cmd/zpool/zpool_vdev.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 /*
  23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  24  */
  25
  26 /*
  27  * Functions to convert between a list of vdevs and an nvlist representing the
  28  * configuration.  Each entry in the list can be one of:
  29  *
  30  *      Device vdevs
  31  *              disk=(path=..., devid=...)
  32  *              file=(path=...)
  33  *
  34  *      Group vdevs
  35  *              raidz[1|2]=(...)
  36  *              mirror=(...)
  37  *
  38  *      Hot spares
  39  *
  40  * While the underlying implementation supports it, group vdevs cannot contain
  41  * other group vdevs.  All userland verification of devices is contained within
  42  * this file.  If successful, the nvlist returned can be passed directly to the
  43  * kernel; we've done as much verification as possible in userland.
  44  *
  45  * Hot spares are a special case, and passed down as an array of disk vdevs, at
  46  * the same level as the root of the vdev tree.
  47  *
  48  * The only function exported by this file is 'make_root_vdev'.  The
  49  * function performs several passes:
  50  *
  51  *      1. Construct the vdev specification.  Performs syntax validation and
  52  *         makes sure each device is valid.
  53  *      2. Check for devices in use.  Using libblkid to make sure that no
  54  *         devices are also in use.  Some can be overridden using the 'force'
  55  *         flag, others cannot.
  56  *      3. Check for replication errors if the 'force' flag is not specified.
  57  *         validates that the replication level is consistent across the
  58  *         entire pool.
  59  *      4. Call libzfs to label any whole disks with an EFI label.
  60  */
  61
  62 #include <assert.h>
  63 #include <ctype.h>
  64 #include <devid.h>
  65 #include <errno.h>
  66 #include <fcntl.h>
  67 #include <libintl.h>
  68 #include <libnvpair.h>
  69 #include <limits.h>
  70 #include <stdio.h>
  71 #include <string.h>
  72 #include <unistd.h>
  73 #include <sys/efi_partition.h>
  74 #include <sys/stat.h>
  75 #include <sys/vtoc.h>
  76 #include <sys/mntent.h>
  77 #include <uuid/uuid.h>
  78 #ifdef HAVE_LIBBLKID
  79 #include <blkid/blkid.h>
  80 #else
  81 #define blkid_cache void *
  82 #endif /* HAVE_LIBBLKID */
  83
  84 #include "zpool_util.h"
  85
  86 /*
  87  * For any given vdev specification, we can have multiple errors.  The
  88  * vdev_error() function keeps track of whether we have seen an error yet, and
  89  * prints out a header if its the first error we've seen.
  90  */
  91 boolean_t error_seen;
  92 boolean_t is_force;
  93
  94 /*PRINTFLIKE1*/
  95 static void
  96 vdev_error(const char *fmt, ...)
  97 {
  98         va_list ap;
  99
 100         if (!error_seen) {
 101                 (void) fprintf(stderr, gettext("invalid vdev specification\n"));
 102                 if (!is_force)
 103                         (void) fprintf(stderr, gettext("use '-f' to override "
 104                             "the following errors:\n"));
 105                 else
 106                         (void) fprintf(stderr, gettext("the following errors "
 107                             "must be manually repaired:\n"));
 108                 error_seen = B_TRUE;
 109         }
 110
 111         va_start(ap, fmt);
 112         (void) vfprintf(stderr, fmt, ap);
 113         va_end(ap);
 114 }
 115
 116 /*
 117  * Check that a file is valid.  All we can do in this case is check that it's
 118  * not in use by another pool, and not in use by swap.
 119  */
 120 static int
 121 check_file(const char *file, boolean_t force, boolean_t isspare)
 122 {
 123         char  *name;
 124         int fd;
 125         int ret = 0;
 126         pool_state_t state;
 127         boolean_t inuse;
 128
 129         if ((fd = open(file, O_RDONLY)) < 0)
 130                 return (0);
 131
 132         if (zpool_in_use(g_zfs, fd, &state, &name, &inuse) == 0 && inuse) {
 133                 const char *desc;
 134
 135                 switch (state) {
 136                 case POOL_STATE_ACTIVE:
 137                         desc = gettext("active");
 138                         break;
 139
 140                 case POOL_STATE_EXPORTED:
 141                         desc = gettext("exported");
 142                         break;
 143
 144                 case POOL_STATE_POTENTIALLY_ACTIVE:
 145                         desc = gettext("potentially active");
 146                         break;
 147
 148                 default:
 149                         desc = gettext("unknown");
 150                         break;
 151                 }
 152
 153                 /*
 154                  * Allow hot spares to be shared between pools.
 155                  */
 156                 if (state == POOL_STATE_SPARE && isspare)
 157                         return (0);
 158
 159                 if (state == POOL_STATE_ACTIVE ||
 160                     state == POOL_STATE_SPARE || !force) {
 161                         switch (state) {
 162                         case POOL_STATE_SPARE:
 163                                 vdev_error(gettext("%s is reserved as a hot "
 164                                     "spare for pool %s\n"), file, name);
 165                                 break;
 166                         default:
 167                                 vdev_error(gettext("%s is part of %s pool "
 168                                     "'%s'\n"), file, desc, name);
 169                                 break;
 170                         }
 171                         ret = -1;
 172                 }
 173
 174                 free(name);
 175         }
 176
 177         (void) close(fd);
 178         return (ret);
 179 }
 180
 181 static void
 182 check_error(int err)
 183 {
 184         (void) fprintf(stderr, gettext("warning: device in use checking "
 185             "failed: %s\n"), strerror(err));
 186 }
 187
 188 static int
 189 check_slice(const char *path, blkid_cache cache, int force, boolean_t isspare)
 190 {
 191         int err;
 192 #ifdef HAVE_LIBBLKID
 193         char *value;
 194
 195         /* No valid type detected device is safe to use */
 196         value = blkid_get_tag_value(cache, "TYPE", path);
 197         if (value == NULL)
 198                 return (0);
 199
 200         /*
 201          * If libblkid detects a ZFS device, we check the device
 202          * using check_file() to see if it's safe.  The one safe
 203          * case is a spare device shared between multiple pools.
 204          */
 205         if (strcmp(value, "zfs") == 0) {
 206                 err = check_file(path, force, isspare);
 207         } else {
 208                 if (force) {
 209                         err = 0;
 210                 } else {
 211                         err = -1;
 212                         vdev_error(gettext("%s contains a filesystem of "
 213                                    "type '%s'\n"), path, value);
 214                 }
 215         }
 216
 217         free(value);
 218 #else
 219         err = check_file(path, force, isspare);
 220 #endif /* HAVE_LIBBLKID */
 221
 222         return (err);
 223 }
 224
 225 /*
 226  * Validate a whole disk.  Iterate over all slices on the disk and make sure
 227  * that none is in use by calling check_slice().
