arch/x86/boot/compressed/kaslr.c

   1 /*
   2  * kaslr.c
   3  *
   4  * This contains the routines needed to generate a reasonable level of
   5  * entropy to choose a randomized kernel base address offset in support
   6  * of Kernel Address Space Layout Randomization (KASLR). Additionally
   7  * handles walking the physical memory maps (and tracking memory regions
   8  * to avoid) in order to select a physical memory location that can
   9  * contain the entire properly aligned running kernel image.
  10  *
  11  */
  12 #include "misc.h"
  13 #include "error.h"
  14
  15 #include <asm/msr.h>
  16 #include <asm/archrandom.h>
  17 #include <asm/e820.h>
  18
  19 #include <generated/compile.h>
  20 #include <linux/module.h>
  21 #include <linux/uts.h>
  22 #include <linux/utsname.h>
  23 #include <generated/utsrelease.h>
  24
  25 /* Simplified build-specific string for starting entropy. */
  26 static const char build_str[] = UTS_RELEASE " (" LINUX_COMPILE_BY "@"
  27                 LINUX_COMPILE_HOST ") (" LINUX_COMPILER ") " UTS_VERSION;
  28
  29 #define I8254_PORT_CONTROL      0x43
  30 #define I8254_PORT_COUNTER0     0x40
  31 #define I8254_CMD_READBACK      0xC0
  32 #define I8254_SELECT_COUNTER0   0x02
  33 #define I8254_STATUS_NOTREADY   0x40
  34 static inline u16 i8254(void)
  35 {
  36         u16 status, timer;
  37
  38         do {
  39                 outb(I8254_PORT_CONTROL,
  40                      I8254_CMD_READBACK | I8254_SELECT_COUNTER0);
  41                 status = inb(I8254_PORT_COUNTER0);
  42                 timer  = inb(I8254_PORT_COUNTER0);
  43                 timer |= inb(I8254_PORT_COUNTER0) << 8;
  44         } while (status & I8254_STATUS_NOTREADY);
  45
  46         return timer;
  47 }
  48
  49 static unsigned long rotate_xor(unsigned long hash, const void *area,
  50                                 size_t size)
  51 {
  52         size_t i;
  53         unsigned long *ptr = (unsigned long *)area;
  54
  55         for (i = 0; i < size / sizeof(hash); i++) {
  56                 /* Rotate by odd number of bits and XOR. */
  57                 hash = (hash << ((sizeof(hash) * 8) - 7)) | (hash >> 7);
  58                 hash ^= ptr[i];
  59         }
  60
  61         return hash;
  62 }
  63
  64 /* Attempt to create a simple but unpredictable starting entropy. */
  65 static unsigned long get_random_boot(void)
  66 {
  67         unsigned long hash = 0;
  68
  69         hash = rotate_xor(hash, build_str, sizeof(build_str));
  70         hash = rotate_xor(hash, boot_params, sizeof(*boot_params));
  71
  72         return hash;
  73 }
  74
  75 static unsigned long get_random_long(void)
  76 {
  77 #ifdef CONFIG_X86_64
  78         const unsigned long mix_const = 0x5d6008cbf3848dd3UL;
  79 #else
  80         const unsigned long mix_const = 0x3f39e593UL;
  81 #endif
  82         unsigned long raw, random = get_random_boot();
  83         bool use_i8254 = true;
  84
  85         debug_putstr("KASLR using");
  86
  87         if (has_cpuflag(X86_FEATURE_RDRAND)) {
  88                 debug_putstr(" RDRAND");
  89                 if (rdrand_long(&raw)) {
  90                         random ^= raw;
  91                         use_i8254 = false;
  92                 }
  93         }
  94
  95         if (has_cpuflag(X86_FEATURE_TSC)) {
  96                 debug_putstr(" RDTSC");
  97                 raw = rdtsc();
  98
  99                 random ^= raw;
 100                 use_i8254 = false;
 101         }
 102
 103         if (use_i8254) {
 104                 debug_putstr(" i8254");
 105                 random ^= i8254();
 106         }
 107
 108         /* Circular multiply for better bit diffusion */
 109         asm("mul %3"
 110             : "=a" (random), "=d" (raw)
 111             : "a" (random), "rm" (mix_const));
 112         random += raw;
 113
 114         debug_putstr("...\n");
 115
 116         return random;
 117 }
 118
 119 struct mem_vector {
 120         unsigned long start;
 121         unsigned long size;
 122 };
 123
 124 enum mem_avoid_index {
 125         MEM_AVOID_ZO_RANGE = 0,
 126         MEM_AVOID_INITRD,
 127         MEM_AVOID_CMDLINE,
 128         MEM_AVOID_BOOTPARAMS,
 129         MEM_AVOID_MAX,
 130 };
 131
 132 static struct mem_vector mem_avoid[MEM_AVOID_MAX];
 133
 134 static bool mem_contains(struct mem_vector *region, struct mem_vector *item)
 135 {
 136         /* Item at least partially before region. */
 137         if (item->start < region->start)
 138                 return false;
 139         /* Item at least partially after region. */
 140         if (item->start + item->size > region->start + region->size)
 141                 return false;
 142         return true;
 143 }
 144
 145 static bool mem_overlaps(struct mem_vector *one, struct mem_vector *two)
 146 {
 147         /* Item one is entirely before item two. */
 148         if (one->start + one->size <= two->start)
 149                 return false;
 150         /* Item one is entirely after item two. */
 151         if (one->start >= two->start + two->size)
 152                 return false;
 153         return true;
 154 }
 155
 156 /*
 157  * In theory, KASLR can put the kernel anywhere in the range of [16M, 64T).
