cddl/contrib/opensolaris/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  * Copyright (c) 2013 by Delphix. All rights reserved.
  25  * Copyright 2016 Igor Kozhukhov <ikozhukhov@gmail.com>.
  26  */
  27
  28 /*
  29  * Functions to convert between a list of vdevs and an nvlist representing the
  30  * configuration.  Each entry in the list can be one of:
  31  *
  32  *      Device vdevs
  33  *              disk=(path=..., devid=...)
  34  *              file=(path=...)
  35  *
  36  *      Group vdevs
  37  *              raidz[1|2]=(...)
  38  *              mirror=(...)
  39  *
  40  *      Hot spares
  41  *
  42  * While the underlying implementation supports it, group vdevs cannot contain
  43  * other group vdevs.  All userland verification of devices is contained within
  44  * this file.  If successful, the nvlist returned can be passed directly to the
  45  * kernel; we've done as much verification as possible in userland.
  46  *
  47  * Hot spares are a special case, and passed down as an array of disk vdevs, at
  48  * the same level as the root of the vdev tree.
  49  *
  50  * The only function exported by this file is 'make_root_vdev'.  The
  51  * function performs several passes:
  52  *
  53  *      1. Construct the vdev specification.  Performs syntax validation and
  54  *         makes sure each device is valid.
  55  *      2. Check for devices in use.  Using libdiskmgt, makes sure that no
  56  *         devices are also in use.  Some can be overridden using the 'force'
  57  *         flag, others cannot.
  58  *      3. Check for replication errors if the 'force' flag is not specified.
  59  *         validates that the replication level is consistent across the
  60  *         entire pool.
  61  *      4. Call libzfs to label any whole disks with an EFI label.
  62  */
  63
  64 #include <assert.h>
  65 #include <devid.h>
  66 #include <errno.h>
  67 #include <fcntl.h>
  68 #include <libintl.h>
  69 #include <libnvpair.h>
  70 #include <limits.h>
  71 #include <stdio.h>
  72 #include <string.h>
  73 #include <unistd.h>
  74 #include <paths.h>
  75 #include <sys/stat.h>
  76 #include <sys/disk.h>
  77 #include <sys/mntent.h>
  78 #include <libgeom.h>
  79
  80 #include "zpool_util.h"
  81
  82 #define BACKUP_SLICE    "s2"
  83
  84 /*
  85  * For any given vdev specification, we can have multiple errors.  The
  86  * vdev_error() function keeps track of whether we have seen an error yet, and
  87  * prints out a header if its the first error we've seen.
  88  */
  89 boolean_t error_seen;
  90 boolean_t is_force;
  91
  92 /*PRINTFLIKE1*/
  93 static void
  94 vdev_error(const char *fmt, ...)
  95 {
  96         va_list ap;
  97
  98         if (!error_seen) {
  99                 (void) fprintf(stderr, gettext("invalid vdev specification\n"));
 100                 if (!is_force)
 101                         (void) fprintf(stderr, gettext("use '-f' to override "
 102                             "the following errors:\n"));
 103                 else
 104                         (void) fprintf(stderr, gettext("the following errors "
 105                             "must be manually repaired:\n"));
 106                 error_seen = B_TRUE;
 107         }
 108
 109         va_start(ap, fmt);
 110         (void) vfprintf(stderr, fmt, ap);
 111         va_end(ap);
 112 }
 113
 114 #ifdef illumos
 115 static void
 116 libdiskmgt_error(int error)
 117 {
 118         /*
 119          * ENXIO/ENODEV is a valid error message if the device doesn't live in
 120          * /dev/dsk.  Don't bother printing an error message in this case.
 121          */
 122         if (error == ENXIO || error == ENODEV)
 123                 return;
 124
 125         (void) fprintf(stderr, gettext("warning: device in use checking "
 126             "failed: %s\n"), strerror(error));
 127 }
 128
 129 /*
 130  * Validate a device, passing the bulk of the work off to libdiskmgt.
 131  */
 132 static int
 133 check_slice(const char *path, int force, boolean_t wholedisk, boolean_t isspare)
 134 {
 135         char *msg;
 136         int error = 0;
 137         dm_who_type_t who;
 138
 139         if (force)
 140                 who = DM_WHO_ZPOOL_FORCE;
 141         else if (isspare)
 142                 who = DM_WHO_ZPOOL_SPARE;
 143         else
 144                 who = DM_WHO_ZPOOL;
 145
 146         if (dm_inuse((char *)path, &msg, who, &error) || error) {
 147                 if (error != 0) {
 148                         libdiskmgt_error(error);
 149                         return (0);
 150                 } else {
 151                         vdev_error("%s", msg);
 152                         free(msg);
 153                         return (-1);
 154                 }
 155         }
 156
 157         /*
 158          * If we're given a whole disk, ignore overlapping slices since we're
 159          * about to label it anyway.
 160          */
 161         error = 0;
 162         if (!wholedisk && !force &&
 163             (dm_isoverlapping((char *)path, &msg, &error) || error)) {
 164                 if (error == 0) {
 165                         /* dm_isoverlapping returned -1 */
 166                         vdev_error(gettext("%s overlaps with %s\n"), path, msg);
 167                         free(msg);
 168                         return (-1);
 169                 } else if (error != ENODEV) {
 170                         /* libdiskmgt's devcache only handles physical drives */
 171                         libdiskmgt_error(error);
 172                         return (0);
 173                 }
 174         }
 175
 176         return (0);
 177 }
 178
 179
 180 /*
 181  * Validate a whole disk.  Iterate over all slices on the disk and make sure
 182  * that none is in use by calling check_slice().
 183  */
 184 static int
 185 check_disk(const char *name, dm_descriptor_t disk, int force, int isspare)
 186 {
 187         dm_descriptor_t *drive, *media, *slice;
 188         int err = 0;
 189         int i;
 190         int ret;
 191
 192         /*
 193          * Get the drive associated with this disk.  This should never fail,
 194          * because we already have an alias handle open for the device.
 195          */
 196         if ((drive = dm_get_associated_descriptors(disk, DM_DRIVE,
 197             &err)) == NULL || *drive == NULL) {
 198                 if (err)
 199                         libdiskmgt_error(err);
 200                 return (0);
 201         }
 202
 203         if ((media = dm_get_associated_descriptors(*drive, DM_MEDIA,
 204             &err)) == NULL) {
 205                 dm_free_descriptors(drive);
 206                 if (err)
 207                         libdiskmgt_error(err);
 208                 return (0);
 209         }
 210
 211         dm_free_descriptors(drive);
 212
 213         /*
 214          * It is possible that the user has specified a removable media drive,
 215          * and the media is not present.
