                 priv = (struct consumer_priv_t*)&cp->private;
 893                 if (cp->private == NULL)
 894                         SLIST_INIT(priv);
 895                 elem = g_malloc(sizeof(*elem), M_WAITOK|M_ZERO);
 896                 elem->vd = vd;
 897                 SLIST_INSERT_HEAD(priv, elem, elems);
 898
 899                 spamode = spa_mode(vd->vdev_spa);
 900                 if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 901                     !ISP2(cp->provider->sectorsize)) {
 902                         ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 903                             cp->provider->name);
 904
 905                         vdev_geom_close_locked(vd);
 906                         error = EINVAL;
 907                         cp = NULL;
 908                 } else if (cp->acw == 0 && (spamode & FWRITE) != 0) {
 909                         int i;
 910
 911                         for (i = 0; i < 5; i++) {
 912                                 error = g_access(cp, 0, 1, 0);
 913                                 if (error == 0)
 914                                         break;
 915                                 g_topology_unlock();
 916                                 tsleep(vd, 0, "vdev", hz / 2);
 917                                 g_topology_lock();
 918                         }
 919                         if (error != 0) {
 920                                 printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 921                                     cp->provider->name, error);
 922                                 vdev_geom_close_locked(vd);
 923                                 cp = NULL;
 924                         }
 925                 }
 926         }
 927
 928         /* Fetch initial physical path information for this device. */
 929         if (cp != NULL) {
 930                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 931
 932                 /* Set other GEOM characteristics */
 933                 vdev_geom_set_physpath(vd, cp, /*do_null_update*/B_FALSE);
 934                 vdev_geom_set_rotation_rate(vd, cp);
 935         }
 936
 937         g_topology_unlock();
 938         PICKUP_GIANT();
 939         if (cp == NULL) {
 940                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 941                 vdev_dbgmsg(vd, "vdev_geom_open: failed to open [error=%d]",
 942                     error);
 943                 return (error);
 944         }
 945 skip_open:
 946         pp = cp->provider;
 947
 948         /*
 949          * Determine the actual size of the device.
 950          */
 951         *max_psize = *psize = pp->mediasize;
 952
 953         /*
 954          * Determine the device's minimum transfer size and preferred
 955          * transfer size.
 956          */
 957         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 958         *physical_ashift = 0;
 959         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 960             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 961                 *physical_ashift = highbit(pp->stripesize) - 1;
 962
 963         /*
 964          * Clear the nowritecache settings, so that on a vdev_reopen()
 965          * we will try again.
 966          */
 967         vd->vdev_nowritecache = B_FALSE;
 968
 969         return (0);
 970 }
 971
 972 static void
 973 vdev_geom_close(vdev_t *vd)
 974 {
 975         struct g_consumer *cp;
 976
 977         cp = vd->vdev_tsd;
 978
 979         DROP_GIANT();
 980         g_topology_lock();
 981
 982         if (!vd->vdev_reopening ||
 983             (cp != NULL && ((cp->flags & G_CF_ORPHAN) != 0 ||
 984             (cp->provider != NULL && cp->provider->error != 0))))
 985                 vdev_geom_close_locked(vd);
 986
 987         g_topology_unlock();
 988         PICKUP_GIANT();
 989 }
 990
 991 static void
 992 vdev_geom_io_intr(struct bio *bp)
 993 {
 994         vdev_t *vd;
 995         zio_t *zio;
 996
 997         zio = bp->bio_caller1;
 998         vd = zio->io_vd;
 999         zio->io_error = bp->bio_error;
1000         if (zio->io_error == 0 && bp->bio_resid != 0)
1001                 zio->io_error = SET_ERROR(EIO);
1002
1003         switch(zio->io_error) {
1004         case ENOTSUP:
1005                 /*
1006                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
1007                  * that future attempts will never succeed. In this case
1008                  * we set a persistent flag so that we don't bother with
1009                  * requests in the future.
1010                  */
1011                 switch(bp->bio_cmd) {
1012                 case BIO_FLUSH:
1013                         vd->vdev_nowritecache = B_TRUE;
1014                         break;
1015                 case BIO_DELETE:
1016                         vd->vdev_notrim = B_TRUE;
1017                         break;
1018                 }
1019                 break;
1020         case ENXIO:
1021                 if (!vd->vdev_remove_wanted) {
1022                         /*
1023                          * If provider's error is set we assume it is being
1024                          * removed.
1025                          */
1026                         if (bp->bio_to->error != 0) {
1027                                 vd->vdev_remove_wanted = B_TRUE;
1028                                 spa_async_request(zio->io_spa,
1029                                     SPA_ASYNC_REMOVE);
1030                         } else if (!vd->vdev_delayed_close) {
1031                                 vd->vdev_delayed_close = B_TRUE;
1032                         }
1033                 }
1034                 break;
1035         }
1036
1037         /*
1038          * We have to split bio freeing into two parts, because the ABD code
1039          * cannot be called in this context and vdev_op_io_done is not called
1040          * for ZIO_TYPE_IOCTL zio-s.
