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         const char *vdpath;
 696         enum match match, best_match;
 697
 698         g_topology_assert();
 699
 700         vdpath = vd->vdev_path + sizeof("/dev/") - 1;
 701         cp = NULL;
 702         best_pp = NULL;
 703         best_match = NO_MATCH;
 704         LIST_FOREACH(mp, &g_classes, class) {
 705                 if (mp == &zfs_vdev_class)
 706                         continue;
 707                 LIST_FOREACH(gp, &mp->geom, geom) {
 708                         if (gp->flags & G_GEOM_WITHER)
 709                                 continue;
 710                         LIST_FOREACH(pp, &gp->provider, provider) {
 711                                 match = vdev_attach_ok(vd, pp);
 712                                 if (match > best_match) {
 713                                         best_match = match;
 714                                         best_pp = pp;
 715                                 } else if (match == best_match) {
 716                                         if (strcmp(pp->name, vdpath) == 0) {
 717                                                 best_pp = pp;
 718                                         }
 719                                 }
 720                                 if (match == FULL_MATCH)
 721                                         goto out;
 722                         }
 723                 }
 724         }
 725
 726 out:
 727         if (best_pp) {
 728                 cp = vdev_geom_attach(best_pp, vd, B_TRUE);
 729                 if (cp == NULL) {
 730                         printf("ZFS WARNING: Unable to attach to %s.\n",
 731                             best_pp->name);
 732                 }
 733         }
 734         return (cp);
 735 }
 736
 737 static struct g_consumer *
 738 vdev_geom_open_by_guids(vdev_t *vd)
 739 {
 740         struct g_consumer *cp;
 741         char *buf;
 742         size_t len;
 743
 744         g_topology_assert();
 745
 746         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 747                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 748         cp = vdev_geom_attach_by_guids(vd);
 749         if (cp != NULL) {
 750                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 751                 buf = kmem_alloc(len, KM_SLEEP);
 752
 753                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 754                 spa_strfree(vd->vdev_path);
 755                 vd->vdev_path = buf;
 756
 757                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 758                     (uintmax_t)spa_guid(vd->vdev_spa),
 759                     (uintmax_t)vd->vdev_guid, cp->provider->name);
 760         } else {
 761                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 762                     (uintmax_t)spa_guid(vd->vdev_spa),
 763                     (uintmax_t)vd->vdev_guid);
 764         }
 765
 766         return (cp);
 767 }
 768
 769 static struct g_consumer *
 770 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 771 {
 772         struct g_provider *pp;
 773         struct g_consumer *cp;
 774
 775         g_topology_assert();
 776
 777         cp = NULL;
 778         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 779         if (pp != NULL) {
 780                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 781                 if (!check_guid || vdev_attach_ok(vd, pp) == FULL_MATCH)
 782                         cp = vdev_geom_attach(pp, vd, B_FALSE);
 783         }
 784
 785         return (cp);
 786 }
 787
 788 static int
 789 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 790     uint64_t *logical_ashift, uint64_t *physical_ashift)
 791 {
 792         struct g_provider *pp;
 793         struct g_consumer *cp;
 794         size_t bufsize;
 795         int error;
 796
 797         /* Set the TLS to indicate downstack that we should not access zvols*/
 798         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 799
 800         /*
 801          * We must have a pathname, and it must be absolute.
 802          */
 803         if (vd->vdev_path == NULL || strncmp(vd->vdev_path, "/dev/", 5) != 0) {
 804                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 805                 return (EINVAL);
 806         }
 807
 808         /*
 809          * Reopen the device if it's not currently open. Otherwise,
 810          * just update the physical size of the device.
 811          */
 812         if ((cp = vd->vdev_tsd) != NULL) {
 813                 ASSERT(vd->vdev_reopening);
 814                 goto skip_open;
 815         }
 816
 817         DROP_GIANT();
 818         g_topology_lock();
 819         error = 0;
 820
 821         if (vd->vdev_spa->spa_splitting_newspa ||
 822             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 823              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 824              vd->vdev_spa->spa_load_state == SPA_LOAD_CREATE)) {
 825                 /*
 826                  * We are dealing with a vdev that hasn't been previously
 827                  * opened (since boot), and we are not loading an
 828                  * existing pool configuration.  This looks like a
 829                  * vdev add operation to a new or existing pool.
