cmd/zpool/zpool_vdev.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21
  22 /*
  23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  24  */
  25
  26 /*
  27  * Functions to convert between a list of vdevs and an nvlist representing the
  28  * configuration.  Each entry in the list can be one of:
  29  *
  30  *      Device vdevs
  31  *              disk=(path=..., devid=...)
  32  *              file=(path=...)
  33  *
  34  *      Group vdevs
  35  *              raidz[1|2]=(...)
  36  *              mirror=(...)
  37  *
  38  *      Hot spares
  39  *
  40  * While the underlying implementation supports it, group vdevs cannot contain
  41  * other group vdevs.  All userland verification of devices is contained within
  42  * this file.  If successful, the nvlist returned can be passed directly to the
  43  * kernel; we've done as much verification as possible in userland.
  44  *
  45  * Hot spares are a special case, and passed down as an array of disk vdevs, at
  46  * the same level as the root of the vdev tree.
  47  *
  48  * The only function exported by this file is 'make_root_vdev'.  The
  49  * function performs several passes:
  50  *
  51  *      1. Construct the vdev specification.  Performs syntax validation and
  52  *         makes sure each device is valid.
  53  *      2. Check for devices in use.  Using libblkid to make sure that no
  54  *         devices are also in use.  Some can be overridden using the 'force'
  55  *         flag, others cannot.
  56  *      3. Check for replication errors if the 'force' flag is not specified.
  57  *         validates that the replication level is consistent across the
  58  *         entire pool.
  59  *      4. Call libzfs to label any whole disks with an EFI label.
  60  */
  61
  62 #include <assert.h>
  63 #include <ctype.h>
  64 #include <devid.h>
  65 #include <errno.h>
  66 #include <fcntl.h>
  67 #include <libintl.h>
  68 #include <libnvpair.h>
  69 #include <limits.h>
  70 #include <stdio.h>
  71 #include <string.h>
  72 #include <unistd.h>
  73 #include <sys/efi_partition.h>
  74 #include <sys/stat.h>
  75 #include <sys/vtoc.h>
  76 #include <sys/mntent.h>
  77 #include <uuid/uuid.h>
  78 #ifdef HAVE_LIBBLKID
  79 #include <blkid/blkid.h>
  80 #else
  81 #define blkid_cache void *
  82 #endif /* HAVE_LIBBLKID */
  83
  84 #include "zpool_util.h"
  85
  86 /*
  87  * For any given vdev specification, we can have multiple errors.  The
  88  * vdev_error() function keeps track of whether we have seen an error yet, and
  89  * prints out a header if its the first error we've seen.
  90  */
  91 boolean_t error_seen;
  92 boolean_t is_force;
  93
  94 /*PRINTFLIKE1*/
  95 static void
  96 vdev_error(const char *fmt, ...)
  97 {
  98         va_list ap;
  99
 100         if (!error_seen) {
 101                 (void) fprintf(stderr, gettext("invalid vdev specification\n"));
 102                 if (!is_force)
 103                         (void) fprintf(stderr, gettext("use '-f' to override "
 104                             "the following errors:\n"));
 105                 else
 106                         (void) fprintf(stderr, gettext("the following errors "
 107                             "must be manually repaired:\n"));
 108                 error_seen = B_TRUE;
 109         }
 110
 111         va_start(ap, fmt);
 112         (void) vfprintf(stderr, fmt, ap);
 113         va_end(ap);
 114 }
 115
 116 /*
 117  * Check that a file is valid.  All we can do in this case is check that it's
 118  * not in use by another pool, and not in use by swap.
 119  */
 120 static int
 121 check_file(const char *file, boolean_t force, boolean_t isspare)
 122 {
 123         char  *name;
 124         int fd;
 125         int ret = 0;
 126         pool_state_t state;
 127         boolean_t inuse;
 128
 129         if ((fd = open(file, O_RDONLY)) < 0)
 130                 return (0);
 131
 132         if (zpool_in_use(g_zfs, fd, &state, &name, &inuse) == 0 && inuse) {
 133                 const char *desc;
 134
 135                 switch (state) {
 136                 case POOL_STATE_ACTIVE:
 137                         desc = gettext("active");
 138                         break;
 139
 140                 case POOL_STATE_EXPORTED:
 141                         desc = gettext("exported");
 142                         break;
 143
 144                 case POOL_STATE_POTENTIALLY_ACTIVE:
 145                         desc = gettext("potentially active");
 146                         break;
 147
 148                 default:
 149                         desc = gettext("unknown");
 150                         break;
 151                 }
 152
 153                 /*
 154                  * Allow hot spares to be shared between pools.
 155                  */
 156                 if (state == POOL_STATE_SPARE && isspare)
 157                         return (0);
 158
 159                 if (state == POOL_STATE_ACTIVE ||
 160                     state == POOL_STATE_SPARE || !force) {
 161                         switch (state) {
 162                         case POOL_STATE_SPARE:
 163                                 vdev_error(gettext("%s is reserved as a hot "
 164                                     "spare for pool %s\n"), file, name);
 165                                 break;
 166                         default:
 167                                 vdev_error(gettext("%s is part of %s pool "
 168                                     "'%s'\n"), file, desc, name);
 169                                 break;
 170                         }
 171                         ret = -1;
 172                 }
 173
 174                 free(name);
 175         }
 176
 177         (void) close(fd);
 178         return (ret);
 179 }
 180
 181 static void
 182 check_error(int err)
 183 {
 184         (void) fprintf(stderr, gettext("warning: device in use checking "
 185             "failed: %s\n"), strerror(err));
 186 }
 187
 188 static int
 189 check_slice(const char *path, blkid_cache cache, int force, boolean_t isspare)
 190 {
 191         struct stat64 statbuf;
 192         int err;
 193 #ifdef HAVE_LIBBLKID
 194         char *value;
 195 #endif /* HAVE_LIBBLKID */
 196
 197         if (stat64(path, &statbuf) != 0) {
 198                 vdev_error(gettext("cannot stat %s: %s\n"),
 199                            path, strerror(errno));
 200                 return (-1);
 201         }
 202
 203 #ifdef HAVE_LIBBLKID
 204         /* No valid type detected device is safe to use */
 205         value = blkid_get_tag_value(cache, "TYPE", path);
 206         if (value == NULL)
 207                 return (0);
 208
 209         /*
 210          * If libblkid detects a ZFS device, we check the device
 211          * using check_file() to see if it's safe.  The one safe
 212          * case is a spare device shared between multiple pools.
