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