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 SYSCTL_DECL(_vfs_zfs_vdev);
  53 /* Don't send BIO_FLUSH. */
  54 static int vdev_geom_bio_flush_disable = 0;
  55 TUNABLE_INT("vfs.zfs.vdev.bio_flush_disable", &vdev_geom_bio_flush_disable);
  56 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_flush_disable, CTLFLAG_RW,
  57     &vdev_geom_bio_flush_disable, 0, "Disable BIO_FLUSH");
  58 /* Don't send BIO_DELETE. */
  59 static int vdev_geom_bio_delete_disable = 0;
  60 TUNABLE_INT("vfs.zfs.vdev.bio_delete_disable", &vdev_geom_bio_delete_disable);
  61 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_delete_disable, CTLFLAG_RW,
  62     &vdev_geom_bio_delete_disable, 0, "Disable BIO_DELETE");
  63
  64 static void
  65 vdev_geom_orphan(struct g_consumer *cp)
  66 {
  67         vdev_t *vd;
  68
  69         g_topology_assert();
  70
  71         vd = cp->private;
  72
  73         /*
  74          * Orphan callbacks occur from the GEOM event thread.
  75          * Concurrent with this call, new I/O requests may be
  76          * working their way through GEOM about to find out
  77          * (only once executed by the g_down thread) that we've
  78          * been orphaned from our disk provider.  These I/Os
  79          * must be retired before we can detach our consumer.
  80          * This is most easily achieved by acquiring the
  81          * SPA ZIO configuration lock as a writer, but doing
  82          * so with the GEOM topology lock held would cause
  83          * a lock order reversal.  Instead, rely on the SPA's
  84          * async removal support to invoke a close on this
  85          * vdev once it is safe to do so.
  86          */
  87         zfs_post_remove(vd->vdev_spa, vd);
  88         vd->vdev_remove_wanted = B_TRUE;
  89         spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
  90 }
  91
  92 static struct g_consumer *
  93 vdev_geom_attach(struct g_provider *pp)
  94 {
  95         struct g_geom *gp;
  96         struct g_consumer *cp;
  97
  98         g_topology_assert();
  99
 100         ZFS_LOG(1, "Attaching to %s.", pp->name);
 101         /* Do we have geom already? No? Create one. */
 102         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 103                 if (gp->flags & G_GEOM_WITHER)
 104                         continue;
 105                 if (strcmp(gp->name, "zfs::vdev") != 0)
 106                         continue;
 107                 break;
 108         }
 109         if (gp == NULL) {
 110                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 111                 gp->orphan = vdev_geom_orphan;
 112                 cp = g_new_consumer(gp);
 113                 if (g_attach(cp, pp) != 0) {
 114                         g_wither_geom(gp, ENXIO);
 115                         return (NULL);
 116                 }
 117                 if (g_access(cp, 1, 0, 1) != 0) {
 118                         g_wither_geom(gp, ENXIO);
 119                         return (NULL);
 120                 }
 121                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 122         } else {
 123                 /* Check if we are already connected to this provider. */
 124                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 125                         if (cp->provider == pp) {
 126                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 127                                 break;
 128                         }
 129                 }
 130                 if (cp == NULL) {
 131                         cp = g_new_consumer(gp);
 132                         if (g_attach(cp, pp) != 0) {
 133                                 g_destroy_consumer(cp);
 134                                 return (NULL);
 135                         }
 136                         if (g_access(cp, 1, 0, 1) != 0) {
 137                                 g_detach(cp);
 138                                 g_destroy_consumer(cp);
 139                                 return (NULL);
 140                         }
 141                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 142                 } else {
 143                         if (g_access(cp, 1, 0, 1) != 0)
 144                                 return (NULL);
 145                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 146                 }
 147         }
 148         return (cp);
 149 }
 150
 151 static void
 152 vdev_geom_detach(void *arg, int flag __unused)
 153 {
 154         struct g_geom *gp;
 155         struct g_consumer *cp;
 156
 157         g_topology_assert();
 158         cp = arg;
 159         gp = cp->geom;
 160
 161         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 162         g_access(cp, -1, 0, -1);
 163         /* Destroy consumer on last close. */
 164         if (cp->acr == 0 && cp->ace == 0) {
