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\n", __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 consumer. Provider %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 to %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 static void
 342 nvlist_get_guids(nvlist_t *list, uint64_t *pguid, uint64_t *vguid)
 343 {
 344
 345         (void) nvlist_lookup_uint64(list, ZPOOL_CONFIG_GUID, vguid);
 346         (void) nvlist_lookup_uint64(list, ZPOOL_CONFIG_POOL_GUID, pguid);
 347 }
 348
 349 /*
 350  * Issue one or more bios to the vdev in parallel
 351  * cmds, datas, offsets, errors, and sizes are arrays of length ncmds.  Each IO
 352  * operation is described by parallel entries from each array.  There may be
 353  * more bios actually issued than entries in the array
 354  */
 355 static void
 356 vdev_geom_io(struct g_consumer *cp, int *cmds, void **datas, off_t *offsets,
 357     off_t *sizes, int *errors, int ncmds)
 358 {
 359         struct bio **bios;
 360         u_char *p;
 361         off_t off, maxio, s, end;
 362         int i, n_bios, j;
 363         size_t bios_size;
 364
 365         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 366         n_bios = 0;
 367
 368         /* How many bios are required for all commands ? */
 369         for (i = 0; i < ncmds; i++)
 370                 n_bios += (sizes[i] + maxio - 1) / maxio;
 371
 372         /* Allocate memory for the bios */
 373         bios_size = n_bios * sizeof(struct bio*);
 374         bios = kmem_zalloc(bios_size, KM_SLEEP);
 375
 376         /* Prepare and issue all of the bios */
 377         for (i = j = 0; i < ncmds; i++) {
 378                 off = offsets[i];
 379                 p = datas[i];
 380                 s = sizes[i];
 381                 end = off + s;
 382                 ASSERT((off % cp->provider->sectorsize) == 0);
 383                 ASSERT((s % cp->provider->sectorsize) == 0);
 384
 385                 for (; off < end; off += maxio, p += maxio, s -= maxio, j++) {
 386                         bios[j] = g_alloc_bio();
 387                         bios[j]->bio_cmd = cmds[i];
 388                         bios[j]->bio_done = NULL;
 389                         bios[j]->bio_offset = off;
 390                         bios[j]->bio_length = MIN(s, maxio);
 391                         bios[j]->bio_data = p;
 392                         g_io_request(bios[j], cp);
 393                 }
 394         }
 395         ASSERT(j == n_bios);
 396
 397         /* Wait for all of the bios to complete, and clean them up */
 398         for (i = j = 0; i < ncmds; i++) {
 399                 off = offsets[i];
 400                 s = sizes[i];
 401                 end = off + s;
 402
 403                 for (; off < end; off += maxio, s -= maxio, j++) {
 404                         errors[i] = biowait(bios[j], "vdev_geom_io") || errors[i];
 405                         g_destroy_bio(bios[j]);
 406                 }
 407         }
 408         kmem_free(bios, bios_size);
 409 }
 410
 411 static int
 412 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 413 {
 414         struct g_provider *pp;
 415         vdev_phys_t *vdev_lists[VDEV_LABELS];
 416         char *p, *buf;
 417         size_t buflen;
 418         uint64_t psize, state, txg;
 419         off_t offsets[VDEV_LABELS];
 420         off_t size;
 421         off_t sizes[VDEV_LABELS];
 422         int cmds[VDEV_LABELS];
 423         int errors[VDEV_LABELS];
 424         int l, len;
 425
 426         g_topology_assert_not();
 427
 428         pp = cp->provider;
 429         ZFS_LOG(1, "Reading config from %s...", pp->name);
 430
 431         psize = pp->mediasize;
 432         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 433
 434         size = sizeof(*vdev_lists[0]) + pp->sectorsize -
 435             ((sizeof(*vdev_lists[0]) - 1) % pp->sectorsize) - 1;
 436
 437         buflen = sizeof(vdev_lists[0]->vp_nvlist);
 438
 439         *config = NULL;
 440         /* Create all of the IO requests */
 441         for (l = 0; l < VDEV_LABELS; l++) {
 442                 cmds[l] = BIO_READ;
 443                 vdev_lists[l] = kmem_alloc(size, KM_SLEEP);
 444                 offsets[l] = vdev_label_offset(psize, l, 0) + VDEV_SKIP_SIZE;
 445                 sizes[l] = size;
 446                 errors[l] = 0;
 447                 ASSERT(offsets[l] % pp->sectorsize == 0);
 448         }
 449
 450         /* Issue the IO requests */
 451         vdev_geom_io(cp, cmds, (void**)vdev_lists, offsets, sizes, errors,
