arch/x86/kernel/kexec-bzimage64.c

   1 /*
   2  * Kexec bzImage loader
   3  *
   4  * Copyright (C) 2014 Red Hat Inc.
   5  * Authors:
   6  *      Vivek Goyal <vgoyal@redhat.com>
   7  *
   8  * This source code is licensed under the GNU General Public License,
   9  * Version 2.  See the file COPYING for more details.
  10  */
  11
  12 #define pr_fmt(fmt)     "kexec-bzImage64: " fmt
  13
  14 #include <linux/string.h>
  15 #include <linux/printk.h>
  16 #include <linux/errno.h>
  17 #include <linux/slab.h>
  18 #include <linux/kexec.h>
  19 #include <linux/kernel.h>
  20 #include <linux/mm.h>
  21 #include <linux/efi.h>
  22 #include <linux/verification.h>
  23
  24 #include <asm/bootparam.h>
  25 #include <asm/setup.h>
  26 #include <asm/crash.h>
  27 #include <asm/efi.h>
  28 #include <asm/e820/api.h>
  29 #include <asm/kexec-bzimage64.h>
  30
  31 #define MAX_ELFCOREHDR_STR_LEN  30      /* elfcorehdr=0x<64bit-value> */
  32
  33 /*
  34  * Defines lowest physical address for various segments. Not sure where
  35  * exactly these limits came from. Current bzimage64 loader in kexec-tools
  36  * uses these so I am retaining it. It can be changed over time as we gain
  37  * more insight.
  38  */
  39 #define MIN_PURGATORY_ADDR      0x3000
  40 #define MIN_BOOTPARAM_ADDR      0x3000
  41 #define MIN_KERNEL_LOAD_ADDR    0x100000
  42 #define MIN_INITRD_LOAD_ADDR    0x1000000
  43
  44 /*
  45  * This is a place holder for all boot loader specific data structure which
  46  * gets allocated in one call but gets freed much later during cleanup
  47  * time. Right now there is only one field but it can grow as need be.
  48  */
  49 struct bzimage64_data {
  50         /*
  51          * Temporary buffer to hold bootparams buffer. This should be
  52          * freed once the bootparam segment has been loaded.
  53          */
  54         void *bootparams_buf;
  55 };
  56
  57 static int setup_initrd(struct boot_params *params,
  58                 unsigned long initrd_load_addr, unsigned long initrd_len)
  59 {
  60         params->hdr.ramdisk_image = initrd_load_addr & 0xffffffffUL;
  61         params->hdr.ramdisk_size = initrd_len & 0xffffffffUL;
  62
  63         params->ext_ramdisk_image = initrd_load_addr >> 32;
  64         params->ext_ramdisk_size = initrd_len >> 32;
  65
  66         return 0;
  67 }
  68
  69 static int setup_cmdline(struct kimage *image, struct boot_params *params,
  70                          unsigned long bootparams_load_addr,
  71                          unsigned long cmdline_offset, char *cmdline,
  72                          unsigned long cmdline_len)
  73 {
  74         char *cmdline_ptr = ((char *)params) + cmdline_offset;
  75         unsigned long cmdline_ptr_phys, len = 0;
  76         uint32_t cmdline_low_32, cmdline_ext_32;
  77
  78         if (image->type == KEXEC_TYPE_CRASH) {
  79                 len = sprintf(cmdline_ptr,
  80                         "elfcorehdr=0x%lx ", image->arch.elf_load_addr);
  81         }
  82         memcpy(cmdline_ptr + len, cmdline, cmdline_len);
  83         cmdline_len += len;
  84
  85         cmdline_ptr[cmdline_len - 1] = '\0';
  86
  87         pr_debug("Final command line is: %s\n", cmdline_ptr);
  88         cmdline_ptr_phys = bootparams_load_addr + cmdline_offset;
  89         cmdline_low_32 = cmdline_ptr_phys & 0xffffffffUL;
  90         cmdline_ext_32 = cmdline_ptr_phys >> 32;