 216          */
 217         if (*media == NULL) {
 218                 dm_free_descriptors(media);
 219                 vdev_error(gettext("'%s' has no media in drive\n"), name);
 220                 return (-1);
 221         }
 222
 223         if ((slice = dm_get_associated_descriptors(*media, DM_SLICE,
 224             &err)) == NULL) {
 225                 dm_free_descriptors(media);
 226                 if (err)
 227                         libdiskmgt_error(err);
 228                 return (0);
 229         }
 230
 231         dm_free_descriptors(media);
 232
 233         ret = 0;
 234
 235         /*
 236          * Iterate over all slices and report any errors.  We don't care about
 237          * overlapping slices because we are using the whole disk.
 238          */
 239         for (i = 0; slice[i] != NULL; i++) {
 240                 char *name = dm_get_name(slice[i], &err);
 241
 242                 if (check_slice(name, force, B_TRUE, isspare) != 0)
 243                         ret = -1;
 244
 245                 dm_free_name(name);
 246         }
 247
 248         dm_free_descriptors(slice);
 249         return (ret);
 250 }
 251
 252 /*
 253  * Validate a device.
 254  */
 255 static int
 256 check_device(const char *path, boolean_t force, boolean_t isspare)
 257 {
 258         dm_descriptor_t desc;
 259         int err;
 260         char *dev;
 261
 262         /*
 263          * For whole disks, libdiskmgt does not include the leading dev path.
 264          */
 265         dev = strrchr(path, '/');
 266         assert(dev != NULL);
 267         dev++;
 268         if ((desc = dm_get_descriptor_by_name(DM_ALIAS, dev, &err)) != NULL) {
 269                 err = check_disk(path, desc, force, isspare);
 270                 dm_free_descriptor(desc);
 271                 return (err);
 272         }
 273
 274         return (check_slice(path, force, B_FALSE, isspare));
 275 }
 276 #endif  /* illumos */
 277
 278 /*
 279  * Check that a file is valid.  All we can do in this case is check that it's
 280  * not in use by another pool, and not in use by swap.
 281  */
 282 static int
 283 check_file(const char *file, boolean_t force, boolean_t isspare)
 284 {
 285         char  *name;
 286         int fd;
 287         int ret = 0;
 288         int err;
 289         pool_state_t state;
 290         boolean_t inuse;
 291
 292 #ifdef illumos
 293         if (dm_inuse_swap(file, &err)) {
 294                 if (err)
 295                         libdiskmgt_error(err);
 296                 else
 297                         vdev_error(gettext("%s is currently used by swap. "
 298                             "Please see swap(1M).\n"), file);
 299                 return (-1);
 300         }
 301 #endif
 302
 303         if ((fd = open(file, O_RDONLY)) < 0)
 304                 return (0);
 305
 306         if (zpool_in_use(g_zfs, fd, &state, &name, &inuse) == 0 && inuse) {
 307                 const char *desc;
 308
 309                 switch (state) {
 310                 case POOL_STATE_ACTIVE:
 311                         desc = gettext("active");
 312                         break;
 313
 314                 case POOL_STATE_EXPORTED:
 315                         desc = gettext("exported");
 316                         break;
 317
 318                 case POOL_STATE_POTENTIALLY_ACTIVE:
 319                         desc = gettext("potentially active");
 320                         break;
 321
 322                 default:
 323                         desc = gettext("unknown");
 324                         break;
 325                 }
 326
 327                 /*
 328                  * Allow hot spares to be shared between pools.
 329                  */
 330                 if (state == POOL_STATE_SPARE && isspare)
 331                         return (0);
 332
 333                 if (state == POOL_STATE_ACTIVE ||
 334                     state == POOL_STATE_SPARE || !force) {
 335                         switch (state) {
 336                         case POOL_STATE_SPARE:
 337                                 vdev_error(gettext("%s is reserved as a hot "
 338                                     "spare for pool %s\n"), file, name);
 339                                 break;
 340                         default:
 341                                 vdev_error(gettext("%s is part of %s pool "
 342                                     "'%s'\n"), file, desc, name);
 343                                 break;
 344                         }
 345                         ret = -1;
 346                 }
 347
 348                 free(name);
 349         }
 350
 351         (void) close(fd);
 352         return (ret);
 353 }
 354
 355 static int
 356 check_device(const char *name, boolean_t force, boolean_t isspare)
 357 {
 358         char path[MAXPATHLEN];
 359
 360         if (strncmp(name, _PATH_DEV, sizeof(_PATH_DEV) - 1) != 0)
 361                 snprintf(path, sizeof(path), "%s%s", _PATH_DEV, name);
 362         else
 363                 strlcpy(path, name, sizeof(path));
 364
 365         return (check_file(path, force, isspare));
 366 }
 367
 368 /*
 369  * By "whole disk" we mean an entire physical disk (something we can
 370  * label, toggle the write cache on, etc.) as opposed to the full
 371  * capacity of a pseudo-device such as lofi or did.  We act as if we
 372  * are labeling the disk, which should be a pretty good test of whether
 373  * it's a viable device or not.  Returns B_TRUE if it is and B_FALSE if
 374  * it isn't.
 375  */
 376 static boolean_t
 377 is_whole_disk(const char *arg)
 378 {
 379 #ifdef illumos
 380         struct dk_gpt *label;
 381         int     fd;
 382         char    path[MAXPATHLEN];
 383
 384         (void) snprintf(path, sizeof (path), "%s%s%s",
 385             ZFS_RDISK_ROOT, strrchr(arg, '/'), BACKUP_SLICE);
 386         if ((fd = open(path, O_RDWR | O_NDELAY)) < 0)
 387                 return (B_FALSE);
 388         if (efi_alloc_and_init(fd, EFI_NUMPAR, &label) != 0) {
 389                 (void) close(fd);
 390                 return (B_FALSE);
 391         }
 392         efi_free(label);
 393         (void) close(fd);
 394         return (B_TRUE);
 395 #else
 396         int fd;
 397
 398         fd = g_open(arg, 0);
 399         if (fd >= 0) {
 400                 g_close(fd);
 401                 return (B_TRUE);
 402         }
 403         return (B_FALSE);
 404 #endif
 405 }
 406
 407 /*
 408  * Create a leaf vdev.  Determine if this is a file or a device.  If it's a
 409  * device, fill in the device id to make a complete nvlist.  Valid forms for a
 410  * leaf vdev are:
 411  *
 412  *      /dev/dsk/xxx    Complete disk path
 413  *      /xxx            Full path to file
 414  *      xxx             Shorthand for /dev/dsk/xxx
 415  */
 416 static nvlist_t *
 417 make_leaf_vdev(const char *arg, uint64_t is_log)
 418 {
 419         char path[MAXPATHLEN];
 420         struct stat64 statbuf;
 421         nvlist_t *vdev = NULL;
 422         char *type = NULL;
 423         boolean_t wholedisk = B_FALSE;
 424
 425         /*
 426          * Determine what type of vdev this is, and put the full path into
 427          * 'path'.  We detect whether this is a device of file afterwards by
 428          * checking the st_mode of the file.
