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(nvlist_t *props, 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         if (props != NULL) {
 503                 uint64_t ashift = 0;
 504                 char *value = NULL;
 505
 506                 if (nvlist_lookup_string(props,
 507                     zpool_prop_to_name(ZPOOL_PROP_ASHIFT), &value) == 0)
 508                         zfs_nicestrtonum(NULL, value, &ashift);
 509
 510                 if (ashift > 0)
 511                         verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_ASHIFT,
 512                             ashift) == 0);
 513         }
 514
 515         return (vdev);
 516 }
 517
 518 /*
 519  * Go through and verify the replication level of the pool is consistent.
 520  * Performs the following checks:
 521  *
 522  *      For the new spec, verifies that devices in mirrors and raidz are the
 523  *      same size.
 524  *
 525  *      If the current configuration already has inconsistent replication
 526  *      levels, ignore any other potential problems in the new spec.
 527  *
 528  *      Otherwise, make sure that the current spec (if there is one) and the new
 529  *      spec have consistent replication levels.
 530  */
 531 typedef struct replication_level {
 532         char *zprl_type;
 533         uint64_t zprl_children;
 534         uint64_t zprl_parity;
 535 } replication_level_t;
 536
 537 #define ZPOOL_FUZZ      (16 * 1024 * 1024)
 538
 539 /*
 540  * Given a list of toplevel vdevs, return the current replication level.  If
 541  * the config is inconsistent, then NULL is returned.  If 'fatal' is set, then
 542  * an error message will be displayed for each self-inconsistent vdev.
 543  */
 544 static replication_level_t *
 545 get_replication(nvlist_t *nvroot, boolean_t fatal)
 546 {
 547         nvlist_t **top;
 548         uint_t t, toplevels;
 549         nvlist_t **child;
 550         uint_t c, children;
 551         nvlist_t *nv;
 552         char *type;
 553         replication_level_t lastrep = { 0 }, rep, *ret;
 554         boolean_t dontreport;
 555
 556         ret = safe_malloc(sizeof (replication_level_t));
 557
 558         verify(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
 559             &top, &toplevels) == 0);
 560
 561         lastrep.zprl_type = NULL;
 562         for (t = 0; t < toplevels; t++) {
 563                 uint64_t is_log = B_FALSE;
 564
 565                 nv = top[t];
 566
 567                 /*
 568                  * For separate logs we ignore the top level vdev replication
 569                  * constraints.
 570                  */
 571                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &is_log);
 572                 if (is_log)
 573                         continue;
 574
 575                 verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE,
 576                     &type) == 0);
 577                 if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 578                     &child, &children) != 0) {
 579                         /*
 580                          * This is a 'file' or 'disk' vdev.
 581                          */
 582                         rep.zprl_type = type;
 583                         rep.zprl_children = 1;
 584                         rep.zprl_parity = 0;
 585                 } else {
 586                         uint64_t vdev_size;
 587
 588                         /*
 589                          * This is a mirror or RAID-Z vdev.  Go through and make
 590                          * sure the contents are all the same (files vs. disks),
 591                          * keeping track of the number of elements in the
 592                          * process.
 593                          *
 594                          * We also check that the size of each vdev (if it can
 595                          * be determined) is the same.
 596                          */
 597                         rep.zprl_type = type;
 598                         rep.zprl_children = 0;
 599
 600                         if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
 601                                 verify(nvlist_lookup_uint64(nv,
 602                                     ZPOOL_CONFIG_NPARITY,
 603                                     &rep.zprl_parity) == 0);
 604                                 assert(rep.zprl_parity != 0);
 605                         } else {
 606                                 rep.zprl_parity = 0;
 607                         }
 608
 609                         /*
 610                          * The 'dontreport' variable indicates that we've
 611                          * already reported an error for this spec, so don't
 612                          * bother doing it again.
 613                          */
 614                         type = NULL;
 615                         dontreport = 0;
 616                         vdev_size = -1ULL;
 617                         for (c = 0; c < children; c++) {
 618                                 nvlist_t *cnv = child[c];
 619                                 char *path;
 620                                 struct stat64 statbuf;
 621                                 uint64_t size = -1ULL;
 622                                 char *childtype;
 623                                 int fd, err;
 624
 625                                 rep.zprl_children++;
 626
 627                                 verify(nvlist_lookup_string(cnv,
 628                                     ZPOOL_CONFIG_TYPE, &childtype) == 0);
 629
 630                                 /*
 631                                  * If this is a replacing or spare vdev, then
 632                                  * get the real first child of the vdev.
