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
23 * _ctype[] in lib/ctype.c is needed by isspace() of linux/ctype.h.
24 * While both lib/ctype.c and lib/cmdline.c will bring EXPORT_SYMBOL
25 * which is meaningless and will cause compiling error in some cases.
26 * So do not include linux/export.h and define EXPORT_SYMBOL(sym)
29 #define _LINUX_EXPORT_H
30 #define EXPORT_SYMBOL(sym)
34 #include "../string.h"
36 #include <generated/compile.h>
37 #include <linux/module.h>
38 #include <linux/uts.h>
39 #include <linux/utsname.h>
40 #include <linux/ctype.h>
41 #include <linux/efi.h>
42 #include <generated/utsrelease.h>
45 /* Macros used by the included decompressor code below. */
47 #include <linux/decompress/mm.h>
49 #ifdef CONFIG_X86_5LEVEL
50 unsigned int pgtable_l5_enabled __ro_after_init = 1;
51 unsigned int pgdir_shift __ro_after_init = 48;
52 unsigned int ptrs_per_p4d __ro_after_init = 512;
55 extern unsigned long get_cmd_line_ptr(void);
57 /* Simplified build-specific string for starting entropy. */
58 static const char build_str[] = UTS_RELEASE " (" LINUX_COMPILE_BY "@"
59 LINUX_COMPILE_HOST ") (" LINUX_COMPILER ") " UTS_VERSION;
61 static unsigned long rotate_xor(unsigned long hash, const void *area,
65 unsigned long *ptr = (unsigned long *)area;
67 for (i = 0; i < size / sizeof(hash); i++) {
68 /* Rotate by odd number of bits and XOR. */
69 hash = (hash << ((sizeof(hash) * 8) - 7)) | (hash >> 7);
76 /* Attempt to create a simple but unpredictable starting entropy. */
77 static unsigned long get_boot_seed(void)
79 unsigned long hash = 0;
81 hash = rotate_xor(hash, build_str, sizeof(build_str));
82 hash = rotate_xor(hash, boot_params, sizeof(*boot_params));
87 #define KASLR_COMPRESSED_BOOT
88 #include "../../lib/kaslr.c"
91 unsigned long long start;
92 unsigned long long size;
95 /* Only supporting at most 4 unusable memmap regions with kaslr */
96 #define MAX_MEMMAP_REGIONS 4
98 static bool memmap_too_large;
101 /* Store memory limit specified by "mem=nn[KMG]" or "memmap=nn[KMG]" */
102 unsigned long long mem_limit = ULLONG_MAX;
105 enum mem_avoid_index {
106 MEM_AVOID_ZO_RANGE = 0,
109 MEM_AVOID_BOOTPARAMS,
110 MEM_AVOID_MEMMAP_BEGIN,
111 MEM_AVOID_MEMMAP_END = MEM_AVOID_MEMMAP_BEGIN + MAX_MEMMAP_REGIONS - 1,
115 static struct mem_vector mem_avoid[MEM_AVOID_MAX];
117 static bool mem_overlaps(struct mem_vector *one, struct mem_vector *two)
119 /* Item one is entirely before item two. */
120 if (one->start + one->size <= two->start)
122 /* Item one is entirely after item two. */
123 if (one->start >= two->start + two->size)
128 char *skip_spaces(const char *str)
130 while (isspace(*str))
134 #include "../../../../lib/ctype.c"
135 #include "../../../../lib/cmdline.c"
138 parse_memmap(char *p, unsigned long long *start, unsigned long long *size)
145 /* We don't care about this option here */
146 if (!strncmp(p, "exactmap", 8))
150 *size = memparse(p, &p);
158 *start = memparse(p + 1, &p);
161 /* memmap=nn@ss specifies usable region, should be skipped */
166 * If w/o offset, only size specified, memmap=nn[KMG] has the
167 * same behaviour as mem=nn[KMG]. It limits the max address
168 * system can use. Region above the limit should be avoided.