 158  * The mem_avoid array is used to store the ranges that need to be avoided
 159  * when KASLR searches for an appropriate random address. We must avoid any
 160  * regions that are unsafe to overlap with during decompression, and other
 161  * things like the initrd, cmdline and boot_params. This comment seeks to
 162  * explain mem_avoid as clearly as possible since incorrect mem_avoid
 163  * memory ranges lead to really hard to debug boot failures.
 164  *
 165  * The initrd, cmdline, and boot_params are trivial to identify for
 166  * avoiding. The are MEM_AVOID_INITRD, MEM_AVOID_CMDLINE, and
 167  * MEM_AVOID_BOOTPARAMS respectively below.
 168  *
 169  * What is not obvious how to avoid is the range of memory that is used
 170  * during decompression (MEM_AVOID_ZO_RANGE below). This range must cover
 171  * the compressed kernel (ZO) and its run space, which is used to extract
 172  * the uncompressed kernel (VO) and relocs.
 173  *
 174  * ZO's full run size sits against the end of the decompression buffer, so
 175  * we can calculate where text, data, bss, etc of ZO are positioned more
 176  * easily.
 177  *
 178  * For additional background, the decompression calculations can be found
 179  * in header.S, and the memory diagram is based on the one found in misc.c.
 180  *
 181  * The following conditions are already enforced by the image layouts and
 182  * associated code:
 183  *  - input + input_size >= output + output_size
 184  *  - kernel_total_size <= init_size
 185  *  - kernel_total_size <= output_size (see Note below)
 186  *  - output + init_size >= output + output_size
 187  *
 188  * (Note that kernel_total_size and output_size have no fundamental
 189  * relationship, but output_size is passed to choose_random_location
 190  * as a maximum of the two. The diagram is showing a case where
 191  * kernel_total_size is larger than output_size, but this case is
 192  * handled by bumping output_size.)
 193  *
 194  * The above conditions can be illustrated by a diagram:
 195  *
 196  * 0   output            input            input+input_size    output+init_size
 197  * |     |                 |                             |             |
 198  * |     |                 |                             |             |
 199  * |-----|--------|--------|--------------|-----------|--|-------------|
 200  *                |                       |           |
 201  *                |                       |           |
 202  * output+init_size-ZO_INIT_SIZE  output+output_size  output+kernel_total_size
 203  *
 204  * [output, output+init_size) is the entire memory range used for
 205  * extracting the compressed image.
 206  *
 207  * [output, output+kernel_total_size) is the range needed for the
 208  * uncompressed kernel (VO) and its run size (bss, brk, etc).
 209  *
 210  * [output, output+output_size) is VO plus relocs (i.e. the entire
 211  * uncompressed payload contained by ZO). This is the area of the buffer
 212  * written to during decompression.
 213  *
 214  * [output+init_size-ZO_INIT_SIZE, output+init_size) is the worst-case
 215  * range of the copied ZO and decompression code. (i.e. the range
 216  * covered backwards of size ZO_INIT_SIZE, starting from output+init_size.)
 217  *
 218  * [input, input+input_size) is the original copied compressed image (ZO)
 219  * (i.e. it does not include its run size). This range must be avoided
 220  * because it contains the data used for decompression.
 221  *
 222  * [input+input_size, output+init_size) is [_text, _end) for ZO. This
 223  * range includes ZO's heap and stack, and must be avoided since it
 224  * performs the decompression.