 228  */
 229 static int
 230 check_disk(const char *path, blkid_cache cache, int force,
 231            boolean_t isspare, boolean_t iswholedisk)
 232 {
 233         struct dk_gpt *vtoc;
 234         char slice_path[MAXPATHLEN];
 235         int err = 0;
 236         int fd, i;
 237
 238         /* This is not a wholedisk we only check the given partition */
 239         if (!iswholedisk)
 240                 return check_slice(path, cache, force, isspare);
 241
 242         /*
 243          * When the device is a whole disk try to read the efi partition
 244          * label.  If this is successful we safely check the all of the
 245          * partitions.  However, when it fails it may simply be because
 246          * the disk is partitioned via the MBR.  Since we currently can
 247          * not easily decode the MBR return a failure and prompt to the
 248          * user to use force option since we cannot check the partitions.
 249          */
 250         if ((fd = open(path, O_RDONLY|O_DIRECT)) < 0) {
 251                 check_error(errno);
 252                 return -1;
 253         }
 254
 255         if ((err = efi_alloc_and_read(fd, &vtoc)) != 0) {
 256                 (void) close(fd);
 257
 258                 if (force) {
 259                         return 0;
 260                 } else {
 261                         vdev_error(gettext("%s does not contain an EFI "
 262                             "label but it may contain partition\n"
 263                             "information in the MBR.\n"), path);
 264                         return -1;
 265                 }
 266         }
 267
 268         /*
 269          * The primary efi partition label is damaged however the secondary
 270          * label at the end of the device is intact.  Rather than use this
 271          * label we should play it safe and treat this as a non efi device.
 272          */
 273         if (vtoc->efi_flags & EFI_GPT_PRIMARY_CORRUPT) {
 274                 efi_free(vtoc);
 275                 (void) close(fd);
 276
 277                 if (force) {
 278                         /* Partitions will no be created using the backup */
 279                         return 0;
 280                 } else {
 281                         vdev_error(gettext("%s contains a corrupt primary "
 282                             "EFI label.\n"), path);
 283                         return -1;
 284                 }
 285         }
 286
 287         for (i = 0; i < vtoc->efi_nparts; i++) {
 288
 289                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED ||
 290                     uuid_is_null((uchar_t *)&vtoc->efi_parts[i].p_guid))
 291                         continue;
 292
 293                 if (strncmp(path, UDISK_ROOT, strlen(UDISK_ROOT)) == 0)
 294                         (void) snprintf(slice_path, sizeof (slice_path),
 295                             "%s%s%d", path, "-part", i+1);
 296                 else
 297                         (void) snprintf(slice_path, sizeof (slice_path),
 298                             "%s%s%d", path, isdigit(path[strlen(path)-1]) ?
 299                             "p" : "", i+1);
 300
 301                 err = check_slice(slice_path, cache, force, isspare);
 302                 if (err)
 303                         break;
 304         }
 305
 306         efi_free(vtoc);
 307         (void) close(fd);
 308
 309         return (err);
 310 }
 311
 312 static int
 313 check_device(const char *path, boolean_t force,
 314              boolean_t isspare, boolean_t iswholedisk)
 315 {
 316         static blkid_cache cache = NULL;
 317
 318 #ifdef HAVE_LIBBLKID
 319         /*
 320          * There is no easy way to add a correct blkid_put_cache() call,
 321          * memory will be reclaimed when the command exits.
 322          */
 323         if (cache == NULL) {
 324                 int err;
 325
 326                 if ((err = blkid_get_cache(&cache, NULL)) != 0) {
 327                         check_error(err);
 328                         return -1;
 329                 }
 330
 331                 if ((err = blkid_probe_all(cache)) != 0) {
 332                         blkid_put_cache(cache);
 333                         check_error(err);
 334                         return -1;
 335                 }
 336         }
 337 #endif /* HAVE_LIBBLKID */
 338
 339         return check_disk(path, cache, force, isspare, iswholedisk);
 340 }
 341
 342 /*
 343  * By "whole disk" we mean an entire physical disk (something we can
 344  * label, toggle the write cache on, etc.) as opposed to the full
 345  * capacity of a pseudo-device such as lofi or did.  We act as if we
 346  * are labeling the disk, which should be a pretty good test of whether
 347  * it's a viable device or not.  Returns B_TRUE if it is and B_FALSE if
 348  * it isn't.
 349  */
 350 static boolean_t
 351 is_whole_disk(const char *path)
 352 {
 353         struct dk_gpt *label;
 354         int     fd;
 355
 356         if ((fd = open(path, O_RDONLY|O_DIRECT)) < 0)
 357                 return (B_FALSE);
 358         if (efi_alloc_and_init(fd, EFI_NUMPAR, &label) != 0) {
 359                 (void) close(fd);
 360                 return (B_FALSE);
 361         }
 362         efi_free(label);
 363         (void) close(fd);
 364         return (B_TRUE);
 365 }
 366
 367 /*
 368  * This may be a shorthand device path or it could be total gibberish.
 369  * Check to see if it is a known device available in zfs_vdev_paths.
 370  * As part of this check, see if we've been given an entire disk
 371  * (minus the slice number).
 372  */
 373 static int
 374 is_shorthand_path(const char *arg, char *path,
 375                   struct stat64 *statbuf, boolean_t *wholedisk)
 376 {
 377         int error;
 378
 379         error = zfs_resolve_shortname(arg, path, MAXPATHLEN);
 380         if (error == 0) {
 381                 *wholedisk = is_whole_disk(path);
 382                 if (*wholedisk || (stat64(path, statbuf) == 0))
 383                         return (0);
 384         }
 385
 386         strlcpy(path, arg, sizeof(path));
 387         memset(statbuf, 0, sizeof(*statbuf));
 388         *wholedisk = B_FALSE;
 389
 390         return (error);
 391 }
 392
 393 /*
 394  * Determine if the given path is a hot spare within the given configuration.
 395  * If no configuration is given we rely solely on the label.
