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 struct g_class zfs_vdev_class = {
  46         .name = "ZFS::VDEV",
  47         .version = G_VERSION,
  48 };
  49
  50 DECLARE_GEOM_CLASS(zfs_vdev_class, zfs_vdev);
  51
  52 /*
  53  * Don't send BIO_FLUSH.
  54  */
  55 static int vdev_geom_bio_flush_disable = 0;
  56 TUNABLE_INT("vfs.zfs.vdev.bio_flush_disable", &vdev_geom_bio_flush_disable);
  57 SYSCTL_DECL(_vfs_zfs_vdev);
  58 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_flush_disable, CTLFLAG_RW,
  59     &vdev_geom_bio_flush_disable, 0, "Disable BIO_FLUSH");
  60
  61 static void
  62 vdev_geom_orphan(struct g_consumer *cp)
  63 {
  64         vdev_t *vd;
  65
  66         g_topology_assert();
  67
  68         vd = cp->private;
  69
  70         /*
  71          * Orphan callbacks occur from the GEOM event thread.
  72          * Concurrent with this call, new I/O requests may be
  73          * working their way through GEOM about to find out
  74          * (only once executed by the g_down thread) that we've
  75          * been orphaned from our disk provider.  These I/Os
  76          * must be retired before we can detach our consumer.
  77          * This is most easily achieved by acquiring the
  78          * SPA ZIO configuration lock as a writer, but doing
  79          * so with the GEOM topology lock held would cause
  80          * a lock order reversal.  Instead, rely on the SPA's
  81          * async removal support to invoke a close on this
  82          * vdev once it is safe to do so.
  83          */
  84         zfs_post_remove(vd->vdev_spa, vd);
  85         vd->vdev_remove_wanted = B_TRUE;
  86         spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
  87 }
  88
  89 static struct g_consumer *
  90 vdev_geom_attach(struct g_provider *pp)
  91 {
  92         struct g_geom *gp;
  93         struct g_consumer *cp;
  94
  95         g_topology_assert();
  96
  97         ZFS_LOG(1, "Attaching to %s.", pp->name);
  98         /* Do we have geom already? No? Create one. */
  99         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 100                 if (gp->flags & G_GEOM_WITHER)
 101                         continue;
 102                 if (strcmp(gp->name, "zfs::vdev") != 0)
 103                         continue;
 104                 break;
 105         }
 106         if (gp == NULL) {
 107                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 108                 gp->orphan = vdev_geom_orphan;
 109                 cp = g_new_consumer(gp);
 110                 if (g_attach(cp, pp) != 0) {
 111                         g_wither_geom(gp, ENXIO);
 112                         return (NULL);
 113                 }
 114                 if (g_access(cp, 1, 0, 1) != 0) {
 115                         g_wither_geom(gp, ENXIO);
 116                         return (NULL);
 117                 }
 118                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 119         } else {
 120                 /* Check if we are already connected to this provider. */
 121                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 122                         if (cp->provider == pp) {
 123                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 124                                 break;
 125                         }
 126                 }
 127                 if (cp == NULL) {
 128                         cp = g_new_consumer(gp);
 129                         if (g_attach(cp, pp) != 0) {
 130                                 g_destroy_consumer(cp);
 131                                 return (NULL);
 132                         }
 133                         if (g_access(cp, 1, 0, 1) != 0) {
 134                                 g_detach(cp);
 135                                 g_destroy_consumer(cp);
 136                                 return (NULL);
 137                         }
 138                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 139                 } else {
 140                         if (g_access(cp, 1, 0, 1) != 0)
 141                                 return (NULL);
 142                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 143                 }
 144         }
 145         return (cp);
 146 }
 147
 148 static void
 149 vdev_geom_detach(void *arg, int flag __unused)
 150 {
 151         struct g_geom *gp;
 152         struct g_consumer *cp;
 153
 154         g_topology_assert();
 155         cp = arg;
 156         gp = cp->geom;
 157
 158         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 159         g_access(cp, -1, 0, -1);