 452             VDEV_LABELS);
 453
 454         /* Parse the labels */
 455         for (l = 0; l < VDEV_LABELS; l++) {
 456                 if (errors[l] != 0)
 457                         continue;
 458
 459                 buf = vdev_lists[l]->vp_nvlist;
 460
 461                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 462                         continue;
 463
 464                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 465                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 466                         nvlist_free(*config);
 467                         *config = NULL;
 468                         continue;
 469                 }
 470
 471                 if (state != POOL_STATE_SPARE &&
 472                     state != POOL_STATE_L2CACHE &&
 473                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 474                     &txg) != 0 || txg == 0)) {
 475                         nvlist_free(*config);
 476                         *config = NULL;
 477                         continue;
 478                 }
 479
 480                 break;
 481         }
 482
 483         /* Free the label storage */
 484         for (l = 0; l < VDEV_LABELS; l++)
 485                 kmem_free(vdev_lists[l], size);
 486
 487         return (*config == NULL ? ENOENT : 0);
 488 }
 489
 490 static void
 491 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 492 {
 493         nvlist_t **new_configs;
 494         uint64_t i;
 495
 496         if (id < *count)
 497                 return;
 498         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 499             KM_SLEEP);
 500         for (i = 0; i < *count; i++)
 501                 new_configs[i] = (*configs)[i];
 502         if (*configs != NULL)
 503                 kmem_free(*configs, *count * sizeof(void *));
 504         *configs = new_configs;
 505         *count = id + 1;
 506 }
 507
 508 static void
 509 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 510     const char *name, uint64_t* known_pool_guid)
 511 {
 512         nvlist_t *vdev_tree;
 513         uint64_t pool_guid;
 514         uint64_t vdev_guid, known_guid;
 515         uint64_t id, txg, known_txg;
 516         char *pname;
 517         int i;
 518
 519         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 520             strcmp(pname, name) != 0)
 521                 goto ignore;
 522
 523         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 524                 goto ignore;
 525
 526         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 527                 goto ignore;
 528
 529         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 530                 goto ignore;
 531
 532         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 533                 goto ignore;
 534
 535         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 536
 537         if (*known_pool_guid != 0) {
 538                 if (pool_guid != *known_pool_guid)
 539                         goto ignore;
 540         } else
 541                 *known_pool_guid = pool_guid;
 542
 543         resize_configs(configs, count, id);
 544
 545         if ((*configs)[id] != NULL) {
 546                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 547                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 548                 if (txg <= known_txg)
 549                         goto ignore;
 550                 nvlist_free((*configs)[id]);
 551         }
 552
 553         (*configs)[id] = cfg;
 554         return;
 555
 556 ignore:
 557         nvlist_free(cfg);
 558 }
 559
 560 int
 561 vdev_geom_read_pool_label(const char *name,
 562     nvlist_t ***configs, uint64_t *count)
 563 {
 564         struct g_class *mp;
 565         struct g_geom *gp;
 566         struct g_provider *pp;
 567         struct g_consumer *zcp;
 568         nvlist_t *vdev_cfg;
 569         uint64_t pool_guid;
 570         int error;
 571
 572         DROP_GIANT();
 573         g_topology_lock();
 574
 575         *configs = NULL;
 576         *count = 0;
 577         pool_guid = 0;
 578         LIST_FOREACH(mp, &g_classes, class) {
 579                 if (mp == &zfs_vdev_class)
 580                         continue;
 581                 LIST_FOREACH(gp, &mp->geom, geom) {
 582                         if (gp->flags & G_GEOM_WITHER)
 583                                 continue;
 584                         LIST_FOREACH(pp, &gp->provider, provider) {
