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/vdev_impl.h>
  33 #include <sys/fs/zfs.h>
  34 #include <sys/zio.h>
  35 #include <geom/geom.h>
  36 #include <geom/geom_int.h>
  37
  38 /*
  39  * Virtual device vector for GEOM.
  40  */
  41
  42 struct g_class zfs_vdev_class = {
  43         .name = "ZFS::VDEV",
  44         .version = G_VERSION,
  45 };
  46
  47 DECLARE_GEOM_CLASS(zfs_vdev_class, zfs_vdev);
  48
  49 typedef struct vdev_geom_ctx {
  50         struct g_consumer *gc_consumer;
  51         int gc_state;
  52         struct bio_queue_head gc_queue;
  53         struct mtx gc_queue_mtx;
  54 } vdev_geom_ctx_t;
  55
  56 static void
  57 vdev_geom_release(vdev_t *vd)
  58 {
  59         vdev_geom_ctx_t *ctx;
  60
  61         ctx = vd->vdev_tsd;
  62         vd->vdev_tsd = NULL;
  63
  64         mtx_lock(&ctx->gc_queue_mtx);
  65         ctx->gc_state = 1;
  66         wakeup_one(&ctx->gc_queue);
  67         while (ctx->gc_state != 2)
  68                 msleep(&ctx->gc_state, &ctx->gc_queue_mtx, 0, "vgeom:w", 0);
  69         mtx_unlock(&ctx->gc_queue_mtx);
  70         mtx_destroy(&ctx->gc_queue_mtx);
  71         kmem_free(ctx, sizeof(*ctx));
  72 }
  73
  74 static void
  75 vdev_geom_orphan(struct g_consumer *cp)
  76 {
  77         struct g_geom *gp;
  78         vdev_t *vd;
  79         int error;
  80
  81         g_topology_assert();
  82
  83         vd = cp->private;
  84         gp = cp->geom;
  85         error = cp->provider->error;
  86
  87         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
  88         if (cp->acr + cp->acw + cp->ace > 0)
  89                 g_access(cp, -cp->acr, -cp->acw, -cp->ace);
  90         ZFS_LOG(1, "Destroyed consumer to %s.", cp->provider->name);
  91         g_detach(cp);
  92         g_destroy_consumer(cp);
  93         /* Destroy geom if there are no consumers left. */
  94         if (LIST_EMPTY(&gp->consumer)) {
  95                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
  96                 g_wither_geom(gp, error);
  97         }
  98         vdev_geom_release(vd);
  99         /* Both methods below work, but in a bit different way. */
 100 #if 0
 101         vd->vdev_reopen_wanted = 1;
 102 #else
 103         vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 104         vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN, vd->vdev_stat.vs_aux);
 105 #endif
 106 }
 107
 108 static struct g_consumer *
 109 vdev_geom_attach(struct g_provider *pp, int write)
 110 {
 111         struct g_geom *gp;
 112         struct g_consumer *cp;
 113
 114         g_topology_assert();
 115
 116         ZFS_LOG(1, "Attaching to %s.", pp->name);
 117         /* Do we have geom already? No? Create one. */
 118         LIST_FOREACH(gp, &zfs_vdev_class.geom, geom) {
 119                 if (gp->flags & G_GEOM_WITHER)
 120                         continue;
 121                 if (strcmp(gp->name, "zfs::vdev") != 0)
 122                         continue;
 123                 break;
 124         }
 125         if (gp == NULL) {
 126                 gp = g_new_geomf(&zfs_vdev_class, "zfs::vdev");
 127                 gp->orphan = vdev_geom_orphan;
 128                 cp = g_new_consumer(gp);
 129                 if (g_attach(cp, pp) != 0) {
 130                         g_wither_geom(gp, ENXIO);
 131                         return (NULL);
 132                 }
 133                 if (g_access(cp, 1, write, 1) != 0) {
 134                         g_wither_geom(gp, ENXIO);
 135                         return (NULL);
 136                 }
 137                 ZFS_LOG(1, "Created geom and consumer for %s.", pp->name);
 138         } else {
 139                 /* Check if we are already connected to this provider. */