177 static void mem_avoid_memmap(char *str)
181 if (i >= MAX_MEMMAP_REGIONS)
184 while (str && (i < MAX_MEMMAP_REGIONS)) {
186 unsigned long long start, size;
187 char *k = strchr(str, ',');
192 rc = parse_memmap(str, &start, &size);
198 /* Store the specified memory limit if size > 0 */
205 mem_avoid[MEM_AVOID_MEMMAP_BEGIN + i].start = start;
206 mem_avoid[MEM_AVOID_MEMMAP_BEGIN + i].size = size;
210 /* More than 4 memmaps, fail kaslr */
211 if ((i >= MAX_MEMMAP_REGIONS) && str)
212 memmap_too_large = true;
215 static int handle_mem_memmap(void)
217 char *args = (char *)get_cmd_line_ptr();
218 size_t len = strlen((char *)args);
223 if (!strstr(args, "memmap=") && !strstr(args, "mem="))
226 tmp_cmdline = malloc(len + 1);
228 error("Failed to allocate space for tmp_cmdline");
230 memcpy(tmp_cmdline, args, len);
231 tmp_cmdline[len] = 0;
234 /* Chew leading spaces */
235 args = skip_spaces(args);
238 args = next_arg(args, ¶m, &val);
240 if (!val && strcmp(param, "--") == 0) {
241 warn("Only '--' specified in cmdline");
246 if (!strcmp(param, "memmap")) {
247 mem_avoid_memmap(val);
248 } else if (!strcmp(param, "mem")) {
251 if (!strcmp(p, "nopentium"))
253 mem_size = memparse(p, &p);
258 mem_limit = mem_size;
267 * In theory, KASLR can put the kernel anywhere in the range of [16M, 64T).
268 * The mem_avoid array is used to store the ranges that need to be avoided
269 * when KASLR searches for an appropriate random address. We must avoid any
270 * regions that are unsafe to overlap with during decompression, and other
271 * things like the initrd, cmdline and boot_params. This comment seeks to
272 * explain mem_avoid as clearly as possible since incorrect mem_avoid
273 * memory ranges lead to really hard to debug boot failures.
275 * The initrd, cmdline, and boot_params are trivial to identify for
276 * avoiding. They are MEM_AVOID_INITRD, MEM_AVOID_CMDLINE, and
277 * MEM_AVOID_BOOTPARAMS respectively below.
279 * What is not obvious how to avoid is the range of memory that is used
280 * during decompression (MEM_AVOID_ZO_RANGE below). This range must cover
281 * the compressed kernel (ZO) and its run space, which is used to extract
282 * the uncompressed kernel (VO) and relocs.
284 * ZO's full run size sits against the end of the decompression buffer, so
285 * we can calculate where text, data, bss, etc of ZO are positioned more
288 * For additional background, the decompression calculations can be found
289 * in header.S, and the memory diagram is based on the one found in misc.c.
291 * The following conditions are already enforced by the image layouts and
293 * - input + input_size >= output + output_size
294 * - kernel_total_size <= init_size
295 * - kernel_total_size <= output_size (see Note below)
296 * - output + init_size >= output + output_size
298 * (Note that kernel_total_size and output_size have no fundamental
299 * relationship, but output_size is passed to choose_random_location
300 * as a maximum of the two. The diagram is showing a case where
301 * kernel_total_size is larger than output_size, but this case is
302 * handled by bumping output_size.)
304 * The above conditions can be illustrated by a diagram:
306 * 0 output input input+input_size output+init_size
309 * |-----|--------|--------|--------------|-----------|--|-------------|
312 * output+init_size-ZO_INIT_SIZE output+output_size output+kernel_total_size
314 * [output, output+init_size) is the entire memory range used for
315 * extracting the compressed image.
317 * [output, output+kernel_total_size) is the range needed for the
318 * uncompressed kernel (VO) and its run size (bss, brk, etc).
320 * [output, output+output_size) is VO plus relocs (i.e. the entire
321 * uncompressed payload contained by ZO). This is the area of the buffer
322 * written to during decompression.
324 * [output+init_size-ZO_INIT_SIZE, output+init_size) is the worst-case
325 * range of the copied ZO and decompression code. (i.e. the range
326 * covered backwards of size ZO_INIT_SIZE, starting from output+init_size.)
328 * [input, input+input_size) is the original copied compressed image (ZO)
329 * (i.e. it does not include its run size). This range must be avoided
330 * because it contains the data used for decompression.
332 * [input+input_size, output+init_size) is [_text, _end) for ZO. This
333 * range includes ZO's heap and stack, and must be avoided since it
334 * performs the decompression.
336 * Since the above two ranges need to be avoided and they are adjacent,
337 * they can be merged, resulting in: [input, output+init_size) which
338 * becomes the MEM_AVOID_ZO_RANGE below.