 633                                  */
 634                                 if (strcmp(childtype,
 635                                     VDEV_TYPE_REPLACING) == 0 ||
 636                                     strcmp(childtype, VDEV_TYPE_SPARE) == 0) {
 637                                         nvlist_t **rchild;
 638                                         uint_t rchildren;
 639
 640                                         verify(nvlist_lookup_nvlist_array(cnv,
 641                                             ZPOOL_CONFIG_CHILDREN, &rchild,
 642                                             &rchildren) == 0);
 643                                         assert(rchildren == 2);
 644                                         cnv = rchild[0];
 645
 646                                         verify(nvlist_lookup_string(cnv,
 647                                             ZPOOL_CONFIG_TYPE,
 648                                             &childtype) == 0);
 649                                 }
 650
 651                                 verify(nvlist_lookup_string(cnv,
 652                                     ZPOOL_CONFIG_PATH, &path) == 0);
 653
 654                                 /*
 655                                  * If we have a raidz/mirror that combines disks
 656                                  * with files, report it as an error.
 657                                  */
 658                                 if (!dontreport && type != NULL &&
 659                                     strcmp(type, childtype) != 0) {
 660                                         if (ret != NULL)
 661                                                 free(ret);
 662                                         ret = NULL;
 663                                         if (fatal)
 664                                                 vdev_error(gettext(
 665                                                     "mismatched replication "
 666                                                     "level: %s contains both "
 667                                                     "files and devices\n"),
 668                                                     rep.zprl_type);
 669                                         else
 670                                                 return (NULL);
 671                                         dontreport = B_TRUE;
 672                                 }
 673
 674                                 /*
 675                                  * According to stat(2), the value of 'st_size'
 676                                  * is undefined for block devices and character
 677                                  * devices.  But there is no effective way to
 678                                  * determine the real size in userland.
 679                                  *
 680                                  * Instead, we'll take advantage of an
 681                                  * implementation detail of spec_size().  If the
 682                                  * device is currently open, then we (should)
 683                                  * return a valid size.
 684                                  *
 685                                  * If we still don't get a valid size (indicated
 686                                  * by a size of 0 or MAXOFFSET_T), then ignore
 687                                  * this device altogether.
 688                                  */
 689                                 if ((fd = open(path, O_RDONLY)) >= 0) {
 690                                         err = fstat64(fd, &statbuf);
 691                                         (void) close(fd);
 692                                 } else {
 693                                         err = stat64(path, &statbuf);
 694                                 }
 695
 696                                 if (err != 0 ||
 697                                     statbuf.st_size == 0 ||
 698                                     statbuf.st_size == MAXOFFSET_T)
 699                                         continue;
 700
 701                                 size = statbuf.st_size;
 702
 703                                 /*
 704                                  * Also make sure that devices and
 705                                  * slices have a consistent size.  If
 706                                  * they differ by a significant amount
 707                                  * (~16MB) then report an error.
 708                                  */
 709                                 if (!dontreport &&
 710                                     (vdev_size != -1ULL &&
 711                                     (labs(size - vdev_size) >
 712                                     ZPOOL_FUZZ))) {
 713                                         if (ret != NULL)
 714                                                 free(ret);
 715                                         ret = NULL;
 716                                         if (fatal)
 717                                                 vdev_error(gettext(
 718                                                     "%s contains devices of "
 719                                                     "different sizes\n"),
 720                                                     rep.zprl_type);
 721                                         else
 722                                                 return (NULL);
 723                                         dontreport = B_TRUE;
 724                                 }
 725
 726                                 type = childtype;
 727                                 vdev_size = size;
 728                         }
 729                 }
 730
 731                 /*
 732                  * At this point, we have the replication of the last toplevel
 733                  * vdev in 'rep'.  Compare it to 'lastrep' to see if its
 734                  * different.