 140                 LIST_FOREACH(cp, &gp->consumer, consumer) {
 141                         if (cp->provider == pp) {
 142                                 ZFS_LOG(1, "Found consumer for %s.", pp->name);
 143                                 break;
 144                         }
 145                 }
 146                 if (cp == NULL) {
 147                         cp = g_new_consumer(gp);
 148                         if (g_attach(cp, pp) != 0) {
 149                                 g_destroy_consumer(cp);
 150                                 return (NULL);
 151                         }
 152                         if (g_access(cp, 1, write, 1) != 0) {
 153                                 g_detach(cp);
 154                                 g_destroy_consumer(cp);
 155                                 return (NULL);
 156                         }
 157                         ZFS_LOG(1, "Created consumer for %s.", pp->name);
 158                 } else {
 159                         if (g_access(cp, 1, cp->acw > 0 ? 0 : write, 1) != 0)
 160                                 return (NULL);
 161                         ZFS_LOG(1, "Used existing consumer for %s.", pp->name);
 162                 }
 163         }
 164         return (cp);
 165 }
 166
 167 static void
 168 vdev_geom_detach(void *arg, int flag __unused)
 169 {
 170         struct g_geom *gp;
 171         struct g_consumer *cp;
 172
 173         g_topology_assert();
 174         cp = arg;
 175         gp = cp->geom;
 176
 177         ZFS_LOG(1, "Closing access to %s.", cp->provider->name);
 178         g_access(cp, -1, 0, -1);
 179         /* Destroy consumer on last close. */
 180         if (cp->acr == 0 && cp->ace == 0) {
 181                 ZFS_LOG(1, "Destroyed consumer to %s.", cp->provider->name);
 182                 if (cp->acw > 0)
 183                         g_access(cp, 0, -cp->acw, 0);
 184                 g_detach(cp);
 185                 g_destroy_consumer(cp);
 186         }
 187         /* Destroy geom if there are no consumers left. */
 188         if (LIST_EMPTY(&gp->consumer)) {
 189                 ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 190                 g_wither_geom(gp, ENXIO);
 191         }
 192 }
 193
 194 static void
 195 vdev_geom_worker(void *arg)
 196 {
 197         vdev_geom_ctx_t *ctx;
 198         zio_t *zio;
 199         struct bio *bp;
 200
 201         ctx = arg;
 202         for (;;) {
 203                 mtx_lock(&ctx->gc_queue_mtx);
 204                 bp = bioq_takefirst(&ctx->gc_queue);
 205                 if (bp == NULL) {
 206                         if (ctx->gc_state == 1) {
 207                                 ctx->gc_state = 2;
 208                                 wakeup_one(&ctx->gc_state);
 209                                 mtx_unlock(&ctx->gc_queue_mtx);
 210                                 kthread_exit(0);
 211                         }
 212                         msleep(&ctx->gc_queue, &ctx->gc_queue_mtx,
 213                             PRIBIO | PDROP, "vgeom:io", 0);
 214                         continue;
 215                 }
 216                 mtx_unlock(&ctx->gc_queue_mtx);
 217                 zio = bp->bio_caller1;
 218                 zio->io_error = bp->bio_error;
 219                 if (bp->bio_cmd == BIO_FLUSH && bp->bio_error == ENOTSUP) {
 220                         vdev_t *vd;
 221
 222                         /*
 223                          * If we get ENOTSUP, we know that no future
 224                          * attempts will ever succeed.  In this case we
 225                          * set a persistent bit so that we don't bother
 226                          * with the ioctl in the future.
