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 **config)
 419 {
 420         struct g_provider *pp;
 421         vdev_phys_t *vdev_lists[VDEV_LABELS];
 422         char *buf;
 423         size_t buflen;
 424         uint64_t psize, state, txg;
 425         off_t offsets[VDEV_LABELS];
 426         off_t size;
 427         off_t sizes[VDEV_LABELS];
 428         int cmds[VDEV_LABELS];
 429         int errors[VDEV_LABELS];
 430         int l, nlabels;
 431
 432         g_topology_assert_not();
 433
 434         pp = cp->provider;
 435         ZFS_LOG(1, "Reading config from %s...", pp->name);
 436
 437         psize = pp->mediasize;
 438         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 439
 440         size = sizeof(*vdev_lists[0]) + pp->sectorsize -
 441             ((sizeof(*vdev_lists[0]) - 1) % pp->sectorsize) - 1;
 442
 443         buflen = sizeof(vdev_lists[0]->vp_nvlist);
 444
 445         *config = NULL;
 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         nlabels = 0;
 462         for (l = 0; l < VDEV_LABELS; l++) {
 463                 if (errors[l] != 0)
 464                         continue;
 465
 466                 buf = vdev_lists[l]->vp_nvlist;
 467
 468                 if (nvlist_unpack(buf, buflen, config, 0) != 0)
 469                         continue;
 470
 471                 if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
 472                     &state) != 0 || state > POOL_STATE_L2CACHE) {
 473                         nvlist_free(*config);
 474                         *config = NULL;
 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                         *config = NULL;
 484                         continue;
 485                 }
 486
 487                 nlabels++;
 488         }
 489
 490         /* Free the label storage */
 491         for (l = 0; l < VDEV_LABELS; l++)
 492                 kmem_free(vdev_lists[l], size);
 493
 494         return (nlabels);
 495 }
 496
 497 static void
 498 resize_configs(nvlist_t ***configs, uint64_t *count, uint64_t id)
 499 {
 500         nvlist_t **new_configs;
 501         uint64_t i;
 502
 503         if (id < *count)
 504                 return;
 505         new_configs = kmem_zalloc((id + 1) * sizeof(nvlist_t *),
 506             KM_SLEEP);
 507         for (i = 0; i < *count; i++)
 508                 new_configs[i] = (*configs)[i];
 509         if (*configs != NULL)
 510                 kmem_free(*configs, *count * sizeof(void *));
 511         *configs = new_configs;
 512         *count = id + 1;
 513 }
 514
 515 static void
 516 process_vdev_config(nvlist_t ***configs, uint64_t *count, nvlist_t *cfg,
 517     const char *name, uint64_t* known_pool_guid)
 518 {
 519         nvlist_t *vdev_tree;
 520         uint64_t pool_guid;
 521         uint64_t vdev_guid, known_guid;
 522         uint64_t id, txg, known_txg;
 523         char *pname;
 524         int i;
 525
 526         if (nvlist_lookup_string(cfg, ZPOOL_CONFIG_POOL_NAME, &pname) != 0 ||
 527             strcmp(pname, name) != 0)
 528                 goto ignore;
 529
 530         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_GUID, &pool_guid) != 0)
 531                 goto ignore;
 532
 533         if (nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_TOP_GUID, &vdev_guid) != 0)
 534                 goto ignore;
 535
 536         if (nvlist_lookup_nvlist(cfg, ZPOOL_CONFIG_VDEV_TREE, &vdev_tree) != 0)
 537                 goto ignore;
 538
 539         if (nvlist_lookup_uint64(vdev_tree, ZPOOL_CONFIG_ID, &id) != 0)
 540                 goto ignore;
 541
 542         VERIFY(nvlist_lookup_uint64(cfg, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
 543
 544         if (*known_pool_guid != 0) {
 545                 if (pool_guid != *known_pool_guid)
 546                         goto ignore;
 547         } else
 548                 *known_pool_guid = pool_guid;
 549
 550         resize_configs(configs, count, id);
 551
 552         if ((*configs)[id] != NULL) {
 553                 VERIFY(nvlist_lookup_uint64((*configs)[id],
 554                     ZPOOL_CONFIG_POOL_TXG, &known_txg) == 0);
 555                 if (txg <= known_txg)
 556                         goto ignore;
 557                 nvlist_free((*configs)[id]);
 558         }
 559
 560         (*configs)[id] = cfg;
 561         return;
 562
 563 ignore:
 564         nvlist_free(cfg);
 565 }
 566
 567 int
 568 vdev_geom_read_pool_label(const char *name,