 735                  */
 736                 if (lastrep.zprl_type != NULL) {
 737                         if (strcmp(lastrep.zprl_type, rep.zprl_type) != 0) {
 738                                 if (ret != NULL)
 739                                         free(ret);
 740                                 ret = NULL;
 741                                 if (fatal)
 742                                         vdev_error(gettext(
 743                                             "mismatched replication level: "
 744                                             "both %s and %s vdevs are "
 745                                             "present\n"),
 746                                             lastrep.zprl_type, rep.zprl_type);
 747                                 else
 748                                         return (NULL);
 749                         } else if (lastrep.zprl_parity != rep.zprl_parity) {
 750                                 if (ret)
 751                                         free(ret);
 752                                 ret = NULL;
 753                                 if (fatal)
 754                                         vdev_error(gettext(
 755                                             "mismatched replication level: "
 756                                             "both %llu and %llu device parity "
 757                                             "%s vdevs are present\n"),
 758                                             lastrep.zprl_parity,
 759                                             rep.zprl_parity,
 760                                             rep.zprl_type);
 761                                 else
 762                                         return (NULL);
 763                         } else if (lastrep.zprl_children != rep.zprl_children) {
 764                                 if (ret)
 765                                         free(ret);
 766                                 ret = NULL;
 767                                 if (fatal)
 768                                         vdev_error(gettext(
 769                                             "mismatched replication level: "
 770                                             "both %llu-way and %llu-way %s "
 771                                             "vdevs are present\n"),
 772                                             lastrep.zprl_children,
 773                                             rep.zprl_children,
 774                                             rep.zprl_type);
 775                                 else
 776                                         return (NULL);
 777                         }
 778                 }
 779                 lastrep = rep;
 780         }
 781
 782         if (ret != NULL)
 783                 *ret = rep;
 784
 785         return (ret);
 786 }
 787
 788 /*
 789  * Check the replication level of the vdev spec against the current pool.  Calls
 790  * get_replication() to make sure the new spec is self-consistent.  If the pool
 791  * has a consistent replication level, then we ignore any errors.  Otherwise,
 792  * report any difference between the two.
 793  */
 794 static int
 795 check_replication(nvlist_t *config, nvlist_t *newroot)
 796 {
 797         nvlist_t **child;
 798         uint_t  children;
 799         replication_level_t *current = NULL, *new;
 800         int ret;
 801
 802         /*
 803          * If we have a current pool configuration, check to see if it's
 804          * self-consistent.  If not, simply return success.
 805          */
 806         if (config != NULL) {
 807                 nvlist_t *nvroot;
 808
 809                 verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
 810                     &nvroot) == 0);
 811                 if ((current = get_replication(nvroot, B_FALSE)) == NULL)
 812                         return (0);
 813         }
 814         /*
 815          * for spares there may be no children, and therefore no
 816          * replication level to check
 817          */
 818         if ((nvlist_lookup_nvlist_array(newroot, ZPOOL_CONFIG_CHILDREN,
 819             &child, &children) != 0) || (children == 0)) {
 820                 free(current);
 821                 return (0);
 822         }
 823
 824         /*
 825          * If all we have is logs then there's no replication level to check.
 826          */
 827         if (num_logs(newroot) == children) {
 828                 free(current);
 829                 return (0);
 830         }
 831
 832         /*
 833          * Get the replication level of the new vdev spec, reporting any
 834          * inconsistencies found.
 835          */
 836         if ((new = get_replication(newroot, B_TRUE)) == NULL) {
 837                 free(current);
 838                 return (-1);
 839         }
 840
 841         /*
 842          * Check to see if the new vdev spec matches the replication level of
 843          * the current pool.
 844          */
 845         ret = 0;
 846         if (current != NULL) {
 847                 if (strcmp(current->zprl_type, new->zprl_type) != 0) {
 848                         vdev_error(gettext(
 849                             "mismatched replication level: pool uses %s "
 850                             "and new vdev is %s\n"),
 851                             current->zprl_type, new->zprl_type);
 852                         ret = -1;
 853                 } else if (current->zprl_parity != new->zprl_parity) {
 854                         vdev_error(gettext(
 855                             "mismatched replication level: pool uses %llu "
 856                             "device parity and new vdev uses %llu\n"),
 857                             current->zprl_parity, new->zprl_parity);
 858                         ret = -1;
 859                 } else if (current->zprl_children != new->zprl_children) {
 860                         vdev_error(gettext(
 861                             "mismatched replication level: pool uses %llu-way "
 862                             "%s and new vdev uses %llu-way %s\n"),
 863                             current->zprl_children, current->zprl_type,
 864                             new->zprl_children, new->zprl_type);
 865                         ret = -1;
 866                 }
 867         }
 868
 869         free(new);
 870         if (current != NULL)
 871                 free(current);
 872
 873         return (ret);
 874 }
 875
 876 static int
 877 zero_label(char *path)
 878 {
 879         const int size = 4096;
 880         char buf[size];
 881         int err, fd;
 882
 883         if ((fd = open(path, O_WRONLY|O_EXCL)) < 0) {
 884                 (void) fprintf(stderr, gettext("cannot open '%s': %s\n"),
 885                     path, strerror(errno));
 886                 return (-1);
 887         }
 888
 889         memset(buf, 0, size);
 890         err = write(fd, buf, size);
 891         (void) fdatasync(fd);
 892         (void) close(fd);
 893
 894         if (err == -1) {
 895                 (void) fprintf(stderr, gettext("cannot zero first %d bytes "
 896                     "of '%s': %s\n"), size, path, strerror(errno));
 897                 return (-1);
 898         }
 899
 900         if (err != size) {
 901                 (void) fprintf(stderr, gettext("could only zero %d/%d bytes "
 902                     "of '%s'\n"), err, size, path);
 903                 return (-1);
 904         }
 905
 906         return 0;
 907 }
 908
 909 /*
 910  * Go through and find any whole disks in the vdev specification, labelling them
 911  * as appropriate.  When constructing the vdev spec, we were unable to open this
 912  * device in order to provide a devid.  Now that we have labelled the disk and
 913  * know that slice 0 is valid, we can construct the devid now.
