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;
  67 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_flush_disable, CTLFLAG_RWTUN,
  68     &vdev_geom_bio_flush_disable, 0, "Disable BIO_FLUSH");
  69 /* Don't send BIO_DELETE. */
  70 static int vdev_geom_bio_delete_disable;
  71 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, bio_delete_disable, CTLFLAG_RWTUN,
  72     &vdev_geom_bio_delete_disable, 0, "Disable BIO_DELETE");
  73
  74 /* Declare local functions */
  75 static void vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read);
  76
  77 /*
  78  * Thread local storage used to indicate when a thread is probing geoms
  79  * for their guids.  If NULL, this thread is not tasting geoms.  If non NULL,
  80  * it is looking for a replacement for the vdev_t* that is its value.
  81  */
  82 uint_t zfs_geom_probe_vdev_key;
  83
  84 static void
  85 vdev_geom_set_rotation_rate(vdev_t *vd, struct g_consumer *cp)
  86 {
  87         int error;
  88         uint16_t rate;
  89
  90         error = g_getattr("GEOM::rotation_rate", cp, &rate);
  91         if (error == 0)
  92                 vd->vdev_rotation_rate = rate;
  93         else
  94                 vd->vdev_rotation_rate = VDEV_RATE_UNKNOWN;
  95 }
  96
  97 static void
  98 vdev_geom_set_physpath(vdev_t *vd, struct g_consumer *cp,
  99                        boolean_t do_null_update)
 100 {
 101         boolean_t needs_update = B_FALSE;
 102         char *physpath;
 103         int error, physpath_len;
 104
 105         physpath_len = MAXPATHLEN;
 106         physpath = g_malloc(physpath_len, M_WAITOK|M_ZERO);
 107         error = g_io_getattr("GEOM::physpath", cp, &physpath_len, physpath);
 108         if (error == 0) {
 109                 char *old_physpath;
 110
 111                 /* g_topology lock ensures that vdev has not been closed */
 112                 g_topology_assert();
 113                 old_physpath = vd->vdev_physpath;
 114                 vd->vdev_physpath = spa_strdup(physpath);
 115
 116                 if (old_physpath != NULL) {
 117                         needs_update = (strcmp(old_physpath,
 118                                                 vd->vdev_physpath) != 0);
 119                         spa_strfree(old_physpath);
 120                 } else
 121                         needs_update = do_null_update;
 122         }
 123         g_free(physpath);
 124
 125         /*
 126          * If the physical path changed, update the config.
 127          * Only request an update for previously unset physpaths if
 128          * requested by the caller.
 129          */
 130         if (needs_update)
 131                 spa_async_request(vd->vdev_spa, SPA_ASYNC_CONFIG_UPDATE);
 132
 133 }
 134
 135 static void
 136 vdev_geom_attrchanged(struct g_consumer *cp, const char *attr)
 137 {
 138         char *old_physpath;
 139         struct consumer_priv_t *priv;
 140         struct consumer_vdev_elem *elem;
 141         int error;
 142
 143         priv = (struct consumer_priv_t*)&cp->private;
 144         if (SLIST_EMPTY(priv))
 145                 return;
 146
 147         SLIST_FOREACH(elem, priv, elems) {
 148                 vdev_t *vd = elem->vd;
 149                 if (strcmp(attr, "GEOM::rotation_rate") == 0) {
 150                         vdev_geom_set_rotation_rate(vd, cp);
 151                         return;
 152                 }
 153                 if (strcmp(attr, "GEOM::physpath") == 0) {
 154                         vdev_geom_set_physpath(vd, cp, /*null_update*/B_TRUE);
 155                         return;
 156                 }
 157         }
 158 }
 159
 160 static void
 161 vdev_geom_orphan(struct g_consumer *cp)
 162 {
 163         struct consumer_priv_t *priv;
 164         struct consumer_vdev_elem *elem;
 165
 166         g_topology_assert();
 167
 168         priv = (struct consumer_priv_t*)&cp->private;
 169         if (SLIST_EMPTY(priv))
 170                 /* Vdev close in progress.  Ignore the event. */
 171                 return;
 172
 173         /*
 174          * Orphan callbacks occur from the GEOM event thread.
 175          * Concurrent with this call, new I/O requests may be
 176          * working their way through GEOM about to find out
 177          * (only once executed by the g_down thread) that we've
 178          * been orphaned from our disk provider.  These I/Os
 179          * must be retired before we can detach our consumer.
 180          * This is most easily achieved by acquiring the
 181          * SPA ZIO configuration lock as a writer, but doing
 182          * so with the GEOM topology lock held would cause
 183          * a lock order reversal.  Instead, rely on the SPA's
 184          * async removal support to invoke a close on this
 185          * vdev once it is safe to do so.