 165                 ZFS_LOG(1, "Destroyed consumer to %s.", cp->provider->name);
 166                 if (cp->acw > 0)
 167                         g_access(cp, 0, -cp->acw, 0);
 168                 g_detach(cp);
 169                 g_destroy_consumer(cp);
 170         }
 171         /* Destroy geom if there are no consumers left. */
 172         if (LIST_EMPTY(&gp->consumer)) {
 173                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 174                 g_wither_geom(gp, ENXIO);
 175         }
 176 }
 177
 178 static uint64_t
 179 nvlist_get_guid(nvlist_t *list)
 180 {
 181         uint64_t value;
 182
 183         value = 0;
 184         nvlist_lookup_uint64(list, ZPOOL_CONFIG_GUID, &value);
 185         return (value);
 186 }
 187
 188 static int
 189 vdev_geom_io(struct g_consumer *cp, int cmd, void *data, off_t offset, off_t size)
 190 {
 191         struct bio *bp;
 192         u_char *p;
 193         off_t off, maxio;
 194         int error;
 195
 196         ASSERT((offset % cp->provider->sectorsize) == 0);
 197         ASSERT((size % cp->provider->sectorsize) == 0);
 198
 199         bp = g_alloc_bio();
 200         off = offset;
 201         offset += size;
 202         p = data;
 203         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 204         error = 0;
 205
 206         for (; off < offset; off += maxio, p += maxio, size -= maxio) {
 207                 bzero(bp, sizeof(*bp));
 208                 bp->bio_cmd = cmd;
 209                 bp->bio_done = NULL;
 210                 bp->bio_offset = off;
 211                 bp->bio_length = MIN(size, maxio);
 212                 bp->bio_data = p;
 213                 g_io_request(bp, cp);
 214                 error = biowait(bp, "vdev_geom_io");
 215                 if (error != 0)
 216                         break;
 217         }
 218
 219         g_destroy_bio(bp);
 220         return (error);
 221 }
 222
 223 static void
 224 vdev_geom_taste_orphan(struct g_consumer *cp)
 225 {
 226
 227         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
 228             cp->provider->name));
 229 }
 230
 231 static int
 232 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 233 {
 234         struct g_provider *pp;
 235         vdev_label_t *label;
 236         char *p, *buf;
 237         size_t buflen;
 238         uint64_t psize;
 239         off_t offset, size;
 240         uint64_t guid, state, txg;
 241         int error, l, len;
 242
 243         g_topology_assert_not();
 244
 245         pp = cp->provider;
 246         ZFS_LOG(1, "Reading config from %s...", pp->name);
 247
 248         psize = pp->mediasize;
 249         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 250
 251         size = sizeof(*label) + pp->sectorsize -
 252             ((sizeof(*label) - 1) % pp->sectorsize) - 1;
 253
 254         guid = 0;
 255         label = kmem_alloc(size, KM_SLEEP);
 256         buflen = sizeof(label->vl_vdev_phys.vp_nvlist);
 257
 258         *config = NULL;
 259         for (l = 0; l < VDEV_LABELS; l++) {
 260
 261                 offset = vdev_label_offset(psize, l, 0);
 262                 if ((offset % pp->sectorsize) != 0)
 263                         continue;
 264
 265                 if (vdev_geom_io(cp, BIO_READ, label, offset, size) != 0)
 266                         continue;
 267                 buf = label->vl_vdev_phys.vp_nvlist;
 268
 269                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 270                         continue;
 271
 272                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 273                     &state) != 0 || state >= POOL_STATE_DESTROYED) {
 274                         nvlist_free(*config);
 275                         *config = NULL;
 276                         continue;
 277                 }
 278
 279                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 280                     &txg) != 0 || txg == 0) {
 281                         nvlist_free(*config);
 282                         *config = NULL;
 283                         continue;
 284                 }
 285
 286                 break;
 287         }
 288
 289         kmem_free(label, size);
 290         return (*config == NULL ? ENOENT : 0);
 291 }
 292
 293 static int
 294 vdev_geom_check_config(nvlist_t *config, const char *name, uint64_t *best_txg)
 295 {
 296         uint64_t vdev_guid;
 297         uint64_t txg;
 298         char *pname;
 299
 300         if (nvlist_lookup_string(config, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 301             strcmp(pname, name) != 0)
 302                 return (ENOENT);
 303
 304         ZFS_LOG(1, "found pool: %s", pname);
 305
 306         txg = 0;
 307         nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG, &txg);
 308         if (txg <= *best_txg)
 309                 return (ENOENT);
 310         *best_txg = txg;