 830                  * Assume the user knows what he/she is doing and find
 831                  * GEOM provider by its name, ignoring GUID mismatches.
 832                  *
 833                  * XXPOLICY: It would be safer to only allow a device
 834                  *           that is unlabeled or labeled but missing
 835                  *           GUID information to be opened in this fashion,
 836                  *           unless we are doing a split, in which case we
 837                  *           should allow any guid.
 838                  */
 839                 cp = vdev_geom_open_by_path(vd, 0);
 840         } else {
 841                 /*
 842                  * Try using the recorded path for this device, but only
 843                  * accept it if its label data contains the expected GUIDs.
 844                  */
 845                 cp = vdev_geom_open_by_path(vd, 1);
 846                 if (cp == NULL) {
 847                         /*
 848                          * The device at vd->vdev_path doesn't have the
 849                          * expected GUIDs. The disks might have merely
 850                          * moved around so try all other GEOM providers
 851                          * to find one with the right GUIDs.
 852                          */
 853                         cp = vdev_geom_open_by_guids(vd);
 854                 }
 855         }
 856
 857         /* Clear the TLS now that tasting is done */
 858         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 859
 860         if (cp == NULL) {
 861                 ZFS_LOG(1, "Vdev %s not found.", vd->vdev_path);
 862                 error = ENOENT;
 863         } else {
 864                 struct consumer_priv_t *priv;
 865                 struct consumer_vdev_elem *elem;
 866                 int spamode;
 867
 868                 priv = (struct consumer_priv_t*)&cp->private;
 869                 if (cp->private == NULL)
 870                         SLIST_INIT(priv);
 871                 elem = g_malloc(sizeof(*elem), M_WAITOK|M_ZERO);
 872                 elem->vd = vd;
 873                 SLIST_INSERT_HEAD(priv, elem, elems);
 874
 875                 spamode = spa_mode(vd->vdev_spa);
 876                 if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 877                     !ISP2(cp->provider->sectorsize)) {
 878                         ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 879                             cp->provider->name);
 880
 881                         vdev_geom_close_locked(vd);
 882                         error = EINVAL;
 883                         cp = NULL;
 884                 } else if (cp->acw == 0 && (spamode & FWRITE) != 0) {
 885                         int i;
 886
 887                         for (i = 0; i < 5; i++) {
 888                                 error = g_access(cp, 0, 1, 0);
 889                                 if (error == 0)
 890                                         break;
 891                                 g_topology_unlock();
 892                                 tsleep(vd, 0, "vdev", hz / 2);
 893                                 g_topology_lock();
 894                         }
 895                         if (error != 0) {
 896                                 printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 897                                     cp->provider->name, error);
 898                                 vdev_geom_close_locked(vd);
 899                                 cp = NULL;
 900                         }
 901                 }
 902         }
 903
 904         /* Fetch initial physical path information for this device. */
 905         if (cp != NULL) {
 906                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 907
 908                 /* Set other GEOM characteristics */
 909                 vdev_geom_set_physpath(vd, cp, /*do_null_update*/B_FALSE);
 910                 vdev_geom_set_rotation_rate(vd, cp);
 911         }
 912
 913         g_topology_unlock();
 914         PICKUP_GIANT();
 915         if (cp == NULL) {
 916                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 917                 vdev_dbgmsg(vd, "vdev_geom_open: failed to open [error=%d]",
 918                     error);
 919                 return (error);
 920         }
 921 skip_open:
 922         pp = cp->provider;
 923
 924         /*
 925          * Determine the actual size of the device.
 926          */
 927         *max_psize = *psize = pp->mediasize;
 928
 929         /*
 930          * Determine the device's minimum transfer size and preferred
 931          * transfer size.