  91
  92         params->hdr.cmd_line_ptr = cmdline_low_32;
  93         if (cmdline_ext_32)
  94                 params->ext_cmd_line_ptr = cmdline_ext_32;
  95
  96         return 0;
  97 }
  98
  99 static int setup_e820_entries(struct boot_params *params)
 100 {
 101         unsigned int nr_e820_entries;
 102
 103         nr_e820_entries = e820_table_kexec->nr_entries;
 104
 105         /* TODO: Pass entries more than E820_MAX_ENTRIES_ZEROPAGE in bootparams setup data */
 106         if (nr_e820_entries > E820_MAX_ENTRIES_ZEROPAGE)
 107                 nr_e820_entries = E820_MAX_ENTRIES_ZEROPAGE;
 108
 109         params->e820_entries = nr_e820_entries;
 110         memcpy(&params->e820_table, &e820_table_kexec->entries, nr_e820_entries*sizeof(struct e820_entry));
 111
 112         return 0;
 113 }
 114
 115 #ifdef CONFIG_EFI
 116 static int setup_efi_info_memmap(struct boot_params *params,
 117                                   unsigned long params_load_addr,
 118                                   unsigned int efi_map_offset,
 119                                   unsigned int efi_map_sz)
 120 {
 121         void *efi_map = (void *)params + efi_map_offset;
 122         unsigned long efi_map_phys_addr = params_load_addr + efi_map_offset;
 123         struct efi_info *ei = &params->efi_info;
 124
 125         if (!efi_map_sz)
 126                 return 0;
 127
 128         efi_runtime_map_copy(efi_map, efi_map_sz);
 129
 130         ei->efi_memmap = efi_map_phys_addr & 0xffffffff;
 131         ei->efi_memmap_hi = efi_map_phys_addr >> 32;
 132         ei->efi_memmap_size = efi_map_sz;
 133
 134         return 0;
 135 }
 136
 137 static int
 138 prepare_add_efi_setup_data(struct boot_params *params,
 139                        unsigned long params_load_addr,
 140                        unsigned int efi_setup_data_offset)
 141 {
 142         unsigned long setup_data_phys;
 143         struct setup_data *sd = (void *)params + efi_setup_data_offset;
 144         struct efi_setup_data *esd = (void *)sd + sizeof(struct setup_data);
 145
 146         esd->fw_vendor = efi.fw_vendor;
 147         esd->runtime = efi.runtime;
 148         esd->tables = efi.config_table;
 149         esd->smbios = efi.smbios;
 150
 151         sd->type = SETUP_EFI;
 152         sd->len = sizeof(struct efi_setup_data);
 153
 154         /* Add setup data */
 155         setup_data_phys = params_load_addr + efi_setup_data_offset;
 156         sd->next = params->hdr.setup_data;
 157         params->hdr.setup_data = setup_data_phys;
 158
 159         return 0;
 160 }
 161
 162 static int
 163 setup_efi_state(struct boot_params *params, unsigned long params_load_addr,
 164                 unsigned int efi_map_offset, unsigned int efi_map_sz,
 165                 unsigned int efi_setup_data_offset)
 166 {
 167         struct efi_info *current_ei = &boot_params.efi_info;
 168         struct efi_info *ei = &params->efi_info;
 169
 170         if (!efi_enabled(EFI_RUNTIME_SERVICES))
 171                 return 0;
 172
 173         if (!current_ei->efi_memmap_size)
 174                 return 0;
 175
 176         /*
 177          * If 1:1 mapping is not enabled, second kernel can not setup EFI
 178          * and use EFI run time services. User space will have to pass
 179          * acpi_rsdp=<addr> on kernel command line to make second kernel boot
 180          * without efi.
