lib/libefi/rdwr_efi.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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Copyright 2012 Nexenta Systems, Inc.  All rights reserved.
  25  */
  26
  27 #include <stdio.h>
  28 #include <stdlib.h>
  29 #include <errno.h>
  30 #include <string.h>
  31 #include <strings.h>
  32 #include <unistd.h>
  33 #include <uuid/uuid.h>
  34 #include <zlib.h>
  35 #include <libintl.h>
  36 #include <sys/types.h>
  37 #include <sys/dkio.h>
  38 #include <sys/vtoc.h>
  39 #include <sys/mhd.h>
  40 #include <sys/param.h>
  41 #include <sys/dktp/fdisk.h>
  42 #include <sys/efi_partition.h>
  43 #include <sys/byteorder.h>
  44 #if defined(__linux__)
  45 #include <linux/fs.h>
  46 #endif
  47
  48 static struct uuid_to_ptag {
  49         struct uuid     uuid;
  50 } conversion_array[] = {
  51         { EFI_UNUSED },
  52         { EFI_BOOT },
  53         { EFI_ROOT },
  54         { EFI_SWAP },
  55         { EFI_USR },
  56         { EFI_BACKUP },
  57         { EFI_UNUSED },         /* STAND is never used */
  58         { EFI_VAR },
  59         { EFI_HOME },
  60         { EFI_ALTSCTR },
  61         { EFI_UNUSED },         /* CACHE (cachefs) is never used */
  62         { EFI_RESERVED },
  63         { EFI_SYSTEM },
  64         { EFI_LEGACY_MBR },
  65         { EFI_SYMC_PUB },
  66         { EFI_SYMC_CDS },
  67         { EFI_MSFT_RESV },
  68         { EFI_DELL_BASIC },
  69         { EFI_DELL_RAID },
  70         { EFI_DELL_SWAP },
  71         { EFI_DELL_LVM },
  72         { EFI_DELL_RESV },
  73         { EFI_AAPL_HFS },
  74         { EFI_AAPL_UFS },
  75         { EFI_FREEBSD_BOOT },
  76         { EFI_FREEBSD_SWAP },
  77         { EFI_FREEBSD_UFS },
  78         { EFI_FREEBSD_VINUM },
  79         { EFI_FREEBSD_ZFS },
  80         { EFI_BIOS_BOOT },
  81         { EFI_INTC_RS },
  82         { EFI_SNE_BOOT },
  83         { EFI_LENOVO_BOOT },
  84         { EFI_MSFT_LDMM },
  85         { EFI_MSFT_LDMD },
  86         { EFI_MSFT_RE },
  87         { EFI_IBM_GPFS },
  88         { EFI_MSFT_STORAGESPACES },
  89         { EFI_HPQ_DATA },
  90         { EFI_HPQ_SVC },
  91         { EFI_RHT_DATA },
  92         { EFI_RHT_HOME },
  93         { EFI_RHT_SRV },
  94         { EFI_RHT_DMCRYPT },
  95         { EFI_RHT_LUKS },
  96         { EFI_FREEBSD_DISKLABEL },
  97         { EFI_AAPL_RAID },
  98         { EFI_AAPL_RAIDOFFLINE },
  99         { EFI_AAPL_BOOT },
 100         { EFI_AAPL_LABEL },
 101         { EFI_AAPL_TVRECOVERY },
 102         { EFI_AAPL_CORESTORAGE },
 103         { EFI_NETBSD_SWAP },
 104         { EFI_NETBSD_FFS },
 105         { EFI_NETBSD_LFS },
 106         { EFI_NETBSD_RAID },
 107         { EFI_NETBSD_CAT },
 108         { EFI_NETBSD_CRYPT },
 109         { EFI_GOOG_KERN },
 110         { EFI_GOOG_ROOT },
 111         { EFI_GOOG_RESV },
 112         { EFI_HAIKU_BFS },
 113         { EFI_MIDNIGHTBSD_BOOT },
 114         { EFI_MIDNIGHTBSD_DATA },
 115         { EFI_MIDNIGHTBSD_SWAP },
 116         { EFI_MIDNIGHTBSD_UFS },
 117         { EFI_MIDNIGHTBSD_VINUM },
 118         { EFI_MIDNIGHTBSD_ZFS },
 119         { EFI_CEPH_JOURNAL },
 120         { EFI_CEPH_DMCRYPTJOURNAL },
 121         { EFI_CEPH_OSD },
 122         { EFI_CEPH_DMCRYPTOSD },
 123         { EFI_CEPH_CREATE },
 124         { EFI_CEPH_DMCRYPTCREATE },
 125         { EFI_OPENBSD_DISKLABEL },
 126         { EFI_BBRY_QNX },
 127         { EFI_BELL_PLAN9 },
 128         { EFI_VMW_KCORE },
 129         { EFI_VMW_VMFS },
 130         { EFI_VMW_RESV },
 131         { EFI_RHT_ROOTX86 },
 132         { EFI_RHT_ROOTAMD64 },
 133         { EFI_RHT_ROOTARM },
 134         { EFI_RHT_ROOTARM64 },
 135         { EFI_ACRONIS_SECUREZONE },
 136         { EFI_ONIE_BOOT },
 137         { EFI_ONIE_CONFIG },
 138         { EFI_IBM_PPRPBOOT },
 139         { EFI_FREEDESKTOP_BOOT }
 140 };
 141
 142 /*
 143  * Default vtoc information for non-SVr4 partitions
 144  */
 145 struct dk_map2  default_vtoc_map[NDKMAP] = {
 146         {       V_ROOT,         0       },              /* a - 0 */
 147         {       V_SWAP,         V_UNMNT },              /* b - 1 */
 148         {       V_BACKUP,       V_UNMNT },              /* c - 2 */
 149         {       V_UNASSIGNED,   0       },              /* d - 3 */
 150         {       V_UNASSIGNED,   0       },              /* e - 4 */
 151         {       V_UNASSIGNED,   0       },              /* f - 5 */
 152         {       V_USR,          0       },              /* g - 6 */
 153         {       V_UNASSIGNED,   0       },              /* h - 7 */
 154
 155 #if defined(_SUNOS_VTOC_16)
 156
 157 #if defined(i386) || defined(__amd64) || defined(__arm) || \
 158     defined(__powerpc) || defined(__sparc) || defined(__s390__) || \
 159     defined(__mips__)
 160         {       V_BOOT,         V_UNMNT },              /* i - 8 */
 161         {       V_ALTSCTR,      0       },              /* j - 9 */
 162
 163 #else
 164 #error No VTOC format defined.
 165 #endif                  /* defined(i386) */
 166
 167         {       V_UNASSIGNED,   0       },              /* k - 10 */
 168         {       V_UNASSIGNED,   0       },              /* l - 11 */
 169         {       V_UNASSIGNED,   0       },              /* m - 12 */
 170         {       V_UNASSIGNED,   0       },              /* n - 13 */
 171         {       V_UNASSIGNED,   0       },              /* o - 14 */
 172         {       V_UNASSIGNED,   0       },              /* p - 15 */
 173 #endif                  /* defined(_SUNOS_VTOC_16) */
 174 };
 175
 176 int efi_debug = 0;
 177
 178 static int efi_read(int, struct dk_gpt *);
 179
 180 /*
 181  * Return a 32-bit CRC of the contents of the buffer.  Pre-and-post
 182  * one's conditioning will be handled by crc32() internally.
 183  */
 184 static uint32_t
 185 efi_crc32(const unsigned char *buf, unsigned int size)
 186 {
 187         uint32_t crc = crc32(0, Z_NULL, 0);
 188
 189         crc = crc32(crc, buf, size);
 190
 191         return (crc);
 192 }
 193
 194 static int
 195 read_disk_info(int fd, diskaddr_t *capacity, uint_t *lbsize)
 196 {
 197         int sector_size;
 198         unsigned long long capacity_size;
 199
 200         if (ioctl(fd, BLKSSZGET, &sector_size) < 0)
 201                 return (-1);
 202
 203         if (ioctl(fd, BLKGETSIZE64, &capacity_size) < 0)
 204                 return (-1);
 205
 206         *lbsize = (uint_t)sector_size;
 207         *capacity = (diskaddr_t)(capacity_size / sector_size);
 208
 209         return (0);
 210 }
 211
 212 static int
 213 efi_get_info(int fd, struct dk_cinfo *dki_info)
 214 {
 215 #if defined(__linux__)
 216         char *path;
 217         char *dev_path;
 218         int rval = 0;
 219
 220         memset(dki_info, 0, sizeof (*dki_info));
 221
 222         path = calloc(1, PATH_MAX);
 223         if (path == NULL)
 224                 goto error;
 225
 226         /*
 227          * The simplest way to get the partition number under linux is
 228          * to parse it out of the /dev/<disk><parition> block device name.