 213          */
 214         if (strcmp(value, "zfs") == 0) {
 215                 err = check_file(path, force, isspare);
 216         } else {
 217                 if (force) {
 218                         err = 0;
 219                 } else {
 220                         err = -1;
 221                         vdev_error(gettext("%s contains a filesystem of "
 222                                    "type '%s'\n"), path, value);
 223                 }
 224         }
 225
 226         free(value);
 227 #else
 228         err = check_file(path, force, isspare);
 229 #endif /* HAVE_LIBBLKID */
 230
 231         return (err);
 232 }
 233
 234 /*
 235  * Validate a whole disk.  Iterate over all slices on the disk and make sure
 236  * that none is in use by calling check_slice().
 237  */
 238 static int
 239 check_disk(const char *path, blkid_cache cache, int force,
 240            boolean_t isspare, boolean_t iswholedisk)
 241 {
 242         struct dk_gpt *vtoc;
 243         char slice_path[MAXPATHLEN];
 244         int err = 0;
 245         int fd, i;
 246
 247         /* This is not a wholedisk we only check the given partition */
 248         if (!iswholedisk)
 249                 return check_slice(path, cache, force, isspare);
 250
 251         /*
 252          * When the device is a whole disk try to read the efi partition
 253          * label.  If this is successful we safely check the all of the
 254          * partitions.  However, when it fails it may simply be because
 255          * the disk is partitioned via the MBR.  Since we currently can
 256          * not easily decode the MBR return a failure and prompt to the
 257          * user to use force option since we cannot check the partitions.
 258          */
 259         if ((fd = open(path, O_RDWR|O_DIRECT|O_EXCL)) < 0) {
 260                 check_error(errno);
 261                 return -1;
 262         }
 263
 264         if ((err = efi_alloc_and_read(fd, &vtoc)) != 0) {
 265                 (void) close(fd);
 266
 267                 if (force) {
 268                         return 0;
 269                 } else {
 270                         vdev_error(gettext("%s does not contain an EFI "
 271                             "label but it may contain partition\n"
 272                             "information in the MBR.\n"), path);
 273                         return -1;
 274                 }
 275         }
 276
 277         /*
 278          * The primary efi partition label is damaged however the secondary
 279          * label at the end of the device is intact.  Rather than use this
 280          * label we should play it safe and treat this as a non efi device.
 281          */
 282         if (vtoc->efi_flags & EFI_GPT_PRIMARY_CORRUPT) {
 283                 efi_free(vtoc);
 284                 (void) close(fd);
 285
 286                 if (force) {
 287                         /* Partitions will no be created using the backup */
 288                         return 0;
 289                 } else {
 290                         vdev_error(gettext("%s contains a corrupt primary "
 291                             "EFI label.\n"), path);
 292                         return -1;
 293                 }
 294         }
 295
 296         for (i = 0; i < vtoc->efi_nparts; i++) {
 297
 298                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED ||
 299                     uuid_is_null((uchar_t *)&vtoc->efi_parts[i].p_guid))
 300                         continue;
 301
 302                 if (strncmp(path, UDISK_ROOT, strlen(UDISK_ROOT)) == 0)
 303                         (void) snprintf(slice_path, sizeof (slice_path),
 304                             "%s%s%d", path, "-part", i+1);
 305                 else
 306                         (void) snprintf(slice_path, sizeof (slice_path),
 307                             "%s%s%d", path, isdigit(path[strlen(path)-1]) ?
 308                             "p" : "", i+1);
 309
 310                 err = check_slice(slice_path, cache, force, isspare);
 311                 if (err)
 312                         break;
 313         }
 314
 315         efi_free(vtoc);
 316         (void) close(fd);
 317
 318         return (err);
 319 }
 320
 321 static int
 322 check_device(const char *path, boolean_t force,
 323              boolean_t isspare, boolean_t iswholedisk)
 324 {
 325         static blkid_cache cache = NULL;
 326
 327 #ifdef HAVE_LIBBLKID
 328         /*
 329          * There is no easy way to add a correct blkid_put_cache() call,
 330          * memory will be reclaimed when the command exits.
 331          */
 332         if (cache == NULL) {
 333                 int err;
 334
 335                 if ((err = blkid_get_cache(&cache, NULL)) != 0) {
 336                         check_error(err);
 337                         return -1;
 338                 }
 339
 340                 if ((err = blkid_probe_all(cache)) != 0) {
 341                         blkid_put_cache(cache);
 342                         check_error(err);
 343                         return -1;
 344                 }
 345         }
 346 #endif /* HAVE_LIBBLKID */
 347
 348         return check_disk(path, cache, force, isspare, iswholedisk);
 349 }
 350
 351 /*
 352  * By "whole disk" we mean an entire physical disk (something we can
 353  * label, toggle the write cache on, etc.) as opposed to the full
 354  * capacity of a pseudo-device such as lofi or did.  We act as if we
 355  * are labeling the disk, which should be a pretty good test of whether
 356  * it's a viable device or not.  Returns B_TRUE if it is and B_FALSE if
 357  * it isn't.
 358  */
 359 static boolean_t
 360 is_whole_disk(const char *path)
 361 {
 362         struct dk_gpt *label;
 363         int     fd;
 364
 365         if ((fd = open(path, O_RDWR|O_DIRECT|O_EXCL)) < 0)
 366                 return (B_FALSE);
 367         if (efi_alloc_and_init(fd, EFI_NUMPAR, &label) != 0) {
 368                 (void) close(fd);
 369                 return (B_FALSE);
 370         }
 371         efi_free(label);
 372         (void) close(fd);
 373         return (B_TRUE);
 374 }
 375
 376 /*
 377  * This may be a shorthand device path or it could be total gibberish.
 378  * Check to see if it's a known device in /dev/, /dev/disk/by-id,
 379  * /dev/disk/by-label, /dev/disk/by-path, /dev/disk/by-uuid, or
 380  * /dev/disk/zpool/.  As part of this check, see if we've been given
 381  * an entire disk (minus the slice number).
 382  */
 383 static int
 384 is_shorthand_path(const char *arg, char *path,
 385                   struct stat64 *statbuf, boolean_t *wholedisk)
 386 {
 387         if (zfs_resolve_shortname(arg, path, MAXPATHLEN) == 0) {
 388                 *wholedisk = is_whole_disk(path);
 389                 if (*wholedisk || (stat64(path, statbuf) == 0))
 390                         return (0);
 391         }
 392
 393         strlcpy(path, arg, sizeof(path));
 394         memset(statbuf, 0, sizeof(*statbuf));
 395         *wholedisk = B_FALSE;
 396
 397         return (ENOENT);
 398 }
 399
 400 /*
 401  * Create a leaf vdev.  Determine if this is a file or a device.  If it's a
 402  * device, fill in the device id to make a complete nvlist.  Valid forms for a
 403  * leaf vdev are:
 404  *
 405  *      /dev/xxx        Complete disk path
 406  *      /xxx            Full path to file
 407  *      xxx             Shorthand for /dev/disk/yyy/xxx
 408  */
 409 static nvlist_t *
 410 make_leaf_vdev(const char *arg, uint64_t is_log)
 411 {
 412         char path[MAXPATHLEN];
 413         struct stat64 statbuf;
 414         nvlist_t *vdev = NULL;
 415         char *type = NULL;
 416         boolean_t wholedisk = B_FALSE;
 417         int err;
 418
 419         /*
 420          * Determine what type of vdev this is, and put the full path into
 421          * 'path'.  We detect whether this is a device of file afterwards by
 422          * checking the st_mode of the file.
