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  * Copyright (c) 2018 by Delphix. All rights reserved.
  26  */
  27
  28 #include <stdio.h>
  29 #include <stdlib.h>
  30 #include <errno.h>
  31 #include <string.h>
  32 #include <strings.h>
  33 #include <unistd.h>
  34 #include <uuid/uuid.h>
  35 #include <zlib.h>
  36 #include <libintl.h>
  37 #include <sys/types.h>
  38 #include <sys/dkio.h>
  39 #include <sys/vtoc.h>
  40 #include <sys/mhd.h>
  41 #include <sys/param.h>
  42 #include <sys/dktp/fdisk.h>
  43 #include <sys/efi_partition.h>
  44 #include <sys/byteorder.h>
  45 #include <sys/vdev_disk.h>
  46 #include <linux/fs.h>
  47 #include <linux/blkpg.h>
  48
  49 static struct uuid_to_ptag {
  50         struct uuid     uuid;
  51 } conversion_array[] = {
  52         { EFI_UNUSED },
  53         { EFI_BOOT },
  54         { EFI_ROOT },
  55         { EFI_SWAP },
  56         { EFI_USR },
  57         { EFI_BACKUP },
  58         { EFI_UNUSED },         /* STAND is never used */
  59         { EFI_VAR },
  60         { EFI_HOME },
  61         { EFI_ALTSCTR },
  62         { EFI_UNUSED },         /* CACHE (cachefs) is never used */
  63         { EFI_RESERVED },
  64         { EFI_SYSTEM },
  65         { EFI_LEGACY_MBR },
  66         { EFI_SYMC_PUB },
  67         { EFI_SYMC_CDS },
  68         { EFI_MSFT_RESV },
  69         { EFI_DELL_BASIC },
  70         { EFI_DELL_RAID },
  71         { EFI_DELL_SWAP },
  72         { EFI_DELL_LVM },
  73         { EFI_DELL_RESV },
  74         { EFI_AAPL_HFS },
  75         { EFI_AAPL_UFS },
  76         { EFI_FREEBSD_BOOT },
  77         { EFI_FREEBSD_SWAP },
  78         { EFI_FREEBSD_UFS },
  79         { EFI_FREEBSD_VINUM },
  80         { EFI_FREEBSD_ZFS },
  81         { EFI_BIOS_BOOT },
  82         { EFI_INTC_RS },
  83         { EFI_SNE_BOOT },
  84         { EFI_LENOVO_BOOT },
  85         { EFI_MSFT_LDMM },
  86         { EFI_MSFT_LDMD },
  87         { EFI_MSFT_RE },
  88         { EFI_IBM_GPFS },
  89         { EFI_MSFT_STORAGESPACES },
  90         { EFI_HPQ_DATA },
  91         { EFI_HPQ_SVC },
  92         { EFI_RHT_DATA },
  93         { EFI_RHT_HOME },
  94         { EFI_RHT_SRV },
  95         { EFI_RHT_DMCRYPT },
  96         { EFI_RHT_LUKS },
  97         { EFI_FREEBSD_DISKLABEL },
  98         { EFI_AAPL_RAID },
  99         { EFI_AAPL_RAIDOFFLINE },
 100         { EFI_AAPL_BOOT },
 101         { EFI_AAPL_LABEL },
 102         { EFI_AAPL_TVRECOVERY },
 103         { EFI_AAPL_CORESTORAGE },
 104         { EFI_NETBSD_SWAP },
 105         { EFI_NETBSD_FFS },
 106         { EFI_NETBSD_LFS },
 107         { EFI_NETBSD_RAID },
 108         { EFI_NETBSD_CAT },
 109         { EFI_NETBSD_CRYPT },
 110         { EFI_GOOG_KERN },
 111         { EFI_GOOG_ROOT },
 112         { EFI_GOOG_RESV },
 113         { EFI_HAIKU_BFS },
 114         { EFI_MIDNIGHTBSD_BOOT },
 115         { EFI_MIDNIGHTBSD_DATA },
 116         { EFI_MIDNIGHTBSD_SWAP },
 117         { EFI_MIDNIGHTBSD_UFS },
 118         { EFI_MIDNIGHTBSD_VINUM },
 119         { EFI_MIDNIGHTBSD_ZFS },
 120         { EFI_CEPH_JOURNAL },
 121         { EFI_CEPH_DMCRYPTJOURNAL },
 122         { EFI_CEPH_OSD },
 123         { EFI_CEPH_DMCRYPTOSD },
 124         { EFI_CEPH_CREATE },
 125         { EFI_CEPH_DMCRYPTCREATE },
 126         { EFI_OPENBSD_DISKLABEL },
 127         { EFI_BBRY_QNX },
 128         { EFI_BELL_PLAN9 },
 129         { EFI_VMW_KCORE },
 130         { EFI_VMW_VMFS },
 131         { EFI_VMW_RESV },
 132         { EFI_RHT_ROOTX86 },
 133         { EFI_RHT_ROOTAMD64 },
 134         { EFI_RHT_ROOTARM },
 135         { EFI_RHT_ROOTARM64 },
 136         { EFI_ACRONIS_SECUREZONE },
 137         { EFI_ONIE_BOOT },
 138         { EFI_ONIE_CONFIG },
 139         { EFI_IBM_PPRPBOOT },
 140         { EFI_FREEDESKTOP_BOOT }
 141 };
 142
 143 /*
 144  * Default vtoc information for non-SVr4 partitions
 145  */
 146 struct dk_map2  default_vtoc_map[NDKMAP] = {
 147         {       V_ROOT,         0       },              /* a - 0 */
 148         {       V_SWAP,         V_UNMNT },              /* b - 1 */
 149         {       V_BACKUP,       V_UNMNT },              /* c - 2 */
 150         {       V_UNASSIGNED,   0       },              /* d - 3 */
 151         {       V_UNASSIGNED,   0       },              /* e - 4 */
 152         {       V_UNASSIGNED,   0       },              /* f - 5 */
 153         {       V_USR,          0       },              /* g - 6 */
 154         {       V_UNASSIGNED,   0       },              /* h - 7 */
 155
 156 #if defined(_SUNOS_VTOC_16)
 157
 158 #if defined(i386) || defined(__amd64) || defined(__arm) || \
 159     defined(__powerpc) || defined(__sparc) || defined(__s390__) || \
 160     defined(__mips__) || defined(__rv64g__)
 161         {       V_BOOT,         V_UNMNT },              /* i - 8 */
 162         {       V_ALTSCTR,      0       },              /* j - 9 */
 163
 164 #else
 165 #error No VTOC format defined.
 166 #endif                  /* defined(i386) */
 167
 168         {       V_UNASSIGNED,   0       },              /* k - 10 */
 169         {       V_UNASSIGNED,   0       },              /* l - 11 */
 170         {       V_UNASSIGNED,   0       },              /* m - 12 */
 171         {       V_UNASSIGNED,   0       },              /* n - 13 */
 172         {       V_UNASSIGNED,   0       },              /* o - 14 */
 173         {       V_UNASSIGNED,   0       },              /* p - 15 */
 174 #endif                  /* defined(_SUNOS_VTOC_16) */
 175 };
 176
 177 int efi_debug = 0;
 178
 179 static int efi_read(int, struct dk_gpt *);
 180
 181 /*
 182  * Return a 32-bit CRC of the contents of the buffer.  Pre-and-post
 183  * one's conditioning will be handled by crc32() internally.
 184  */
 185 static uint32_t
 186 efi_crc32(const unsigned char *buf, unsigned int size)
 187 {
 188         uint32_t crc = crc32(0, Z_NULL, 0);
 189
 190         crc = crc32(crc, buf, size);
 191
 192         return (crc);
 193 }
 194
 195 static int
 196 read_disk_info(int fd, diskaddr_t *capacity, uint_t *lbsize)
 197 {
 198         int sector_size;
 199         unsigned long long capacity_size;
 200
 201         if (ioctl(fd, BLKSSZGET, &sector_size) < 0)
 202                 return (-1);
 203
 204         if (ioctl(fd, BLKGETSIZE64, &capacity_size) < 0)
 205                 return (-1);
 206
 207         *lbsize = (uint_t)sector_size;
 208         *capacity = (diskaddr_t)(capacity_size / sector_size);
 209
 210         return (0);
 211 }
 212
 213 /*
 214  * Return back the device name associated with the file descriptor. The
 215  * caller is responsible for freeing the memory associated with the
 216  * returned string.
 217  */
 218 static char *
 219 efi_get_devname(int fd)
 220 {
 221         char *path;
 222         char *dev_name;
 223
 224         path = calloc(1, PATH_MAX);
 225         if (path == NULL)
 226                 return (NULL);
 227
 228         /*
 229          * The libefi API only provides the open fd and not the file path.
 230          * To handle this realpath(3) is used to resolve the block device
 231          * name from /proc/self/fd/<fd>.
 232          */
 233         (void) sprintf(path, "/proc/self/fd/%d", fd);
 234         dev_name = realpath(path, NULL);
 235         free(path);
 236         return (dev_name);
 237 }
 238
 239 static int
 240 efi_get_info(int fd, struct dk_cinfo *dki_info)
 241 {
 242         char *dev_path;
 243         int rval = 0;
 244
 245         memset(dki_info, 0, sizeof (*dki_info));
 246
 247         /*
 248          * The simplest way to get the partition number under linux is
 249          * to parse it out of the /dev/<disk><partition> block device name.