 914  *
 915  * If the disk was already labeled with an EFI label, we will have gotten the
 916  * devid already (because we were able to open the whole disk).  Otherwise, we
 917  * need to get the devid after we label the disk.
 918  */
 919 static int
 920 make_disks(zpool_handle_t *zhp, nvlist_t *nv)
 921 {
 922         nvlist_t **child;
 923         uint_t c, children;
 924         char *type, *path, *diskname;
 925         char devpath[MAXPATHLEN];
 926         char udevpath[MAXPATHLEN];
 927         uint64_t wholedisk;
 928         struct stat64 statbuf;
 929         int ret;
 930
 931         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
 932
 933         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 934             &child, &children) != 0) {
 935
 936                 if (strcmp(type, VDEV_TYPE_DISK) != 0)
 937                         return (0);
 938
 939                 /*
 940                  * We have a disk device.  If this is a whole disk write
 941                  * out the efi partition table, otherwise write zero's to
 942                  * the first 4k of the partition.  This is to ensure that
 943                  * libblkid will not misidentify the partition due to a
 944                  * magic value left by the previous filesystem.
 945                  */
 946                 verify(!nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path));
 947                 verify(!nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
 948                     &wholedisk));
 949
 950                 if (!wholedisk) {
 951                         ret = zero_label(path);
 952                         return (ret);
 953                 }
 954
 955                 if (realpath(path, devpath) == NULL) {
 956                         ret = errno;
 957                         (void) fprintf(stderr,
 958                             gettext("cannot resolve path '%s'\n"), path);
 959                         return (ret);
 960                 }
 961
 962                 /*
 963                  * Remove any previously existing symlink from a udev path to
 964                  * the device before labeling the disk.  This makes
 965                  * zpool_label_disk_wait() truly wait for the new link to show
 966                  * up instead of returning if it finds an old link still in
 967                  * place.  Otherwise there is a window between when udev
 968                  * deletes and recreates the link during which access attempts
 969                  * will fail with ENOENT.
 970                  */
 971                 zfs_append_partition(path, udevpath, sizeof (udevpath));
 972                 if ((strncmp(udevpath, UDISK_ROOT, strlen(UDISK_ROOT)) == 0) &&
 973                     (lstat64(udevpath, &statbuf) == 0) &&
 974                     S_ISLNK(statbuf.st_mode))
 975                         (void) unlink(udevpath);
 976
 977                 diskname = strrchr(devpath, '/');
 978                 assert(diskname != NULL);
 979                 diskname++;
 980                 if (zpool_label_disk(g_zfs, zhp, diskname) == -1)
 981                         return (-1);
 982
 983                 /*
 984                  * Now we've labeled the disk and the partitions have been
 985                  * created.  We still need to wait for udev to create the
 986                  * symlinks to those partitions.
 987                  */
 988                 if ((ret = zpool_label_disk_wait(udevpath, 1000)) != 0) {
 989                         (void) fprintf(stderr,
 990                             gettext( "cannot resolve path '%s'\n"), udevpath);
 991                         return (-1);
 992                 }
 993
 994                 /*
 995                  * Update the path to refer to FIRST_SLICE.  The presence of
 996                  * the 'whole_disk' field indicates to the CLI that we should
 997                  * chop off the slice number when displaying the device in
 998                  * future output.