 160         /* Destroy consumer on last close. */
 161         if (cp->acr == 0 && cp->ace == 0) {
 162                 ZFS_LOG(1, "Destroyed consumer to %s.", cp->provider->name);
 163                 if (cp->acw > 0)
 164                         g_access(cp, 0, -cp->acw, 0);
 165                 g_detach(cp);
 166                 g_destroy_consumer(cp);
 167         }
 168         /* Destroy geom if there are no consumers left. */
 169         if (LIST_EMPTY(&gp->consumer)) {
 170                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 171                 g_wither_geom(gp, ENXIO);
 172         }
 173 }
 174
 175 static uint64_t
 176 nvlist_get_guid(nvlist_t *list)
 177 {
 178         uint64_t value;
 179
 180         value = 0;
 181         nvlist_lookup_uint64(list, ZPOOL_CONFIG_GUID, &value);
 182         return (value);
 183 }
 184
 185 static int
 186 vdev_geom_io(struct g_consumer *cp, int cmd, void *data, off_t offset, off_t size)
 187 {
 188         struct bio *bp;
 189         u_char *p;
 190         off_t off, maxio;
 191         int error;
 192
 193         ASSERT((offset % cp->provider->sectorsize) == 0);
 194         ASSERT((size % cp->provider->sectorsize) == 0);
 195
 196         bp = g_alloc_bio();
 197         off = offset;
 198         offset += size;
 199         p = data;
 200         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 201         error = 0;
 202
 203         for (; off < offset; off += maxio, p += maxio, size -= maxio) {
 204                 bzero(bp, sizeof(*bp));
 205                 bp->bio_cmd = cmd;
 206                 bp->bio_done = NULL;
 207                 bp->bio_offset = off;
 208                 bp->bio_length = MIN(size, maxio);
 209                 bp->bio_data = p;
 210                 g_io_request(bp, cp);
 211                 error = biowait(bp, "vdev_geom_io");
 212                 if (error != 0)
 213                         break;
 214         }
 215
 216         g_destroy_bio(bp);
 217         return (error);
 218 }
 219
 220 static void
 221 vdev_geom_taste_orphan(struct g_consumer *cp)
 222 {
 223
 224         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
 225             cp->provider->name));
 226 }
 227
 228 static int
 229 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 230 {
 231         struct g_provider *pp;
 232         vdev_label_t *label;
 233         char *p, *buf;
 234         size_t buflen;
 235         uint64_t psize;
 236         off_t offset, size;
 237         uint64_t guid, state, txg;
 238         int error, l, len;
 239
 240         g_topology_assert_not();
 241
 242         pp = cp->provider;
 243         ZFS_LOG(1, "Reading config from %s...", pp->name);
 244
 245         psize = pp->mediasize;
 246         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 247
 248         size = sizeof(*label) + pp->sectorsize -
 249             ((sizeof(*label) - 1) % pp->sectorsize) - 1;
 250
 251         guid = 0;
 252         label = kmem_alloc(size, KM_SLEEP);
 253         buflen = sizeof(label->vl_vdev_phys.vp_nvlist);
 254
 255         *config = NULL;
 256         for (l = 0; l < VDEV_LABELS; l++) {
 257
 258                 offset = vdev_label_offset(psize, l, 0);
 259                 if ((offset % pp->sectorsize) != 0)
 260                         continue;
 261
 262                 if (vdev_geom_io(cp, BIO_READ, label, offset, size) != 0)
 263                         continue;
 264                 buf = label->vl_vdev_phys.vp_nvlist;
 265
 266                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 267                         continue;
 268
 269                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 270                     &state) != 0 || state == POOL_STATE_DESTROYED ||
 271                     state > POOL_STATE_L2CACHE) {
 272                         nvlist_free(*config);
 273                         *config = NULL;
 274                         continue;
 275                 }
 276
 277                 if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
 278                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 279                     &txg) != 0 || txg == 0)) {
 280                         nvlist_free(*config);
 281                         *config = NULL;
 282                         continue;
 283                 }
 284
 285                 break;
 286         }
 287
 288         kmem_free(label, size);
 289         return (*config == NULL ? ENOENT : 0);
 290 }
 291
 292 static int
 293 vdev_geom_check_config(nvlist_t *config, const char *name, uint64_t *best_txg)