 186          */
 187         SLIST_FOREACH(elem, priv, elems) {
 188                 vdev_t *vd = elem->vd;
 189
 190                 vd->vdev_remove_wanted = B_TRUE;
 191                 spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
 192         }
 193 }
 194
 195 static struct g_consumer *
 196 vdev_geom_attach(struct g_provider *pp, vdev_t *vd)
 197 {
 198         struct g_geom *gp;
 199         struct g_consumer *cp;
 200         int error;
 201
 202         g_topology_assert();
 203
 204         ZFS_LOG(1, "Attaching to %s.", pp->name);
 205
 206         if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize)) {
 207                 ZFS_LOG(1, "Failing attach of %s. Incompatible sectorsize %d\n",
 208                     pp->name, pp->sectorsize);
 209                 return (NULL);
 210         } else if (pp->mediasize < SPA_MINDEVSIZE) {
 211                 ZFS_LOG(1, "Failing attach of %s. Incompatible mediasize %ju\n",
 212                     pp->name, pp->mediasize);
 213                 return (NULL);
 214         }
 215
 216         /* Do we have geom already? No? Create one. */
 217         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 218                 if (gp->flags & G_GEOM_WITHER)
 219                         continue;
 220                 if (strcmp(gp->name, "zfs::vdev") != 0)
 221                         continue;
 222                 break;
 223         }
 224         if (gp == NULL) {
 225                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 226                 gp->orphan = vdev_geom_orphan;
 227                 gp->attrchanged = vdev_geom_attrchanged;
 228                 cp = g_new_consumer(gp);
 229                 error = g_attach(cp, pp);
 230                 if (error != 0) {
 231                         ZFS_LOG(1, "%s(%d): g_attach failed: %d\n", __func__,
 232                             __LINE__, error);
 233                         vdev_geom_detach(cp, B_FALSE);
 234                         return (NULL);
 235                 }
 236                 error = g_access(cp, 1, 0, 1);
 237                 if (error != 0) {
 238                         ZFS_LOG(1, "%s(%d): g_access failed: %d", __func__,
 239                                __LINE__, error);
 240                         vdev_geom_detach(cp, B_FALSE);
 241                         return (NULL);
 242                 }
 243                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 244         } else {
 245                 /* Check if we are already connected to this provider. */
 246                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 247                         if (cp->provider == pp) {
 248                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 249                                 break;
 250                         }
 251                 }
 252                 if (cp == NULL) {
 253                         cp = g_new_consumer(gp);
 254                         error = g_attach(cp, pp);
 255                         if (error != 0) {
 256                                 ZFS_LOG(1, "%s(%d): g_attach failed: %d\n",
 257                                     __func__, __LINE__, error);
 258                                 vdev_geom_detach(cp, B_FALSE);
 259                                 return (NULL);
 260                         }
 261                         error = g_access(cp, 1, 0, 1);
 262                         if (error != 0) {
 263                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 264                                     __func__, __LINE__, error);
 265                                 vdev_geom_detach(cp, B_FALSE);
 266                                 return (NULL);
 267                         }
 268                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 269                 } else {
 270                         error = g_access(cp, 1, 0, 1);
 271                         if (error != 0) {
 272                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 273                                     __func__, __LINE__, error);
 274                                 return (NULL);
 275                         }
 276                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 277                 }
 278         }
 279
 280         if (vd != NULL)
 281                 vd->vdev_tsd = cp;
 282
 283         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
 284         return (cp);
 285 }
 286
 287 static void
 288 vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read)
 289 {
 290         struct g_geom *gp;
 291
 292         g_topology_assert();
 293
 294         ZFS_LOG(1, "Detaching from %s.",