 569     nvlist_t ***configs, uint64_t *count)
 570 {
 571         struct g_class *mp;
 572         struct g_geom *gp;
 573         struct g_provider *pp;
 574         struct g_consumer *zcp;
 575         nvlist_t *vdev_cfg;
 576         uint64_t pool_guid;
 577         int error, nlabels;
 578
 579         DROP_GIANT();
 580         g_topology_lock();
 581
 582         *configs = NULL;
 583         *count = 0;
 584         pool_guid = 0;
 585         LIST_FOREACH(mp, &g_classes, class) {
 586                 if (mp == &zfs_vdev_class)
 587                         continue;
 588                 LIST_FOREACH(gp, &mp->geom, geom) {
 589                         if (gp->flags & G_GEOM_WITHER)
 590                                 continue;
 591                         LIST_FOREACH(pp, &gp->provider, provider) {
 592                                 if (pp->flags & G_PF_WITHER)
 593                                         continue;
 594                                 zcp = vdev_geom_attach(pp, NULL, B_TRUE);
 595                                 if (zcp == NULL)
 596                                         continue;
 597                                 g_topology_unlock();
 598                                 nlabels = vdev_geom_read_config(zcp, &vdev_cfg);
 599                                 g_topology_lock();
 600                                 vdev_geom_detach(zcp, B_TRUE);
 601                                 if (nlabels == 0)
 602                                         continue;
 603                                 ZFS_LOG(1, "successfully read vdev config");
 604
 605                                 process_vdev_config(configs, count,
 606                                     vdev_cfg, name, &pool_guid);
 607                         }
 608                 }
 609         }
 610         g_topology_unlock();
 611         PICKUP_GIANT();
 612
 613         return (*count > 0 ? 0 : ENOENT);
 614 }
 615
 616 enum match {
 617         NO_MATCH = 0,           /* No matching labels found */
 618         TOPGUID_MATCH = 1,      /* Labels match top guid, not vdev guid*/
 619         ZERO_MATCH = 1,         /* Should never be returned */
 620         ONE_MATCH = 2,          /* 1 label matching the vdev_guid */
 621         TWO_MATCH = 3,          /* 2 label matching the vdev_guid */
 622         THREE_MATCH = 4,        /* 3 label matching the vdev_guid */
 623         FULL_MATCH = 5          /* all labels match the vdev_guid */
 624 };
 625
 626 static enum match
 627 vdev_attach_ok(vdev_t *vd, struct g_provider *pp)
 628 {
 629         nvlist_t *config;
 630         uint64_t pool_guid, top_guid, vdev_guid;
 631         struct g_consumer *cp;
 632         int nlabels;
 633
 634         cp = vdev_geom_attach(pp, NULL, B_TRUE);
 635         if (cp == NULL) {
 636                 ZFS_LOG(1, "Unable to attach tasting instance to %s.",
 637                     pp->name);
 638                 return (NO_MATCH);
 639         }
 640         g_topology_unlock();
 641         nlabels = vdev_geom_read_config(cp, &config);
 642         g_topology_lock();
 643         vdev_geom_detach(cp, B_TRUE);
 644         if (nlabels == 0) {
 645                 ZFS_LOG(1, "Unable to read config from %s.", pp->name);
 646                 return (NO_MATCH);
 647         }
 648
 649         pool_guid = 0;
 650         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid);
 651         top_guid = 0;
 652         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TOP_GUID, &top_guid);
 653         vdev_guid = 0;
 654         (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_GUID, &vdev_guid);
 655         nvlist_free(config);
 656
 657         /*
 658          * Check that the label's pool guid matches the desired guid.
 659          * Inactive spares and L2ARCs do not have any pool guid in the label.
 660          */
 661         if (pool_guid != 0 && pool_guid != spa_guid(vd->vdev_spa)) {
 662                 ZFS_LOG(1, "pool guid mismatch for provider %s: %ju != %ju.",
 663                     pp->name,
 664                     (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)pool_guid);
 665                 return (NO_MATCH);
 666         }
 667
 668         /*
 669          * Check that the label's vdev guid matches the desired guid.
 670          * The second condition handles possible race on vdev detach, when
 671          * remaining vdev receives GUID of destroyed top level mirror vdev.