 585                                 if (pp->flags & G_PF_WITHER)
 586                                         continue;
 587                                 zcp = vdev_geom_attach(pp, NULL);
 588                                 if (zcp == NULL)
 589                                         continue;
 590                                 g_topology_unlock();
 591                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 592                                 g_topology_lock();
 593                                 vdev_geom_detach(zcp, B_TRUE);
 594                                 if (error)
 595                                         continue;
 596                                 ZFS_LOG(1, "successfully read vdev config");
 597
 598                                 process_vdev_config(configs, count,
 599                                     vdev_cfg, name, &pool_guid);
 600                         }
 601                 }
 602         }
 603         g_topology_unlock();
 604         PICKUP_GIANT();
 605
 606         return (*count > 0 ? 0 : ENOENT);
 607 }
 608
 609 static void
 610 vdev_geom_read_guids(struct g_consumer *cp, uint64_t *pguid, uint64_t *vguid)
 611 {
 612         nvlist_t *config;
 613
 614         g_topology_assert_not();
 615
 616         *pguid = 0;
 617         *vguid = 0;
 618         if (vdev_geom_read_config(cp, &config) == 0) {
 619                 nvlist_get_guids(config, pguid, vguid);
 620                 nvlist_free(config);
 621         }
 622 }
 623
 624 static boolean_t
 625 vdev_attach_ok(vdev_t *vd, struct g_provider *pp)
 626 {
 627         uint64_t pool_guid;
 628         uint64_t vdev_guid;
 629         struct g_consumer *zcp;
 630         boolean_t pool_ok;
 631         boolean_t vdev_ok;
 632
 633         zcp = vdev_geom_attach(pp, NULL);
 634         if (zcp == NULL) {
 635                 ZFS_LOG(1, "Unable to attach tasting instance to %s.",
 636                     pp->name);
 637                 return (B_FALSE);
 638         }
 639         g_topology_unlock();
 640         vdev_geom_read_guids(zcp, &pool_guid, &vdev_guid);
 641         g_topology_lock();
 642         vdev_geom_detach(zcp, B_TRUE);
 643
 644         /*
 645          * Check that the label's vdev guid matches the desired guid.  If the
 646          * label has a pool guid, check that it matches too. (Inactive spares
 647          * and L2ARCs do not have any pool guid in the label.)
 648          */
 649         if ((pool_guid == 0 || pool_guid == spa_guid(vd->vdev_spa)) &&
 650             vdev_guid == vd->vdev_guid) {
 651                 ZFS_LOG(1, "guids match for provider %s.", vd->vdev_path);
 652                 return (B_TRUE);
 653         } else {
 654                 ZFS_LOG(1, "guid mismatch for provider %s: "
 655                     "%ju:%ju != %ju:%ju.", vd->vdev_path,
 656                     (uintmax_t)spa_guid(vd->vdev_spa),
 657                     (uintmax_t)vd->vdev_guid,
 658                     (uintmax_t)pool_guid, (uintmax_t)vdev_guid);
 659                 return (B_FALSE);
 660         }
 661 }
 662
 663 static struct g_consumer *
 664 vdev_geom_attach_by_guids(vdev_t *vd)
 665 {
 666         struct g_class *mp;
 667         struct g_geom *gp;
 668         struct g_provider *pp;
 669         struct g_consumer *cp;
 670
 671         g_topology_assert();
 672
 673         cp = NULL;
 674         LIST_FOREACH(mp, &g_classes, class) {
 675                 if (mp == &zfs_vdev_class)
 676                         continue;
 677                 LIST_FOREACH(gp, &mp->geom, geom) {
 678                         if (gp->flags & G_GEOM_WITHER)
 679                                 continue;
 680                         LIST_FOREACH(pp, &gp->provider, provider) {
 681                                 if (!vdev_attach_ok(vd, pp))
 682                                         continue;
 683                                 cp = vdev_geom_attach(pp, vd);
 684                                 if (cp == NULL) {
 685                                         printf("ZFS WARNING: Unable to "
 686                                             "attach to %s.\n", pp->name);
 687                                         continue;
 688                                 }
 689                                 break;
 690                         }
 691                         if (cp != NULL)
 692                                 break;
 693                 }
 694                 if (cp != NULL)
 695                         break;