 295             cp->provider && cp->provider->name ? cp->provider->name : "NULL");
 296
 297         gp = cp->geom;
 298         if (open_for_read)
 299                 g_access(cp, -1, 0, -1);
 300         /* Destroy consumer on last close. */
 301         if (cp->acr == 0 && cp->ace == 0) {
 302                 if (cp->acw > 0)
 303                         g_access(cp, 0, -cp->acw, 0);
 304                 if (cp->provider != NULL) {
 305                         ZFS_LOG(1, "Destroying consumer for %s.",
 306                             cp->provider->name ? cp->provider->name : "NULL");
 307                         g_detach(cp);
 308                 }
 309                 g_destroy_consumer(cp);
 310         }
 311         /* Destroy geom if there are no consumers left. */
 312         if (LIST_EMPTY(&gp->consumer)) {
 313                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 314                 g_wither_geom(gp, ENXIO);
 315         }
 316 }
 317
 318 static void
 319 vdev_geom_close_locked(vdev_t *vd)
 320 {
 321         struct g_consumer *cp;
 322         struct consumer_priv_t *priv;
 323         struct consumer_vdev_elem *elem, *elem_temp;
 324
 325         g_topology_assert();
 326
 327         cp = vd->vdev_tsd;
 328         vd->vdev_delayed_close = B_FALSE;
 329         if (cp == NULL)
 330                 return;
 331
 332         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 333         KASSERT(cp->private != NULL, ("%s: cp->private is NULL", __func__));
 334         priv = (struct consumer_priv_t*)&cp->private;
 335         vd->vdev_tsd = NULL;
 336         SLIST_FOREACH_SAFE(elem, priv, elems, elem_temp) {
 337                 if (elem->vd == vd) {
 338                         SLIST_REMOVE(priv, elem, consumer_vdev_elem, elems);
 339                         g_free(elem);
 340                 }
 341         }
 342
 343         vdev_geom_detach(cp, B_TRUE);
 344 }
 345
 346 /*
 347  * Issue one or more bios to the vdev in parallel
 348  * cmds, datas, offsets, errors, and sizes are arrays of length ncmds.  Each IO
 349  * operation is described by parallel entries from each array.  There may be
 350  * more bios actually issued than entries in the array
 351  */
 352 static void
 353 vdev_geom_io(struct g_consumer *cp, int *cmds, void **datas, off_t *offsets,
 354     off_t *sizes, int *errors, int ncmds)
 355 {
 356         struct bio **bios;
 357         u_char *p;
 358         off_t off, maxio, s, end;
 359         int i, n_bios, j;
 360         size_t bios_size;
 361
 362         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 363         n_bios = 0;
 364
 365         /* How many bios are required for all commands ? */
 366         for (i = 0; i < ncmds; i++)
 367                 n_bios += (sizes[i] + maxio - 1) / maxio;
 368
 369         /* Allocate memory for the bios */
 370         bios_size = n_bios * sizeof(struct bio*);
 371         bios = kmem_zalloc(bios_size, KM_SLEEP);
 372
 373         /* Prepare and issue all of the bios */
 374         for (i = j = 0; i < ncmds; i++) {
 375                 off = offsets[i];
 376                 p = datas[i];
 377                 s = sizes[i];
 378                 end = off + s;
 379                 ASSERT((off % cp->provider->sectorsize) == 0);
 380                 ASSERT((s % cp->provider->sectorsize) == 0);
 381
 382                 for (; off < end; off += maxio, p += maxio, s -= maxio, j++) {
 383                         bios[j] = g_alloc_bio();
 384                         bios[j]->bio_cmd = cmds[i];
 385                         bios[j]->bio_done = NULL;
 386                         bios[j]->bio_offset = off;
 387                         bios[j]->bio_length = MIN(s, maxio);
 388                         bios[j]->bio_data = p;
 389                         g_io_request(bios[j], cp);
 390                 }
 391         }
 392         ASSERT(j == n_bios);
 393
 394         /* Wait for all of the bios to complete, and clean them up */
 395         for (i = j = 0; i < ncmds; i++) {
 396                 off = offsets[i];
 397                 s = sizes[i];
 398                 end = off + s;
 399
 400                 for (; off < end; off += maxio, s -= maxio, j++) {
 401                         errors[i] = biowait(bios[j], "vdev_geom_io") || errors[i];
 402                         g_destroy_bio(bios[j]);
 403                 }
 404         }
 405         kmem_free(bios, bios_size);
 406 }
 407
 408 /*
 409  * Read the vdev config from a device.  Return the number of valid labels that
 410  * were found.  The vdev config will be returned in config if and only if at
 411  * least one valid label was found.