 672          */
 673         if (vdev_guid == vd->vdev_guid) {
 674                 ZFS_LOG(1, "guids match for provider %s.", pp->name);
 675                 return (ZERO_MATCH + nlabels);
 676         } else if (top_guid == vd->vdev_guid && vd == vd->vdev_top) {
 677                 ZFS_LOG(1, "top vdev guid match for provider %s.", pp->name);
 678                 return (TOPGUID_MATCH);
 679         }
 680         ZFS_LOG(1, "vdev guid mismatch for provider %s: %ju != %ju.",
 681             pp->name, (uintmax_t)vd->vdev_guid, (uintmax_t)vdev_guid);
 682         return (NO_MATCH);
 683 }
 684
 685 static struct g_consumer *
 686 vdev_geom_attach_by_guids(vdev_t *vd)
 687 {
 688         struct g_class *mp;
 689         struct g_geom *gp;
 690         struct g_provider *pp, *best_pp;
 691         struct g_consumer *cp;
 692         enum match match, best_match;
 693
 694         g_topology_assert();
 695
 696         cp = NULL;
 697         best_pp = NULL;
 698         best_match = NO_MATCH;
 699         LIST_FOREACH(mp, &g_classes, class) {
 700                 if (mp == &zfs_vdev_class)
 701                         continue;
 702                 LIST_FOREACH(gp, &mp->geom, geom) {
 703                         if (gp->flags & G_GEOM_WITHER)
 704                                 continue;
 705                         LIST_FOREACH(pp, &gp->provider, provider) {
 706                                 match = vdev_attach_ok(vd, pp);
 707                                 if (match > best_match) {
 708                                         best_match = match;
 709                                         best_pp = pp;
 710                                 }
 711                                 if (match == FULL_MATCH)
 712                                         goto out;
 713                         }
 714                 }
 715         }
 716
 717 out:
 718         if (best_pp) {
 719                 cp = vdev_geom_attach(best_pp, vd, B_TRUE);
 720                 if (cp == NULL) {
 721                         printf("ZFS WARNING: Unable to attach to %s.\n",
 722                             best_pp->name);
 723                 }
 724         }
 725         return (cp);
 726 }
 727
 728 static struct g_consumer *
 729 vdev_geom_open_by_guids(vdev_t *vd)
 730 {
 731         struct g_consumer *cp;
 732         char *buf;
 733         size_t len;
 734
 735         g_topology_assert();
 736
 737         ZFS_LOG(1, "Searching by guids [%ju:%ju].",
 738                 (uintmax_t)spa_guid(vd->vdev_spa), (uintmax_t)vd->vdev_guid);
 739         cp = vdev_geom_attach_by_guids(vd);
 740         if (cp != NULL) {
 741                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 742                 buf = kmem_alloc(len, KM_SLEEP);
 743
 744                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 745                 spa_strfree(vd->vdev_path);
 746                 vd->vdev_path = buf;
 747
 748                 ZFS_LOG(1, "Attach by guid [%ju:%ju] succeeded, provider %s.",
 749                     (uintmax_t)spa_guid(vd->vdev_spa),
 750                     (uintmax_t)vd->vdev_guid, cp->provider->name);
 751         } else {
 752                 ZFS_LOG(1, "Search by guid [%ju:%ju] failed.",
 753                     (uintmax_t)spa_guid(vd->vdev_spa),
 754                     (uintmax_t)vd->vdev_guid);
 755         }
 756
 757         return (cp);
 758 }
 759
 760 static struct g_consumer *
 761 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 762 {
 763         struct g_provider *pp;
 764         struct g_consumer *cp;
 765
 766         g_topology_assert();
 767
 768         cp = NULL;
 769         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 770         if (pp != NULL) {
 771                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 772                 if (!check_guid || vdev_attach_ok(vd, pp) == FULL_MATCH)
 773                         cp = vdev_geom_attach(pp, vd, B_FALSE);
 774         }
 775
 776         return (cp);
 777 }
 778
 779 static int
 780 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 781     uint64_t *logical_ashift, uint64_t *physical_ashift)
 782 {
 783         struct g_provider *pp;
 784         struct g_consumer *cp;
 785         size_t bufsize;
 786         int error;
 787
 788         /* Set the TLS to indicate downstack that we should not access zvols*/
 789         VERIFY(tsd_set(zfs_geom_probe_vdev_key, vd) == 0);
 790
 791         /*
 792          * We must have a pathname, and it must be absolute.
 793          */
 794         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 795                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 796                 return (EINVAL);
 797         }
 798
 799         /*
 800          * Reopen the device if it's not currently open. Otherwise,
 801          * just update the physical size of the device.
