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 struct consumer_vdev_elem {
  53         SLIST_ENTRY(consumer_vdev_elem) elems;
  54         vdev_t                          *vd;
  55 };
  56
  57 SLIST_HEAD(consumer_priv_t, consumer_vdev_elem);
  58 _Static_assert(sizeof(((struct g_consumer*)NULL)->private)
  59     == sizeof(struct consumer_priv_t*),
  60     "consumer_priv_t* can't be stored in g_consumer.private");
  61
  62 DECLARE_GEOM_CLASS(zfs_vdev_class, zfs_vdev);
  63
  64 SYSCTL_DECL(_vfs_zfs_vdev);
  65 /* Don't send BIO_FLUSH. */
  66 static int vdev_geom_bio_flush_disable = 0;
  67 TUNABLE_INT("vfs.zfs.vdev.bio_flush_disable", &vdev_geom_bio_flush_disable);
  68 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_flush_disable, CTLFLAG_RW,
  69     &vdev_geom_bio_flush_disable, 0, "Disable BIO_FLUSH");
  70 /* Don't send BIO_DELETE. */
  71 static int vdev_geom_bio_delete_disable = 0;
  72 TUNABLE_INT("vfs.zfs.vdev.bio_delete_disable", &vdev_geom_bio_delete_disable);
  73 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_delete_disable, CTLFLAG_RW,
  74     &vdev_geom_bio_delete_disable, 0, "Disable BIO_DELETE");
  75
  76 /* Declare local functions */
  77 static void vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read);
  78
  79 /*
  80  * Thread local storage used to indicate when a thread is probing geoms
  81  * for their guids.  If NULL, this thread is not tasting geoms.  If non NULL,
  82  * it is looking for a replacement for the vdev_t* that is its value.
  83  */
  84 uint_t zfs_geom_probe_vdev_key;
  85
  86 static void
  87 vdev_geom_set_rotation_rate(vdev_t *vd, struct g_consumer *cp)
  88 {
  89         int error;
  90         uint16_t rate;
  91
  92         error = g_getattr("GEOM::rotation_rate", cp, &rate);
  93         if (error == 0)
  94                 vd->vdev_rotation_rate = rate;
  95         else
  96                 vd->vdev_rotation_rate = VDEV_RATE_UNKNOWN;
  97 }
  98
  99 static void
 100 vdev_geom_set_physpath(vdev_t *vd, struct g_consumer *cp,
 101                        boolean_t do_null_update)
 102 {
 103         boolean_t needs_update = B_FALSE;
 104         char *physpath;
 105         int error, physpath_len;
 106
 107         physpath_len = MAXPATHLEN;
 108         physpath = g_malloc(physpath_len, M_WAITOK|M_ZERO);
 109         error = g_io_getattr("GEOM::physpath", cp, &physpath_len, physpath);
 110         if (error == 0) {
 111                 char *old_physpath;
 112
 113                 /* g_topology lock ensures that vdev has not been closed */
 114                 g_topology_assert();
 115                 old_physpath = vd->vdev_physpath;
 116                 vd->vdev_physpath = spa_strdup(physpath);
 117
 118                 if (old_physpath != NULL) {
 119                         needs_update = (strcmp(old_physpath,
 120                                                 vd->vdev_physpath) != 0);
 121                         spa_strfree(old_physpath);
 122                 } else
 123                         needs_update = do_null_update;
 124         }
 125         g_free(physpath);
 126
 127         /*
 128          * If the physical path changed, update the config.
 129          * Only request an update for previously unset physpaths if
 130          * requested by the caller.
 131          */
 132         if (needs_update)
 133                 spa_async_request(vd->vdev_spa, SPA_ASYNC_CONFIG_UPDATE);
 134
 135 }
 136
 137 static void
 138 vdev_geom_attrchanged(struct g_consumer *cp, const char *attr)
 139 {
 140         char *old_physpath;
 141         struct consumer_priv_t *priv;
 142         struct consumer_vdev_elem *elem;
 143         int error;
 144
 145         priv = (struct consumer_priv_t*)&cp->private;
 146         if (SLIST_EMPTY(priv))
 147                 return;
 148
 149         SLIST_FOREACH(elem, priv, elems) {
 150                 vdev_t *vd = elem->vd;
 151                 if (strcmp(attr, "GEOM::rotation_rate") == 0) {
 152                         vdev_geom_set_rotation_rate(vd, cp);
 153                         return;
 154                 }
 155                 if (strcmp(attr, "GEOM::physpath") == 0) {
 156                         vdev_geom_set_physpath(vd, cp, /*null_update*/B_TRUE);
 157                         return;
 158                 }
 159         }
 160 }
 161
 162 static void
 163 vdev_geom_orphan(struct g_consumer *cp)
 164 {
 165         struct consumer_priv_t *priv;
 166         struct consumer_vdev_elem *elem;
 167
 168         g_topology_assert();
 169
 170         priv = (struct consumer_priv_t*)&cp->private;
 171         if (SLIST_EMPTY(priv))
 172                 /* Vdev close in progress.  Ignore the event. */
 173                 return;
 174
 175         /*
 176          * Orphan callbacks occur from the GEOM event thread.