 250          * The kernel creates this using the partition number when it
 251          * populates /dev/ so it may be trusted.  The tricky bit here is
 252          * that the naming convention is based on the block device type.
 253          * So we need to take this in to account when parsing out the
 254          * partition information.  Aside from the partition number we collect
 255          * some additional device info.
 256          */
 257         dev_path = efi_get_devname(fd);
 258         if (dev_path == NULL)
 259                 goto error;
 260
 261         if ((strncmp(dev_path, "/dev/sd", 7) == 0)) {
 262                 strcpy(dki_info->dki_cname, "sd");
 263                 dki_info->dki_ctype = DKC_SCSI_CCS;
 264                 rval = sscanf(dev_path, "/dev/%[a-zA-Z]%hu",
 265                     dki_info->dki_dname,
 266                     &dki_info->dki_partition);
 267         } else if ((strncmp(dev_path, "/dev/hd", 7) == 0)) {
 268                 strcpy(dki_info->dki_cname, "hd");
 269                 dki_info->dki_ctype = DKC_DIRECT;
 270                 rval = sscanf(dev_path, "/dev/%[a-zA-Z]%hu",
 271                     dki_info->dki_dname,
 272                     &dki_info->dki_partition);
 273         } else if ((strncmp(dev_path, "/dev/md", 7) == 0)) {
 274                 strcpy(dki_info->dki_cname, "pseudo");
 275                 dki_info->dki_ctype = DKC_MD;
 276                 strcpy(dki_info->dki_dname, "md");
 277                 rval = sscanf(dev_path, "/dev/md%[0-9]p%hu",
 278                     dki_info->dki_dname + 2,
 279                     &dki_info->dki_partition);
 280         } else if ((strncmp(dev_path, "/dev/vd", 7) == 0)) {
 281                 strcpy(dki_info->dki_cname, "vd");
 282                 dki_info->dki_ctype = DKC_MD;
 283                 rval = sscanf(dev_path, "/dev/%[a-zA-Z]%hu",
 284                     dki_info->dki_dname,
 285                     &dki_info->dki_partition);
 286         } else if ((strncmp(dev_path, "/dev/xvd", 8) == 0)) {
 287                 strcpy(dki_info->dki_cname, "xvd");
 288                 dki_info->dki_ctype = DKC_MD;
 289                 rval = sscanf(dev_path, "/dev/%[a-zA-Z]%hu",
 290                     dki_info->dki_dname,
 291                     &dki_info->dki_partition);
 292         } else if ((strncmp(dev_path, "/dev/zd", 7) == 0)) {
 293                 strcpy(dki_info->dki_cname, "zd");
 294                 dki_info->dki_ctype = DKC_MD;
 295                 strcpy(dki_info->dki_dname, "zd");
 296                 rval = sscanf(dev_path, "/dev/zd%[0-9]p%hu",
 297                     dki_info->dki_dname + 2,
 298                     &dki_info->dki_partition);
 299         } else if ((strncmp(dev_path, "/dev/dm-", 8) == 0)) {
 300                 strcpy(dki_info->dki_cname, "pseudo");
 301                 dki_info->dki_ctype = DKC_VBD;
 302                 strcpy(dki_info->dki_dname, "dm-");
 303                 rval = sscanf(dev_path, "/dev/dm-%[0-9]p%hu",
 304                     dki_info->dki_dname + 3,
 305                     &dki_info->dki_partition);
 306         } else if ((strncmp(dev_path, "/dev/ram", 8) == 0)) {
 307                 strcpy(dki_info->dki_cname, "pseudo");
 308                 dki_info->dki_ctype = DKC_PCMCIA_MEM;
 309                 strcpy(dki_info->dki_dname, "ram");
 310                 rval = sscanf(dev_path, "/dev/ram%[0-9]p%hu",
 311                     dki_info->dki_dname + 3,
 312                     &dki_info->dki_partition);
 313         } else if ((strncmp(dev_path, "/dev/loop", 9) == 0)) {
 314                 strcpy(dki_info->dki_cname, "pseudo");
 315                 dki_info->dki_ctype = DKC_VBD;
 316                 strcpy(dki_info->dki_dname, "loop");
 317                 rval = sscanf(dev_path, "/dev/loop%[0-9]p%hu",
 318                     dki_info->dki_dname + 4,
 319                     &dki_info->dki_partition);
 320         } else if ((strncmp(dev_path, "/dev/nvme", 9) == 0)) {
 321                 strcpy(dki_info->dki_cname, "nvme");
 322                 dki_info->dki_ctype = DKC_SCSI_CCS;
 323                 strcpy(dki_info->dki_dname, "nvme");
 324                 (void) sscanf(dev_path, "/dev/nvme%[0-9]",
 325                     dki_info->dki_dname + 4);
 326                 size_t controller_length = strlen(
 327                     dki_info->dki_dname);
 328                 strcpy(dki_info->dki_dname + controller_length,
 329                     "n");
 330                 rval = sscanf(dev_path,
 331                     "/dev/nvme%*[0-9]n%[0-9]p%hu",
 332                     dki_info->dki_dname + controller_length + 1,
 333                     &dki_info->dki_partition);
 334         } else {
 335                 strcpy(dki_info->dki_dname, "unknown");
 336                 strcpy(dki_info->dki_cname, "unknown");
 337                 dki_info->dki_ctype = DKC_UNKNOWN;
 338         }
 339
 340         switch (rval) {
 341         case 0:
 342                 errno = EINVAL;
 343                 goto error;
 344         case 1:
 345                 dki_info->dki_partition = 0;
 346         }
 347
 348         free(dev_path);
 349
 350         return (0);
 351 error:
 352         if (efi_debug)
 353                 (void) fprintf(stderr, "DKIOCINFO errno 0x%x\n", errno);
 354
 355         switch (errno) {
 356         case EIO:
 357                 return (VT_EIO);
 358         case EINVAL:
 359                 return (VT_EINVAL);
 360         default:
 361                 return (VT_ERROR);
 362         }
 363 }
 364
 365 /*
 366  * the number of blocks the EFI label takes up (round up to nearest
 367  * block)
 368  */
 369 #define NBLOCKS(p, l)   (1 + ((((p) * (int)sizeof (efi_gpe_t))  + \
 370                                 ((l) - 1)) / (l)))
 371 /* number of partitions -- limited by what we can malloc */
 372 #define MAX_PARTS       ((4294967295UL - sizeof (struct dk_gpt)) / \
 373                             sizeof (struct dk_part))
 374
 375 int
 376 efi_alloc_and_init(int fd, uint32_t nparts, struct dk_gpt **vtoc)
 377 {
 378         diskaddr_t      capacity = 0;
 379         uint_t          lbsize = 0;
 380         uint_t          nblocks;
 381         size_t          length;
 382         struct dk_gpt   *vptr;
 383         struct uuid     uuid;
 384         struct dk_cinfo dki_info;
 385
 386         if (read_disk_info(fd, &capacity, &lbsize) != 0)
 387                 return (-1);
 388
 389         if (efi_get_info(fd, &dki_info) != 0)
 390                 return (-1);
 391
 392         if (dki_info.dki_partition != 0)
 393                 return (-1);
 394
 395         if ((dki_info.dki_ctype == DKC_PCMCIA_MEM) ||
 396             (dki_info.dki_ctype == DKC_VBD) ||
 397             (dki_info.dki_ctype == DKC_UNKNOWN))
 398                 return (-1);
 399
 400         nblocks = NBLOCKS(nparts, lbsize);
 401         if ((nblocks * lbsize) < EFI_MIN_ARRAY_SIZE + lbsize) {
 402                 /* 16K plus one block for the GPT */
 403                 nblocks = EFI_MIN_ARRAY_SIZE / lbsize + 1;
 404         }
 405
 406         if (nparts > MAX_PARTS) {
 407                 if (efi_debug) {
 408                         (void) fprintf(stderr,
 409                         "the maximum number of partitions supported is %lu\n",
 410                             MAX_PARTS);
 411                 }
 412                 return (-1);
 413         }
 414
 415         length = sizeof (struct dk_gpt) +
 416             sizeof (struct dk_part) * (nparts - 1);
 417
 418         vptr = calloc(1, length);
 419         if (vptr == NULL)
 420                 return (-1);
 421
 422         *vtoc = vptr;
 423
 424         vptr->efi_version = EFI_VERSION_CURRENT;
 425         vptr->efi_lbasize = lbsize;
 426         vptr->efi_nparts = nparts;
 427         /*
 428          * add one block here for the PMBR; on disks with a 512 byte
 429          * block size and 128 or fewer partitions, efi_first_u_lba
 430          * should work out to "34"
 431          */
 432         vptr->efi_first_u_lba = nblocks + 1;
 433         vptr->efi_last_lba = capacity - 1;
 434         vptr->efi_altern_lba = capacity -1;
 435         vptr->efi_last_u_lba = vptr->efi_last_lba - nblocks;
 436
 437         (void) uuid_generate((uchar_t *)&uuid);
 438         UUID_LE_CONVERT(vptr->efi_disk_uguid, uuid);
 439         return (0);
 440 }
 441
 442 /*
 443  * Read EFI - return partition number upon success.