 423          */
 424         if (arg[0] == '/') {
 425                 /*
 426                  * Complete device or file path.  Exact type is determined by
 427                  * examining the file descriptor afterwards.  Symbolic links
 428                  * are resolved to their real paths for the is_whole_disk()
 429                  * and S_ISBLK/S_ISREG type checks.  However, we are careful
 430                  * to store the given path as ZPOOL_CONFIG_PATH to ensure we
 431                  * can leverage udev's persistent device labels.
 432                  */
 433                 if (realpath(arg, path) == NULL) {
 434                         (void) fprintf(stderr,
 435                             gettext("cannot resolve path '%s'\n"), arg);
 436                         return (NULL);
 437                 }
 438
 439                 wholedisk = is_whole_disk(path);
 440                 if (!wholedisk && (stat64(path, &statbuf) != 0)) {
 441                         (void) fprintf(stderr,
 442                             gettext("cannot open '%s': %s\n"),
 443                             path, strerror(errno));
 444                         return (NULL);
 445                 }
 446
 447                 /* After is_whole_disk() check restore original passed path */
 448                 strlcpy(path, arg, MAXPATHLEN);
 449         } else {
 450                 err = is_shorthand_path(arg, path, &statbuf, &wholedisk);
 451                 if (err != 0) {
 452                         /*
 453                          * If we got ENOENT, then the user gave us
 454                          * gibberish, so try to direct them with a
 455                          * reasonable error message.  Otherwise,
 456                          * regurgitate strerror() since it's the best we
 457                          * can do.
 458                          */
 459                         if (err == ENOENT) {
 460                                 (void) fprintf(stderr,
 461                                     gettext("cannot open '%s': no such "
 462                                     "device in %s\n"), arg, DISK_ROOT);
 463                                 (void) fprintf(stderr,
 464                                     gettext("must be a full path or "
 465                                     "shorthand device name\n"));
 466                                 return (NULL);
 467                         } else {
 468                                 (void) fprintf(stderr,
 469                                     gettext("cannot open '%s': %s\n"),
 470                                     path, strerror(errno));
 471                                 return (NULL);
 472                         }
 473                 }
 474         }
 475
 476         /*
 477          * Determine whether this is a device or a file.
 478          */
 479         if (wholedisk || S_ISBLK(statbuf.st_mode)) {
 480                 type = VDEV_TYPE_DISK;
 481         } else if (S_ISREG(statbuf.st_mode)) {
 482                 type = VDEV_TYPE_FILE;
 483         } else {
 484                 (void) fprintf(stderr, gettext("cannot use '%s': must be a "
 485                     "block device or regular file\n"), path);
 486                 return (NULL);
 487         }
 488
 489         /*
 490          * Finally, we have the complete device or file, and we know that it is
 491          * acceptable to use.  Construct the nvlist to describe this vdev.  All
 492          * vdevs have a 'path' element, and devices also have a 'devid' element.
 493          */
 494         verify(nvlist_alloc(&vdev, NV_UNIQUE_NAME, 0) == 0);
 495         verify(nvlist_add_string(vdev, ZPOOL_CONFIG_PATH, path) == 0);
 496         verify(nvlist_add_string(vdev, ZPOOL_CONFIG_TYPE, type) == 0);
 497         verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_IS_LOG, is_log) == 0);
 498         if (strcmp(type, VDEV_TYPE_DISK) == 0)
 499                 verify(nvlist_add_uint64(vdev, ZPOOL_CONFIG_WHOLE_DISK,
 500                     (uint64_t)wholedisk) == 0);
 501
 502 #if defined(__sun__) || defined(__sun)
 503         /*
 504          * For a whole disk, defer getting its devid until after labeling it.
 505          */
 506         if (S_ISBLK(statbuf.st_mode) && !wholedisk) {
 507                 /*
 508                  * Get the devid for the device.
 509                  */
 510                 int fd;
 511                 ddi_devid_t devid;
 512                 char *minor = NULL, *devid_str = NULL;
 513
 514                 if ((fd = open(path, O_RDONLY|O_EXCL)) < 0) {
 515                         (void) fprintf(stderr, gettext("cannot open '%s': "
 516                             "%s\n"), path, strerror(errno));
 517                         nvlist_free(vdev);
 518                         return (NULL);
 519                 }
 520
 521                 if (devid_get(fd, &devid) == 0) {
 522                         if (devid_get_minor_name(fd, &minor) == 0 &&
 523                             (devid_str = devid_str_encode(devid, minor)) !=
 524                             NULL) {
 525                                 verify(nvlist_add_string(vdev,
 526                                     ZPOOL_CONFIG_DEVID, devid_str) == 0);
 527                         }
 528                         if (devid_str != NULL)
 529                                 devid_str_free(devid_str);
 530                         if (minor != NULL)
 531                                 devid_str_free(minor);
 532                         devid_free(devid);
 533                 }
 534
 535                 (void) close(fd);
 536         }
 537 #endif
 538
 539         return (vdev);
 540 }
 541
 542 /*
 543  * Go through and verify the replication level of the pool is consistent.
 544  * Performs the following checks:
 545  *
 546  *      For the new spec, verifies that devices in mirrors and raidz are the
 547  *      same size.
 548  *
 549  *      If the current configuration already has inconsistent replication
 550  *      levels, ignore any other potential problems in the new spec.
 551  *
 552  *      Otherwise, make sure that the current spec (if there is one) and the new
 553  *      spec have consistent replication levels.
 554  */
 555 typedef struct replication_level {
 556         char *zprl_type;
 557         uint64_t zprl_children;
 558         uint64_t zprl_parity;
 559 } replication_level_t;
 560
 561 #define ZPOOL_FUZZ      (16 * 1024 * 1024)
 562
 563 /*
 564  * Given a list of toplevel vdevs, return the current replication level.  If
 565  * the config is inconsistent, then NULL is returned.  If 'fatal' is set, then
 566  * an error message will be displayed for each self-inconsistent vdev.