 229          * The kernel creates this using the partition number when it
 230          * populates /dev/ so it may be trusted.  The tricky bit here is
 231          * that the naming convention is based on the block device type.
 232          * So we need to take this in to account when parsing out the
 233          * partition information.  Another issue is that the libefi API
 234          * API only provides the open fd and not the file path.  To handle
 235          * this realpath(3) is used to resolve the block device name from
 236          * /proc/self/fd/<fd>.  Aside from the partition number we collect
 237          * some additional device info.
 238          */
 239         (void) sprintf(path, "/proc/self/fd/%d", fd);
 240         dev_path = realpath(path, NULL);
 241         free(path);
 242
 243         if (dev_path == NULL)
 244                 goto error;
 245
 246         if ((strncmp(dev_path, "/dev/sd", 7) == 0)) {
 247                 strcpy(dki_info->dki_cname, "sd");
 248                 dki_info->dki_ctype = DKC_SCSI_CCS;
 249                 rval = sscanf(dev_path, "/dev/%[a-zA-Z]%hu",
 250                     dki_info->dki_dname,
 251                     &dki_info->dki_partition);
 252         } else if ((strncmp(dev_path, "/dev/hd", 7) == 0)) {
 253                 strcpy(dki_info->dki_cname, "hd");
 254                 dki_info->dki_ctype = DKC_DIRECT;
 255                 rval = sscanf(dev_path, "/dev/%[a-zA-Z]%hu",
 256                     dki_info->dki_dname,
 257                     &dki_info->dki_partition);
 258         } else if ((strncmp(dev_path, "/dev/md", 7) == 0)) {
 259                 strcpy(dki_info->dki_cname, "pseudo");
 260                 dki_info->dki_ctype = DKC_MD;
 261                 strcpy(dki_info->dki_dname, "md");
 262                 rval = sscanf(dev_path, "/dev/md%[0-9]p%hu",
 263                     dki_info->dki_dname + 2,
 264                     &dki_info->dki_partition);
 265         } else if ((strncmp(dev_path, "/dev/vd", 7) == 0)) {
 266                 strcpy(dki_info->dki_cname, "vd");
 267                 dki_info->dki_ctype = DKC_MD;
 268                 rval = sscanf(dev_path, "/dev/%[a-zA-Z]%hu",
 269                     dki_info->dki_dname,
 270                     &dki_info->dki_partition);
 271         } else if ((strncmp(dev_path, "/dev/xvd", 8) == 0)) {
 272                 strcpy(dki_info->dki_cname, "xvd");
 273                 dki_info->dki_ctype = DKC_MD;
 274                 rval = sscanf(dev_path, "/dev/%[a-zA-Z]%hu",
 275                     dki_info->dki_dname,
 276                     &dki_info->dki_partition);
 277         } else if ((strncmp(dev_path, "/dev/zd", 7) == 0)) {
 278                 strcpy(dki_info->dki_cname, "zd");
 279                 dki_info->dki_ctype = DKC_MD;
 280                 rval = sscanf(dev_path, "/dev/%[a-zA-Z]%hu",
 281                     dki_info->dki_dname,
 282                     &dki_info->dki_partition);
 283         } else if ((strncmp(dev_path, "/dev/dm-", 8) == 0)) {
 284                 strcpy(dki_info->dki_cname, "pseudo");
 285                 dki_info->dki_ctype = DKC_VBD;
 286                 strcpy(dki_info->dki_dname, "dm-");
 287                 rval = sscanf(dev_path, "/dev/dm-%[0-9]p%hu",
 288                     dki_info->dki_dname + 3,
 289                     &dki_info->dki_partition);
 290         } else if ((strncmp(dev_path, "/dev/ram", 8) == 0)) {
 291                 strcpy(dki_info->dki_cname, "pseudo");
 292                 dki_info->dki_ctype = DKC_PCMCIA_MEM;
 293                 strcpy(dki_info->dki_dname, "ram");
 294                 rval = sscanf(dev_path, "/dev/ram%[0-9]p%hu",
 295                     dki_info->dki_dname + 3,
 296                     &dki_info->dki_partition);
 297         } else if ((strncmp(dev_path, "/dev/loop", 9) == 0)) {
 298                 strcpy(dki_info->dki_cname, "pseudo");
 299                 dki_info->dki_ctype = DKC_VBD;
 300                 strcpy(dki_info->dki_dname, "loop");
 301                 rval = sscanf(dev_path, "/dev/loop%[0-9]p%hu",
 302                     dki_info->dki_dname + 4,
 303                     &dki_info->dki_partition);
 304         } else if ((strncmp(dev_path, "/dev/nvme", 9) == 0)) {
 305                 strcpy(dki_info->dki_cname, "nvme");
 306                 dki_info->dki_ctype = DKC_SCSI_CCS;
 307                 strcpy(dki_info->dki_dname, "nvme");
 308                 (void) sscanf(dev_path, "/dev/nvme%[0-9]",
 309                     dki_info->dki_dname + 4);
 310                 size_t controller_length = strlen(
 311                     dki_info->dki_dname);
 312                 strcpy(dki_info->dki_dname + controller_length,
 313                     "n");
 314                 rval = sscanf(dev_path,
 315                     "/dev/nvme%*[0-9]n%[0-9]p%hu",
 316                     dki_info->dki_dname + controller_length + 1,
 317                     &dki_info->dki_partition);
 318         } else {
 319                 strcpy(dki_info->dki_dname, "unknown");
 320                 strcpy(dki_info->dki_cname, "unknown");
 321                 dki_info->dki_ctype = DKC_UNKNOWN;
 322         }
 323
 324         switch (rval) {
 325         case 0:
 326                 errno = EINVAL;
 327                 goto error;
 328         case 1:
 329                 dki_info->dki_partition = 0;
 330         }
 331
 332         free(dev_path);
 333 #else
 334         if (ioctl(fd, DKIOCINFO, (caddr_t)dki_info) == -1)
 335                 goto error;
 336 #endif
 337         return (0);
 338 error:
 339         if (efi_debug)
 340                 (void) fprintf(stderr, "DKIOCINFO errno 0x%x\n", errno);
 341
 342         switch (errno) {
 343         case EIO:
 344                 return (VT_EIO);
 345         case EINVAL:
 346                 return (VT_EINVAL);
 347         default:
 348                 return (VT_ERROR);
 349         }
 350 }
 351
 352 /*
 353  * the number of blocks the EFI label takes up (round up to nearest
 354  * block)
 355  */
 356 #define NBLOCKS(p, l)   (1 + ((((p) * (int)sizeof (efi_gpe_t))  + \
 357                                 ((l) - 1)) / (l)))
 358 /* number of partitions -- limited by what we can malloc */
 359 #define MAX_PARTS       ((4294967295UL - sizeof (struct dk_gpt)) / \
 360                             sizeof (struct dk_part))
 361
 362 int
 363 efi_alloc_and_init(int fd, uint32_t nparts, struct dk_gpt **vtoc)
 364 {
 365         diskaddr_t      capacity = 0;
 366         uint_t          lbsize = 0;
 367         uint_t          nblocks;
 368         size_t          length;
 369         struct dk_gpt   *vptr;
 370         struct uuid     uuid;
 371         struct dk_cinfo dki_info;
 372
 373         if (read_disk_info(fd, &capacity, &lbsize) != 0)
 374                 return (-1);
 375
 376 #if defined(__linux__)
 377         if (efi_get_info(fd, &dki_info) != 0)
 378                 return (-1);
 379
 380         if (dki_info.dki_partition != 0)
 381                 return (-1);
 382
 383         if ((dki_info.dki_ctype == DKC_PCMCIA_MEM) ||
 384             (dki_info.dki_ctype == DKC_VBD) ||
 385             (dki_info.dki_ctype == DKC_UNKNOWN))
 386                 return (-1);
 387 #endif
 388
 389         nblocks = NBLOCKS(nparts, lbsize);
 390         if ((nblocks * lbsize) < EFI_MIN_ARRAY_SIZE + lbsize) {
 391                 /* 16K plus one block for the GPT */
 392                 nblocks = EFI_MIN_ARRAY_SIZE / lbsize + 1;
 393         }
 394
 395         if (nparts > MAX_PARTS) {
 396                 if (efi_debug) {
 397                         (void) fprintf(stderr,
 398                         "the maximum number of partitions supported is %lu\n",
 399                             MAX_PARTS);
 400                 }
 401                 return (-1);
 402         }
 403
 404         length = sizeof (struct dk_gpt) +
 405             sizeof (struct dk_part) * (nparts - 1);
 406
 407         if ((*vtoc = calloc(1, length)) == NULL)
 408                 return (-1);
 409
 410         vptr = *vtoc;
 411
 412         vptr->efi_version = EFI_VERSION_CURRENT;
 413         vptr->efi_lbasize = lbsize;
 414         vptr->efi_nparts = nparts;
 415         /*
 416          * add one block here for the PMBR; on disks with a 512 byte
 417          * block size and 128 or fewer partitions, efi_first_u_lba
 418          * should work out to "34"
 419          */
 420         vptr->efi_first_u_lba = nblocks + 1;
 421         vptr->efi_last_lba = capacity - 1;
 422         vptr->efi_altern_lba = capacity -1;
 423         vptr->efi_last_u_lba = vptr->efi_last_lba - nblocks;
 424
 425         (void) uuid_generate((uchar_t *)&uuid);
 426         UUID_LE_CONVERT(vptr->efi_disk_uguid, uuid);
 427         return (0);
 428 }
 429
 430 /*
 431  * Read EFI - return partition number upon success.