 444  */
 445 int
 446 efi_alloc_and_read(int fd, struct dk_gpt **vtoc)
 447 {
 448         int                     rval;
 449         uint32_t                nparts;
 450         int                     length;
 451         struct dk_gpt           *vptr;
 452
 453         /* figure out the number of entries that would fit into 16K */
 454         nparts = EFI_MIN_ARRAY_SIZE / sizeof (efi_gpe_t);
 455         length = (int) sizeof (struct dk_gpt) +
 456             (int) sizeof (struct dk_part) * (nparts - 1);
 457         vptr = calloc(1, length);
 458
 459         if (vptr == NULL)
 460                 return (VT_ERROR);
 461
 462         vptr->efi_nparts = nparts;
 463         rval = efi_read(fd, vptr);
 464
 465         if ((rval == VT_EINVAL) && vptr->efi_nparts > nparts) {
 466                 void *tmp;
 467                 length = (int) sizeof (struct dk_gpt) +
 468                     (int) sizeof (struct dk_part) * (vptr->efi_nparts - 1);
 469                 nparts = vptr->efi_nparts;
 470                 if ((tmp = realloc(vptr, length)) == NULL) {
 471                         free(vptr);
 472                         *vtoc = NULL;
 473                         return (VT_ERROR);
 474                 } else {
 475                         vptr = tmp;
 476                         rval = efi_read(fd, vptr);
 477                 }
 478         }
 479
 480         if (rval < 0) {
 481                 if (efi_debug) {
 482                         (void) fprintf(stderr,
 483                             "read of EFI table failed, rval=%d\n", rval);
 484                 }
 485                 free(vptr);
 486                 *vtoc = NULL;
 487         } else {
 488                 *vtoc = vptr;
 489         }
 490
 491         return (rval);
 492 }
 493
 494 static int
 495 efi_ioctl(int fd, int cmd, dk_efi_t *dk_ioc)
 496 {
 497         void *data = dk_ioc->dki_data;
 498         int error;
 499         diskaddr_t capacity;
 500         uint_t lbsize;
 501
 502         /*
 503          * When the IO is not being performed in kernel as an ioctl we need
 504          * to know the sector size so we can seek to the proper byte offset.
 505          */
 506         if (read_disk_info(fd, &capacity, &lbsize) == -1) {
 507                 if (efi_debug)
 508                         fprintf(stderr, "unable to read disk info: %d", errno);
 509
 510                 errno = EIO;
 511                 return (-1);
 512         }
 513
 514         switch (cmd) {
 515         case DKIOCGETEFI:
 516                 if (lbsize == 0) {
 517                         if (efi_debug)
 518                                 (void) fprintf(stderr, "DKIOCGETEFI assuming "
 519                                     "LBA %d bytes\n", DEV_BSIZE);
 520
 521                         lbsize = DEV_BSIZE;
 522                 }
 523
 524                 error = lseek(fd, dk_ioc->dki_lba * lbsize, SEEK_SET);
 525                 if (error == -1) {
 526                         if (efi_debug)
 527                                 (void) fprintf(stderr, "DKIOCGETEFI lseek "
 528                                     "error: %d\n", errno);
 529                         return (error);
 530                 }
 531
 532                 error = read(fd, data, dk_ioc->dki_length);
 533                 if (error == -1) {
 534                         if (efi_debug)
 535                                 (void) fprintf(stderr, "DKIOCGETEFI read "
 536                                     "error: %d\n", errno);
 537                         return (error);
 538                 }
 539
 540                 if (error != dk_ioc->dki_length) {
 541                         if (efi_debug)
 542                                 (void) fprintf(stderr, "DKIOCGETEFI short "
 543                                     "read of %d bytes\n", error);
 544                         errno = EIO;
 545                         return (-1);
 546                 }
 547                 error = 0;
 548                 break;
 549
 550         case DKIOCSETEFI:
 551                 if (lbsize == 0) {
 552                         if (efi_debug)
 553                                 (void) fprintf(stderr, "DKIOCSETEFI unknown "
 554                                     "LBA size\n");
 555                         errno = EIO;
 556                         return (-1);
 557                 }
 558
 559                 error = lseek(fd, dk_ioc->dki_lba * lbsize, SEEK_SET);
 560                 if (error == -1) {
 561                         if (efi_debug)
 562                                 (void) fprintf(stderr, "DKIOCSETEFI lseek "
 563                                     "error: %d\n", errno);
 564                         return (error);
 565                 }
 566
 567                 error = write(fd, data, dk_ioc->dki_length);
 568                 if (error == -1) {
 569                         if (efi_debug)
 570                                 (void) fprintf(stderr, "DKIOCSETEFI write "
 571                                     "error: %d\n", errno);
 572                         return (error);
 573                 }
 574
 575                 if (error != dk_ioc->dki_length) {
 576                         if (efi_debug)
 577                                 (void) fprintf(stderr, "DKIOCSETEFI short "
 578                                     "write of %d bytes\n", error);
 579                         errno = EIO;
 580                         return (-1);
 581                 }
 582
 583                 /* Sync the new EFI table to disk */
 584                 error = fsync(fd);
 585                 if (error == -1)
 586                         return (error);
 587
 588                 /* Ensure any local disk cache is also flushed */
 589                 if (ioctl(fd, BLKFLSBUF, 0) == -1)
 590                         return (error);
 591
 592                 error = 0;
 593                 break;
 594
 595         default:
 596                 if (efi_debug)
 597                         (void) fprintf(stderr, "unsupported ioctl()\n");
 598
 599                 errno = EIO;
 600                 return (-1);
 601         }
 602
 603         return (error);
 604 }
 605
 606 int
 607 efi_rescan(int fd)
 608 {
 609         int retry = 10;
 610         int error;
 611
 612         /* Notify the kernel a devices partition table has been updated */
 613         while ((error = ioctl(fd, BLKRRPART)) != 0) {
 614                 if ((--retry == 0) || (errno != EBUSY)) {
 615                         (void) fprintf(stderr, "the kernel failed to rescan "
 616                             "the partition table: %d\n", errno);
 617                         return (-1);
 618                 }
 619                 usleep(50000);
 620         }
 621
 622         return (0);
 623 }
 624
 625 static int
 626 check_label(int fd, dk_efi_t *dk_ioc)
 627 {
 628         efi_gpt_t               *efi;
 629         uint_t                  crc;
 630
 631         if (efi_ioctl(fd, DKIOCGETEFI, dk_ioc) == -1) {
 632                 switch (errno) {
 633                 case EIO:
 634                         return (VT_EIO);
 635                 default:
 636                         return (VT_ERROR);
 637                 }
 638         }
 639         efi = dk_ioc->dki_data;
 640         if (efi->efi_gpt_Signature != LE_64(EFI_SIGNATURE)) {
 641                 if (efi_debug)
 642                         (void) fprintf(stderr,
 643                             "Bad EFI signature: 0x%llx != 0x%llx\n",
 644                             (long long)efi->efi_gpt_Signature,
 645                             (long long)LE_64(EFI_SIGNATURE));
 646                 return (VT_EINVAL);
 647         }
 648
 649         /*
 650          * check CRC of the header; the size of the header should
 651          * never be larger than one block
 652          */
 653         crc = efi->efi_gpt_HeaderCRC32;
 654         efi->efi_gpt_HeaderCRC32 = 0;
 655         len_t headerSize = (len_t)LE_32(efi->efi_gpt_HeaderSize);
 656
 657         if (headerSize < EFI_MIN_LABEL_SIZE || headerSize > EFI_LABEL_SIZE) {
 658                 if (efi_debug)
 659                         (void) fprintf(stderr,
 660                             "Invalid EFI HeaderSize %llu.  Assuming %d.\n",
 661                             headerSize, EFI_MIN_LABEL_SIZE);
 662         }
 663
 664         if ((headerSize > dk_ioc->dki_length) ||
 665             crc != LE_32(efi_crc32((unsigned char *)efi, headerSize))) {
 666                 if (efi_debug)
 667                         (void) fprintf(stderr,
 668                             "Bad EFI CRC: 0x%x != 0x%x\n",
 669                             crc, LE_32(efi_crc32((unsigned char *)efi,
 670                             headerSize)));
 671                 return (VT_EINVAL);
 672         }
 673
 674         return (0);
 675 }
 676
 677 static int
 678 efi_read(int fd, struct dk_gpt *vtoc)
 679 {
 680         int                     i, j;
 681         int                     label_len;
 682         int                     rval = 0;
 683         int                     md_flag = 0;
 684         int                     vdc_flag = 0;
 685         diskaddr_t              capacity = 0;
 686         uint_t                  lbsize = 0;
 687         struct dk_minfo         disk_info;
 688         dk_efi_t                dk_ioc;
 689         efi_gpt_t               *efi;
 690         efi_gpe_t               *efi_parts;
 691         struct dk_cinfo         dki_info;
 692         uint32_t                user_length;
 693         boolean_t               legacy_label = B_FALSE;
 694
 695         /*
 696          * get the partition number for this file descriptor.