 567  */
 568 static replication_level_t *
 569 get_replication(nvlist_t *nvroot, boolean_t fatal)
 570 {
 571         nvlist_t **top;
 572         uint_t t, toplevels;
 573         nvlist_t **child;
 574         uint_t c, children;
 575         nvlist_t *nv;
 576         char *type;
 577         replication_level_t lastrep = { 0 }, rep, *ret;
 578         boolean_t dontreport;
 579
 580         ret = safe_malloc(sizeof (replication_level_t));
 581
 582         verify(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
 583             &top, &toplevels) == 0);
 584
 585         lastrep.zprl_type = NULL;
 586         for (t = 0; t < toplevels; t++) {
 587                 uint64_t is_log = B_FALSE;
 588
 589                 nv = top[t];
 590
 591                 /*
 592                  * For separate logs we ignore the top level vdev replication
 593                  * constraints.
 594                  */
 595                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &is_log);
 596                 if (is_log)
 597                         continue;
 598
 599                 verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE,
 600                     &type) == 0);
 601                 if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 602                     &child, &children) != 0) {
 603                         /*
 604                          * This is a 'file' or 'disk' vdev.
 605                          */
 606                         rep.zprl_type = type;
 607                         rep.zprl_children = 1;
 608                         rep.zprl_parity = 0;
 609                 } else {
 610                         uint64_t vdev_size;
 611
 612                         /*
 613                          * This is a mirror or RAID-Z vdev.  Go through and make
 614                          * sure the contents are all the same (files vs. disks),
 615                          * keeping track of the number of elements in the
 616                          * process.
 617                          *
 618                          * We also check that the size of each vdev (if it can
 619                          * be determined) is the same.
 620                          */
 621                         rep.zprl_type = type;
 622                         rep.zprl_children = 0;
 623
 624                         if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
 625                                 verify(nvlist_lookup_uint64(nv,
 626                                     ZPOOL_CONFIG_NPARITY,
 627                                     &rep.zprl_parity) == 0);
 628                                 assert(rep.zprl_parity != 0);
 629                         } else {
 630                                 rep.zprl_parity = 0;
 631                         }
 632
 633                         /*
 634                          * The 'dontreport' variable indicates that we've
 635                          * already reported an error for this spec, so don't
 636                          * bother doing it again.
 637                          */
 638                         type = NULL;
 639                         dontreport = 0;
 640                         vdev_size = -1ULL;
 641                         for (c = 0; c < children; c++) {
 642                                 nvlist_t *cnv = child[c];
 643                                 char *path;
 644                                 struct stat64 statbuf;
 645                                 uint64_t size = -1ULL;
 646                                 char *childtype;
 647                                 int fd, err;
 648
 649                                 rep.zprl_children++;
 650
 651                                 verify(nvlist_lookup_string(cnv,
 652                                     ZPOOL_CONFIG_TYPE, &childtype) == 0);
 653
 654                                 /*
 655                                  * If this is a replacing or spare vdev, then
 656                                  * get the real first child of the vdev.
 657                                  */
 658                                 if (strcmp(childtype,
 659                                     VDEV_TYPE_REPLACING) == 0 ||
 660                                     strcmp(childtype, VDEV_TYPE_SPARE) == 0) {
 661                                         nvlist_t **rchild;
 662                                         uint_t rchildren;
 663
 664                                         verify(nvlist_lookup_nvlist_array(cnv,
 665                                             ZPOOL_CONFIG_CHILDREN, &rchild,
 666                                             &rchildren) == 0);
 667                                         assert(rchildren == 2);
 668                                         cnv = rchild[0];
 669
 670                                         verify(nvlist_lookup_string(cnv,
 671                                             ZPOOL_CONFIG_TYPE,
 672                                             &childtype) == 0);
 673                                 }
 674
 675                                 verify(nvlist_lookup_string(cnv,
 676                                     ZPOOL_CONFIG_PATH, &path) == 0);
 677
 678                                 /*
 679                                  * If we have a raidz/mirror that combines disks
 680                                  * with files, report it as an error.
 681                                  */
 682                                 if (!dontreport && type != NULL &&
 683                                     strcmp(type, childtype) != 0) {
 684                                         if (ret != NULL)
 685                                                 free(ret);
 686                                         ret = NULL;
 687                                         if (fatal)
 688                                                 vdev_error(gettext(
 689                                                     "mismatched replication "
 690                                                     "level: %s contains both "
 691                                                     "files and devices\n"),
 692                                                     rep.zprl_type);
 693                                         else
 694                                                 return (NULL);
 695                                         dontreport = B_TRUE;
 696                                 }
 697
 698                                 /*
 699                                  * According to stat(2), the value of 'st_size'
 700                                  * is undefined for block devices and character
 701                                  * devices.  But there is no effective way to
 702                                  * determine the real size in userland.
 703                                  *
 704                                  * Instead, we'll take advantage of an
 705                                  * implementation detail of spec_size().  If the
 706                                  * device is currently open, then we (should)
 707                                  * return a valid size.
 708                                  *
 709                                  * If we still don't get a valid size (indicated
 710                                  * by a size of 0 or MAXOFFSET_T), then ignore
 711                                  * this device altogether.
 712                                  */
 713                                 if ((fd = open(path, O_RDONLY)) >= 0) {
 714                                         err = fstat64(fd, &statbuf);
 715                                         (void) close(fd);
 716                                 } else {
 717                                         err = stat64(path, &statbuf);
 718                                 }
 719
 720                                 if (err != 0 ||
 721                                     statbuf.st_size == 0 ||
 722                                     statbuf.st_size == MAXOFFSET_T)
 723                                         continue;
 724
 725                                 size = statbuf.st_size;
 726
 727                                 /*
 728                                  * Also make sure that devices and
 729                                  * slices have a consistent size.  If
 730                                  * they differ by a significant amount
 731                                  * (~16MB) then report an error.
 732                                  */
 733                                 if (!dontreport &&
 734                                     (vdev_size != -1ULL &&
 735                                     (labs(size - vdev_size) >
 736                                     ZPOOL_FUZZ))) {
 737                                         if (ret != NULL)
 738                                                 free(ret);
 739                                         ret = NULL;
 740                                         if (fatal)
 741                                                 vdev_error(gettext(
 742                                                     "%s contains devices of "
 743                                                     "different sizes\n"),
 744                                                     rep.zprl_type);
 745                                         else
 746                                                 return (NULL);
 747                                         dontreport = B_TRUE;
 748                                 }
 749
 750                                 type = childtype;
 751                                 vdev_size = size;
 752                         }
 753                 }
 754
 755                 /*
 756                  * At this point, we have the replication of the last toplevel
 757                  * vdev in 'rep'.  Compare it to 'lastrep' to see if its
 758                  * different.
