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         struct stat64 statbuf;
 192         int err;
 193 #ifdef HAVE_LIBBLKID
 194         char *value;
 195 #endif /* HAVE_LIBBLKID */
 196
 197         if (stat64(path, &statbuf) != 0) {
 198                 vdev_error(gettext("cannot stat %s: %s\n"),
 199                            path, strerror(errno));
 200                 return (-1);
 201         }
 202
 203 #ifdef HAVE_LIBBLKID
 204         /* No valid type detected device is safe to use */
 205         value = blkid_get_tag_value(cache, "TYPE", path);
 206         if (value == NULL)
 207                 return (0);
 208
 209         /*
 210          * If libblkid detects a ZFS device, we check the device
 211          * using check_file() to see if it's safe.  The one safe
 212          * case is a spare device shared between multiple pools.
 213          */
 214         if (strcmp(value, "zfs") == 0) {
 215                 err = check_file(path, force, isspare);
 216         } else {
 217                 if (force) {
 218                         err = 0;
 219                 } else {
 220                         err = -1;
 221                         vdev_error(gettext("%s contains a filesystem of "
 222                                    "type '%s'\n"), path, value);
 223                 }
 224         }
 225
 226         free(value);
 227 #else
 228         err = check_file(path, force, isspare);
 229 #endif /* HAVE_LIBBLKID */
 230
 231         return (err);
 232 }
 233
 234 /*
 235  * Validate a whole disk.  Iterate over all slices on the disk and make sure
 236  * that none is in use by calling check_slice().
 237  */
 238 static int
 239 check_disk(const char *path, blkid_cache cache, int force,
 240            boolean_t isspare, boolean_t iswholedisk)
 241 {
 242         struct dk_gpt *vtoc;
 243         char slice_path[MAXPATHLEN];
 244         int err = 0;
 245         int fd, i;
 246
 247         /* This is not a wholedisk we only check the given partition */
 248         if (!iswholedisk)
 249                 return check_slice(path, cache, force, isspare);
 250
 251         /*
 252          * When the device is a whole disk try to read the efi partition
 253          * label.  If this is successful we safely check the all of the
 254          * partitions.  However, when it fails it may simply be because
 255          * the disk is partitioned via the MBR.  Since we currently can
 256          * not easily decode the MBR return a failure and prompt to the
 257          * user to use force option since we cannot check the partitions.
 258          */
 259         if ((fd = open(path, O_RDWR|O_DIRECT|O_EXCL)) < 0) {
 260                 check_error(errno);
 261                 return -1;
 262         }
 263
 264         if ((err = efi_alloc_and_read(fd, &vtoc)) != 0) {
 265                 (void) close(fd);
 266
 267                 if (force) {
 268                         return 0;
 269                 } else {
 270                         vdev_error(gettext("%s does not contain an EFI "
 271                             "label but it may contain partition\n"
 272                             "information in the MBR.\n"), path);
 273                         return -1;
 274                 }
 275         }
 276
 277         /*
 278          * The primary efi partition label is damaged however the secondary
 279          * label at the end of the device is intact.  Rather than use this
 280          * label we should play it safe and treat this as a non efi device.
 281          */
 282         if (vtoc->efi_flags & EFI_GPT_PRIMARY_CORRUPT) {
 283                 efi_free(vtoc);
 284                 (void) close(fd);
 285
 286                 if (force) {
 287                         /* Partitions will no be created using the backup */
 288                         return 0;
 289                 } else {
 290                         vdev_error(gettext("%s contains a corrupt primary "
 291                             "EFI label.\n"), path);
 292                         return -1;
 293                 }
 294         }
 295
 296         for (i = 0; i < vtoc->efi_nparts; i++) {
 297
 298                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED ||
 299                     uuid_is_null((uchar_t *)&vtoc->efi_parts[i].p_guid))
 300                         continue;
 301
 302                 if (strncmp(path, UDISK_ROOT, strlen(UDISK_ROOT)) == 0)
 303                         (void) snprintf(slice_path, sizeof (slice_path),
 304                             "%s%s%d", path, "-part", i+1);
 305                 else
 306                         (void) snprintf(slice_path, sizeof (slice_path),
 307                             "%s%s%d", path, isdigit(path[strlen(path)-1]) ?
 308                             "p" : "", i+1);
 309
 310                 err = check_slice(slice_path, cache, force, isspare);
 311                 if (err)
 312                         break;
 313         }
 314
 315         efi_free(vtoc);
 316         (void) close(fd);
 317
 318         return (err);
 319 }
 320
 321 static int
 322 check_device(const char *path, boolean_t force,
 323              boolean_t isspare, boolean_t iswholedisk)
 324 {
 325         static blkid_cache cache = NULL;
 326
 327 #ifdef HAVE_LIBBLKID
 328         /*
 329          * There is no easy way to add a correct blkid_put_cache() call,
 330          * memory will be reclaimed when the command exits.
 331          */
 332         if (cache == NULL) {
 333                 int err;
 334
 335                 if ((err = blkid_get_cache(&cache, NULL)) != 0) {
 336                         check_error(err);
 337                         return -1;
 338                 }
 339
 340                 if ((err = blkid_probe_all(cache)) != 0) {
 341                         blkid_put_cache(cache);
 342                         check_error(err);
 343                         return -1;
 344                 }
 345         }
 346 #endif /* HAVE_LIBBLKID */
 347
 348         return check_disk(path, cache, force, isspare, iswholedisk);
 349 }
 350
 351 /*
 352  * By "whole disk" we mean an entire physical disk (something we can
 353  * label, toggle the write cache on, etc.) as opposed to the full
 354  * capacity of a pseudo-device such as lofi or did.  We act as if we
 355  * are labeling the disk, which should be a pretty good test of whether
 356  * it's a viable device or not.  Returns B_TRUE if it is and B_FALSE if
 357  * it isn't.