 697          */
 698         if ((rval = efi_get_info(fd, &dki_info)) != 0)
 699                 return (rval);
 700
 701         if ((strncmp(dki_info.dki_cname, "pseudo", 7) == 0) &&
 702             (strncmp(dki_info.dki_dname, "md", 3) == 0)) {
 703                 md_flag++;
 704         } else if ((strncmp(dki_info.dki_cname, "vdc", 4) == 0) &&
 705             (strncmp(dki_info.dki_dname, "vdc", 4) == 0)) {
 706                 /*
 707                  * The controller and drive name "vdc" (virtual disk client)
 708                  * indicates a LDoms virtual disk.
 709                  */
 710                 vdc_flag++;
 711         }
 712
 713         /* get the LBA size */
 714         if (read_disk_info(fd, &capacity, &lbsize) == -1) {
 715                 if (efi_debug) {
 716                         (void) fprintf(stderr,
 717                             "unable to read disk info: %d",
 718                             errno);
 719                 }
 720                 return (VT_EINVAL);
 721         }
 722
 723         disk_info.dki_lbsize = lbsize;
 724         disk_info.dki_capacity = capacity;
 725
 726         if (disk_info.dki_lbsize == 0) {
 727                 if (efi_debug) {
 728                         (void) fprintf(stderr,
 729                             "efi_read: assuming LBA 512 bytes\n");
 730                 }
 731                 disk_info.dki_lbsize = DEV_BSIZE;
 732         }
 733         /*
 734          * Read the EFI GPT to figure out how many partitions we need
 735          * to deal with.
 736          */
 737         dk_ioc.dki_lba = 1;
 738         if (NBLOCKS(vtoc->efi_nparts, disk_info.dki_lbsize) < 34) {
 739                 label_len = EFI_MIN_ARRAY_SIZE + disk_info.dki_lbsize;
 740         } else {
 741                 label_len = vtoc->efi_nparts * (int) sizeof (efi_gpe_t) +
 742                     disk_info.dki_lbsize;
 743                 if (label_len % disk_info.dki_lbsize) {
 744                         /* pad to physical sector size */
 745                         label_len += disk_info.dki_lbsize;
 746                         label_len &= ~(disk_info.dki_lbsize - 1);
 747                 }
 748         }
 749
 750         if (posix_memalign((void **)&dk_ioc.dki_data,
 751             disk_info.dki_lbsize, label_len))
 752                 return (VT_ERROR);
 753
 754         memset(dk_ioc.dki_data, 0, label_len);
 755         dk_ioc.dki_length = disk_info.dki_lbsize;
 756         user_length = vtoc->efi_nparts;
 757         efi = dk_ioc.dki_data;
 758         if (md_flag) {
 759                 dk_ioc.dki_length = label_len;
 760                 if (efi_ioctl(fd, DKIOCGETEFI, &dk_ioc) == -1) {
 761                         switch (errno) {
 762                         case EIO:
 763                                 return (VT_EIO);
 764                         default:
 765                                 return (VT_ERROR);
 766                         }
 767                 }
 768         } else if ((rval = check_label(fd, &dk_ioc)) == VT_EINVAL) {
 769                 /*
 770                  * No valid label here; try the alternate. Note that here
 771                  * we just read GPT header and save it into dk_ioc.data,
 772                  * Later, we will read GUID partition entry array if we
 773                  * can get valid GPT header.
 774                  */
 775
 776                 /*
 777                  * This is a workaround for legacy systems. In the past, the
 778                  * last sector of SCSI disk was invisible on x86 platform. At
 779                  * that time, backup label was saved on the next to the last
 780                  * sector. It is possible for users to move a disk from previous
 781                  * solaris system to present system. Here, we attempt to search
 782                  * legacy backup EFI label first.
 783                  */
 784                 dk_ioc.dki_lba = disk_info.dki_capacity - 2;
 785                 dk_ioc.dki_length = disk_info.dki_lbsize;
 786                 rval = check_label(fd, &dk_ioc);
 787                 if (rval == VT_EINVAL) {
 788                         /*
 789                          * we didn't find legacy backup EFI label, try to
 790                          * search backup EFI label in the last block.
 791                          */
 792                         dk_ioc.dki_lba = disk_info.dki_capacity - 1;
 793                         dk_ioc.dki_length = disk_info.dki_lbsize;
 794                         rval = check_label(fd, &dk_ioc);
 795                         if (rval == 0) {
 796                                 legacy_label = B_TRUE;
 797                                 if (efi_debug)
 798                                         (void) fprintf(stderr,
 799                                             "efi_read: primary label corrupt; "
 800                                             "using EFI backup label located on"
 801                                             " the last block\n");
 802                         }
 803                 } else {
 804                         if ((efi_debug) && (rval == 0))
 805                                 (void) fprintf(stderr, "efi_read: primary label"
 806                                     " corrupt; using legacy EFI backup label "
 807                                     " located on the next to last block\n");
 808                 }
 809
 810                 if (rval == 0) {
 811                         dk_ioc.dki_lba = LE_64(efi->efi_gpt_PartitionEntryLBA);
 812                         vtoc->efi_flags |= EFI_GPT_PRIMARY_CORRUPT;
 813                         vtoc->efi_nparts =
 814                             LE_32(efi->efi_gpt_NumberOfPartitionEntries);
 815                         /*
 816                          * Partition tables are between backup GPT header
 817                          * table and ParitionEntryLBA (the starting LBA of
 818                          * the GUID partition entries array). Now that we
 819                          * already got valid GPT header and saved it in
 820                          * dk_ioc.dki_data, we try to get GUID partition
 821                          * entry array here.
 822                          */
 823                         /* LINTED */
 824                         dk_ioc.dki_data = (efi_gpt_t *)((char *)dk_ioc.dki_data
 825                             + disk_info.dki_lbsize);
 826                         if (legacy_label)
 827                                 dk_ioc.dki_length = disk_info.dki_capacity - 1 -
 828                                     dk_ioc.dki_lba;
 829                         else
 830                                 dk_ioc.dki_length = disk_info.dki_capacity - 2 -
 831                                     dk_ioc.dki_lba;
 832                         dk_ioc.dki_length *= disk_info.dki_lbsize;
 833                         if (dk_ioc.dki_length >
 834                             ((len_t)label_len - sizeof (*dk_ioc.dki_data))) {
 835                                 rval = VT_EINVAL;
 836                         } else {
 837                                 /*
 838                                  * read GUID partition entry array
 839                                  */
 840                                 rval = efi_ioctl(fd, DKIOCGETEFI, &dk_ioc);
 841                         }
 842                 }
 843
 844         } else if (rval == 0) {
 845
 846                 dk_ioc.dki_lba = LE_64(efi->efi_gpt_PartitionEntryLBA);
 847                 /* LINTED */
 848                 dk_ioc.dki_data = (efi_gpt_t *)((char *)dk_ioc.dki_data
 849                     + disk_info.dki_lbsize);
 850                 dk_ioc.dki_length = label_len - disk_info.dki_lbsize;
 851                 rval = efi_ioctl(fd, DKIOCGETEFI, &dk_ioc);
 852
 853         } else if (vdc_flag && rval == VT_ERROR && errno == EINVAL) {
 854                 /*
 855                  * When the device is a LDoms virtual disk, the DKIOCGETEFI
 856                  * ioctl can fail with EINVAL if the virtual disk backend
 857                  * is a ZFS volume serviced by a domain running an old version
 858                  * of Solaris. This is because the DKIOCGETEFI ioctl was
 859                  * initially incorrectly implemented for a ZFS volume and it
 860                  * expected the GPT and GPE to be retrieved with a single ioctl.
 861                  * So we try to read the GPT and the GPE using that old style
 862                  * ioctl.
 863                  */
 864                 dk_ioc.dki_lba = 1;
 865                 dk_ioc.dki_length = label_len;
 866                 rval = check_label(fd, &dk_ioc);
 867         }
 868
 869         if (rval < 0) {
 870                 free(efi);
 871                 return (rval);
 872         }
 873
 874         /* LINTED -- always longlong aligned */
 875         efi_parts = (efi_gpe_t *)(((char *)efi) + disk_info.dki_lbsize);
 876
 877         /*
 878          * Assemble this into a "dk_gpt" struct for easier
 879          * digestibility by applications.
