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, boolean_t sanity)
 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 (sanity) {
 207                 if (pp->sectorsize > VDEV_PAD_SIZE || !ISP2(pp->sectorsize)) {
 208                         ZFS_LOG(1, "Failing attach of %s. "
 209                                    "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. "
 214                                    "Incompatible mediasize %ju\n",
 215                             pp->name, pp->mediasize);
 216                         return (NULL);
 217                 }
 218         }
 219
 220         /* Do we have geom already? No? Create one. */
 221         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 222                 if (gp->flags & G_GEOM_WITHER)
 223                         continue;
 224                 if (strcmp(gp->name, "zfs::vdev") != 0)
 225                         continue;
 226                 break;
 227         }
 228         if (gp == NULL) {
 229                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 230                 gp->orphan = vdev_geom_orphan;
 231                 gp->attrchanged = vdev_geom_attrchanged;
 232                 cp = g_new_consumer(gp);
 233                 error = g_attach(cp, pp);
 234                 if (error != 0) {
 235                         ZFS_LOG(1, "%s(%d): g_attach failed: %d\n", __func__,
 236                             __LINE__, error);
 237                         vdev_geom_detach(cp, B_FALSE);
 238                         return (NULL);
 239                 }
 240                 error = g_access(cp, 1, 0, 1);
 241                 if (error != 0) {
 242                         ZFS_LOG(1, "%s(%d): g_access failed: %d\n", __func__,
 243                                __LINE__, error);
 244                         vdev_geom_detach(cp, B_FALSE);
 245                         return (NULL);
 246                 }
 247                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 248         } else {
 249                 /* Check if we are already connected to this provider. */
 250                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 251                         if (cp->provider == pp) {
 252                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 253                                 break;
 254                         }
 255                 }
 256                 if (cp == NULL) {
 257                         cp = g_new_consumer(gp);
 258                         error = g_attach(cp, pp);
 259                         if (error != 0) {
 260                                 ZFS_LOG(1, "%s(%d): g_attach failed: %d\n",
 261                                     __func__, __LINE__, error);
 262                                 vdev_geom_detach(cp, B_FALSE);
 263                                 return (NULL);
 264                         }
 265                         error = g_access(cp, 1, 0, 1);
 266                         if (error != 0) {
 267                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 268                                     __func__, __LINE__, error);
 269                                 vdev_geom_detach(cp, B_FALSE);
 270                                 return (NULL);
 271                         }
 272                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 273                 } else {
 274                         error = g_access(cp, 1, 0, 1);
 275                         if (error != 0) {
 276                                 ZFS_LOG(1, "%s(%d): g_access failed: %d\n",
 277                                     __func__, __LINE__, error);
 278                                 return (NULL);
 279                         }
 280                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 281                 }
 282         }
 283
 284         if (vd != NULL)
 285                 vd->vdev_tsd = cp;
 286
 287         cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
 288         return (cp);
 289 }
 290
 291 static void
 292 vdev_geom_detach(struct g_consumer *cp, boolean_t open_for_read)
 293 {
 294         struct g_geom *gp;
 295
 296         g_topology_assert();
 297
 298         ZFS_LOG(1, "Detaching from %s.",
 299             cp->provider && cp->provider->name ? cp->provider->name : "NULL");
 300
 301         gp = cp->geom;
 302         if (open_for_read)
 303                 g_access(cp, -1, 0, -1);