 429          */
 430         if (arg[0] == '/') {
 431                 /*
 432                  * Complete device or file path.  Exact type is determined by
 433                  * examining the file descriptor afterwards.
 434                  */
 435                 wholedisk = is_whole_disk(arg);
 436                 if (!wholedisk && (stat64(arg, &statbuf) != 0)) {
 437                         (void) fprintf(stderr,
 438                             gettext("cannot open '%s': %s\n"),
 439                             arg, strerror(errno));
 440                         return (NULL);
 441                 }
 442
 443                 (void) strlcpy(path, arg, sizeof (path));
 444         } else {
 445                 /*
 446                  * This may be a short path for a device, or it could be total
 447                  * gibberish.  Check to see if it's a known device in
 448                  * /dev/dsk/.  As part of this check, see if we've been given a
 449                  * an entire disk (minus the slice number).
 450                  */
 451                 if (strncmp(arg, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
 452                         strlcpy(path, arg, sizeof (path));
 453                 else
 454                         snprintf(path, sizeof (path), "%s%s", _PATH_DEV, arg);
 455                 wholedisk = is_whole_disk(path);
 456                 if (!wholedisk && (stat64(path, &statbuf) != 0)) {
 457                         /*
 458                          * If we got ENOENT, then the user gave us
 459                          * gibberish, so try to direct them with a
 460                          * reasonable error message.  Otherwise,
 461                          * regurgitate strerror() since it's the best we
 462                          * can do.
 463                          */
 464                         if (errno == ENOENT) {
 465                                 (void) fprintf(stderr,
 466                                     gettext("cannot open '%s': no such "
 467                                     "GEOM provider\n"), arg);
 468                                 (void) fprintf(stderr,
 469                                     gettext("must be a full path or "
 470                                     "shorthand device name\n"));
 471                                 return (NULL);
 472                         } else {
 473                                 (void) fprintf(stderr,
 474                                     gettext("cannot open '%s': %s\n"),
 475                                     path, strerror(errno));
 476                                 return (NULL);
 477                         }
 478                 }
 479         }
 480
 481 #ifdef __FreeBSD__
 482         if (S_ISCHR(statbuf.st_mode)) {
 483                 statbuf.st_mode &= ~S_IFCHR;
 484                 statbuf.st_mode |= S_IFBLK;
 485                 wholedisk = B_FALSE;
 486         }
 487 #endif
 488
 489         /*
 490          * Determine whether this is a device or a file.
 491          */
 492         if (wholedisk || S_ISBLK(statbuf.st_mode)) {
 493                 type = VDEV_TYPE_DISK;
 494         } else if (S_ISREG(statbuf.st_mode)) {
 495                 type = VDEV_TYPE_FILE;
 496         } else {
 497                 (void) fprintf(stderr, gettext("cannot use '%s': must be a "
 498                     "GEOM provider or regular file\n"), path);
 499                 return (NULL);
 500         }
 501
 502         /*
 503          * Finally, we have the complete device or file, and we know that it is
 504          * acceptable to use.  Construct the nvlist to describe this vdev.  All
 505          * vdevs have a 'path' element, and devices also have a 'devid' element.
 506          */
 507         verify(nvlist_alloc(&vdev, NV_UNIQUE_NAME, 0) == 0);
 508         verify(nvlist_add_string(vdev, ZPOOL_CONFIG_PATH, path) == 0);
 509         verify(nvlist_add_string(vdev, ZPOOL_CONFIG_TYPE, type) == 0);
 510         verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_IS_LOG, is_log) == 0);
 511         if (strcmp(type, VDEV_TYPE_DISK) == 0)
 512                 verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_WHOLE_DISK,
 513                     (uint64_t)wholedisk) == 0);
 514
 515 #ifdef have_devid
 516         /*
 517          * For a whole disk, defer getting its devid until after labeling it.
 518          */
 519         if (S_ISBLK(statbuf.st_mode) && !wholedisk) {
 520                 /*
 521                  * Get the devid for the device.
 522                  */
 523                 int fd;
 524                 ddi_devid_t devid;
 525                 char *minor = NULL, *devid_str = NULL;
 526
 527                 if ((fd = open(path, O_RDONLY)) < 0) {
 528                         (void) fprintf(stderr, gettext("cannot open '%s': "
 529                             "%s\n"), path, strerror(errno));
 530                         nvlist_free(vdev);
 531                         return (NULL);
 532                 }
 533
 534                 if (devid_get(fd, &devid) == 0) {
 535                         if (devid_get_minor_name(fd, &minor) == 0 &&
 536                             (devid_str = devid_str_encode(devid, minor)) !=
 537                             NULL) {
 538                                 verify(nvlist_add_string(vdev,
 539                                     ZPOOL_CONFIG_DEVID, devid_str) == 0);
 540                         }
 541                         if (devid_str != NULL)
 542                                 devid_str_free(devid_str);
 543                         if (minor != NULL)
 544                                 devid_str_free(minor);
 545                         devid_free(devid);
 546                 }
 547
 548                 (void) close(fd);
 549         }
 550 #endif
 551
 552         return (vdev);
 553 }
 554
 555 /*
 556  * Go through and verify the replication level of the pool is consistent.
 557  * Performs the following checks:
 558  *
 559  *      For the new spec, verifies that devices in mirrors and raidz are the
 560  *      same size.
 561  *
 562  *      If the current configuration already has inconsistent replication
 563  *      levels, ignore any other potential problems in the new spec.
 564  *
 565  *      Otherwise, make sure that the current spec (if there is one) and the new
 566  *      spec have consistent replication levels.
 567  */
 568 typedef struct replication_level {
 569         char *zprl_type;
 570         uint64_t zprl_children;
 571         uint64_t zprl_parity;
 572 } replication_level_t;
 573
 574 #define ZPOOL_FUZZ      (16 * 1024 * 1024)
 575
 576 /*
 577  * Given a list of toplevel vdevs, return the current replication level.  If
 578  * the config is inconsistent, then NULL is returned.  If 'fatal' is set, then
 579  * an error message will be displayed for each self-inconsistent vdev.