 225  *
 226  * Since the above two ranges need to be avoided and they are adjacent,
 227  * they can be merged, resulting in: [input, output+init_size) which
 228  * becomes the MEM_AVOID_ZO_RANGE below.
 229  */
 230 static void mem_avoid_init(unsigned long input, unsigned long input_size,
 231                            unsigned long output)
 232 {
 233         unsigned long init_size = boot_params->hdr.init_size;
 234         u64 initrd_start, initrd_size;
 235         u64 cmd_line, cmd_line_size;
 236         char *ptr;
 237
 238         /*
 239          * Avoid the region that is unsafe to overlap during
 240          * decompression.
 241          */
 242         mem_avoid[MEM_AVOID_ZO_RANGE].start = input;
 243         mem_avoid[MEM_AVOID_ZO_RANGE].size = (output + init_size) - input;
 244         add_identity_map(mem_avoid[MEM_AVOID_ZO_RANGE].start,
 245                          mem_avoid[MEM_AVOID_ZO_RANGE].size);
 246
 247         /* Avoid initrd. */
 248         initrd_start  = (u64)boot_params->ext_ramdisk_image << 32;
 249         initrd_start |= boot_params->hdr.ramdisk_image;
 250         initrd_size  = (u64)boot_params->ext_ramdisk_size << 32;
 251         initrd_size |= boot_params->hdr.ramdisk_size;
 252         mem_avoid[MEM_AVOID_INITRD].start = initrd_start;
 253         mem_avoid[MEM_AVOID_INITRD].size = initrd_size;
 254         /* No need to set mapping for initrd, it will be handled in VO. */
 255
 256         /* Avoid kernel command line. */
 257         cmd_line  = (u64)boot_params->ext_cmd_line_ptr << 32;
 258         cmd_line |= boot_params->hdr.cmd_line_ptr;
 259         /* Calculate size of cmd_line. */
 260         ptr = (char *)(unsigned long)cmd_line;
 261         for (cmd_line_size = 0; ptr[cmd_line_size++]; )
 262                 ;
 263         mem_avoid[MEM_AVOID_CMDLINE].start = cmd_line;
 264         mem_avoid[MEM_AVOID_CMDLINE].size = cmd_line_size;
 265         add_identity_map(mem_avoid[MEM_AVOID_CMDLINE].start,
 266                          mem_avoid[MEM_AVOID_CMDLINE].size);
 267
 268         /* Avoid boot parameters. */
 269         mem_avoid[MEM_AVOID_BOOTPARAMS].start = (unsigned long)boot_params;
 270         mem_avoid[MEM_AVOID_BOOTPARAMS].size = sizeof(*boot_params);
 271         add_identity_map(mem_avoid[MEM_AVOID_BOOTPARAMS].start,
 272                          mem_avoid[MEM_AVOID_BOOTPARAMS].size);
 273
 274         /* We don't need to set a mapping for setup_data. */
 275
 276 #ifdef CONFIG_X86_VERBOSE_BOOTUP
 277         /* Make sure video RAM can be used. */
 278         add_identity_map(0, PMD_SIZE);
 279 #endif
 280 }
 281
 282 /* Does this memory vector overlap a known avoided area? */
 283 static bool mem_avoid_overlap(struct mem_vector *img)
 284 {
 285         int i;
 286         struct setup_data *ptr;
 287
 288         for (i = 0; i < MEM_AVOID_MAX; i++) {
 289                 if (mem_overlaps(img, &mem_avoid[i]))
 290                         return true;
 291         }
 292
 293         /* Avoid all entries in the setup_data linked list. */
 294         ptr = (struct setup_data *)(unsigned long)boot_params->hdr.setup_data;
 295         while (ptr) {
 296                 struct mem_vector avoid;
 297
 298                 avoid.start = (unsigned long)ptr;
 299                 avoid.size = sizeof(*ptr) + ptr->len;
 300
 301                 if (mem_overlaps(img, &avoid))
 302                         return true;
 303
 304                 ptr = (struct setup_data *)(unsigned long)ptr->next;
 305         }
 306
 307         return false;
 308 }
 309
 310 static unsigned long slots[KERNEL_IMAGE_SIZE / CONFIG_PHYSICAL_ALIGN];
 311 static unsigned long slot_max;
 312
 313 static void slots_append(unsigned long addr)
 314 {
 315         /* Overflowing the slots list should be impossible. */
 316         if (slot_max >= KERNEL_IMAGE_SIZE / CONFIG_PHYSICAL_ALIGN)
 317                 return;
 318
 319         slots[slot_max++] = addr;
 320 }
 321
 322 static unsigned long slots_fetch_random(void)
 323 {
 324         /* Handle case of no slots stored. */
 325         if (slot_max == 0)
 326                 return 0;
 327
 328         return slots[get_random_long() % slot_max];
 329 }
 330
 331 static void process_e820_entry(struct e820entry *entry,