 802          */
 803         if ((cp = vd->vdev_tsd) != NULL) {
 804                 ASSERT(vd->vdev_reopening);
 805                 goto skip_open;
 806         }
 807
 808         DROP_GIANT();
 809         g_topology_lock();
 810         error = 0;
 811
 812         if (vd->vdev_spa->spa_splitting_newspa ||
 813             (vd->vdev_prevstate == VDEV_STATE_UNKNOWN &&
 814              vd->vdev_spa->spa_load_state == SPA_LOAD_NONE ||
 815              vd->vdev_spa->spa_load_state == SPA_LOAD_CREATE)) {
 816                 /*
 817                  * We are dealing with a vdev that hasn't been previously
 818                  * opened (since boot), and we are not loading an
 819                  * existing pool configuration.  This looks like a
 820                  * vdev add operation to a new or existing pool.
 821                  * Assume the user knows what he/she is doing and find
 822                  * GEOM provider by its name, ignoring GUID mismatches.
 823                  *
 824                  * XXPOLICY: It would be safer to only allow a device
 825                  *           that is unlabeled or labeled but missing
 826                  *           GUID information to be opened in this fashion,
 827                  *           unless we are doing a split, in which case we
 828                  *           should allow any guid.
 829                  */
 830                 cp = vdev_geom_open_by_path(vd, 0);
 831         } else {
 832                 /*
 833                  * Try using the recorded path for this device, but only
 834                  * accept it if its label data contains the expected GUIDs.
 835                  */
 836                 cp = vdev_geom_open_by_path(vd, 1);
 837                 if (cp == NULL) {
 838                         /*
 839                          * The device at vd->vdev_path doesn't have the
 840                          * expected GUIDs. The disks might have merely
 841                          * moved around so try all other GEOM providers
 842                          * to find one with the right GUIDs.
 843                          */
 844                         cp = vdev_geom_open_by_guids(vd);
 845                 }
 846         }
 847
 848         /* Clear the TLS now that tasting is done */
 849         VERIFY(tsd_set(zfs_geom_probe_vdev_key, NULL) == 0);
 850
 851         if (cp == NULL) {
 852                 ZFS_LOG(1, "Vdev %s not found.", vd->vdev_path);
 853                 error = ENOENT;
 854         } else {
 855                 struct consumer_priv_t *priv;
 856                 struct consumer_vdev_elem *elem;
 857                 int spamode;
 858
 859                 priv = (struct consumer_priv_t*)&cp->private;
 860                 if (cp->private == NULL)
 861                         SLIST_INIT(priv);
 862                 elem = g_malloc(sizeof(*elem), M_WAITOK|M_ZERO);
 863                 elem->vd = vd;
 864                 SLIST_INSERT_HEAD(priv, elem, elems);
 865
 866                 spamode = spa_mode(vd->vdev_spa);
 867                 if (cp->provider->sectorsize > VDEV_PAD_SIZE ||
 868                     !ISP2(cp->provider->sectorsize)) {
 869                         ZFS_LOG(1, "Provider %s has unsupported sectorsize.",
 870                             cp->provider->name);
 871
 872                         vdev_geom_close_locked(vd);
 873                         error = EINVAL;
 874                         cp = NULL;
 875                 } else if (cp->acw == 0 && (spamode & FWRITE) != 0) {
 876                         int i;
 877
 878                         for (i = 0; i < 5; i++) {
 879                                 error = g_access(cp, 0, 1, 0);
 880                                 if (error == 0)
 881                                         break;
 882                                 g_topology_unlock();
 883                                 tsleep(vd, 0, "vdev", hz / 2);
 884                                 g_topology_lock();
 885                         }
 886                         if (error != 0) {
 887                                 printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 888                                     cp->provider->name, error);
 889                                 vdev_geom_close_locked(vd);
 890                                 cp = NULL;
 891                         }
 892                 }
 893         }
 894
 895         /* Fetch initial physical path information for this device. */
 896         if (cp != NULL) {
 897                 vdev_geom_attrchanged(cp, "GEOM::physpath");
 898
 899                 /* Set other GEOM characteristics */
 900                 vdev_geom_set_physpath(vd, cp, /*do_null_update*/B_FALSE);
 901                 vdev_geom_set_rotation_rate(vd, cp);
 902         }
 903
 904         g_topology_unlock();
 905         PICKUP_GIANT();
 906         if (cp == NULL) {
 907                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 908                 vdev_dbgmsg(vd, "vdev_geom_open: failed to open [error=%d]",
 909                     error);
 910                 return (error);
 911         }
 912 skip_open:
 913         pp = cp->provider;
 914
 915         /*
 916          * Determine the actual size of the device.