 759                  */
 760                 if (lastrep.zprl_type != NULL) {
 761                         if (strcmp(lastrep.zprl_type, rep.zprl_type) != 0) {
 762                                 if (ret != NULL)
 763                                         free(ret);
 764                                 ret = NULL;
 765                                 if (fatal)
 766                                         vdev_error(gettext(
 767                                             "mismatched replication level: "
 768                                             "both %s and %s vdevs are "
 769                                             "present\n"),
 770                                             lastrep.zprl_type, rep.zprl_type);
 771                                 else
 772                                         return (NULL);
 773                         } else if (lastrep.zprl_parity != rep.zprl_parity) {
 774                                 if (ret)
 775                                         free(ret);
 776                                 ret = NULL;
 777                                 if (fatal)
 778                                         vdev_error(gettext(
 779                                             "mismatched replication level: "
 780                                             "both %llu and %llu device parity "
 781                                             "%s vdevs are present\n"),
 782                                             lastrep.zprl_parity,
 783                                             rep.zprl_parity,
 784                                             rep.zprl_type);
 785                                 else
 786                                         return (NULL);
 787                         } else if (lastrep.zprl_children != rep.zprl_children) {
 788                                 if (ret)
 789                                         free(ret);
 790                                 ret = NULL;
 791                                 if (fatal)
 792                                         vdev_error(gettext(
 793                                             "mismatched replication level: "
 794                                             "both %llu-way and %llu-way %s "
 795                                             "vdevs are present\n"),
 796                                             lastrep.zprl_children,
 797                                             rep.zprl_children,
 798                                             rep.zprl_type);
 799                                 else
 800                                         return (NULL);
 801                         }
 802                 }
 803                 lastrep = rep;
 804         }
 805
 806         if (ret != NULL)
 807                 *ret = rep;
 808
 809         return (ret);
 810 }
 811
 812 /*
 813  * Check the replication level of the vdev spec against the current pool.  Calls
 814  * get_replication() to make sure the new spec is self-consistent.  If the pool
 815  * has a consistent replication level, then we ignore any errors.  Otherwise,
 816  * report any difference between the two.
 817  */
 818 static int
 819 check_replication(nvlist_t *config, nvlist_t *newroot)
 820 {
 821         nvlist_t **child;
 822         uint_t  children;
 823         replication_level_t *current = NULL, *new;
 824         int ret;
 825
 826         /*
 827          * If we have a current pool configuration, check to see if it's
 828          * self-consistent.  If not, simply return success.
 829          */
 830         if (config != NULL) {
 831                 nvlist_t *nvroot;
 832
 833                 verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
 834                     &nvroot) == 0);
 835                 if ((current = get_replication(nvroot, B_FALSE)) == NULL)
 836                         return (0);
 837         }
 838         /*
 839          * for spares there may be no children, and therefore no
 840          * replication level to check
 841          */
 842         if ((nvlist_lookup_nvlist_array(newroot, ZPOOL_CONFIG_CHILDREN,
 843             &child, &children) != 0) || (children == 0)) {
 844                 free(current);
 845                 return (0);
 846         }
 847
 848         /*
 849          * If all we have is logs then there's no replication level to check.
 850          */
 851         if (num_logs(newroot) == children) {
 852                 free(current);
 853                 return (0);
 854         }
 855
 856         /*
 857          * Get the replication level of the new vdev spec, reporting any
 858          * inconsistencies found.
 859          */
 860         if ((new = get_replication(newroot, B_TRUE)) == NULL) {
 861                 free(current);
 862                 return (-1);
 863         }
 864
 865         /*
 866          * Check to see if the new vdev spec matches the replication level of
 867          * the current pool.
 868          */
 869         ret = 0;
 870         if (current != NULL) {
 871                 if (strcmp(current->zprl_type, new->zprl_type) != 0) {
 872                         vdev_error(gettext(
 873                             "mismatched replication level: pool uses %s "
 874                             "and new vdev is %s\n"),
 875                             current->zprl_type, new->zprl_type);
 876                         ret = -1;
 877                 } else if (current->zprl_parity != new->zprl_parity) {
 878                         vdev_error(gettext(
 879                             "mismatched replication level: pool uses %llu "
 880                             "device parity and new vdev uses %llu\n"),
 881                             current->zprl_parity, new->zprl_parity);
 882                         ret = -1;
 883                 } else if (current->zprl_children != new->zprl_children) {
 884                         vdev_error(gettext(
 885                             "mismatched replication level: pool uses %llu-way "
 886                             "%s and new vdev uses %llu-way %s\n"),
 887                             current->zprl_children, current->zprl_type,
 888                             new->zprl_children, new->zprl_type);
 889                         ret = -1;
 890                 }
 891         }
 892
 893         free(new);
 894         if (current != NULL)
 895                 free(current);
 896
 897         return (ret);
 898 }
 899
 900 static int
 901 zero_label(char *path)
 902 {
 903         const int size = 4096;
 904         char buf[size];
 905         int err, fd;
 906
 907         if ((fd = open(path, O_WRONLY|O_EXCL)) < 0) {
 908                 (void) fprintf(stderr, gettext("cannot open '%s': %s\n"),
 909                     path, strerror(errno));
 910                 return (-1);
 911         }
 912
 913         memset(buf, 0, size);
 914         err = write(fd, buf, size);
 915         (void) fdatasync(fd);
 916         (void) close(fd);
 917
 918         if (err == -1) {
 919                 (void) fprintf(stderr, gettext("cannot zero first %d bytes "
 920                     "of '%s': %s\n"), size, path, strerror(errno));
 921                 return (-1);
 922         }
 923
 924         if (err != size) {
 925                 (void) fprintf(stderr, gettext("could only zero %d/%d bytes "
 926                     "of '%s'\n"), err, size, path);
 927                 return (-1);
 928         }
 929
 930         return 0;
 931 }
 932
 933 /*
 934  * Go through and find any whole disks in the vdev specification, labelling them
 935  * as appropriate.  When constructing the vdev spec, we were unable to open this
 936  * device in order to provide a devid.  Now that we have labelled the disk and
 937  * know that slice 0 is valid, we can construct the devid now.
 938  *
 939  * If the disk was already labeled with an EFI label, we will have gotten the
 940  * devid already (because we were able to open the whole disk).  Otherwise, we
 941  * need to get the devid after we label the disk.
 942  */
 943 static int
 944 make_disks(zpool_handle_t *zhp, nvlist_t *nv)
 945 {
 946         nvlist_t **child;
 947         uint_t c, children;
 948         char *type, *path, *diskname;
 949         char buf[MAXPATHLEN];
 950         uint64_t wholedisk;
 951         int ret;
 952
 953         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
 954
 955         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
 956             &child, &children) != 0) {
 957
 958                 if (strcmp(type, VDEV_TYPE_DISK) != 0)
 959                         return (0);
 960
 961                 /*
 962                  * We have a disk device.  If this is a whole disk write
 963                  * out the efi partition table, otherwise write zero's to
 964                  * the first 4k of the partition.  This is to ensure that
 965                  * libblkid will not misidentify the partition due to a
 966                  * magic value left by the previous filesystem.