 432  */
 433 int
 434 efi_alloc_and_read(int fd, struct dk_gpt **vtoc)
 435 {
 436         int                     rval;
 437         uint32_t                nparts;
 438         int                     length;
 439
 440         /* figure out the number of entries that would fit into 16K */
 441         nparts = EFI_MIN_ARRAY_SIZE / sizeof (efi_gpe_t);
 442         length = (int) sizeof (struct dk_gpt) +
 443             (int) sizeof (struct dk_part) * (nparts - 1);
 444         if ((*vtoc = calloc(1, length)) == NULL)
 445                 return (VT_ERROR);
 446
 447         (*vtoc)->efi_nparts = nparts;
 448         rval = efi_read(fd, *vtoc);
 449
 450         if ((rval == VT_EINVAL) && (*vtoc)->efi_nparts > nparts) {
 451                 void *tmp;
 452                 length = (int) sizeof (struct dk_gpt) +
 453                     (int) sizeof (struct dk_part) *
 454                     ((*vtoc)->efi_nparts - 1);
 455                 nparts = (*vtoc)->efi_nparts;
 456                 if ((tmp = realloc(*vtoc, length)) == NULL) {
 457                         free (*vtoc);
 458                         *vtoc = NULL;
 459                         return (VT_ERROR);
 460                 } else {
 461                         *vtoc = tmp;
 462                         rval = efi_read(fd, *vtoc);
 463                 }
 464         }
 465
 466         if (rval < 0) {
 467                 if (efi_debug) {
 468                         (void) fprintf(stderr,
 469                             "read of EFI table failed, rval=%d\n", rval);
 470                 }
 471                 free (*vtoc);
 472                 *vtoc = NULL;
 473         }
 474
 475         return (rval);
 476 }
 477
 478 static int
 479 efi_ioctl(int fd, int cmd, dk_efi_t *dk_ioc)
 480 {
 481         void *data = dk_ioc->dki_data;
 482         int error;
 483 #if defined(__linux__)
 484         diskaddr_t capacity;
 485         uint_t lbsize;
 486
 487         /*
 488          * When the IO is not being performed in kernel as an ioctl we need
 489          * to know the sector size so we can seek to the proper byte offset.
 490          */
 491         if (read_disk_info(fd, &capacity, &lbsize) == -1) {
 492                 if (efi_debug)
 493                         fprintf(stderr, "unable to read disk info: %d", errno);
 494
 495                 errno = EIO;
 496                 return (-1);
 497         }
 498
 499         switch (cmd) {
 500         case DKIOCGETEFI:
 501                 if (lbsize == 0) {
 502                         if (efi_debug)
 503                                 (void) fprintf(stderr, "DKIOCGETEFI assuming "
 504                                     "LBA %d bytes\n", DEV_BSIZE);
 505
 506                         lbsize = DEV_BSIZE;
 507                 }
 508
 509                 error = lseek(fd, dk_ioc->dki_lba * lbsize, SEEK_SET);
 510                 if (error == -1) {
 511                         if (efi_debug)
 512                                 (void) fprintf(stderr, "DKIOCGETEFI lseek "
 513                                     "error: %d\n", errno);
 514                         return (error);
 515                 }
 516
 517                 error = read(fd, data, dk_ioc->dki_length);
 518                 if (error == -1) {
 519                         if (efi_debug)
 520                                 (void) fprintf(stderr, "DKIOCGETEFI read "
 521                                     "error: %d\n", errno);
 522                         return (error);
 523                 }
 524
 525                 if (error != dk_ioc->dki_length) {
 526                         if (efi_debug)
 527                                 (void) fprintf(stderr, "DKIOCGETEFI short "
 528                                     "read of %d bytes\n", error);
 529                         errno = EIO;
 530                         return (-1);
 531                 }
 532                 error = 0;
 533                 break;
 534
 535         case DKIOCSETEFI:
 536                 if (lbsize == 0) {
 537                         if (efi_debug)
 538                                 (void) fprintf(stderr, "DKIOCSETEFI unknown "
 539                                     "LBA size\n");
 540                         errno = EIO;
 541                         return (-1);
 542                 }
 543
 544                 error = lseek(fd, dk_ioc->dki_lba * lbsize, SEEK_SET);
 545                 if (error == -1) {
 546                         if (efi_debug)
 547                                 (void) fprintf(stderr, "DKIOCSETEFI lseek "
 548                                     "error: %d\n", errno);
 549                         return (error);
 550                 }
 551
 552                 error = write(fd, data, dk_ioc->dki_length);
 553                 if (error == -1) {
 554                         if (efi_debug)
 555                                 (void) fprintf(stderr, "DKIOCSETEFI write "
 556                                     "error: %d\n", errno);
 557                         return (error);
 558                 }
 559
 560                 if (error != dk_ioc->dki_length) {
 561                         if (efi_debug)
 562                                 (void) fprintf(stderr, "DKIOCSETEFI short "
 563                                     "write of %d bytes\n", error);
 564                         errno = EIO;
 565                         return (-1);
 566                 }
 567
 568                 /* Sync the new EFI table to disk */
 569                 error = fsync(fd);
 570                 if (error == -1)
 571                         return (error);
 572
 573                 /* Ensure any local disk cache is also flushed */
 574                 if (ioctl(fd, BLKFLSBUF, 0) == -1)
 575                         return (error);
 576
 577                 error = 0;
 578                 break;
 579
 580         default:
 581                 if (efi_debug)
 582                         (void) fprintf(stderr, "unsupported ioctl()\n");
 583
 584                 errno = EIO;
 585                 return (-1);
 586         }
 587 #else
 588         dk_ioc->dki_data_64 = (uint64_t)(uintptr_t)data;
 589         error = ioctl(fd, cmd, (void *)dk_ioc);
 590         dk_ioc->dki_data = data;
 591 #endif
 592         return (error);
 593 }
 594
 595 int
 596 efi_rescan(int fd)
 597 {
 598 #if defined(__linux__)
 599         int retry = 10;
 600         int error;
 601
 602         /* Notify the kernel a devices partition table has been updated */
 603         while ((error = ioctl(fd, BLKRRPART)) != 0) {
 604                 if ((--retry == 0) || (errno != EBUSY)) {
 605                         (void) fprintf(stderr, "the kernel failed to rescan "
 606                             "the partition table: %d\n", errno);
 607                         return (-1);
 608                 }
 609                 usleep(50000);
 610         }
 611 #endif
 612
 613         return (0);
 614 }
 615
 616 static int
 617 check_label(int fd, dk_efi_t *dk_ioc)
 618 {
 619         efi_gpt_t               *efi;
 620         uint_t                  crc;
 621
 622         if (efi_ioctl(fd, DKIOCGETEFI, dk_ioc) == -1) {
 623                 switch (errno) {
 624                 case EIO:
 625                         return (VT_EIO);
 626                 default:
 627                         return (VT_ERROR);
 628                 }
 629         }
 630         efi = dk_ioc->dki_data;
 631         if (efi->efi_gpt_Signature != LE_64(EFI_SIGNATURE)) {
 632                 if (efi_debug)
 633                         (void) fprintf(stderr,
 634                             "Bad EFI signature: 0x%llx != 0x%llx\n",
 635                             (long long)efi->efi_gpt_Signature,
 636                             (long long)LE_64(EFI_SIGNATURE));
 637                 return (VT_EINVAL);
 638         }
 639
 640         /*
 641          * check CRC of the header; the size of the header should
 642          * never be larger than one block
 643          */
 644         crc = efi->efi_gpt_HeaderCRC32;