 880          */
 881         vtoc->efi_version = LE_32(efi->efi_gpt_Revision);
 882         vtoc->efi_nparts = LE_32(efi->efi_gpt_NumberOfPartitionEntries);
 883         vtoc->efi_part_size = LE_32(efi->efi_gpt_SizeOfPartitionEntry);
 884         vtoc->efi_lbasize = disk_info.dki_lbsize;
 885         vtoc->efi_last_lba = disk_info.dki_capacity - 1;
 886         vtoc->efi_first_u_lba = LE_64(efi->efi_gpt_FirstUsableLBA);
 887         vtoc->efi_last_u_lba = LE_64(efi->efi_gpt_LastUsableLBA);
 888         vtoc->efi_altern_lba = LE_64(efi->efi_gpt_AlternateLBA);
 889         UUID_LE_CONVERT(vtoc->efi_disk_uguid, efi->efi_gpt_DiskGUID);
 890
 891         /*
 892          * If the array the user passed in is too small, set the length
 893          * to what it needs to be and return
 894          */
 895         if (user_length < vtoc->efi_nparts) {
 896                 return (VT_EINVAL);
 897         }
 898
 899         for (i = 0; i < vtoc->efi_nparts; i++) {
 900
 901                 UUID_LE_CONVERT(vtoc->efi_parts[i].p_guid,
 902                     efi_parts[i].efi_gpe_PartitionTypeGUID);
 903
 904                 for (j = 0;
 905                     j < sizeof (conversion_array)
 906                     / sizeof (struct uuid_to_ptag); j++) {
 907
 908                         if (bcmp(&vtoc->efi_parts[i].p_guid,
 909                             &conversion_array[j].uuid,
 910                             sizeof (struct uuid)) == 0) {
 911                                 vtoc->efi_parts[i].p_tag = j;
 912                                 break;
 913                         }
 914                 }
 915                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED)
 916                         continue;
 917                 vtoc->efi_parts[i].p_flag =
 918                     LE_16(efi_parts[i].efi_gpe_Attributes.PartitionAttrs);
 919                 vtoc->efi_parts[i].p_start =
 920                     LE_64(efi_parts[i].efi_gpe_StartingLBA);
 921                 vtoc->efi_parts[i].p_size =
 922                     LE_64(efi_parts[i].efi_gpe_EndingLBA) -
 923                     vtoc->efi_parts[i].p_start + 1;
 924                 for (j = 0; j < EFI_PART_NAME_LEN; j++) {
 925                         vtoc->efi_parts[i].p_name[j] =
 926                             (uchar_t)LE_16(
 927                             efi_parts[i].efi_gpe_PartitionName[j]);
 928                 }
 929
 930                 UUID_LE_CONVERT(vtoc->efi_parts[i].p_uguid,
 931                     efi_parts[i].efi_gpe_UniquePartitionGUID);
 932         }
 933         free(efi);
 934
 935         return (dki_info.dki_partition);
 936 }
 937
 938 /* writes a "protective" MBR */
 939 static int
 940 write_pmbr(int fd, struct dk_gpt *vtoc)
 941 {
 942         dk_efi_t        dk_ioc;
 943         struct mboot    mb;
 944         uchar_t         *cp;
 945         diskaddr_t      size_in_lba;
 946         uchar_t         *buf;
 947         int             len;
 948
 949         len = (vtoc->efi_lbasize == 0) ? sizeof (mb) : vtoc->efi_lbasize;
 950         if (posix_memalign((void **)&buf, len, len))
 951                 return (VT_ERROR);
 952
 953         /*
 954          * Preserve any boot code and disk signature if the first block is
 955          * already an MBR.
 956          */
 957         memset(buf, 0, len);
 958         dk_ioc.dki_lba = 0;
 959         dk_ioc.dki_length = len;
 960         /* LINTED -- always longlong aligned */
 961         dk_ioc.dki_data = (efi_gpt_t *)buf;
 962         if (efi_ioctl(fd, DKIOCGETEFI, &dk_ioc) == -1) {
 963                 (void) memcpy(&mb, buf, sizeof (mb));
 964                 bzero(&mb, sizeof (mb));
 965                 mb.signature = LE_16(MBB_MAGIC);
 966         } else {
 967                 (void) memcpy(&mb, buf, sizeof (mb));
 968                 if (mb.signature != LE_16(MBB_MAGIC)) {
 969                         bzero(&mb, sizeof (mb));
 970                         mb.signature = LE_16(MBB_MAGIC);
 971                 }
 972         }
 973
 974         bzero(&mb.parts, sizeof (mb.parts));
 975         cp = (uchar_t *)&mb.parts[0];
 976         /* bootable or not */
 977         *cp++ = 0;
 978         /* beginning CHS; 0xffffff if not representable */
 979         *cp++ = 0xff;
 980         *cp++ = 0xff;
 981         *cp++ = 0xff;
 982         /* OS type */
 983         *cp++ = EFI_PMBR;
 984         /* ending CHS; 0xffffff if not representable */
 985         *cp++ = 0xff;
 986         *cp++ = 0xff;
 987         *cp++ = 0xff;
 988         /* starting LBA: 1 (little endian format) by EFI definition */
 989         *cp++ = 0x01;
 990         *cp++ = 0x00;
 991         *cp++ = 0x00;
 992         *cp++ = 0x00;
 993         /* ending LBA: last block on the disk (little endian format) */
 994         size_in_lba = vtoc->efi_last_lba;
 995         if (size_in_lba < 0xffffffff) {
 996                 *cp++ = (size_in_lba & 0x000000ff);
 997                 *cp++ = (size_in_lba & 0x0000ff00) >> 8;
 998                 *cp++ = (size_in_lba & 0x00ff0000) >> 16;
 999                 *cp++ = (size_in_lba & 0xff000000) >> 24;
1000         } else {
1001                 *cp++ = 0xff;
1002                 *cp++ = 0xff;
1003                 *cp++ = 0xff;
1004                 *cp++ = 0xff;
1005         }
1006
1007         (void) memcpy(buf, &mb, sizeof (mb));
1008         /* LINTED -- always longlong aligned */
1009         dk_ioc.dki_data = (efi_gpt_t *)buf;
1010         dk_ioc.dki_lba = 0;
1011         dk_ioc.dki_length = len;
1012         if (efi_ioctl(fd, DKIOCSETEFI, &dk_ioc) == -1) {
1013                 free(buf);
1014                 switch (errno) {
1015                 case EIO:
1016                         return (VT_EIO);
1017                 case EINVAL:
1018                         return (VT_EINVAL);
1019                 default:
1020                         return (VT_ERROR);
1021                 }
1022         }
1023         free(buf);
1024         return (0);
1025 }
1026
1027 /* make sure the user specified something reasonable */
1028 static int
1029 check_input(struct dk_gpt *vtoc)
1030 {
1031         int                     resv_part = -1;
1032         int                     i, j;
1033         diskaddr_t              istart, jstart, isize, jsize, endsect;
1034
1035         /*
1036          * Sanity-check the input (make sure no partitions overlap)
1037          */
1038         for (i = 0; i < vtoc->efi_nparts; i++) {
1039                 /* It can't be unassigned and have an actual size */
1040                 if ((vtoc->efi_parts[i].p_tag == V_UNASSIGNED) &&
1041                     (vtoc->efi_parts[i].p_size != 0)) {
1042                         if (efi_debug) {
1043                                 (void) fprintf(stderr, "partition %d is "
1044                                     "\"unassigned\" but has a size of %llu",
1045                                     i, vtoc->efi_parts[i].p_size);
1046                         }
1047                         return (VT_EINVAL);
1048                 }
1049                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED) {
1050                         if (uuid_is_null((uchar_t *)&vtoc->efi_parts[i].p_guid))
1051                                 continue;
1052                         /* we have encountered an unknown uuid */
1053                         vtoc->efi_parts[i].p_tag = 0xff;
1054                 }
1055                 if (vtoc->efi_parts[i].p_tag == V_RESERVED) {
1056                         if (resv_part != -1) {
1057                                 if (efi_debug) {
1058                                         (void) fprintf(stderr, "found "
1059                                             "duplicate reserved partition "
1060                                             "at %d\n", i);
1061                                 }
1062                                 return (VT_EINVAL);
1063                         }
1064                         resv_part = i;
1065                 }
1066                 if ((vtoc->efi_parts[i].p_start < vtoc->efi_first_u_lba) ||
1067                     (vtoc->efi_parts[i].p_start > vtoc->efi_last_u_lba)) {
1068                         if (efi_debug) {
1069                                 (void) fprintf(stderr,
1070                                     "Partition %d starts at %llu.  ",
1071                                     i,
1072                                     vtoc->efi_parts[i].p_start);
1073                                 (void) fprintf(stderr,
1074                                     "It must be between %llu and %llu.\n",
1075                                     vtoc->efi_first_u_lba,
1076                                     vtoc->efi_last_u_lba);
1077                         }
1078                         return (VT_EINVAL);
1079                 }
1080                 if ((vtoc->efi_parts[i].p_start +
1081                     vtoc->efi_parts[i].p_size <
1082                     vtoc->efi_first_u_lba) ||
1083                     (vtoc->efi_parts[i].p_start +
1084                     vtoc->efi_parts[i].p_size >
1085                     vtoc->efi_last_u_lba + 1)) {
1086                         if (efi_debug) {
1087                                 (void) fprintf(stderr,
1088                                     "Partition %d ends at %llu.  ",
1089                                     i,
1090                                     vtoc->efi_parts[i].p_start +
1091                                     vtoc->efi_parts[i].p_size);
1092                                 (void) fprintf(stderr,
1093                                     "It must be between %llu and %llu.\n",
1094                                     vtoc->efi_first_u_lba,
1095                                     vtoc->efi_last_u_lba);
1096                         }
1097                         return (VT_EINVAL);
1098                 }
1099
1100                 for (j = 0; j < vtoc->efi_nparts; j++) {
1101                         isize = vtoc->efi_parts[i].p_size;
1102                         jsize = vtoc->efi_parts[j].p_size;
1103                         istart = vtoc->efi_parts[i].p_start;