 311         ZFS_LOG(1, "txg: %ju", (uintmax_t)*best_txg);
 312
 313         return (0);
 314 }
 315
 316 static int
 317 vdev_geom_attach_taster(struct g_consumer *cp, struct g_provider *pp)
 318 {
 319         int error;
 320
 321         if (pp->flags & G_PF_WITHER)
 322                 return (EINVAL);
 323         if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize))
 324                 return (EINVAL);
 325         g_attach(cp, pp);
 326         error = g_access(cp, 1, 0, 0);
 327         if (error != 0)
 328                 g_detach(cp);
 329         return (error);
 330 }
 331
 332 static void
 333 vdev_geom_dettach_taster(struct g_consumer *cp)
 334 {
 335         g_access(cp, -1, 0, 0);
 336         g_detach(cp);
 337 }
 338
 339 int
 340 vdev_geom_read_pool_label(const char *name, nvlist_t **config)
 341 {
 342         struct g_class *mp;
 343         struct g_geom *gp, *zgp;
 344         struct g_provider *pp;
 345         struct g_consumer *zcp;
 346         nvlist_t *vdev_cfg;
 347         uint64_t best_txg;
 348         int error;
 349
 350         DROP_GIANT();
 351         g_topology_lock();
 352
 353         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 354         /* This orphan function should be never called. */
 355         zgp->orphan = vdev_geom_taste_orphan;
 356         zcp = g_new_consumer(zgp);
 357
 358         best_txg = 0;
 359         *config = NULL;
 360         LIST_FOREACH(mp, &g_classes, class) {
 361                 if (mp == &zfs_vdev_class)
 362                         continue;
 363                 LIST_FOREACH(gp, &mp->geom, geom) {
 364                         if (gp->flags & G_GEOM_WITHER)
 365                                 continue;
 366                         LIST_FOREACH(pp, &gp->provider, provider) {
 367                                 if (pp->flags & G_PF_WITHER)
 368                                         continue;
 369                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 370                                         continue;
 371                                 g_topology_unlock();
 372                                 error = vdev_geom_read_config(zcp, &vdev_cfg);
 373                                 g_topology_lock();
 374                                 vdev_geom_dettach_taster(zcp);
 375                                 if (error)
 376                                         continue;
 377                                 ZFS_LOG(1, "successfully read vdev config");
 378
 379                                 error = vdev_geom_check_config(vdev_cfg, name,
 380                                     &best_txg);
 381                                 if (error != 0) {
 382                                         nvlist_free(vdev_cfg);
 383                                         continue;
 384                                 }
 385                                 nvlist_free(*config);
 386                                 *config = vdev_cfg;
 387                         }
 388                 }
 389         }
 390
 391         g_destroy_consumer(zcp);
 392         g_destroy_geom(zgp);
 393         g_topology_unlock();
 394         PICKUP_GIANT();
 395         return (*config == NULL ? ENOENT : 0);
 396 }
 397
 398 static uint64_t
 399 vdev_geom_read_guid(struct g_consumer *cp)
 400 {
 401         nvlist_t *config;
 402         uint64_t guid;
 403
 404         g_topology_assert_not();
 405
 406         guid = 0;
 407         if (vdev_geom_read_config(cp, &config) == 0) {
 408                 guid = nvlist_get_guid(config);
 409                 nvlist_free(config);
 410         }
 411         return (guid);
 412 }
 413
 414 static struct g_consumer *
 415 vdev_geom_attach_by_guid(uint64_t guid)
 416 {
 417         struct g_class *mp;
 418         struct g_geom *gp, *zgp;
 419         struct g_provider *pp;
 420         struct g_consumer *cp, *zcp;
 421         uint64_t pguid;
 422
 423         g_topology_assert();
 424
 425         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 426         /* This orphan function should be never called. */
 427         zgp->orphan = vdev_geom_taste_orphan;
 428         zcp = g_new_consumer(zgp);
 429
 430         cp = NULL;
 431         LIST_FOREACH(mp, &g_classes, class) {
 432                 if (mp == &zfs_vdev_class)
 433                         continue;
 434                 LIST_FOREACH(gp, &mp->geom, geom) {
 435                         if (gp->flags & G_GEOM_WITHER)
 436                                 continue;
 437                         LIST_FOREACH(pp, &gp->provider, provider) {
 438                                 if (vdev_geom_attach_taster(zcp, pp) != 0)
 439                                         continue;
 440                                 g_topology_unlock();
 441                                 pguid = vdev_geom_read_guid(zcp);
 442                                 g_topology_lock();