 645         efi->efi_gpt_HeaderCRC32 = 0;
 646         len_t headerSize = (len_t)LE_32(efi->efi_gpt_HeaderSize);
 647
 648         if (headerSize < EFI_MIN_LABEL_SIZE || headerSize > EFI_LABEL_SIZE) {
 649                 if (efi_debug)
 650                         (void) fprintf(stderr,
 651                             "Invalid EFI HeaderSize %llu.  Assuming %d.\n",
 652                             headerSize, EFI_MIN_LABEL_SIZE);
 653         }
 654
 655         if ((headerSize > dk_ioc->dki_length) ||
 656             crc != LE_32(efi_crc32((unsigned char *)efi, headerSize))) {
 657                 if (efi_debug)
 658                         (void) fprintf(stderr,
 659                             "Bad EFI CRC: 0x%x != 0x%x\n",
 660                             crc, LE_32(efi_crc32((unsigned char *)efi,
 661                             headerSize)));
 662                 return (VT_EINVAL);
 663         }
 664
 665         return (0);
 666 }
 667
 668 static int
 669 efi_read(int fd, struct dk_gpt *vtoc)
 670 {
 671         int                     i, j;
 672         int                     label_len;
 673         int                     rval = 0;
 674         int                     md_flag = 0;
 675         int                     vdc_flag = 0;
 676         diskaddr_t              capacity = 0;
 677         uint_t                  lbsize = 0;
 678         struct dk_minfo         disk_info;
 679         dk_efi_t                dk_ioc;
 680         efi_gpt_t               *efi;
 681         efi_gpe_t               *efi_parts;
 682         struct dk_cinfo         dki_info;
 683         uint32_t                user_length;
 684         boolean_t               legacy_label = B_FALSE;
 685
 686         /*
 687          * get the partition number for this file descriptor.
 688          */
 689         if ((rval = efi_get_info(fd, &dki_info)) != 0)
 690                 return (rval);
 691
 692         if ((strncmp(dki_info.dki_cname, "pseudo", 7) == 0) &&
 693             (strncmp(dki_info.dki_dname, "md", 3) == 0)) {
 694                 md_flag++;
 695         } else if ((strncmp(dki_info.dki_cname, "vdc", 4) == 0) &&
 696             (strncmp(dki_info.dki_dname, "vdc", 4) == 0)) {
 697                 /*
 698                  * The controller and drive name "vdc" (virtual disk client)
 699                  * indicates a LDoms virtual disk.
 700                  */
 701                 vdc_flag++;
 702         }
 703
 704         /* get the LBA size */
 705         if (read_disk_info(fd, &capacity, &lbsize) == -1) {
 706                 if (efi_debug) {
 707                         (void) fprintf(stderr,
 708                             "unable to read disk info: %d",
 709                             errno);
 710                 }
 711                 return (VT_EINVAL);
 712         }
 713
 714         disk_info.dki_lbsize = lbsize;
 715         disk_info.dki_capacity = capacity;
 716
 717         if (disk_info.dki_lbsize == 0) {
 718                 if (efi_debug) {
 719                         (void) fprintf(stderr,
 720                             "efi_read: assuming LBA 512 bytes\n");
 721                 }
 722                 disk_info.dki_lbsize = DEV_BSIZE;
 723         }
 724         /*
 725          * Read the EFI GPT to figure out how many partitions we need
 726          * to deal with.
 727          */
 728         dk_ioc.dki_lba = 1;
 729         if (NBLOCKS(vtoc->efi_nparts, disk_info.dki_lbsize) < 34) {
 730                 label_len = EFI_MIN_ARRAY_SIZE + disk_info.dki_lbsize;
 731         } else {
 732                 label_len = vtoc->efi_nparts * (int) sizeof (efi_gpe_t) +
 733                     disk_info.dki_lbsize;
 734                 if (label_len % disk_info.dki_lbsize) {
 735                         /* pad to physical sector size */
 736                         label_len += disk_info.dki_lbsize;
 737                         label_len &= ~(disk_info.dki_lbsize - 1);
 738                 }
 739         }
 740
 741         if (posix_memalign((void **)&dk_ioc.dki_data,
 742             disk_info.dki_lbsize, label_len))
 743                 return (VT_ERROR);
 744
 745         memset(dk_ioc.dki_data, 0, label_len);
 746         dk_ioc.dki_length = disk_info.dki_lbsize;
 747         user_length = vtoc->efi_nparts;
 748         efi = dk_ioc.dki_data;
 749         if (md_flag) {
 750                 dk_ioc.dki_length = label_len;
 751                 if (efi_ioctl(fd, DKIOCGETEFI, &dk_ioc) == -1) {
 752                         switch (errno) {
 753                         case EIO:
 754                                 return (VT_EIO);
 755                         default:
 756                                 return (VT_ERROR);
 757                         }
 758                 }
 759         } else if ((rval = check_label(fd, &dk_ioc)) == VT_EINVAL) {
 760                 /*
 761                  * No valid label here; try the alternate. Note that here
 762                  * we just read GPT header and save it into dk_ioc.data,
 763                  * Later, we will read GUID partition entry array if we
 764                  * can get valid GPT header.
 765                  */
 766
 767                 /*
 768                  * This is a workaround for legacy systems. In the past, the
 769                  * last sector of SCSI disk was invisible on x86 platform. At
 770                  * that time, backup label was saved on the next to the last
 771                  * sector. It is possible for users to move a disk from previous
 772                  * solaris system to present system. Here, we attempt to search
 773                  * legacy backup EFI label first.
 774                  */
 775                 dk_ioc.dki_lba = disk_info.dki_capacity - 2;
 776                 dk_ioc.dki_length = disk_info.dki_lbsize;
 777                 rval = check_label(fd, &dk_ioc);
 778                 if (rval == VT_EINVAL) {
 779                         /*
 780                          * we didn't find legacy backup EFI label, try to
 781                          * search backup EFI label in the last block.
 782                          */
 783                         dk_ioc.dki_lba = disk_info.dki_capacity - 1;
 784                         dk_ioc.dki_length = disk_info.dki_lbsize;
 785                         rval = check_label(fd, &dk_ioc);
 786                         if (rval == 0) {
 787                                 legacy_label = B_TRUE;
 788                                 if (efi_debug)
 789                                         (void) fprintf(stderr,
 790                                             "efi_read: primary label corrupt; "
 791                                             "using EFI backup label located on"
 792                                             " the last block\n");
 793                         }
 794                 } else {
 795                         if ((efi_debug) && (rval == 0))
 796                                 (void) fprintf(stderr, "efi_read: primary label"
 797                                     " corrupt; using legacy EFI backup label "
 798                                     " located on the next to last block\n");
 799                 }
 800
 801                 if (rval == 0) {
 802                         dk_ioc.dki_lba = LE_64(efi->efi_gpt_PartitionEntryLBA);
 803                         vtoc->efi_flags |= EFI_GPT_PRIMARY_CORRUPT;
 804                         vtoc->efi_nparts =
 805                             LE_32(efi->efi_gpt_NumberOfPartitionEntries);
 806                         /*
 807                          * Partition tables are between backup GPT header
 808                          * table and ParitionEntryLBA (the starting LBA of
 809                          * the GUID partition entries array). Now that we
 810                          * already got valid GPT header and saved it in
 811                          * dk_ioc.dki_data, we try to get GUID partition
 812                          * entry array here.