 181          */
 182         if (efi_enabled(EFI_OLD_MEMMAP))
 183                 return 0;
 184
 185         ei->efi_loader_signature = current_ei->efi_loader_signature;
 186         ei->efi_systab = current_ei->efi_systab;
 187         ei->efi_systab_hi = current_ei->efi_systab_hi;
 188
 189         ei->efi_memdesc_version = current_ei->efi_memdesc_version;
 190         ei->efi_memdesc_size = efi_get_runtime_map_desc_size();
 191
 192         setup_efi_info_memmap(params, params_load_addr, efi_map_offset,
 193                               efi_map_sz);
 194         prepare_add_efi_setup_data(params, params_load_addr,
 195                                    efi_setup_data_offset);
 196         return 0;
 197 }
 198 #endif /* CONFIG_EFI */
 199
 200 static int
 201 setup_boot_parameters(struct kimage *image, struct boot_params *params,
 202                       unsigned long params_load_addr,
 203                       unsigned int efi_map_offset, unsigned int efi_map_sz,
 204                       unsigned int efi_setup_data_offset)
 205 {
 206         unsigned int nr_e820_entries;
 207         unsigned long long mem_k, start, end;
 208         int i, ret = 0;
 209
 210         /* Get subarch from existing bootparams */
 211         params->hdr.hardware_subarch = boot_params.hdr.hardware_subarch;
 212
 213         /* Copying screen_info will do? */
 214         memcpy(&params->screen_info, &boot_params.screen_info,
 215                                 sizeof(struct screen_info));
 216
 217         /* Fill in memsize later */
 218         params->screen_info.ext_mem_k = 0;
 219         params->alt_mem_k = 0;
 220
 221         /* Always fill in RSDP: it is either 0 or a valid value */
 222         params->acpi_rsdp_addr = boot_params.acpi_rsdp_addr;
 223
 224         /* Default APM info */
 225         memset(&params->apm_bios_info, 0, sizeof(params->apm_bios_info));
 226
 227         /* Default drive info */
 228         memset(&params->hd0_info, 0, sizeof(params->hd0_info));
 229         memset(&params->hd1_info, 0, sizeof(params->hd1_info));
 230
 231         if (image->type == KEXEC_TYPE_CRASH) {
 232                 ret = crash_setup_memmap_entries(image, params);
 233                 if (ret)
 234                         return ret;
 235         } else
 236                 setup_e820_entries(params);
 237
 238         nr_e820_entries = params->e820_entries;
 239
 240         for (i = 0; i < nr_e820_entries; i++) {
 241                 if (params->e820_table[i].type != E820_TYPE_RAM)
 242                         continue;
 243                 start = params->e820_table[i].addr;
 244                 end = params->e820_table[i].addr + params->e820_table[i].size - 1;
 245
 246                 if ((start <= 0x100000) && end > 0x100000) {
 247                         mem_k = (end >> 10) - (0x100000 >> 10);
 248                         params->screen_info.ext_mem_k = mem_k;
 249                         params->alt_mem_k = mem_k;
 250                         if (mem_k > 0xfc00)
 251                                 params->screen_info.ext_mem_k = 0xfc00; /* 64M*/
 252                         if (mem_k > 0xffffffff)
 253                                 params->alt_mem_k = 0xffffffff;
 254                 }
 255         }
 256
 257 #ifdef CONFIG_EFI
 258         /* Setup EFI state */
 259         setup_efi_state(params, params_load_addr, efi_map_offset, efi_map_sz,
 260                         efi_setup_data_offset);
 261 #endif
 262         /* Setup EDD info */
 263         memcpy(params->eddbuf, boot_params.eddbuf,
 264                                 EDDMAXNR * sizeof(struct edd_info));
 265         params->eddbuf_entries = boot_params.eddbuf_entries;
 266
 267         memcpy(params->edd_mbr_sig_buffer, boot_params.edd_mbr_sig_buffer,
 268                EDD_MBR_SIG_MAX * sizeof(unsigned int));
 269
 270         return ret;
 271 }
 272
 273 static int bzImage64_probe(const char *buf, unsigned long len)
 274 {
 275         int ret = -ENOEXEC;
 276         struct setup_header *header;
 277
 278         /* kernel should be at least two sectors long */
 279         if (len < 2 * 512) {
 280                 pr_err("File is too short to be a bzImage\n");
 281                 return ret;