 412  */
 413 static int
 414 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **config)
 415 {
 416         struct g_provider *pp;
 417         vdev_phys_t *vdev_lists[VDEV_LABELS];
 418         char *buf;
 419         size_t buflen;
 420         uint64_t psize, state, txg;
 421         off_t offsets[VDEV_LABELS];
 422         off_t size;
 423         off_t sizes[VDEV_LABELS];
 424         int cmds[VDEV_LABELS];
 425         int errors[VDEV_LABELS];
 426         int l, nlabels;
 427
 428         g_topology_assert_not();
 429
 430         pp = cp->provider;
 431         ZFS_LOG(1, "Reading config from %s...", pp->name);
 432
 433         psize = pp->mediasize;
 434         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 435
 436         size = sizeof(*vdev_lists[0]) + pp->sectorsize -
 437             ((sizeof(*vdev_lists[0]) - 1) % pp->sectorsize) - 1;
 438
 439         buflen = sizeof(vdev_lists[0]->vp_nvlist);
 440
 441         *config = NULL;
 442         /* Create all of the IO requests */
 443         for (l = 0; l < VDEV_LABELS; l++) {
 444                 cmds[l] = BIO_READ;
 445                 vdev_lists[l] = kmem_alloc(size, KM_SLEEP);
 446                 offsets[l] = vdev_label_offset(psize, l, 0) + VDEV_SKIP_SIZE;
 447                 sizes[l] = size;
 448                 errors[l] = 0;
 449                 ASSERT(offsets[l] % pp->sectorsize == 0);
 450         }
 451
 452         /* Issue the IO requests */
 453         vdev_geom_io(cp, cmds, (void**)vdev_lists, offsets, sizes, errors,
 454             VDEV_LABELS);
 455
 456         /* Parse the labels */
 457         nlabels = 0;
 458         for (l = 0; l < VDEV_LABELS; l++) {
 459                 if (errors[l] != 0)
 460                         continue;
 461
 462                 buf = vdev_lists[l]->vp_nvlist;
 463
 464                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 465                         continue;
 466
 467                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 468                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 469                         nvlist_free(*config);
 470                         *config = NULL;
 471                         continue;
 472                 }
 473
 474                 if (state != POOL_STATE_SPARE &&
 475                     state != POOL_STATE_L2CACHE &&
 476                     (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
 477                     &txg) != 0 || txg == 0)) {
 478                         nvlist_free(*config);
 479                         *config = NULL;
 480                         continue;
 481                 }
 482
 483                 nlabels++;
 484         }
 485
 486         /* Free the label storage */
 487         for (l = 0; l < VDEV_LABELS; l++)
 488                 kmem_free(vdev_lists[l], size);
 489
 490         return (nlabels);
 491 }
 492
 493 static void
 494 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 495 {
 496         nvlist_t **new_configs;
 497         uint64_t i;
 498
 499         if (id < *count)
 500                 return;
 501         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 502             KM_SLEEP);
 503         for (i = 0; i < *count; i++)
 504                 new_configs[i] = (*configs)[i];
 505         if (*configs != NULL)
 506                 kmem_free(*configs, *count * sizeof(void *));
 507         *configs = new_configs;
 508         *count = id + 1;
 509 }
 510
 511 static void
 512 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 513     const char *name, uint64_t* known_pool_guid)
 514 {
 515         nvlist_t *vdev_tree;
 516         uint64_t pool_guid;
 517         uint64_t vdev_guid, known_guid;
 518         uint64_t id, txg, known_txg;
 519         char *pname;
 520         int i;
 521
 522         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 523             strcmp(pname, name) != 0)
 524                 goto ignore;
 525
 526         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 527                 goto ignore;
 528
 529         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 530                 goto ignore;
 531
 532         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 533                 goto ignore;
 534
 535         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 536                 goto ignore;
 537
 538         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 539
 540         if (*known_pool_guid != 0) {
 541                 if (pool_guid != *known_pool_guid)
 542                         goto ignore;
 543         } else
 544                 *known_pool_guid = pool_guid;
 545
 546         resize_configs(configs, count, id);
 547
 548         if ((*configs)[id] != NULL) {
 549                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 550                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 551                 if (txg <= known_txg)
 552                         goto ignore;
 553                 nvlist_free((*configs)[id]);
 554         }
 555
 556         (*configs)[id] = cfg;
 557         return;
 558
 559 ignore:
 560         nvlist_free(cfg);
 561 }
 562
 563 int
 564 vdev_geom_read_pool_label(const char *name,
 565     nvlist_t ***configs, uint64_t *count)
 566 {
 567         struct g_class *mp;
 568         struct g_geom *gp;
 569         struct g_provider *pp;
 570         struct g_consumer *zcp;
 571         nvlist_t *vdev_cfg;
 572         uint64_t pool_guid;
 573         int error, nlabels;
 574
 575         DROP_GIANT();
 576         g_topology_lock();