 358  */
 359 static boolean_t
 360 is_whole_disk(const char *path)
 361 {
 362         struct dk_gpt *label;
 363         int     fd;
 364
 365         if ((fd = open(path, O_RDWR|O_DIRECT|O_EXCL)) < 0)
 366                 return (B_FALSE);
 367         if (efi_alloc_and_init(fd, EFI_NUMPAR, &label) != 0) {
 368                 (void) close(fd);
 369                 return (B_FALSE);
 370         }
 371         efi_free(label);
 372         (void) close(fd);
 373         return (B_TRUE);
 374 }
 375
 376 /*
 377  * This may be a shorthand device path or it could be total gibberish.
 378  * Check to see if it's a known device in /dev/, /dev/disk/by-id,
 379  * /dev/disk/by-label, /dev/disk/by-path, /dev/disk/by-uuid, or
 380  * /dev/disk/zpool/.  As part of this check, see if we've been given
 381  * an entire disk (minus the slice number).
 382  */
 383 static int
 384 is_shorthand_path(const char *arg, char *path,
 385                   struct stat64 *statbuf, boolean_t *wholedisk)
 386 {
 387         if (zfs_resolve_shortname(arg, path, MAXPATHLEN) == 0) {
 388                 *wholedisk = is_whole_disk(path);
 389                 if (*wholedisk || (stat64(path, statbuf) == 0))
 390                         return (0);
 391         }
 392
 393         strlcpy(path, arg, sizeof(path));
 394         memset(statbuf, 0, sizeof(*statbuf));
 395         *wholedisk = B_FALSE;
 396
 397         return (ENOENT);
 398 }
 399
 400 /*
 401  * Create a leaf vdev.  Determine if this is a file or a device.  If it's a
 402  * device, fill in the device id to make a complete nvlist.  Valid forms for a
 403  * leaf vdev are:
 404  *
 405  *      /dev/xxx        Complete disk path
 406  *      /xxx            Full path to file
 407  *      xxx             Shorthand for /dev/disk/yyy/xxx
 408  */
 409 static nvlist_t *
 410 make_leaf_vdev(const char *arg, uint64_t is_log)
 411 {
 412         char path[MAXPATHLEN];
 413         struct stat64 statbuf;
 414         nvlist_t *vdev = NULL;
 415         char *type = NULL;
 416         boolean_t wholedisk = B_FALSE;
 417         int err;
 418
 419         /*
 420          * Determine what type of vdev this is, and put the full path into
 421          * 'path'.  We detect whether this is a device of file afterwards by
 422          * checking the st_mode of the file.
 423          */
 424         if (arg[0] == '/') {
 425                 /*
 426                  * Complete device or file path.  Exact type is determined by
 427                  * examining the file descriptor afterwards.  Symbolic links
 428                  * are resolved to their real paths for the is_whole_disk()
 429                  * and S_ISBLK/S_ISREG type checks.  However, we are careful
 430                  * to store the given path as ZPOOL_CONFIG_PATH to ensure we
 431                  * can leverage udev's persistent device labels.
 432                  */
 433                 if (realpath(arg, path) == NULL) {
 434                         (void) fprintf(stderr,
 435                             gettext("cannot resolve path '%s'\n"), arg);
 436                         return (NULL);
 437                 }
 438
 439                 wholedisk = is_whole_disk(path);
 440                 if (!wholedisk && (stat64(path, &statbuf) != 0)) {
 441                         (void) fprintf(stderr,
 442                             gettext("cannot open '%s': %s\n"),
 443                             path, strerror(errno));
 444                         return (NULL);
 445                 }
 446
 447                 /* After is_whole_disk() check restore original passed path */
 448                 strlcpy(path, arg, MAXPATHLEN);
 449         } else {
 450                 err = is_shorthand_path(arg, path, &statbuf, &wholedisk);
 451                 if (err != 0) {
 452                         /*
 453                          * If we got ENOENT, then the user gave us
 454                          * gibberish, so try to direct them with a
 455                          * reasonable error message.  Otherwise,
 456                          * regurgitate strerror() since it's the best we
 457                          * can do.
 458                          */
 459                         if (err == ENOENT) {
 460                                 (void) fprintf(stderr,
 461                                     gettext("cannot open '%s': no such "
 462                                     "device in %s\n"), arg, DISK_ROOT);
 463                                 (void) fprintf(stderr,
 464                                     gettext("must be a full path or "
 465                                     "shorthand device name\n"));
 466                                 return (NULL);
 467                         } else {
 468                                 (void) fprintf(stderr,
 469                                     gettext("cannot open '%s': %s\n"),
 470                                     path, strerror(errno));
 471                                 return (NULL);
 472                         }
 473                 }
 474         }
 475
 476         /*
 477          * Determine whether this is a device or a file.
 478          */
 479         if (wholedisk || S_ISBLK(statbuf.st_mode)) {
 480                 type = VDEV_TYPE_DISK;
 481         } else if (S_ISREG(statbuf.st_mode)) {
 482                 type = VDEV_TYPE_FILE;
 483         } else {
 484                 (void) fprintf(stderr, gettext("cannot use '%s': must be a "
 485                     "block device or regular file\n"), path);
 486                 return (NULL);
 487         }
 488
 489         /*
 490          * Finally, we have the complete device or file, and we know that it is
 491          * acceptable to use.  Construct the nvlist to describe this vdev.  All
 492          * vdevs have a 'path' element, and devices also have a 'devid' element.
 493          */
 494         verify(nvlist_alloc(&vdev, NV_UNIQUE_NAME, 0) == 0);
 495         verify(nvlist_add_string(vdev, ZPOOL_CONFIG_PATH, path) == 0);
 496         verify(nvlist_add_string(vdev, ZPOOL_CONFIG_TYPE, type) == 0);
 497         verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_IS_LOG, is_log) == 0);
 498         if (strcmp(type, VDEV_TYPE_DISK) == 0)
 499                 verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_WHOLE_DISK,
 500                     (uint64_t)wholedisk) == 0);
 501
 502         return (vdev);
 503 }
 504
 505 /*
 506  * Go through and verify the replication level of the pool is consistent.