 282         }
 283
 284         header = (struct setup_header *)(buf + offsetof(struct boot_params, hdr));
 285         if (memcmp((char *)&header->header, "HdrS", 4) != 0) {
 286                 pr_err("Not a bzImage\n");
 287                 return ret;
 288         }
 289
 290         if (header->boot_flag != 0xAA55) {
 291                 pr_err("No x86 boot sector present\n");
 292                 return ret;
 293         }
 294
 295         if (header->version < 0x020C) {
 296                 pr_err("Must be at least protocol version 2.12\n");
 297                 return ret;
 298         }
 299
 300         if (!(header->loadflags & LOADED_HIGH)) {
 301                 pr_err("zImage not a bzImage\n");
 302                 return ret;
 303         }
 304
 305         if (!(header->xloadflags & XLF_KERNEL_64)) {
 306                 pr_err("Not a bzImage64. XLF_KERNEL_64 is not set.\n");
 307                 return ret;
 308         }
 309
 310         if (!(header->xloadflags & XLF_CAN_BE_LOADED_ABOVE_4G)) {
 311                 pr_err("XLF_CAN_BE_LOADED_ABOVE_4G is not set.\n");
 312                 return ret;
 313         }
 314
 315         /*
 316          * Can't handle 32bit EFI as it does not allow loading kernel
 317          * above 4G. This should be handled by 32bit bzImage loader
 318          */
 319         if (efi_enabled(EFI_RUNTIME_SERVICES) && !efi_enabled(EFI_64BIT)) {
 320                 pr_debug("EFI is 32 bit. Can't load kernel above 4G.\n");
 321                 return ret;
 322         }
 323
 324         /* I've got a bzImage */
 325         pr_debug("It's a relocatable bzImage64\n");
 326         ret = 0;
 327
 328         return ret;
 329 }
 330
 331 static void *bzImage64_load(struct kimage *image, char *kernel,
 332                             unsigned long kernel_len, char *initrd,
 333                             unsigned long initrd_len, char *cmdline,
 334                             unsigned long cmdline_len)
 335 {
 336
 337         struct setup_header *header;
 338         int setup_sects, kern16_size, ret = 0;
 339         unsigned long setup_header_size, params_cmdline_sz;
 340         struct boot_params *params;
 341         unsigned long bootparam_load_addr, kernel_load_addr, initrd_load_addr;
 342         struct bzimage64_data *ldata;
 343         struct kexec_entry64_regs regs64;
 344         void *stack;
 345         unsigned int setup_hdr_offset = offsetof(struct boot_params, hdr);
 346         unsigned int efi_map_offset, efi_map_sz, efi_setup_data_offset;
 347         struct kexec_buf kbuf = { .image = image, .buf_max = ULONG_MAX,
 348                                   .top_down = true };
 349         struct kexec_buf pbuf = { .image = image, .buf_min = MIN_PURGATORY_ADDR,
 350                                   .buf_max = ULONG_MAX, .top_down = true };
 351
 352         header = (struct setup_header *)(kernel + setup_hdr_offset);
 353         setup_sects = header->setup_sects;
 354         if (setup_sects == 0)
 355                 setup_sects = 4;
 356
 357         kern16_size = (setup_sects + 1) * 512;
 358         if (kernel_len < kern16_size) {
 359                 pr_err("bzImage truncated\n");
 360                 return ERR_PTR(-ENOEXEC);
 361         }
 362
 363         if (cmdline_len > header->cmdline_size) {
 364                 pr_err("Kernel command line too long\n");
 365                 return ERR_PTR(-EINVAL);
 366         }
 367
 368         /*
 369          * In case of crash dump, we will append elfcorehdr=<addr> to
 370          * command line. Make sure it does not overflow
 371          */
 372         if (cmdline_len + MAX_ELFCOREHDR_STR_LEN > header->cmdline_size) {
 373                 pr_debug("Appending elfcorehdr=<addr> to command line exceeds maximum allowed length\n");
 374                 return ERR_PTR(-EINVAL);
 375         }
 376
 377         /* Allocate and load backup region */
 378         if (image->type == KEXEC_TYPE_CRASH) {
 379                 ret = crash_load_segments(image);
 380                 if (ret)
 381                         return ERR_PTR(ret);
 382         }
 383
 384         /*
 385          * Load purgatory. For 64bit entry point, purgatory  code can be
 386          * anywhere.