 967                  */
 968                 verify(!nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path));
 969                 verify(!nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
 970                     &wholedisk));
 971
 972                 if (!wholedisk) {
 973                         ret = zero_label(path);
 974                         return (ret);
 975                 }
 976
 977                 if (realpath(path, buf) == NULL) {
 978                         ret = errno;
 979                         (void) fprintf(stderr,
 980                             gettext("cannot resolve path '%s'\n"), path);
 981                         return (ret);
 982                 }
 983
 984                 diskname = strrchr(buf, '/');
 985                 assert(diskname != NULL);
 986                 diskname++;
 987                 if (zpool_label_disk(g_zfs, zhp, diskname) == -1)
 988                         return (-1);
 989
 990                 /*
 991                  * Now the we've labeled the disk and the partitions have
 992                  * been created.  We still need to wait for udev to create
 993                  * the symlinks to those partitions.  If we are accessing
 994                  * the devices via a udev disk path, /dev/disk, then wait
 995                  * for *-part# to be created.  Otherwise just use the normal
 996                  * syntax for devices in /dev.
 997                  */
 998                 zfs_append_partition(path, buf, sizeof (buf));
 999
1000                 if ((ret = zpool_label_disk_wait(buf, 1000)) != 0) {
1001                         (void) fprintf(stderr,
1002                             gettext( "cannot resolve path '%s'\n"), buf);
1003                         return (-1);
1004                 }
1005
1006                 /*
1007                  * Update the path to refer to FIRST_SLICE.  The presence of
1008                  * the 'whole_disk' field indicates to the CLI that we should
1009                  * chop off the slice number when displaying the device in
1010                  * future output.
1011                  */
1012                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_PATH, buf) == 0);
1013
1014                 /* Just in case this partition already existed. */
1015                 (void) zero_label(buf);
1016
1017                 return (0);
1018         }
1019
1020         for (c = 0; c < children; c++)
1021                 if ((ret = make_disks(zhp, child[c])) != 0)
1022                         return (ret);
1023
1024         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_SPARES,
1025             &child, &children) == 0)
1026                 for (c = 0; c < children; c++)
1027                         if ((ret = make_disks(zhp, child[c])) != 0)
1028                                 return (ret);
1029
1030         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1031             &child, &children) == 0)
1032                 for (c = 0; c < children; c++)
1033                         if ((ret = make_disks(zhp, child[c])) != 0)
1034                                 return (ret);
1035
1036         return (0);
1037 }
1038
1039 /*
1040  * Determine if the given path is a hot spare within the given configuration.
1041  */
1042 static boolean_t
1043 is_spare(nvlist_t *config, const char *path)
1044 {
1045         int fd;
1046         pool_state_t state;
1047         char *name = NULL;
1048         nvlist_t *label;
1049         uint64_t guid, spareguid;
1050         nvlist_t *nvroot;
1051         nvlist_t **spares;
1052         uint_t i, nspares;
1053         boolean_t inuse;
1054
1055         if ((fd = open(path, O_RDONLY|O_EXCL)) < 0)
1056                 return (B_FALSE);
1057
1058         if (zpool_in_use(g_zfs, fd, &state, &name, &inuse) != 0 ||
1059             !inuse ||
1060             state != POOL_STATE_SPARE ||
1061             zpool_read_label(fd, &label) != 0) {
1062                 free(name);
1063                 (void) close(fd);
1064                 return (B_FALSE);
1065         }
1066         free(name);
1067         (void) close(fd);
1068
1069         verify(nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) == 0);
1070         nvlist_free(label);
1071
1072         verify(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
1073             &nvroot) == 0);
1074         if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1075             &spares, &nspares) == 0) {
1076                 for (i = 0; i < nspares; i++) {
1077                         verify(nvlist_lookup_uint64(spares[i],
1078                             ZPOOL_CONFIG_GUID, &spareguid) == 0);
1079                         if (spareguid == guid)
1080                                 return (B_TRUE);
1081                 }
1082         }
1083
1084         return (B_FALSE);
1085 }
1086
1087 /*
1088  * Go through and find any devices that are in use.  We rely on libdiskmgt for
1089  * the majority of this task.
1090  */
1091 static int
1092 check_in_use(nvlist_t *config, nvlist_t *nv, boolean_t force,
1093     boolean_t replacing, boolean_t isspare)
1094 {
1095         nvlist_t **child;
1096         uint_t c, children;
1097         char *type, *path;
1098         int ret = 0;
1099         char buf[MAXPATHLEN];
1100         uint64_t wholedisk = B_FALSE;
1101
1102         verify(nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) == 0);
1103
1104         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
1105             &child, &children) != 0) {
1106
1107                 verify(!nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &path));
1108                 if (strcmp(type, VDEV_TYPE_DISK) == 0)
1109                         verify(!nvlist_lookup_uint64(nv,
1110                                ZPOOL_CONFIG_WHOLE_DISK, &wholedisk));
1111
1112                 /*
1113                  * As a generic check, we look to see if this is a replace of a
1114                  * hot spare within the same pool.  If so, we allow it
1115                  * regardless of what libblkid or zpool_in_use() says.