 577
 578         *configs = NULL;
 579         *count = 0;
 580         pool_guid = 0;
 581         LIST_FOREACH(mp, &g_classes, class) {
 582                 if (mp == &zfs_vdev_class)
 583                         continue;
 584                 LIST_FOREACH(gp, &mp->geom, geom) {
 585                         if (gp->flags & G_GEOM_WITHER)
 586                                 continue;
 587                         LIST_FOREACH(pp, &gp->provider, provider) {
 588                                 if (pp->flags & G_PF_WITHER)
 589                                         continue;
 590                                 zcp = vdev_geom_attach(pp, NULL);
 591                                 if (zcp == NULL)
 592                                         continue;
 593                                 g_topology_unlock();
 594                                 nlabels = vdev_geom_read_config(zcp, &vdev_cfg);
 595                                 g_topology_lock();
 596                                 vdev_geom_detach(zcp, B_TRUE);
 597                                 if (nlabels == 0)
 598                                         continue;
 599                                 ZFS_LOG(1, "successfully read vdev config");
 600
 601                                 process_vdev_config(configs, count,
 602                                     vdev_cfg, name, &pool_guid);
 603                         }
 604                 }
 605         }
 606         g_topology_unlock();
 607         PICKUP_GIANT();
 608
 609         return (*count > 0 ? 0 : ENOENT);
 610 }
 611
 612 enum match {
 613         NO_MATCH = 0,           /* No matching labels found */
 614         TOPGUID_MATCH = 1,      /* Labels match top guid, not vdev guid*/
 615         ZERO_MATCH = 1,         /* Should never be returned */
 616         ONE_MATCH = 2,          /* 1 label matching the vdev_guid */
 617         TWO_MATCH = 3,          /* 2 label matching the vdev_guid */
 618         THREE_MATCH = 4,        /* 3 label matching the vdev_guid */
 619         FULL_MATCH = 5          /* all labels match the vdev_guid */
 620 };
 621
 622 static enum match
 623 vdev_attach_ok(vdev_t *vd, struct g_provider *pp)
 624 {
 625         nvlist_t *config;
 626         uint64_t pool_guid, top_guid, vdev_guid;
 627         struct g_consumer *cp;
 628         int nlabels;
 629
 630         cp = vdev_geom_attach(pp, NULL);
 631         if (cp == NULL) {
 632                 ZFS_LOG(1, "Unable to attach tasting instance to %s.",
 633                     pp->name);
 634                 return (NO_MATCH);
 635         }
 636         g_topology_unlock();
 637         nlabels = vdev_geom_read_config(cp, &config);
 638         if (nlabels == 0) {
 639                 g_topology_lock();
 640                 vdev_geom_detach(cp, B_TRUE);
 641                 ZFS_LOG(1, "Unable to read config from %s.", pp->name);
 642                 return (NO_MATCH);
 643         }
 644         g_topology_lock();
 645         vdev_geom_detach(cp, B_TRUE);
 646
 647         pool_guid = 0;
 648         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid);
 649         top_guid = 0;
 650         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TOP_GUID, &top_guid);
 651         vdev_guid = 0;
 652         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_GUID, &vdev_guid);
 653         nvlist_free(config);
 654
 655         /*
 656          * Check that the label's pool guid matches the desired guid.
 657          * Inactive spares and L2ARCs do not have any pool guid in the label.
 658          */
 659         if (pool_guid != 0 && pool_guid != spa_guid(vd->vdev_spa)) {
 660                 ZFS_LOG(1, "pool guid mismatch for provider %s: %ju != %ju.",
 661                     pp->name,
 662                     (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)pool_guid);
 663                 return (NO_MATCH);
 664         }
 665
 666         /*
 667          * Check that the label's vdev guid matches the desired guid.
 668          * The second condition handles possible race on vdev detach, when
 669          * remaining vdev receives GUID of destroyed top level mirror vdev.
 670          */
 671         if (vdev_guid == vd->vdev_guid) {
 672                 ZFS_LOG(1, "guids match for provider %s.", pp->name);
 673                 return (ZERO_MATCH + nlabels);
 674         } else if (top_guid == vd->vdev_guid && vd == vd->vdev_top) {
 675                 ZFS_LOG(1, "top vdev guid match for provider %s.", pp->name);
 676                 return (TOPGUID_MATCH);
 677         }
 678         ZFS_LOG(1, "vdev guid mismatch for provider %s: %ju != %ju.",
 679             pp->name, (uintmax_t)vd->vdev_guid, (uintmax_t)vdev_guid);
 680         return (NO_MATCH);
 681 }
 682
 683 static struct g_consumer *
 684 vdev_geom_attach_by_guids(vdev_t *vd)
 685 {
 686         struct g_class *mp;
 687         struct g_geom *gp;
 688         struct g_provider *pp, *best_pp;
 689         struct g_consumer *cp;
 690         enum match match, best_match;
 691
 692         g_topology_assert();
 693
 694         cp = NULL;
 695         best_pp = NULL;
 696         best_match = NO_MATCH;
 697         LIST_FOREACH(mp, &g_classes, class) {
 698                 if (mp == &zfs_vdev_class)
 699                         continue;
 700                 LIST_FOREACH(gp, &mp->geom, geom) {
 701                         if (gp->flags & G_GEOM_WITHER)
 702                                 continue;
 703                         LIST_FOREACH(pp, &gp->provider, provider) {