 387          */
 388         ret = kexec_load_purgatory(image, &pbuf);
 389         if (ret) {
 390                 pr_err("Loading purgatory failed\n");
 391                 return ERR_PTR(ret);
 392         }
 393
 394         pr_debug("Loaded purgatory at 0x%lx\n", pbuf.mem);
 395
 396
 397         /*
 398          * Load Bootparams and cmdline and space for efi stuff.
 399          *
 400          * Allocate memory together for multiple data structures so
 401          * that they all can go in single area/segment and we don't
 402          * have to create separate segment for each. Keeps things
 403          * little bit simple
 404          */
 405         efi_map_sz = efi_get_runtime_map_size();
 406         params_cmdline_sz = sizeof(struct boot_params) + cmdline_len +
 407                                 MAX_ELFCOREHDR_STR_LEN;
 408         params_cmdline_sz = ALIGN(params_cmdline_sz, 16);
 409         kbuf.bufsz = params_cmdline_sz + ALIGN(efi_map_sz, 16) +
 410                                 sizeof(struct setup_data) +
 411                                 sizeof(struct efi_setup_data);
 412
 413         params = kzalloc(kbuf.bufsz, GFP_KERNEL);
 414         if (!params)
 415                 return ERR_PTR(-ENOMEM);
 416         efi_map_offset = params_cmdline_sz;
 417         efi_setup_data_offset = efi_map_offset + ALIGN(efi_map_sz, 16);
 418
 419         /* Copy setup header onto bootparams. Documentation/x86/boot.txt */
 420         setup_header_size = 0x0202 + kernel[0x0201] - setup_hdr_offset;
 421
 422         /* Is there a limit on setup header size? */
 423         memcpy(&params->hdr, (kernel + setup_hdr_offset), setup_header_size);
 424
 425         kbuf.buffer = params;
 426         kbuf.memsz = kbuf.bufsz;
 427         kbuf.buf_align = 16;
 428         kbuf.buf_min = MIN_BOOTPARAM_ADDR;
 429         ret = kexec_add_buffer(&kbuf);
 430         if (ret)
 431                 goto out_free_params;
 432         bootparam_load_addr = kbuf.mem;
 433         pr_debug("Loaded boot_param, command line and misc at 0x%lx bufsz=0x%lx memsz=0x%lx\n",
 434                  bootparam_load_addr, kbuf.bufsz, kbuf.bufsz);
 435
 436         /* Load kernel */
 437         kbuf.buffer = kernel + kern16_size;
 438         kbuf.bufsz =  kernel_len - kern16_size;
 439         kbuf.memsz = PAGE_ALIGN(header->init_size);
 440         kbuf.buf_align = header->kernel_alignment;
 441         kbuf.buf_min = MIN_KERNEL_LOAD_ADDR;
 442         kbuf.mem = KEXEC_BUF_MEM_UNKNOWN;
 443         ret = kexec_add_buffer(&kbuf);
 444         if (ret)
 445                 goto out_free_params;
 446         kernel_load_addr = kbuf.mem;
 447
 448         pr_debug("Loaded 64bit kernel at 0x%lx bufsz=0x%lx memsz=0x%lx\n",
 449                  kernel_load_addr, kbuf.bufsz, kbuf.memsz);
 450
 451         /* Load initrd high */
 452         if (initrd) {
 453                 kbuf.buffer = initrd;
 454                 kbuf.bufsz = kbuf.memsz = initrd_len;
 455                 kbuf.buf_align = PAGE_SIZE;
 456                 kbuf.buf_min = MIN_INITRD_LOAD_ADDR;
 457                 kbuf.mem = KEXEC_BUF_MEM_UNKNOWN;
 458                 ret = kexec_add_buffer(&kbuf);
 459                 if (ret)
 460                         goto out_free_params;
 461                 initrd_load_addr = kbuf.mem;
 462
 463                 pr_debug("Loaded initrd at 0x%lx bufsz=0x%lx memsz=0x%lx\n",
