1 // SPDX-License-Identifier: GPL-2.0
5 * This contains the routines needed to generate a reasonable level of
6 * entropy to choose a randomized kernel base address offset in support
7 * of Kernel Address Space Layout Randomization (KASLR). Additionally
8 * handles walking the physical memory maps (and tracking memory regions
9 * to avoid) in order to select a physical memory location that can
10 * contain the entire properly aligned running kernel image.
15 * isspace() in linux/ctype.h is expected by next_args() to filter
16 * out "space/lf/tab". While boot/ctype.h conflicts with linux/ctype.h,
17 * since isdigit() is implemented in both of them. Hence disable it
24 #include "../string.h"
27 #include <generated/compile.h>
28 #include <generated/utsversion.h>
29 #include <generated/utsrelease.h>
32 #include <asm/setup.h> /* For COMMAND_LINE_SIZE */
35 extern unsigned long get_cmd_line_ptr(void);
37 /* Simplified build-specific string for starting entropy. */
38 static const char build_str[] = UTS_RELEASE " (" LINUX_COMPILE_BY "@"
39 LINUX_COMPILE_HOST ") (" LINUX_COMPILER ") " UTS_VERSION;
41 static unsigned long rotate_xor(unsigned long hash, const void *area,
45 unsigned long *ptr = (unsigned long *)area;
47 for (i = 0; i < size / sizeof(hash); i++) {
48 /* Rotate by odd number of bits and XOR. */
49 hash = (hash << ((sizeof(hash) * 8) - 7)) | (hash >> 7);
56 /* Attempt to create a simple but unpredictable starting entropy. */
57 static unsigned long get_boot_seed(void)
59 unsigned long hash = 0;
61 hash = rotate_xor(hash, build_str, sizeof(build_str));
62 hash = rotate_xor(hash, boot_params_ptr, sizeof(*boot_params_ptr));
67 #define KASLR_COMPRESSED_BOOT
68 #include "../../lib/kaslr.c"
71 /* Only supporting at most 4 unusable memmap regions with kaslr */
72 #define MAX_MEMMAP_REGIONS 4
74 static bool memmap_too_large;
78 * Store memory limit: MAXMEM on 64-bit and KERNEL_IMAGE_SIZE on 32-bit.
79 * It may be reduced by "mem=nn[KMG]" or "memmap=nn[KMG]" command line options.
83 /* Number of immovable memory regions */
84 static int num_immovable_mem;
86 enum mem_avoid_index {
87 MEM_AVOID_ZO_RANGE = 0,
91 MEM_AVOID_MEMMAP_BEGIN,
92 MEM_AVOID_MEMMAP_END = MEM_AVOID_MEMMAP_BEGIN + MAX_MEMMAP_REGIONS - 1,
96 static struct mem_vector mem_avoid[MEM_AVOID_MAX];
98 static bool mem_overlaps(struct mem_vector *one, struct mem_vector *two)
100 /* Item one is entirely before item two. */
101 if (one->start + one->size <= two->start)
103 /* Item one is entirely after item two. */
104 if (one->start >= two->start + two->size)
109 char *skip_spaces(const char *str)
111 while (isspace(*str))
115 #include "../../../../lib/ctype.c"
116 #include "../../../../lib/cmdline.c"
119 parse_memmap(char *p, u64 *start, u64 *size)
126 /* We don't care about this option here */
127 if (!strncmp(p, "exactmap", 8))
131 *size = memparse(p, &p);
139 *start = memparse(p + 1, &p);
143 * memmap=nn@ss specifies usable region, should
150 * If w/o offset, only size specified, memmap=nn[KMG] has the
151 * same behaviour as mem=nn[KMG]. It limits the max address
152 * system can use. Region above the limit should be avoided.