 507  * Performs the following checks:
 508  *
 509  *      For the new spec, verifies that devices in mirrors and raidz are the
 510  *      same size.
 511  *
 512  *      If the current configuration already has inconsistent replication
 513  *      levels, ignore any other potential problems in the new spec.
 514  *
 515  *      Otherwise, make sure that the current spec (if there is one) and the new
 516  *      spec have consistent replication levels.
 517  */
 518 typedef struct replication_level {
 519         char *zprl_type;
 520         uint64_t zprl_children;
 521         uint64_t zprl_parity;
 522 } replication_level_t;
 523
 524 #define ZPOOL_FUZZ      (16 * 1024 * 1024)
 525
 526 /*
 527  * Given a list of toplevel vdevs, return the current replication level.  If
 528  * the config is inconsistent, then NULL is returned.  If 'fatal' is set, then
 529  * an error message will be displayed for each self-inconsistent vdev.
 530  */
 531 static replication_level_t *
 532 get_replication(nvlist_t *nvroot, boolean_t fatal)
 533 {
 534         nvlist_t **top;
 535         uint_t t, toplevels;
 536         nvlist_t **child;
 537         uint_t c, children;
 538         nvlist_t *nv;
 539         char *type;
 540         replication_level_t lastrep = { 0 }, rep, *ret;
 541         boolean_t dontreport;
 542
 543         ret = safe_malloc(sizeof (replication_level_t));
 544
 545         verify(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
 546             &top, &toplevels) == 0);
 547
 548         lastrep.zprl_type = NULL;
 549         for (t = 0; t < toplevels; t++) {
 550                 uint64_t is_log = B_FALSE;
 551
 552                 nv = top[t];
 553
 554                 /*
 555                  * For separate logs we ignore the top level vdev replication
 556                  * constraints.
 557                  */
 558                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &is_log);
 559                 if (is_log)
 560                         continue;
 561
 562                 verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE,
 563                     &type) == 0);
 564                 if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 565                     &child, &children) != 0) {
 566                         /*
 567                          * This is a 'file' or 'disk' vdev.
 568                          */
 569                         rep.zprl_type = type;
 570                         rep.zprl_children = 1;
 571                         rep.zprl_parity = 0;
 572                 } else {
 573                         uint64_t vdev_size;
 574
 575                         /*
 576                          * This is a mirror or RAID-Z vdev.  Go through and make
 577                          * sure the contents are all the same (files vs. disks),
 578                          * keeping track of the number of elements in the
 579                          * process.
 580                          *
 581                          * We also check that the size of each vdev (if it can
 582                          * be determined) is the same.
 583                          */
 584                         rep.zprl_type = type;
 585                         rep.zprl_children = 0;
 586
 587                         if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
 588                                 verify(nvlist_lookup_uint64(nv,
 589                                     ZPOOL_CONFIG_NPARITY,
 590                                     &rep.zprl_parity) == 0);
 591                                 assert(rep.zprl_parity != 0);
 592                         } else {
 593                                 rep.zprl_parity = 0;
 594                         }
 595
 596                         /*
 597                          * The 'dontreport' variable indicates that we've
 598                          * already reported an error for this spec, so don't
 599                          * bother doing it again.
 600                          */
 601                         type = NULL;
 602                         dontreport = 0;
 603                         vdev_size = -1ULL;
 604                         for (c = 0; c < children; c++) {
 605                                 nvlist_t *cnv = child[c];
 606                                 char *path;
 607                                 struct stat64 statbuf;
 608                                 uint64_t size = -1ULL;
 609                                 char *childtype;
 610                                 int fd, err;
 611
 612                                 rep.zprl_children++;
 613
 614                                 verify(nvlist_lookup_string(cnv,
 615                                     ZPOOL_CONFIG_TYPE, &childtype) == 0);
 616
 617                                 /*
 618                                  * If this is a replacing or spare vdev, then
 619                                  * get the real first child of the vdev.
 620                                  */
 621                                 if (strcmp(childtype,
 622                                     VDEV_TYPE_REPLACING) == 0 ||
 623                                     strcmp(childtype, VDEV_TYPE_SPARE) == 0) {
 624                                         nvlist_t **rchild;
 625                                         uint_t rchildren;
 626
 627                                         verify(nvlist_lookup_nvlist_array(cnv,
 628                                             ZPOOL_CONFIG_CHILDREN, &rchild,
 629                                             &rchildren) == 0);
 630                                         assert(rchildren == 2);
 631                                         cnv = rchild[0];
 632
 633                                         verify(nvlist_lookup_string(cnv,
 634                                             ZPOOL_CONFIG_TYPE,
 635                                             &childtype) == 0);
 636                                 }
 637
 638                                 verify(nvlist_lookup_string(cnv,
 639                                     ZPOOL_CONFIG_PATH, &path) == 0);
 640
 641                                 /*
 642                                  * If we have a raidz/mirror that combines disks
 643                                  * with files, report it as an error.
 644                                  */
 645                                 if (!dontreport && type != NULL &&
 646                                     strcmp(type, childtype) != 0) {
 647                                         if (ret != NULL)
 648                                                 free(ret);
 649                                         ret = NULL;
 650                                         if (fatal)
 651                                                 vdev_error(gettext(
 652                                                     "mismatched replication "
 653                                                     "level: %s contains both "
 654                                                     "files and devices\n"),
 655                                                     rep.zprl_type);
 656                                         else
 657                                                 return (NULL);
 658                                         dontreport = B_TRUE;
 659                                 }
 660
 661                                 /*
 662                                  * According to stat(2), the value of 'st_size'
 663                                  * is undefined for block devices and character
 664                                  * devices.  But there is no effective way to
 665                                  * determine the real size in userland.