 704                                 match = vdev_attach_ok(vd, pp);
 705                                 if (match > best_match) {
 706                                         best_match = match;
 707                                         best_pp = pp;
 708                                 }
 709                                 if (match == FULL_MATCH)
 710                                         goto out;
 711                         }
 712                 }
 713         }
 714
 715 out:
 716         if (best_pp) {
 717                 cp = vdev_geom_attach(best_pp, vd);
 718                 if (cp == NULL) {
 719                         printf("ZFS WARNING: Unable to attach to %s.\n",
 720                             best_pp->name);
 721                 }
 722         }
 723         return (cp);
 724 }
 725
 726 static struct g_consumer *
 727 vdev_geom_open_by_guids(vdev_t *vd)
 728 {
 729         struct g_consumer *cp;
 730         char *buf;
 731         size_t len;
 732
 733         g_topology_assert();
 734
 735         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 736                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 737         cp = vdev_geom_attach_by_guids(vd);
 738         if (cp != NULL) {
 739                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 740                 buf = kmem_alloc(len, KM_SLEEP);
 741
 742                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 743                 spa_strfree(vd->vdev_path);
 744                 vd->vdev_path = buf;
 745
 746                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 747                     (uintmax_t)spa_guid(vd->vdev_spa),
 748                     (uintmax_t)vd->vdev_guid, cp->provider->name);
 749         } else {
 750                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 751                     (uintmax_t)spa_guid(vd->vdev_spa),
 752                     (uintmax_t)vd->vdev_guid);
 753         }
 754
 755         return (cp);
 756 }
 757
 758 static struct g_consumer *
 759 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 760 {
 761         struct g_provider *pp;
 762         struct g_consumer *cp;
 763
 764         g_topology_assert();
 765
 766         cp = NULL;
 767         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 768         if (pp != NULL) {
 769                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 770                 if (!check_guid || vdev_attach_ok(vd, pp) == FULL_MATCH)
 771                         cp = vdev_geom_attach(pp, vd);
 772         }
 773
 774         return (cp);
 775 }
 776
 777 static int
 778 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 779     uint64_t *logical_ashift, uint64_t *physical_ashift)
 780 {
 781         struct g_provider *pp;
 782         struct g_consumer *cp;
 783         size_t bufsize;
 784         int error;
 785
 786         /* Set the TLS to indicate downstack that we should not access zvols*/
 787         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 788
 789         /*
 790          * We must have a pathname, and it must be absolute.
 791          */
 792         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 793                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 794                 return (EINVAL);
 795         }
 796
 797         /*
 798          * Reopen the device if it's not currently open. Otherwise,
 799          * just update the physical size of the device.
 800          */
 801         if ((cp = vd->vdev_tsd) != NULL) {
 802                 ASSERT(vd->vdev_reopening);
 803                 goto skip_open;
 804         }
 805
 806         DROP_GIANT();
 807         g_topology_lock();
 808         error = 0;
 809
 810         if (vd->vdev_spa->spa_splitting_newspa ||
 811             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 812              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 813              vd->vdev_spa->spa_load_state == SPA_LOAD_CREATE)) {
 814                 /*
 815                  * We are dealing with a vdev that hasn't been previously
 816                  * opened (since boot), and we are not loading an
 817                  * existing pool configuration.  This looks like a
 818                  * vdev add operation to a new or existing pool.
 819                  * Assume the user knows what he/she is doing and find
 820                  * GEOM provider by its name, ignoring GUID mismatches.
 821                  *
 822                  * XXPOLICY: It would be safer to only allow a device
 823                  *           that is unlabeled or labeled but missing
 824                  *           GUID information to be opened in this fashion,
 825                  *           unless we are doing a split, in which case we
 826                  *           should allow any guid.
 827                  */
 828                 cp = vdev_geom_open_by_path(vd, 0);
 829         } else {
 830                 /*
 831                  * Try using the recorded path for this device, but only
 832                  * accept it if its label data contains the expected GUIDs.
 833                  */
 834                 cp = vdev_geom_open_by_path(vd, 1);
 835                 if (cp == NULL) {
 836                         /*
 837                          * The device at vd->vdev_path doesn't have the
 838                          * expected GUIDs. The disks might have merely
 839                          * moved around so try all other GEOM providers
 840                          * to find one with the right GUIDs.