 177          * Concurrent with this call, new I/O requests may be
 178          * working their way through GEOM about to find out
 179          * (only once executed by the g_down thread) that we've
 180          * been orphaned from our disk provider.  These I/Os
 181          * must be retired before we can detach our consumer.
 182          * This is most easily achieved by acquiring the
 183          * SPA ZIO configuration lock as a writer, but doing
 184          * so with the GEOM topology lock held would cause
 185          * a lock order reversal.  Instead, rely on the SPA's
 186          * async removal support to invoke a close on this
 187          * vdev once it is safe to do so.
 188          */
 189         SLIST_FOREACH(elem, priv, elems) {
 190                 vdev_t *vd = elem->vd;
 191
 192                 vd->vdev_remove_wanted = B_TRUE;
 193                 spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
 194         }
 195 }
 196
 197 static struct g_consumer *
 198 vdev_geom_attach(struct g_provider *pp, vdev_t *vd, boolean_t sanity)
 199 {
 200         struct g_geom *gp;
 201         struct g_consumer *cp;
 202         int error;
 203
 204         g_topology_assert();
 205
 206         ZFS_LOG(1, "Attaching to %s.", pp->name);
 207
 208         if (sanity) {
 209                 if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize)) {
 210                         ZFS_LOG(1, "Failing attach of %s. "
 211                                    "Incompatible sectorsize %d\n",
 212                             pp->name, pp->sectorsize);
 213                         return (NULL);
 214                 } else if (pp->mediasize < SPA_MINDEVSIZE) {
 215                         ZFS_LOG(1, "Failing attach of %s. "
 216                                    "Incompatible mediasize %ju\n",
 217                             pp->name, pp->mediasize);
 218                         return (NULL);
 219                 }
 220         }
 221
 222         /* Do we have geom already? No? Create one. */
 223         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 224                 if (gp->flags & G_GEOM_WITHER)
 225                         continue;
 226                 if (strcmp(gp->name, "zfs::vdev") != 0)
 227                         continue;
 228                 break;
 229         }
 230         if (gp == NULL) {
 231                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 232                 gp->orphan = vdev_geom_orphan;
 233                 gp->attrchanged = vdev_geom_attrchanged;
 234                 cp = g_new_consumer(gp);
 235                 error = g_attach(cp, pp);
 236                 if (error != 0) {
 237                         ZFS_LOG(1, "%s(%d): g_attach failed: %d\n", __func__,
 238                             __LINE__, error);
 239                         vdev_geom_detach(cp, B_FALSE);
 240                         return (NULL);
 241                 }
 242                 error = g_access(cp, 1, 0, 1);
 243                 if (error != 0) {
 244                         ZFS_LOG(1, "%s(%d): g_access failed: %d", __func__,
 245                                __LINE__, error);
 246                         vdev_geom_detach(cp, B_FALSE);
 247                         return (NULL);
 248                 }
 249                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 250         } else {
 251                 /* Check if we are already connected to this provider. */
 252                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 253                         if (cp->provider == pp) {
 254                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 255                                 break;
 256                         }
 257                 }
 258                 if (cp == NULL) {
 259                         cp = g_new_consumer(gp);
 260                         error = g_attach(cp, pp);
 261                         if (error != 0) {
 262                                 ZFS_LOG(1, "%s(%d): g_attach failed: %d\n",
 263                                     __func__, __LINE__, error);
 264                                 vdev_geom_detach(cp, B_FALSE);
 265                                 return (NULL);
 266                         }
 267                         error = g_access(cp, 1, 0, 1);
 268                         if (error != 0) {
 269                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 270                                     __func__, __LINE__, error);
 271                                 vdev_geom_detach(cp, B_FALSE);
 272                                 return (NULL);
 273                         }
 274                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 275                 } else {
 276                         error = g_access(cp, 1, 0, 1);
 277                         if (error != 0) {
 278                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 279                                     __func__, __LINE__, error);
 280                                 return (NULL);
 281                         }