 917          */
 918         *max_psize = *psize = pp->mediasize;
 919
 920         /*
 921          * Determine the device's minimum transfer size and preferred
 922          * transfer size.
 923          */
 924         *logical_ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 925         *physical_ashift = 0;
 926         if (pp->stripesize > (1 << *logical_ashift) && ISP2(pp->stripesize) &&
 927             pp->stripesize <= (1 << SPA_MAXASHIFT) && pp->stripeoffset == 0)
 928                 *physical_ashift = highbit(pp->stripesize) - 1;
 929
 930         /*
 931          * Clear the nowritecache settings, so that on a vdev_reopen()
 932          * we will try again.
 933          */
 934         vd->vdev_nowritecache = B_FALSE;
 935
 936         return (0);
 937 }
 938
 939 static void
 940 vdev_geom_close(vdev_t *vd)
 941 {
 942         struct g_consumer *cp;
 943
 944         cp = vd->vdev_tsd;
 945
 946         DROP_GIANT();
 947         g_topology_lock();
 948
 949         if (!vd->vdev_reopening ||
 950             (cp != NULL && ((cp->flags & G_CF_ORPHAN) != 0 ||
 951             (cp->provider != NULL && cp->provider->error != 0))))
 952                 vdev_geom_close_locked(vd);
 953
 954         g_topology_unlock();
 955         PICKUP_GIANT();
 956 }
 957
 958 static void
 959 vdev_geom_io_intr(struct bio *bp)
 960 {
 961         vdev_t *vd;
 962         zio_t *zio;
 963
 964         zio = bp->bio_caller1;
 965         vd = zio->io_vd;
 966         zio->io_error = bp->bio_error;
 967         if (zio->io_error == 0 && bp->bio_resid != 0)
 968                 zio->io_error = SET_ERROR(EIO);
 969
 970         switch(zio->io_error) {
 971         case ENOTSUP:
 972                 /*
 973                  * If we get ENOTSUP for BIO_FLUSH or BIO_DELETE we know
 974                  * that future attempts will never succeed. In this case
 975                  * we set a persistent flag so that we don't bother with
 976                  * requests in the future.
 977                  */
 978                 switch(bp->bio_cmd) {
 979                 case BIO_FLUSH:
 980                         vd->vdev_nowritecache = B_TRUE;
 981                         break;
 982                 case BIO_DELETE:
 983                         vd->vdev_notrim = B_TRUE;
 984                         break;
 985                 }
 986                 break;
 987         case ENXIO:
 988                 if (!vd->vdev_remove_wanted) {
 989                         /*
 990                          * If provider's error is set we assume it is being
 991                          * removed.
 992                          */
 993                         if (bp->bio_to->error != 0) {
 994                                 vd->vdev_remove_wanted = B_TRUE;
 995                                 spa_async_request(zio->io_spa,
 996                                     SPA_ASYNC_REMOVE);
 997                         } else if (!vd->vdev_delayed_close) {
 998                                 vd->vdev_delayed_close = B_TRUE;
 999                         }
1000                 }
1001                 break;
1002         }
1003
1004         /*
1005          * We have to split bio freeing into two parts, because the ABD code
1006          * cannot be called in this context and vdev_op_io_done is not called
1007          * for ZIO_TYPE_IOCTL zio-s.