 396  */
 397 static boolean_t
 398 is_spare(nvlist_t *config, const char *path)
 399 {
 400         int fd;
 401         pool_state_t state;
 402         char *name = NULL;
 403         nvlist_t *label;
 404         uint64_t guid, spareguid;
 405         nvlist_t *nvroot;
 406         nvlist_t **spares;
 407         uint_t i, nspares;
 408         boolean_t inuse;
 409
 410         if ((fd = open(path, O_RDONLY)) < 0)
 411                 return (B_FALSE);
 412
 413         if (zpool_in_use(g_zfs, fd, &state, &name, &inuse) != 0 ||
 414             !inuse ||
 415             state != POOL_STATE_SPARE ||
 416             zpool_read_label(fd, &label) != 0) {
 417                 free(name);
 418                 (void) close(fd);
 419                 return (B_FALSE);
 420         }
 421         free(name);
 422         (void) close(fd);
 423
 424         if (config == NULL)
 425                 return (B_TRUE);
 426
 427         verify(nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) == 0);
 428         nvlist_free(label);
 429
 430         verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
 431             &nvroot) == 0);
 432         if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
 433             &spares, &nspares) == 0) {
 434                 for (i = 0; i < nspares; i++) {
 435                         verify(nvlist_lookup_uint64(spares[i],
 436                             ZPOOL_CONFIG_GUID, &spareguid) == 0);
 437                         if (spareguid == guid)
 438                                 return (B_TRUE);
 439                 }
 440         }
 441
 442         return (B_FALSE);
 443 }
 444
 445 /*
 446  * Create a leaf vdev.  Determine if this is a file or a device.  If it's a
 447  * device, fill in the device id to make a complete nvlist.  Valid forms for a
 448  * leaf vdev are:
 449  *
 450  *      /dev/xxx        Complete disk path
 451  *      /xxx            Full path to file
 452  *      xxx             Shorthand for <zfs_vdev_paths>/xxx
 453  */
 454 static nvlist_t *
 455 make_leaf_vdev(nvlist_t *props, const char *arg, uint64_t is_log)
 456 {
 457         char path[MAXPATHLEN];
 458         struct stat64 statbuf;
 459         nvlist_t *vdev = NULL;
 460         char *type = NULL;
 461         boolean_t wholedisk = B_FALSE;
 462         int err;
 463
 464         /*
 465          * Determine what type of vdev this is, and put the full path into
 466          * 'path'.  We detect whether this is a device of file afterwards by
 467          * checking the st_mode of the file.
 468          */
 469         if (arg[0] == '/') {
 470                 /*
 471                  * Complete device or file path.  Exact type is determined by
 472                  * examining the file descriptor afterwards.  Symbolic links
 473                  * are resolved to their real paths for the is_whole_disk()
 474                  * and S_ISBLK/S_ISREG type checks.  However, we are careful
 475                  * to store the given path as ZPOOL_CONFIG_PATH to ensure we
 476                  * can leverage udev's persistent device labels.
 477                  */
 478                 if (realpath(arg, path) == NULL) {
 479                         (void) fprintf(stderr,
 480                             gettext("cannot resolve path '%s'\n"), arg);
 481                         return (NULL);
 482                 }
 483
 484                 wholedisk = is_whole_disk(path);
 485                 if (!wholedisk && (stat64(path, &statbuf) != 0)) {
 486                         (void) fprintf(stderr,
 487                             gettext("cannot open '%s': %s\n"),
 488                             path, strerror(errno));
 489                         return (NULL);
 490                 }
 491
 492                 /* After is_whole_disk() check restore original passed path */
 493                 strlcpy(path, arg, MAXPATHLEN);
 494         } else {
 495                 err = is_shorthand_path(arg, path, &statbuf, &wholedisk);
 496                 if (err != 0) {
 497                         /*
 498                          * If we got ENOENT, then the user gave us
 499                          * gibberish, so try to direct them with a
 500                          * reasonable error message.  Otherwise,
 501                          * regurgitate strerror() since it's the best we
 502                          * can do.
 503                          */
 504                         if (err == ENOENT) {
 505                                 (void) fprintf(stderr,
 506                                     gettext("cannot open '%s': no such "
 507                                     "device in %s\n"), arg, DISK_ROOT);
 508                                 (void) fprintf(stderr,
 509                                     gettext("must be a full path or "
 510                                     "shorthand device name\n"));
 511                                 return (NULL);
 512                         } else {
 513                                 (void) fprintf(stderr,
 514                                     gettext("cannot open '%s': %s\n"),
 515                                     path, strerror(errno));
 516                                 return (NULL);
 517                         }
 518                 }
 519         }
 520
 521         /*
 522          * Determine whether this is a device or a file.
 523          */
 524         if (wholedisk || S_ISBLK(statbuf.st_mode)) {
 525                 type = VDEV_TYPE_DISK;
 526         } else if (S_ISREG(statbuf.st_mode)) {
 527                 type = VDEV_TYPE_FILE;
 528         } else {
 529                 (void) fprintf(stderr, gettext("cannot use '%s': must be a "
 530                     "block device or regular file\n"), path);
 531                 return (NULL);
 532         }
 533
 534         /*
 535          * Finally, we have the complete device or file, and we know that it is
 536          * acceptable to use.  Construct the nvlist to describe this vdev.  All
 537          * vdevs have a 'path' element, and devices also have a 'devid' element.
 538          */
 539         verify(nvlist_alloc(&vdev, NV_UNIQUE_NAME, 0) == 0);
 540         verify(nvlist_add_string(vdev, ZPOOL_CONFIG_PATH, path) == 0);
 541         verify(nvlist_add_string(vdev, ZPOOL_CONFIG_TYPE, type) == 0);
 542         verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_IS_LOG, is_log) == 0);
 543         if (strcmp(type, VDEV_TYPE_DISK) == 0)
 544                 verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_WHOLE_DISK,
 545                     (uint64_t)wholedisk) == 0);
 546
 547         if (props != NULL) {
 548                 uint64_t ashift = 0;
 549                 char *value = NULL;
 550
 551                 if (nvlist_lookup_string(props,
 552                     zpool_prop_to_name(ZPOOL_PROP_ASHIFT), &value) == 0)
 553                         zfs_nicestrtonum(NULL, value, &ashift);
 554
 555                 if (ashift > 0)
 556                         verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_ASHIFT,
 557                             ashift) == 0);
 558         }
 559
 560         return (vdev);
 561 }
 562
 563 /*
 564  * Go through and verify the replication level of the pool is consistent.
 565  * Performs the following checks:
 566  *
 567  *      For the new spec, verifies that devices in mirrors and raidz are the
 568  *      same size.
 569  *
 570  *      If the current configuration already has inconsistent replication
 571  *      levels, ignore any other potential problems in the new spec.
 572  *
 573  *      Otherwise, make sure that the current spec (if there is one) and the new
 574  *      spec have consistent replication levels.
 575  */
 576 typedef struct replication_level {
 577         char *zprl_type;
 578         uint64_t zprl_children;
 579         uint64_t zprl_parity;
 580 } replication_level_t;
 581
 582 #define ZPOOL_FUZZ      (16 * 1024 * 1024)
 583
 584 /*
 585  * Given a list of toplevel vdevs, return the current replication level.  If
 586  * the config is inconsistent, then NULL is returned.  If 'fatal' is set, then
 587  * an error message will be displayed for each self-inconsistent vdev.