 999                  */
1000                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_PATH, udevpath) == 0);
1001
1002                 /* Just in case this partition already existed. */
1003                 (void) zero_label(udevpath);
1004
1005                 return (0);
1006         }
1007
1008         for (c = 0; c < children; c++)
1009                 if ((ret = make_disks(zhp, child[c])) != 0)
1010                         return (ret);
1011
1012         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_SPARES,
1013             &child, &children) == 0)
1014                 for (c = 0; c < children; c++)
1015                         if ((ret = make_disks(zhp, child[c])) != 0)
1016                                 return (ret);
1017
1018         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1019             &child, &children) == 0)
1020                 for (c = 0; c < children; c++)
1021                         if ((ret = make_disks(zhp, child[c])) != 0)
1022                                 return (ret);
1023
1024         return (0);
1025 }
1026
1027 /*
1028  * Determine if the given path is a hot spare within the given configuration.
1029  */
1030 static boolean_t
1031 is_spare(nvlist_t *config, const char *path)
1032 {
1033         int fd;
1034         pool_state_t state;
1035         char *name = NULL;
1036         nvlist_t *label;
1037         uint64_t guid, spareguid;
1038         nvlist_t *nvroot;
1039         nvlist_t **spares;
1040         uint_t i, nspares;
1041         boolean_t inuse;
1042
1043         if ((fd = open(path, O_RDONLY|O_EXCL)) < 0)
1044                 return (B_FALSE);
1045
1046         if (zpool_in_use(g_zfs, fd, &state, &name, &inuse) != 0 ||
1047             !inuse ||
1048             state != POOL_STATE_SPARE ||
1049             zpool_read_label(fd, &label) != 0) {
1050                 free(name);
1051                 (void) close(fd);
1052                 return (B_FALSE);
1053         }
1054         free(name);
1055         (void) close(fd);
1056
1057         verify(nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) == 0);
1058         nvlist_free(label);
1059
1060         verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
1061             &nvroot) == 0);
1062         if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1063             &spares, &nspares) == 0) {
1064                 for (i = 0; i < nspares; i++) {
1065                         verify(nvlist_lookup_uint64(spares[i],
1066                             ZPOOL_CONFIG_GUID, &spareguid) == 0);
1067                         if (spareguid == guid)
1068                                 return (B_TRUE);
1069                 }
1070         }
1071
1072         return (B_FALSE);
1073 }
1074
1075 /*
1076  * Go through and find any devices that are in use.  We rely on libdiskmgt for
1077  * the majority of this task.
1078  */
1079 static int
1080 check_in_use(nvlist_t *config, nvlist_t *nv, boolean_t force,
1081     boolean_t replacing, boolean_t isspare)
1082 {
1083         nvlist_t **child;
1084         uint_t c, children;
1085         char *type, *path;
1086         int ret = 0;
1087         char buf[MAXPATHLEN];
1088         uint64_t wholedisk = B_FALSE;
1089
1090         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
1091
1092         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
1093             &child, &children) != 0) {
1094
1095                 verify(!nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path));
1096                 if (strcmp(type, VDEV_TYPE_DISK) == 0)
1097                         verify(!nvlist_lookup_uint64(nv,
1098                                ZPOOL_CONFIG_WHOLE_DISK, &wholedisk));
1099
1100                 /*
1101                  * As a generic check, we look to see if this is a replace of a
1102                  * hot spare within the same pool.  If so, we allow it
1103                  * regardless of what libblkid or zpool_in_use() says.
1104                  */
1105                 if (replacing) {
1106                         if (wholedisk)
1107                                 (void) snprintf(buf, sizeof (buf), "%ss0",
1108                                     path);
1109                         else
1110                                 (void) strlcpy(buf, path, sizeof (buf));
1111
1112                         if (is_spare(config, buf))
1113                                 return (0);
1114                 }
1115
1116                 if (strcmp(type, VDEV_TYPE_DISK) == 0)
1117                         ret = check_device(path, force, isspare, wholedisk);
1118
1119                 if (strcmp(type, VDEV_TYPE_FILE) == 0)
1120                         ret = check_file(path, force, isspare);
1121
1122                 return (ret);
1123         }
1124
1125         for (c = 0; c < children; c++)
1126                 if ((ret = check_in_use(config, child[c], force,
1127                     replacing, B_FALSE)) != 0)
1128                         return (ret);
1129
1130         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_SPARES,
1131             &child, &children) == 0)
1132                 for (c = 0; c < children; c++)
1133                         if ((ret = check_in_use(config, child[c], force,