1116                  */
1117                 if (replacing) {
1118                         if (wholedisk)
1119                                 (void) snprintf(buf, sizeof (buf), "%ss0",
1120                                     path);
1121                         else
1122                                 (void) strlcpy(buf, path, sizeof (buf));
1123
1124                         if (is_spare(config, buf))
1125                                 return (0);
1126                 }
1127
1128                 if (strcmp(type, VDEV_TYPE_DISK) == 0)
1129                         ret = check_device(path, force, isspare, wholedisk);
1130
1131                 if (strcmp(type, VDEV_TYPE_FILE) == 0)
1132                         ret = check_file(path, force, isspare);
1133
1134                 return (ret);
1135         }
1136
1137         for (c = 0; c < children; c++)
1138                 if ((ret = check_in_use(config, child[c], force,
1139                     replacing, B_FALSE)) != 0)
1140                         return (ret);
1141
1142         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_SPARES,
1143             &child, &children) == 0)
1144                 for (c = 0; c < children; c++)
1145                         if ((ret = check_in_use(config, child[c], force,
1146                             replacing, B_TRUE)) != 0)
1147                                 return (ret);
1148
1149         if (nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_L2CACHE,
1150             &child, &children) == 0)
1151                 for (c = 0; c < children; c++)
1152                         if ((ret = check_in_use(config, child[c], force,
1153                             replacing, B_FALSE)) != 0)
1154                                 return (ret);
1155
1156         return (0);
1157 }
1158
1159 static const char *
1160 is_grouping(const char *type, int *mindev, int *maxdev)
1161 {
1162         if (strncmp(type, "raidz", 5) == 0) {
1163                 const char *p = type + 5;
1164                 char *end;
1165                 long nparity;
1166
1167                 if (*p == '\0') {
1168                         nparity = 1;
1169                 } else if (*p == '0') {
1170                         return (NULL); /* no zero prefixes allowed */
1171                 } else {
1172                         errno = 0;
1173                         nparity = strtol(p, &end, 10);
1174                         if (errno != 0 || nparity < 1 || nparity >= 255 ||
1175                             *end != '\0')
1176                                 return (NULL);
1177                 }
1178
1179                 if (mindev != NULL)
1180                         *mindev = nparity + 1;
1181                 if (maxdev != NULL)
1182                         *maxdev = 255;
1183                 return (VDEV_TYPE_RAIDZ);
1184         }
1185
1186         if (maxdev != NULL)
1187                 *maxdev = INT_MAX;
1188
1189         if (strcmp(type, "mirror") == 0) {
1190                 if (mindev != NULL)
1191                         *mindev = 2;
1192                 return (VDEV_TYPE_MIRROR);
1193         }
1194
1195         if (strcmp(type, "spare") == 0) {
1196                 if (mindev != NULL)
1197                         *mindev = 1;
1198                 return (VDEV_TYPE_SPARE);
1199         }
1200
1201         if (strcmp(type, "log") == 0) {
1202                 if (mindev != NULL)
1203                         *mindev = 1;
1204                 return (VDEV_TYPE_LOG);
1205         }
1206
1207         if (strcmp(type, "cache") == 0) {
1208                 if (mindev != NULL)
1209                         *mindev = 1;
1210                 return (VDEV_TYPE_L2CACHE);
1211         }
1212
1213         return (NULL);
1214 }
1215
1216 /*
1217  * Construct a syntactically valid vdev specification,
1218  * and ensure that all devices and files exist and can be opened.
1219  * Note: we don't bother freeing anything in the error paths
1220  * because the program is just going to exit anyway.
1221  */
1222 nvlist_t *
1223 construct_spec(int argc, char **argv)
1224 {
1225         nvlist_t *nvroot, *nv, **top, **spares, **l2cache;
1226         int t, toplevels, mindev, maxdev, nspares, nlogs, nl2cache;
1227         const char *type;
1228         uint64_t is_log;
1229         boolean_t seen_logs;
1230
1231         top = NULL;
1232         toplevels = 0;
1233         spares = NULL;
1234         l2cache = NULL;
1235         nspares = 0;
1236         nlogs = 0;
1237         nl2cache = 0;
1238         is_log = B_FALSE;
1239         seen_logs = B_FALSE;
1240
1241         while (argc > 0) {
1242                 nv = NULL;
1243
1244                 /*
1245                  * If it's a mirror or raidz, the subsequent arguments are
1246                  * its leaves -- until we encounter the next mirror or raidz.
1247                  */
1248                 if ((type = is_grouping(argv[0], &mindev, &maxdev)) != NULL) {
1249                         nvlist_t **child = NULL;
1250                         int c, children = 0;
1251
1252                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1253                                 if (spares != NULL) {
1254                                         (void) fprintf(stderr,
1255                                             gettext("invalid vdev "
1256                                             "specification: 'spare' can be "
1257                                             "specified only once\n"));
1258                                         return (NULL);
1259                                 }
1260                                 is_log = B_FALSE;
1261                         }
1262
1263                         if (strcmp(type, VDEV_TYPE_LOG) == 0) {
1264                                 if (seen_logs) {
1265                                         (void) fprintf(stderr,
1266                                             gettext("invalid vdev "
1267                                             "specification: 'log' can be "
1268                                             "specified only once\n"));
1269                                         return (NULL);
1270                                 }
1271                                 seen_logs = B_TRUE;
1272                                 is_log = B_TRUE;
1273                                 argc--;
1274                                 argv++;
1275                                 /*
1276                                  * A log is not a real grouping device.
1277                                  * We just set is_log and continue.
1278                                  */
1279                                 continue;
1280                         }
1281
1282                         if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1283                                 if (l2cache != NULL) {
1284                                         (void) fprintf(stderr,
1285                                             gettext("invalid vdev "
1286                                             "specification: 'cache' can be "
1287                                             "specified only once\n"));
1288                                         return (NULL);
1289                                 }
1290                                 is_log = B_FALSE;
1291                         }
1292
1293                         if (is_log) {
1294                                 if (strcmp(type, VDEV_TYPE_MIRROR) != 0) {
1295                                         (void) fprintf(stderr,
1296                                             gettext("invalid vdev "
1297                                             "specification: unsupported 'log' "
1298                                             "device: %s\n"), type);
1299                                         return (NULL);
1300                                 }
1301                                 nlogs++;
1302                         }
1303
1304                         for (c = 1; c < argc; c++) {
1305                                 if (is_grouping(argv[c], NULL, NULL) != NULL)
1306                                         break;
1307                                 children++;
1308                                 child = realloc(child,
1309                                     children * sizeof (nvlist_t *));
1310                                 if (child == NULL)
1311                                         zpool_no_memory();
1312                                 if ((nv = make_leaf_vdev(argv[c], B_FALSE))
1313                                     == NULL)
1314                                         return (NULL);
1315                                 child[children - 1] = nv;
1316                         }
1317
1318                         if (children < mindev) {
1319                                 (void) fprintf(stderr, gettext("invalid vdev "
1320                                     "specification: %s requires at least %d "
1321                                     "devices\n"), argv[0], mindev);
1322                                 return (NULL);
1323                         }
1324
1325                         if (children > maxdev) {
1326                                 (void) fprintf(stderr, gettext("invalid vdev "
1327                                     "specification: %s supports no more than "
1328                                     "%d devices\n"), argv[0], maxdev);
1329                                 return (NULL);
1330                         }
1331
1332                         argc -= c;
1333                         argv += c;
1334
1335                         if (strcmp(type, VDEV_TYPE_SPARE) == 0) {
1336                                 spares = child;
1337                                 nspares = children;
1338                                 continue;
1339                         } else if (strcmp(type, VDEV_TYPE_L2CACHE) == 0) {
1340                                 l2cache = child;
1341                                 nl2cache = children;
1342                                 continue;
1343                         } else {
1344                                 verify(nvlist_alloc(&nv, NV_UNIQUE_NAME,
1345                                     0) == 0);
1346                                 verify(nvlist_add_string(nv, ZPOOL_CONFIG_TYPE,
1347                                     type) == 0);
1348                                 verify(nvlist_add_uint64(nv,
1349                                     ZPOOL_CONFIG_IS_LOG, is_log) == 0);
1350                                 if (strcmp(type, VDEV_TYPE_RAIDZ) == 0) {
1351                                         verify(nvlist_add_uint64(nv,
1352                                             ZPOOL_CONFIG_NPARITY,
1353                                             mindev - 1) == 0);
1354                                 }
1355                                 verify(nvlist_add_nvlist_array(nv,
1356                                     ZPOOL_CONFIG_CHILDREN, child,
1357                                     children) == 0);
1358
1359                                 for (c = 0; c < children; c++)
1360                                         nvlist_free(child[c]);
1361                                 free(child);
1362                         }
1363                 } else {
1364                         /*
1365                          * We have a device.  Pass off to make_leaf_vdev() to
1366                          * construct the appropriate nvlist describing the vdev.