 580  */
 581 static replication_level_t *
 582 get_replication(nvlist_t *nvroot, boolean_t fatal)
 583 {
 584         nvlist_t **top;
 585         uint_t t, toplevels;
 586         nvlist_t **child;
 587         uint_t c, children;
 588         nvlist_t *nv;
 589         char *type;
 590         replication_level_t lastrep = {0};
 591         replication_level_t rep;
 592         replication_level_t *ret;
 593         boolean_t dontreport;
 594
 595         ret = safe_malloc(sizeof (replication_level_t));
 596
 597         verify(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
 598             &top, &toplevels) == 0);
 599
 600         lastrep.zprl_type = NULL;
 601         for (t = 0; t < toplevels; t++) {
 602                 uint64_t is_log = B_FALSE;
 603
 604                 nv = top[t];
 605
 606                 /*
 607                  * For separate logs we ignore the top level vdev replication
 608                  * constraints.
 609                  */
 610                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &is_log);
 611                 if (is_log)
 612                         continue;
 613
 614                 verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE,
 615                     &type) == 0);
 616                 if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 617                     &child, &children) != 0) {
 618                         /*
 619                          * This is a 'file' or 'disk' vdev.
 620                          */
 621                         rep.zprl_type = type;
 622                         rep.zprl_children = 1;
 623                         rep.zprl_parity = 0;
 624                 } else {
 625                         uint64_t vdev_size;
 626
 627                         /*
 628                          * This is a mirror or RAID-Z vdev.  Go through and make
 629                          * sure the contents are all the same (files vs. disks),
 630                          * keeping track of the number of elements in the
 631                          * process.
 632                          *
 633                          * We also check that the size of each vdev (if it can
 634                          * be determined) is the same.
 635                          */
 636                         rep.zprl_type = type;
 637                         rep.zprl_children = 0;
 638
 639                         if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
 640                                 verify(nvlist_lookup_uint64(nv,
 641                                     ZPOOL_CONFIG_NPARITY,
 642                                     &rep.zprl_parity) == 0);
 643                                 assert(rep.zprl_parity != 0);
 644                         } else {
 645                                 rep.zprl_parity = 0;
 646                         }
 647
 648                         /*
 649                          * The 'dontreport' variable indicates that we've
 650                          * already reported an error for this spec, so don't
 651                          * bother doing it again.
 652                          */
 653                         type = NULL;
 654                         dontreport = 0;
 655                         vdev_size = -1ULL;
 656                         for (c = 0; c < children; c++) {
 657                                 nvlist_t *cnv = child[c];
 658                                 char *path;
 659                                 struct stat64 statbuf;
 660                                 uint64_t size = -1ULL;
 661                                 char *childtype;
 662                                 int fd, err;
 663
 664                                 rep.zprl_children++;
 665
 666                                 verify(nvlist_lookup_string(cnv,
 667                                     ZPOOL_CONFIG_TYPE, &childtype) == 0);
 668
 669                                 /*
 670                                  * If this is a replacing or spare vdev, then
 671                                  * get the real first child of the vdev.
 672                                  */
 673                                 if (strcmp(childtype,
 674                                     VDEV_TYPE_REPLACING) == 0 ||
 675                                     strcmp(childtype, VDEV_TYPE_SPARE) == 0) {
 676                                         nvlist_t **rchild;
 677                                         uint_t rchildren;
 678
 679                                         verify(nvlist_lookup_nvlist_array(cnv,
 680                                             ZPOOL_CONFIG_CHILDREN, &rchild,
 681                                             &rchildren) == 0);
 682                                         assert(rchildren == 2);
 683                                         cnv = rchild[0];
 684
 685                                         verify(nvlist_lookup_string(cnv,
 686                                             ZPOOL_CONFIG_TYPE,
 687                                             &childtype) == 0);
 688                                 }
 689
 690                                 verify(nvlist_lookup_string(cnv,
 691                                     ZPOOL_CONFIG_PATH, &path) == 0);
 692
 693                                 /*
 694                                  * If we have a raidz/mirror that combines disks
 695                                  * with files, report it as an error.
 696                                  */
 697                                 if (!dontreport && type != NULL &&
 698                                     strcmp(type, childtype) != 0) {
 699                                         if (ret != NULL)
 700                                                 free(ret);
 701                                         ret = NULL;
 702                                         if (fatal)
 703                                                 vdev_error(gettext(
 704                                                     "mismatched replication "
 705                                                     "level: %s contains both "
 706                                                     "files and devices\n"),
 707                                                     rep.zprl_type);
 708                                         else
 709                                                 return (NULL);
 710                                         dontreport = B_TRUE;
 711                                 }
 712
 713                                 /*
 714                                  * According to stat(2), the value of 'st_size'
 715                                  * is undefined for block devices and character
 716                                  * devices.  But there is no effective way to
 717                                  * determine the real size in userland.
 718                                  *
 719                                  * Instead, we'll take advantage of an
 720                                  * implementation detail of spec_size().  If the
 721                                  * device is currently open, then we (should)
 722                                  * return a valid size.
 723                                  *
 724                                  * If we still don't get a valid size (indicated
 725                                  * by a size of 0 or MAXOFFSET_T), then ignore
 726                                  * this device altogether.
 727                                  */
 728                                 if ((fd = open(path, O_RDONLY)) >= 0) {
 729                                         err = fstat64(fd, &statbuf);
 730                                         (void) close(fd);
 731                                 } else {
 732                                         err = stat64(path, &statbuf);
 733                                 }
 734
 735                                 if (err != 0 ||
 736                                     statbuf.st_size == 0 ||
 737                                     statbuf.st_size == MAXOFFSET_T)
 738                                         continue;
 739
 740                                 size = statbuf.st_size;
 741
 742                                 /*
 743                                  * Also make sure that devices and
 744                                  * slices have a consistent size.  If
 745                                  * they differ by a significant amount
 746                                  * (~16MB) then report an error.