 841                          */
 842                         cp = vdev_geom_open_by_guids(vd);
 843                 }
 844         }
 845
 846         /* Clear the TLS now that tasting is done */
 847         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 848
 849         if (cp == NULL) {
 850                 ZFS_LOG(1, "Vdev %s not found.", vd->vdev_path);
 851                 error = ENOENT;
 852         } else if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 853             !ISP2(cp->provider->sectorsize)) {
 854                 ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 855                     cp->provider->name);
 856
 857                 vdev_geom_close_locked(vd);
 858                 error = EINVAL;
 859                 cp = NULL;
 860         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 861                 int i;
 862
 863                 for (i = 0; i < 5; i++) {
 864                         error = g_access(cp, 0, 1, 0);
 865                         if (error == 0)
 866                                 break;
 867                         g_topology_unlock();
 868                         tsleep(vd, 0, "vdev", hz / 2);
 869                         g_topology_lock();
 870                 }
 871                 if (error != 0) {
 872                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 873                             cp->provider->name, error);
 874                         vdev_geom_close_locked(vd);
 875                         cp = NULL;
 876                 }
 877         }
 878         if (cp != NULL) {
 879                 struct consumer_priv_t *priv;
 880                 struct consumer_vdev_elem *elem;
 881
 882                 priv = (struct consumer_priv_t*)&cp->private;
 883                 if (cp->private == NULL)
 884                         SLIST_INIT(priv);
 885                 elem = g_malloc(sizeof(*elem), M_WAITOK|M_ZERO);
 886                 elem->vd = vd;
 887                 SLIST_INSERT_HEAD(priv, elem, elems);
 888         }
 889
 890         /* Fetch initial physical path information for this device. */
 891         if (cp != NULL) {
 892                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 893
 894                 /* Set other GEOM characteristics */
 895                 vdev_geom_set_physpath(vd, cp, /*do_null_update*/B_FALSE);
 896                 vdev_geom_set_rotation_rate(vd, cp);
 897         }
 898
 899         g_topology_unlock();
 900         PICKUP_GIANT();
 901         if (cp == NULL) {
 902                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 903                 return (error);
 904         }
 905 skip_open:
 906         pp = cp->provider;
 907
 908         /*
 909          * Determine the actual size of the device.
 910          */
 911         *max_psize = *psize = pp->mediasize;
 912
 913         /*
 914          * Determine the device's minimum transfer size and preferred
 915          * transfer size.
 916          */
 917         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 918         *physical_ashift = 0;
 919         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 920             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 921                 *physical_ashift = highbit(pp->stripesize) - 1;
 922
 923         /*
 924          * Clear the nowritecache settings, so that on a vdev_reopen()
 925          * we will try again.
 926          */
 927         vd->vdev_nowritecache = B_FALSE;
 928
 929         return (0);
 930 }
 931
 932 static void
 933 vdev_geom_close(vdev_t *vd)
 934 {
 935
 936         if (vd->vdev_reopening)
 937                 return;
 938
 939         DROP_GIANT();
 940         g_topology_lock();
 941         vdev_geom_close_locked(vd);
 942         g_topology_unlock();
 943         PICKUP_GIANT();
 944 }
 945
 946 static void
 947 vdev_geom_io_intr(struct bio *bp)
 948 {
 949         vdev_t *vd;
 950         zio_t *zio;
 951
 952         zio = bp->bio_caller1;
 953         vd = zio->io_vd;
 954         zio->io_error = bp->bio_error;
 955         if (zio->io_error == 0 && bp->bio_resid != 0)
 956                 zio->io_error = SET_ERROR(EIO);
 957
 958         switch(zio->io_error) {
 959         case ENOTSUP:
 960                 /*
 961                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 962                  * that future attempts will never succeed. In this case
 963                  * we set a persistent flag so that we don't bother with
 964                  * requests in the future.
 965                  */
 966                 switch(bp->bio_cmd) {
 967                 case BIO_FLUSH:
 968                         vd->vdev_nowritecache = B_TRUE;
 969                         break;
 970                 case BIO_DELETE:
 971                         vd->vdev_notrim = B_TRUE;
 972                         break;
 973                 }
 974                 break;
 975         case ENXIO:
 976                 if (!vd->vdev_remove_wanted) {
 977                         /*
 978                          * If provider's error is set we assume it is being
 979                          * removed.
 980                          */
 981                         if (bp->bio_to->error != 0) {
 982                                 vd->vdev_remove_wanted = B_TRUE;
 983                                 spa_async_request(zio->io_spa,
 984                                     SPA_ASYNC_REMOVE);
 985                         } else if (!vd->vdev_delayed_close) {
 986                                 vd->vdev_delayed_close = B_TRUE;
 987                         }
 988                 }
 989                 break;
 990         }
 991
 992         /*
 993          * We have to split bio freeing into two parts, because the ABD code
 994          * cannot be called in this context and vdev_op_io_done is not called
 995          * for ZIO_TYPE_IOCTL zio-s.