1367                          */
1368                         if ((nv = make_leaf_vdev(argv[0], is_log)) == NULL)
1369                                 return (NULL);
1370                         if (is_log)
1371                                 nlogs++;
1372                         argc--;
1373                         argv++;
1374                 }
1375
1376                 toplevels++;
1377                 top = realloc(top, toplevels * sizeof (nvlist_t *));
1378                 if (top == NULL)
1379                         zpool_no_memory();
1380                 top[toplevels - 1] = nv;
1381         }
1382
1383         if (toplevels == 0 && nspares == 0 && nl2cache == 0) {
1384                 (void) fprintf(stderr, gettext("invalid vdev "
1385                     "specification: at least one toplevel vdev must be "
1386                     "specified\n"));
1387                 return (NULL);
1388         }
1389
1390         if (seen_logs && nlogs == 0) {
1391                 (void) fprintf(stderr, gettext("invalid vdev specification: "
1392                     "log requires at least 1 device\n"));
1393                 return (NULL);
1394         }
1395
1396         /*
1397          * Finally, create nvroot and add all top-level vdevs to it.
1398          */
1399         verify(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, 0) == 0);
1400         verify(nvlist_add_string(nvroot, ZPOOL_CONFIG_TYPE,
1401             VDEV_TYPE_ROOT) == 0);
1402         verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
1403             top, toplevels) == 0);
1404         if (nspares != 0)
1405                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
1406                     spares, nspares) == 0);
1407         if (nl2cache != 0)
1408                 verify(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
1409                     l2cache, nl2cache) == 0);
1410
1411         for (t = 0; t < toplevels; t++)
1412                 nvlist_free(top[t]);
1413         for (t = 0; t < nspares; t++)
1414                 nvlist_free(spares[t]);
1415         for (t = 0; t < nl2cache; t++)
1416                 nvlist_free(l2cache[t]);
1417         if (spares)
1418                 free(spares);
1419         if (l2cache)
1420                 free(l2cache);
1421         free(top);
1422
1423         return (nvroot);
1424 }
1425
1426 nvlist_t *
1427 split_mirror_vdev(zpool_handle_t *zhp, char *newname, nvlist_t *props,
1428     splitflags_t flags, int argc, char **argv)
1429 {
1430         nvlist_t *newroot = NULL, **child;
1431         uint_t c, children;
1432
1433         if (argc > 0) {
1434                 if ((newroot = construct_spec(argc, argv)) == NULL) {
1435                         (void) fprintf(stderr, gettext("Unable to build a "
1436                             "pool from the specified devices\n"));
1437                         return (NULL);
1438                 }
1439
1440                 if (!flags.dryrun && make_disks(zhp, newroot) != 0) {
1441                         nvlist_free(newroot);
1442                         return (NULL);
1443                 }
1444
1445                 /* avoid any tricks in the spec */
1446                 verify(nvlist_lookup_nvlist_array(newroot,
1447                     ZPOOL_CONFIG_CHILDREN, &child, &children) == 0);
1448                 for (c = 0; c < children; c++) {
1449                         char *path;
1450                         const char *type;
1451                         int min, max;
1452
1453                         verify(nvlist_lookup_string(child[c],
1454                             ZPOOL_CONFIG_PATH, &path) == 0);
1455                         if ((type = is_grouping(path, &min, &max)) != NULL) {
1456                                 (void) fprintf(stderr, gettext("Cannot use "
1457                                     "'%s' as a device for splitting\n"), type);
1458                                 nvlist_free(newroot);
1459                                 return (NULL);
1460                         }
1461                 }
1462         }
1463
1464         if (zpool_vdev_split(zhp, newname, &newroot, props, flags) != 0) {
1465                 if (newroot != NULL)
1466                         nvlist_free(newroot);
1467                 return (NULL);
1468         }
1469
1470         return (newroot);
1471 }
1472
1473 /*
1474  * Get and validate the contents of the given vdev specification.  This ensures
1475  * that the nvlist returned is well-formed, that all the devices exist, and that
1476  * they are not currently in use by any other known consumer.  The 'poolconfig'
1477  * parameter is the current configuration of the pool when adding devices
1478  * existing pool, and is used to perform additional checks, such as changing the
1479  * replication level of the pool.  It can be 'NULL' to indicate that this is a
1480  * new pool.  The 'force' flag controls whether devices should be forcefully
1481  * added, even if they appear in use.
1482  */
1483 nvlist_t *
1484 make_root_vdev(zpool_handle_t *zhp, int force, int check_rep,
1485     boolean_t replacing, boolean_t dryrun, int argc, char **argv)
1486 {
1487         nvlist_t *newroot;
1488         nvlist_t *poolconfig = NULL;
1489         is_force = force;
1490
1491         /*
1492          * Construct the vdev specification.  If this is successful, we know
1493          * that we have a valid specification, and that all devices can be
1494          * opened.
1495          */
1496         if ((newroot = construct_spec(argc, argv)) == NULL)
1497                 return (NULL);
1498
1499         if (zhp && ((poolconfig = zpool_get_config(zhp, NULL)) == NULL))
1500                 return (NULL);
1501
1502         /*
1503          * Validate each device to make sure that its not shared with another
1504          * subsystem.  We do this even if 'force' is set, because there are some
1505          * uses (such as a dedicated dump device) that even '-f' cannot
1506          * override.
1507          */
1508         if (check_in_use(poolconfig, newroot, force, replacing, B_FALSE) != 0) {
1509                 nvlist_free(newroot);
1510                 return (NULL);
1511         }
1512
1513         /*
1514          * Check the replication level of the given vdevs and report any errors
1515          * found.  We include the existing pool spec, if any, as we need to
1516          * catch changes against the existing replication level.
1517          */
1518         if (check_rep && check_replication(poolconfig, newroot) != 0) {
1519                 nvlist_free(newroot);
1520                 return (NULL);
1521         }
1522
1523         /*
1524          * Run through the vdev specification and label any whole disks found.
1525          */
1526         if (!dryrun && make_disks(zhp, newroot) != 0) {
1527                 nvlist_free(newroot);
1528                 return (NULL);
1529         }
1530
1531         return (newroot);
1532 }