 294 {
 295         uint64_t vdev_guid;
 296         uint64_t txg;
 297         char *pname;
 298
 299         if (nvlist_lookup_string(config, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 300             strcmp(pname, name) != 0)
 301                 return (ENOENT);
 302
 303         ZFS_LOG(1, "found pool: %s", pname);
 304
 305         txg = 0;
 306         nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG, &txg);
 307         if (txg <= *best_txg)
 308                 return (ENOENT);
 309         *best_txg = txg;
 310         ZFS_LOG(1, "txg: %ju", (uintmax_t)*best_txg);
 311
 312         return (0);
 313 }
 314
 315 static int
 316 vdev_geom_attach_taster(struct g_consumer *cp, struct g_provider *pp)
 317 {
 318         int error;
 319
 320         if (pp->flags & G_PF_WITHER)
 321                 return (EINVAL);
 322         if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize))
 323                 return (EINVAL);
 324         g_attach(cp, pp);
 325         error = g_access(cp, 1, 0, 0);
 326         if (error != 0)
 327                 g_detach(cp);
 328         return (error);
 329 }
 330
 331 static void
 332 vdev_geom_dettach_taster(struct g_consumer *cp)
 333 {
 334         g_access(cp, -1, 0, 0);
 335         g_detach(cp);
 336 }
 337
 338 int
 339 vdev_geom_read_pool_label(const char *name, nvlist_t **config)
 340 {
 341         struct g_class *mp;
 342         struct g_geom *gp, *zgp;
 343         struct g_provider *pp;
 344         struct g_consumer *zcp;
 345         nvlist_t *vdev_cfg;
 346         uint64_t best_txg;
 347         int error;
 348
 349         DROP_GIANT();
 350         g_topology_lock();
 351
 352         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 353         /* This orphan function should be never called. */
 354         zgp->orphan = vdev_geom_taste_orphan;
 355         zcp = g_new_consumer(zgp);
 356
 357         best_txg = 0;
 358         *config = NULL;
 359         LIST_FOREACH(mp, &g_classes, class) {
 360                 if (mp == &zfs_vdev_class)
 361                         continue;
 362                 LIST_FOREACH(gp, &mp->geom, geom) {
 363                         if (gp->flags & G_GEOM_WITHER)
 364                                 continue;
 365                         LIST_FOREACH(pp, &gp->provider, provider) {
 366                                 if (pp->flags & G_PF_WITHER)
 367                                         continue;
 368                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 369                                         continue;
 370                                 g_topology_unlock();
 371                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 372                                 g_topology_lock();
 373                                 vdev_geom_dettach_taster(zcp);
 374                                 if (error)
 375                                         continue;
 376                                 ZFS_LOG(1, "successfully read vdev config");
 377
 378                                 error = vdev_geom_check_config(vdev_cfg, name,
 379                                     &best_txg);
 380                                 if (error != 0) {
 381                                         nvlist_free(vdev_cfg);
 382                                         continue;
 383                                 }
 384                                 nvlist_free(*config);
 385                                 *config = vdev_cfg;
 386                         }
 387                 }
 388         }
 389
 390         g_destroy_consumer(zcp);
 391         g_destroy_geom(zgp);
 392         g_topology_unlock();
 393         PICKUP_GIANT();
 394         return (*config == NULL ? ENOENT : 0);
 395 }
 396
 397 static uint64_t
 398 vdev_geom_read_guid(struct g_consumer *cp)
 399 {
 400         nvlist_t *config;
 401         uint64_t guid;
 402
 403         g_topology_assert_not();
 404
 405         guid = 0;
 406         if (vdev_geom_read_config(cp, &config) == 0) {
 407                 guid = nvlist_get_guid(config);
 408                 nvlist_free(config);
 409         }
 410         return (guid);
 411 }
 412
 413 static struct g_consumer *
 414 vdev_geom_attach_by_guid(uint64_t guid)
 415 {
 416         struct g_class *mp;
 417         struct g_geom *gp, *zgp;
 418         struct g_provider *pp;
 419         struct g_consumer *cp, *zcp;
 420         uint64_t pguid;
 421
 422         g_topology_assert();
 423
 424         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 425         /* This orphan function should be never called. */