 666                                  *
 667                                  * Instead, we'll take advantage of an
 668                                  * implementation detail of spec_size().  If the
 669                                  * device is currently open, then we (should)
 670                                  * return a valid size.
 671                                  *
 672                                  * If we still don't get a valid size (indicated
 673                                  * by a size of 0 or MAXOFFSET_T), then ignore
 674                                  * this device altogether.
 675                                  */
 676                                 if ((fd = open(path, O_RDONLY)) >= 0) {
 677                                         err = fstat64(fd, &statbuf);
 678                                         (void) close(fd);
 679                                 } else {
 680                                         err = stat64(path, &statbuf);
 681                                 }
 682
 683                                 if (err != 0 ||
 684                                     statbuf.st_size == 0 ||
 685                                     statbuf.st_size == MAXOFFSET_T)
 686                                         continue;
 687
 688                                 size = statbuf.st_size;
 689
 690                                 /*
 691                                  * Also make sure that devices and
 692                                  * slices have a consistent size.  If
 693                                  * they differ by a significant amount
 694                                  * (~16MB) then report an error.
 695                                  */
 696                                 if (!dontreport &&
 697                                     (vdev_size != -1ULL &&
 698                                     (labs(size - vdev_size) >
 699                                     ZPOOL_FUZZ))) {
 700                                         if (ret != NULL)
 701                                                 free(ret);
 702                                         ret = NULL;
 703                                         if (fatal)
 704                                                 vdev_error(gettext(
 705                                                     "%s contains devices of "
 706                                                     "different sizes\n"),
 707                                                     rep.zprl_type);
 708                                         else
 709                                                 return (NULL);
 710                                         dontreport = B_TRUE;
 711                                 }
 712
 713                                 type = childtype;
 714                                 vdev_size = size;
 715                         }
 716                 }
 717
 718                 /*
 719                  * At this point, we have the replication of the last toplevel
 720                  * vdev in 'rep'.  Compare it to 'lastrep' to see if its
 721                  * different.
 722                  */
 723                 if (lastrep.zprl_type != NULL) {
 724                         if (strcmp(lastrep.zprl_type, rep.zprl_type) != 0) {
 725                                 if (ret != NULL)
 726                                         free(ret);
 727                                 ret = NULL;
 728                                 if (fatal)
 729                                         vdev_error(gettext(
 730                                             "mismatched replication level: "
 731                                             "both %s and %s vdevs are "
 732                                             "present\n"),
 733                                             lastrep.zprl_type, rep.zprl_type);
 734                                 else
 735                                         return (NULL);
 736                         } else if (lastrep.zprl_parity != rep.zprl_parity) {
 737                                 if (ret)
 738                                         free(ret);
 739                                 ret = NULL;
 740                                 if (fatal)
 741                                         vdev_error(gettext(
 742                                             "mismatched replication level: "
 743                                             "both %llu and %llu device parity "
 744                                             "%s vdevs are present\n"),
 745                                             lastrep.zprl_parity,
 746                                             rep.zprl_parity,
 747                                             rep.zprl_type);
 748                                 else
 749                                         return (NULL);
 750                         } else if (lastrep.zprl_children != rep.zprl_children) {
 751                                 if (ret)
 752                                         free(ret);
 753                                 ret = NULL;
 754                                 if (fatal)
 755                                         vdev_error(gettext(
 756                                             "mismatched replication level: "
 757                                             "both %llu-way and %llu-way %s "
 758                                             "vdevs are present\n"),
 759                                             lastrep.zprl_children,
 760                                             rep.zprl_children,
 761                                             rep.zprl_type);
 762                                 else
 763                                         return (NULL);
 764                         }
 765                 }
 766                 lastrep = rep;
 767         }
 768
 769         if (ret != NULL)
 770                 *ret = rep;
 771
 772         return (ret);
 773 }
 774
 775 /*
 776  * Check the replication level of the vdev spec against the current pool.  Calls
 777  * get_replication() to make sure the new spec is self-consistent.  If the pool
 778  * has a consistent replication level, then we ignore any errors.  Otherwise,
 779  * report any difference between the two.
 780  */
 781 static int
 782 check_replication(nvlist_t *config, nvlist_t *newroot)
 783 {
 784         nvlist_t **child;
 785         uint_t  children;
 786         replication_level_t *current = NULL, *new;
 787         int ret;
 788
 789         /*
 790          * If we have a current pool configuration, check to see if it's
 791          * self-consistent.  If not, simply return success.
 792          */
 793         if (config != NULL) {
 794                 nvlist_t *nvroot;
 795
 796                 verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
 797                     &nvroot) == 0);
 798                 if ((current = get_replication(nvroot, B_FALSE)) == NULL)
 799                         return (0);
 800         }
 801         /*
 802          * for spares there may be no children, and therefore no
 803          * replication level to check
 804          */
 805         if ((nvlist_lookup_nvlist_array(newroot, ZPOOL_CONFIG_CHILDREN,
 806             &child, &children) != 0) || (children == 0)) {
 807                 free(current);
 808                 return (0);
 809         }
 810
 811         /*
 812          * If all we have is logs then there's no replication level to check.
 813          */
 814         if (num_logs(newroot) == children) {
 815                 free(current);
 816                 return (0);
 817         }
 818
 819         /*
 820          * Get the replication level of the new vdev spec, reporting any
 821          * inconsistencies found.
 822          */
 823         if ((new = get_replication(newroot, B_TRUE)) == NULL) {
 824                 free(current);
 825                 return (-1);
 826         }
 827
 828         /*
 829          * Check to see if the new vdev spec matches the replication level of
 830          * the current pool.
