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
  26 #include <sys/zfs_context.h>
  27 #include <sys/param.h>
  28 #include <sys/kernel.h>
  29 #include <sys/bio.h>
  30 #include <sys/disk.h>
  31 #include <sys/spa.h>
  32 #include <sys/spa_impl.h>
  33 #include <sys/vdev_impl.h>
  34 #include <sys/fs/zfs.h>
  35 #include <sys/zio.h>
  36 #include <geom/geom.h>
  37 #include <geom/geom_int.h>
  38
  39 /*
  40  * Virtual device vector for GEOM.
  41  */
  42
  43 struct g_class zfs_vdev_class = {
  44         .name = "ZFS::VDEV",
  45         .version = G_VERSION,
  46 };
  47
  48 DECLARE_GEOM_CLASS(zfs_vdev_class, zfs_vdev);
  49
  50 static void
  51 vdev_geom_orphan(struct g_consumer *cp)
  52 {
  53         struct g_geom *gp;
  54         vdev_t *vd;
  55         int error;
  56
  57         g_topology_assert();
  58
  59         vd = cp->private;
  60         gp = cp->geom;
  61         error = cp->provider->error;
  62
  63         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
  64         if (cp->acr + cp->acw + cp->ace > 0)
  65                 g_access(cp, -cp->acr, -cp->acw, -cp->ace);
  66         ZFS_LOG(1, "Destroyed consumer to %s.", cp->provider->name);
  67         g_detach(cp);
  68         g_destroy_consumer(cp);
  69         /* Destroy geom if there are no consumers left. */
  70         if (LIST_EMPTY(&gp->consumer)) {
  71                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
  72                 g_wither_geom(gp, error);
  73         }
  74         vd->vdev_tsd = NULL;
  75         vd->vdev_remove_wanted = B_TRUE;
  76         spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
  77 }
  78
  79 static struct g_consumer *
  80 vdev_geom_attach(struct g_provider *pp)
  81 {
  82         struct g_geom *gp;
  83         struct g_consumer *cp;
  84
  85         g_topology_assert();
  86
  87         ZFS_LOG(1, "Attaching to %s.", pp->name);
  88         /* Do we have geom already? No? Create one. */
  89         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
  90                 if (gp->flags & G_GEOM_WITHER)
  91                         continue;
  92                 if (strcmp(gp->name, "zfs::vdev") != 0)
  93                         continue;
  94                 break;
  95         }
  96         if (gp == NULL) {
  97                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
  98                 gp->orphan = vdev_geom_orphan;
  99                 cp = g_new_consumer(gp);
 100                 if (g_attach(cp, pp) != 0) {
 101                         g_wither_geom(gp, ENXIO);
 102                         return (NULL);
 103                 }
 104                 if (g_access(cp, 1, 0, 1) != 0) {
 105                         g_wither_geom(gp, ENXIO);
 106                         return (NULL);
 107                 }
 108                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 109         } else {
 110                 /* Check if we are already connected to this provider. */
 111                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 112                         if (cp->provider == pp) {
 113                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 114                                 break;
 115                         }
 116                 }
 117                 if (cp == NULL) {
 118                         cp = g_new_consumer(gp);
 119                         if (g_attach(cp, pp) != 0) {
 120                                 g_destroy_consumer(cp);
 121                                 return (NULL);
 122                         }
 123                         if (g_access(cp, 1, 0, 1) != 0) {
 124                                 g_detach(cp);
 125                                 g_destroy_consumer(cp);
 126                                 return (NULL);
 127                         }
 128                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 129                 } else {
 130                         if (g_access(cp, 1, 0, 1) != 0)
 131                                 return (NULL);
 132                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 133                 }
 134         }
 135         return (cp);
 136 }
 137
 138 static void
 139 vdev_geom_detach(void *arg, int flag __unused)
 140 {
 141         struct g_geom *gp;
 142         struct g_consumer *cp;
 143
 144         g_topology_assert();
 145         cp = arg;