 227                          */
 228                         vd = zio->io_vd;
 229                         vd->vdev_nowritecache = B_TRUE;
 230                 }
 231                 g_destroy_bio(bp);
 232                 zio_next_stage_async(zio);
 233         }
 234 }
 235
 236 static char *
 237 vdev_geom_get_id(struct g_consumer *cp)
 238 {
 239         char *id;
 240         int len;
 241
 242         g_topology_assert_not();
 243         len = DISK_IDENT_SIZE;
 244         id = kmem_zalloc(len, KM_SLEEP);
 245         if (g_io_getattr("GEOM::ident", cp, &len, id) != 0) {
 246                 kmem_free(id, DISK_IDENT_SIZE);
 247                 return (NULL);
 248         }
 249         return (id);
 250 }
 251
 252 static void
 253 vdev_geom_free_id(char *id)
 254 {
 255
 256         if (id != NULL)
 257                 kmem_free(id, DISK_IDENT_SIZE);
 258 }
 259
 260 struct vdev_geom_find {
 261         const char *id;
 262         int write;
 263         struct g_consumer *cp;
 264 };
 265
 266 static void
 267 vdev_geom_taste_orphan(struct g_consumer *cp)
 268 {
 269
 270         KASSERT(1 == 0, ("%s called while tasting %s.", __func__,
 271             cp->provider->name));
 272 }
 273
 274 static void
 275 vdev_geom_attach_by_id_event(void *arg, int flags __unused)
 276 {
 277         struct vdev_geom_find *ap;
 278         struct g_class *mp;
 279         struct g_geom *gp, *zgp;
 280         struct g_provider *pp;
 281         struct g_consumer *zcp;
 282         char *id;
 283
 284         g_topology_assert();
 285
 286         ap = arg;
 287
 288         zgp = g_new_geomf(&zfs_vdev_class, "zfs::vdev::taste");
 289         /* This orphan function should be never called. */
 290         zgp->orphan = vdev_geom_taste_orphan;
 291         zcp = g_new_consumer(zgp);
 292
 293         LIST_FOREACH(mp, &g_classes, class) {
 294                 if (mp == &zfs_vdev_class)
 295                         continue;
 296                 LIST_FOREACH(gp, &mp->geom, geom) {
 297                         if (gp->flags & G_GEOM_WITHER)
 298                                 continue;
 299                         LIST_FOREACH(pp, &gp->provider, provider) {
 300                                 if (pp->flags & G_PF_WITHER)
 301                                         continue;
 302                                 g_attach(zcp, pp);
 303                                 if (g_access(zcp, 1, 0, 0) != 0) {
 304                                         g_detach(zcp);
 305                                         continue;
 306                                 }
 307                                 g_topology_unlock();
 308                                 id = vdev_geom_get_id(zcp);
 309                                 g_topology_lock();
 310                                 g_access(zcp, -1, 0, 0);
 311                                 g_detach(zcp);
 312                                 if (id == NULL || strcmp(id, ap->id) != 0) {
 313                                         vdev_geom_free_id(id);
 314                                         continue;
 315                                 }
 316                                 vdev_geom_free_id(id);
 317                                 ap->cp = vdev_geom_attach(pp, ap->write);
 318                                 if (ap->cp == NULL) {
 319                                         printf("ZFS WARNING: Cannot open %s "
 320                                             "for writting.\n", pp->name);
 321                                         continue;
 322                                 }
 323                                 goto end;
 324                         }
 325                 }
 326         }
 327         ap->cp = NULL;
 328 end:
 329         g_destroy_consumer(zcp);
 330         g_destroy_geom(zgp);
 331 }
 332
 333 static struct g_consumer *
 334 vdev_geom_attach_by_id(const char *id, int write)
 335 {
 336         struct vdev_geom_find *ap;
 337         struct g_consumer *cp;
 338
 339         ap = kmem_zalloc(sizeof(*ap), KM_SLEEP);
 340         ap->id = id;
 341         ap->write = write;
 342         g_waitfor_event(vdev_geom_attach_by_id_event, ap, M_WAITOK, NULL);
 343         cp = ap->cp;
 344         kmem_free(ap, sizeof(*ap));
 345         return (cp);
 346 }
 347
 348 static int
 349 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *ashift)
 350 {
 351         vdev_geom_ctx_t *ctx;
 352         struct g_provider *pp;
 353         struct g_consumer *cp;
 354         char *id = NULL;
 355         int owned;
 356
 357         /*
 358          * We must have a pathname, and it must be absolute.