 747                                  */
 748                                 if (!dontreport &&
 749                                     (vdev_size != -1ULL &&
 750                                     (labs(size - vdev_size) >
 751                                     ZPOOL_FUZZ))) {
 752                                         if (ret != NULL)
 753                                                 free(ret);
 754                                         ret = NULL;
 755                                         if (fatal)
 756                                                 vdev_error(gettext(
 757                                                     "%s contains devices of "
 758                                                     "different sizes\n"),
 759                                                     rep.zprl_type);
 760                                         else
 761                                                 return (NULL);
 762                                         dontreport = B_TRUE;
 763                                 }
 764
 765                                 type = childtype;
 766                                 vdev_size = size;
 767                         }
 768                 }
 769
 770                 /*
 771                  * At this point, we have the replication of the last toplevel
 772                  * vdev in 'rep'.  Compare it to 'lastrep' to see if its
 773                  * different.
 774                  */
 775                 if (lastrep.zprl_type != NULL) {
 776                         if (strcmp(lastrep.zprl_type, rep.zprl_type) != 0) {
 777                                 if (ret != NULL)
 778                                         free(ret);
 779                                 ret = NULL;
 780                                 if (fatal)
 781                                         vdev_error(gettext(
 782                                             "mismatched replication level: "
 783                                             "both %s and %s vdevs are "
 784                                             "present\n"),
 785                                             lastrep.zprl_type, rep.zprl_type);
 786                                 else
 787                                         return (NULL);
 788                         } else if (lastrep.zprl_parity != rep.zprl_parity) {
 789                                 if (ret)
 790                                         free(ret);
 791                                 ret = NULL;
 792                                 if (fatal)
 793                                         vdev_error(gettext(
 794                                             "mismatched replication level: "
 795                                             "both %llu and %llu device parity "
 796                                             "%s vdevs are present\n"),
 797                                             lastrep.zprl_parity,
 798                                             rep.zprl_parity,
 799                                             rep.zprl_type);
 800                                 else
 801                                         return (NULL);
 802                         } else if (lastrep.zprl_children != rep.zprl_children) {
 803                                 if (ret)
 804                                         free(ret);
 805                                 ret = NULL;
 806                                 if (fatal)
 807                                         vdev_error(gettext(
 808                                             "mismatched replication level: "
 809                                             "both %llu-way and %llu-way %s "
 810                                             "vdevs are present\n"),
 811                                             lastrep.zprl_children,
 812                                             rep.zprl_children,
 813                                             rep.zprl_type);
 814                                 else
 815                                         return (NULL);
 816                         }
 817                 }
 818                 lastrep = rep;
 819         }
 820
 821         if (ret != NULL)
 822                 *ret = rep;
 823
 824         return (ret);
 825 }
 826
 827 /*
 828  * Check the replication level of the vdev spec against the current pool.  Calls
 829  * get_replication() to make sure the new spec is self-consistent.  If the pool
 830  * has a consistent replication level, then we ignore any errors.  Otherwise,
 831  * report any difference between the two.
 832  */
 833 static int
 834 check_replication(nvlist_t *config, nvlist_t *newroot)
 835 {
 836         nvlist_t **child;
 837         uint_t  children;
 838         replication_level_t *current = NULL, *new;
 839         int ret;
 840
 841         /*
 842          * If we have a current pool configuration, check to see if it's
 843          * self-consistent.  If not, simply return success.
 844          */
 845         if (config != NULL) {
 846                 nvlist_t *nvroot;
 847
 848                 verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
 849                     &nvroot) == 0);
 850                 if ((current = get_replication(nvroot, B_FALSE)) == NULL)
 851                         return (0);
 852         }
 853         /*
 854          * for spares there may be no children, and therefore no
 855          * replication level to check
 856          */
 857         if ((nvlist_lookup_nvlist_array(newroot, ZPOOL_CONFIG_CHILDREN,
 858             &child, &children) != 0) || (children == 0)) {
 859                 free(current);
 860                 return (0);
 861         }
 862
 863         /*
 864          * If all we have is logs then there's no replication level to check.
 865          */
 866         if (num_logs(newroot) == children) {
 867                 free(current);
 868                 return (0);
 869         }
 870
 871         /*
 872          * Get the replication level of the new vdev spec, reporting any
 873          * inconsistencies found.
 874          */
 875         if ((new = get_replication(newroot, B_TRUE)) == NULL) {
 876                 free(current);
 877                 return (-1);
 878         }
 879
 880         /*
 881          * Check to see if the new vdev spec matches the replication level of
 882          * the current pool.
 883          */
 884         ret = 0;
 885         if (current != NULL) {
 886                 if (strcmp(current->zprl_type, new->zprl_type) != 0) {
 887                         vdev_error(gettext(
 888                             "mismatched replication level: pool uses %s "
 889                             "and new vdev is %s\n"),
 890                             current->zprl_type, new->zprl_type);
 891                         ret = -1;
 892                 } else if (current->zprl_parity != new->zprl_parity) {
 893                         vdev_error(gettext(
 894                             "mismatched replication level: pool uses %llu "
 895                             "device parity and new vdev uses %llu\n"),
 896                             current->zprl_parity, new->zprl_parity);
 897                         ret = -1;
 898                 } else if (current->zprl_children != new->zprl_children) {
 899                         vdev_error(gettext(
 900                             "mismatched replication level: pool uses %llu-way "
 901                             "%s and new vdev uses %llu-way %s\n"),
 902                             current->zprl_children, current->zprl_type,
 903                             new->zprl_children, new->zprl_type);
 904                         ret = -1;
 905                 }
 906         }
 907
 908         free(new);
 909         if (current != NULL)
 910                 free(current);
 911
 912         return (ret);
 913 }
 914
 915 #ifdef illumos
 916 /*
 917  * Go through and find any whole disks in the vdev specification, labelling them
 918  * as appropriate.  When constructing the vdev spec, we were unable to open this
 919  * device in order to provide a devid.  Now that we have labelled the disk and
 920  * know that slice 0 is valid, we can construct the devid now.
 921  *
 922  * If the disk was already labeled with an EFI label, we will have gotten the
 923  * devid already (because we were able to open the whole disk).  Otherwise, we
 924  * need to get the devid after we label the disk.
 925  */
 926 static int
 927 make_disks(zpool_handle_t *zhp, nvlist_t *nv)
 928 {
 929         nvlist_t **child;
 930         uint_t c, children;
 931         char *type, *path, *diskname;
 932         char buf[MAXPATHLEN];
 933         uint64_t wholedisk;
 934         int fd;
 935         int ret;
 936         ddi_devid_t devid;
 937         char *minor = NULL, *devid_str = NULL;
 938
 939         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
 940
 941         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 942             &child, &children) != 0) {
 943
 944                 if (strcmp(type, VDEV_TYPE_DISK) != 0)
 945                         return (0);
 946
 947                 /*
 948                  * We have a disk device.  Get the path to the device
 949                  * and see if it's a whole disk by appending the backup
 950                  * slice and stat()ing the device.