 696         }
 697 end:
 698         return (cp);
 699 }
 700
 701 static struct g_consumer *
 702 vdev_geom_open_by_guids(vdev_t *vd)
 703 {
 704         struct g_consumer *cp;
 705         char *buf;
 706         size_t len;
 707
 708         g_topology_assert();
 709
 710         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 711                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 712         cp = vdev_geom_attach_by_guids(vd);
 713         if (cp != NULL) {
 714                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 715                 buf = kmem_alloc(len, KM_SLEEP);
 716
 717                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 718                 spa_strfree(vd->vdev_path);
 719                 vd->vdev_path = buf;
 720
 721                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 722                     (uintmax_t)spa_guid(vd->vdev_spa),
 723                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 724         } else {
 725                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 726                     (uintmax_t)spa_guid(vd->vdev_spa),
 727                     (uintmax_t)vd->vdev_guid);
 728         }
 729
 730         return (cp);
 731 }
 732
 733 static struct g_consumer *
 734 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 735 {
 736         struct g_provider *pp;
 737         struct g_consumer *cp;
 738
 739         g_topology_assert();
 740
 741         cp = NULL;
 742         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 743         if (pp != NULL) {
 744                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 745                 if (!check_guid || vdev_attach_ok(vd, pp))
 746                         cp = vdev_geom_attach(pp, vd);
 747         }
 748
 749         return (cp);
 750 }
 751
 752 static int
 753 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 754     uint64_t *logical_ashift, uint64_t *physical_ashift)
 755 {
 756         struct g_provider *pp;
 757         struct g_consumer *cp;
 758         size_t bufsize;
 759         int error;
 760
 761         /* Set the TLS to indicate downstack that we should not access zvols*/
 762         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 763
 764         /*
 765          * We must have a pathname, and it must be absolute.
 766          */
 767         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 768                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 769                 return (EINVAL);
 770         }
 771
 772         vd->vdev_tsd = NULL;
 773
 774         DROP_GIANT();
 775         g_topology_lock();
 776         error = 0;
 777
 778         if (vd->vdev_spa->spa_splitting_newspa ||
 779             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 780              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE)) {
 781                 /*
 782                  * We are dealing with a vdev that hasn't been previously
 783                  * opened (since boot), and we are not loading an
 784                  * existing pool configuration.  This looks like a
 785                  * vdev add operation to a new or existing pool.
 786                  * Assume the user knows what he/she is doing and find
 787                  * GEOM provider by its name, ignoring GUID mismatches.
 788                  *
 789                  * XXPOLICY: It would be safer to only allow a device
 790                  *           that is unlabeled or labeled but missing
 791                  *           GUID information to be opened in this fashion,
 792                  *           unless we are doing a split, in which case we
 793                  *           should allow any guid.
 794                  */
 795                 cp = vdev_geom_open_by_path(vd, 0);
 796         } else {
 797                 /*
 798                  * Try using the recorded path for this device, but only
 799                  * accept it if its label data contains the expected GUIDs.
 800                  */
 801                 cp = vdev_geom_open_by_path(vd, 1);
 802                 if (cp == NULL) {
 803                         /*
 804                          * The device at vd->vdev_path doesn't have the
 805                          * expected GUIDs. The disks might have merely
 806                          * moved around so try all other GEOM providers
 807                          * to find one with the right GUIDs.