 282                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 283                 }
 284         }
 285
 286         if (vd != NULL)
 287                 vd->vdev_tsd = cp;
 288
 289         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
 290         return (cp);
 291 }
 292
 293 static void
 294 vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read)
 295 {
 296         struct g_geom *gp;
 297
 298         g_topology_assert();
 299
 300         ZFS_LOG(1, "Detaching from %s.",
 301             cp->provider && cp->provider->name ? cp->provider->name : "NULL");
 302
 303         gp = cp->geom;
 304         if (open_for_read)
 305                 g_access(cp, -1, 0, -1);
 306         /* Destroy consumer on last close. */
 307         if (cp->acr == 0 && cp->ace == 0) {
 308                 if (cp->acw > 0)
 309                         g_access(cp, 0, -cp->acw, 0);
 310                 if (cp->provider != NULL) {
 311                         ZFS_LOG(1, "Destroying consumer for %s.",
 312                             cp->provider->name ? cp->provider->name : "NULL");
 313                         g_detach(cp);
 314                 }
 315                 g_destroy_consumer(cp);
 316         }
 317         /* Destroy geom if there are no consumers left. */
 318         if (LIST_EMPTY(&gp->consumer)) {
 319                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 320                 g_wither_geom(gp, ENXIO);
 321         }
 322 }
 323
 324 static void
 325 vdev_geom_close_locked(vdev_t *vd)
 326 {
 327         struct g_consumer *cp;
 328         struct consumer_priv_t *priv;
 329         struct consumer_vdev_elem *elem, *elem_temp;
 330
 331         g_topology_assert();
 332
 333         cp = vd->vdev_tsd;
 334         vd->vdev_delayed_close = B_FALSE;
 335         if (cp == NULL)
 336                 return;
 337
 338         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 339         KASSERT(cp->private != NULL, ("%s: cp->private is NULL", __func__));
 340         priv = (struct consumer_priv_t*)&cp->private;
 341         vd->vdev_tsd = NULL;
 342         SLIST_FOREACH_SAFE(elem, priv, elems, elem_temp) {
 343                 if (elem->vd == vd) {
 344                         SLIST_REMOVE(priv, elem, consumer_vdev_elem, elems);
 345                         g_free(elem);
 346                 }
 347         }
 348
 349         vdev_geom_detach(cp, B_TRUE);
 350 }
 351
 352 /*
 353  * Issue one or more bios to the vdev in parallel
 354  * cmds, datas, offsets, errors, and sizes are arrays of length ncmds.  Each IO
 355  * operation is described by parallel entries from each array.  There may be
 356  * more bios actually issued than entries in the array
 357  */
 358 static void
 359 vdev_geom_io(struct g_consumer *cp, int *cmds, void **datas, off_t *offsets,
 360     off_t *sizes, int *errors, int ncmds)
 361 {
 362         struct bio **bios;
 363         u_char *p;
 364         off_t off, maxio, s, end;
 365         int i, n_bios, j;
 366         size_t bios_size;
 367
 368         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 369         n_bios = 0;
 370
 371         /* How many bios are required for all commands ? */
 372         for (i = 0; i < ncmds; i++)
 373                 n_bios += (sizes[i] + maxio - 1) / maxio;
 374
 375         /* Allocate memory for the bios */
 376         bios_size = n_bios * sizeof(struct bio*);
 377         bios = kmem_zalloc(bios_size, KM_SLEEP);
 378
 379         /* Prepare and issue all of the bios */
 380         for (i = j = 0; i < ncmds; i++) {
 381                 off = offsets[i];
 382                 p = datas[i];
 383                 s = sizes[i];
 384                 end = off + s;
 385                 ASSERT((off % cp->provider->sectorsize) == 0);
 386                 ASSERT((s % cp->provider->sectorsize) == 0);
 387
 388                 for (; off < end; off += maxio, p += maxio, s -= maxio, j++) {
 389                         bios[j] = g_alloc_bio();
 390                         bios[j]->bio_cmd = cmds[i];
 391                         bios[j]->bio_done = NULL;
 392                         bios[j]->bio_offset = off;
 393                         bios[j]->bio_length = MIN(s, maxio);
 394                         bios[j]->bio_data = p;
 395                         g_io_request(bios[j], cp);
 396                 }
 397         }
 398         ASSERT(j == n_bios);
 399
 400         /* Wait for all of the bios to complete, and clean them up */
 401         for (i = j = 0; i < ncmds; i++) {
 402                 off = offsets[i];
 403                 s = sizes[i];
 404                 end = off + s;
 405
 406                 for (; off < end; off += maxio, s -= maxio, j++) {
 407                         errors[i] = biowait(bios[j], "vdev_geom_io") || errors[i];
 408                         g_destroy_bio(bios[j]);
 409                 }
 410         }
 411         kmem_free(bios, bios_size);
 412 }
 413
 414 /*
 415  * Read the vdev config from a device.  Return the number of valid labels that
 416  * were found.  The vdev config will be returned in config if and only if at
 417  * least one valid label was found.