 443                                 vdev_geom_dettach_taster(zcp);
 444                                 if (pguid != guid)
 445                                         continue;
 446                                 cp = vdev_geom_attach(pp);
 447                                 if (cp == NULL) {
 448                                         printf("ZFS WARNING: Unable to attach to %s.\n",
 449                                             pp->name);
 450                                         continue;
 451                                 }
 452                                 break;
 453                         }
 454                         if (cp != NULL)
 455                                 break;
 456                 }
 457                 if (cp != NULL)
 458                         break;
 459         }
 460 end:
 461         g_destroy_consumer(zcp);
 462         g_destroy_geom(zgp);
 463         return (cp);
 464 }
 465
 466 static struct g_consumer *
 467 vdev_geom_open_by_guid(vdev_t *vd)
 468 {
 469         struct g_consumer *cp;
 470         char *buf;
 471         size_t len;
 472
 473         g_topology_assert();
 474
 475         ZFS_LOG(1, "Searching by guid [%ju].", (uintmax_t)vd->vdev_guid);
 476         cp = vdev_geom_attach_by_guid(vd->vdev_guid);
 477         if (cp != NULL) {
 478                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 479                 buf = kmem_alloc(len, KM_SLEEP);
 480
 481                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 482                 spa_strfree(vd->vdev_path);
 483                 vd->vdev_path = buf;
 484
 485                 ZFS_LOG(1, "Attach by guid [%ju] succeeded, provider %s.",
 486                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 487         } else {
 488                 ZFS_LOG(1, "Search by guid [%ju] failed.",
 489                     (uintmax_t)vd->vdev_guid);
 490         }
 491
 492         return (cp);
 493 }
 494
 495 static struct g_consumer *
 496 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 497 {
 498         struct g_provider *pp;
 499         struct g_consumer *cp;
 500         uint64_t guid;
 501
 502         g_topology_assert();
 503
 504         cp = NULL;
 505         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 506         if (pp != NULL) {
 507                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 508                 cp = vdev_geom_attach(pp);
 509                 if (cp != NULL && check_guid && ISP2(pp->sectorsize) &&
 510                     pp->sectorsize <= VDEV_PAD_SIZE) {
 511                         g_topology_unlock();
 512                         guid = vdev_geom_read_guid(cp);
 513                         g_topology_lock();
 514                         if (guid != vd->vdev_guid) {
 515                                 vdev_geom_detach(cp, 0);
 516                                 cp = NULL;
 517                                 ZFS_LOG(1, "guid mismatch for provider %s: "
 518                                     "%ju != %ju.", vd->vdev_path,
 519                                     (uintmax_t)vd->vdev_guid, (uintmax_t)guid);
 520                         } else {
 521                                 ZFS_LOG(1, "guid match for provider %s.",
 522                                     vd->vdev_path);
 523                         }
 524                 }
 525         }
 526
 527         return (cp);
 528 }
 529
 530 static int
 531 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 532     uint64_t *ashift)
 533 {
 534         struct g_provider *pp;
 535         struct g_consumer *cp;
 536         size_t bufsize;
 537         int error;
 538
 539         /*
 540          * We must have a pathname, and it must be absolute.
 541          */
 542         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 543                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 544                 return (EINVAL);
 545         }
 546
 547         vd->vdev_tsd = NULL;
 548
 549         DROP_GIANT();
 550         g_topology_lock();
 551         error = 0;
 552
 553         /*
 554          * If we're creating or splitting a pool, just find the GEOM provider
 555          * by its name and ignore GUID mismatches.
 556          */
 557         if (vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 558             vd->vdev_spa->spa_splitting_newspa == B_TRUE)
 559                 cp = vdev_geom_open_by_path(vd, 0);
 560         else {
 561                 cp = vdev_geom_open_by_path(vd, 1);
 562                 if (cp == NULL) {
 563                         /*
 564                          * The device at vd->vdev_path doesn't have the
 565                          * expected guid. The disks might have merely
 566                          * moved around so try all other GEOM providers
 567                          * to find one with the right guid.