 588  */
 589 static replication_level_t *
 590 get_replication(nvlist_t *nvroot, boolean_t fatal)
 591 {
 592         nvlist_t **top;
 593         uint_t t, toplevels;
 594         nvlist_t **child;
 595         uint_t c, children;
 596         nvlist_t *nv;
 597         char *type;
 598         replication_level_t lastrep = { 0 }, rep, *ret;
 599         boolean_t dontreport;
 600
 601         ret = safe_malloc(sizeof (replication_level_t));
 602
 603         verify(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
 604             &top, &toplevels) == 0);
 605
 606         lastrep.zprl_type = NULL;
 607         for (t = 0; t < toplevels; t++) {
 608                 uint64_t is_log = B_FALSE;
 609
 610                 nv = top[t];
 611
 612                 /*
 613                  * For separate logs we ignore the top level vdev replication
 614                  * constraints.
 615                  */
 616                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &is_log);
 617                 if (is_log)
 618                         continue;
 619
 620                 verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE,
 621                     &type) == 0);
 622                 if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 623                     &child, &children) != 0) {
 624                         /*
 625                          * This is a 'file' or 'disk' vdev.
 626                          */
 627                         rep.zprl_type = type;
 628                         rep.zprl_children = 1;
 629                         rep.zprl_parity = 0;
 630                 } else {
 631                         uint64_t vdev_size;
 632
 633                         /*
 634                          * This is a mirror or RAID-Z vdev.  Go through and make
 635                          * sure the contents are all the same (files vs. disks),
 636                          * keeping track of the number of elements in the
 637                          * process.
 638                          *
 639                          * We also check that the size of each vdev (if it can
 640                          * be determined) is the same.
 641                          */
 642                         rep.zprl_type = type;
 643                         rep.zprl_children = 0;
 644
 645                         if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
 646                                 verify(nvlist_lookup_uint64(nv,
 647                                     ZPOOL_CONFIG_NPARITY,
 648                                     &rep.zprl_parity) == 0);
 649                                 assert(rep.zprl_parity != 0);
 650                         } else {
 651                                 rep.zprl_parity = 0;
 652                         }
 653
 654                         /*
 655                          * The 'dontreport' variable indicates that we've
 656                          * already reported an error for this spec, so don't
 657                          * bother doing it again.
 658                          */
 659                         type = NULL;
 660                         dontreport = 0;
 661                         vdev_size = -1ULL;
 662                         for (c = 0; c < children; c++) {
 663                                 nvlist_t *cnv = child[c];
 664                                 char *path;
 665                                 struct stat64 statbuf;
 666                                 uint64_t size = -1ULL;
 667                                 char *childtype;
 668                                 int fd, err;
 669
 670                                 rep.zprl_children++;
 671
 672                                 verify(nvlist_lookup_string(cnv,
 673                                     ZPOOL_CONFIG_TYPE, &childtype) == 0);
 674
 675                                 /*
 676                                  * If this is a replacing or spare vdev, then
 677                                  * get the real first child of the vdev.
 678                                  */
 679                                 if (strcmp(childtype,
 680                                     VDEV_TYPE_REPLACING) == 0 ||
 681                                     strcmp(childtype, VDEV_TYPE_SPARE) == 0) {
 682                                         nvlist_t **rchild;
 683                                         uint_t rchildren;
 684
 685                                         verify(nvlist_lookup_nvlist_array(cnv,
 686                                             ZPOOL_CONFIG_CHILDREN, &rchild,
 687                                             &rchildren) == 0);
 688                                         assert(rchildren == 2);
 689                                         cnv = rchild[0];
 690
 691                                         verify(nvlist_lookup_string(cnv,
 692                                             ZPOOL_CONFIG_TYPE,
 693                                             &childtype) == 0);
 694                                 }
 695
 696                                 verify(nvlist_lookup_string(cnv,
 697                                     ZPOOL_CONFIG_PATH, &path) == 0);
 698
 699                                 /*
 700                                  * If we have a raidz/mirror that combines disks
 701                                  * with files, report it as an error.
 702                                  */
 703                                 if (!dontreport && type != NULL &&
 704                                     strcmp(type, childtype) != 0) {
 705                                         if (ret != NULL)
 706                                                 free(ret);
 707                                         ret = NULL;
 708                                         if (fatal)
 709                                                 vdev_error(gettext(
 710                                                     "mismatched replication "
 711                                                     "level: %s contains both "
 712                                                     "files and devices\n"),
 713                                                     rep.zprl_type);
 714                                         else
 715                                                 return (NULL);
 716                                         dontreport = B_TRUE;
 717                                 }
 718
 719                                 /*
 720                                  * According to stat(2), the value of 'st_size'
 721                                  * is undefined for block devices and character
 722                                  * devices.  But there is no effective way to
 723                                  * determine the real size in userland.
 724                                  *
 725                                  * Instead, we'll take advantage of an
 726                                  * implementation detail of spec_size().  If the
 727                                  * device is currently open, then we (should)
 728                                  * return a valid size.
 729                                  *
 730                                  * If we still don't get a valid size (indicated
 731                                  * by a size of 0 or MAXOFFSET_T), then ignore
 732                                  * this device altogether.
 733                                  */
 734                                 if ((fd = open(path, O_RDONLY)) >= 0) {
 735                                         err = fstat64(fd, &statbuf);
 736                                         (void) close(fd);
 737                                 } else {
 738                                         err = stat64(path, &statbuf);
 739                                 }
 740
 741                                 if (err != 0 ||
 742                                     statbuf.st_size == 0 ||
 743                                     statbuf.st_size == MAXOFFSET_T)
 744                                         continue;
 745
 746                                 size = statbuf.st_size;
 747
 748                                 /*
 749                                  * Also make sure that devices and
 750                                  * slices have a consistent size.  If
 751                                  * they differ by a significant amount
 752                                  * (~16MB) then report an error.
 753                                  */
 754                                 if (!dontreport &&
 755                                     (vdev_size != -1ULL &&
 756                                     (labs(size - vdev_size) >
 757                                     ZPOOL_FUZZ))) {
 758                                         if (ret != NULL)
 759                                                 free(ret);
 760                                         ret = NULL;
 761                                         if (fatal)
 762                                                 vdev_error(gettext(
 763                                                     "%s contains devices of "
 764                                                     "different sizes\n"),
 765                                                     rep.zprl_type);
 766                                         else
 767                                                 return (NULL);
 768                                         dontreport = B_TRUE;
 769                                 }
 770
 771                                 type = childtype;
 772                                 vdev_size = size;
 773                         }
 774                 }
 775
 776                 /*
 777                  * At this point, we have the replication of the last toplevel
 778                  * vdev in 'rep'.  Compare it to 'lastrep' to see if its
 779                  * different.