 146         gp = cp->geom;
 147
 148         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 149         g_access(cp, -1, 0, -1);
 150         /* Destroy consumer on last close. */
 151         if (cp->acr == 0 && cp->ace == 0) {
 152                 ZFS_LOG(1, "Destroyed consumer to %s.", cp->provider->name);
 153                 if (cp->acw > 0)
 154                         g_access(cp, 0, -cp->acw, 0);
 155                 g_detach(cp);
 156                 g_destroy_consumer(cp);
 157         }
 158         /* Destroy geom if there are no consumers left. */
 159         if (LIST_EMPTY(&gp->consumer)) {
 160                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 161                 g_wither_geom(gp, ENXIO);
 162         }
 163 }
 164
 165 static uint64_t
 166 nvlist_get_guid(nvlist_t *list)
 167 {
 168         nvpair_t *elem = NULL;
 169         uint64_t value;
 170
 171         while ((elem = nvlist_next_nvpair(list, elem)) != NULL) {
 172                 if (nvpair_type(elem) == DATA_TYPE_UINT64 &&
 173                     strcmp(nvpair_name(elem), "guid") == 0) {
 174                         VERIFY(nvpair_value_uint64(elem, &value) == 0);
 175                         return (value);
 176                 }
 177         }
 178         return (0);
 179 }
 180
 181 static int
 182 vdev_geom_io(struct g_consumer *cp, int cmd, void *data, off_t offset, off_t size)
 183 {
 184         struct bio *bp;
 185         u_char *p;
 186         off_t off, maxio;
 187         int error;
 188
 189         ASSERT((offset % cp->provider->sectorsize) == 0);
 190         ASSERT((size % cp->provider->sectorsize) == 0);
 191
 192         bp = g_alloc_bio();
 193         off = offset;
 194         offset += size;
 195         p = data;
 196         maxio = MAXPHYS - (MAXPHYS % cp->provider->sectorsize);
 197         error = 0;
 198
 199         for (; off < offset; off += maxio, p += maxio, size -= maxio) {
 200                 bzero(bp, sizeof(*bp));
 201                 bp->bio_cmd = cmd;
 202                 bp->bio_done = NULL;
 203                 bp->bio_offset = off;
 204                 bp->bio_length = MIN(size, maxio);
 205                 bp->bio_data = p;
 206                 g_io_request(bp, cp);
 207                 error = biowait(bp, "vdev_geom_io");
 208                 if (error != 0)
 209                         break;
 210         }
 211
 212         g_destroy_bio(bp);
 213         return (error);
 214 }
 215
 216 static uint64_t
 217 vdev_geom_read_guid(struct g_consumer *cp)
 218 {
 219         struct g_provider *pp;
 220         vdev_label_t *label;
 221         char *p, *buf;
 222         size_t buflen;
 223         uint64_t psize;
 224         off_t offset, size;
 225         uint64_t guid;
 226         int error, l, len, iszvol;
 227
 228         g_topology_assert_not();
 229
 230         pp = cp->provider;
 231         ZFS_LOG(1, "Reading guid from %s...", pp->name);
 232         if (g_getattr("ZFS::iszvol", cp, &iszvol) == 0 && iszvol) {
 233                 ZFS_LOG(1, "Skipping ZVOL-based provider %s.", pp->name);
 234                 return (0);
 235         }
 236
 237         psize = pp->mediasize;
 238         psize = P2ALIGN(psize, (uint64_t)sizeof(vdev_label_t));
 239
 240         size = sizeof(*label) + pp->sectorsize -
 241             ((sizeof(*label) - 1) % pp->sectorsize) - 1;
 242
 243         guid = 0;
 244         label = kmem_alloc(size, KM_SLEEP);
 245         buflen = sizeof(label->vl_vdev_phys.vp_nvlist);
 246
 247         for (l = 0; l < VDEV_LABELS; l++) {
 248                 nvlist_t *config = NULL;
 249
 250                 offset = vdev_label_offset(psize, l, 0);
 251                 if ((offset % pp->sectorsize) != 0)
 252                         continue;
 253
 254                 if (vdev_geom_io(cp, BIO_READ, label, offset, size) != 0)
 255                         continue;
 256                 buf = label->vl_vdev_phys.vp_nvlist;
 257
 258                 if (nvlist_unpack(buf, buflen, &config, 0) != 0)
 259                         continue;
 260
 261                 guid = nvlist_get_guid(config);
 262                 nvlist_free(config);
 263                 if (guid != 0)
 264                         break;
 265         }
 266
 267         kmem_free(label, size);
 268         if (guid != 0)
 269                 ZFS_LOG(1, "guid for %s is %ju", pp->name, (uintmax_t)guid);