340 static void mem_avoid_init(unsigned long input, unsigned long input_size,
341 unsigned long output)
343 unsigned long init_size = boot_params->hdr.init_size;
344 u64 initrd_start, initrd_size;
345 u64 cmd_line, cmd_line_size;
349 * Avoid the region that is unsafe to overlap during
352 mem_avoid[MEM_AVOID_ZO_RANGE].start = input;
353 mem_avoid[MEM_AVOID_ZO_RANGE].size = (output + init_size) - input;
354 add_identity_map(mem_avoid[MEM_AVOID_ZO_RANGE].start,
355 mem_avoid[MEM_AVOID_ZO_RANGE].size);
358 initrd_start = (u64)boot_params->ext_ramdisk_image << 32;
359 initrd_start |= boot_params->hdr.ramdisk_image;
360 initrd_size = (u64)boot_params->ext_ramdisk_size << 32;
361 initrd_size |= boot_params->hdr.ramdisk_size;
362 mem_avoid[MEM_AVOID_INITRD].start = initrd_start;
363 mem_avoid[MEM_AVOID_INITRD].size = initrd_size;
364 /* No need to set mapping for initrd, it will be handled in VO. */
366 /* Avoid kernel command line. */
367 cmd_line = (u64)boot_params->ext_cmd_line_ptr << 32;
368 cmd_line |= boot_params->hdr.cmd_line_ptr;
369 /* Calculate size of cmd_line. */
370 ptr = (char *)(unsigned long)cmd_line;
371 for (cmd_line_size = 0; ptr[cmd_line_size++];)
373 mem_avoid[MEM_AVOID_CMDLINE].start = cmd_line;
374 mem_avoid[MEM_AVOID_CMDLINE].size = cmd_line_size;
375 add_identity_map(mem_avoid[MEM_AVOID_CMDLINE].start,
376 mem_avoid[MEM_AVOID_CMDLINE].size);
378 /* Avoid boot parameters. */
379 mem_avoid[MEM_AVOID_BOOTPARAMS].start = (unsigned long)boot_params;
380 mem_avoid[MEM_AVOID_BOOTPARAMS].size = sizeof(*boot_params);
381 add_identity_map(mem_avoid[MEM_AVOID_BOOTPARAMS].start,
382 mem_avoid[MEM_AVOID_BOOTPARAMS].size);
384 /* We don't need to set a mapping for setup_data. */
386 /* Mark the memmap regions we need to avoid */
389 #ifdef CONFIG_X86_VERBOSE_BOOTUP
390 /* Make sure video RAM can be used. */
391 add_identity_map(0, PMD_SIZE);
396 * Does this memory vector overlap a known avoided area? If so, record the
397 * overlap region with the lowest address.
399 static bool mem_avoid_overlap(struct mem_vector *img,
400 struct mem_vector *overlap)
403 struct setup_data *ptr;
404 unsigned long earliest = img->start + img->size;
405 bool is_overlapping = false;
407 for (i = 0; i < MEM_AVOID_MAX; i++) {
408 if (mem_overlaps(img, &mem_avoid[i]) &&
409 mem_avoid[i].start < earliest) {
410 *overlap = mem_avoid[i];
411 earliest = overlap->start;
412 is_overlapping = true;
416 /* Avoid all entries in the setup_data linked list. */
417 ptr = (struct setup_data *)(unsigned long)boot_params->hdr.setup_data;
419 struct mem_vector avoid;
421 avoid.start = (unsigned long)ptr;
422 avoid.size = sizeof(*ptr) + ptr->len;
424 if (mem_overlaps(img, &avoid) && (avoid.start < earliest)) {
426 earliest = overlap->start;
427 is_overlapping = true;
430 ptr = (struct setup_data *)(unsigned long)ptr->next;
433 return is_overlapping;
441 #define MAX_SLOT_AREA 100
443 static struct slot_area slot_areas[MAX_SLOT_AREA];
445 static unsigned long slot_max;
447 static unsigned long slot_area_index;
449 static void store_slot_info(struct mem_vector *region, unsigned long image_size)
451 struct slot_area slot_area;
453 if (slot_area_index == MAX_SLOT_AREA)
456 slot_area.addr = region->start;
457 slot_area.num = (region->size - image_size) /
458 CONFIG_PHYSICAL_ALIGN + 1;
460 if (slot_area.num > 0) {
461 slot_areas[slot_area_index++] = slot_area;
462 slot_max += slot_area.num;
466 static unsigned long slots_fetch_random(void)
471 /* Handle case of no slots stored. */
475 slot = kaslr_get_random_long("Physical") % slot_max;
477 for (i = 0; i < slot_area_index; i++) {
478 if (slot >= slot_areas[i].num) {
479 slot -= slot_areas[i].num;
482 return slot_areas[i].addr + slot * CONFIG_PHYSICAL_ALIGN;
485 if (i == slot_area_index)
486 debug_putstr("slots_fetch_random() failed!?\n");