1104                         jstart = vtoc->efi_parts[j].p_start;
1105                         if ((i != j) && (isize != 0) && (jsize != 0)) {
1106                                 endsect = jstart + jsize -1;
1107                                 if ((jstart <= istart) &&
1108                                     (istart <= endsect)) {
1109                                         if (efi_debug) {
1110                                                 (void) fprintf(stderr,
1111                                                     "Partition %d overlaps "
1112                                                     "partition %d.", i, j);
1113                                         }
1114                                         return (VT_EINVAL);
1115                                 }
1116                         }
1117                 }
1118         }
1119         /* just a warning for now */
1120         if ((resv_part == -1) && efi_debug) {
1121                 (void) fprintf(stderr,
1122                     "no reserved partition found\n");
1123         }
1124         return (0);
1125 }
1126
1127 static int
1128 call_blkpg_ioctl(int fd, int command, diskaddr_t start,
1129     diskaddr_t size, uint_t pno)
1130 {
1131         struct blkpg_ioctl_arg ioctl_arg;
1132         struct blkpg_partition  linux_part;
1133         memset(&linux_part, 0, sizeof (linux_part));
1134
1135         char *path = efi_get_devname(fd);
1136         if (path == NULL) {
1137                 (void) fprintf(stderr, "failed to retrieve device name\n");
1138                 return (VT_EINVAL);
1139         }
1140
1141         linux_part.start = start;
1142         linux_part.length = size;
1143         linux_part.pno = pno;
1144         snprintf(linux_part.devname, BLKPG_DEVNAMELTH - 1, "%s%u", path, pno);
1145         linux_part.devname[BLKPG_DEVNAMELTH - 1] = '\0';
1146         free(path);
1147
1148         ioctl_arg.op = command;
1149         ioctl_arg.flags = 0;
1150         ioctl_arg.datalen = sizeof (struct blkpg_partition);
1151         ioctl_arg.data = &linux_part;
1152
1153         return (ioctl(fd, BLKPG, &ioctl_arg));
1154 }
1155
1156 /*
1157  * add all the unallocated space to the current label
1158  */
1159 int
1160 efi_use_whole_disk(int fd)
1161 {
1162         struct dk_gpt *efi_label = NULL;
1163         int rval;
1164         int i;
1165         uint_t resv_index = 0, data_index = 0;
1166         diskaddr_t resv_start = 0, data_start = 0;
1167         diskaddr_t data_size, limit, difference;
1168         boolean_t sync_needed = B_FALSE;
1169         uint_t nblocks;
1170
1171         rval = efi_alloc_and_read(fd, &efi_label);
1172         if (rval < 0) {
1173                 if (efi_label != NULL)
1174                         efi_free(efi_label);
1175                 return (rval);
1176         }
1177
1178         /*
1179          * Find the last physically non-zero partition.
1180          * This should be the reserved partition.
1181          */
1182         for (i = 0; i < efi_label->efi_nparts; i ++) {
1183                 if (resv_start < efi_label->efi_parts[i].p_start) {
1184                         resv_start = efi_label->efi_parts[i].p_start;
1185                         resv_index = i;
1186                 }
1187         }
1188
1189         /*
1190          * Find the last physically non-zero partition before that.
1191          * This is the data partition.
1192          */
1193         for (i = 0; i < resv_index; i ++) {
1194                 if (data_start < efi_label->efi_parts[i].p_start) {
1195                         data_start = efi_label->efi_parts[i].p_start;
1196                         data_index = i;
1197                 }
1198         }
1199         data_size = efi_label->efi_parts[data_index].p_size;
1200
1201         /*
1202          * See the "efi_alloc_and_init" function for more information
1203          * about where this "nblocks" value comes from.
1204          */
1205         nblocks = efi_label->efi_first_u_lba - 1;
1206
1207         /*
1208          * Determine if the EFI label is out of sync. We check that:
1209          *
1210          * 1. the data partition ends at the limit we set, and
1211          * 2. the reserved partition starts at the limit we set.
1212          *
1213          * If either of these conditions is not met, then we need to
1214          * resync the EFI label.
1215          *
1216          * The limit is the last usable LBA, determined by the last LBA
1217          * and the first usable LBA fields on the EFI label of the disk
1218          * (see the lines directly above). Additionally, we factor in
1219          * EFI_MIN_RESV_SIZE (per its use in "zpool_label_disk") and
1220          * P2ALIGN it to ensure the partition boundaries are aligned
1221          * (for performance reasons). The alignment should match the
1222          * alignment used by the "zpool_label_disk" function.
1223          */
1224         limit = P2ALIGN(efi_label->efi_last_lba - nblocks - EFI_MIN_RESV_SIZE,
1225             PARTITION_END_ALIGNMENT);
1226         if (data_start + data_size != limit || resv_start != limit)
1227                 sync_needed = B_TRUE;
1228
1229         if (efi_debug && sync_needed)
1230                 (void) fprintf(stderr, "efi_use_whole_disk: sync needed\n");
1231
1232         /*
1233          * If alter_lba is 1, we are using the backup label.
1234          * Since we can locate the backup label by disk capacity,
1235          * there must be no unallocated space.
1236          */
1237         if ((efi_label->efi_altern_lba == 1) || (efi_label->efi_altern_lba
1238             >= efi_label->efi_last_lba && !sync_needed)) {
1239                 if (efi_debug) {
1240                         (void) fprintf(stderr,
1241                             "efi_use_whole_disk: requested space not found\n");
1242                 }
1243                 efi_free(efi_label);
1244                 return (VT_ENOSPC);
1245         }
1246
1247         /*
1248          * Verify that we've found the reserved partition by checking
1249          * that it looks the way it did when we created it in zpool_label_disk.
1250          * If we've found the incorrect partition, then we know that this
1251          * device was reformatted and no longer is solely used by ZFS.
1252          */
1253         if ((efi_label->efi_parts[resv_index].p_size != EFI_MIN_RESV_SIZE) ||
1254             (efi_label->efi_parts[resv_index].p_tag != V_RESERVED) ||
1255             (resv_index != 8)) {
1256                 if (efi_debug) {
1257                         (void) fprintf(stderr,
1258                             "efi_use_whole_disk: wholedisk not available\n");
1259                 }
1260                 efi_free(efi_label);
1261                 return (VT_ENOSPC);
1262         }
1263
1264         if (data_start + data_size != resv_start) {
1265                 if (efi_debug) {
1266                         (void) fprintf(stderr,
1267                             "efi_use_whole_disk: "
1268                             "data_start (%lli) + "
1269                             "data_size (%lli) != "
1270                             "resv_start (%lli)\n",
1271                             data_start, data_size, resv_start);
1272                 }
1273
1274                 return (VT_EINVAL);
1275         }
1276
1277         if (limit < resv_start) {
1278                 if (efi_debug) {
1279                         (void) fprintf(stderr,
1280                             "efi_use_whole_disk: "
1281                             "limit (%lli) < resv_start (%lli)\n",
1282                             limit, resv_start);
1283                 }
1284
1285                 return (VT_EINVAL);
1286         }
1287
1288         difference = limit - resv_start;
1289
1290         if (efi_debug)
1291                 (void) fprintf(stderr,
1292                     "efi_use_whole_disk: difference is %lli\n", difference);
1293
1294         /*
1295          * Move the reserved partition. There is currently no data in
1296          * here except fabricated devids (which get generated via
1297          * efi_write()). So there is no need to copy data.
1298          */
1299         efi_label->efi_parts[data_index].p_size += difference;
1300         efi_label->efi_parts[resv_index].p_start += difference;
1301         efi_label->efi_last_u_lba = efi_label->efi_last_lba - nblocks;
1302
1303         /*
1304          * Rescanning the partition table in the kernel can result
1305          * in the device links to be removed (see comment in vdev_disk_open).
1306          * If BLKPG_RESIZE_PARTITION is available, then we can resize
1307          * the partition table online and avoid having to remove the device
1308          * links used by the pool. This provides a very deterministic
1309          * approach to resizing devices and does not require any
1310          * loops waiting for devices to reappear.
1311          */
1312 #ifdef BLKPG_RESIZE_PARTITION
1313         /*
1314          * Delete the reserved partition since we're about to expand
1315          * the data partition and it would overlap with the reserved
1316          * partition.
1317          * NOTE: The starting index for the ioctl is 1 while for the
1318          * EFI partitions it's 0. For that reason we have to add one
1319          * whenever we make an ioctl call.
1320          */
1321         rval = call_blkpg_ioctl(fd, BLKPG_DEL_PARTITION, 0, 0, resv_index + 1);
1322         if (rval != 0)
1323                 goto out;
1324
1325         /*
1326          * Expand the data partition
1327          */
1328         rval = call_blkpg_ioctl(fd, BLKPG_RESIZE_PARTITION,
1329             efi_label->efi_parts[data_index].p_start * efi_label->efi_lbasize,
1330             efi_label->efi_parts[data_index].p_size * efi_label->efi_lbasize,
1331             data_index + 1);
1332         if (rval != 0) {
1333                 (void) fprintf(stderr, "Unable to resize data "
1334                     "partition:  %d\n", rval);
1335                 /*
1336                  * Since we failed to resize, we need to reset the start
1337                  * of the reserve partition and re-create it.