 831          */
 832         ret = 0;
 833         if (current != NULL) {
 834                 if (strcmp(current->zprl_type, new->zprl_type) != 0) {
 835                         vdev_error(gettext(
 836                             "mismatched replication level: pool uses %s "
 837                             "and new vdev is %s\n"),
 838                             current->zprl_type, new->zprl_type);
 839                         ret = -1;
 840                 } else if (current->zprl_parity != new->zprl_parity) {
 841                         vdev_error(gettext(
 842                             "mismatched replication level: pool uses %llu "
 843                             "device parity and new vdev uses %llu\n"),
 844                             current->zprl_parity, new->zprl_parity);
 845                         ret = -1;
 846                 } else if (current->zprl_children != new->zprl_children) {
 847                         vdev_error(gettext(
 848                             "mismatched replication level: pool uses %llu-way "
 849                             "%s and new vdev uses %llu-way %s\n"),
 850                             current->zprl_children, current->zprl_type,
 851                             new->zprl_children, new->zprl_type);
 852                         ret = -1;
 853                 }
 854         }
 855
 856         free(new);
 857         if (current != NULL)
 858                 free(current);
 859
 860         return (ret);
 861 }
 862
 863 static int
 864 zero_label(char *path)
 865 {
 866         const int size = 4096;
 867         char buf[size];
 868         int err, fd;
 869
 870         if ((fd = open(path, O_WRONLY|O_EXCL)) < 0) {
 871                 (void) fprintf(stderr, gettext("cannot open '%s': %s\n"),
 872                     path, strerror(errno));
 873                 return (-1);
 874         }
 875
 876         memset(buf, 0, size);
 877         err = write(fd, buf, size);
 878         (void) fdatasync(fd);
 879         (void) close(fd);
 880
 881         if (err == -1) {
 882                 (void) fprintf(stderr, gettext("cannot zero first %d bytes "
 883                     "of '%s': %s\n"), size, path, strerror(errno));
 884                 return (-1);
 885         }
 886
 887         if (err != size) {
 888                 (void) fprintf(stderr, gettext("could only zero %d/%d bytes "
 889                     "of '%s'\n"), err, size, path);
 890                 return (-1);
 891         }
 892
 893         return 0;
 894 }
 895
 896 /*
 897  * Go through and find any whole disks in the vdev specification, labelling them
 898  * as appropriate.  When constructing the vdev spec, we were unable to open this
 899  * device in order to provide a devid.  Now that we have labelled the disk and
 900  * know that slice 0 is valid, we can construct the devid now.
 901  *
 902  * If the disk was already labeled with an EFI label, we will have gotten the
 903  * devid already (because we were able to open the whole disk).  Otherwise, we
 904  * need to get the devid after we label the disk.
 905  */
 906 static int
 907 make_disks(zpool_handle_t *zhp, nvlist_t *nv)
 908 {
 909         nvlist_t **child;
 910         uint_t c, children;
 911         char *type, *path, *diskname;
 912         char devpath[MAXPATHLEN];
 913         char udevpath[MAXPATHLEN];
 914         uint64_t wholedisk;
 915         struct stat64 statbuf;
 916         int ret;
 917
 918         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
 919
 920         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 921             &child, &children) != 0) {
 922
 923                 if (strcmp(type, VDEV_TYPE_DISK) != 0)
 924                         return (0);
 925
 926                 /*
 927                  * We have a disk device.  If this is a whole disk write
 928                  * out the efi partition table, otherwise write zero's to
 929                  * the first 4k of the partition.  This is to ensure that
 930                  * libblkid will not misidentify the partition due to a
 931                  * magic value left by the previous filesystem.
 932                  */
 933                 verify(!nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path));
 934                 verify(!nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
 935                     &wholedisk));
 936
 937                 if (!wholedisk) {
 938                         ret = zero_label(path);
 939                         return (ret);
 940                 }
 941
 942                 if (realpath(path, devpath) == NULL) {
 943                         ret = errno;
 944                         (void) fprintf(stderr,
 945                             gettext("cannot resolve path '%s'\n"), path);
 946                         return (ret);
 947                 }
 948
 949                 /*
 950                  * Remove any previously existing symlink from a udev path to
 951                  * the device before labeling the disk.  This makes
 952                  * zpool_label_disk_wait() truly wait for the new link to show
 953                  * up instead of returning if it finds an old link still in
 954                  * place.  Otherwise there is a window between when udev
 955                  * deletes and recreates the link during which access attempts
 956                  * will fail with ENOENT.
 957                  */
 958                 zfs_append_partition(path, udevpath, sizeof (udevpath));
 959                 if ((strncmp(udevpath, UDISK_ROOT, strlen(UDISK_ROOT)) == 0) &&
 960                     (lstat64(udevpath, &statbuf) == 0) &&
 961                     S_ISLNK(statbuf.st_mode))
 962                         (void) unlink(udevpath);
 963
 964                 diskname = strrchr(devpath, '/');
 965                 assert(diskname != NULL);
 966                 diskname++;
 967                 if (zpool_label_disk(g_zfs, zhp, diskname) == -1)
 968                         return (-1);
 969
 970                 /*
 971                  * Now we've labeled the disk and the partitions have been
 972                  * created.  We still need to wait for udev to create the
 973                  * symlinks to those partitions.
 974                  */
 975                 if ((ret = zpool_label_disk_wait(udevpath, 1000)) != 0) {
 976                         (void) fprintf(stderr,
 977                             gettext( "cannot resolve path '%s'\n"), udevpath);
 978                         return (-1);
 979                 }
 980
 981                 /*
 982                  * Update the path to refer to FIRST_SLICE.  The presence of
 983                  * the 'whole_disk' field indicates to the CLI that we should
 984                  * chop off the slice number when displaying the device in
 985                  * future output.
 986                  */
 987                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_PATH, udevpath) == 0);
 988
 989                 /* Just in case this partition already existed. */
 990                 (void) zero_label(udevpath);
 991
 992                 return (0);
 993         }
 994
 995         for (c = 0; c < children; c++)
 996                 if ((ret = make_disks(zhp, child[c])) != 0)
 997                         return (ret);
 998
 999         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_SPARES,
1000             &child, &children) == 0)
1001                 for (c = 0; c < children; c++)
1002                         if ((ret = make_disks(zhp, child[c])) != 0)
1003                                 return (ret);
1004
1005         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1006             &child, &children) == 0)
1007                 for (c = 0; c < children; c++)
1008                         if ((ret = make_disks(zhp, child[c])) != 0)
1009                                 return (ret);
1010
1011         return (0);
1012 }
1013
1014 /*
1015  * Determine if the given path is a hot spare within the given configuration.