 780                  */
 781                 if (lastrep.zprl_type != NULL) {
 782                         if (strcmp(lastrep.zprl_type, rep.zprl_type) != 0) {
 783                                 if (ret != NULL)
 784                                         free(ret);
 785                                 ret = NULL;
 786                                 if (fatal)
 787                                         vdev_error(gettext(
 788                                             "mismatched replication level: "
 789                                             "both %s and %s vdevs are "
 790                                             "present\n"),
 791                                             lastrep.zprl_type, rep.zprl_type);
 792                                 else
 793                                         return (NULL);
 794                         } else if (lastrep.zprl_parity != rep.zprl_parity) {
 795                                 if (ret)
 796                                         free(ret);
 797                                 ret = NULL;
 798                                 if (fatal)
 799                                         vdev_error(gettext(
 800                                             "mismatched replication level: "
 801                                             "both %llu and %llu device parity "
 802                                             "%s vdevs are present\n"),
 803                                             lastrep.zprl_parity,
 804                                             rep.zprl_parity,
 805                                             rep.zprl_type);
 806                                 else
 807                                         return (NULL);
 808                         } else if (lastrep.zprl_children != rep.zprl_children) {
 809                                 if (ret)
 810                                         free(ret);
 811                                 ret = NULL;
 812                                 if (fatal)
 813                                         vdev_error(gettext(
 814                                             "mismatched replication level: "
 815                                             "both %llu-way and %llu-way %s "
 816                                             "vdevs are present\n"),
 817                                             lastrep.zprl_children,
 818                                             rep.zprl_children,
 819                                             rep.zprl_type);
 820                                 else
 821                                         return (NULL);
 822                         }
 823                 }
 824                 lastrep = rep;
 825         }
 826
 827         if (ret != NULL)
 828                 *ret = rep;
 829
 830         return (ret);
 831 }
 832
 833 /*
 834  * Check the replication level of the vdev spec against the current pool.  Calls
 835  * get_replication() to make sure the new spec is self-consistent.  If the pool
 836  * has a consistent replication level, then we ignore any errors.  Otherwise,
 837  * report any difference between the two.
 838  */
 839 static int
 840 check_replication(nvlist_t *config, nvlist_t *newroot)
 841 {
 842         nvlist_t **child;
 843         uint_t  children;
 844         replication_level_t *current = NULL, *new;
 845         int ret;
 846
 847         /*
 848          * If we have a current pool configuration, check to see if it's
 849          * self-consistent.  If not, simply return success.
 850          */
 851         if (config != NULL) {
 852                 nvlist_t *nvroot;
 853
 854                 verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
 855                     &nvroot) == 0);
 856                 if ((current = get_replication(nvroot, B_FALSE)) == NULL)
 857                         return (0);
 858         }
 859         /*
 860          * for spares there may be no children, and therefore no
 861          * replication level to check
 862          */
 863         if ((nvlist_lookup_nvlist_array(newroot, ZPOOL_CONFIG_CHILDREN,
 864             &child, &children) != 0) || (children == 0)) {
 865                 free(current);
 866                 return (0);
 867         }
 868
 869         /*
 870          * If all we have is logs then there's no replication level to check.
 871          */
 872         if (num_logs(newroot) == children) {
 873                 free(current);
 874                 return (0);
 875         }
 876
 877         /*
 878          * Get the replication level of the new vdev spec, reporting any
 879          * inconsistencies found.
 880          */
 881         if ((new = get_replication(newroot, B_TRUE)) == NULL) {
 882                 free(current);
 883                 return (-1);
 884         }
 885
 886         /*
 887          * Check to see if the new vdev spec matches the replication level of
 888          * the current pool.
 889          */
 890         ret = 0;
 891         if (current != NULL) {
 892                 if (strcmp(current->zprl_type, new->zprl_type) != 0) {
 893                         vdev_error(gettext(
 894                             "mismatched replication level: pool uses %s "
 895                             "and new vdev is %s\n"),
 896                             current->zprl_type, new->zprl_type);
 897                         ret = -1;
 898                 } else if (current->zprl_parity != new->zprl_parity) {
 899                         vdev_error(gettext(
 900                             "mismatched replication level: pool uses %llu "
 901                             "device parity and new vdev uses %llu\n"),
 902                             current->zprl_parity, new->zprl_parity);
 903                         ret = -1;
 904                 } else if (current->zprl_children != new->zprl_children) {
 905                         vdev_error(gettext(
 906                             "mismatched replication level: pool uses %llu-way "
 907                             "%s and new vdev uses %llu-way %s\n"),
 908                             current->zprl_children, current->zprl_type,
 909                             new->zprl_children, new->zprl_type);
 910                         ret = -1;
 911                 }
 912         }
 913
 914         free(new);
 915         if (current != NULL)
 916                 free(current);
 917
 918         return (ret);
 919 }
 920
 921 static int
 922 zero_label(char *path)
 923 {
 924         const int size = 4096;
 925         char buf[size];
 926         int err, fd;
 927
 928         if ((fd = open(path, O_WRONLY|O_EXCL)) < 0) {
 929                 (void) fprintf(stderr, gettext("cannot open '%s': %s\n"),
 930                     path, strerror(errno));
 931                 return (-1);
 932         }
 933
 934         memset(buf, 0, size);
 935         err = write(fd, buf, size);
 936         (void) fdatasync(fd);
 937         (void) close(fd);
 938
 939         if (err == -1) {
 940                 (void) fprintf(stderr, gettext("cannot zero first %d bytes "
 941                     "of '%s': %s\n"), size, path, strerror(errno));
 942                 return (-1);
 943         }
 944
 945         if (err != size) {
 946                 (void) fprintf(stderr, gettext("could only zero %d/%d bytes "
 947                     "of '%s'\n"), err, size, path);
 948                 return (-1);
 949         }
 950
 951         return 0;
 952 }
 953
 954 /*
 955  * Go through and find any whole disks in the vdev specification, labelling them
 956  * as appropriate.  When constructing the vdev spec, we were unable to open this
 957  * device in order to provide a devid.  Now that we have labelled the disk and
 958  * know that slice 0 is valid, we can construct the devid now.
 959  *
 960  * If the disk was already labeled with an EFI label, we will have gotten the
 961  * devid already (because we were able to open the whole disk).  Otherwise, we
 962  * need to get the devid after we label the disk.