 418  */
 419 static int
 420 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **configp)
 421 {
 422         struct g_provider *pp;
 423         nvlist_t *config;
 424         vdev_phys_t *vdev_lists[VDEV_LABELS];
 425         char *buf;
 426         size_t buflen;
 427         uint64_t psize, state, txg;
 428         off_t offsets[VDEV_LABELS];
 429         off_t size;
 430         off_t sizes[VDEV_LABELS];
 431         int cmds[VDEV_LABELS];
 432         int errors[VDEV_LABELS];
 433         int l, nlabels;
 434
 435         g_topology_assert_not();
 436
 437         pp = cp->provider;
 438         ZFS_LOG(1, "Reading config from %s...", pp->name);
 439
 440         psize = pp->mediasize;
 441         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 442
 443         size = sizeof(*vdev_lists[0]) + pp->sectorsize -
 444             ((sizeof(*vdev_lists[0]) - 1) % pp->sectorsize) - 1;
 445
 446         buflen = sizeof(vdev_lists[0]->vp_nvlist);
 447
 448         /* Create all of the IO requests */
 449         for (l = 0; l < VDEV_LABELS; l++) {
 450                 cmds[l] = BIO_READ;
 451                 vdev_lists[l] = kmem_alloc(size, KM_SLEEP);
 452                 offsets[l] = vdev_label_offset(psize, l, 0) + VDEV_SKIP_SIZE;
 453                 sizes[l] = size;
 454                 errors[l] = 0;
 455                 ASSERT(offsets[l] % pp->sectorsize == 0);
 456         }
 457
 458         /* Issue the IO requests */
 459         vdev_geom_io(cp, cmds, (void**)vdev_lists, offsets, sizes, errors,
 460             VDEV_LABELS);
 461
 462         /* Parse the labels */
 463         config = *configp = NULL;
 464         nlabels = 0;
 465         for (l = 0; l < VDEV_LABELS; l++) {
 466                 if (errors[l] != 0)
 467                         continue;
 468
 469                 buf = vdev_lists[l]->vp_nvlist;
 470
 471                 if (nvlist_unpack(buf, buflen, &config, 0) != 0)
 472                         continue;
 473
 474                 if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE,
 475                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 476                         nvlist_free(config);
 477                         continue;
 478                 }
 479
 480                 if (state != POOL_STATE_SPARE &&
 481                     state != POOL_STATE_L2CACHE &&
 482                     (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG,
 483                     &txg) != 0 || txg == 0)) {
 484                         nvlist_free(config);
 485                         continue;
 486                 }
 487
 488                 if (*configp != NULL)
 489                         nvlist_free(*configp);
 490                 *configp = config;
 491
 492                 nlabels++;
 493         }
 494
 495         /* Free the label storage */
 496         for (l = 0; l < VDEV_LABELS; l++)
 497                 kmem_free(vdev_lists[l], size);
 498
 499         return (nlabels);
 500 }
 501
 502 static void
 503 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 504 {
 505         nvlist_t **new_configs;
 506         uint64_t i;
 507
 508         if (id < *count)
 509                 return;
 510         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 511             KM_SLEEP);
 512         for (i = 0; i < *count; i++)
 513                 new_configs[i] = (*configs)[i];
 514         if (*configs != NULL)
 515                 kmem_free(*configs, *count * sizeof(void *));
 516         *configs = new_configs;
 517         *count = id + 1;
 518 }
 519
 520 static void
 521 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 522     const char *name, uint64_t* known_pool_guid)
 523 {
 524         nvlist_t *vdev_tree;
 525         uint64_t pool_guid;
 526         uint64_t vdev_guid, known_guid;
 527         uint64_t id, txg, known_txg;
 528         char *pname;
 529         int i;
 530
 531         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 532             strcmp(pname, name) != 0)
 533                 goto ignore;
 534
 535         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 536                 goto ignore;
 537
 538         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 539                 goto ignore;
 540
 541         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 542                 goto ignore;
 543
 544         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 545                 goto ignore;
 546
 547         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 548
 549         if (*known_pool_guid != 0) {
 550                 if (pool_guid != *known_pool_guid)
 551                         goto ignore;
 552         } else
 553                 *known_pool_guid = pool_guid;
 554
 555         resize_configs(configs, count, id);
 556
 557         if ((*configs)[id] != NULL) {