1338                  */
1339                 efi_label->efi_parts[resv_index].p_start -= difference;
1340         }
1341
1342         /*
1343          * Re-add the reserved partition. If we've expanded the data partition
1344          * then we'll move the reserve partition to the end of the data
1345          * partition. Otherwise, we'll recreate the partition in its original
1346          * location. Note that we do this as best-effort and ignore any
1347          * errors that may arise here. This will ensure that we finish writing
1348          * the EFI label.
1349          */
1350         (void) call_blkpg_ioctl(fd, BLKPG_ADD_PARTITION,
1351             efi_label->efi_parts[resv_index].p_start * efi_label->efi_lbasize,
1352             efi_label->efi_parts[resv_index].p_size * efi_label->efi_lbasize,
1353             resv_index + 1);
1354 #endif
1355
1356         /*
1357          * We're now ready to write the EFI label.
1358          */
1359         if (rval == 0) {
1360                 rval = efi_write(fd, efi_label);
1361                 if (rval < 0 && efi_debug) {
1362                         (void) fprintf(stderr, "efi_use_whole_disk:fail "
1363                             "to write label, rval=%d\n", rval);
1364                 }
1365         }
1366
1367 out:
1368         efi_free(efi_label);
1369         return (rval);
1370 }
1371
1372 /*
1373  * write EFI label and backup label
1374  */
1375 int
1376 efi_write(int fd, struct dk_gpt *vtoc)
1377 {
1378         dk_efi_t                dk_ioc;
1379         efi_gpt_t               *efi;
1380         efi_gpe_t               *efi_parts;
1381         int                     i, j;
1382         struct dk_cinfo         dki_info;
1383         int                     rval;
1384         int                     md_flag = 0;
1385         int                     nblocks;
1386         diskaddr_t              lba_backup_gpt_hdr;
1387
1388         if ((rval = efi_get_info(fd, &dki_info)) != 0)
1389                 return (rval);
1390
1391         /* check if we are dealing with a metadevice */
1392         if ((strncmp(dki_info.dki_cname, "pseudo", 7) == 0) &&
1393             (strncmp(dki_info.dki_dname, "md", 3) == 0)) {
1394                 md_flag = 1;
1395         }
1396
1397         if (check_input(vtoc)) {
1398                 /*
1399                  * not valid; if it's a metadevice just pass it down
1400                  * because SVM will do its own checking
1401                  */
1402                 if (md_flag == 0) {
1403                         return (VT_EINVAL);
1404                 }
1405         }
1406
1407         dk_ioc.dki_lba = 1;
1408         if (NBLOCKS(vtoc->efi_nparts, vtoc->efi_lbasize) < 34) {
1409                 dk_ioc.dki_length = EFI_MIN_ARRAY_SIZE + vtoc->efi_lbasize;
1410         } else {
1411                 dk_ioc.dki_length = NBLOCKS(vtoc->efi_nparts,
1412                     vtoc->efi_lbasize) *
1413                     vtoc->efi_lbasize;
1414         }
1415
1416         /*
1417          * the number of blocks occupied by GUID partition entry array
1418          */
1419         nblocks = dk_ioc.dki_length / vtoc->efi_lbasize - 1;
1420
1421         /*
1422          * Backup GPT header is located on the block after GUID
1423          * partition entry array. Here, we calculate the address
1424          * for backup GPT header.
1425          */
1426         lba_backup_gpt_hdr = vtoc->efi_last_u_lba + 1 + nblocks;
1427         if (posix_memalign((void **)&dk_ioc.dki_data,
1428             vtoc->efi_lbasize, dk_ioc.dki_length))
1429                 return (VT_ERROR);
1430
1431         memset(dk_ioc.dki_data, 0, dk_ioc.dki_length);
1432         efi = dk_ioc.dki_data;
1433
1434         /* stuff user's input into EFI struct */
1435         efi->efi_gpt_Signature = LE_64(EFI_SIGNATURE);
1436         efi->efi_gpt_Revision = LE_32(vtoc->efi_version); /* 0x02000100 */
1437         efi->efi_gpt_HeaderSize = LE_32(sizeof (struct efi_gpt) - LEN_EFI_PAD);
1438         efi->efi_gpt_Reserved1 = 0;
1439         efi->efi_gpt_MyLBA = LE_64(1ULL);
1440         efi->efi_gpt_AlternateLBA = LE_64(lba_backup_gpt_hdr);
1441         efi->efi_gpt_FirstUsableLBA = LE_64(vtoc->efi_first_u_lba);
1442         efi->efi_gpt_LastUsableLBA = LE_64(vtoc->efi_last_u_lba);
1443         efi->efi_gpt_PartitionEntryLBA = LE_64(2ULL);
1444         efi->efi_gpt_NumberOfPartitionEntries = LE_32(vtoc->efi_nparts);
1445         efi->efi_gpt_SizeOfPartitionEntry = LE_32(sizeof (struct efi_gpe));
1446         UUID_LE_CONVERT(efi->efi_gpt_DiskGUID, vtoc->efi_disk_uguid);
1447
1448         /* LINTED -- always longlong aligned */
1449         efi_parts = (efi_gpe_t *)((char *)dk_ioc.dki_data + vtoc->efi_lbasize);
1450
1451         for (i = 0; i < vtoc->efi_nparts; i++) {
1452                 for (j = 0;
1453                     j < sizeof (conversion_array) /
1454                     sizeof (struct uuid_to_ptag); j++) {
1455
1456                         if (vtoc->efi_parts[i].p_tag == j) {
1457                                 UUID_LE_CONVERT(
1458                                     efi_parts[i].efi_gpe_PartitionTypeGUID,
1459                                     conversion_array[j].uuid);
1460                                 break;
1461                         }
1462                 }
1463
1464                 if (j == sizeof (conversion_array) /
1465                     sizeof (struct uuid_to_ptag)) {
1466                         /*
1467                          * If we didn't have a matching uuid match, bail here.
1468                          * Don't write a label with unknown uuid.
1469                          */
1470                         if (efi_debug) {
1471                                 (void) fprintf(stderr,
1472                                     "Unknown uuid for p_tag %d\n",
1473                                     vtoc->efi_parts[i].p_tag);
1474                         }
1475                         return (VT_EINVAL);
1476                 }
1477
1478                 /* Zero's should be written for empty partitions */
1479                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED)
1480                         continue;
1481
1482                 efi_parts[i].efi_gpe_StartingLBA =
1483                     LE_64(vtoc->efi_parts[i].p_start);
1484                 efi_parts[i].efi_gpe_EndingLBA =
1485                     LE_64(vtoc->efi_parts[i].p_start +
1486                     vtoc->efi_parts[i].p_size - 1);
1487                 efi_parts[i].efi_gpe_Attributes.PartitionAttrs =
1488                     LE_16(vtoc->efi_parts[i].p_flag);
1489                 for (j = 0; j < EFI_PART_NAME_LEN; j++) {
1490                         efi_parts[i].efi_gpe_PartitionName[j] =
1491                             LE_16((ushort_t)vtoc->efi_parts[i].p_name[j]);
1492                 }
1493                 if ((vtoc->efi_parts[i].p_tag != V_UNASSIGNED) &&
1494                     uuid_is_null((uchar_t *)&vtoc->efi_parts[i].p_uguid)) {
1495                         (void) uuid_generate((uchar_t *)
1496                             &vtoc->efi_parts[i].p_uguid);
1497                 }
1498                 bcopy(&vtoc->efi_parts[i].p_uguid,
1499                     &efi_parts[i].efi_gpe_UniquePartitionGUID,
1500                     sizeof (uuid_t));
1501         }
1502         efi->efi_gpt_PartitionEntryArrayCRC32 =
1503             LE_32(efi_crc32((unsigned char *)efi_parts,
1504             vtoc->efi_nparts * (int)sizeof (struct efi_gpe)));
1505         efi->efi_gpt_HeaderCRC32 =
1506             LE_32(efi_crc32((unsigned char *)efi,
1507             LE_32(efi->efi_gpt_HeaderSize)));
1508
1509         if (efi_ioctl(fd, DKIOCSETEFI, &dk_ioc) == -1) {
1510                 free(dk_ioc.dki_data);
1511                 switch (errno) {
1512                 case EIO:
1513                         return (VT_EIO);
1514                 case EINVAL:
1515                         return (VT_EINVAL);
1516                 default:
1517                         return (VT_ERROR);
1518                 }
1519         }
1520         /* if it's a metadevice we're done */
1521         if (md_flag) {
1522                 free(dk_ioc.dki_data);
1523                 return (0);
1524         }
1525
1526         /* write backup partition array */
1527         dk_ioc.dki_lba = vtoc->efi_last_u_lba + 1;
1528         dk_ioc.dki_length -= vtoc->efi_lbasize;
1529         /* LINTED */
1530         dk_ioc.dki_data = (efi_gpt_t *)((char *)dk_ioc.dki_data +
1531             vtoc->efi_lbasize);
1532
1533         if (efi_ioctl(fd, DKIOCSETEFI, &dk_ioc) == -1) {
1534                 /*
1535                  * we wrote the primary label okay, so don't fail
1536                  */
1537                 if (efi_debug) {
1538                         (void) fprintf(stderr,
1539                             "write of backup partitions to block %llu "
1540                             "failed, errno %d\n",
1541                             vtoc->efi_last_u_lba + 1,
1542                             errno);
1543                 }
1544         }
1545         /*
1546          * now swap MyLBA and AlternateLBA fields and write backup
1547          * partition table header
1548          */
1549         dk_ioc.dki_lba = lba_backup_gpt_hdr;
1550         dk_ioc.dki_length = vtoc->efi_lbasize;
1551         /* LINTED */
1552         dk_ioc.dki_data = (efi_gpt_t *)((char *)dk_ioc.dki_data -
1553             vtoc->efi_lbasize);
1554         efi->efi_gpt_AlternateLBA = LE_64(1ULL);
1555         efi->efi_gpt_MyLBA = LE_64(lba_backup_gpt_hdr);
1556         efi->efi_gpt_PartitionEntryLBA = LE_64(vtoc->efi_last_u_lba + 1);
1557         efi->efi_gpt_HeaderCRC32 = 0;
1558         efi->efi_gpt_HeaderCRC32 =
1559             LE_32(efi_crc32((unsigned char *)dk_ioc.dki_data,
1560             LE_32(efi->efi_gpt_HeaderSize)));
1561
1562         if (efi_ioctl(fd, DKIOCSETEFI, &dk_ioc) == -1) {
1563                 if (efi_debug) {
1564                         (void) fprintf(stderr,
1565                             "write of backup header to block %llu failed, "
1566                             "errno %d\n",
1567                             lba_backup_gpt_hdr,
1568                             errno);
1569                 }
1570         }
1571         /* write the PMBR */
1572         (void) write_pmbr(fd, vtoc);
1573         free(dk_ioc.dki_data);
1574
1575         return (0);
1576 }
1577
1578 void
1579 efi_free(struct dk_gpt *ptr)
1580 {
1581         free(ptr);
1582 }
1583
1584 /*
1585  * Input: File descriptor
1586  * Output: 1 if disk has an EFI label, or > 2TB with no VTOC or legacy MBR.