 813                          */
 814                         /* LINTED */
 815                         dk_ioc.dki_data = (efi_gpt_t *)((char *)dk_ioc.dki_data
 816                             + disk_info.dki_lbsize);
 817                         if (legacy_label)
 818                                 dk_ioc.dki_length = disk_info.dki_capacity - 1 -
 819                                     dk_ioc.dki_lba;
 820                         else
 821                                 dk_ioc.dki_length = disk_info.dki_capacity - 2 -
 822                                     dk_ioc.dki_lba;
 823                         dk_ioc.dki_length *= disk_info.dki_lbsize;
 824                         if (dk_ioc.dki_length >
 825                             ((len_t)label_len - sizeof (*dk_ioc.dki_data))) {
 826                                 rval = VT_EINVAL;
 827                         } else {
 828                                 /*
 829                                  * read GUID partition entry array
 830                                  */
 831                                 rval = efi_ioctl(fd, DKIOCGETEFI, &dk_ioc);
 832                         }
 833                 }
 834
 835         } else if (rval == 0) {
 836
 837                 dk_ioc.dki_lba = LE_64(efi->efi_gpt_PartitionEntryLBA);
 838                 /* LINTED */
 839                 dk_ioc.dki_data = (efi_gpt_t *)((char *)dk_ioc.dki_data
 840                     + disk_info.dki_lbsize);
 841                 dk_ioc.dki_length = label_len - disk_info.dki_lbsize;
 842                 rval = efi_ioctl(fd, DKIOCGETEFI, &dk_ioc);
 843
 844         } else if (vdc_flag && rval == VT_ERROR && errno == EINVAL) {
 845                 /*
 846                  * When the device is a LDoms virtual disk, the DKIOCGETEFI
 847                  * ioctl can fail with EINVAL if the virtual disk backend
 848                  * is a ZFS volume serviced by a domain running an old version
 849                  * of Solaris. This is because the DKIOCGETEFI ioctl was
 850                  * initially incorrectly implemented for a ZFS volume and it
 851                  * expected the GPT and GPE to be retrieved with a single ioctl.
 852                  * So we try to read the GPT and the GPE using that old style
 853                  * ioctl.
 854                  */
 855                 dk_ioc.dki_lba = 1;
 856                 dk_ioc.dki_length = label_len;
 857                 rval = check_label(fd, &dk_ioc);
 858         }
 859
 860         if (rval < 0) {
 861                 free(efi);
 862                 return (rval);
 863         }
 864
 865         /* LINTED -- always longlong aligned */
 866         efi_parts = (efi_gpe_t *)(((char *)efi) + disk_info.dki_lbsize);
 867
 868         /*
 869          * Assemble this into a "dk_gpt" struct for easier
 870          * digestibility by applications.
 871          */
 872         vtoc->efi_version = LE_32(efi->efi_gpt_Revision);
 873         vtoc->efi_nparts = LE_32(efi->efi_gpt_NumberOfPartitionEntries);
 874         vtoc->efi_part_size = LE_32(efi->efi_gpt_SizeOfPartitionEntry);
 875         vtoc->efi_lbasize = disk_info.dki_lbsize;
 876         vtoc->efi_last_lba = disk_info.dki_capacity - 1;
 877         vtoc->efi_first_u_lba = LE_64(efi->efi_gpt_FirstUsableLBA);
 878         vtoc->efi_last_u_lba = LE_64(efi->efi_gpt_LastUsableLBA);
 879         vtoc->efi_altern_lba = LE_64(efi->efi_gpt_AlternateLBA);
 880         UUID_LE_CONVERT(vtoc->efi_disk_uguid, efi->efi_gpt_DiskGUID);
 881
 882         /*
 883          * If the array the user passed in is too small, set the length
 884          * to what it needs to be and return
 885          */
 886         if (user_length < vtoc->efi_nparts) {
 887                 return (VT_EINVAL);
 888         }
 889
 890         for (i = 0; i < vtoc->efi_nparts; i++) {
 891
 892                 UUID_LE_CONVERT(vtoc->efi_parts[i].p_guid,
 893                     efi_parts[i].efi_gpe_PartitionTypeGUID);
 894
 895                 for (j = 0;
 896                     j < sizeof (conversion_array)
 897                     / sizeof (struct uuid_to_ptag); j++) {
 898
 899                         if (bcmp(&vtoc->efi_parts[i].p_guid,
 900                             &conversion_array[j].uuid,
 901                             sizeof (struct uuid)) == 0) {
 902                                 vtoc->efi_parts[i].p_tag = j;
 903                                 break;
 904                         }
 905                 }
 906                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED)
 907                         continue;
 908                 vtoc->efi_parts[i].p_flag =
 909                     LE_16(efi_parts[i].efi_gpe_Attributes.PartitionAttrs);
 910                 vtoc->efi_parts[i].p_start =
 911                     LE_64(efi_parts[i].efi_gpe_StartingLBA);
 912                 vtoc->efi_parts[i].p_size =
 913                     LE_64(efi_parts[i].efi_gpe_EndingLBA) -
 914                     vtoc->efi_parts[i].p_start + 1;
 915                 for (j = 0; j < EFI_PART_NAME_LEN; j++) {
 916                         vtoc->efi_parts[i].p_name[j] =
 917                             (uchar_t)LE_16(
 918                             efi_parts[i].efi_gpe_PartitionName[j]);
 919                 }
 920
 921                 UUID_LE_CONVERT(vtoc->efi_parts[i].p_uguid,
 922                     efi_parts[i].efi_gpe_UniquePartitionGUID);
 923         }
 924         free(efi);
 925
 926         return (dki_info.dki_partition);
 927 }
 928
 929 /* writes a "protective" MBR */
 930 static int
 931 write_pmbr(int fd, struct dk_gpt *vtoc)
 932 {
 933         dk_efi_t        dk_ioc;
 934         struct mboot    mb;
 935         uchar_t         *cp;
 936         diskaddr_t      size_in_lba;
 937         uchar_t         *buf;
 938         int             len;
 939
 940         len = (vtoc->efi_lbasize == 0) ? sizeof (mb) : vtoc->efi_lbasize;
 941         if (posix_memalign((void **)&buf, len, len))
 942                 return (VT_ERROR);
 943
 944         /*
 945          * Preserve any boot code and disk signature if the first block is
 946          * already an MBR.