 963  */
 964 static int
 965 make_disks(zpool_handle_t *zhp, nvlist_t *nv)
 966 {
 967         nvlist_t **child;
 968         uint_t c, children;
 969         char *type, *path;
 970         char devpath[MAXPATHLEN];
 971         char udevpath[MAXPATHLEN];
 972         uint64_t wholedisk;
 973         struct stat64 statbuf;
 974         int is_exclusive = 0;
 975         int fd;
 976         int ret;
 977
 978         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
 979
 980         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 981             &child, &children) != 0) {
 982
 983                 if (strcmp(type, VDEV_TYPE_DISK) != 0)
 984                         return (0);
 985
 986                 /*
 987                  * We have a disk device.  If this is a whole disk write
 988                  * out the efi partition table, otherwise write zero's to
 989                  * the first 4k of the partition.  This is to ensure that
 990                  * libblkid will not misidentify the partition due to a
 991                  * magic value left by the previous filesystem.
 992                  */
 993                 verify(!nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path));
 994                 verify(!nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
 995                     &wholedisk));
 996
 997                 if (!wholedisk) {
 998                         (void) zero_label(path);
 999                         return (0);
1000                 }
1001
1002                 if (realpath(path, devpath) == NULL) {
1003                         ret = errno;
1004                         (void) fprintf(stderr,
1005                             gettext("cannot resolve path '%s'\n"), path);
1006                         return (ret);
1007                 }
1008
1009                 /*
1010                  * Remove any previously existing symlink from a udev path to
1011                  * the device before labeling the disk.  This makes
1012                  * zpool_label_disk_wait() truly wait for the new link to show
1013                  * up instead of returning if it finds an old link still in
1014                  * place.  Otherwise there is a window between when udev
1015                  * deletes and recreates the link during which access attempts
1016                  * will fail with ENOENT.
1017                  */
1018                 strncpy(udevpath, path, MAXPATHLEN);
1019                 (void) zfs_append_partition(udevpath, MAXPATHLEN);
1020
1021                 fd = open(devpath, O_RDWR|O_EXCL);
1022                 if (fd == -1) {
1023                         if (errno == EBUSY)
1024                                 is_exclusive = 1;
1025                 } else {
1026                         (void) close(fd);
1027                 }
1028
1029                 /*
1030                  * If the partition exists, contains a valid spare label,
1031                  * and is opened exclusively there is no need to partition
1032                  * it.  Hot spares have already been partitioned and are
1033                  * held open exclusively by the kernel as a safety measure.
1034                  *
1035                  * If the provided path is for a /dev/disk/ device its
1036                  * symbolic link will be removed, partition table created,
1037                  * and then block until udev creates the new link.
1038                  */
1039                 if (!is_exclusive || !is_spare(NULL, udevpath)) {
1040                         ret = strncmp(udevpath,UDISK_ROOT,strlen(UDISK_ROOT));
1041                         if (ret == 0) {
1042                                 ret = lstat64(udevpath, &statbuf);
1043                                 if (ret == 0 && S_ISLNK(statbuf.st_mode))
1044                                         (void) unlink(udevpath);
1045                         }
1046
1047                         if (zpool_label_disk(g_zfs, zhp,
1048                             strrchr(devpath, '/') + 1) == -1)
1049                                 return (-1);
1050
1051                         ret = zpool_label_disk_wait(udevpath, 1000);
1052                         if (ret) {
1053                                 (void) fprintf(stderr, gettext("cannot "
1054                                     "resolve path '%s': %d\n"), udevpath, ret);
1055                                 return (-1);
1056                         }
1057
1058                         (void) zero_label(udevpath);
1059                 }
1060
1061                 /*
1062                  * Update the path to refer to the partition.  The presence of
1063                  * the 'whole_disk' field indicates to the CLI that we should
1064                  * chop off the partition number when displaying the device in
1065                  * future output.
1066                  */
1067                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_PATH, udevpath) == 0);
1068
1069                 return (0);
1070         }
1071
1072         for (c = 0; c < children; c++)
1073                 if ((ret = make_disks(zhp, child[c])) != 0)
1074                         return (ret);
1075
1076         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_SPARES,
1077             &child, &children) == 0)
1078                 for (c = 0; c < children; c++)
1079                         if ((ret = make_disks(zhp, child[c])) != 0)
1080                                 return (ret);
1081
1082         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1083             &child, &children) == 0)
1084                 for (c = 0; c < children; c++)
1085                         if ((ret = make_disks(zhp, child[c])) != 0)
1086                                 return (ret);
1087
1088         return (0);
1089 }
1090
1091 /*
1092  * Go through and find any devices that are in use.  We rely on libdiskmgt for
1093  * the majority of this task.
1094  */
1095 static int
1096 check_in_use(nvlist_t *config, nvlist_t *nv, boolean_t force,
1097     boolean_t replacing, boolean_t isspare)
1098 {
1099         nvlist_t **child;
1100         uint_t c, children;
1101         char *type, *path;
1102         int ret = 0;
1103         char buf[MAXPATHLEN];
1104         uint64_t wholedisk = B_FALSE;
1105
1106         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
1107
1108         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
1109             &child, &children) != 0) {
1110
1111                 verify(!nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path));
1112                 if (strcmp(type, VDEV_TYPE_DISK) == 0)
1113                         verify(!nvlist_lookup_uint64(nv,
1114                                ZPOOL_CONFIG_WHOLE_DISK, &wholedisk));
1115
1116                 /*
1117                  * As a generic check, we look to see if this is a replace of a
1118                  * hot spare within the same pool.  If so, we allow it
1119                  * regardless of what libblkid or zpool_in_use() says.