 951                  */
 952                 verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path) == 0);
 953                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
 954                     &wholedisk) != 0 || !wholedisk)
 955                         return (0);
 956
 957                 diskname = strrchr(path, '/');
 958                 assert(diskname != NULL);
 959                 diskname++;
 960                 if (zpool_label_disk(g_zfs, zhp, diskname) == -1)
 961                         return (-1);
 962
 963                 /*
 964                  * Fill in the devid, now that we've labeled the disk.
 965                  */
 966                 (void) snprintf(buf, sizeof (buf), "%ss0", path);
 967                 if ((fd = open(buf, O_RDONLY)) < 0) {
 968                         (void) fprintf(stderr,
 969                             gettext("cannot open '%s': %s\n"),
 970                             buf, strerror(errno));
 971                         return (-1);
 972                 }
 973
 974                 if (devid_get(fd, &devid) == 0) {
 975                         if (devid_get_minor_name(fd, &minor) == 0 &&
 976                             (devid_str = devid_str_encode(devid, minor)) !=
 977                             NULL) {
 978                                 verify(nvlist_add_string(nv,
 979                                     ZPOOL_CONFIG_DEVID, devid_str) == 0);
 980                         }
 981                         if (devid_str != NULL)
 982                                 devid_str_free(devid_str);
 983                         if (minor != NULL)
 984                                 devid_str_free(minor);
 985                         devid_free(devid);
 986                 }
 987
 988                 /*
 989                  * Update the path to refer to the 's0' slice.  The presence of
 990                  * the 'whole_disk' field indicates to the CLI that we should
 991                  * chop off the slice number when displaying the device in
 992                  * future output.
 993                  */
 994                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_PATH, buf) == 0);
 995
 996                 (void) close(fd);
 997
 998                 return (0);
 999         }
1000
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_SPARES,
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         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1012             &child, &children) == 0)
1013                 for (c = 0; c < children; c++)
1014                         if ((ret = make_disks(zhp, child[c])) != 0)
1015                                 return (ret);
1016
1017         return (0);
1018 }
1019 #endif  /* illumos */
1020
1021 /*
1022  * Determine if the given path is a hot spare within the given configuration.
1023  */
1024 static boolean_t
1025 is_spare(nvlist_t *config, const char *path)
1026 {
1027         int fd;
1028         pool_state_t state;
1029         char *name = NULL;
1030         nvlist_t *label;
1031         uint64_t guid, spareguid;
1032         nvlist_t *nvroot;
1033         nvlist_t **spares;
1034         uint_t i, nspares;
1035         boolean_t inuse;
1036
1037         if ((fd = open(path, O_RDONLY)) < 0)
1038                 return (B_FALSE);
1039
1040         if (zpool_in_use(g_zfs, fd, &state, &name, &inuse) != 0 ||
1041             !inuse ||
1042             state != POOL_STATE_SPARE ||
1043             zpool_read_label(fd, &label) != 0) {
1044                 free(name);
1045                 (void) close(fd);
1046                 return (B_FALSE);
1047         }
1048         free(name);
1049         (void) close(fd);
1050
1051         verify(nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) == 0);
1052         nvlist_free(label);
1053
1054         verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
1055             &nvroot) == 0);
1056         if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1057             &spares, &nspares) == 0) {
1058                 for (i = 0; i < nspares; i++) {
1059                         verify(nvlist_lookup_uint64(spares[i],
1060                             ZPOOL_CONFIG_GUID, &spareguid) == 0);
1061                         if (spareguid == guid)
1062                                 return (B_TRUE);
1063                 }
1064         }
1065
1066         return (B_FALSE);
1067 }
1068
1069 /*
1070  * Go through and find any devices that are in use.  We rely on libdiskmgt for
1071  * the majority of this task.
1072  */
1073 static boolean_t
1074 is_device_in_use(nvlist_t *config, nvlist_t *nv, boolean_t force,
1075     boolean_t replacing, boolean_t isspare)
1076 {
1077         nvlist_t **child;
1078         uint_t c, children;
1079         char *type, *path;
1080         int ret = 0;
1081         char buf[MAXPATHLEN];
1082         uint64_t wholedisk;
1083         boolean_t anyinuse = B_FALSE;
1084
1085         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
1086
1087         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
1088             &child, &children) != 0) {
1089
1090                 verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path) == 0);
1091
1092                 /*
1093                  * As a generic check, we look to see if this is a replace of a
1094                  * hot spare within the same pool.  If so, we allow it
1095                  * regardless of what libdiskmgt or zpool_in_use() says.
1096                  */
1097                 if (replacing) {
1098 #ifdef illumos
1099                         if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
1100                             &wholedisk) == 0 && wholedisk)
1101                                 (void) snprintf(buf, sizeof (buf), "%ss0",
1102                                     path);
1103                         else
1104 #endif
1105                                 (void) strlcpy(buf, path, sizeof (buf));
1106
1107                         if (is_spare(config, buf))
1108                                 return (B_FALSE);
1109                 }
1110
1111                 if (strcmp(type, VDEV_TYPE_DISK) == 0)
1112                         ret = check_device(path, force, isspare);
1113                 else if (strcmp(type, VDEV_TYPE_FILE) == 0)
1114                         ret = check_file(path, force, isspare);
1115
1116                 return (ret != 0);
1117         }
1118
1119         for (c = 0; c < children; c++)
1120                 if (is_device_in_use(config, child[c], force, replacing,
1121                     B_FALSE))
1122                         anyinuse = B_TRUE;
1123
1124         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_SPARES,
1125             &child, &children) == 0)
1126                 for (c = 0; c < children; c++)
1127                         if (is_device_in_use(config, child[c], force, replacing,
1128                             B_TRUE))
1129                                 anyinuse = B_TRUE;
1130
1131         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1132             &child, &children) == 0)
1133                 for (c = 0; c < children; c++)
1134                         if (is_device_in_use(config, child[c], force, replacing,
1135                             B_FALSE))
1136                                 anyinuse = B_TRUE;
1137
1138         return (anyinuse);
1139 }
1140
1141 static const char *
1142 is_grouping(const char *type, int *mindev, int *maxdev)
1143 {
1144         if (strncmp(type, "raidz", 5) == 0) {
1145                 const char *p = type + 5;
1146                 char *end;
1147                 long nparity;
1148
1149                 if (*p == '\0') {
1150                         nparity = 1;
1151                 } else if (*p == '0') {
1152                         return (NULL); /* no zero prefixes allowed */
1153                 } else {
1154                         errno = 0;
1155                         nparity = strtol(p, &end, 10);
1156                         if (errno != 0 || nparity < 1 || nparity >= 255 ||
1157                             *end != '\0')
1158                                 return (NULL);
1159                 }
1160
1161                 if (mindev != NULL)
1162                         *mindev = nparity + 1;
1163                 if (maxdev != NULL)
1164                         *maxdev = 255;
1165                 return (VDEV_TYPE_RAIDZ);
1166         }
1167
1168         if (maxdev != NULL)
1169                 *maxdev = INT_MAX;
1170
1171         if (strcmp(type, "mirror") == 0) {
1172                 if (mindev != NULL)
1173                         *mindev = 2;
1174                 return (VDEV_TYPE_MIRROR);
1175         }
1176
1177         if (strcmp(type, "spare") == 0) {
1178                 if (mindev != NULL)
1179                         *mindev = 1;
1180                 return (VDEV_TYPE_SPARE);
1181         }
1182
1183         if (strcmp(type, "log") == 0) {
1184                 if (mindev != NULL)
1185                         *mindev = 1;
1186                 return (VDEV_TYPE_LOG);
1187         }
1188
1189         if (strcmp(type, "cache") == 0) {
1190                 if (mindev != NULL)
1191                         *mindev = 1;
1192                 return (VDEV_TYPE_L2CACHE);
1193         }
1194
1195         return (NULL);
1196 }
1197
1198 /*
1199  * Construct a syntactically valid vdev specification,
1200  * and ensure that all devices and files exist and can be opened.