 464                                 initrd_load_addr, initrd_len, initrd_len);
 465
 466                 setup_initrd(params, initrd_load_addr, initrd_len);
 467         }
 468
 469         setup_cmdline(image, params, bootparam_load_addr,
 470                       sizeof(struct boot_params), cmdline, cmdline_len);
 471
 472         /* bootloader info. Do we need a separate ID for kexec kernel loader? */
 473         params->hdr.type_of_loader = 0x0D << 4;
 474         params->hdr.loadflags = 0;
 475
 476         /* Setup purgatory regs for entry */
 477         ret = kexec_purgatory_get_set_symbol(image, "entry64_regs", &regs64,
 478                                              sizeof(regs64), 1);
 479         if (ret)
 480                 goto out_free_params;
 481
 482         regs64.rbx = 0; /* Bootstrap Processor */
 483         regs64.rsi = bootparam_load_addr;
 484         regs64.rip = kernel_load_addr + 0x200;
 485         stack = kexec_purgatory_get_symbol_addr(image, "stack_end");
 486         if (IS_ERR(stack)) {
 487                 pr_err("Could not find address of symbol stack_end\n");
 488                 ret = -EINVAL;
 489                 goto out_free_params;
 490         }
 491
 492         regs64.rsp = (unsigned long)stack;
 493         ret = kexec_purgatory_get_set_symbol(image, "entry64_regs", &regs64,
 494                                              sizeof(regs64), 0);
 495         if (ret)
 496                 goto out_free_params;
 497
 498         ret = setup_boot_parameters(image, params, bootparam_load_addr,
 499                                     efi_map_offset, efi_map_sz,
 500                                     efi_setup_data_offset);
 501         if (ret)
 502                 goto out_free_params;
 503
 504         /* Allocate loader specific data */
 505         ldata = kzalloc(sizeof(struct bzimage64_data), GFP_KERNEL);
 506         if (!ldata) {
 507                 ret = -ENOMEM;
 508                 goto out_free_params;
 509         }
 510
 511         /*
 512          * Store pointer to params so that it could be freed after loading
 513          * params segment has been loaded and contents have been copied
 514          * somewhere else.
 515          */
 516         ldata->bootparams_buf = params;
 517         return ldata;
 518
 519 out_free_params:
 520         kfree(params);
 521         return ERR_PTR(ret);
 522 }
 523
 524 /* This cleanup function is called after various segments have been loaded */
 525 static int bzImage64_cleanup(void *loader_data)
 526 {
 527         struct bzimage64_data *ldata = loader_data;
 528
 529         if (!ldata)
 530                 return 0;
 531
 532         kfree(ldata->bootparams_buf);
 533         ldata->bootparams_buf = NULL;
 534
 535         return 0;
 536 }
 537
 538 #ifdef CONFIG_KEXEC_BZIMAGE_VERIFY_SIG
 539 static int bzImage64_verify_sig(const char *kernel, unsigned long kernel_len)
 540 {
 541         return verify_pefile_signature(kernel, kernel_len,
 542                                        VERIFY_USE_SECONDARY_KEYRING,
 543                                        VERIFYING_KEXEC_PE_SIGNATURE);
 544 }
 545 #endif
 546
 547 const struct kexec_file_ops kexec_bzImage64_ops = {
 548         .probe = bzImage64_probe,
 549         .load = bzImage64_load,
 550         .cleanup = bzImage64_cleanup,
 551 #ifdef CONFIG_KEXEC_BZIMAGE_VERIFY_SIG
 552         .verify_sig = bzImage64_verify_sig,
 553 #endif
 554 };