 996          */
 997         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
 998                 g_destroy_bio(bp);
 999                 zio->io_bio = NULL;
1000         }
1001         zio_delay_interrupt(zio);
1002 }
1003
1004 static void
1005 vdev_geom_io_start(zio_t *zio)
1006 {
1007         vdev_t *vd;
1008         struct g_consumer *cp;
1009         struct bio *bp;
1010         int error;
1011
1012         vd = zio->io_vd;
1013
1014         switch (zio->io_type) {
1015         case ZIO_TYPE_IOCTL:
1016                 /* XXPOLICY */
1017                 if (!vdev_readable(vd)) {
1018                         zio->io_error = SET_ERROR(ENXIO);
1019                         zio_interrupt(zio);
1020                         return;
1021                 } else {
1022                         switch (zio->io_cmd) {
1023                         case DKIOCFLUSHWRITECACHE:
1024                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
1025                                         break;
1026                                 if (vd->vdev_nowritecache) {
1027                                         zio->io_error = SET_ERROR(ENOTSUP);
1028                                         break;
1029                                 }
1030                                 goto sendreq;
1031                         default:
1032                                 zio->io_error = SET_ERROR(ENOTSUP);
1033                         }
1034                 }
1035
1036                 zio_execute(zio);
1037                 return;
1038         case ZIO_TYPE_FREE:
1039                 if (vd->vdev_notrim) {
1040                         zio->io_error = SET_ERROR(ENOTSUP);
1041                 } else if (!vdev_geom_bio_delete_disable) {
1042                         goto sendreq;
1043                 }
1044                 zio_execute(zio);
1045                 return;
1046         }
1047 sendreq:
1048         ASSERT(zio->io_type == ZIO_TYPE_READ ||
1049             zio->io_type == ZIO_TYPE_WRITE ||
1050             zio->io_type == ZIO_TYPE_FREE ||
1051             zio->io_type == ZIO_TYPE_IOCTL);
1052
1053         cp = vd->vdev_tsd;
1054         if (cp == NULL) {
1055                 zio->io_error = SET_ERROR(ENXIO);
1056                 zio_interrupt(zio);
1057                 return;
1058         }
1059         bp = g_alloc_bio();
1060         bp->bio_caller1 = zio;
1061         switch (zio->io_type) {
1062         case ZIO_TYPE_READ:
1063         case ZIO_TYPE_WRITE:
1064                 zio->io_target_timestamp = zio_handle_io_delay(zio);
1065                 bp->bio_offset = zio->io_offset;
1066                 bp->bio_length = zio->io_size;
1067                 if (zio->io_type == ZIO_TYPE_READ) {
1068                         bp->bio_cmd = BIO_READ;
1069                         bp->bio_data =
1070                             abd_borrow_buf(zio->io_abd, zio->io_size);
1071                 } else {
1072                         bp->bio_cmd = BIO_WRITE;
1073                         bp->bio_data =
1074                             abd_borrow_buf_copy(zio->io_abd, zio->io_size);
1075                 }
1076                 break;
1077         case ZIO_TYPE_FREE:
1078                 bp->bio_cmd = BIO_DELETE;
1079                 bp->bio_data = NULL;
1080                 bp->bio_offset = zio->io_offset;
1081                 bp->bio_length = zio->io_size;
1082                 break;
1083         case ZIO_TYPE_IOCTL:
1084                 bp->bio_cmd = BIO_FLUSH;
1085                 bp->bio_flags |= BIO_ORDERED;
1086                 bp->bio_data = NULL;
1087                 bp->bio_offset = cp->provider->mediasize;
1088                 bp->bio_length = 0;
1089                 break;
1090         }
1091         bp->bio_done = vdev_geom_io_intr;
1092         zio->io_bio = bp;
1093
1094         g_io_request(bp, cp);
1095 }
1096
1097 static void
1098 vdev_geom_io_done(zio_t *zio)
1099 {
1100         struct bio *bp = zio->io_bio;
1101
1102         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
1103                 ASSERT(bp == NULL);
1104                 return;
1105         }
1106
1107         if (bp == NULL) {
1108                 ASSERT3S(zio->io_error, ==, ENXIO);
1109                 return;
1110         }
1111
1112         if (zio->io_type == ZIO_TYPE_READ)
1113                 abd_return_buf_copy(zio->io_abd, bp->bio_data, zio->io_size);
1114         else
1115                 abd_return_buf(zio->io_abd, bp->bio_data, zio->io_size);
1116
1117         g_destroy_bio(bp);
1118         zio->io_bio = NULL;
1119 }
1120
1121 static void
1122 vdev_geom_hold(vdev_t *vd)
1123 {
1124 }
1125
1126 static void
1127 vdev_geom_rele(vdev_t *vd)
1128 {
1129 }
1130
1131 vdev_ops_t vdev_geom_ops = {
1132         vdev_geom_open,
1133         vdev_geom_close,
1134         vdev_default_asize,
1135         vdev_geom_io_start,
1136         vdev_geom_io_done,
1137         NULL,
1138         vdev_geom_hold,
1139         vdev_geom_rele,
1140         VDEV_TYPE_DISK,         /* name of this vdev type */
1141         B_TRUE                  /* leaf vdev */
1142 };