 426         zgp->orphan = vdev_geom_taste_orphan;
 427         zcp = g_new_consumer(zgp);
 428
 429         cp = NULL;
 430         LIST_FOREACH(mp, &g_classes, class) {
 431                 if (mp == &zfs_vdev_class)
 432                         continue;
 433                 LIST_FOREACH(gp, &mp->geom, geom) {
 434                         if (gp->flags & G_GEOM_WITHER)
 435                                 continue;
 436                         LIST_FOREACH(pp, &gp->provider, provider) {
 437                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 438                                         continue;
 439                                 g_topology_unlock();
 440                                 pguid = vdev_geom_read_guid(zcp);
 441                                 g_topology_lock();
 442                                 vdev_geom_dettach_taster(zcp);
 443                                 if (pguid != guid)
 444                                         continue;
 445                                 cp = vdev_geom_attach(pp);
 446                                 if (cp == NULL) {
 447                                         printf("ZFS WARNING: Unable to attach to %s.\n",
 448                                             pp->name);
 449                                         continue;
 450                                 }
 451                                 break;
 452                         }
 453                         if (cp != NULL)
 454                                 break;
 455                 }
 456                 if (cp != NULL)
 457                         break;
 458         }
 459 end:
 460         g_destroy_consumer(zcp);
 461         g_destroy_geom(zgp);
 462         return (cp);
 463 }
 464
 465 static struct g_consumer *
 466 vdev_geom_open_by_guid(vdev_t *vd)
 467 {
 468         struct g_consumer *cp;
 469         char *buf;
 470         size_t len;
 471
 472         g_topology_assert();
 473
 474         ZFS_LOG(1, "Searching by guid [%ju].", (uintmax_t)vd->vdev_guid);
 475         cp = vdev_geom_attach_by_guid(vd->vdev_guid);
 476         if (cp != NULL) {
 477                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 478                 buf = kmem_alloc(len, KM_SLEEP);
 479
 480                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 481                 spa_strfree(vd->vdev_path);
 482                 vd->vdev_path = buf;
 483
 484                 ZFS_LOG(1, "Attach by guid [%ju] succeeded, provider %s.",
 485                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 486         } else {
 487                 ZFS_LOG(1, "Search by guid [%ju] failed.",
 488                     (uintmax_t)vd->vdev_guid);
 489         }
 490
 491         return (cp);
 492 }
 493
 494 static struct g_consumer *
 495 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 496 {
 497         struct g_provider *pp;
 498         struct g_consumer *cp;
 499         uint64_t guid;
 500
 501         g_topology_assert();
 502
 503         cp = NULL;
 504         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 505         if (pp != NULL) {
 506                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 507                 cp = vdev_geom_attach(pp);
 508                 if (cp != NULL && check_guid && ISP2(pp->sectorsize) &&
 509                     pp->sectorsize <= VDEV_PAD_SIZE) {
 510                         g_topology_unlock();
 511                         guid = vdev_geom_read_guid(cp);
 512                         g_topology_lock();
 513                         if (guid != vd->vdev_guid) {
 514                                 vdev_geom_detach(cp, 0);
 515                                 cp = NULL;
 516                                 ZFS_LOG(1, "guid mismatch for provider %s: "
 517                                     "%ju != %ju.", vd->vdev_path,
 518                                     (uintmax_t)vd->vdev_guid, (uintmax_t)guid);
 519                         } else {
 520                                 ZFS_LOG(1, "guid match for provider %s.",
 521                                     vd->vdev_path);
 522                         }
 523                 }
 524         }
 525
 526         return (cp);
 527 }
 528
 529 static int
 530 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 531     uint64_t *ashift)
 532 {
 533         struct g_provider *pp;
 534         struct g_consumer *cp;
 535         size_t bufsize;
 536         int error;
 537
 538         /*
 539          * We must have a pathname, and it must be absolute.