 947          */
 948         memset(buf, 0, len);
 949         dk_ioc.dki_lba = 0;
 950         dk_ioc.dki_length = len;
 951         /* LINTED -- always longlong aligned */
 952         dk_ioc.dki_data = (efi_gpt_t *)buf;
 953         if (efi_ioctl(fd, DKIOCGETEFI, &dk_ioc) == -1) {
 954                 (void) memcpy(&mb, buf, sizeof (mb));
 955                 bzero(&mb, sizeof (mb));
 956                 mb.signature = LE_16(MBB_MAGIC);
 957         } else {
 958                 (void) memcpy(&mb, buf, sizeof (mb));
 959                 if (mb.signature != LE_16(MBB_MAGIC)) {
 960                         bzero(&mb, sizeof (mb));
 961                         mb.signature = LE_16(MBB_MAGIC);
 962                 }
 963         }
 964
 965         bzero(&mb.parts, sizeof (mb.parts));
 966         cp = (uchar_t *)&mb.parts[0];
 967         /* bootable or not */
 968         *cp++ = 0;
 969         /* beginning CHS; 0xffffff if not representable */
 970         *cp++ = 0xff;
 971         *cp++ = 0xff;
 972         *cp++ = 0xff;
 973         /* OS type */
 974         *cp++ = EFI_PMBR;
 975         /* ending CHS; 0xffffff if not representable */
 976         *cp++ = 0xff;
 977         *cp++ = 0xff;
 978         *cp++ = 0xff;
 979         /* starting LBA: 1 (little endian format) by EFI definition */
 980         *cp++ = 0x01;
 981         *cp++ = 0x00;
 982         *cp++ = 0x00;
 983         *cp++ = 0x00;
 984         /* ending LBA: last block on the disk (little endian format) */
 985         size_in_lba = vtoc->efi_last_lba;
 986         if (size_in_lba < 0xffffffff) {
 987                 *cp++ = (size_in_lba & 0x000000ff);
 988                 *cp++ = (size_in_lba & 0x0000ff00) >> 8;
 989                 *cp++ = (size_in_lba & 0x00ff0000) >> 16;
 990                 *cp++ = (size_in_lba & 0xff000000) >> 24;
 991         } else {
 992                 *cp++ = 0xff;
 993                 *cp++ = 0xff;
 994                 *cp++ = 0xff;
 995                 *cp++ = 0xff;
 996         }
 997
 998         (void) memcpy(buf, &mb, sizeof (mb));
 999         /* LINTED -- always longlong aligned */
1000         dk_ioc.dki_data = (efi_gpt_t *)buf;
1001         dk_ioc.dki_lba = 0;
1002         dk_ioc.dki_length = len;
1003         if (efi_ioctl(fd, DKIOCSETEFI, &dk_ioc) == -1) {
1004                 free(buf);
1005                 switch (errno) {
1006                 case EIO:
1007                         return (VT_EIO);
1008                 case EINVAL:
1009                         return (VT_EINVAL);
1010                 default:
1011                         return (VT_ERROR);
1012                 }
1013         }
1014         free(buf);
1015         return (0);
1016 }
1017
1018 /* make sure the user specified something reasonable */
1019 static int
1020 check_input(struct dk_gpt *vtoc)
1021 {
1022         int                     resv_part = -1;
1023         int                     i, j;
1024         diskaddr_t              istart, jstart, isize, jsize, endsect;
1025
1026         /*
1027          * Sanity-check the input (make sure no partitions overlap)
1028          */
1029         for (i = 0; i < vtoc->efi_nparts; i++) {
1030                 /* It can't be unassigned and have an actual size */
1031                 if ((vtoc->efi_parts[i].p_tag == V_UNASSIGNED) &&
1032                     (vtoc->efi_parts[i].p_size != 0)) {
1033                         if (efi_debug) {
1034                                 (void) fprintf(stderr, "partition %d is "
1035                                     "\"unassigned\" but has a size of %llu",
1036                                     i, vtoc->efi_parts[i].p_size);
1037                         }
1038                         return (VT_EINVAL);
1039                 }
1040                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED) {
1041                         if (uuid_is_null((uchar_t *)&vtoc->efi_parts[i].p_guid))
1042                                 continue;
1043                         /* we have encountered an unknown uuid */
1044                         vtoc->efi_parts[i].p_tag = 0xff;
1045                 }
1046                 if (vtoc->efi_parts[i].p_tag == V_RESERVED) {
1047                         if (resv_part != -1) {
1048                                 if (efi_debug) {
1049                                         (void) fprintf(stderr, "found "
1050                                             "duplicate reserved partition "
1051                                             "at %d\n", i);
1052                                 }
1053                                 return (VT_EINVAL);
1054                         }
1055                         resv_part = i;
1056                 }
1057                 if ((vtoc->efi_parts[i].p_start < vtoc->efi_first_u_lba) ||
1058                     (vtoc->efi_parts[i].p_start > vtoc->efi_last_u_lba)) {
1059                         if (efi_debug) {
1060                                 (void) fprintf(stderr,
1061                                     "Partition %d starts at %llu.  ",
1062                                     i,
1063                                     vtoc->efi_parts[i].p_start);
1064                                 (void) fprintf(stderr,
1065                                     "It must be between %llu and %llu.\n",
1066                                     vtoc->efi_first_u_lba,
1067                                     vtoc->efi_last_u_lba);
1068                         }
1069                         return (VT_EINVAL);
1070                 }
1071                 if ((vtoc->efi_parts[i].p_start +
1072                     vtoc->efi_parts[i].p_size <
1073                     vtoc->efi_first_u_lba) ||
1074                     (vtoc->efi_parts[i].p_start +
1075                     vtoc->efi_parts[i].p_size >
1076                     vtoc->efi_last_u_lba + 1)) {
1077                         if (efi_debug) {
1078                                 (void) fprintf(stderr,
1079                                     "Partition %d ends at %llu.  ",
1080                                     i,
1081                                     vtoc->efi_parts[i].p_start +
1082                                     vtoc->efi_parts[i].p_size);
1083                                 (void) fprintf(stderr,
1084                                     "It must be between %llu and %llu.\n",
1085                                     vtoc->efi_first_u_lba,
1086                                     vtoc->efi_last_u_lba);
1087                         }
1088                         return (VT_EINVAL);
1089                 }
1090
1091                 for (j = 0; j < vtoc->efi_nparts; j++) {
1092                         isize = vtoc->efi_parts[i].p_size;
1093                         jsize = vtoc->efi_parts[j].p_size;
1094                         istart = vtoc->efi_parts[i].p_start;
1095                         jstart = vtoc->efi_parts[j].p_start;
1096                         if ((i != j) && (isize != 0) && (jsize != 0)) {
1097                                 endsect = jstart + jsize -1;
1098                                 if ((jstart <= istart) &&
1099                                     (istart <= endsect)) {
1100                                         if (efi_debug) {
1101                                                 (void) fprintf(stderr,
1102                                                     "Partition %d overlaps "
1103                                                     "partition %d.", i, j);
1104                                         }
1105                                         return (VT_EINVAL);
1106                                 }
1107                         }
1108                 }
1109         }
1110         /* just a warning for now */
1111         if ((resv_part == -1) && efi_debug) {
1112                 (void) fprintf(stderr,
1113                     "no reserved partition found\n");
1114         }
1115         return (0);
1116 }
1117
1118 /*
1119  * add all the unallocated space to the current label
1120  */
1121 int
1122 efi_use_whole_disk(int fd)
1123 {
1124         struct dk_gpt           *efi_label = NULL;
1125         int                     rval;
1126         int                     i;
1127         uint_t                  resv_index = 0, data_index = 0;
1128         diskaddr_t              resv_start = 0, data_start = 0;
1129         diskaddr_t              difference;
1130
1131         rval = efi_alloc_and_read(fd, &efi_label);
1132         if (rval < 0) {
1133                 if (efi_label != NULL)
1134                         efi_free(efi_label);
1135                 return (rval);
1136         }
1137
1138         /*
1139          * If alter_lba is 1, we are using the backup label.
1140          * Since we can locate the backup label by disk capacity,
1141          * there must be no unallocated space.
1142          */
1143         if ((efi_label->efi_altern_lba == 1) || (efi_label->efi_altern_lba
1144             >= efi_label->efi_last_lba)) {
1145                 if (efi_debug) {
1146                         (void) fprintf(stderr,
1147                             "efi_use_whole_disk: requested space not found\n");
1148                 }
1149                 efi_free(efi_label);
1150                 return (VT_ENOSPC);
1151         }
1152
1153         difference = efi_label->efi_last_lba - efi_label->efi_altern_lba;
1154
1155         /*
1156          * Find the last physically non-zero partition.
1157          * This is the reserved partition.
1158          */
1159         for (i = 0; i < efi_label->efi_nparts; i ++) {
1160                 if (resv_start < efi_label->efi_parts[i].p_start) {
1161                         resv_start = efi_label->efi_parts[i].p_start;
1162                         resv_index = i;
1163                 }
1164         }
1165
1166         /*
1167          * Find the last physically non-zero partition before that.
1168          * This is the data partition.
1169          */
1170         for (i = 0; i < resv_index; i ++) {
1171                 if (data_start < efi_label->efi_parts[i].p_start) {
1172                         data_start = efi_label->efi_parts[i].p_start;
1173                         data_index = i;
1174                 }
1175         }
1176
1177         /*
1178          * Move the reserved partition. There is currently no data in
1179          * here except fabricated devids (which get generated via
1180          * efi_write()). So there is no need to copy data.