490 static void process_mem_region(struct mem_vector *entry,
491 unsigned long minimum,
492 unsigned long image_size)
494 struct mem_vector region, overlap;
495 struct slot_area slot_area;
496 unsigned long start_orig, end;
497 struct mem_vector cur_entry;
499 /* On 32-bit, ignore entries entirely above our maximum. */
500 if (IS_ENABLED(CONFIG_X86_32) && entry->start >= KERNEL_IMAGE_SIZE)
503 /* Ignore entries entirely below our minimum. */
504 if (entry->start + entry->size < minimum)
507 /* Ignore entries above memory limit */
508 end = min(entry->size + entry->start, mem_limit);
509 if (entry->start >= end)
511 cur_entry.start = entry->start;
512 cur_entry.size = end - entry->start;
514 region.start = cur_entry.start;
515 region.size = cur_entry.size;
517 /* Give up if slot area array is full. */
518 while (slot_area_index < MAX_SLOT_AREA) {
519 start_orig = region.start;
521 /* Potentially raise address to minimum location. */
522 if (region.start < minimum)
523 region.start = minimum;
525 /* Potentially raise address to meet alignment needs. */
526 region.start = ALIGN(region.start, CONFIG_PHYSICAL_ALIGN);
528 /* Did we raise the address above the passed in memory entry? */
529 if (region.start > cur_entry.start + cur_entry.size)
532 /* Reduce size by any delta from the original address. */
533 region.size -= region.start - start_orig;
535 /* On 32-bit, reduce region size to fit within max size. */
536 if (IS_ENABLED(CONFIG_X86_32) &&
537 region.start + region.size > KERNEL_IMAGE_SIZE)
538 region.size = KERNEL_IMAGE_SIZE - region.start;
540 /* Return if region can't contain decompressed kernel */
541 if (region.size < image_size)
544 /* If nothing overlaps, store the region and return. */
545 if (!mem_avoid_overlap(®ion, &overlap)) {
546 store_slot_info(®ion, image_size);
550 /* Store beginning of region if holds at least image_size. */
551 if (overlap.start > region.start + image_size) {
552 struct mem_vector beginning;
554 beginning.start = region.start;
555 beginning.size = overlap.start - region.start;
556 store_slot_info(&beginning, image_size);
559 /* Return if overlap extends to or past end of region. */
560 if (overlap.start + overlap.size >= region.start + region.size)
563 /* Clip off the overlapping region and start over. */
564 region.size -= overlap.start - region.start + overlap.size;
565 region.start = overlap.start + overlap.size;
571 * Returns true if mirror region found (and must have been processed
575 process_efi_entries(unsigned long minimum, unsigned long image_size)
577 struct efi_info *e = &boot_params->efi_info;
578 bool efi_mirror_found = false;
579 struct mem_vector region;
580 efi_memory_desc_t *md;
586 signature = (char *)&e->efi_loader_signature;
587 if (strncmp(signature, EFI32_LOADER_SIGNATURE, 4) &&
588 strncmp(signature, EFI64_LOADER_SIGNATURE, 4))
592 /* Can't handle data above 4GB at this time */
593 if (e->efi_memmap_hi) {
594 warn("EFI memmap is above 4GB, can't be handled now on x86_32. EFI should be disabled.\n");
597 pmap = e->efi_memmap;
599 pmap = (e->efi_memmap | ((__u64)e->efi_memmap_hi << 32));
602 nr_desc = e->efi_memmap_size / e->efi_memdesc_size;
603 for (i = 0; i < nr_desc; i++) {
604 md = efi_early_memdesc_ptr(pmap, e->efi_memdesc_size, i);
605 if (md->attribute & EFI_MEMORY_MORE_RELIABLE) {
606 efi_mirror_found = true;
611 for (i = 0; i < nr_desc; i++) {
612 md = efi_early_memdesc_ptr(pmap, e->efi_memdesc_size, i);
615 * Here we are more conservative in picking free memory than
616 * the EFI spec allows:
618 * According to the spec, EFI_BOOT_SERVICES_{CODE|DATA} are also
619 * free memory and thus available to place the kernel image into,
620 * but in practice there's firmware where using that memory leads
623 * Only EFI_CONVENTIONAL_MEMORY is guaranteed to be free.