 558                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 559                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 560                 if (txg <= known_txg)
 561                         goto ignore;
 562                 nvlist_free((*configs)[id]);
 563         }
 564
 565         (*configs)[id] = cfg;
 566         return;
 567
 568 ignore:
 569         nvlist_free(cfg);
 570 }
 571
 572 int
 573 vdev_geom_read_pool_label(const char *name,
 574     nvlist_t ***configs, uint64_t *count)
 575 {
 576         struct g_class *mp;
 577         struct g_geom *gp;
 578         struct g_provider *pp;
 579         struct g_consumer *zcp;
 580         nvlist_t *vdev_cfg;
 581         uint64_t pool_guid;
 582         int error, nlabels;
 583
 584         DROP_GIANT();
 585         g_topology_lock();
 586
 587         *configs = NULL;
 588         *count = 0;
 589         pool_guid = 0;
 590         LIST_FOREACH(mp, &g_classes, class) {
 591                 if (mp == &zfs_vdev_class)
 592                         continue;
 593                 LIST_FOREACH(gp, &mp->geom, geom) {
 594                         if (gp->flags & G_GEOM_WITHER)
 595                                 continue;
 596                         LIST_FOREACH(pp, &gp->provider, provider) {
 597                                 if (pp->flags & G_PF_WITHER)
 598                                         continue;
 599                                 zcp = vdev_geom_attach(pp, NULL, B_TRUE);
 600                                 if (zcp == NULL)
 601                                         continue;
 602                                 g_topology_unlock();
 603                                 nlabels = vdev_geom_read_config(zcp, &vdev_cfg);
 604                                 g_topology_lock();
 605                                 vdev_geom_detach(zcp, B_TRUE);
 606                                 if (nlabels == 0)
 607                                         continue;
 608                                 ZFS_LOG(1, "successfully read vdev config");
 609
 610                                 process_vdev_config(configs, count,
 611                                     vdev_cfg, name, &pool_guid);
 612                         }
 613                 }
 614         }
 615         g_topology_unlock();
 616         PICKUP_GIANT();
 617
 618         return (*count > 0 ? 0 : ENOENT);
 619 }
 620
 621 enum match {
 622         NO_MATCH = 0,           /* No matching labels found */
 623         TOPGUID_MATCH = 1,      /* Labels match top guid, not vdev guid*/
 624         ZERO_MATCH = 1,         /* Should never be returned */
 625         ONE_MATCH = 2,          /* 1 label matching the vdev_guid */
 626         TWO_MATCH = 3,          /* 2 label matching the vdev_guid */
 627         THREE_MATCH = 4,        /* 3 label matching the vdev_guid */
 628         FULL_MATCH = 5          /* all labels match the vdev_guid */
 629 };
 630
 631 static enum match
 632 vdev_attach_ok(vdev_t *vd, struct g_provider *pp)
 633 {
 634         nvlist_t *config;
 635         uint64_t pool_guid, top_guid, vdev_guid;
 636         struct g_consumer *cp;
 637         int nlabels;
 638
 639         cp = vdev_geom_attach(pp, NULL, B_TRUE);
 640         if (cp == NULL) {
 641                 ZFS_LOG(1, "Unable to attach tasting instance to %s.",
 642                     pp->name);
 643                 return (NO_MATCH);
 644         }
 645         g_topology_unlock();
 646         nlabels = vdev_geom_read_config(cp, &config);
 647         g_topology_lock();
 648         vdev_geom_detach(cp, B_TRUE);
 649         if (nlabels == 0) {
 650                 ZFS_LOG(1, "Unable to read config from %s.", pp->name);
 651                 return (NO_MATCH);
 652         }
 653
 654         pool_guid = 0;
 655         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid);
 656         top_guid = 0;
 657         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TOP_GUID, &top_guid);
 658         vdev_guid = 0;
 659         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_GUID, &vdev_guid);
 660         nvlist_free(config);
 661
 662         /*
 663          * Check that the label's pool guid matches the desired guid.
 664          * Inactive spares and L2ARCs do not have any pool guid in the label.
 665          */
 666         if (pool_guid != 0 && pool_guid != spa_guid(vd->vdev_spa)) {
 667                 ZFS_LOG(1, "pool guid mismatch for provider %s: %ju != %ju.",
 668                     pp->name,
 669                     (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)pool_guid);
 670                 return (NO_MATCH);
 671         }
 672
 673         /*
 674          * Check that the label's vdev guid matches the desired guid.
 675          * The second condition handles possible race on vdev detach, when
 676          * remaining vdev receives GUID of destroyed top level mirror vdev.