1008          */
1009         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
1010                 g_destroy_bio(bp);
1011                 zio->io_bio = NULL;
1012         }
1013         zio_delay_interrupt(zio);
1014 }
1015
1016 static void
1017 vdev_geom_io_start(zio_t *zio)
1018 {
1019         vdev_t *vd;
1020         struct g_consumer *cp;
1021         struct bio *bp;
1022         int error;
1023
1024         vd = zio->io_vd;
1025
1026         switch (zio->io_type) {
1027         case ZIO_TYPE_IOCTL:
1028                 /* XXPOLICY */
1029                 if (!vdev_readable(vd)) {
1030                         zio->io_error = SET_ERROR(ENXIO);
1031                         zio_interrupt(zio);
1032                         return;
1033                 } else {
1034                         switch (zio->io_cmd) {
1035                         case DKIOCFLUSHWRITECACHE:
1036                                 if (zfs_nocacheflush || vdev_geom_bio_flush_disable)
1037                                         break;
1038                                 if (vd->vdev_nowritecache) {
1039                                         zio->io_error = SET_ERROR(ENOTSUP);
1040                                         break;
1041                                 }
1042                                 goto sendreq;
1043                         default:
1044                                 zio->io_error = SET_ERROR(ENOTSUP);
1045                         }
1046                 }
1047
1048                 zio_execute(zio);
1049                 return;
1050         case ZIO_TYPE_FREE:
1051                 if (vd->vdev_notrim) {
1052                         zio->io_error = SET_ERROR(ENOTSUP);
1053                 } else if (!vdev_geom_bio_delete_disable) {
1054                         goto sendreq;
1055                 }
1056                 zio_execute(zio);
1057                 return;
1058         }
1059 sendreq:
1060         ASSERT(zio->io_type == ZIO_TYPE_READ ||
1061             zio->io_type == ZIO_TYPE_WRITE ||
1062             zio->io_type == ZIO_TYPE_FREE ||
1063             zio->io_type == ZIO_TYPE_IOCTL);
1064
1065         cp = vd->vdev_tsd;
1066         if (cp == NULL) {
1067                 zio->io_error = SET_ERROR(ENXIO);
1068                 zio_interrupt(zio);
1069                 return;
1070         }
1071         bp = g_alloc_bio();
1072         bp->bio_caller1 = zio;
1073         switch (zio->io_type) {
1074         case ZIO_TYPE_READ:
1075         case ZIO_TYPE_WRITE:
1076                 zio->io_target_timestamp = zio_handle_io_delay(zio);
1077                 bp->bio_offset = zio->io_offset;
1078                 bp->bio_length = zio->io_size;
1079                 if (zio->io_type == ZIO_TYPE_READ) {
1080                         bp->bio_cmd = BIO_READ;
1081                         bp->bio_data =
1082                             abd_borrow_buf(zio->io_abd, zio->io_size);
1083                 } else {
1084                         bp->bio_cmd = BIO_WRITE;
1085                         bp->bio_data =
1086                             abd_borrow_buf_copy(zio->io_abd, zio->io_size);
1087                 }
1088                 break;
1089         case ZIO_TYPE_FREE:
1090                 bp->bio_cmd = BIO_DELETE;
1091                 bp->bio_data = NULL;
1092                 bp->bio_offset = zio->io_offset;
1093                 bp->bio_length = zio->io_size;
1094                 break;
1095         case ZIO_TYPE_IOCTL:
1096                 bp->bio_cmd = BIO_FLUSH;
1097                 bp->bio_flags |= BIO_ORDERED;
1098                 bp->bio_data = NULL;
1099                 bp->bio_offset = cp->provider->mediasize;
1100                 bp->bio_length = 0;
1101                 break;
1102         }
1103         bp->bio_done = vdev_geom_io_intr;
1104         zio->io_bio = bp;
1105
1106         g_io_request(bp, cp);
1107 }
1108
1109 static void
1110 vdev_geom_io_done(zio_t *zio)
1111 {
1112         struct bio *bp = zio->io_bio;
1113
1114         if (zio->io_type != ZIO_TYPE_READ && zio->io_type != ZIO_TYPE_WRITE) {
1115                 ASSERT(bp == NULL);
1116                 return;
1117         }
1118
1119         if (bp == NULL) {
1120                 ASSERT3S(zio->io_error, ==, ENXIO);
1121                 return;
1122         }
1123
1124         if (zio->io_type == ZIO_TYPE_READ)
1125                 abd_return_buf_copy(zio->io_abd, bp->bio_data, zio->io_size);
1126         else
1127                 abd_return_buf(zio->io_abd, bp->bio_data, zio->io_size);
1128
1129         g_destroy_bio(bp);
1130         zio->io_bio = NULL;
1131 }
1132
1133 static void
1134 vdev_geom_hold(vdev_t *vd)
1135 {
1136 }
1137
1138 static void
1139 vdev_geom_rele(vdev_t *vd)
1140 {
1141 }
1142
1143 vdev_ops_t vdev_geom_ops = {
1144         vdev_geom_open,
1145         vdev_geom_close,
1146         vdev_default_asize,
1147         vdev_geom_io_start,
1148         vdev_geom_io_done,
1149         NULL,
1150         NULL,
1151         vdev_geom_hold,
1152         vdev_geom_rele,
1153         NULL,
1154         vdev_default_xlate,
1155         VDEV_TYPE_DISK,         /* name of this vdev type */
1156         B_TRUE                  /* leaf vdev */
1157 };