 540          */
 541         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 542                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 543                 return (EINVAL);
 544         }
 545
 546         vd->vdev_tsd = NULL;
 547
 548         DROP_GIANT();
 549         g_topology_lock();
 550         error = 0;
 551
 552         /*
 553          * If we're creating or splitting a pool, just find the GEOM provider
 554          * by its name and ignore GUID mismatches.
 555          */
 556         if (vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 557             vd->vdev_spa->spa_splitting_newspa == B_TRUE)
 558                 cp = vdev_geom_open_by_path(vd, 0);
 559         else {
 560                 cp = vdev_geom_open_by_path(vd, 1);
 561                 if (cp == NULL) {
 562                         /*
 563                          * The device at vd->vdev_path doesn't have the
 564                          * expected guid. The disks might have merely
 565                          * moved around so try all other GEOM providers
 566                          * to find one with the right guid.
 567                          */
 568                         cp = vdev_geom_open_by_guid(vd);
 569                 }
 570         }
 571
 572         if (cp == NULL) {
 573                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 574                 error = ENOENT;
 575         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 576             !ISP2(cp->provider->sectorsize)) {
 577                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 578                     vd->vdev_path);
 579                 vdev_geom_detach(cp, 0);
 580                 error = EINVAL;
 581                 cp = NULL;
 582         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 583                 int i;
 584
 585                 for (i = 0; i < 5; i++) {
 586                         error = g_access(cp, 0, 1, 0);
 587                         if (error == 0)
 588                                 break;
 589                         g_topology_unlock();
 590                         tsleep(vd, 0, "vdev", hz / 2);
 591                         g_topology_lock();
 592                 }
 593                 if (error != 0) {
 594                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 595                             vd->vdev_path, error);
 596                         vdev_geom_detach(cp, 0);
 597                         cp = NULL;
 598                 }
 599         }
 600         g_topology_unlock();
 601         PICKUP_GIANT();
 602         if (cp == NULL) {
 603                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 604                 return (error);
 605         }
 606
 607         cp->private = vd;
 608         vd->vdev_tsd = cp;
 609         pp = cp->provider;
 610
 611         /*
 612          * Determine the actual size of the device.
 613          */
 614         *max_psize = *psize = pp->mediasize;
 615
 616         /*
 617          * Determine the device's minimum transfer size.
 618          */
 619         *ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 620
 621         /*
 622          * Clear the nowritecache bit, so that on a vdev_reopen() we will
 623          * try again.
 624          */
 625         vd->vdev_nowritecache = B_FALSE;
 626
 627         if (vd->vdev_physpath != NULL)
 628                 spa_strfree(vd->vdev_physpath);
 629         bufsize = sizeof("/dev/") + strlen(pp->name);
 630         vd->vdev_physpath = kmem_alloc(bufsize, KM_SLEEP);
 631         snprintf(vd->vdev_physpath, bufsize, "/dev/%s", pp->name);
 632
 633         return (0);
 634 }
 635
 636 static void
 637 vdev_geom_close(vdev_t *vd)
 638 {
 639         struct g_consumer *cp;
 640
 641         cp = vd->vdev_tsd;
 642         if (cp == NULL)
 643                 return;
 644         vd->vdev_tsd = NULL;
 645         vd->vdev_delayed_close = B_FALSE;
 646         g_post_event(vdev_geom_detach, cp, M_WAITOK, NULL);
 647 }
 648
 649 static void
 650 vdev_geom_io_intr(struct bio *bp)
 651 {
 652         vdev_t *vd;
 653         zio_t *zio;
 654
 655         zio = bp->bio_caller1;
 656         vd = zio->io_vd;
 657         zio->io_error = bp->bio_error;
 658         if (zio->io_error == 0 && bp->bio_resid != 0)
 659                 zio->io_error = EIO;
 660         if (bp->bio_cmd == BIO_FLUSH && bp->bio_error == ENOTSUP) {
 661                 /*
 662                  * If we get ENOTSUP, we know that no future
 663                  * attempts will ever succeed.  In this case we
 664                  * set a persistent bit so that we don't bother
 665                  * with the ioctl in the future.