161 static void mem_avoid_memmap(char *str)
165 if (i >= MAX_MEMMAP_REGIONS)
168 while (str && (i < MAX_MEMMAP_REGIONS)) {
171 char *k = strchr(str, ',');
176 rc = parse_memmap(str, &start, &size);
182 /* Store the specified memory limit if size > 0 */
183 if (size > 0 && size < mem_limit)
189 mem_avoid[MEM_AVOID_MEMMAP_BEGIN + i].start = start;
190 mem_avoid[MEM_AVOID_MEMMAP_BEGIN + i].size = size;
194 /* More than 4 memmaps, fail kaslr */
195 if ((i >= MAX_MEMMAP_REGIONS) && str)
196 memmap_too_large = true;
199 /* Store the number of 1GB huge pages which users specified: */
200 static unsigned long max_gb_huge_pages;
202 static void parse_gb_huge_pages(char *param, char *val)
204 static bool gbpage_sz;
207 if (!strcmp(param, "hugepagesz")) {
209 if (memparse(p, &p) != PUD_SIZE) {
215 warn("Repeatedly set hugeTLB page size of 1G!\n");
220 if (!strcmp(param, "hugepages") && gbpage_sz) {
222 max_gb_huge_pages = simple_strtoull(p, &p, 0);
227 static void handle_mem_options(void)
229 char *args = (char *)get_cmd_line_ptr();
238 len = strnlen(args, COMMAND_LINE_SIZE-1);
239 tmp_cmdline = malloc(len + 1);
241 error("Failed to allocate space for tmp_cmdline");
243 memcpy(tmp_cmdline, args, len);
244 tmp_cmdline[len] = 0;
247 /* Chew leading spaces */
248 args = skip_spaces(args);
251 args = next_arg(args, ¶m, &val);
253 if (!val && strcmp(param, "--") == 0)
256 if (!strcmp(param, "memmap")) {
257 mem_avoid_memmap(val);
258 } else if (IS_ENABLED(CONFIG_X86_64) && strstr(param, "hugepages")) {
259 parse_gb_huge_pages(param, val);
260 } else if (!strcmp(param, "mem")) {
263 if (!strcmp(p, "nopentium"))
265 mem_size = memparse(p, &p);
269 if (mem_size < mem_limit)
270 mem_limit = mem_size;
279 * In theory, KASLR can put the kernel anywhere in the range of [16M, MAXMEM)
280 * on 64-bit, and [16M, KERNEL_IMAGE_SIZE) on 32-bit.
282 * The mem_avoid array is used to store the ranges that need to be avoided
283 * when KASLR searches for an appropriate random address. We must avoid any
284 * regions that are unsafe to overlap with during decompression, and other
285 * things like the initrd, cmdline and boot_params. This comment seeks to
286 * explain mem_avoid as clearly as possible since incorrect mem_avoid
287 * memory ranges lead to really hard to debug boot failures.
289 * The initrd, cmdline, and boot_params are trivial to identify for
290 * avoiding. They are MEM_AVOID_INITRD, MEM_AVOID_CMDLINE, and
291 * MEM_AVOID_BOOTPARAMS respectively below.
293 * What is not obvious how to avoid is the range of memory that is used
294 * during decompression (MEM_AVOID_ZO_RANGE below). This range must cover
295 * the compressed kernel (ZO) and its run space, which is used to extract
296 * the uncompressed kernel (VO) and relocs.
298 * ZO's full run size sits against the end of the decompression buffer, so
299 * we can calculate where text, data, bss, etc of ZO are positioned more
302 * For additional background, the decompression calculations can be found
303 * in header.S, and the memory diagram is based on the one found in misc.c.
305 * The following conditions are already enforced by the image layouts and
307 * - input + input_size >= output + output_size
308 * - kernel_total_size <= init_size
309 * - kernel_total_size <= output_size (see Note below)
310 * - output + init_size >= output + output_size
312 * (Note that kernel_total_size and output_size have no fundamental
313 * relationship, but output_size is passed to choose_random_location
314 * as a maximum of the two. The diagram is showing a case where
315 * kernel_total_size is larger than output_size, but this case is
316 * handled by bumping output_size.)
318 * The above conditions can be illustrated by a diagram:
320 * 0 output input input+input_size output+init_size
323 * |-----|--------|--------|--------------|-----------|--|-------------|
326 * output+init_size-ZO_INIT_SIZE output+output_size output+kernel_total_size
328 * [output, output+init_size) is the entire memory range used for
329 * extracting the compressed image.