 304         /* Destroy consumer on last close. */
 305         if (cp->acr == 0 && cp->ace == 0) {
 306                 if (cp->acw > 0)
 307                         g_access(cp, 0, -cp->acw, 0);
 308                 if (cp->provider != NULL) {
 309                         ZFS_LOG(1, "Destroying consumer for %s.",
 310                             cp->provider->name ? cp->provider->name : "NULL");
 311                         g_detach(cp);
 312                 }
 313                 g_destroy_consumer(cp);
 314         }
 315         /* Destroy geom if there are no consumers left. */
 316         if (LIST_EMPTY(&gp->consumer)) {
 317                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 318                 g_wither_geom(gp, ENXIO);
 319         }
 320 }
 321
 322 static void
 323 vdev_geom_close_locked(vdev_t *vd)
 324 {
 325         struct g_consumer *cp;
 326         struct consumer_priv_t *priv;
 327         struct consumer_vdev_elem *elem, *elem_temp;
 328
 329         g_topology_assert();
 330
 331         cp = vd->vdev_tsd;
 332         vd->vdev_delayed_close = B_FALSE;
 333         if (cp == NULL)
 334                 return;
 335
 336         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 337         KASSERT(cp->private != NULL, ("%s: cp->private is NULL", __func__));
 338         priv = (struct consumer_priv_t*)&cp->private;
 339         vd->vdev_tsd = NULL;
 340         SLIST_FOREACH_SAFE(elem, priv, elems, elem_temp) {
 341                 if (elem->vd == vd) {
 342                         SLIST_REMOVE(priv, elem, consumer_vdev_elem, elems);
 343                         g_free(elem);
 344                 }
 345         }
 346
 347         vdev_geom_detach(cp, B_TRUE);
 348 }
 349
 350 /*
 351  * Issue one or more bios to the vdev in parallel
 352  * cmds, datas, offsets, errors, and sizes are arrays of length ncmds.  Each IO
 353  * operation is described by parallel entries from each array.  There may be
 354  * more bios actually issued than entries in the array
 355  */
 356 static void
 357 vdev_geom_io(struct g_consumer *cp, int *cmds, void **datas, off_t *offsets,
 358     off_t *sizes, int *errors, int ncmds)
 359 {
 360         struct bio **bios;
 361         u_char *p;
 362         off_t off, maxio, s, end;
 363         int i, n_bios, j;
 364         size_t bios_size;
 365
 366         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 367         n_bios = 0;
 368
 369         /* How many bios are required for all commands ? */
 370         for (i = 0; i < ncmds; i++)
 371                 n_bios += (sizes[i] + maxio - 1) / maxio;
 372
 373         /* Allocate memory for the bios */
 374         bios_size = n_bios * sizeof(struct bio*);
 375         bios = kmem_zalloc(bios_size, KM_SLEEP);
 376
 377         /* Prepare and issue all of the bios */
 378         for (i = j = 0; i < ncmds; i++) {
 379                 off = offsets[i];
 380                 p = datas[i];
 381                 s = sizes[i];
 382                 end = off + s;
 383                 ASSERT((off % cp->provider->sectorsize) == 0);
 384                 ASSERT((s % cp->provider->sectorsize) == 0);
 385
 386                 for (; off < end; off += maxio, p += maxio, s -= maxio, j++) {
 387                         bios[j] = g_alloc_bio();
 388                         bios[j]->bio_cmd = cmds[i];
 389                         bios[j]->bio_done = NULL;
 390                         bios[j]->bio_offset = off;
 391                         bios[j]->bio_length = MIN(s, maxio);
 392                         bios[j]->bio_data = p;
 393                         g_io_request(bios[j], cp);
 394                 }
 395         }
 396         ASSERT(j == n_bios);
 397
 398         /* Wait for all of the bios to complete, and clean them up */
 399         for (i = j = 0; i < ncmds; i++) {
 400                 off = offsets[i];
 401                 s = sizes[i];
 402                 end = off + s;
 403
 404                 for (; off < end; off += maxio, s -= maxio, j++) {
 405                         errors[i] = biowait(bios[j], "vdev_geom_io") || errors[i];
 406                         g_destroy_bio(bios[j]);
 407                 }
 408         }
 409         kmem_free(bios, bios_size);
 410 }
 411
 412 /*
 413  * Read the vdev config from a device.  Return the number of valid labels that
 414  * were found.  The vdev config will be returned in config if and only if at
 415  * least one valid label was found.