 332                                unsigned long minimum,
 333                                unsigned long image_size)
 334 {
 335         struct mem_vector region, img;
 336
 337         /* Skip non-RAM entries. */
 338         if (entry->type != E820_RAM)
 339                 return;
 340
 341         /* Ignore entries entirely above our maximum. */
 342         if (entry->addr >= KERNEL_IMAGE_SIZE)
 343                 return;
 344
 345         /* Ignore entries entirely below our minimum. */
 346         if (entry->addr + entry->size < minimum)
 347                 return;
 348
 349         region.start = entry->addr;
 350         region.size = entry->size;
 351
 352         /* Potentially raise address to minimum location. */
 353         if (region.start < minimum)
 354                 region.start = minimum;
 355
 356         /* Potentially raise address to meet alignment requirements. */
 357         region.start = ALIGN(region.start, CONFIG_PHYSICAL_ALIGN);
 358
 359         /* Did we raise the address above the bounds of this e820 region? */
 360         if (region.start > entry->addr + entry->size)
 361                 return;
 362
 363         /* Reduce size by any delta from the original address. */
 364         region.size -= region.start - entry->addr;
 365
 366         /* Reduce maximum size to fit end of image within maximum limit. */
 367         if (region.start + region.size > KERNEL_IMAGE_SIZE)
 368                 region.size = KERNEL_IMAGE_SIZE - region.start;
 369
 370         /* Walk each aligned slot and check for avoided areas. */
 371         for (img.start = region.start, img.size = image_size ;
 372              mem_contains(&region, &img) ;
 373              img.start += CONFIG_PHYSICAL_ALIGN) {
 374                 if (mem_avoid_overlap(&img))
 375                         continue;
 376                 slots_append(img.start);
 377         }
 378 }
 379
 380 static unsigned long find_random_addr(unsigned long minimum,
 381                                       unsigned long size)
 382 {
 383         int i;
 384         unsigned long addr;
 385
 386         /* Make sure minimum is aligned. */
 387         minimum = ALIGN(minimum, CONFIG_PHYSICAL_ALIGN);
 388
 389         /* Verify potential e820 positions, appending to slots list. */
 390         for (i = 0; i < boot_params->e820_entries; i++) {
 391                 process_e820_entry(&boot_params->e820_map[i], minimum, size);
 392         }
 393
 394         return slots_fetch_random();
 395 }
 396
 397 /*
 398  * Since this function examines addresses much more numerically,
 399  * it takes the input and output pointers as 'unsigned long'.
 400  */
 401 unsigned char *choose_random_location(unsigned long input,
 402                                       unsigned long input_size,
 403                                       unsigned long output,
 404                                       unsigned long output_size)
 405 {
 406         unsigned long choice = output;
 407         unsigned long random_addr;
 408
 409 #ifdef CONFIG_HIBERNATION
 410         if (!cmdline_find_option_bool("kaslr")) {
 411                 warn("KASLR disabled: 'kaslr' not on cmdline (hibernation selected).");
 412                 goto out;
 413         }
 414 #else
 415         if (cmdline_find_option_bool("nokaslr")) {
 416                 warn("KASLR disabled: 'nokaslr' on cmdline.");
 417                 goto out;
 418         }
 419 #endif
 420
 421         boot_params->hdr.loadflags |= KASLR_FLAG;
 422
 423         /* Record the various known unsafe memory ranges. */
 424         mem_avoid_init(input, input_size, output);
 425
 426         /* Walk e820 and find a random address. */
 427         random_addr = find_random_addr(output, output_size);
 428         if (!random_addr) {
 429                 warn("KASLR disabled: could not find suitable E820 region!");
 430                 goto out;
 431         }
 432
 433         /* Always enforce the minimum. */
 434         if (random_addr < choice)
 435                 goto out;
 436
 437         choice = random_addr;
 438
 439         add_identity_map(choice, output_size);
 440
 441         /* This actually loads the identity pagetable on x86_64. */
 442         finalize_identity_maps();
 443 out:
 444         return (unsigned char *)choice;
 445 }