1134                             replacing, B_TRUE)) != 0)
1135                                 return (ret);
1136
1137         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1138             &child, &children) == 0)
1139                 for (c = 0; c < children; c++)
1140                         if ((ret = check_in_use(config, child[c], force,
1141                             replacing, B_FALSE)) != 0)
1142                                 return (ret);
1143
1144         return (0);
1145 }
1146
1147 static const char *
1148 is_grouping(const char *type, int *mindev, int *maxdev)
1149 {
1150         if (strncmp(type, "raidz", 5) == 0) {
1151                 const char *p = type + 5;
1152                 char *end;
1153                 long nparity;
1154
1155                 if (*p == '\0') {
1156                         nparity = 1;
1157                 } else if (*p == '0') {
1158                         return (NULL); /* no zero prefixes allowed */
1159                 } else {
1160                         errno = 0;
1161                         nparity = strtol(p, &end, 10);
1162                         if (errno != 0 || nparity < 1 || nparity >= 255 ||
1163                             *end != '\0')
1164                                 return (NULL);
1165                 }
1166
1167                 if (mindev != NULL)
1168                         *mindev = nparity + 1;
1169                 if (maxdev != NULL)
1170                         *maxdev = 255;
1171                 return (VDEV_TYPE_RAIDZ);
1172         }
1173
1174         if (maxdev != NULL)
1175                 *maxdev = INT_MAX;
1176
1177         if (strcmp(type, "mirror") == 0) {
1178                 if (mindev != NULL)
1179                         *mindev = 2;
1180                 return (VDEV_TYPE_MIRROR);
1181         }
1182
1183         if (strcmp(type, "spare") == 0) {
1184                 if (mindev != NULL)
1185                         *mindev = 1;
1186                 return (VDEV_TYPE_SPARE);
1187         }
1188
1189         if (strcmp(type, "log") == 0) {
1190                 if (mindev != NULL)
1191                         *mindev = 1;
1192                 return (VDEV_TYPE_LOG);
1193         }
1194
1195         if (strcmp(type, "cache") == 0) {
1196                 if (mindev != NULL)
1197                         *mindev = 1;
1198                 return (VDEV_TYPE_L2CACHE);
1199         }
1200
1201         return (NULL);
1202 }
1203
1204 /*
1205  * Construct a syntactically valid vdev specification,
1206  * and ensure that all devices and files exist and can be opened.
1207  * Note: we don't bother freeing anything in the error paths
1208  * because the program is just going to exit anyway.
1209  */
1210 nvlist_t *
1211 construct_spec(nvlist_t *props, int argc, char **argv)
1212 {
1213         nvlist_t *nvroot, *nv, **top, **spares, **l2cache;
1214         int t, toplevels, mindev, maxdev, nspares, nlogs, nl2cache;
1215         const char *type;
1216         uint64_t is_log;
1217         boolean_t seen_logs;
1218
1219         top = NULL;
1220         toplevels = 0;
1221         spares = NULL;
1222         l2cache = NULL;
1223         nspares = 0;
1224         nlogs = 0;
1225         nl2cache = 0;
1226         is_log = B_FALSE;
1227         seen_logs = B_FALSE;
1228
1229         while (argc > 0) {
1230                 nv = NULL;
1231
1232                 /*
1233                  * If it's a mirror or raidz, the subsequent arguments are
1234                  * its leaves -- until we encounter the next mirror or raidz.
1235                  */
1236                 if ((type = is_grouping(argv[0], &mindev, &maxdev)) != NULL) {
1237                         nvlist_t **child = NULL;
1238                         int c, children = 0;
1239
1240                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1241                                 if (spares != NULL) {
1242                                         (void) fprintf(stderr,
1243                                             gettext("invalid vdev "
1244                                             "specification: 'spare' can be "
1245                                             "specified only once\n"));
1246                                         return (NULL);
1247                                 }
1248                                 is_log = B_FALSE;
1249                         }
1250
1251                         if (strcmp(type, VDEV_TYPE_LOG) == 0) {
1252                                 if (seen_logs) {
1253                                         (void) fprintf(stderr,
1254                                             gettext("invalid vdev "
1255                                             "specification: 'log' can be "
1256                                             "specified only once\n"));
1257                                         return (NULL);
1258                                 }
1259                                 seen_logs = B_TRUE;
1260                                 is_log = B_TRUE;
1261                                 argc--;
1262                                 argv++;
1263                                 /*
1264                                  * A log is not a real grouping device.
1265                                  * We just set is_log and continue.