 677          */
 678         if (vdev_guid == vd->vdev_guid) {
 679                 ZFS_LOG(1, "guids match for provider %s.", pp->name);
 680                 return (ZERO_MATCH + nlabels);
 681         } else if (top_guid == vd->vdev_guid && vd == vd->vdev_top) {
 682                 ZFS_LOG(1, "top vdev guid match for provider %s.", pp->name);
 683                 return (TOPGUID_MATCH);
 684         }
 685         ZFS_LOG(1, "vdev guid mismatch for provider %s: %ju != %ju.",
 686             pp->name, (uintmax_t)vd->vdev_guid, (uintmax_t)vdev_guid);
 687         return (NO_MATCH);
 688 }
 689
 690 static struct g_consumer *
 691 vdev_geom_attach_by_guids(vdev_t *vd)
 692 {
 693         struct g_class *mp;
 694         struct g_geom *gp;
 695         struct g_provider *pp, *best_pp;
 696         struct g_consumer *cp;
 697         enum match match, best_match;
 698
 699         g_topology_assert();
 700
 701         cp = NULL;
 702         best_pp = NULL;
 703         best_match = NO_MATCH;
 704         LIST_FOREACH(mp, &g_classes, class) {
 705                 if (mp == &zfs_vdev_class)
 706                         continue;
 707                 LIST_FOREACH(gp, &mp->geom, geom) {
 708                         if (gp->flags & G_GEOM_WITHER)
 709                                 continue;
 710                         LIST_FOREACH(pp, &gp->provider, provider) {
 711                                 match = vdev_attach_ok(vd, pp);
 712                                 if (match > best_match) {
 713                                         best_match = match;
 714                                         best_pp = pp;
 715                                 }
 716                                 if (match == FULL_MATCH)
 717                                         goto out;
 718                         }
 719                 }
 720         }
 721
 722 out:
 723         if (best_pp) {
 724                 cp = vdev_geom_attach(best_pp, vd, B_TRUE);
 725                 if (cp == NULL) {
 726                         printf("ZFS WARNING: Unable to attach to %s.\n",
 727                             best_pp->name);
 728                 }
 729         }
 730         return (cp);
 731 }
 732
 733 static struct g_consumer *
 734 vdev_geom_open_by_guids(vdev_t *vd)
 735 {
 736         struct g_consumer *cp;
 737         char *buf;
 738         size_t len;
 739
 740         g_topology_assert();
 741
 742         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 743                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 744         cp = vdev_geom_attach_by_guids(vd);
 745         if (cp != NULL) {
 746                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 747                 buf = kmem_alloc(len, KM_SLEEP);
 748
 749                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 750                 spa_strfree(vd->vdev_path);
 751                 vd->vdev_path = buf;
 752
 753                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 754                     (uintmax_t)spa_guid(vd->vdev_spa),
 755                     (uintmax_t)vd->vdev_guid, cp->provider->name);
 756         } else {
 757                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 758                     (uintmax_t)spa_guid(vd->vdev_spa),
 759                     (uintmax_t)vd->vdev_guid);
 760         }
 761
 762         return (cp);
 763 }
 764
 765 static struct g_consumer *
 766 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 767 {
 768         struct g_provider *pp;
 769         struct g_consumer *cp;
 770
 771         g_topology_assert();
 772
 773         cp = NULL;
 774         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 775         if (pp != NULL) {
 776                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 777                 if (!check_guid || vdev_attach_ok(vd, pp) == FULL_MATCH)
 778                         cp = vdev_geom_attach(pp, vd, B_FALSE);
 779         }
 780
 781         return (cp);
 782 }
 783
 784 static int
 785 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 786     uint64_t *logical_ashift, uint64_t *physical_ashift)
 787 {
 788         struct g_provider *pp;
 789         struct g_consumer *cp;
 790         size_t bufsize;
 791         int error;
 792
 793         /* Set the TLS to indicate downstack that we should not access zvols*/
 794         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 795
 796         /*
 797          * We must have a pathname, and it must be absolute.
 798          */
 799         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 800                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 801                 return (EINVAL);
 802         }
 803
 804         /*
 805          * Reopen the device if it's not currently open. Otherwise,
 806          * just update the physical size of the device.
 807          */
 808         if ((cp = vd->vdev_tsd) != NULL) {
 809                 ASSERT(vd->vdev_reopening);
 810                 goto skip_open;
 811         }
 812
 813         DROP_GIANT();
 814         g_topology_lock();
 815         error = 0;
 816
 817         if (vd->vdev_spa->spa_splitting_newspa ||
 818             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 819              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 820              vd->vdev_spa->spa_load_state == SPA_LOAD_CREATE)) {
 821                 /*
 822                  * We are dealing with a vdev that hasn't been previously
 823                  * opened (since boot), and we are not loading an
 824                  * existing pool configuration.  This looks like a
 825                  * vdev add operation to a new or existing pool.
 826                  * Assume the user knows what he/she is doing and find
 827                  * GEOM provider by its name, ignoring GUID mismatches.
 828                  *
 829                  * XXPOLICY: It would be safer to only allow a device
 830                  *           that is unlabeled or labeled but missing
 831                  *           GUID information to be opened in this fashion,
 832                  *           unless we are doing a split, in which case we
 833                  *           should allow any guid.
 834                  */
 835                 cp = vdev_geom_open_by_path(vd, 0);
 836         } else {
 837                 /*
 838                  * Try using the recorded path for this device, but only
 839                  * accept it if its label data contains the expected GUIDs.