 568                          */
 569                         cp = vdev_geom_open_by_guid(vd);
 570                 }
 571         }
 572
 573         if (cp == NULL) {
 574                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 575                 error = ENOENT;
 576         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 577             !ISP2(cp->provider->sectorsize)) {
 578                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 579                     vd->vdev_path);
 580                 vdev_geom_detach(cp, 0);
 581                 error = EINVAL;
 582                 cp = NULL;
 583         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 584                 int i;
 585
 586                 for (i = 0; i < 5; i++) {
 587                         error = g_access(cp, 0, 1, 0);
 588                         if (error == 0)
 589                                 break;
 590                         g_topology_unlock();
 591                         tsleep(vd, 0, "vdev", hz / 2);
 592                         g_topology_lock();
 593                 }
 594                 if (error != 0) {
 595                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 596                             vd->vdev_path, error);
 597                         vdev_geom_detach(cp, 0);
 598                         cp = NULL;
 599                 }
 600         }
 601         g_topology_unlock();
 602         PICKUP_GIANT();
 603         if (cp == NULL) {
 604                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 605                 return (error);
 606         }
 607
 608         cp->private = vd;
 609         vd->vdev_tsd = cp;
 610         pp = cp->provider;
 611
 612         /*
 613          * Determine the actual size of the device.
 614          */
 615         *max_psize = *psize = pp->mediasize;
 616
 617         /*
 618          * Determine the device's minimum transfer size.
 619          */
 620         *ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 621
 622         /*
 623          * Clear the nowritecache settings, so that on a vdev_reopen()
 624          * we will try again.
 625          */
 626         vd->vdev_nowritecache = B_FALSE;
 627
 628         if (vd->vdev_physpath != NULL)
 629                 spa_strfree(vd->vdev_physpath);
 630         bufsize = sizeof("/dev/") + strlen(pp->name);
 631         vd->vdev_physpath = kmem_alloc(bufsize, KM_SLEEP);
 632         snprintf(vd->vdev_physpath, bufsize, "/dev/%s", pp->name);
 633
 634         return (0);
 635 }
 636
 637 static void
 638 vdev_geom_close(vdev_t *vd)
 639 {
 640         struct g_consumer *cp;
 641
 642         cp = vd->vdev_tsd;
 643         if (cp == NULL)
 644                 return;
 645         vd->vdev_tsd = NULL;
 646         vd->vdev_delayed_close = B_FALSE;
 647         g_post_event(vdev_geom_detach, cp, M_WAITOK, NULL);
 648 }
 649
 650 static void
 651 vdev_geom_io_intr(struct bio *bp)
 652 {
 653         vdev_t *vd;
 654         zio_t *zio;
 655
 656         zio = bp->bio_caller1;
 657         vd = zio->io_vd;
 658         zio->io_error = bp->bio_error;
 659         if (zio->io_error == 0 && bp->bio_resid != 0)
 660                 zio->io_error = EIO;
 661         if (bp->bio_cmd == BIO_FLUSH && bp->bio_error == ENOTSUP) {
 662                 /*
 663                  * If we get ENOTSUP, we know that no future
 664                  * attempts will ever succeed.  In this case we
 665                  * set a persistent bit so that we don't bother
 666                  * with the ioctl in the future.
 667                  */
 668                 vd->vdev_nowritecache = B_TRUE;
 669         }
 670         if (bp->bio_cmd == BIO_DELETE && bp->bio_error == ENOTSUP) {
 671                 /*
 672                  * If we get ENOTSUP, we know that no future
 673                  * attempts will ever succeed.  In this case we
 674                  * set a persistent bit so that we don't bother
 675                  * with the ioctl in the future.
 676                  */
 677                 vd->vdev_notrim = B_TRUE;
 678         }
 679         if (zio->io_error == EIO && !vd->vdev_remove_wanted) {
 680                 /*
 681                  * If provider's error is set we assume it is being
 682                  * removed.
