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

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