1120                  */
1121                 if (replacing) {
1122                         (void) strlcpy(buf, path, sizeof (buf));
1123                         if (wholedisk) {
1124                                 ret = zfs_append_partition(buf,  sizeof (buf));
1125                                 if (ret == -1)
1126                                         return (-1);
1127                         }
1128
1129                         if (is_spare(config, buf))
1130                                 return (0);
1131                 }
1132
1133                 if (strcmp(type, VDEV_TYPE_DISK) == 0)
1134                         ret = check_device(path, force, isspare, wholedisk);
1135
1136                 if (strcmp(type, VDEV_TYPE_FILE) == 0)
1137                         ret = check_file(path, force, isspare);
1138
1139                 return (ret);
1140         }
1141
1142         for (c = 0; c < children; c++)
1143                 if ((ret = check_in_use(config, child[c], force,
1144                     replacing, B_FALSE)) != 0)
1145                         return (ret);
1146
1147         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_SPARES,
1148             &child, &children) == 0)
1149                 for (c = 0; c < children; c++)
1150                         if ((ret = check_in_use(config, child[c], force,
1151                             replacing, B_TRUE)) != 0)
1152                                 return (ret);
1153
1154         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1155             &child, &children) == 0)
1156                 for (c = 0; c < children; c++)
1157                         if ((ret = check_in_use(config, child[c], force,
1158                             replacing, B_FALSE)) != 0)
1159                                 return (ret);
1160
1161         return (0);
1162 }
1163
1164 static const char *
1165 is_grouping(const char *type, int *mindev, int *maxdev)
1166 {
1167         if (strncmp(type, "raidz", 5) == 0) {
1168                 const char *p = type + 5;
1169                 char *end;
1170                 long nparity;
1171
1172                 if (*p == '\0') {
1173                         nparity = 1;
1174                 } else if (*p == '0') {
1175                         return (NULL); /* no zero prefixes allowed */
1176                 } else {
1177                         errno = 0;
1178                         nparity = strtol(p, &end, 10);
1179                         if (errno != 0 || nparity < 1 || nparity >= 255 ||
1180                             *end != '\0')
1181                                 return (NULL);
1182                 }
1183
1184                 if (mindev != NULL)
1185                         *mindev = nparity + 1;
1186                 if (maxdev != NULL)
1187                         *maxdev = 255;
1188                 return (VDEV_TYPE_RAIDZ);
1189         }
1190
1191         if (maxdev != NULL)
1192                 *maxdev = INT_MAX;
1193
1194         if (strcmp(type, "mirror") == 0) {
1195                 if (mindev != NULL)
1196                         *mindev = 2;
1197                 return (VDEV_TYPE_MIRROR);
1198         }
1199
1200         if (strcmp(type, "spare") == 0) {
1201                 if (mindev != NULL)
1202                         *mindev = 1;
1203                 return (VDEV_TYPE_SPARE);
1204         }
1205
1206         if (strcmp(type, "log") == 0) {
1207                 if (mindev != NULL)
1208                         *mindev = 1;
1209                 return (VDEV_TYPE_LOG);
1210         }
1211
1212         if (strcmp(type, "cache") == 0) {
1213                 if (mindev != NULL)
1214                         *mindev = 1;
1215                 return (VDEV_TYPE_L2CACHE);
1216         }
1217
1218         return (NULL);
1219 }
1220
1221 /*
1222  * Construct a syntactically valid vdev specification,
1223  * and ensure that all devices and files exist and can be opened.
1224  * Note: we don't bother freeing anything in the error paths
1225  * because the program is just going to exit anyway.
1226  */
1227 nvlist_t *
1228 construct_spec(nvlist_t *props, int argc, char **argv)
1229 {
1230         nvlist_t *nvroot, *nv, **top, **spares, **l2cache;
1231         int t, toplevels, mindev, maxdev, nspares, nlogs, nl2cache;
1232         const char *type;
1233         uint64_t is_log;
1234         boolean_t seen_logs;
1235
1236         top = NULL;
1237         toplevels = 0;
1238         spares = NULL;
1239         l2cache = NULL;
1240         nspares = 0;
1241         nlogs = 0;
1242         nl2cache = 0;
1243         is_log = B_FALSE;
1244         seen_logs = B_FALSE;
1245
1246         while (argc > 0) {
1247                 nv = NULL;
1248
1249                 /*
1250                  * If it's a mirror or raidz, the subsequent arguments are
1251                  * its leaves -- until we encounter the next mirror or raidz.
1252                  */
1253                 if ((type = is_grouping(argv[0], &mindev, &maxdev)) != NULL) {
1254                         nvlist_t **child = NULL;
1255                         int c, children = 0;
1256
1257                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1258                                 if (spares != NULL) {
1259                                         (void) fprintf(stderr,
1260                                             gettext("invalid vdev "
1261                                             "specification: 'spare' can be "
1262                                             "specified only once\n"));
1263                                         return (NULL);
1264                                 }
1265                                 is_log = B_FALSE;
1266                         }
1267
1268                         if (strcmp(type, VDEV_TYPE_LOG) == 0) {
1269                                 if (seen_logs) {
1270                                         (void) fprintf(stderr,
1271                                             gettext("invalid vdev "
1272                                             "specification: 'log' can be "
1273                                             "specified only once\n"));
1274                                         return (NULL);
1275                                 }
1276                                 seen_logs = B_TRUE;
1277                                 is_log = B_TRUE;
1278                                 argc--;
1279                                 argv++;
1280                                 /*
1281                                  * A log is not a real grouping device.
1282                                  * We just set is_log and continue.
1283                                  */
1284                                 continue;
1285                         }
1286
1287                         if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1288                                 if (l2cache != NULL) {
1289                                         (void) fprintf(stderr,
1290                                             gettext("invalid vdev "
1291                                             "specification: 'cache' can be "
1292                                             "specified only once\n"));
1293                                         return (NULL);
1294                                 }
1295                                 is_log = B_FALSE;
1296                         }
1297
1298                         if (is_log) {
1299                                 if (strcmp(type, VDEV_TYPE_MIRROR) != 0) {
1300                                         (void) fprintf(stderr,
1301                                             gettext("invalid vdev "
1302                                             "specification: unsupported 'log' "
1303                                             "device: %s\n"), type);
1304                                         return (NULL);
1305                                 }
1306                                 nlogs++;
1307                         }
1308
1309                         for (c = 1; c < argc; c++) {
1310                                 if (is_grouping(argv[c], NULL, NULL) != NULL)
1311                                         break;
1312                                 children++;
1313                                 child = realloc(child,
1314                                     children * sizeof (nvlist_t *));
1315                                 if (child == NULL)
1316                                         zpool_no_memory();
1317                                 if ((nv = make_leaf_vdev(props, argv[c], B_FALSE))
1318                                     == NULL)
1319                                         return (NULL);
1320                                 child[children - 1] = nv;
1321                         }
1322
1323                         if (children < mindev) {
1324                                 (void) fprintf(stderr, gettext("invalid vdev "
1325                                     "specification: %s requires at least %d "
1326                                     "devices\n"), argv[0], mindev);
1327                                 return (NULL);
1328                         }
1329
1330                         if (children > maxdev) {
1331                                 (void) fprintf(stderr, gettext("invalid vdev "
1332                                     "specification: %s supports no more than "
1333                                     "%d devices\n"), argv[0], maxdev);
1334                                 return (NULL);
1335                         }
1336
1337                         argc -= c;
1338                         argv += c;
1339
1340                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1341                                 spares = child;
1342                                 nspares = children;
1343                                 continue;
1344                         } else if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1345                                 l2cache = child;
1346                                 nl2cache = children;
1347                                 continue;
1348                         } else {
1349                                 verify(nvlist_alloc(&nv, NV_UNIQUE_NAME,
1350                                     0) == 0);
1351                                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_TYPE,
1352                                     type) == 0);
1353                                 verify(nvlist_add_uint64(nv,
1354                                     ZPOOL_CONFIG_IS_LOG, is_log) == 0);
1355                                 if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
1356                                         verify(nvlist_add_uint64(nv,
1357                                             ZPOOL_CONFIG_NPARITY,
1358                                             mindev - 1) == 0);
1359                                 }
1360                                 verify(nvlist_add_nvlist_array(nv,
1361                                     ZPOOL_CONFIG_CHILDREN, child,
1362                                     children) == 0);
1363
1364                                 for (c = 0; c < children; c++)
1365                                         nvlist_free(child[c]);
1366                                 free(child);
1367                         }
1368                 } else {
1369                         /*
1370                          * We have a device.  Pass off to make_leaf_vdev() to
1371                          * construct the appropriate nvlist describing the vdev.