1266                                  */
1267                                 continue;
1268                         }
1269
1270                         if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1271                                 if (l2cache != NULL) {
1272                                         (void) fprintf(stderr,
1273                                             gettext("invalid vdev "
1274                                             "specification: 'cache' can be "
1275                                             "specified only once\n"));
1276                                         return (NULL);
1277                                 }
1278                                 is_log = B_FALSE;
1279                         }
1280
1281                         if (is_log) {
1282                                 if (strcmp(type, VDEV_TYPE_MIRROR) != 0) {
1283                                         (void) fprintf(stderr,
1284                                             gettext("invalid vdev "
1285                                             "specification: unsupported 'log' "
1286                                             "device: %s\n"), type);
1287                                         return (NULL);
1288                                 }
1289                                 nlogs++;
1290                         }
1291
1292                         for (c = 1; c < argc; c++) {
1293                                 if (is_grouping(argv[c], NULL, NULL) != NULL)
1294                                         break;
1295                                 children++;
1296                                 child = realloc(child,
1297                                     children * sizeof (nvlist_t *));
1298                                 if (child == NULL)
1299                                         zpool_no_memory();
1300                                 if ((nv = make_leaf_vdev(props, argv[c], B_FALSE))
1301                                     == NULL)
1302                                         return (NULL);
1303                                 child[children - 1] = nv;
1304                         }
1305
1306                         if (children < mindev) {
1307                                 (void) fprintf(stderr, gettext("invalid vdev "
1308                                     "specification: %s requires at least %d "
1309                                     "devices\n"), argv[0], mindev);
1310                                 return (NULL);
1311                         }
1312
1313                         if (children > maxdev) {
1314                                 (void) fprintf(stderr, gettext("invalid vdev "
1315                                     "specification: %s supports no more than "
1316                                     "%d devices\n"), argv[0], maxdev);
1317                                 return (NULL);
1318                         }
1319
1320                         argc -= c;
1321                         argv += c;
1322
1323                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1324                                 spares = child;
1325                                 nspares = children;
1326                                 continue;
1327                         } else if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1328                                 l2cache = child;
1329                                 nl2cache = children;
1330                                 continue;
1331                         } else {
1332                                 verify(nvlist_alloc(&nv, NV_UNIQUE_NAME,
1333                                     0) == 0);
1334                                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_TYPE,
1335                                     type) == 0);
1336                                 verify(nvlist_add_uint64(nv,
1337                                     ZPOOL_CONFIG_IS_LOG, is_log) == 0);
1338                                 if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
1339                                         verify(nvlist_add_uint64(nv,
1340                                             ZPOOL_CONFIG_NPARITY,
1341                                             mindev - 1) == 0);
1342                                 }
1343                                 verify(nvlist_add_nvlist_array(nv,
1344                                     ZPOOL_CONFIG_CHILDREN, child,
1345                                     children) == 0);
1346
1347                                 for (c = 0; c < children; c++)
1348                                         nvlist_free(child[c]);
1349                                 free(child);
1350                         }
1351                 } else {
1352                         /*
1353                          * We have a device.  Pass off to make_leaf_vdev() to
1354                          * construct the appropriate nvlist describing the vdev.
1355                          */
1356                         if ((nv = make_leaf_vdev(props, argv[0], is_log)) == NULL)
1357                                 return (NULL);
1358                         if (is_log)
1359                                 nlogs++;
1360                         argc--;
1361                         argv++;
1362                 }
1363
1364                 toplevels++;
1365                 top = realloc(top, toplevels * sizeof (nvlist_t *));
1366                 if (top == NULL)
1367                         zpool_no_memory();
1368                 top[toplevels - 1] = nv;
1369         }
1370
1371         if (toplevels == 0 && nspares == 0 && nl2cache == 0) {
1372                 (void) fprintf(stderr, gettext("invalid vdev "
1373                     "specification: at least one toplevel vdev must be "
1374                     "specified\n"));
1375                 return (NULL);
1376         }
1377
1378         if (seen_logs && nlogs == 0) {
1379                 (void) fprintf(stderr, gettext("invalid vdev specification: "
1380                     "log requires at least 1 device\n"));
1381                 return (NULL);
1382         }
1383
1384         /*
1385          * Finally, create nvroot and add all top-level vdevs to it.