 416  */
 417 static int
 418 vdev_geom_read_config(struct g_consumer *cp, nvlist_t **configp)
 419 {
 420         struct g_provider *pp;
 421         nvlist_t *config;
 422         vdev_phys_t *vdev_lists[VDEV_LABELS];
 423         char *buf;
 424         size_t buflen;
 425         uint64_t psize, state, txg;
 426         off_t offsets[VDEV_LABELS];
 427         off_t size;
 428         off_t sizes[VDEV_LABELS];
 429         int cmds[VDEV_LABELS];
 430         int errors[VDEV_LABELS];
 431         int l, nlabels;
 432
 433         g_topology_assert_not();
 434
 435         pp = cp->provider;
 436         ZFS_LOG(1, "Reading config from %s...", pp->name);
 437
 438         psize = pp->mediasize;
 439         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 440
 441         size = sizeof(*vdev_lists[0]) + pp->sectorsize -
 442             ((sizeof(*vdev_lists[0]) - 1) % pp->sectorsize) - 1;
 443
 444         buflen = sizeof(vdev_lists[0]->vp_nvlist);
 445
 446         /* Create all of the IO requests */
 447         for (l = 0; l < VDEV_LABELS; l++) {
 448                 cmds[l] = BIO_READ;
 449                 vdev_lists[l] = kmem_alloc(size, KM_SLEEP);
 450                 offsets[l] = vdev_label_offset(psize, l, 0) + VDEV_SKIP_SIZE;
 451                 sizes[l] = size;
 452                 errors[l] = 0;
 453                 ASSERT(offsets[l] % pp->sectorsize == 0);
 454         }
 455
 456         /* Issue the IO requests */
 457         vdev_geom_io(cp, cmds, (void**)vdev_lists, offsets, sizes, errors,
 458             VDEV_LABELS);
 459
 460         /* Parse the labels */
 461         config = *configp = NULL;
 462         nlabels = 0;
 463         for (l = 0; l < VDEV_LABELS; l++) {
 464                 if (errors[l] != 0)
 465                         continue;
 466
 467                 buf = vdev_lists[l]->vp_nvlist;
 468
 469                 if (nvlist_unpack(buf, buflen, &config, 0) != 0)
 470                         continue;
 471
 472                 if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE,
 473                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 474                         nvlist_free(config);
 475                         continue;
 476                 }
 477
 478                 if (state != POOL_STATE_SPARE &&
 479                     state != POOL_STATE_L2CACHE &&
 480                     (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG,
 481                     &txg) != 0 || txg == 0)) {
 482                         nvlist_free(config);
 483                         continue;
 484                 }
 485
 486                 if (*configp != NULL)
 487                         nvlist_free(*configp);
 488                 *configp = config;
 489
 490                 nlabels++;
 491         }
 492
 493         /* Free the label storage */
 494         for (l = 0; l < VDEV_LABELS; l++)
 495                 kmem_free(vdev_lists[l], size);
 496
 497         return (nlabels);
 498 }
 499
 500 static void
 501 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 502 {
 503         nvlist_t **new_configs;
 504         uint64_t i;
 505
 506         if (id < *count)
 507                 return;
 508         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 509             KM_SLEEP);
 510         for (i = 0; i < *count; i++)
 511                 new_configs[i] = (*configs)[i];
 512         if (*configs != NULL)
 513                 kmem_free(*configs, *count * sizeof(void *));
 514         *configs = new_configs;
 515         *count = id + 1;
 516 }
 517
 518 static void
 519 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 520     const char *name, uint64_t* known_pool_guid)
 521 {
 522         nvlist_t *vdev_tree;
 523         uint64_t pool_guid;
 524         uint64_t vdev_guid, known_guid;
 525         uint64_t id, txg, known_txg;
 526         char *pname;
 527         int i;
 528
 529         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 530             strcmp(pname, name) != 0)
 531                 goto ignore;
 532
 533         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 534                 goto ignore;
 535
 536         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 537                 goto ignore;
 538
 539         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 540                 goto ignore;
 541
 542         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 543                 goto ignore;
 544
 545         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 546
 547         if (*known_pool_guid != 0) {
 548                 if (pool_guid != *known_pool_guid)
 549                         goto ignore;
 550         } else
 551                 *known_pool_guid = pool_guid;
 552
 553         resize_configs(configs, count, id);
 554
 555         if ((*configs)[id] != NULL) {