 359          */
 360         if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
 361                 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
 362                 return (EINVAL);
 363         }
 364
 365         if ((owned = mtx_owned(&Giant)))
 366                 mtx_unlock(&Giant);
 367         cp = NULL;
 368         g_topology_lock();
 369         pp = g_provider_by_name(vd->vdev_path + sizeof("/dev/") - 1);
 370         if (pp != NULL) {
 371                 ZFS_LOG(1, "Found provider by name %s.", vd->vdev_path);
 372                 cp = vdev_geom_attach(pp, !!(spa_mode & FWRITE));
 373                 if (cp != NULL && vd->vdev_devid != NULL) {
 374                         g_topology_unlock();
 375                         id = vdev_geom_get_id(cp);
 376                         g_topology_lock();
 377                         if (id == NULL || strcmp(id, vd->vdev_devid) != 0) {
 378                                 vdev_geom_detach(cp, 0);
 379                                 cp = NULL;
 380                                 ZFS_LOG(1, "ID mismatch for provider %s: "
 381                                     "[%s]!=[%s].", vd->vdev_path,
 382                                     vd->vdev_devid, id);
 383                                 goto next;
 384                         }
 385                         ZFS_LOG(1, "ID match for provider %s.", vd->vdev_path);
 386                 }
 387         }
 388 next:
 389         g_topology_unlock();
 390         vdev_geom_free_id(id);
 391         if (cp == NULL && vd->vdev_devid != NULL) {
 392                 ZFS_LOG(1, "Searching by ID [%s].", vd->vdev_devid);
 393                 cp = vdev_geom_attach_by_id(vd->vdev_devid,
 394                     !!(spa_mode & FWRITE));
 395                 if (cp != NULL) {
 396                         size_t len = strlen(cp->provider->name) + 6; /* 6 == strlen("/dev/") + 1 */
 397                         char *buf = kmem_alloc(len, KM_SLEEP);
 398
 399                         snprintf(buf, len, "/dev/%s", cp->provider->name);
 400                         spa_strfree(vd->vdev_path);
 401                         vd->vdev_path = buf;
 402
 403                         ZFS_LOG(1, "Attach by ID [%s] succeeded, provider %s.",
 404                             vd->vdev_devid, vd->vdev_path);
 405                 }
 406         }
 407         if (owned)
 408                 mtx_lock(&Giant);
 409         if (cp == NULL) {
 410                 ZFS_LOG(1, "Provider %s (id=[%s]) not found.", vd->vdev_path,
 411                     vd->vdev_devid);
 412                 vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 413                 return (EACCES);
 414         }
 415         pp = cp->provider;
 416
 417         /*
 418          * Determine the actual size of the device.
 419          */
 420         *psize = pp->mediasize;
 421
 422         /*
 423          * Determine the device's minimum transfer size.
 424          */
 425         *ashift = highbit(MAX(pp->sectorsize, SPA_MINBLOCKSIZE)) - 1;
 426
 427         /*
 428          * Clear the nowritecache bit, so that on a vdev_reopen() we will
 429          * try again.