 270         return (guid);
 271 }
 272
 273 struct vdev_geom_find {
 274         uint64_t guid;
 275         struct g_consumer *cp;
 276 };
 277
 278 static void
 279 vdev_geom_taste_orphan(struct g_consumer *cp)
 280 {
 281
 282         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
 283             cp->provider->name));
 284 }
 285
 286 static void
 287 vdev_geom_attach_by_guid_event(void *arg, int flags __unused)
 288 {
 289         struct vdev_geom_find *ap;
 290         struct g_class *mp;
 291         struct g_geom *gp, *zgp;
 292         struct g_provider *pp;
 293         struct g_consumer *zcp;
 294         uint64_t guid;
 295
 296         g_topology_assert();
 297
 298         ap = arg;
 299
 300         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 301         /* This orphan function should be never called. */
 302         zgp->orphan = vdev_geom_taste_orphan;
 303         zcp = g_new_consumer(zgp);
 304
 305         LIST_FOREACH(mp, &g_classes, class) {
 306                 if (mp == &zfs_vdev_class)
 307                         continue;
 308                 LIST_FOREACH(gp, &mp->geom, geom) {
 309                         if (gp->flags & G_GEOM_WITHER)
 310                                 continue;
 311                         LIST_FOREACH(pp, &gp->provider, provider) {
 312                                 if (pp->flags & G_PF_WITHER)
 313                                         continue;
 314                                 g_attach(zcp, pp);
 315                                 if (g_access(zcp, 1, 0, 0) != 0) {
 316                                         g_detach(zcp);
 317                                         continue;
 318                                 }
 319                                 g_topology_unlock();
 320                                 guid = vdev_geom_read_guid(zcp);
 321                                 g_topology_lock();
 322                                 g_access(zcp, -1, 0, 0);
 323                                 g_detach(zcp);
 324                                 if (guid != ap->guid)
 325                                         continue;
 326                                 ap->cp = vdev_geom_attach(pp);
 327                                 if (ap->cp == NULL) {
 328                                         printf("ZFS WARNING: Unable to attach to %s.\n",
 329                                             pp->name);
 330                                         continue;
 331                                 }
 332                                 goto end;
 333                         }
 334                 }
 335         }
 336         ap->cp = NULL;
 337 end:
 338         g_destroy_consumer(zcp);
 339         g_destroy_geom(zgp);
 340 }
 341
 342 static struct g_consumer *
 343 vdev_geom_attach_by_guid(uint64_t guid)
 344 {
 345         struct vdev_geom_find *ap;
 346         struct g_consumer *cp;
 347
 348         ap = kmem_zalloc(sizeof(*ap), KM_SLEEP);
 349         ap->guid = guid;
 350         g_waitfor_event(vdev_geom_attach_by_guid_event, ap, M_WAITOK, NULL);
 351         cp = ap->cp;
 352         kmem_free(ap, sizeof(*ap));
 353         return (cp);
 354 }
 355
 356 static struct g_consumer *
 357 vdev_geom_open_by_guid(vdev_t *vd)
 358 {
 359         struct g_consumer *cp;
 360         char *buf;
 361         size_t len;
 362
 363         ZFS_LOG(1, "Searching by guid [%ju].", (uintmax_t)vd->vdev_guid);
 364         cp = vdev_geom_attach_by_guid(vd->vdev_guid);
 365         if (cp != NULL) {
 366                 len = strlen(cp->provider->name) + strlen("/dev/") + 1;
 367                 buf = kmem_alloc(len, KM_SLEEP);
 368
 369                 snprintf(buf, len, "/dev/%s", cp->provider->name);
 370                 spa_strfree(vd->vdev_path);
 371                 vd->vdev_path = buf;
 372
 373                 ZFS_LOG(1, "Attach by guid [%ju] succeeded, provider %s.",
 374                     (uintmax_t)vd->vdev_guid, vd->vdev_path);
 375         } else {
 376                 ZFS_LOG(1, "Search by guid [%ju] failed.",
 377                     (uintmax_t)vd->vdev_guid);
 378         }
 379
 380         return (cp);
 381 }
 382
 383 static struct g_consumer *
 384 vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 385 {
 386         struct g_provider *pp;
 387         struct g_consumer *cp;
 388         uint64_t guid;
 389
 390         cp = NULL;
 391         g_topology_lock();