1016  */
1017 static boolean_t
1018 is_spare(nvlist_t *config, const char *path)
1019 {
1020         int fd;
1021         pool_state_t state;
1022         char *name = NULL;
1023         nvlist_t *label;
1024         uint64_t guid, spareguid;
1025         nvlist_t *nvroot;
1026         nvlist_t **spares;
1027         uint_t i, nspares;
1028         boolean_t inuse;
1029
1030         if ((fd = open(path, O_RDONLY|O_EXCL)) < 0)
1031                 return (B_FALSE);
1032
1033         if (zpool_in_use(g_zfs, fd, &state, &name, &inuse) != 0 ||
1034             !inuse ||
1035             state != POOL_STATE_SPARE ||
1036             zpool_read_label(fd, &label) != 0) {
1037                 free(name);
1038                 (void) close(fd);
1039                 return (B_FALSE);
1040         }
1041         free(name);
1042         (void) close(fd);
1043
1044         verify(nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) == 0);
1045         nvlist_free(label);
1046
1047         verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
1048             &nvroot) == 0);
1049         if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1050             &spares, &nspares) == 0) {
1051                 for (i = 0; i < nspares; i++) {
1052                         verify(nvlist_lookup_uint64(spares[i],
1053                             ZPOOL_CONFIG_GUID, &spareguid) == 0);
1054                         if (spareguid == guid)
1055                                 return (B_TRUE);
1056                 }
1057         }
1058
1059         return (B_FALSE);
1060 }
1061
1062 /*
1063  * Go through and find any devices that are in use.  We rely on libdiskmgt for
1064  * the majority of this task.
1065  */
1066 static int
1067 check_in_use(nvlist_t *config, nvlist_t *nv, boolean_t force,
1068     boolean_t replacing, boolean_t isspare)
1069 {
1070         nvlist_t **child;
1071         uint_t c, children;
1072         char *type, *path;
1073         int ret = 0;
1074         char buf[MAXPATHLEN];
1075         uint64_t wholedisk = B_FALSE;
1076
1077         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
1078
1079         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
1080             &child, &children) != 0) {
1081
1082                 verify(!nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path));
1083                 if (strcmp(type, VDEV_TYPE_DISK) == 0)
1084                         verify(!nvlist_lookup_uint64(nv,
1085                                ZPOOL_CONFIG_WHOLE_DISK, &wholedisk));
1086
1087                 /*
1088                  * As a generic check, we look to see if this is a replace of a
1089                  * hot spare within the same pool.  If so, we allow it
1090                  * regardless of what libblkid or zpool_in_use() says.
1091                  */
1092                 if (replacing) {
1093                         if (wholedisk)
1094                                 (void) snprintf(buf, sizeof (buf), "%ss0",
1095                                     path);
1096                         else
1097                                 (void) strlcpy(buf, path, sizeof (buf));
1098
1099                         if (is_spare(config, buf))
1100                                 return (0);
1101                 }
1102
1103                 if (strcmp(type, VDEV_TYPE_DISK) == 0)
1104                         ret = check_device(path, force, isspare, wholedisk);
1105
1106                 if (strcmp(type, VDEV_TYPE_FILE) == 0)
1107                         ret = check_file(path, force, isspare);
1108
1109                 return (ret);
1110         }
1111
1112         for (c = 0; c < children; c++)
1113                 if ((ret = check_in_use(config, child[c], force,
1114                     replacing, B_FALSE)) != 0)
1115                         return (ret);
1116
1117         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_SPARES,
1118             &child, &children) == 0)
1119                 for (c = 0; c < children; c++)
1120                         if ((ret = check_in_use(config, child[c], force,
1121                             replacing, B_TRUE)) != 0)
1122                                 return (ret);
1123
1124         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1125             &child, &children) == 0)
1126                 for (c = 0; c < children; c++)
1127                         if ((ret = check_in_use(config, child[c], force,
1128                             replacing, B_FALSE)) != 0)
1129                                 return (ret);
1130
1131         return (0);
1132 }
1133
1134 static const char *
1135 is_grouping(const char *type, int *mindev, int *maxdev)
1136 {
1137         if (strncmp(type, "raidz", 5) == 0) {
1138                 const char *p = type + 5;
1139                 char *end;
1140                 long nparity;
1141
1142                 if (*p == '\0') {
1143                         nparity = 1;
1144                 } else if (*p == '0') {
1145                         return (NULL); /* no zero prefixes allowed */
1146                 } else {
1147                         errno = 0;
1148                         nparity = strtol(p, &end, 10);
1149                         if (errno != 0 || nparity < 1 || nparity >= 255 ||
1150                             *end != '\0')
1151                                 return (NULL);
1152                 }
1153
1154                 if (mindev != NULL)
1155                         *mindev = nparity + 1;
1156                 if (maxdev != NULL)
1157                         *maxdev = 255;
1158                 return (VDEV_TYPE_RAIDZ);
1159         }
1160
1161         if (maxdev != NULL)
1162                 *maxdev = INT_MAX;
1163
1164         if (strcmp(type, "mirror") == 0) {
1165                 if (mindev != NULL)
1166                         *mindev = 2;
1167                 return (VDEV_TYPE_MIRROR);
1168         }
1169
1170         if (strcmp(type, "spare") == 0) {
1171                 if (mindev != NULL)
1172                         *mindev = 1;
1173                 return (VDEV_TYPE_SPARE);
1174         }
1175
1176         if (strcmp(type, "log") == 0) {
1177                 if (mindev != NULL)
1178                         *mindev = 1;
1179                 return (VDEV_TYPE_LOG);
1180         }
1181
1182         if (strcmp(type, "cache") == 0) {
1183                 if (mindev != NULL)
1184                         *mindev = 1;
1185                 return (VDEV_TYPE_L2CACHE);
1186         }
1187
1188         return (NULL);
1189 }
1190
1191 /*
1192  * Construct a syntactically valid vdev specification,
1193  * and ensure that all devices and files exist and can be opened.