 430          */
 431         vd->vdev_nowritecache = B_FALSE;
 432
 433         cp->private = vd;
 434
 435         ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP);
 436         bioq_init(&ctx->gc_queue);
 437         mtx_init(&ctx->gc_queue_mtx, "zfs:vdev:geom:queue", NULL, MTX_DEF);
 438         ctx->gc_consumer = cp;
 439         ctx->gc_state = 0;
 440
 441         vd->vdev_tsd = ctx;
 442
 443         kthread_create(vdev_geom_worker, ctx, NULL, 0, 0, "vdev:worker %s",
 444             pp->name);
 445
 446         return (0);
 447 }
 448
 449 static void
 450 vdev_geom_close(vdev_t *vd)
 451 {
 452         vdev_geom_ctx_t *ctx;
 453         struct g_consumer *cp;
 454
 455         if ((ctx = vd->vdev_tsd) == NULL)
 456                 return;
 457         if ((cp = ctx->gc_consumer) == NULL)
 458                 return;
 459         vdev_geom_release(vd);
 460         g_post_event(vdev_geom_detach, cp, M_WAITOK, NULL);
 461 }
 462
 463 static void
 464 vdev_geom_io_intr(struct bio *bp)
 465 {
 466         vdev_geom_ctx_t *ctx;
 467         zio_t *zio;
 468
 469         zio = bp->bio_caller1;
 470         ctx = zio->io_vd->vdev_tsd;
 471
 472         mtx_lock(&ctx->gc_queue_mtx);
 473         bioq_insert_tail(&ctx->gc_queue, bp);
 474         wakeup_one(&ctx->gc_queue);
 475         mtx_unlock(&ctx->gc_queue_mtx);
 476 }
 477
 478 static void
 479 vdev_geom_io_start(zio_t *zio)
 480 {
 481         vdev_t *vd;
 482         vdev_geom_ctx_t *ctx;
 483         struct g_consumer *cp;
 484         struct bio *bp;
 485         int error;
 486
 487         cp = NULL;
 488
 489         vd = zio->io_vd;
 490         ctx = vd->vdev_tsd;
 491         if (ctx != NULL)
 492                 cp = ctx->gc_consumer;
 493
 494         if (zio->io_type == ZIO_TYPE_IOCTL) {
 495                 zio_vdev_io_bypass(zio);
 496
 497                 /* XXPOLICY */
 498                 if (vdev_is_dead(vd)) {
 499                         zio->io_error = ENXIO;
 500                         zio_next_stage_async(zio);
 501                         return;
 502                 }
 503
 504                 switch (zio->io_cmd) {
 505
 506                 case DKIOCFLUSHWRITECACHE:
 507                         if (vd->vdev_nowritecache) {
 508                                 zio->io_error = ENOTSUP;
 509                                 break;
 510                         }
 511
 512                         goto sendreq;
 513                 default:
 514                         zio->io_error = ENOTSUP;
 515                 }
 516
 517                 zio_next_stage_async(zio);
 518                 return;
 519         }
 520
 521         if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
 522                 return;
 523
 524         if ((zio = vdev_queue_io(zio)) == NULL)
 525                 return;
 526
 527 sendreq:
 528
 529         error = vdev_is_dead(vd) ? ENXIO : vdev_error_inject(vd, zio);
 530         if (error == 0 && cp == NULL)
 531                 error = ENXIO;
 532         if (error) {
 533                 zio->io_error = error;
 534                 zio_next_stage_async(zio);
 535                 return;
 536         }
 537
 538         bp = g_alloc_bio();
 539         bp->bio_caller1 = zio;
 540         switch (zio->io_type) {
 541         case ZIO_TYPE_READ:
 542         case ZIO_TYPE_WRITE:
 543                 bp->bio_cmd = zio->io_type == ZIO_TYPE_READ ? BIO_READ : BIO_WRITE;
 544                 bp->bio_data = zio->io_data;
 545                 bp->bio_offset = zio->io_offset;
 546                 bp->bio_length = zio->io_size;
 547                 break;
 548         case ZIO_TYPE_IOCTL:
 549                 bp->bio_cmd = BIO_FLUSH;
 550                 bp->bio_data = NULL;
 551                 bp->bio_offset = cp->provider->mediasize;
 552                 bp->bio_length = 0;
 553                 break;
 554         }
 555         bp->bio_done = vdev_geom_io_intr;
 556
 557         g_io_request(bp, cp);
 558 }
 559
 560 static void
 561 vdev_geom_io_done(zio_t *zio)
 562 {
 563         vdev_queue_io_done(zio);
 564
 565         if (zio->io_type == ZIO_TYPE_WRITE)
 566                 vdev_cache_write(zio);
 567
 568         if (zio_injection_enabled && zio->io_error == 0)
 569                 zio->io_error = zio_handle_device_injection(zio->io_vd, EIO);
 570
 571         zio_next_stage(zio);
 572 }
 573
 574 vdev_ops_t vdev_geom_ops = {
 575         vdev_geom_open,
 576         vdev_geom_close,
 577         vdev_default_asize,
 578         vdev_geom_io_start,
 579         vdev_geom_io_done,
 580         NULL,
 581         VDEV_TYPE_DISK,         /* name of this vdev type */
 582         B_TRUE                  /* leaf vdev */
 583 };