1587  * Otherwise 0.
1588  */
1589 int
1590 efi_type(int fd)
1591 {
1592 #if 0
1593         struct vtoc vtoc;
1594         struct extvtoc extvtoc;
1595
1596         if (ioctl(fd, DKIOCGEXTVTOC, &extvtoc) == -1) {
1597                 if (errno == ENOTSUP)
1598                         return (1);
1599                 else if (errno == ENOTTY) {
1600                         if (ioctl(fd, DKIOCGVTOC, &vtoc) == -1)
1601                                 if (errno == ENOTSUP)
1602                                         return (1);
1603                 }
1604         }
1605         return (0);
1606 #else
1607         return (ENOSYS);
1608 #endif
1609 }
1610
1611 void
1612 efi_err_check(struct dk_gpt *vtoc)
1613 {
1614         int                     resv_part = -1;
1615         int                     i, j;
1616         diskaddr_t              istart, jstart, isize, jsize, endsect;
1617         int                     overlap = 0;
1618
1619         /*
1620          * make sure no partitions overlap
1621          */
1622         for (i = 0; i < vtoc->efi_nparts; i++) {
1623                 /* It can't be unassigned and have an actual size */
1624                 if ((vtoc->efi_parts[i].p_tag == V_UNASSIGNED) &&
1625                     (vtoc->efi_parts[i].p_size != 0)) {
1626                         (void) fprintf(stderr,
1627                             "partition %d is \"unassigned\" but has a size "
1628                             "of %llu\n", i, vtoc->efi_parts[i].p_size);
1629                 }
1630                 if (vtoc->efi_parts[i].p_tag == V_UNASSIGNED) {
1631                         continue;
1632                 }
1633                 if (vtoc->efi_parts[i].p_tag == V_RESERVED) {
1634                         if (resv_part != -1) {
1635                                 (void) fprintf(stderr,
1636                                     "found duplicate reserved partition at "
1637                                     "%d\n", i);
1638                         }
1639                         resv_part = i;
1640                         if (vtoc->efi_parts[i].p_size != EFI_MIN_RESV_SIZE)
1641                                 (void) fprintf(stderr,
1642                                     "Warning: reserved partition size must "
1643                                     "be %d sectors\n", EFI_MIN_RESV_SIZE);
1644                 }
1645                 if ((vtoc->efi_parts[i].p_start < vtoc->efi_first_u_lba) ||
1646                     (vtoc->efi_parts[i].p_start > vtoc->efi_last_u_lba)) {
1647                         (void) fprintf(stderr,
1648                             "Partition %d starts at %llu\n",
1649                             i,
1650                             vtoc->efi_parts[i].p_start);
1651                         (void) fprintf(stderr,
1652                             "It must be between %llu and %llu.\n",
1653                             vtoc->efi_first_u_lba,
1654                             vtoc->efi_last_u_lba);
1655                 }
1656                 if ((vtoc->efi_parts[i].p_start +
1657                     vtoc->efi_parts[i].p_size <
1658                     vtoc->efi_first_u_lba) ||
1659                     (vtoc->efi_parts[i].p_start +
1660                     vtoc->efi_parts[i].p_size >
1661                     vtoc->efi_last_u_lba + 1)) {
1662                         (void) fprintf(stderr,
1663                             "Partition %d ends at %llu\n",
1664                             i,
1665                             vtoc->efi_parts[i].p_start +
1666                             vtoc->efi_parts[i].p_size);
1667                         (void) fprintf(stderr,
1668                             "It must be between %llu and %llu.\n",
1669                             vtoc->efi_first_u_lba,
1670                             vtoc->efi_last_u_lba);
1671                 }
1672
1673                 for (j = 0; j < vtoc->efi_nparts; j++) {
1674                         isize = vtoc->efi_parts[i].p_size;
1675                         jsize = vtoc->efi_parts[j].p_size;
1676                         istart = vtoc->efi_parts[i].p_start;
1677                         jstart = vtoc->efi_parts[j].p_start;
1678                         if ((i != j) && (isize != 0) && (jsize != 0)) {
1679                                 endsect = jstart + jsize -1;
1680                                 if ((jstart <= istart) &&
1681                                     (istart <= endsect)) {
1682                                         if (!overlap) {
1683                                         (void) fprintf(stderr,
1684                                             "label error: EFI Labels do not "
1685                                             "support overlapping partitions\n");
1686                                         }
1687                                         (void) fprintf(stderr,
1688                                             "Partition %d overlaps partition "
1689                                             "%d.\n", i, j);
1690                                         overlap = 1;
1691                                 }
1692                         }
1693                 }
1694         }
1695         /* make sure there is a reserved partition */
1696         if (resv_part == -1) {
1697                 (void) fprintf(stderr,
1698                     "no reserved partition found\n");
1699         }
1700 }
1701
1702 /*
1703  * We need to get information necessary to construct a *new* efi
1704  * label type
1705  */
1706 int
1707 efi_auto_sense(int fd, struct dk_gpt **vtoc)
1708 {
1709
1710         int     i;
1711
1712         /*
1713          * Now build the default partition table
1714          */
1715         if (efi_alloc_and_init(fd, EFI_NUMPAR, vtoc) != 0) {
1716                 if (efi_debug) {
1717                         (void) fprintf(stderr, "efi_alloc_and_init failed.\n");
1718                 }
1719                 return (-1);
1720         }
1721
1722         for (i = 0; i < MIN((*vtoc)->efi_nparts, V_NUMPAR); i++) {
1723                 (*vtoc)->efi_parts[i].p_tag = default_vtoc_map[i].p_tag;
1724                 (*vtoc)->efi_parts[i].p_flag = default_vtoc_map[i].p_flag;
1725                 (*vtoc)->efi_parts[i].p_start = 0;
1726                 (*vtoc)->efi_parts[i].p_size = 0;
1727         }
1728         /*
1729          * Make constants first
1730          * and variable partitions later
1731          */
1732
1733         /* root partition - s0 128 MB */
1734         (*vtoc)->efi_parts[0].p_start = 34;
1735         (*vtoc)->efi_parts[0].p_size = 262144;
1736
1737         /* partition - s1  128 MB */
1738         (*vtoc)->efi_parts[1].p_start = 262178;
1739         (*vtoc)->efi_parts[1].p_size = 262144;
1740
1741         /* partition -s2 is NOT the Backup disk */
1742         (*vtoc)->efi_parts[2].p_tag = V_UNASSIGNED;
1743
1744         /* partition -s6 /usr partition - HOG */
1745         (*vtoc)->efi_parts[6].p_start = 524322;
1746         (*vtoc)->efi_parts[6].p_size = (*vtoc)->efi_last_u_lba - 524322
1747             - (1024 * 16);
1748
1749         /* efi reserved partition - s9 16K */
1750         (*vtoc)->efi_parts[8].p_start = (*vtoc)->efi_last_u_lba - (1024 * 16);
1751         (*vtoc)->efi_parts[8].p_size = (1024 * 16);
1752         (*vtoc)->efi_parts[8].p_tag = V_RESERVED;
1753         return (0);
1754 }