1201  * Note: we don't bother freeing anything in the error paths
1202  * because the program is just going to exit anyway.
1203  */
1204 nvlist_t *
1205 construct_spec(int argc, char **argv)
1206 {
1207         nvlist_t *nvroot, *nv, **top, **spares, **l2cache;
1208         int t, toplevels, mindev, maxdev, nspares, nlogs, nl2cache;
1209         const char *type;
1210         uint64_t is_log;
1211         boolean_t seen_logs;
1212
1213         top = NULL;
1214         toplevels = 0;
1215         spares = NULL;
1216         l2cache = NULL;
1217         nspares = 0;
1218         nlogs = 0;
1219         nl2cache = 0;
1220         is_log = B_FALSE;
1221         seen_logs = B_FALSE;
1222
1223         while (argc > 0) {
1224                 nv = NULL;
1225
1226                 /*
1227                  * If it's a mirror or raidz, the subsequent arguments are
1228                  * its leaves -- until we encounter the next mirror or raidz.
1229                  */
1230                 if ((type = is_grouping(argv[0], &mindev, &maxdev)) != NULL) {
1231                         nvlist_t **child = NULL;
1232                         int c, children = 0;
1233
1234                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1235                                 if (spares != NULL) {
1236                                         (void) fprintf(stderr,
1237                                             gettext("invalid vdev "
1238                                             "specification: 'spare' can be "
1239                                             "specified only once\n"));
1240                                         return (NULL);
1241                                 }
1242                                 is_log = B_FALSE;
1243                         }
1244
1245                         if (strcmp(type, VDEV_TYPE_LOG) == 0) {
1246                                 if (seen_logs) {
1247                                         (void) fprintf(stderr,
1248                                             gettext("invalid vdev "
1249                                             "specification: 'log' can be "
1250                                             "specified only once\n"));
1251                                         return (NULL);
1252                                 }
1253                                 seen_logs = B_TRUE;
1254                                 is_log = B_TRUE;
1255                                 argc--;
1256                                 argv++;
1257                                 /*
1258                                  * A log is not a real grouping device.
1259                                  * We just set is_log and continue.
1260                                  */
1261                                 continue;
1262                         }
1263
1264                         if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1265                                 if (l2cache != NULL) {
1266                                         (void) fprintf(stderr,
1267                                             gettext("invalid vdev "
1268                                             "specification: 'cache' can be "
1269                                             "specified only once\n"));
1270                                         return (NULL);
1271                                 }
1272                                 is_log = B_FALSE;
1273                         }
1274
1275                         if (is_log) {
1276                                 if (strcmp(type, VDEV_TYPE_MIRROR) != 0) {
1277                                         (void) fprintf(stderr,
1278                                             gettext("invalid vdev "
1279                                             "specification: unsupported 'log' "
1280                                             "device: %s\n"), type);
1281                                         return (NULL);
1282                                 }
1283                                 nlogs++;
1284                         }
1285
1286                         for (c = 1; c < argc; c++) {
1287                                 if (is_grouping(argv[c], NULL, NULL) != NULL)
1288                                         break;
1289                                 children++;
1290                                 child = realloc(child,
1291                                     children * sizeof (nvlist_t *));
1292                                 if (child == NULL)
1293                                         zpool_no_memory();
1294                                 if ((nv = make_leaf_vdev(argv[c], B_FALSE))
1295                                     == NULL)
1296                                         return (NULL);
1297                                 child[children - 1] = nv;
1298                         }
1299
1300                         if (children < mindev) {
1301                                 (void) fprintf(stderr, gettext("invalid vdev "
1302                                     "specification: %s requires at least %d "
1303                                     "devices\n"), argv[0], mindev);
1304                                 return (NULL);
1305                         }
1306
1307                         if (children > maxdev) {
1308                                 (void) fprintf(stderr, gettext("invalid vdev "
1309                                     "specification: %s supports no more than "
1310                                     "%d devices\n"), argv[0], maxdev);
1311                                 return (NULL);
1312                         }
1313
1314                         argc -= c;
1315                         argv += c;
1316
1317                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1318                                 spares = child;
1319                                 nspares = children;
1320                                 continue;
1321                         } else if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1322                                 l2cache = child;
1323                                 nl2cache = children;
1324                                 continue;
1325                         } else {
1326                                 verify(nvlist_alloc(&nv, NV_UNIQUE_NAME,
1327                                     0) == 0);
1328                                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_TYPE,
1329                                     type) == 0);
1330                                 verify(nvlist_add_uint64(nv,
1331                                     ZPOOL_CONFIG_IS_LOG, is_log) == 0);
1332                                 if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
1333                                         verify(nvlist_add_uint64(nv,
1334                                             ZPOOL_CONFIG_NPARITY,
1335                                             mindev - 1) == 0);
1336                                 }
1337                                 verify(nvlist_add_nvlist_array(nv,
1338                                     ZPOOL_CONFIG_CHILDREN, child,
1339                                     children) == 0);
1340
1341                                 for (c = 0; c < children; c++)
1342                                         nvlist_free(child[c]);
1343                                 free(child);
1344                         }
1345                 } else {
1346                         /*
1347                          * We have a device.  Pass off to make_leaf_vdev() to
1348                          * construct the appropriate nvlist describing the vdev.