 556                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 557                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 558                 if (txg <= known_txg)
 559                         goto ignore;
 560                 nvlist_free((*configs)[id]);
 561         }
 562
 563         (*configs)[id] = cfg;
 564         return;
 565
 566 ignore:
 567         nvlist_free(cfg);
 568 }
 569
 570 int
 571 vdev_geom_read_pool_label(const char *name,
 572     nvlist_t ***configs, uint64_t *count)
 573 {
 574         struct g_class *mp;
 575         struct g_geom *gp;
 576         struct g_provider *pp;
 577         struct g_consumer *zcp;
 578         nvlist_t *vdev_cfg;
 579         uint64_t pool_guid;
 580         int error, nlabels;
 581
 582         DROP_GIANT();
 583         g_topology_lock();
 584
 585         *configs = NULL;
 586         *count = 0;
 587         pool_guid = 0;
 588         LIST_FOREACH(mp, &g_classes, class) {
 589                 if (mp == &zfs_vdev_class)
 590                         continue;
 591                 LIST_FOREACH(gp, &mp->geom, geom) {
 592                         if (gp->flags & G_GEOM_WITHER)
 593                                 continue;
 594                         LIST_FOREACH(pp, &gp->provider, provider) {
 595                                 if (pp->flags & G_PF_WITHER)
 596                                         continue;
 597                                 zcp = vdev_geom_attach(pp, NULL, B_TRUE);
 598                                 if (zcp == NULL)
 599                                         continue;
 600                                 g_topology_unlock();
 601                                 nlabels = vdev_geom_read_config(zcp, &vdev_cfg);
 602                                 g_topology_lock();
 603                                 vdev_geom_detach(zcp, B_TRUE);
 604                                 if (nlabels == 0)
 605                                         continue;
 606                                 ZFS_LOG(1, "successfully read vdev config");
 607
 608                                 process_vdev_config(configs, count,
 609                                     vdev_cfg, name, &pool_guid);
 610                         }
 611                 }
 612         }
 613         g_topology_unlock();
 614         PICKUP_GIANT();
 615
 616         return (*count > 0 ? 0 : ENOENT);
 617 }
 618
 619 enum match {
 620         NO_MATCH = 0,           /* No matching labels found */
 621         TOPGUID_MATCH = 1,      /* Labels match top guid, not vdev guid*/
 622         ZERO_MATCH = 1,         /* Should never be returned */
 623         ONE_MATCH = 2,          /* 1 label matching the vdev_guid */
 624         TWO_MATCH = 3,          /* 2 label matching the vdev_guid */
 625         THREE_MATCH = 4,        /* 3 label matching the vdev_guid */
 626         FULL_MATCH = 5          /* all labels match the vdev_guid */
 627 };
 628
 629 static enum match
 630 vdev_attach_ok(vdev_t *vd, struct g_provider *pp)
 631 {
 632         nvlist_t *config;
 633         uint64_t pool_guid, top_guid, vdev_guid;
 634         struct g_consumer *cp;
 635         int nlabels;
 636
 637         cp = vdev_geom_attach(pp, NULL, B_TRUE);
 638         if (cp == NULL) {
 639                 ZFS_LOG(1, "Unable to attach tasting instance to %s.",
 640                     pp->name);
 641                 return (NO_MATCH);
 642         }
 643         g_topology_unlock();
 644         nlabels = vdev_geom_read_config(cp, &config);
 645         g_topology_lock();
 646         vdev_geom_detach(cp, B_TRUE);
 647         if (nlabels == 0) {
 648                 ZFS_LOG(1, "Unable to read config from %s.", pp->name);
 649                 return (NO_MATCH);
 650         }
 651
 652         pool_guid = 0;
 653         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid);
 654         top_guid = 0;
 655         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TOP_GUID, &top_guid);
 656         vdev_guid = 0;
 657         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_GUID, &vdev_guid);
 658         nvlist_free(config);
 659
 660         /*
 661          * Check that the label's pool guid matches the desired guid.
 662          * Inactive spares and L2ARCs do not have any pool guid in the label.
 663          */
 664         if (pool_guid != 0 && pool_guid != spa_guid(vd->vdev_spa)) {
 665                 ZFS_LOG(1, "pool guid mismatch for provider %s: %ju != %ju.",
 666                     pp->name,
 667                     (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)pool_guid);
 668                 return (NO_MATCH);
 669         }
 670
 671         /*
 672          * Check that the label's vdev guid matches the desired guid.
 673          * The second condition handles possible race on vdev detach, when
 674          * remaining vdev receives GUID of destroyed top level mirror vdev.