331 * [output, output+kernel_total_size) is the range needed for the
332 * uncompressed kernel (VO) and its run size (bss, brk, etc).
334 * [output, output+output_size) is VO plus relocs (i.e. the entire
335 * uncompressed payload contained by ZO). This is the area of the buffer
336 * written to during decompression.
338 * [output+init_size-ZO_INIT_SIZE, output+init_size) is the worst-case
339 * range of the copied ZO and decompression code. (i.e. the range
340 * covered backwards of size ZO_INIT_SIZE, starting from output+init_size.)
342 * [input, input+input_size) is the original copied compressed image (ZO)
343 * (i.e. it does not include its run size). This range must be avoided
344 * because it contains the data used for decompression.
346 * [input+input_size, output+init_size) is [_text, _end) for ZO. This
347 * range includes ZO's heap and stack, and must be avoided since it
348 * performs the decompression.
350 * Since the above two ranges need to be avoided and they are adjacent,
351 * they can be merged, resulting in: [input, output+init_size) which
352 * becomes the MEM_AVOID_ZO_RANGE below.
354 static void mem_avoid_init(unsigned long input, unsigned long input_size,
355 unsigned long output)
357 unsigned long init_size = boot_params_ptr->hdr.init_size;
358 u64 initrd_start, initrd_size;
359 unsigned long cmd_line, cmd_line_size;
362 * Avoid the region that is unsafe to overlap during
365 mem_avoid[MEM_AVOID_ZO_RANGE].start = input;
366 mem_avoid[MEM_AVOID_ZO_RANGE].size = (output + init_size) - input;
369 initrd_start = (u64)boot_params_ptr->ext_ramdisk_image << 32;
370 initrd_start |= boot_params_ptr->hdr.ramdisk_image;
371 initrd_size = (u64)boot_params_ptr->ext_ramdisk_size << 32;
372 initrd_size |= boot_params_ptr->hdr.ramdisk_size;
373 mem_avoid[MEM_AVOID_INITRD].start = initrd_start;
374 mem_avoid[MEM_AVOID_INITRD].size = initrd_size;
375 /* No need to set mapping for initrd, it will be handled in VO. */
377 /* Avoid kernel command line. */
378 cmd_line = get_cmd_line_ptr();
379 /* Calculate size of cmd_line. */
381 cmd_line_size = strnlen((char *)cmd_line, COMMAND_LINE_SIZE-1) + 1;
382 mem_avoid[MEM_AVOID_CMDLINE].start = cmd_line;
383 mem_avoid[MEM_AVOID_CMDLINE].size = cmd_line_size;
386 /* Avoid boot parameters. */
387 mem_avoid[MEM_AVOID_BOOTPARAMS].start = (unsigned long)boot_params_ptr;
388 mem_avoid[MEM_AVOID_BOOTPARAMS].size = sizeof(*boot_params_ptr);
390 /* We don't need to set a mapping for setup_data. */
392 /* Mark the memmap regions we need to avoid */
393 handle_mem_options();
395 /* Enumerate the immovable memory regions */
396 num_immovable_mem = count_immovable_mem_regions();
400 * Does this memory vector overlap a known avoided area? If so, record the
401 * overlap region with the lowest address.