1194  * Note: we don't bother freeing anything in the error paths
1195  * because the program is just going to exit anyway.
1196  */
1197 nvlist_t *
1198 construct_spec(int argc, char **argv)
1199 {
1200         nvlist_t *nvroot, *nv, **top, **spares, **l2cache;
1201         int t, toplevels, mindev, maxdev, nspares, nlogs, nl2cache;
1202         const char *type;
1203         uint64_t is_log;
1204         boolean_t seen_logs;
1205
1206         top = NULL;
1207         toplevels = 0;
1208         spares = NULL;
1209         l2cache = NULL;
1210         nspares = 0;
1211         nlogs = 0;
1212         nl2cache = 0;
1213         is_log = B_FALSE;
1214         seen_logs = B_FALSE;
1215
1216         while (argc > 0) {
1217                 nv = NULL;
1218
1219                 /*
1220                  * If it's a mirror or raidz, the subsequent arguments are
1221                  * its leaves -- until we encounter the next mirror or raidz.
1222                  */
1223                 if ((type = is_grouping(argv[0], &mindev, &maxdev)) != NULL) {
1224                         nvlist_t **child = NULL;
1225                         int c, children = 0;
1226
1227                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1228                                 if (spares != NULL) {
1229                                         (void) fprintf(stderr,
1230                                             gettext("invalid vdev "
1231                                             "specification: 'spare' can be "
1232                                             "specified only once\n"));
1233                                         return (NULL);
1234                                 }
1235                                 is_log = B_FALSE;
1236                         }
1237
1238                         if (strcmp(type, VDEV_TYPE_LOG) == 0) {
1239                                 if (seen_logs) {
1240                                         (void) fprintf(stderr,
1241                                             gettext("invalid vdev "
1242                                             "specification: 'log' can be "
1243                                             "specified only once\n"));
1244                                         return (NULL);
1245                                 }
1246                                 seen_logs = B_TRUE;
1247                                 is_log = B_TRUE;
1248                                 argc--;
1249                                 argv++;
1250                                 /*
1251                                  * A log is not a real grouping device.
1252                                  * We just set is_log and continue.
1253                                  */
1254                                 continue;
1255                         }
1256
1257                         if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1258                                 if (l2cache != NULL) {
1259                                         (void) fprintf(stderr,
1260                                             gettext("invalid vdev "
1261                                             "specification: 'cache' can be "
1262                                             "specified only once\n"));
1263                                         return (NULL);
1264                                 }
1265                                 is_log = B_FALSE;
1266                         }
1267
1268                         if (is_log) {
1269                                 if (strcmp(type, VDEV_TYPE_MIRROR) != 0) {
1270                                         (void) fprintf(stderr,
1271                                             gettext("invalid vdev "
1272                                             "specification: unsupported 'log' "
1273                                             "device: %s\n"), type);
1274                                         return (NULL);
1275                                 }
1276                                 nlogs++;
1277                         }
1278
1279                         for (c = 1; c < argc; c++) {
1280                                 if (is_grouping(argv[c], NULL, NULL) != NULL)
1281                                         break;
1282                                 children++;
1283                                 child = realloc(child,
1284                                     children * sizeof (nvlist_t *));
1285                                 if (child == NULL)
1286                                         zpool_no_memory();
1287                                 if ((nv = make_leaf_vdev(argv[c], B_FALSE))
1288                                     == NULL)
1289                                         return (NULL);
1290                                 child[children - 1] = nv;
1291                         }
1292
1293                         if (children < mindev) {
1294                                 (void) fprintf(stderr, gettext("invalid vdev "
1295                                     "specification: %s requires at least %d "
1296                                     "devices\n"), argv[0], mindev);
1297                                 return (NULL);
1298                         }
1299
1300                         if (children > maxdev) {
1301                                 (void) fprintf(stderr, gettext("invalid vdev "
1302                                     "specification: %s supports no more than "
1303                                     "%d devices\n"), argv[0], maxdev);
1304                                 return (NULL);
1305                         }
1306
1307                         argc -= c;
1308                         argv += c;
1309
1310                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1311                                 spares = child;
1312                                 nspares = children;
1313                                 continue;
1314                         } else if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1315                                 l2cache = child;
1316                                 nl2cache = children;
1317                                 continue;
1318                         } else {
1319                                 verify(nvlist_alloc(&nv, NV_UNIQUE_NAME,
1320                                     0) == 0);
1321                                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_TYPE,
1322                                     type) == 0);
1323                                 verify(nvlist_add_uint64(nv,
1324                                     ZPOOL_CONFIG_IS_LOG, is_log) == 0);
1325                                 if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
1326                                         verify(nvlist_add_uint64(nv,
1327                                             ZPOOL_CONFIG_NPARITY,
1328                                             mindev - 1) == 0);
1329                                 }
1330                                 verify(nvlist_add_nvlist_array(nv,
1331                                     ZPOOL_CONFIG_CHILDREN, child,
1332                                     children) == 0);
1333
1334                                 for (c = 0; c < children; c++)
1335                                         nvlist_free(child[c]);
1336                                 free(child);
1337                         }
1338                 } else {
1339                         /*
1340                          * We have a device.  Pass off to make_leaf_vdev() to
1341                          * construct the appropriate nvlist describing the vdev.