 392         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 393         if (pp != NULL) {
 394                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 395                 cp = vdev_geom_attach(pp);
 396                 if (cp != NULL && check_guid) {
 397                         g_topology_unlock();
 398                         guid = vdev_geom_read_guid(cp);
 399                         g_topology_lock();
 400                         if (guid != vd->vdev_guid) {
 401                                 vdev_geom_detach(cp, 0);
 402                                 cp = NULL;
 403                                 ZFS_LOG(1, "guid mismatch for provider %s: "
 404                                     "%ju != %ju.", vd->vdev_path,
 405                                     (uintmax_t)vd->vdev_guid, (uintmax_t)guid);
 406                         } else {
 407                                 ZFS_LOG(1, "guid match for provider %s.",
 408                                     vd->vdev_path);
 409                         }
 410                 }
 411         }
 412         g_topology_unlock();
 413
 414         return (cp);
 415 }
 416
 417 static int
 418 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *ashift)
 419 {
 420         struct g_provider *pp;
 421         struct g_consumer *cp;
 422         int error, owned;
 423
 424         /*
 425          * We must have a pathname, and it must be absolute.
 426          */
 427         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 428                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 429                 return (EINVAL);
 430         }
 431
 432         vd->vdev_tsd = NULL;
 433
 434         if ((owned = mtx_owned(&Giant)))
 435                 mtx_unlock(&Giant);
 436         error = 0;
 437
 438         /*
 439          * If we're creating pool, just find GEOM provider by its name
 440          * and ignore GUID mismatches.
 441          */
 442         if (vd->vdev_spa->spa_load_state == SPA_LOAD_NONE)
 443                 cp = vdev_geom_open_by_path(vd, 0);
 444         else {
 445                 cp = vdev_geom_open_by_path(vd, 1);
 446                 if (cp == NULL) {
 447                         /*
 448                          * The device at vd->vdev_path doesn't have the
 449                          * expected guid. The disks might have merely
 450                          * moved around so try all other GEOM providers
 451                          * to find one with the right guid.
 452                          */
 453                         cp = vdev_geom_open_by_guid(vd);
 454                 }
 455         }
 456
 457         if (cp == NULL) {
 458                 ZFS_LOG(1, "Provider %s not found.", vd->vdev_path);
 459                 error = ENOENT;
 460         } else if (cp->acw == 0 && (spa_mode(vd->vdev_spa) & FWRITE) != 0) {
 461                 int i;
 462
 463                 g_topology_lock();
 464                 for (i = 0; i < 5; i++) {
 465                         error = g_access(cp, 0, 1, 0);
 466                         if (error == 0)
 467                                 break;
 468                         g_topology_unlock();
 469                         tsleep(vd, 0, "vdev", hz / 2);
 470                         g_topology_lock();
 471                 }
 472                 if (error != 0) {
 473                         printf("ZFS WARNING: Unable to open %s for writing (error=%d).\n",
 474                             vd->vdev_path, error);
 475                         vdev_geom_detach(cp, 0);
 476                         cp = NULL;
 477                 }
 478                 g_topology_unlock();
 479         }
 480         if (owned)
 481                 mtx_lock(&Giant);
 482         if (cp == NULL) {
 483                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 484                 return (error);
 485         }
 486
 487         cp->private = vd;
 488         vd->vdev_tsd = cp;
 489         pp = cp->provider;
 490
 491         /*
 492          * Determine the actual size of the device.
 493          */
 494         *psize = pp->mediasize;
 495
 496         /*
 497          * Determine the device's minimum transfer size.
 498          */
 499         *ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 500
 501         /*
 502          * Clear the nowritecache bit, so that on a vdev_reopen() we will
 503          * try again.