 675          */
 676         if (vdev_guid == vd->vdev_guid) {
 677                 ZFS_LOG(1, "guids match for provider %s.", pp->name);
 678                 return (ZERO_MATCH + nlabels);
 679         } else if (top_guid == vd->vdev_guid && vd == vd->vdev_top) {
 680                 ZFS_LOG(1, "top vdev guid match for provider %s.", pp->name);
 681                 return (TOPGUID_MATCH);
 682         }
 683         ZFS_LOG(1, "vdev guid mismatch for provider %s: %ju != %ju.",
 684             pp->name, (uintmax_t)vd->vdev_guid, (uintmax_t)vdev_guid);
 685         return (NO_MATCH);
 686 }
 687
 688 static struct g_consumer *
 689 vdev_geom_attach_by_guids(vdev_t *vd)
 690 {
 691         struct g_class *mp;
 692         struct g_geom *gp;
 693         struct g_provider *pp, *best_pp;
 694         struct g_consumer *cp;
 695         enum match match, best_match;
 696
 697         g_topology_assert();
 698
 699         cp = NULL;
 700         best_pp = NULL;
 701         best_match = NO_MATCH;
 702         LIST_FOREACH(mp, &g_classes, class) {
 703                 if (mp == &zfs_vdev_class)
 704                         continue;
 705                 LIST_FOREACH(gp, &mp->geom, geom) {
 706                         if (gp->flags & G_GEOM_WITHER)
 707                                 continue;
 708                         LIST_FOREACH(pp, &gp->provider, provider) {
 709                                 match = vdev_attach_ok(vd, pp);
 710                                 if (match > best_match) {
 711                                         best_match = match;
 712                                         best_pp = pp;
 713                                 }
 714                                 if (match == FULL_MATCH)
 715                                         goto out;
 716                         }
 717                 }
 718         }
 719
 720 out:
 721         if (best_pp) {
 722                 cp = vdev_geom_attach(best_pp, vd, B_TRUE);
 723                 if (cp == NULL) {
 724                         printf("ZFS WARNING: Unable to attach to %s.\n",
 725                             best_pp->name);
 726                 }
 727         }
 728         return (cp);
 729 }
 730
 731 static struct g_consumer *
 732 vdev_geom_open_by_guids(vdev_t *vd)
 733 {
 734         struct g_consumer *cp;
 735         char *buf;
 736         size_t len;
 737
 738         g_topology_assert();
 739
 740         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 741                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 742         cp = vdev_geom_attach_by_guids(vd);
 743         if (cp != NULL) {
 744                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 745                 buf = kmem_alloc(len, KM_SLEEP);
 746
 747                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 748                 spa_strfree(vd->vdev_path);
 749                 vd->vdev_path = buf;
 750
 751                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 752                     (uintmax_t)spa_guid(vd->vdev_spa),
 753                     (uintmax_t)vd->vdev_guid, cp->provider->name);
 754         } else {
 755                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 756                     (uintmax_t)spa_guid(vd->vdev_spa),
 757                     (uintmax_t)vd->vdev_guid);
 758         }
 759
 760         return (cp);
 761 }
 762
 763 static struct g_consumer *
 764 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 765 {
 766         struct g_provider *pp;
 767         struct g_consumer *cp;
 768
 769         g_topology_assert();
 770
 771         cp = NULL;
 772         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 773         if (pp != NULL) {
 774                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 775                 if (!check_guid || vdev_attach_ok(vd, pp) == FULL_MATCH)
 776                         cp = vdev_geom_attach(pp, vd, B_FALSE);
 777         }
 778
 779         return (cp);
 780 }
 781
 782 static int
 783 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 784     uint64_t *logical_ashift, uint64_t *physical_ashift)
 785 {
 786         struct g_provider *pp;
 787         struct g_consumer *cp;
 788         size_t bufsize;
 789         int error;
 790
 791         /* Set the TLS to indicate downstack that we should not access zvols*/
 792         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 793
 794         /*
 795          * We must have a pathname, and it must be absolute.
 796          */
 797         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 798                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 799                 return (EINVAL);
 800         }
 801
 802         /*
 803          * Reopen the device if it's not currently open. Otherwise,
 804          * just update the physical size of the device.
 805          */
 806         if ((cp = vd->vdev_tsd) != NULL) {
 807                 ASSERT(vd->vdev_reopening);
 808                 goto skip_open;
 809         }
 810
 811         DROP_GIANT();
 812         g_topology_lock();
 813         error = 0;
 814
 815         if (vd->vdev_spa->spa_splitting_newspa ||
 816             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 817              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 818              vd->vdev_spa->spa_load_state == SPA_LOAD_CREATE)) {
 819                 /*
 820                  * We are dealing with a vdev that hasn't been previously
 821                  * opened (since boot), and we are not loading an
 822                  * existing pool configuration.  This looks like a
 823                  * vdev add operation to a new or existing pool.
 824                  * Assume the user knows what he/she is doing and find
 825                  * GEOM provider by its name, ignoring GUID mismatches.