 840                  */
 841                 cp = vdev_geom_open_by_path(vd, 1);
 842                 if (cp == NULL) {
 843                         /*
 844                          * The device at vd->vdev_path doesn't have the
 845                          * expected GUIDs. The disks might have merely
 846                          * moved around so try all other GEOM providers
 847                          * to find one with the right GUIDs.
 848                          */
 849                         cp = vdev_geom_open_by_guids(vd);
 850                 }
 851         }
 852
 853         /* Clear the TLS now that tasting is done */
 854         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 855
 856         if (cp == NULL) {
 857                 ZFS_LOG(1, "Vdev %s not found.", vd->vdev_path);
 858                 error = ENOENT;
 859         } else {
 860                 struct consumer_priv_t *priv;
 861                 struct consumer_vdev_elem *elem;
 862                 int spamode;
 863
 864                 priv = (struct consumer_priv_t*)&cp->private;
 865                 if (cp->private == NULL)
 866                         SLIST_INIT(priv);
 867                 elem = g_malloc(sizeof(*elem), M_WAITOK|M_ZERO);
 868                 elem->vd = vd;
 869                 SLIST_INSERT_HEAD(priv, elem, elems);
 870
 871                 spamode = spa_mode(vd->vdev_spa);
 872                 if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 873                     !ISP2(cp->provider->sectorsize)) {
 874                         ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 875                             cp->provider->name);
 876
 877                         vdev_geom_close_locked(vd);
 878                         error = EINVAL;
 879                         cp = NULL;
 880                 } else if (cp->acw == 0 && (spamode & FWRITE) != 0) {
 881                         int i;
 882
 883                         for (i = 0; i < 5; i++) {
 884                                 error = g_access(cp, 0, 1, 0);
 885                                 if (error == 0)
 886                                         break;
 887                                 g_topology_unlock();
 888                                 tsleep(vd, 0, "vdev", hz / 2);
 889                                 g_topology_lock();
 890                         }
 891                         if (error != 0) {
 892                                 printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 893                                     cp->provider->name, error);
 894                                 vdev_geom_close_locked(vd);
 895                                 cp = NULL;
 896                         }
 897                 }
 898         }
 899
 900         /* Fetch initial physical path information for this device. */
 901         if (cp != NULL) {
 902                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 903
 904                 /* Set other GEOM characteristics */
 905                 vdev_geom_set_physpath(vd, cp, /*do_null_update*/B_FALSE);
 906                 vdev_geom_set_rotation_rate(vd, cp);
 907         }
 908
 909         g_topology_unlock();
 910         PICKUP_GIANT();
 911         if (cp == NULL) {
 912                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 913                 return (error);
 914         }
 915 skip_open:
 916         pp = cp->provider;
 917
 918         /*
 919          * Determine the actual size of the device.
 920          */
 921         *max_psize = *psize = pp->mediasize;
 922
 923         /*
 924          * Determine the device's minimum transfer size and preferred
 925          * transfer size.
 926          */
 927         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 928         *physical_ashift = 0;
 929         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 930             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 931                 *physical_ashift = highbit(pp->stripesize) - 1;
 932
 933         /*
 934          * Clear the nowritecache settings, so that on a vdev_reopen()
 935          * we will try again.
 936          */
 937         vd->vdev_nowritecache = B_FALSE;
 938
 939         return (0);
 940 }
 941
 942 static void
 943 vdev_geom_close(vdev_t *vd)
 944 {
 945         struct g_consumer *cp;
 946
 947         cp = vd->vdev_tsd;
 948
 949         DROP_GIANT();
 950         g_topology_lock();
 951
 952         if (!vd->vdev_reopening ||
 953             (cp != NULL && ((cp->flags & G_CF_ORPHAN) != 0 ||
 954             (cp->provider != NULL && cp->provider->error != 0))))
 955                 vdev_geom_close_locked(vd);
 956
 957         g_topology_unlock();
 958         PICKUP_GIANT();
 959 }
 960
 961 static void
 962 vdev_geom_io_intr(struct bio *bp)
 963 {
 964         vdev_t *vd;
 965         zio_t *zio;
 966
 967         zio = bp->bio_caller1;
 968         vd = zio->io_vd;
 969         zio->io_error = bp->bio_error;
 970         if (zio->io_error == 0 && bp->bio_resid != 0)
 971                 zio->io_error = SET_ERROR(EIO);
 972
 973         switch(zio->io_error) {
 974         case ENOTSUP:
 975                 /*
 976                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 977                  * that future attempts will never succeed. In this case
 978                  * we set a persistent flag so that we don't bother with
 979                  * requests in the future.