 504          */
 505         vd->vdev_nowritecache = B_FALSE;
 506
 507         return (0);
 508 }
 509
 510 static void
 511 vdev_geom_close(vdev_t *vd)
 512 {
 513         struct g_consumer *cp;
 514
 515         cp = vd->vdev_tsd;
 516         if (cp == NULL)
 517                 return;
 518         vd->vdev_tsd = NULL;
 519         g_post_event(vdev_geom_detach, cp, M_WAITOK, NULL);
 520 }
 521
 522 static void
 523 vdev_geom_io_intr(struct bio *bp)
 524 {
 525         zio_t *zio;
 526
 527         zio = bp->bio_caller1;
 528         zio->io_error = bp->bio_error;
 529         if (zio->io_error == 0 && bp->bio_resid != 0)
 530                 zio->io_error = EIO;
 531         if (bp->bio_cmd == BIO_FLUSH && bp->bio_error == ENOTSUP) {
 532                 vdev_t *vd;
 533
 534                 /*
 535                  * If we get ENOTSUP, we know that no future
 536                  * attempts will ever succeed.  In this case we
 537                  * set a persistent bit so that we don't bother
 538                  * with the ioctl in the future.
 539                  */
 540                 vd = zio->io_vd;
 541                 vd->vdev_nowritecache = B_TRUE;
 542         }
 543         g_destroy_bio(bp);
 544         zio_interrupt(zio);
 545 }
 546
 547 static int
 548 vdev_geom_io_start(zio_t *zio)
 549 {
 550         vdev_t *vd;
 551         struct g_consumer *cp;
 552         struct bio *bp;
 553         int error;
 554
 555         vd = zio->io_vd;
 556
 557         if (zio->io_type == ZIO_TYPE_IOCTL) {
 558                 /* XXPOLICY */
 559                 if (!vdev_readable(vd)) {
 560                         zio->io_error = ENXIO;
 561                         return (ZIO_PIPELINE_CONTINUE);
 562                 }
 563
 564                 switch (zio->io_cmd) {
 565
 566                 case DKIOCFLUSHWRITECACHE:
 567
 568                         if (zfs_nocacheflush)
 569                                 break;
 570
 571                         if (vd->vdev_nowritecache) {
 572                                 zio->io_error = ENOTSUP;
 573                                 break;
 574                         }
 575
 576                         goto sendreq;
 577                 default:
 578                         zio->io_error = ENOTSUP;
 579                 }
 580
 581                 return (ZIO_PIPELINE_CONTINUE);
 582         }
 583 sendreq:
 584         cp = vd->vdev_tsd;
 585         if (cp == NULL) {
 586                 zio->io_error = ENXIO;
 587                 return (ZIO_PIPELINE_CONTINUE);
 588         }
 589         bp = g_alloc_bio();
 590         bp->bio_caller1 = zio;
 591         switch (zio->io_type) {
 592         case ZIO_TYPE_READ:
 593         case ZIO_TYPE_WRITE:
 594                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 595                 bp->bio_data = zio->io_data;
 596                 bp->bio_offset = zio->io_offset;
 597                 bp->bio_length = zio->io_size;
 598                 break;
 599         case ZIO_TYPE_IOCTL:
 600                 bp->bio_cmd = BIO_FLUSH;
 601                 bp->bio_flags |= BIO_ORDERED;
 602                 bp->bio_data = NULL;
 603                 bp->bio_offset = cp->provider->mediasize;
 604                 bp->bio_length = 0;
 605                 break;
 606         }
 607         bp->bio_done = vdev_geom_io_intr;
 608
 609         g_io_request(bp, cp);
 610
 611         return (ZIO_PIPELINE_STOP);
 612 }
 613
 614 static void
 615 vdev_geom_io_done(zio_t *zio)
 616 {
 617 }
 618
 619 vdev_ops_t vdev_geom_ops = {
 620         vdev_geom_open,
 621         vdev_geom_close,
 622         vdev_default_asize,
 623         vdev_geom_io_start,
 624         vdev_geom_io_done,
 625         NULL,
 626         VDEV_TYPE_DISK,         /* name of this vdev type */
 627         B_TRUE                  /* leaf vdev */
 628 };