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