1349                          */
1350                         if ((nv = make_leaf_vdev(argv[0], is_log)) == NULL)
1351                                 return (NULL);
1352                         if (is_log)
1353                                 nlogs++;
1354                         argc--;
1355                         argv++;
1356                 }
1357
1358                 toplevels++;
1359                 top = realloc(top, toplevels * sizeof (nvlist_t *));
1360                 if (top == NULL)
1361                         zpool_no_memory();
1362                 top[toplevels - 1] = nv;
1363         }
1364
1365         if (toplevels == 0 && nspares == 0 && nl2cache == 0) {
1366                 (void) fprintf(stderr, gettext("invalid vdev "
1367                     "specification: at least one toplevel vdev must be "
1368                     "specified\n"));
1369                 return (NULL);
1370         }
1371
1372         if (seen_logs && nlogs == 0) {
1373                 (void) fprintf(stderr, gettext("invalid vdev specification: "
1374                     "log requires at least 1 device\n"));
1375                 return (NULL);
1376         }
1377
1378         /*
1379          * Finally, create nvroot and add all top-level vdevs to it.
1380          */
1381         verify(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, 0) == 0);
1382         verify(nvlist_add_string(nvroot, ZPOOL_CONFIG_TYPE,
1383             VDEV_TYPE_ROOT) == 0);
1384         verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
1385             top, toplevels) == 0);
1386         if (nspares != 0)
1387                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1388                     spares, nspares) == 0);
1389         if (nl2cache != 0)
1390                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
1391                     l2cache, nl2cache) == 0);
1392
1393         for (t = 0; t < toplevels; t++)
1394                 nvlist_free(top[t]);
1395         for (t = 0; t < nspares; t++)
1396                 nvlist_free(spares[t]);
1397         for (t = 0; t < nl2cache; t++)
1398                 nvlist_free(l2cache[t]);
1399         if (spares)
1400                 free(spares);
1401         if (l2cache)
1402                 free(l2cache);
1403         free(top);
1404
1405         return (nvroot);
1406 }
1407
1408 nvlist_t *
1409 split_mirror_vdev(zpool_handle_t *zhp, char *newname, nvlist_t *props,
1410     splitflags_t flags, int argc, char **argv)
1411 {
1412         nvlist_t *newroot = NULL, **child;
1413         uint_t c, children;
1414
1415         if (argc > 0) {
1416                 if ((newroot = construct_spec(argc, argv)) == NULL) {
1417                         (void) fprintf(stderr, gettext("Unable to build a "
1418                             "pool from the specified devices\n"));
1419                         return (NULL);
1420                 }
1421
1422 #ifdef illumos
1423                 if (!flags.dryrun && make_disks(zhp, newroot) != 0) {
1424                         nvlist_free(newroot);
1425                         return (NULL);
1426                 }
1427 #endif
1428
1429                 /* avoid any tricks in the spec */
1430                 verify(nvlist_lookup_nvlist_array(newroot,
1431                     ZPOOL_CONFIG_CHILDREN, &child, &children) == 0);
1432                 for (c = 0; c < children; c++) {
1433                         char *path;
1434                         const char *type;
1435                         int min, max;
1436
1437                         verify(nvlist_lookup_string(child[c],
1438                             ZPOOL_CONFIG_PATH, &path) == 0);
1439                         if ((type = is_grouping(path, &min, &max)) != NULL) {
1440                                 (void) fprintf(stderr, gettext("Cannot use "
1441                                     "'%s' as a device for splitting\n"), type);
1442                                 nvlist_free(newroot);
1443                                 return (NULL);
1444                         }
1445                 }
1446         }
1447
1448         if (zpool_vdev_split(zhp, newname, &newroot, props, flags) != 0) {
1449                 nvlist_free(newroot);
1450                 return (NULL);
1451         }
1452
1453         return (newroot);
1454 }
1455
1456 /*
1457  * Get and validate the contents of the given vdev specification.  This ensures
1458  * that the nvlist returned is well-formed, that all the devices exist, and that
1459  * they are not currently in use by any other known consumer.  The 'poolconfig'
1460  * parameter is the current configuration of the pool when adding devices
1461  * existing pool, and is used to perform additional checks, such as changing the
1462  * replication level of the pool.  It can be 'NULL' to indicate that this is a
1463  * new pool.  The 'force' flag controls whether devices should be forcefully
1464  * added, even if they appear in use.
1465  */
1466 nvlist_t *
1467 make_root_vdev(zpool_handle_t *zhp, int force, int check_rep,
1468     boolean_t replacing, boolean_t dryrun, int argc, char **argv)
1469 {
1470         nvlist_t *newroot;
1471         nvlist_t *poolconfig = NULL;
1472         is_force = force;
1473
1474         /*
1475          * Construct the vdev specification.  If this is successful, we know
1476          * that we have a valid specification, and that all devices can be
1477          * opened.
1478          */
1479         if ((newroot = construct_spec(argc, argv)) == NULL)
1480                 return (NULL);
1481
1482         if (zhp && ((poolconfig = zpool_get_config(zhp, NULL)) == NULL))
1483                 return (NULL);
1484
1485         /*
1486          * Validate each device to make sure that its not shared with another
1487          * subsystem.  We do this even if 'force' is set, because there are some
1488          * uses (such as a dedicated dump device) that even '-f' cannot
1489          * override.
1490          */
1491         if (is_device_in_use(poolconfig, newroot, force, replacing, B_FALSE)) {
1492                 nvlist_free(newroot);
1493                 return (NULL);
1494         }
1495
1496         /*
1497          * Check the replication level of the given vdevs and report any errors
1498          * found.  We include the existing pool spec, if any, as we need to
1499          * catch changes against the existing replication level.
1500          */
1501         if (check_rep && check_replication(poolconfig, newroot) != 0) {
1502                 nvlist_free(newroot);
1503                 return (NULL);
1504         }
1505
1506 #ifdef illumos
1507         /*
1508          * Run through the vdev specification and label any whole disks found.
1509          */
1510         if (!dryrun && make_disks(zhp, newroot) != 0) {
1511                 nvlist_free(newroot);
1512                 return (NULL);
1513         }
1514 #endif
1515
1516         return (newroot);
1517 }