 932          */
 933         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 934         *physical_ashift = 0;
 935         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 936             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 937                 *physical_ashift = highbit(pp->stripesize) - 1;
 938
 939         /*
 940          * Clear the nowritecache settings, so that on a vdev_reopen()
 941          * we will try again.
 942          */
 943         vd->vdev_nowritecache = B_FALSE;
 944
 945         return (0);
 946 }
 947
 948 static void
 949 vdev_geom_close(vdev_t *vd)
 950 {
 951         struct g_consumer *cp;
 952
 953         cp = vd->vdev_tsd;
 954
 955         DROP_GIANT();
 956         g_topology_lock();
 957
 958         if (!vd->vdev_reopening ||
 959             (cp != NULL && ((cp->flags & G_CF_ORPHAN) != 0 ||
 960             (cp->provider != NULL && cp->provider->error != 0))))
 961                 vdev_geom_close_locked(vd);
 962
 963         g_topology_unlock();
 964         PICKUP_GIANT();
 965 }
 966
 967 static void
 968 vdev_geom_io_intr(struct bio *bp)
 969 {
 970         vdev_t *vd;
 971         zio_t *zio;
 972
 973         zio = bp->bio_caller1;
 974         vd = zio->io_vd;
 975         zio->io_error = bp->bio_error;
 976         if (zio->io_error == 0 && bp->bio_resid != 0)
 977                 zio->io_error = SET_ERROR(EIO);
 978
 979         switch(zio->io_error) {
 980         case ENOTSUP:
 981                 /*
 982                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 983                  * that future attempts will never succeed. In this case
 984                  * we set a persistent flag so that we don't bother with
 985                  * requests in the future.
 986                  */
 987                 switch(bp->bio_cmd) {
 988                 case BIO_FLUSH:
 989                         vd->vdev_nowritecache = B_TRUE;
 990                         break;
 991                 case BIO_DELETE:
 992                         vd->vdev_notrim = B_TRUE;
 993                         break;
 994                 }
 995                 break;
 996         case ENXIO:
 997                 if (!vd->vdev_remove_wanted) {
 998                         /*
 999                          * If provider's error is set we assume it is being
1000                          * removed.
1001                          */
1002                         if (bp->bio_to->error != 0) {
1003                                 vd->vdev_remove_wanted = B_TRUE;
1004                                 spa_async_request(zio->io_spa,
1005                                     SPA_ASYNC_REMOVE);
1006                         } else if (!vd->vdev_delayed_close) {
1007                                 vd->vdev_delayed_close = B_TRUE;
1008                         }
1009                 }
1010                 break;
1011         }
1012
1013         /*
1014          * We have to split bio freeing into two parts, because the ABD code
1015          * cannot be called in this context and vdev_op_io_done is not called
1016          * for ZIO_TYPE_IOCTL zio-s.
1017          */
1018         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
1019                 g_destroy_bio(bp);
1020                 zio->io_bio = NULL;
1021         }
1022         zio_delay_interrupt(zio);
1023 }
1024
1025 static void
1026 vdev_geom_io_start(zio_t *zio)
1027 {
1028         vdev_t *vd;
1029         struct g_consumer *cp;
1030         struct bio *bp;
1031         int error;
1032
1033         vd = zio->io_vd;
1034
1035         switch (zio->io_type) {
1036         case ZIO_TYPE_IOCTL:
1037                 /* XXPOLICY */
1038                 if (!vdev_readable(vd)) {
1039                         zio->io_error = SET_ERROR(ENXIO);
1040                         zio_interrupt(zio);
1041                         return;
1042                 } else {
1043                         switch (zio->io_cmd) {
1044                         case DKIOCFLUSHWRITECACHE:
1045                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
1046                                         break;