403 static bool mem_avoid_overlap(struct mem_vector *img,
404 struct mem_vector *overlap)
407 struct setup_data *ptr;
408 u64 earliest = img->start + img->size;
409 bool is_overlapping = false;
411 for (i = 0; i < MEM_AVOID_MAX; i++) {
412 if (mem_overlaps(img, &mem_avoid[i]) &&
413 mem_avoid[i].start < earliest) {
414 *overlap = mem_avoid[i];
415 earliest = overlap->start;
416 is_overlapping = true;
420 /* Avoid all entries in the setup_data linked list. */
421 ptr = (struct setup_data *)(unsigned long)boot_params_ptr->hdr.setup_data;
423 struct mem_vector avoid;
425 avoid.start = (unsigned long)ptr;
426 avoid.size = sizeof(*ptr) + ptr->len;
428 if (mem_overlaps(img, &avoid) && (avoid.start < earliest)) {
430 earliest = overlap->start;
431 is_overlapping = true;
434 if (ptr->type == SETUP_INDIRECT &&
435 ((struct setup_indirect *)ptr->data)->type != SETUP_INDIRECT) {
436 avoid.start = ((struct setup_indirect *)ptr->data)->addr;
437 avoid.size = ((struct setup_indirect *)ptr->data)->len;
439 if (mem_overlaps(img, &avoid) && (avoid.start < earliest)) {
441 earliest = overlap->start;
442 is_overlapping = true;
446 ptr = (struct setup_data *)(unsigned long)ptr->next;
449 return is_overlapping;
457 #define MAX_SLOT_AREA 100
459 static struct slot_area slot_areas[MAX_SLOT_AREA];
460 static unsigned int slot_area_index;
461 static unsigned long slot_max;
463 static void store_slot_info(struct mem_vector *region, unsigned long image_size)
465 struct slot_area slot_area;
467 if (slot_area_index == MAX_SLOT_AREA)
470 slot_area.addr = region->start;
471 slot_area.num = 1 + (region->size - image_size) / CONFIG_PHYSICAL_ALIGN;
473 slot_areas[slot_area_index++] = slot_area;
474 slot_max += slot_area.num;
478 * Skip as many 1GB huge pages as possible in the passed region
479 * according to the number which users specified:
482 process_gb_huge_pages(struct mem_vector *region, unsigned long image_size)
484 u64 pud_start, pud_end;
485 unsigned long gb_huge_pages;
486 struct mem_vector tmp;
488 if (!IS_ENABLED(CONFIG_X86_64) || !max_gb_huge_pages) {
489 store_slot_info(region, image_size);
493 /* Are there any 1GB pages in the region? */
494 pud_start = ALIGN(region->start, PUD_SIZE);
495 pud_end = ALIGN_DOWN(region->start + region->size, PUD_SIZE);
497 /* No good 1GB huge pages found: */
498 if (pud_start >= pud_end) {
499 store_slot_info(region, image_size);
503 /* Check if the head part of the region is usable. */
504 if (pud_start >= region->start + image_size) {
505 tmp.start = region->start;
506 tmp.size = pud_start - region->start;
507 store_slot_info(&tmp, image_size);
510 /* Skip the good 1GB pages. */
511 gb_huge_pages = (pud_end - pud_start) >> PUD_SHIFT;
512 if (gb_huge_pages > max_gb_huge_pages) {
513 pud_end = pud_start + (max_gb_huge_pages << PUD_SHIFT);
514 max_gb_huge_pages = 0;
516 max_gb_huge_pages -= gb_huge_pages;
519 /* Check if the tail part of the region is usable. */
520 if (region->start + region->size >= pud_end + image_size) {
522 tmp.size = region->start + region->size - pud_end;
523 store_slot_info(&tmp, image_size);
527 static u64 slots_fetch_random(void)
532 /* Handle case of no slots stored. */
536 slot = kaslr_get_random_long("Physical") % slot_max;
538 for (i = 0; i < slot_area_index; i++) {
539 if (slot >= slot_areas[i].num) {
540 slot -= slot_areas[i].num;
543 return slot_areas[i].addr + ((u64)slot * CONFIG_PHYSICAL_ALIGN);
546 if (i == slot_area_index)
547 debug_putstr("slots_fetch_random() failed!?\n");
551 static void __process_mem_region(struct mem_vector *entry,
552 unsigned long minimum,
553 unsigned long image_size)
555 struct mem_vector region, overlap;
558 /* Enforce minimum and memory limit. */
559 region.start = max_t(u64, entry->start, minimum);
560 region_end = min(entry->start + entry->size, mem_limit);
562 /* Give up if slot area array is full. */
563 while (slot_area_index < MAX_SLOT_AREA) {
564 /* Potentially raise address to meet alignment needs. */
565 region.start = ALIGN(region.start, CONFIG_PHYSICAL_ALIGN);
567 /* Did we raise the address above the passed in memory entry? */
568 if (region.start > region_end)
571 /* Reduce size by any delta from the original address. */
572 region.size = region_end - region.start;
574 /* Return if region can't contain decompressed kernel */
575 if (region.size < image_size)
578 /* If nothing overlaps, store the region and return. */
579 if (!mem_avoid_overlap(®ion, &overlap)) {
580 process_gb_huge_pages(®ion, image_size);
584 /* Store beginning of region if holds at least image_size. */
585 if (overlap.start >= region.start + image_size) {
586 region.size = overlap.start - region.start;
587 process_gb_huge_pages(®ion, image_size);
590 /* Clip off the overlapping region and start over. */
591 region.start = overlap.start + overlap.size;
595 static bool process_mem_region(struct mem_vector *region,
596 unsigned long minimum,
597 unsigned long image_size)
601 * If no immovable memory found, or MEMORY_HOTREMOVE disabled,
602 * use @region directly.