 666                  */
 667                 vd->vdev_nowritecache = B_TRUE;
 668         }
 669         if (zio->io_error == EIO && !vd->vdev_remove_wanted) {
 670                 /*
 671                  * If provider's error is set we assume it is being
 672                  * removed.
 673                  */
 674                 if (bp->bio_to->error != 0) {
 675                         /*
 676                          * We post the resource as soon as possible, instead of
 677                          * when the async removal actually happens, because the
 678                          * DE is using this information to discard previous I/O
 679                          * errors.
 680                          */
 681                         /* XXX: zfs_post_remove() can sleep. */
 682                         zfs_post_remove(zio->io_spa, vd);
 683                         vd->vdev_remove_wanted = B_TRUE;
 684                         spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
 685                 } else if (!vd->vdev_delayed_close) {
 686                         vd->vdev_delayed_close = B_TRUE;
 687                 }
 688         }
 689         g_destroy_bio(bp);
 690         zio_interrupt(zio);
 691 }
 692
 693 static int
 694 vdev_geom_io_start(zio_t *zio)
 695 {
 696         vdev_t *vd;
 697         struct g_consumer *cp;
 698         struct bio *bp;
 699         int error;
 700
 701         vd = zio->io_vd;
 702
 703         if (zio->io_type == ZIO_TYPE_IOCTL) {
 704                 /* XXPOLICY */
 705                 if (!vdev_readable(vd)) {
 706                         zio->io_error = ENXIO;
 707                         return (ZIO_PIPELINE_CONTINUE);
 708                 }
 709
 710                 switch (zio->io_cmd) {
 711
 712                 case DKIOCFLUSHWRITECACHE:
 713
 714                         if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 715                                 break;
 716
 717                         if (vd->vdev_nowritecache) {
 718                                 zio->io_error = ENOTSUP;
 719                                 break;
 720                         }
 721
 722                         goto sendreq;
 723                 default:
 724                         zio->io_error = ENOTSUP;
 725                 }
 726
 727                 return (ZIO_PIPELINE_CONTINUE);
 728         }
 729 sendreq:
 730         cp = vd->vdev_tsd;
 731         if (cp == NULL) {
 732                 zio->io_error = ENXIO;
 733                 return (ZIO_PIPELINE_CONTINUE);
 734         }
 735         bp = g_alloc_bio();
 736         bp->bio_caller1 = zio;
 737         switch (zio->io_type) {
 738         case ZIO_TYPE_READ:
 739         case ZIO_TYPE_WRITE:
 740                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 741                 bp->bio_data = zio->io_data;
 742                 bp->bio_offset = zio->io_offset;
 743                 bp->bio_length = zio->io_size;
 744                 break;
 745         case ZIO_TYPE_IOCTL:
 746                 bp->bio_cmd = BIO_FLUSH;
 747                 bp->bio_flags |= BIO_ORDERED;
 748                 bp->bio_data = NULL;
 749                 bp->bio_offset = cp->provider->mediasize;
 750                 bp->bio_length = 0;
 751                 break;
 752         }
 753         bp->bio_done = vdev_geom_io_intr;
 754
 755         g_io_request(bp, cp);
 756
 757         return (ZIO_PIPELINE_STOP);
 758 }
 759
 760 static void
 761 vdev_geom_io_done(zio_t *zio)
 762 {
 763 }
 764
 765 static void
 766 vdev_geom_hold(vdev_t *vd)
 767 {
 768 }
 769
 770 static void
 771 vdev_geom_rele(vdev_t *vd)
 772 {
 773 }
 774
 775 vdev_ops_t vdev_geom_ops = {
 776         vdev_geom_open,
 777         vdev_geom_close,
 778         vdev_default_asize,
 779         vdev_geom_io_start,
 780         vdev_geom_io_done,
 781         NULL,
 782         vdev_geom_hold,
 783         vdev_geom_rele,
 784         VDEV_TYPE_DISK,         /* name of this vdev type */
 785         B_TRUE                  /* leaf vdev */
 786 };