1386          */
1387         verify(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, 0) == 0);
1388         verify(nvlist_add_string(nvroot, ZPOOL_CONFIG_TYPE,
1389             VDEV_TYPE_ROOT) == 0);
1390         verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
1391             top, toplevels) == 0);
1392         if (nspares != 0)
1393                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1394                     spares, nspares) == 0);
1395         if (nl2cache != 0)
1396                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
1397                     l2cache, nl2cache) == 0);
1398
1399         for (t = 0; t < toplevels; t++)
1400                 nvlist_free(top[t]);
1401         for (t = 0; t < nspares; t++)
1402                 nvlist_free(spares[t]);
1403         for (t = 0; t < nl2cache; t++)
1404                 nvlist_free(l2cache[t]);
1405         if (spares)
1406                 free(spares);
1407         if (l2cache)
1408                 free(l2cache);
1409         free(top);
1410
1411         return (nvroot);
1412 }
1413
1414 nvlist_t *
1415 split_mirror_vdev(zpool_handle_t *zhp, char *newname, nvlist_t *props,
1416     splitflags_t flags, int argc, char **argv)
1417 {
1418         nvlist_t *newroot = NULL, **child;
1419         uint_t c, children;
1420
1421         if (argc > 0) {
1422                 if ((newroot = construct_spec(props, argc, argv)) == NULL) {
1423                         (void) fprintf(stderr, gettext("Unable to build a "
1424                             "pool from the specified devices\n"));
1425                         return (NULL);
1426                 }
1427
1428                 if (!flags.dryrun && make_disks(zhp, newroot) != 0) {
1429                         nvlist_free(newroot);
1430                         return (NULL);
1431                 }
1432
1433                 /* avoid any tricks in the spec */
1434                 verify(nvlist_lookup_nvlist_array(newroot,
1435                     ZPOOL_CONFIG_CHILDREN, &child, &children) == 0);
1436                 for (c = 0; c < children; c++) {
1437                         char *path;
1438                         const char *type;
1439                         int min, max;
1440
1441                         verify(nvlist_lookup_string(child[c],
1442                             ZPOOL_CONFIG_PATH, &path) == 0);
1443                         if ((type = is_grouping(path, &min, &max)) != NULL) {
1444                                 (void) fprintf(stderr, gettext("Cannot use "
1445                                     "'%s' as a device for splitting\n"), type);
1446                                 nvlist_free(newroot);
1447                                 return (NULL);
1448                         }
1449                 }
1450         }
1451
1452         if (zpool_vdev_split(zhp, newname, &newroot, props, flags) != 0) {
1453                 if (newroot != NULL)
1454                         nvlist_free(newroot);
1455                 return (NULL);
1456         }
1457
1458         return (newroot);
1459 }
1460
1461 /*
1462  * Get and validate the contents of the given vdev specification.  This ensures
1463  * that the nvlist returned is well-formed, that all the devices exist, and that
1464  * they are not currently in use by any other known consumer.  The 'poolconfig'
1465  * parameter is the current configuration of the pool when adding devices
1466  * existing pool, and is used to perform additional checks, such as changing the
1467  * replication level of the pool.  It can be 'NULL' to indicate that this is a
1468  * new pool.  The 'force' flag controls whether devices should be forcefully
1469  * added, even if they appear in use.
1470  */
1471 nvlist_t *
1472 make_root_vdev(zpool_handle_t *zhp, nvlist_t *props, int force, int check_rep,
1473     boolean_t replacing, boolean_t dryrun, int argc, char **argv)
1474 {
1475         nvlist_t *newroot;
1476         nvlist_t *poolconfig = NULL;
1477         is_force = force;
1478
1479         /*
1480          * Construct the vdev specification.  If this is successful, we know
1481          * that we have a valid specification, and that all devices can be
1482          * opened.
1483          */
1484         if ((newroot = construct_spec(props, argc, argv)) == NULL)
1485                 return (NULL);
1486
1487         if (zhp && ((poolconfig = zpool_get_config(zhp, NULL)) == NULL))
1488                 return (NULL);
1489
1490         /*
1491          * Validate each device to make sure that its not shared with another
1492          * subsystem.  We do this even if 'force' is set, because there are some
1493          * uses (such as a dedicated dump device) that even '-f' cannot
1494          * override.
1495          */
1496         if (check_in_use(poolconfig, newroot, force, replacing, B_FALSE) != 0) {
1497                 nvlist_free(newroot);
1498                 return (NULL);
1499         }
1500
1501         /*
1502          * Check the replication level of the given vdevs and report any errors
1503          * found.  We include the existing pool spec, if any, as we need to
1504          * catch changes against the existing replication level.
1505          */
1506         if (check_rep && check_replication(poolconfig, newroot) != 0) {
1507                 nvlist_free(newroot);
1508                 return (NULL);
1509         }
1510
1511         /*
1512          * Run through the vdev specification and label any whole disks found.
1513          */
1514         if (!dryrun && make_disks(zhp, newroot) != 0) {
1515                 nvlist_free(newroot);
1516                 return (NULL);
1517         }
1518
1519         return (newroot);
1520 }