604 if (!num_immovable_mem) {
605 __process_mem_region(region, minimum, image_size);
607 if (slot_area_index == MAX_SLOT_AREA) {
608 debug_putstr("Aborted e820/efi memmap scan (slot_areas full)!\n");
614 #if defined(CONFIG_MEMORY_HOTREMOVE) && defined(CONFIG_ACPI)
616 * If immovable memory found, filter the intersection between
617 * immovable memory and @region.
619 for (i = 0; i < num_immovable_mem; i++) {
620 u64 start, end, entry_end, region_end;
621 struct mem_vector entry;
623 if (!mem_overlaps(region, &immovable_mem[i]))
626 start = immovable_mem[i].start;
627 end = start + immovable_mem[i].size;
628 region_end = region->start + region->size;
630 entry.start = clamp(region->start, start, end);
631 entry_end = clamp(region_end, start, end);
632 entry.size = entry_end - entry.start;
634 __process_mem_region(&entry, minimum, image_size);
636 if (slot_area_index == MAX_SLOT_AREA) {
637 debug_putstr("Aborted e820/efi memmap scan when walking immovable regions(slot_areas full)!\n");
648 * Only EFI_CONVENTIONAL_MEMORY and EFI_UNACCEPTED_MEMORY (if supported) are
649 * guaranteed to be free.
651 * Pick free memory more conservatively than the EFI spec allows: according to
652 * the spec, EFI_BOOT_SERVICES_{CODE|DATA} are also free memory and thus
653 * available to place the kernel image into, but in practice there's firmware
654 * where using that memory leads to crashes. Buggy vendor EFI code registers
655 * for an event that triggers on SetVirtualAddressMap(). The handler assumes
656 * that EFI_BOOT_SERVICES_DATA memory has not been touched by loader yet, which
657 * is probably true for Windows.
659 * Preserve EFI_BOOT_SERVICES_* regions until after SetVirtualAddressMap().
661 static inline bool memory_type_is_free(efi_memory_desc_t *md)
663 if (md->type == EFI_CONVENTIONAL_MEMORY)
666 if (IS_ENABLED(CONFIG_UNACCEPTED_MEMORY) &&
667 md->type == EFI_UNACCEPTED_MEMORY)
674 * Returns true if we processed the EFI memmap, which we prefer over the E820
675 * table if it is available.