1047                                 if (vd->vdev_nowritecache) {
1048                                         zio->io_error = SET_ERROR(ENOTSUP);
1049                                         break;
1050                                 }
1051                                 goto sendreq;
1052                         default:
1053                                 zio->io_error = SET_ERROR(ENOTSUP);
1054                         }
1055                 }
1056
1057                 zio_execute(zio);
1058                 return;
1059         case ZIO_TYPE_FREE:
1060                 if (vd->vdev_notrim) {
1061                         zio->io_error = SET_ERROR(ENOTSUP);
1062                 } else if (!vdev_geom_bio_delete_disable) {
1063                         goto sendreq;
1064                 }
1065                 zio_execute(zio);
1066                 return;
1067         }
1068 sendreq:
1069         ASSERT(zio->io_type == ZIO_TYPE_READ ||
1070             zio->io_type == ZIO_TYPE_WRITE ||
1071             zio->io_type == ZIO_TYPE_FREE ||
1072             zio->io_type == ZIO_TYPE_IOCTL);
1073
1074         cp = vd->vdev_tsd;
1075         if (cp == NULL) {
1076                 zio->io_error = SET_ERROR(ENXIO);
1077                 zio_interrupt(zio);
1078                 return;
1079         }
1080         bp = g_alloc_bio();
1081         bp->bio_caller1 = zio;
1082         switch (zio->io_type) {
1083         case ZIO_TYPE_READ:
1084         case ZIO_TYPE_WRITE:
1085                 zio->io_target_timestamp = zio_handle_io_delay(zio);
1086                 bp->bio_offset = zio->io_offset;
1087                 bp->bio_length = zio->io_size;
1088                 if (zio->io_type == ZIO_TYPE_READ) {
1089                         bp->bio_cmd = BIO_READ;
1090                         bp->bio_data =
1091                             abd_borrow_buf(zio->io_abd, zio->io_size);
1092                 } else {
1093                         bp->bio_cmd = BIO_WRITE;
1094                         bp->bio_data =
1095                             abd_borrow_buf_copy(zio->io_abd, zio->io_size);
1096                 }
1097                 break;
1098         case ZIO_TYPE_FREE:
1099                 bp->bio_cmd = BIO_DELETE;
1100                 bp->bio_data = NULL;
1101                 bp->bio_offset = zio->io_offset;
1102                 bp->bio_length = zio->io_size;
1103                 break;
1104         case ZIO_TYPE_IOCTL:
1105                 bp->bio_cmd = BIO_FLUSH;
1106                 bp->bio_data = NULL;
1107                 bp->bio_offset = cp->provider->mediasize;
1108                 bp->bio_length = 0;
1109                 break;
1110         }
1111         bp->bio_done = vdev_geom_io_intr;
1112         zio->io_bio = bp;
1113
1114         g_io_request(bp, cp);
1115 }
1116
1117 static void
1118 vdev_geom_io_done(zio_t *zio)
1119 {
1120         struct bio *bp = zio->io_bio;
1121
1122         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
1123                 ASSERT(bp == NULL);
1124                 return;
1125         }
1126
1127         if (bp == NULL) {
1128                 ASSERT3S(zio->io_error, ==, ENXIO);
1129                 return;
1130         }
1131
1132         if (zio->io_type == ZIO_TYPE_READ)
1133                 abd_return_buf_copy(zio->io_abd, bp->bio_data, zio->io_size);
1134         else
1135                 abd_return_buf(zio->io_abd, bp->bio_data, zio->io_size);
1136
1137         g_destroy_bio(bp);
1138         zio->io_bio = NULL;
1139 }
1140
1141 static void
1142 vdev_geom_hold(vdev_t *vd)
1143 {
1144 }
1145
1146 static void
1147 vdev_geom_rele(vdev_t *vd)
1148 {
1149 }
1150
1151 vdev_ops_t vdev_geom_ops = {
1152         vdev_geom_open,
1153         vdev_geom_close,
1154         vdev_default_asize,
1155         vdev_geom_io_start,
1156         vdev_geom_io_done,
1157         NULL,
1158         NULL,
1159         vdev_geom_hold,
1160         vdev_geom_rele,
1161         NULL,
1162         vdev_default_xlate,
1163         VDEV_TYPE_DISK,         /* name of this vdev type */
1164         B_TRUE                  /* leaf vdev */
1165 };