1181          */
1182         efi_label->efi_parts[data_index].p_size += difference;
1183         efi_label->efi_parts[resv_index].p_start += difference;
1184         efi_label->efi_last_u_lba += difference;
1185
1186         rval = efi_write(fd, efi_label);
1187         if (rval < 0) {
1188                 if (efi_debug) {
1189                         (void) fprintf(stderr,
1190                             "efi_use_whole_disk:fail to write label, rval=%d\n",
1191                             rval);
1192                 }
1193                 efi_free(efi_label);
1194                 return (rval);
1195         }
1196
1197         efi_free(efi_label);
1198         return (0);
1199 }
1200
1201
1202 /*
1203  * write EFI label and backup label
1204  */
1205 int
1206 efi_write(int fd, struct dk_gpt *vtoc)
1207 {
1208         dk_efi_t                dk_ioc;
1209         efi_gpt_t               *efi;
1210         efi_gpe_t               *efi_parts;
1211         int                     i, j;
1212         struct dk_cinfo         dki_info;
1213         int                     rval;
1214         int                     md_flag = 0;
1215         int                     nblocks;
1216         diskaddr_t              lba_backup_gpt_hdr;
1217
1218         if ((rval = efi_get_info(fd, &dki_info)) != 0)
1219                 return (rval);
1220
1221         /* check if we are dealing wih a metadevice */
1222         if ((strncmp(dki_info.dki_cname, "pseudo", 7) == 0) &&
1223             (strncmp(dki_info.dki_dname, "md", 3) == 0)) {
1224                 md_flag = 1;
1225         }
1226
1227         if (check_input(vtoc)) {
1228                 /*
1229                  * not valid; if it's a metadevice just pass it down
1230                  * because SVM will do its own checking
1231                  */
1232                 if (md_flag == 0) {
1233                         return (VT_EINVAL);
1234                 }
1235         }
1236
1237         dk_ioc.dki_lba = 1;
1238         if (NBLOCKS(vtoc->efi_nparts, vtoc->efi_lbasize) < 34) {
1239                 dk_ioc.dki_length = EFI_MIN_ARRAY_SIZE + vtoc->efi_lbasize;
1240         } else {
1241                 dk_ioc.dki_length = NBLOCKS(vtoc->efi_nparts,
1242                     vtoc->efi_lbasize) *
1243                     vtoc->efi_lbasize;
1244         }
1245
1246         /*
1247          * the number of blocks occupied by GUID partition entry array
1248          */
1249         nblocks = dk_ioc.dki_length / vtoc->efi_lbasize - 1;
1250
1251         /*
1252          * Backup GPT header is located on the block after GUID
1253          * partition entry array. Here, we calculate the address
1254          * for backup GPT header.
1255          */
1256         lba_backup_gpt_hdr = vtoc->efi_last_u_lba + 1 + nblocks;
1257         if (posix_memalign((void **)&dk_ioc.dki_data,
1258             vtoc->efi_lbasize, dk_ioc.dki_length))
1259                 return (VT_ERROR);
1260
1261         memset(dk_ioc.dki_data, 0, dk_ioc.dki_length);
1262         efi = dk_ioc.dki_data;
1263
1264         /* stuff user's input into EFI struct */
1265         efi->efi_gpt_Signature = LE_64(EFI_SIGNATURE);
1266         efi->efi_gpt_Revision = LE_32(vtoc->efi_version); /* 0x02000100 */
1267         efi->efi_gpt_HeaderSize = LE_32(sizeof (struct efi_gpt) - LEN_EFI_PAD);
1268         efi->efi_gpt_Reserved1 = 0;
1269         efi->efi_gpt_MyLBA = LE_64(1ULL);
1270         efi->efi_gpt_AlternateLBA = LE_64(lba_backup_gpt_hdr);
1271         efi->efi_gpt_FirstUsableLBA = LE_64(vtoc->efi_first_u_lba);
1272         efi->efi_gpt_LastUsableLBA = LE_64(vtoc->efi_last_u_lba);
1273         efi->efi_gpt_PartitionEntryLBA = LE_64(2ULL);
1274         efi->efi_gpt_NumberOfPartitionEntries = LE_32(vtoc->efi_nparts);
1275         efi->efi_gpt_SizeOfPartitionEntry = LE_32(sizeof (struct efi_gpe));
1276         UUID_LE_CONVERT(efi->efi_gpt_DiskGUID, vtoc->efi_disk_uguid);
1277
1278         /* LINTED -- always longlong aligned */
1279         efi_parts = (efi_gpe_t *)((char *)dk_ioc.dki_data + vtoc->efi_lbasize);
1280
1281         for (i = 0; i < vtoc->efi_nparts; i++) {
1282                 for (j = 0;
1283                     j < sizeof (conversion_array) /
1284                     sizeof (struct uuid_to_ptag); j++) {
1285
1286                         if (vtoc->efi_parts[i].p_tag == j) {
1287                                 UUID_LE_CONVERT(
1288                                     efi_parts[i].efi_gpe_PartitionTypeGUID,
1289                                     conversion_array[j].uuid);
1290                                 break;
1291                         }
1292                 }
1293
1294                 if (j == sizeof (conversion_array) /
1295                     sizeof (struct uuid_to_ptag)) {
1296                         /*
1297                          * If we didn't have a matching uuid match, bail here.
1298                          * Don't write a label with unknown uuid.
1299                          */
1300                         if (efi_debug) {
1301                                 (void) fprintf(stderr,
1302                                     "Unknown uuid for p_tag %d\n",
1303                                     vtoc->efi_parts[i].p_tag);
1304                         }
1305                         return (VT_EINVAL);
1306                 }
1307
1308                 /* Zero's should be written for empty partitions */
1309                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED)
1310                         continue;
1311
1312                 efi_parts[i].efi_gpe_StartingLBA =
1313                     LE_64(vtoc->efi_parts[i].p_start);
1314                 efi_parts[i].efi_gpe_EndingLBA =
1315                     LE_64(vtoc->efi_parts[i].p_start +
1316                     vtoc->efi_parts[i].p_size - 1);
1317                 efi_parts[i].efi_gpe_Attributes.PartitionAttrs =
1318                     LE_16(vtoc->efi_parts[i].p_flag);
1319                 for (j = 0; j < EFI_PART_NAME_LEN; j++) {
1320                         efi_parts[i].efi_gpe_PartitionName[j] =
1321                             LE_16((ushort_t)vtoc->efi_parts[i].p_name[j]);
1322                 }
1323                 if ((vtoc->efi_parts[i].p_tag != V_UNASSIGNED) &&
1324                     uuid_is_null((uchar_t *)&vtoc->efi_parts[i].p_uguid)) {
1325                         (void) uuid_generate((uchar_t *)
1326                             &vtoc->efi_parts[i].p_uguid);
1327                 }
1328                 bcopy(&vtoc->efi_parts[i].p_uguid,
1329                     &efi_parts[i].efi_gpe_UniquePartitionGUID,
1330                     sizeof (uuid_t));
1331         }
1332         efi->efi_gpt_PartitionEntryArrayCRC32 =
1333             LE_32(efi_crc32((unsigned char *)efi_parts,
1334             vtoc->efi_nparts * (int)sizeof (struct efi_gpe)));
1335         efi->efi_gpt_HeaderCRC32 =
1336             LE_32(efi_crc32((unsigned char *)efi,
1337             LE_32(efi->efi_gpt_HeaderSize)));
1338
1339         if (efi_ioctl(fd, DKIOCSETEFI, &dk_ioc) == -1) {
1340                 free(dk_ioc.dki_data);
1341                 switch (errno) {
1342                 case EIO:
1343                         return (VT_EIO);
1344                 case EINVAL:
1345                         return (VT_EINVAL);
1346                 default:
1347                         return (VT_ERROR);
1348                 }
1349         }
1350         /* if it's a metadevice we're done */
1351         if (md_flag) {
1352                 free(dk_ioc.dki_data);
1353                 return (0);
1354         }
1355
1356         /* write backup partition array */
1357         dk_ioc.dki_lba = vtoc->efi_last_u_lba + 1;
1358         dk_ioc.dki_length -= vtoc->efi_lbasize;
1359         /* LINTED */
1360         dk_ioc.dki_data = (efi_gpt_t *)((char *)dk_ioc.dki_data +
1361             vtoc->efi_lbasize);
1362
1363         if (efi_ioctl(fd, DKIOCSETEFI, &dk_ioc) == -1) {
1364                 /*
1365                  * we wrote the primary label okay, so don't fail
1366                  */
1367                 if (efi_debug) {
1368                         (void) fprintf(stderr,
1369                             "write of backup partitions to block %llu "
1370                             "failed, errno %d\n",
1371                             vtoc->efi_last_u_lba + 1,
1372                             errno);
1373                 }
1374         }
1375         /*
1376          * now swap MyLBA and AlternateLBA fields and write backup
1377          * partition table header
1378          */
1379         dk_ioc.dki_lba = lba_backup_gpt_hdr;
1380         dk_ioc.dki_length = vtoc->efi_lbasize;
1381         /* LINTED */
1382         dk_ioc.dki_data = (efi_gpt_t *)((char *)dk_ioc.dki_data -
1383             vtoc->efi_lbasize);
1384         efi->efi_gpt_AlternateLBA = LE_64(1ULL);
1385         efi->efi_gpt_MyLBA = LE_64(lba_backup_gpt_hdr);
1386         efi->efi_gpt_PartitionEntryLBA = LE_64(vtoc->efi_last_u_lba + 1);
1387         efi->efi_gpt_HeaderCRC32 = 0;
1388         efi->efi_gpt_HeaderCRC32 =
1389             LE_32(efi_crc32((unsigned char *)dk_ioc.dki_data,
1390             LE_32(efi->efi_gpt_HeaderSize)));
1391
1392         if (efi_ioctl(fd, DKIOCSETEFI, &dk_ioc) == -1) {
1393                 if (efi_debug) {
1394                         (void) fprintf(stderr,
1395                             "write of backup header to block %llu failed, "
1396                             "errno %d\n",
1397                             lba_backup_gpt_hdr,
1398                             errno);
1399                 }
1400         }
1401         /* write the PMBR */
1402         (void) write_pmbr(fd, vtoc);
1403         free(dk_ioc.dki_data);
1404
1405         return (0);
1406 }
1407
1408 void
1409 efi_free(struct dk_gpt *ptr)
1410 {
1411         free(ptr);
1412 }
1413
1414 /*
1415  * Input: File descriptor
1416  * Output: 1 if disk has an EFI label, or > 2TB with no VTOC or legacy MBR.