 980                  */
 981                 switch(bp->bio_cmd) {
 982                 case BIO_FLUSH:
 983                         vd->vdev_nowritecache = B_TRUE;
 984                         break;
 985                 case BIO_DELETE:
 986                         vd->vdev_notrim = B_TRUE;
 987                         break;
 988                 }
 989                 break;
 990         case ENXIO:
 991                 if (!vd->vdev_remove_wanted) {
 992                         /*
 993                          * If provider's error is set we assume it is being
 994                          * removed.
 995                          */
 996                         if (bp->bio_to->error != 0) {
 997                                 vd->vdev_remove_wanted = B_TRUE;
 998                                 spa_async_request(zio->io_spa,
 999                                     SPA_ASYNC_REMOVE);
1000                         } else if (!vd->vdev_delayed_close) {
1001                                 vd->vdev_delayed_close = B_TRUE;
1002                         }
1003                 }
1004                 break;
1005         }
1006         g_destroy_bio(bp);
1007         zio_delay_interrupt(zio);
1008 }
1009
1010 static void
1011 vdev_geom_io_start(zio_t *zio)
1012 {
1013         vdev_t *vd;
1014         struct g_consumer *cp;
1015         struct bio *bp;
1016         int error;
1017
1018         vd = zio->io_vd;
1019
1020         switch (zio->io_type) {
1021         case ZIO_TYPE_IOCTL:
1022                 /* XXPOLICY */
1023                 if (!vdev_readable(vd)) {
1024                         zio->io_error = SET_ERROR(ENXIO);
1025                         zio_interrupt(zio);
1026                         return;
1027                 } else {
1028                         switch (zio->io_cmd) {
1029                         case DKIOCFLUSHWRITECACHE:
1030                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
1031                                         break;
1032                                 if (vd->vdev_nowritecache) {
1033                                         zio->io_error = SET_ERROR(ENOTSUP);
1034                                         break;
1035                                 }
1036                                 goto sendreq;
1037                         default:
1038                                 zio->io_error = SET_ERROR(ENOTSUP);
1039                         }
1040                 }
1041
1042                 zio_execute(zio);
1043                 return;
1044         case ZIO_TYPE_FREE:
1045                 if (vd->vdev_notrim) {
1046                         zio->io_error = SET_ERROR(ENOTSUP);
1047                 } else if (!vdev_geom_bio_delete_disable) {
1048                         goto sendreq;
1049                 }
1050                 zio_execute(zio);
1051                 return;
1052         }
1053 sendreq:
1054         ASSERT(zio->io_type == ZIO_TYPE_READ ||
1055             zio->io_type == ZIO_TYPE_WRITE ||
1056             zio->io_type == ZIO_TYPE_FREE ||
1057             zio->io_type == ZIO_TYPE_IOCTL);
1058
1059         cp = vd->vdev_tsd;
1060         if (cp == NULL) {
1061                 zio->io_error = SET_ERROR(ENXIO);
1062                 zio_interrupt(zio);
1063                 return;
1064         }
1065         bp = g_alloc_bio();
1066         bp->bio_caller1 = zio;
1067         switch (zio->io_type) {
1068         case ZIO_TYPE_READ:
1069         case ZIO_TYPE_WRITE:
1070                 zio->io_target_timestamp = zio_handle_io_delay(zio);
1071                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
1072                 bp->bio_data = zio->io_data;
1073                 bp->bio_offset = zio->io_offset;
1074                 bp->bio_length = zio->io_size;
1075                 break;
1076         case ZIO_TYPE_FREE:
1077                 bp->bio_cmd = BIO_DELETE;
1078                 bp->bio_data = NULL;
1079                 bp->bio_offset = zio->io_offset;
1080                 bp->bio_length = zio->io_size;
1081                 break;
1082         case ZIO_TYPE_IOCTL:
1083                 bp->bio_cmd = BIO_FLUSH;
1084                 bp->bio_flags |= BIO_ORDERED;
1085                 bp->bio_data = NULL;
1086                 bp->bio_offset = cp->provider->mediasize;
1087                 bp->bio_length = 0;
1088                 break;
1089         }
1090         bp->bio_done = vdev_geom_io_intr;
1091
1092         g_io_request(bp, cp);
1093 }
1094
1095 static void
1096 vdev_geom_io_done(zio_t *zio)
1097 {
1098 }
1099
1100 static void
1101 vdev_geom_hold(vdev_t *vd)
1102 {
1103 }
1104
1105 static void
1106 vdev_geom_rele(vdev_t *vd)
1107 {
1108 }
1109
1110 vdev_ops_t vdev_geom_ops = {
1111         vdev_geom_open,
1112         vdev_geom_close,
1113         vdev_default_asize,
1114         vdev_geom_io_start,
1115         vdev_geom_io_done,
1116         NULL,
1117         vdev_geom_hold,
1118         vdev_geom_rele,
1119         VDEV_TYPE_DISK,         /* name of this vdev type */
1120         B_TRUE                  /* leaf vdev */
1121 };