 808                          */
 809                         cp = vdev_geom_open_by_guids(vd);
 810                 }
 811         }
 812
 813         /* Clear the TLS now that tasting is done */
 814         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 815
 816         if (cp == NULL) {
 817                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 818                 error = ENOENT;
 819         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 820             !ISP2(cp->provider->sectorsize)) {
 821                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 822                     vd->vdev_path);
 823
 824                 vdev_geom_close_locked(vd);
 825                 error = EINVAL;
 826                 cp = NULL;
 827         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 828                 int i;
 829
 830                 for (i = 0; i < 5; i++) {
 831                         error = g_access(cp, 0, 1, 0);
 832                         if (error == 0)
 833                                 break;
 834                         g_topology_unlock();
 835                         tsleep(vd, 0, "vdev", hz / 2);
 836                         g_topology_lock();
 837                 }
 838                 if (error != 0) {
 839                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 840                             vd->vdev_path, error);
 841                         vdev_geom_close_locked(vd);
 842                         cp = NULL;
 843                 }
 844         }
 845
 846         /* Fetch initial physical path information for this device. */
 847         if (cp != NULL)
 848                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 849
 850         g_topology_unlock();
 851         PICKUP_GIANT();
 852         if (cp == NULL) {
 853                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 854                 return (error);
 855         }
 856         pp = cp->provider;
 857
 858         /*
 859          * Determine the actual size of the device.
 860          */
 861         *max_psize = *psize = pp->mediasize;
 862
 863         /*
 864          * Determine the device's minimum transfer size and preferred
 865          * transfer size.
 866          */
 867         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 868         *physical_ashift = 0;
 869         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 870             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 871                 *physical_ashift = highbit(pp->stripesize) - 1;
 872
 873         /*
 874          * Clear the nowritecache settings, so that on a vdev_reopen()
 875          * we will try again.
 876          */
 877         vd->vdev_nowritecache = B_FALSE;
 878
 879         /*
 880          * Determine the device's rotation rate.
 881          */
 882         vdev_geom_set_rotation_rate(vd, cp);
 883
 884         return (0);
 885 }
 886
 887 static void
 888 vdev_geom_close(vdev_t *vd)
 889 {
 890
 891         DROP_GIANT();
 892         g_topology_lock();
 893         vdev_geom_close_locked(vd);
 894         g_topology_unlock();
 895         PICKUP_GIANT();
 896 }
 897
 898 static void
 899 vdev_geom_io_intr(struct bio *bp)
 900 {
 901         vdev_t *vd;
 902         zio_t *zio;
 903
 904         zio = bp->bio_caller1;
 905         vd = zio->io_vd;
 906         zio->io_error = bp->bio_error;
 907         if (zio->io_error == 0 && bp->bio_resid != 0)
 908                 zio->io_error = SET_ERROR(EIO);
 909
 910         switch(zio->io_error) {
 911         case ENOTSUP:
 912                 /*
 913                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 914                  * that future attempts will never succeed. In this case
 915                  * we set a persistent flag so that we don't bother with
 916                  * requests in the future.
 917                  */
 918                 switch(bp->bio_cmd) {
 919                 case BIO_FLUSH:
 920                         vd->vdev_nowritecache = B_TRUE;
 921                         break;
 922                 case BIO_DELETE:
 923                         vd->vdev_notrim = B_TRUE;
 924                         break;
 925                 }
 926                 break;
 927         case ENXIO:
 928                 if (!vd->vdev_remove_wanted) {
 929                         /*
 930                          * If provider's error is set we assume it is being
 931                          * removed.