1041          */
1042         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
1043                 g_destroy_bio(bp);
1044                 zio->io_bio = NULL;
1045         }
1046         zio_delay_interrupt(zio);
1047 }
1048
1049 static void
1050 vdev_geom_io_start(zio_t *zio)
1051 {
1052         vdev_t *vd;
1053         struct g_consumer *cp;
1054         struct bio *bp;
1055         int error;
1056
1057         vd = zio->io_vd;
1058
1059         switch (zio->io_type) {
1060         case ZIO_TYPE_IOCTL:
1061                 /* XXPOLICY */
1062                 if (!vdev_readable(vd)) {
1063                         zio->io_error = SET_ERROR(ENXIO);
1064                         zio_interrupt(zio);
1065                         return;
1066                 } else {
1067                         switch (zio->io_cmd) {
1068                         case DKIOCFLUSHWRITECACHE:
1069                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
1070                                         break;
1071                                 if (vd->vdev_nowritecache) {
1072                                         zio->io_error = SET_ERROR(ENOTSUP);
1073                                         break;
1074                                 }
1075                                 goto sendreq;
1076                         default:
1077                                 zio->io_error = SET_ERROR(ENOTSUP);
1078                         }
1079                 }
1080
1081                 zio_execute(zio);
1082                 return;
1083         case ZIO_TYPE_FREE:
1084                 if (vd->vdev_notrim) {
1085                         zio->io_error = SET_ERROR(ENOTSUP);
1086                 } else if (!vdev_geom_bio_delete_disable) {
1087                         goto sendreq;
1088                 }
1089                 zio_execute(zio);
1090                 return;
1091         }
1092 sendreq:
1093         ASSERT(zio->io_type == ZIO_TYPE_READ ||
1094             zio->io_type == ZIO_TYPE_WRITE ||
1095             zio->io_type == ZIO_TYPE_FREE ||
1096             zio->io_type == ZIO_TYPE_IOCTL);
1097
1098         cp = vd->vdev_tsd;
1099         if (cp == NULL) {
1100                 zio->io_error = SET_ERROR(ENXIO);
1101                 zio_interrupt(zio);
1102                 return;
1103         }
1104         bp = g_alloc_bio();
1105         bp->bio_caller1 = zio;
1106         switch (zio->io_type) {
1107         case ZIO_TYPE_READ:
1108         case ZIO_TYPE_WRITE:
1109                 zio->io_target_timestamp = zio_handle_io_delay(zio);
1110                 bp->bio_offset = zio->io_offset;
1111                 bp->bio_length = zio->io_size;
1112                 if (zio->io_type == ZIO_TYPE_READ) {
1113                         bp->bio_cmd = BIO_READ;
1114                         bp->bio_data =
1115                             abd_borrow_buf(zio->io_abd, zio->io_size);
1116                 } else {
1117                         bp->bio_cmd = BIO_WRITE;
1118                         bp->bio_data =
1119                             abd_borrow_buf_copy(zio->io_abd, zio->io_size);
1120                 }
1121                 break;
1122         case ZIO_TYPE_FREE:
1123                 bp->bio_cmd = BIO_DELETE;
1124                 bp->bio_data = NULL;
1125                 bp->bio_offset = zio->io_offset;
1126                 bp->bio_length = zio->io_size;
1127                 break;
1128         case ZIO_TYPE_IOCTL:
1129                 bp->bio_cmd = BIO_FLUSH;
1130                 bp->bio_data = NULL;
1131                 bp->bio_offset = cp->provider->mediasize;
1132                 bp->bio_length = 0;
1133                 break;
1134         }
1135         bp->bio_done = vdev_geom_io_intr;
1136         zio->io_bio = bp;
1137
1138         g_io_request(bp, cp);
1139 }
1140
1141 static void
1142 vdev_geom_io_done(zio_t *zio)
1143 {
1144         struct bio *bp = zio->io_bio;
1145
1146         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
1147                 ASSERT(bp == NULL);
1148                 return;
1149         }
1150
1151         if (bp == NULL) {
1152                 ASSERT3S(zio->io_error, ==, ENXIO);
1153                 return;
1154         }
1155
1156         if (zio->io_type == ZIO_TYPE_READ)
1157                 abd_return_buf_copy(zio->io_abd, bp->bio_data, zio->io_size);
1158         else
1159                 abd_return_buf(zio->io_abd, bp->bio_data, zio->io_size);
1160
1161         g_destroy_bio(bp);
1162         zio->io_bio = NULL;
1163 }
1164
1165 static void
1166 vdev_geom_hold(vdev_t *vd)
1167 {
1168 }
1169
1170 static void
1171 vdev_geom_rele(vdev_t *vd)
1172 {
1173 }
1174
1175 vdev_ops_t vdev_geom_ops = {
1176         vdev_geom_open,
1177         vdev_geom_close,
1178         vdev_default_asize,
1179         vdev_geom_io_start,
1180         vdev_geom_io_done,
1181         NULL,
1182         NULL,
1183         vdev_geom_hold,
1184         vdev_geom_rele,
1185         NULL,
1186         vdev_default_xlate,
1187         VDEV_TYPE_DISK,         /* name of this vdev type */
1188         B_TRUE                  /* leaf vdev */
1189 };