625 if (md->type != EFI_CONVENTIONAL_MEMORY)
628 if (efi_mirror_found &&
629 !(md->attribute & EFI_MEMORY_MORE_RELIABLE))
632 region.start = md->phys_addr;
633 region.size = md->num_pages << EFI_PAGE_SHIFT;
634 process_mem_region(®ion, minimum, image_size);
635 if (slot_area_index == MAX_SLOT_AREA) {
636 debug_putstr("Aborted EFI scan (slot_areas full)!\n");
644 process_efi_entries(unsigned long minimum, unsigned long image_size)
650 static void process_e820_entries(unsigned long minimum,
651 unsigned long image_size)
654 struct mem_vector region;
655 struct boot_e820_entry *entry;
657 /* Verify potential e820 positions, appending to slots list. */
658 for (i = 0; i < boot_params->e820_entries; i++) {
659 entry = &boot_params->e820_table[i];
660 /* Skip non-RAM entries. */
661 if (entry->type != E820_TYPE_RAM)
663 region.start = entry->addr;
664 region.size = entry->size;
665 process_mem_region(®ion, minimum, image_size);
666 if (slot_area_index == MAX_SLOT_AREA) {
667 debug_putstr("Aborted e820 scan (slot_areas full)!\n");
673 static unsigned long find_random_phys_addr(unsigned long minimum,
674 unsigned long image_size)
676 /* Check if we had too many memmaps. */
677 if (memmap_too_large) {
678 debug_putstr("Aborted memory entries scan (more than 4 memmap= args)!\n");
682 /* Make sure minimum is aligned. */
683 minimum = ALIGN(minimum, CONFIG_PHYSICAL_ALIGN);
685 if (process_efi_entries(minimum, image_size))
686 return slots_fetch_random();
688 process_e820_entries(minimum, image_size);
689 return slots_fetch_random();
692 static unsigned long find_random_virt_addr(unsigned long minimum,
693 unsigned long image_size)
695 unsigned long slots, random_addr;
697 /* Make sure minimum is aligned. */
698 minimum = ALIGN(minimum, CONFIG_PHYSICAL_ALIGN);
699 /* Align image_size for easy slot calculations. */
700 image_size = ALIGN(image_size, CONFIG_PHYSICAL_ALIGN);
703 * There are how many CONFIG_PHYSICAL_ALIGN-sized slots
704 * that can hold image_size within the range of minimum to
707 slots = (KERNEL_IMAGE_SIZE - minimum - image_size) /
708 CONFIG_PHYSICAL_ALIGN + 1;
710 random_addr = kaslr_get_random_long("Virtual") % slots;
712 return random_addr * CONFIG_PHYSICAL_ALIGN + minimum;
716 * Since this function examines addresses much more numerically,
717 * it takes the input and output pointers as 'unsigned long'.
719 void choose_random_location(unsigned long input,
720 unsigned long input_size,
721 unsigned long *output,
722 unsigned long output_size,
723 unsigned long *virt_addr)
725 unsigned long random_addr, min_addr;
727 if (cmdline_find_option_bool("nokaslr")) {
728 warn("KASLR disabled: 'nokaslr' on cmdline.");
732 boot_params->hdr.loadflags |= KASLR_FLAG;
734 /* Prepare to add new identity pagetables on demand. */
735 initialize_identity_maps();
737 /* Record the various known unsafe memory ranges. */
738 mem_avoid_init(input, input_size, *output);
741 * Low end of the randomization range should be the
742 * smaller of 512M or the initial kernel image
745 min_addr = min(*output, 512UL << 20);
747 /* Walk available memory entries to find a random address. */
748 random_addr = find_random_phys_addr(min_addr, output_size);
750 warn("Physical KASLR disabled: no suitable memory region!");
752 /* Update the new physical address location. */
753 if (*output != random_addr) {
754 add_identity_map(random_addr, output_size);
755 *output = random_addr;
759 * This loads the identity mapping page table.
760 * This should only be done if a new physical address
761 * is found for the kernel, otherwise we should keep
762 * the old page table to make it be like the "nokaslr"
765 finalize_identity_maps();
769 /* Pick random virtual address starting from LOAD_PHYSICAL_ADDR. */
770 if (IS_ENABLED(CONFIG_X86_64))
771 random_addr = find_random_virt_addr(LOAD_PHYSICAL_ADDR, output_size);
772 *virt_addr = random_addr;