 826                  *
 827                  * XXPOLICY: It would be safer to only allow a device
 828                  *           that is unlabeled or labeled but missing
 829                  *           GUID information to be opened in this fashion,
 830                  *           unless we are doing a split, in which case we
 831                  *           should allow any guid.
 832                  */
 833                 cp = vdev_geom_open_by_path(vd, 0);
 834         } else {
 835                 /*
 836                  * Try using the recorded path for this device, but only
 837                  * accept it if its label data contains the expected GUIDs.
 838                  */
 839                 cp = vdev_geom_open_by_path(vd, 1);
 840                 if (cp == NULL) {
 841                         /*
 842                          * The device at vd->vdev_path doesn't have the
 843                          * expected GUIDs. The disks might have merely
 844                          * moved around so try all other GEOM providers
 845                          * to find one with the right GUIDs.
 846                          */
 847                         cp = vdev_geom_open_by_guids(vd);
 848                 }
 849         }
 850
 851         /* Clear the TLS now that tasting is done */
 852         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 853
 854         if (cp == NULL) {
 855                 ZFS_LOG(1, "Vdev %s not found.", vd->vdev_path);
 856                 error = ENOENT;
 857         } else {
 858                 struct consumer_priv_t *priv;
 859                 struct consumer_vdev_elem *elem;
 860                 int spamode;
 861
 862                 priv = (struct consumer_priv_t*)&cp->private;
 863                 if (cp->private == NULL)
 864                         SLIST_INIT(priv);
 865                 elem = g_malloc(sizeof(*elem), M_WAITOK|M_ZERO);
 866                 elem->vd = vd;
 867                 SLIST_INSERT_HEAD(priv, elem, elems);
 868
 869                 spamode = spa_mode(vd->vdev_spa);
 870                 if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 871                     !ISP2(cp->provider->sectorsize)) {
 872                         ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 873                             cp->provider->name);
 874
 875                         vdev_geom_close_locked(vd);
 876                         error = EINVAL;
 877                         cp = NULL;
 878                 } else if (cp->acw == 0 && (spamode & FWRITE) != 0) {
 879                         int i;
 880
 881                         for (i = 0; i < 5; i++) {
 882                                 error = g_access(cp, 0, 1, 0);
 883                                 if (error == 0)
 884                                         break;
 885                                 g_topology_unlock();
 886                                 tsleep(vd, 0, "vdev", hz / 2);
 887                                 g_topology_lock();
 888                         }
 889                         if (error != 0) {
 890                                 printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 891                                     cp->provider->name, error);
 892                                 vdev_geom_close_locked(vd);
 893                                 cp = NULL;
 894                         }
 895                 }
 896         }
 897
 898         /* Fetch initial physical path information for this device. */
 899         if (cp != NULL) {
 900                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 901
 902                 /* Set other GEOM characteristics */
 903                 vdev_geom_set_physpath(vd, cp, /*do_null_update*/B_FALSE);
 904                 vdev_geom_set_rotation_rate(vd, cp);
 905         }
 906
 907         g_topology_unlock();
 908         PICKUP_GIANT();
 909         if (cp == NULL) {
 910                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 911                 vdev_dbgmsg(vd, "vdev_geom_open: failed to open [error=%d]",
 912                     error);
 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
1007         /*
1008          * We have to split bio freeing into two parts, because the ABD code
1009          * cannot be called in this context and vdev_op_io_done is not called
1010          * for ZIO_TYPE_IOCTL zio-s.