1417  * Otherwise 0.
1418  */
1419 int
1420 efi_type(int fd)
1421 {
1422 #if 0
1423         struct vtoc vtoc;
1424         struct extvtoc extvtoc;
1425
1426         if (ioctl(fd, DKIOCGEXTVTOC, &extvtoc) == -1) {
1427                 if (errno == ENOTSUP)
1428                         return (1);
1429                 else if (errno == ENOTTY) {
1430                         if (ioctl(fd, DKIOCGVTOC, &vtoc) == -1)
1431                                 if (errno == ENOTSUP)
1432                                         return (1);
1433                 }
1434         }
1435         return (0);
1436 #else
1437         return (ENOSYS);
1438 #endif
1439 }
1440
1441 void
1442 efi_err_check(struct dk_gpt *vtoc)
1443 {
1444         int                     resv_part = -1;
1445         int                     i, j;
1446         diskaddr_t              istart, jstart, isize, jsize, endsect;
1447         int                     overlap = 0;
1448
1449         /*
1450          * make sure no partitions overlap
1451          */
1452         for (i = 0; i < vtoc->efi_nparts; i++) {
1453                 /* It can't be unassigned and have an actual size */
1454                 if ((vtoc->efi_parts[i].p_tag == V_UNASSIGNED) &&
1455                     (vtoc->efi_parts[i].p_size != 0)) {
1456                         (void) fprintf(stderr,
1457                             "partition %d is \"unassigned\" but has a size "
1458                             "of %llu\n", i, vtoc->efi_parts[i].p_size);
1459                 }
1460                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED) {
1461                         continue;
1462                 }
1463                 if (vtoc->efi_parts[i].p_tag == V_RESERVED) {
1464                         if (resv_part != -1) {
1465                                 (void) fprintf(stderr,
1466                                     "found duplicate reserved partition at "
1467                                     "%d\n", i);
1468                         }
1469                         resv_part = i;
1470                         if (vtoc->efi_parts[i].p_size != EFI_MIN_RESV_SIZE)
1471                                 (void) fprintf(stderr,
1472                                     "Warning: reserved partition size must "
1473                                     "be %d sectors\n", EFI_MIN_RESV_SIZE);
1474                 }
1475                 if ((vtoc->efi_parts[i].p_start < vtoc->efi_first_u_lba) ||
1476                     (vtoc->efi_parts[i].p_start > vtoc->efi_last_u_lba)) {
1477                         (void) fprintf(stderr,
1478                             "Partition %d starts at %llu\n",
1479                             i,
1480                             vtoc->efi_parts[i].p_start);
1481                         (void) fprintf(stderr,
1482                             "It must be between %llu and %llu.\n",
1483                             vtoc->efi_first_u_lba,
1484                             vtoc->efi_last_u_lba);
1485                 }
1486                 if ((vtoc->efi_parts[i].p_start +
1487                     vtoc->efi_parts[i].p_size <
1488                     vtoc->efi_first_u_lba) ||
1489                     (vtoc->efi_parts[i].p_start +
1490                     vtoc->efi_parts[i].p_size >
1491                     vtoc->efi_last_u_lba + 1)) {
1492                         (void) fprintf(stderr,
1493                             "Partition %d ends at %llu\n",
1494                             i,
1495                             vtoc->efi_parts[i].p_start +
1496                             vtoc->efi_parts[i].p_size);
1497                         (void) fprintf(stderr,
1498                             "It must be between %llu and %llu.\n",
1499                             vtoc->efi_first_u_lba,
1500                             vtoc->efi_last_u_lba);
1501                 }
1502
1503                 for (j = 0; j < vtoc->efi_nparts; j++) {
1504                         isize = vtoc->efi_parts[i].p_size;
1505                         jsize = vtoc->efi_parts[j].p_size;
1506                         istart = vtoc->efi_parts[i].p_start;
1507                         jstart = vtoc->efi_parts[j].p_start;
1508                         if ((i != j) && (isize != 0) && (jsize != 0)) {
1509                                 endsect = jstart + jsize -1;
1510                                 if ((jstart <= istart) &&
1511                                     (istart <= endsect)) {
1512                                         if (!overlap) {
1513                                         (void) fprintf(stderr,
1514                                             "label error: EFI Labels do not "
1515                                             "support overlapping partitions\n");
1516                                         }
1517                                         (void) fprintf(stderr,
1518                                             "Partition %d overlaps partition "
1519                                             "%d.\n", i, j);
1520                                         overlap = 1;
1521                                 }
1522                         }
1523                 }
1524         }
1525         /* make sure there is a reserved partition */
1526         if (resv_part == -1) {
1527                 (void) fprintf(stderr,
1528                     "no reserved partition found\n");
1529         }
1530 }
1531
1532 /*
1533  * We need to get information necessary to construct a *new* efi
1534  * label type
1535  */
1536 int
1537 efi_auto_sense(int fd, struct dk_gpt **vtoc)
1538 {
1539
1540         int     i;
1541
1542         /*
1543          * Now build the default partition table
1544          */
1545         if (efi_alloc_and_init(fd, EFI_NUMPAR, vtoc) != 0) {
1546                 if (efi_debug) {
1547                         (void) fprintf(stderr, "efi_alloc_and_init failed.\n");
1548                 }
1549                 return (-1);
1550         }
1551
1552         for (i = 0; i < MIN((*vtoc)->efi_nparts, V_NUMPAR); i++) {
1553                 (*vtoc)->efi_parts[i].p_tag = default_vtoc_map[i].p_tag;
1554                 (*vtoc)->efi_parts[i].p_flag = default_vtoc_map[i].p_flag;
1555                 (*vtoc)->efi_parts[i].p_start = 0;
1556                 (*vtoc)->efi_parts[i].p_size = 0;
1557         }
1558         /*
1559          * Make constants first
1560          * and variable partitions later
1561          */
1562
1563         /* root partition - s0 128 MB */
1564         (*vtoc)->efi_parts[0].p_start = 34;
1565         (*vtoc)->efi_parts[0].p_size = 262144;
1566
1567         /* partition - s1  128 MB */
1568         (*vtoc)->efi_parts[1].p_start = 262178;
1569         (*vtoc)->efi_parts[1].p_size = 262144;
1570
1571         /* partition -s2 is NOT the Backup disk */
1572         (*vtoc)->efi_parts[2].p_tag = V_UNASSIGNED;
1573
1574         /* partition -s6 /usr partition - HOG */
1575         (*vtoc)->efi_parts[6].p_start = 524322;
1576         (*vtoc)->efi_parts[6].p_size = (*vtoc)->efi_last_u_lba - 524322
1577             - (1024 * 16);
1578
1579         /* efi reserved partition - s9 16K */
1580         (*vtoc)->efi_parts[8].p_start = (*vtoc)->efi_last_u_lba - (1024 * 16);
1581         (*vtoc)->efi_parts[8].p_size = (1024 * 16);
1582         (*vtoc)->efi_parts[8].p_tag = V_RESERVED;
1583         return (0);
1584 }