 683                  */
 684                 if (bp->bio_to->error != 0) {
 685                         /*
 686                          * We post the resource as soon as possible, instead of
 687                          * when the async removal actually happens, because the
 688                          * DE is using this information to discard previous I/O
 689                          * errors.
 690                          */
 691                         /* XXX: zfs_post_remove() can sleep. */
 692                         zfs_post_remove(zio->io_spa, vd);
 693                         vd->vdev_remove_wanted = B_TRUE;
 694                         spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
 695                 } else if (!vd->vdev_delayed_close) {
 696                         vd->vdev_delayed_close = B_TRUE;
 697                 }
 698         }
 699         g_destroy_bio(bp);
 700         zio_interrupt(zio);
 701 }
 702
 703 static int
 704 vdev_geom_io_start(zio_t *zio)
 705 {
 706         vdev_t *vd;
 707         struct g_consumer *cp;
 708         struct bio *bp;
 709         int error;
 710
 711         vd = zio->io_vd;
 712
 713         if (zio->io_type == ZIO_TYPE_IOCTL) {
 714                 /* XXPOLICY */
 715                 if (!vdev_readable(vd)) {
 716                         zio->io_error = ENXIO;
 717                         return (ZIO_PIPELINE_CONTINUE);
 718                 }
 719
 720                 switch (zio->io_cmd) {
 721                 case DKIOCFLUSHWRITECACHE:
 722                         if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
 723                                 break;
 724                         if (vd->vdev_nowritecache) {
 725                                 zio->io_error = ENOTSUP;
 726                                 break;
 727                         }
 728                         goto sendreq;
 729                 case DKIOCTRIM:
 730                         if (vdev_geom_bio_delete_disable)
 731                                 break;
 732                         if (vd->vdev_notrim) {
 733                                 zio->io_error = ENOTSUP;
 734                                 break;
 735                         }
 736                         goto sendreq;
 737                 default:
 738                         zio->io_error = ENOTSUP;
 739                 }
 740
 741                 return (ZIO_PIPELINE_CONTINUE);
 742         }
 743 sendreq:
 744         cp = vd->vdev_tsd;
 745         if (cp == NULL) {
 746                 zio->io_error = ENXIO;
 747                 return (ZIO_PIPELINE_CONTINUE);
 748         }
 749         bp = g_alloc_bio();
 750         bp->bio_caller1 = zio;
 751         switch (zio->io_type) {
 752         case ZIO_TYPE_READ:
 753         case ZIO_TYPE_WRITE:
 754                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 755                 bp->bio_data = zio->io_data;
 756                 bp->bio_offset = zio->io_offset;
 757                 bp->bio_length = zio->io_size;
 758                 break;
 759         case ZIO_TYPE_IOCTL:
 760                 switch (zio->io_cmd) {
 761                 case DKIOCFLUSHWRITECACHE:
 762                         bp->bio_cmd = BIO_FLUSH;
 763                         bp->bio_flags |= BIO_ORDERED;
 764                         bp->bio_data = NULL;
 765                         bp->bio_offset = cp->provider->mediasize;
 766                         bp->bio_length = 0;
 767                         break;
 768                 case DKIOCTRIM:
 769                         bp->bio_cmd = BIO_DELETE;
 770                         bp->bio_data = NULL;
 771                         bp->bio_offset = zio->io_offset;
 772                         bp->bio_length = zio->io_size;
 773                         break;
 774                 }
 775                 break;
 776         }
 777         bp->bio_done = vdev_geom_io_intr;
 778
 779         g_io_request(bp, cp);
 780
 781         return (ZIO_PIPELINE_STOP);
 782 }
 783
 784 static void
 785 vdev_geom_io_done(zio_t *zio)
 786 {
 787 }
 788
 789 static void
 790 vdev_geom_hold(vdev_t *vd)
 791 {
 792 }
 793
 794 static void
 795 vdev_geom_rele(vdev_t *vd)
 796 {
 797 }
 798
 799 vdev_ops_t vdev_geom_ops = {
 800         vdev_geom_open,
 801         vdev_geom_close,
 802         vdev_default_asize,
 803         vdev_geom_io_start,
 804         vdev_geom_io_done,
 805         NULL,
 806         vdev_geom_hold,
 807         vdev_geom_rele,
 808         VDEV_TYPE_DISK,         /* name of this vdev type */
 809         B_TRUE                  /* leaf vdev */
 810 };