1011          */
1012         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
1013                 g_destroy_bio(bp);
1014                 zio->io_bio = NULL;
1015         }
1016         zio_delay_interrupt(zio);
1017 }
1018
1019 static void
1020 vdev_geom_io_start(zio_t *zio)
1021 {
1022         vdev_t *vd;
1023         struct g_consumer *cp;
1024         struct bio *bp;
1025         int error;
1026
1027         vd = zio->io_vd;
1028
1029         switch (zio->io_type) {
1030         case ZIO_TYPE_IOCTL:
1031                 /* XXPOLICY */
1032                 if (!vdev_readable(vd)) {
1033                         zio->io_error = SET_ERROR(ENXIO);
1034                         zio_interrupt(zio);
1035                         return;
1036                 } else {
1037                         switch (zio->io_cmd) {
1038                         case DKIOCFLUSHWRITECACHE:
1039                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
1040                                         break;
1041                                 if (vd->vdev_nowritecache) {
1042                                         zio->io_error = SET_ERROR(ENOTSUP);
1043                                         break;
1044                                 }
1045                                 goto sendreq;
1046                         default:
1047                                 zio->io_error = SET_ERROR(ENOTSUP);
1048                         }
1049                 }
1050
1051                 zio_execute(zio);
1052                 return;
1053         case ZIO_TYPE_FREE:
1054                 if (vd->vdev_notrim) {
1055                         zio->io_error = SET_ERROR(ENOTSUP);
1056                 } else if (!vdev_geom_bio_delete_disable) {
1057                         goto sendreq;
1058                 }
1059                 zio_execute(zio);
1060                 return;
1061         }
1062 sendreq:
1063         ASSERT(zio->io_type == ZIO_TYPE_READ ||
1064             zio->io_type == ZIO_TYPE_WRITE ||
1065             zio->io_type == ZIO_TYPE_FREE ||
1066             zio->io_type == ZIO_TYPE_IOCTL);
1067
1068         cp = vd->vdev_tsd;
1069         if (cp == NULL) {
1070                 zio->io_error = SET_ERROR(ENXIO);
1071                 zio_interrupt(zio);
1072                 return;
1073         }
1074         bp = g_alloc_bio();
1075         bp->bio_caller1 = zio;
1076         switch (zio->io_type) {
1077         case ZIO_TYPE_READ:
1078         case ZIO_TYPE_WRITE:
1079                 zio->io_target_timestamp = zio_handle_io_delay(zio);
1080                 bp->bio_offset = zio->io_offset;
1081                 bp->bio_length = zio->io_size;
1082                 if (zio->io_type == ZIO_TYPE_READ) {
1083                         bp->bio_cmd = BIO_READ;
1084                         bp->bio_data =
1085                             abd_borrow_buf(zio->io_abd, zio->io_size);
1086                 } else {
1087                         bp->bio_cmd = BIO_WRITE;
1088                         bp->bio_data =
1089                             abd_borrow_buf_copy(zio->io_abd, zio->io_size);
1090                 }
1091                 break;
1092         case ZIO_TYPE_FREE:
1093                 bp->bio_cmd = BIO_DELETE;
1094                 bp->bio_data = NULL;
1095                 bp->bio_offset = zio->io_offset;
1096                 bp->bio_length = zio->io_size;
1097                 break;
1098         case ZIO_TYPE_IOCTL:
1099                 bp->bio_cmd = BIO_FLUSH;
1100                 bp->bio_flags |= BIO_ORDERED;
1101                 bp->bio_data = NULL;
1102                 bp->bio_offset = cp->provider->mediasize;
1103                 bp->bio_length = 0;
1104                 break;
1105         }
1106         bp->bio_done = vdev_geom_io_intr;
1107         zio->io_bio = bp;
1108
1109         g_io_request(bp, cp);
1110 }
1111
1112 static void
1113 vdev_geom_io_done(zio_t *zio)
1114 {
1115         struct bio *bp = zio->io_bio;
1116
1117         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
1118                 ASSERT(bp == NULL);
1119                 return;
1120         }
1121
1122         if (bp == NULL) {
1123                 ASSERT3S(zio->io_error, ==, ENXIO);
1124                 return;
1125         }
1126
1127         if (zio->io_type == ZIO_TYPE_READ)
1128                 abd_return_buf_copy(zio->io_abd, bp->bio_data, zio->io_size);
1129         else
1130                 abd_return_buf(zio->io_abd, bp->bio_data, zio->io_size);
1131
1132         g_destroy_bio(bp);
1133         zio->io_bio = NULL;
1134 }
1135
1136 static void
1137 vdev_geom_hold(vdev_t *vd)
1138 {
1139 }
1140
1141 static void
1142 vdev_geom_rele(vdev_t *vd)
1143 {
1144 }
1145
1146 vdev_ops_t vdev_geom_ops = {
1147         vdev_geom_open,
1148         vdev_geom_close,
1149         vdev_default_asize,
1150         vdev_geom_io_start,
1151         vdev_geom_io_done,
1152         NULL,
1153         NULL,
1154         vdev_geom_hold,
1155         vdev_geom_rele,
1156         NULL,
1157         vdev_default_xlate,
1158         VDEV_TYPE_DISK,         /* name of this vdev type */
1159         B_TRUE                  /* leaf vdev */
1160 };