1372                          */
1373                         if ((nv = make_leaf_vdev(props, argv[0], is_log)) == NULL)
1374                                 return (NULL);
1375                         if (is_log)
1376                                 nlogs++;
1377                         argc--;
1378                         argv++;
1379                 }
1380
1381                 toplevels++;
1382                 top = realloc(top, toplevels * sizeof (nvlist_t *));
1383                 if (top == NULL)
1384                         zpool_no_memory();
1385                 top[toplevels - 1] = nv;
1386         }
1387
1388         if (toplevels == 0 && nspares == 0 && nl2cache == 0) {
1389                 (void) fprintf(stderr, gettext("invalid vdev "
1390                     "specification: at least one toplevel vdev must be "
1391                     "specified\n"));
1392                 return (NULL);
1393         }
1394
1395         if (seen_logs && nlogs == 0) {
1396                 (void) fprintf(stderr, gettext("invalid vdev specification: "
1397                     "log requires at least 1 device\n"));
1398                 return (NULL);
1399         }
1400
1401         /*
1402          * Finally, create nvroot and add all top-level vdevs to it.
1403          */
1404         verify(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, 0) == 0);
1405         verify(nvlist_add_string(nvroot, ZPOOL_CONFIG_TYPE,
1406             VDEV_TYPE_ROOT) == 0);
1407         verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
1408             top, toplevels) == 0);
1409         if (nspares != 0)
1410                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1411                     spares, nspares) == 0);
1412         if (nl2cache != 0)
1413                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
1414                     l2cache, nl2cache) == 0);
1415
1416         for (t = 0; t < toplevels; t++)
1417                 nvlist_free(top[t]);
1418         for (t = 0; t < nspares; t++)
1419                 nvlist_free(spares[t]);
1420         for (t = 0; t < nl2cache; t++)
1421                 nvlist_free(l2cache[t]);
1422         if (spares)
1423                 free(spares);
1424         if (l2cache)
1425                 free(l2cache);
1426         free(top);
1427
1428         return (nvroot);
1429 }
1430
1431 nvlist_t *
1432 split_mirror_vdev(zpool_handle_t *zhp, char *newname, nvlist_t *props,
1433     splitflags_t flags, int argc, char **argv)
1434 {
1435         nvlist_t *newroot = NULL, **child;
1436         uint_t c, children;
1437
1438         if (argc > 0) {
1439                 if ((newroot = construct_spec(props, argc, argv)) == NULL) {
1440                         (void) fprintf(stderr, gettext("Unable to build a "
1441                             "pool from the specified devices\n"));
1442                         return (NULL);
1443                 }
1444
1445                 if (!flags.dryrun && make_disks(zhp, newroot) != 0) {
1446                         nvlist_free(newroot);
1447                         return (NULL);
1448                 }
1449
1450                 /* avoid any tricks in the spec */
1451                 verify(nvlist_lookup_nvlist_array(newroot,
1452                     ZPOOL_CONFIG_CHILDREN, &child, &children) == 0);
1453                 for (c = 0; c < children; c++) {
1454                         char *path;
1455                         const char *type;
1456                         int min, max;
1457
1458                         verify(nvlist_lookup_string(child[c],
1459                             ZPOOL_CONFIG_PATH, &path) == 0);
1460                         if ((type = is_grouping(path, &min, &max)) != NULL) {
1461                                 (void) fprintf(stderr, gettext("Cannot use "
1462                                     "'%s' as a device for splitting\n"), type);
1463                                 nvlist_free(newroot);
1464                                 return (NULL);
1465                         }
1466                 }
1467         }
1468
1469         if (zpool_vdev_split(zhp, newname, &newroot, props, flags) != 0) {
1470                 if (newroot != NULL)
1471                         nvlist_free(newroot);
1472                 return (NULL);
1473         }
1474
1475         return (newroot);
1476 }
1477
1478 /*
1479  * Get and validate the contents of the given vdev specification.  This ensures
1480  * that the nvlist returned is well-formed, that all the devices exist, and that
1481  * they are not currently in use by any other known consumer.  The 'poolconfig'
1482  * parameter is the current configuration of the pool when adding devices
1483  * existing pool, and is used to perform additional checks, such as changing the
1484  * replication level of the pool.  It can be 'NULL' to indicate that this is a
1485  * new pool.  The 'force' flag controls whether devices should be forcefully
1486  * added, even if they appear in use.
1487  */
1488 nvlist_t *
1489 make_root_vdev(zpool_handle_t *zhp, nvlist_t *props, int force, int check_rep,
1490     boolean_t replacing, boolean_t dryrun, int argc, char **argv)
1491 {
1492         nvlist_t *newroot;
1493         nvlist_t *poolconfig = NULL;
1494         is_force = force;
1495
1496         /*
1497          * Construct the vdev specification.  If this is successful, we know
1498          * that we have a valid specification, and that all devices can be
1499          * opened.
1500          */
1501         if ((newroot = construct_spec(props, argc, argv)) == NULL)
1502                 return (NULL);
1503
1504         if (zhp && ((poolconfig = zpool_get_config(zhp, NULL)) == NULL))
1505                 return (NULL);
1506
1507         /*
1508          * Validate each device to make sure that its not shared with another
1509          * subsystem.  We do this even if 'force' is set, because there are some
1510          * uses (such as a dedicated dump device) that even '-f' cannot
1511          * override.
1512          */
1513         if (check_in_use(poolconfig, newroot, force, replacing, B_FALSE) != 0) {
1514                 nvlist_free(newroot);
1515                 return (NULL);
1516         }
1517
1518         /*
1519          * Check the replication level of the given vdevs and report any errors
1520          * found.  We include the existing pool spec, if any, as we need to
1521          * catch changes against the existing replication level.
1522          */
1523         if (check_rep && check_replication(poolconfig, newroot) != 0) {
1524                 nvlist_free(newroot);
1525                 return (NULL);
1526         }
1527
1528         /*
1529          * Run through the vdev specification and label any whole disks found.
1530          */
1531         if (!dryrun && make_disks(zhp, newroot) != 0) {
1532                 nvlist_free(newroot);
1533                 return (NULL);
1534         }
1535
1536         return (newroot);
1537 }