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