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