1342                          */
1343                         if ((nv = make_leaf_vdev(argv[0], is_log)) == NULL)
1344                                 return (NULL);
1345                         if (is_log)
1346                                 nlogs++;
1347                         argc--;
1348                         argv++;
1349                 }
1350
1351                 toplevels++;
1352                 top = realloc(top, toplevels * sizeof (nvlist_t *));
1353                 if (top == NULL)
1354                         zpool_no_memory();
1355                 top[toplevels - 1] = nv;
1356         }
1357
1358         if (toplevels == 0 && nspares == 0 && nl2cache == 0) {
1359                 (void) fprintf(stderr, gettext("invalid vdev "
1360                     "specification: at least one toplevel vdev must be "
1361                     "specified\n"));
1362                 return (NULL);
1363         }
1364
1365         if (seen_logs && nlogs == 0) {
1366                 (void) fprintf(stderr, gettext("invalid vdev specification: "
1367                     "log requires at least 1 device\n"));
1368                 return (NULL);
1369         }
1370
1371         /*
1372          * Finally, create nvroot and add all top-level vdevs to it.
1373          */
1374         verify(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, 0) == 0);
1375         verify(nvlist_add_string(nvroot, ZPOOL_CONFIG_TYPE,
1376             VDEV_TYPE_ROOT) == 0);
1377         verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
1378             top, toplevels) == 0);
1379         if (nspares != 0)
1380                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1381                     spares, nspares) == 0);
1382         if (nl2cache != 0)
1383                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
1384                     l2cache, nl2cache) == 0);
1385
1386         for (t = 0; t < toplevels; t++)
1387                 nvlist_free(top[t]);
1388         for (t = 0; t < nspares; t++)
1389                 nvlist_free(spares[t]);
1390         for (t = 0; t < nl2cache; t++)
1391                 nvlist_free(l2cache[t]);
1392         if (spares)
1393                 free(spares);
1394         if (l2cache)
1395                 free(l2cache);
1396         free(top);
1397
1398         return (nvroot);
1399 }
1400
1401 nvlist_t *
1402 split_mirror_vdev(zpool_handle_t *zhp, char *newname, nvlist_t *props,
1403     splitflags_t flags, int argc, char **argv)
1404 {
1405         nvlist_t *newroot = NULL, **child;
1406         uint_t c, children;
1407
1408         if (argc > 0) {
1409                 if ((newroot = construct_spec(argc, argv)) == NULL) {
1410                         (void) fprintf(stderr, gettext("Unable to build a "
1411                             "pool from the specified devices\n"));
1412                         return (NULL);
1413                 }
1414
1415                 if (!flags.dryrun && make_disks(zhp, newroot) != 0) {
1416                         nvlist_free(newroot);
1417                         return (NULL);
1418                 }
1419
1420                 /* avoid any tricks in the spec */
1421                 verify(nvlist_lookup_nvlist_array(newroot,
1422                     ZPOOL_CONFIG_CHILDREN, &child, &children) == 0);
1423                 for (c = 0; c < children; c++) {
1424                         char *path;
1425                         const char *type;
1426                         int min, max;
1427
1428                         verify(nvlist_lookup_string(child[c],
1429                             ZPOOL_CONFIG_PATH, &path) == 0);
1430                         if ((type = is_grouping(path, &min, &max)) != NULL) {
1431                                 (void) fprintf(stderr, gettext("Cannot use "
1432                                     "'%s' as a device for splitting\n"), type);
1433                                 nvlist_free(newroot);
1434                                 return (NULL);
1435                         }
1436                 }
1437         }
1438
1439         if (zpool_vdev_split(zhp, newname, &newroot, props, flags) != 0) {
1440                 if (newroot != NULL)
1441                         nvlist_free(newroot);
1442                 return (NULL);
1443         }
1444
1445         return (newroot);
1446 }
1447
1448 /*
1449  * Get and validate the contents of the given vdev specification.  This ensures
1450  * that the nvlist returned is well-formed, that all the devices exist, and that
1451  * they are not currently in use by any other known consumer.  The 'poolconfig'
1452  * parameter is the current configuration of the pool when adding devices
1453  * existing pool, and is used to perform additional checks, such as changing the
1454  * replication level of the pool.  It can be 'NULL' to indicate that this is a
1455  * new pool.  The 'force' flag controls whether devices should be forcefully
1456  * added, even if they appear in use.
1457  */
1458 nvlist_t *
1459 make_root_vdev(zpool_handle_t *zhp, int force, int check_rep,
1460     boolean_t replacing, boolean_t dryrun, int argc, char **argv)
1461 {
1462         nvlist_t *newroot;
1463         nvlist_t *poolconfig = NULL;
1464         is_force = force;
1465
1466         /*
1467          * Construct the vdev specification.  If this is successful, we know
1468          * that we have a valid specification, and that all devices can be
1469          * opened.
1470          */
1471         if ((newroot = construct_spec(argc, argv)) == NULL)
1472                 return (NULL);
1473
1474         if (zhp && ((poolconfig = zpool_get_config(zhp, NULL)) == NULL))
1475                 return (NULL);
1476
1477         /*
1478          * Validate each device to make sure that its not shared with another
1479          * subsystem.  We do this even if 'force' is set, because there are some
1480          * uses (such as a dedicated dump device) that even '-f' cannot
1481          * override.
1482          */
1483         if (check_in_use(poolconfig, newroot, force, replacing, B_FALSE) != 0) {
1484                 nvlist_free(newroot);
1485                 return (NULL);
1486         }
1487
1488         /*
1489          * Check the replication level of the given vdevs and report any errors
1490          * found.  We include the existing pool spec, if any, as we need to
1491          * catch changes against the existing replication level.
1492          */
1493         if (check_rep && check_replication(poolconfig, newroot) != 0) {
1494                 nvlist_free(newroot);
1495                 return (NULL);
1496         }
1497
1498         /*
1499          * Run through the vdev specification and label any whole disks found.
1500          */
1501         if (!dryrun && make_disks(zhp, newroot) != 0) {
1502                 nvlist_free(newroot);
1503                 return (NULL);
1504         }
1505
1506         return (newroot);
1507 }