678 process_efi_entries(unsigned long minimum, unsigned long image_size)
680 struct efi_info *e = &boot_params_ptr->efi_info;
681 bool efi_mirror_found = false;
682 struct mem_vector region;
683 efi_memory_desc_t *md;
689 signature = (char *)&e->efi_loader_signature;
690 if (strncmp(signature, EFI32_LOADER_SIGNATURE, 4) &&
691 strncmp(signature, EFI64_LOADER_SIGNATURE, 4))
695 /* Can't handle data above 4GB at this time */
696 if (e->efi_memmap_hi) {
697 warn("EFI memmap is above 4GB, can't be handled now on x86_32. EFI should be disabled.\n");
700 pmap = e->efi_memmap;
702 pmap = (e->efi_memmap | ((__u64)e->efi_memmap_hi << 32));
705 nr_desc = e->efi_memmap_size / e->efi_memdesc_size;
706 for (i = 0; i < nr_desc; i++) {
707 md = efi_early_memdesc_ptr(pmap, e->efi_memdesc_size, i);
708 if (md->attribute & EFI_MEMORY_MORE_RELIABLE) {
709 efi_mirror_found = true;
714 for (i = 0; i < nr_desc; i++) {
715 md = efi_early_memdesc_ptr(pmap, e->efi_memdesc_size, i);
717 if (!memory_type_is_free(md))
720 if (efi_soft_reserve_enabled() &&
721 (md->attribute & EFI_MEMORY_SP))
724 if (efi_mirror_found &&
725 !(md->attribute & EFI_MEMORY_MORE_RELIABLE))
728 region.start = md->phys_addr;
729 region.size = md->num_pages << EFI_PAGE_SHIFT;
730 if (process_mem_region(®ion, minimum, image_size))
737 process_efi_entries(unsigned long minimum, unsigned long image_size)
743 static void process_e820_entries(unsigned long minimum,
744 unsigned long image_size)
747 struct mem_vector region;
748 struct boot_e820_entry *entry;
750 /* Verify potential e820 positions, appending to slots list. */
751 for (i = 0; i < boot_params_ptr->e820_entries; i++) {
752 entry = &boot_params_ptr->e820_table[i];
753 /* Skip non-RAM entries. */
754 if (entry->type != E820_TYPE_RAM)
756 region.start = entry->addr;
757 region.size = entry->size;
758 if (process_mem_region(®ion, minimum, image_size))
763 static unsigned long find_random_phys_addr(unsigned long minimum,
764 unsigned long image_size)
768 /* Bail out early if it's impossible to succeed. */
769 if (minimum + image_size > mem_limit)
772 /* Check if we had too many memmaps. */
773 if (memmap_too_large) {
774 debug_putstr("Aborted memory entries scan (more than 4 memmap= args)!\n");
778 if (!process_efi_entries(minimum, image_size))
779 process_e820_entries(minimum, image_size);
781 phys_addr = slots_fetch_random();
783 /* Perform a final check to make sure the address is in range. */
784 if (phys_addr < minimum || phys_addr + image_size > mem_limit) {
785 warn("Invalid physical address chosen!\n");
789 return (unsigned long)phys_addr;
792 static unsigned long find_random_virt_addr(unsigned long minimum,
793 unsigned long image_size)
795 unsigned long slots, random_addr;
798 * There are how many CONFIG_PHYSICAL_ALIGN-sized slots
799 * that can hold image_size within the range of minimum to
802 slots = 1 + (KERNEL_IMAGE_SIZE - minimum - image_size) / CONFIG_PHYSICAL_ALIGN;
804 random_addr = kaslr_get_random_long("Virtual") % slots;
806 return random_addr * CONFIG_PHYSICAL_ALIGN + minimum;
810 * Since this function examines addresses much more numerically,
811 * it takes the input and output pointers as 'unsigned long'.
813 void choose_random_location(unsigned long input,
814 unsigned long input_size,
815 unsigned long *output,
816 unsigned long output_size,
817 unsigned long *virt_addr)
819 unsigned long random_addr, min_addr;
821 if (cmdline_find_option_bool("nokaslr")) {
822 warn("KASLR disabled: 'nokaslr' on cmdline.");
826 boot_params_ptr->hdr.loadflags |= KASLR_FLAG;
828 if (IS_ENABLED(CONFIG_X86_32))
829 mem_limit = KERNEL_IMAGE_SIZE;
833 /* Record the various known unsafe memory ranges. */
834 mem_avoid_init(input, input_size, *output);
837 * Low end of the randomization range should be the
838 * smaller of 512M or the initial kernel image
841 min_addr = min(*output, 512UL << 20);
842 /* Make sure minimum is aligned. */
843 min_addr = ALIGN(min_addr, CONFIG_PHYSICAL_ALIGN);
845 /* Walk available memory entries to find a random address. */
846 random_addr = find_random_phys_addr(min_addr, output_size);
848 warn("Physical KASLR disabled: no suitable memory region!");
850 /* Update the new physical address location. */
851 if (*output != random_addr)
852 *output = random_addr;
856 /* Pick random virtual address starting from LOAD_PHYSICAL_ADDR. */
857 if (IS_ENABLED(CONFIG_X86_64))
858 random_addr = find_random_virt_addr(LOAD_PHYSICAL_ADDR, output_size);
859 *virt_addr = random_addr;