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