 932                          */
 933                         if (bp->bio_to->error != 0) {
 934                                 vd->vdev_remove_wanted = B_TRUE;
 935                                 spa_async_request(zio->io_spa,
 936                                     SPA_ASYNC_REMOVE);
 937                         } else if (!vd->vdev_delayed_close) {
 938                                 vd->vdev_delayed_close = B_TRUE;
 939                         }
 940                 }
 941                 break;
 942         }
 943         g_destroy_bio(bp);
 944         zio_delay_interrupt(zio);
 945 }
 946
 947 static void
 948 vdev_geom_io_start(zio_t *zio)
 949 {
 950         vdev_t *vd;
 951         struct g_consumer *cp;
 952         struct bio *bp;
 953         int error;
 954
 955         vd = zio->io_vd;
 956
 957         switch (zio->io_type) {
 958         case ZIO_TYPE_IOCTL:
 959                 /* XXPOLICY */
 960                 if (!vdev_readable(vd)) {
 961                         zio->io_error = SET_ERROR(ENXIO);
 962                         zio_interrupt(zio);
 963                         return;
 964                 } else {
 965                         switch (zio->io_cmd) {
 966                         case DKIOCFLUSHWRITECACHE:
 967                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 968                                         break;
 969                                 if (vd->vdev_nowritecache) {
 970                                         zio->io_error = SET_ERROR(ENOTSUP);
 971                                         break;
 972                                 }
 973                                 goto sendreq;
 974                         default:
 975                                 zio->io_error = SET_ERROR(ENOTSUP);
 976                         }
 977                 }
 978
 979                 zio_execute(zio);
 980                 return;
 981         case ZIO_TYPE_FREE:
 982                 if (vd->vdev_notrim) {
 983                         zio->io_error = SET_ERROR(ENOTSUP);
 984                 } else if (!vdev_geom_bio_delete_disable) {
 985                         goto sendreq;
 986                 }
 987                 zio_execute(zio);
 988                 return;
 989         }
 990 sendreq:
 991         ASSERT(zio->io_type == ZIO_TYPE_READ ||
 992             zio->io_type == ZIO_TYPE_WRITE ||
 993             zio->io_type == ZIO_TYPE_FREE ||
 994             zio->io_type == ZIO_TYPE_IOCTL);
 995
 996         cp = vd->vdev_tsd;
 997         if (cp == NULL) {
 998                 zio->io_error = SET_ERROR(ENXIO);
 999                 zio_interrupt(zio);
1000                 return;
1001         }
1002         bp = g_alloc_bio();
1003         bp->bio_caller1 = zio;
1004         switch (zio->io_type) {
1005         case ZIO_TYPE_READ:
1006         case ZIO_TYPE_WRITE:
1007                 zio->io_target_timestamp = zio_handle_io_delay(zio);
1008                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
1009                 bp->bio_data = zio->io_data;
1010                 bp->bio_offset = zio->io_offset;
1011                 bp->bio_length = zio->io_size;
1012                 break;
1013         case ZIO_TYPE_FREE:
1014                 bp->bio_cmd = BIO_DELETE;
1015                 bp->bio_data = NULL;
1016                 bp->bio_offset = zio->io_offset;
1017                 bp->bio_length = zio->io_size;
1018                 break;
1019         case ZIO_TYPE_IOCTL:
1020                 bp->bio_cmd = BIO_FLUSH;
1021                 bp->bio_flags |= BIO_ORDERED;
1022                 bp->bio_data = NULL;
1023                 bp->bio_offset = cp->provider->mediasize;
1024                 bp->bio_length = 0;
1025                 break;
1026         }
1027         bp->bio_done = vdev_geom_io_intr;
1028
1029         g_io_request(bp, cp);
1030 }
1031
1032 static void
1033 vdev_geom_io_done(zio_t *zio)
1034 {
1035 }
1036
1037 static void
1038 vdev_geom_hold(vdev_t *vd)
1039 {
1040 }
1041
1042 static void
1043 vdev_geom_rele(vdev_t *vd)
1044 {
1045 }
1046
1047 vdev_ops_t vdev_geom_ops = {
1048         vdev_geom_open,
1049         vdev_geom_close,
1050         vdev_default_asize,
1051         vdev_geom_io_start,
1052         vdev_geom_io_done,
1053         NULL,
1054         vdev_geom_hold,
1055         vdev_geom_rele,
1056         VDEV_TYPE_DISK,         /* name of this vdev type */
1057         B_TRUE                  /* leaf vdev */
1058 };