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 extern unsigned long get_cmd_line_ptr(void);
51 /* Simplified build-specific string for starting entropy. */
52 static const char build_str[] = UTS_RELEASE " (" LINUX_COMPILE_BY "@"
53 LINUX_COMPILE_HOST ") (" LINUX_COMPILER ") " UTS_VERSION;
55 static unsigned long rotate_xor(unsigned long hash, const void *area,
59 unsigned long *ptr = (unsigned long *)area;
61 for (i = 0; i < size / sizeof(hash); i++) {
62 /* Rotate by odd number of bits and XOR. */
63 hash = (hash << ((sizeof(hash) * 8) - 7)) | (hash >> 7);
70 /* Attempt to create a simple but unpredictable starting entropy. */
71 static unsigned long get_boot_seed(void)
73 unsigned long hash = 0;
75 hash = rotate_xor(hash, build_str, sizeof(build_str));
76 hash = rotate_xor(hash, boot_params, sizeof(*boot_params));
81 #define KASLR_COMPRESSED_BOOT
82 #include "../../lib/kaslr.c"
85 unsigned long long start;
86 unsigned long long size;
89 /* Only supporting at most 4 unusable memmap regions with kaslr */
90 #define MAX_MEMMAP_REGIONS 4
92 static bool memmap_too_large;
95 /* Store memory limit specified by "mem=nn[KMG]" or "memmap=nn[KMG]" */
96 unsigned long long mem_limit = ULLONG_MAX;
99 enum mem_avoid_index {
100 MEM_AVOID_ZO_RANGE = 0,
103 MEM_AVOID_BOOTPARAMS,
104 MEM_AVOID_MEMMAP_BEGIN,
105 MEM_AVOID_MEMMAP_END = MEM_AVOID_MEMMAP_BEGIN + MAX_MEMMAP_REGIONS - 1,
109 static struct mem_vector mem_avoid[MEM_AVOID_MAX];
111 static bool mem_overlaps(struct mem_vector *one, struct mem_vector *two)
113 /* Item one is entirely before item two. */
114 if (one->start + one->size <= two->start)
116 /* Item one is entirely after item two. */
117 if (one->start >= two->start + two->size)
122 char *skip_spaces(const char *str)
124 while (isspace(*str))
128 #include "../../../../lib/ctype.c"
129 #include "../../../../lib/cmdline.c"
132 parse_memmap(char *p, unsigned long long *start, unsigned long long *size)
139 /* We don't care about this option here */
140 if (!strncmp(p, "exactmap", 8))
144 *size = memparse(p, &p);
152 *start = memparse(p + 1, &p);
155 /* memmap=nn@ss specifies usable region, should be skipped */
160 * If w/o offset, only size specified, memmap=nn[KMG] has the
161 * same behaviour as mem=nn[KMG]. It limits the max address
162 * system can use. Region above the limit should be avoided.
171 static void mem_avoid_memmap(char *str)
176 if (i >= MAX_MEMMAP_REGIONS)
179 while (str && (i < MAX_MEMMAP_REGIONS)) {
181 unsigned long long start, size;
182 char *k = strchr(str, ',');
187 rc = parse_memmap(str, &start, &size);
193 /* Store the specified memory limit if size > 0 */
200 mem_avoid[MEM_AVOID_MEMMAP_BEGIN + i].start = start;
201 mem_avoid[MEM_AVOID_MEMMAP_BEGIN + i].size = size;
205 /* More than 4 memmaps, fail kaslr */
206 if ((i >= MAX_MEMMAP_REGIONS) && str)
207 memmap_too_large = true;
210 static int handle_mem_memmap(void)
212 char *args = (char *)get_cmd_line_ptr();
213 size_t len = strlen((char *)args);
218 if (!strstr(args, "memmap=") && !strstr(args, "mem="))
221 tmp_cmdline = malloc(len + 1);
223 error("Failed to allocate space for tmp_cmdline");
225 memcpy(tmp_cmdline, args, len);
226 tmp_cmdline[len] = 0;
229 /* Chew leading spaces */
230 args = skip_spaces(args);
233 args = next_arg(args, ¶m, &val);
235 if (!val && strcmp(param, "--") == 0) {
236 warn("Only '--' specified in cmdline");
241 if (!strcmp(param, "memmap")) {
242 mem_avoid_memmap(val);
243 } else if (!strcmp(param, "mem")) {
246 if (!strcmp(p, "nopentium"))
248 mem_size = memparse(p, &p);
253 mem_limit = mem_size;
262 * In theory, KASLR can put the kernel anywhere in the range of [16M, 64T).
263 * The mem_avoid array is used to store the ranges that need to be avoided
264 * when KASLR searches for an appropriate random address. We must avoid any
265 * regions that are unsafe to overlap with during decompression, and other
266 * things like the initrd, cmdline and boot_params. This comment seeks to
267 * explain mem_avoid as clearly as possible since incorrect mem_avoid
268 * memory ranges lead to really hard to debug boot failures.
270 * The initrd, cmdline, and boot_params are trivial to identify for
271 * avoiding. They are MEM_AVOID_INITRD, MEM_AVOID_CMDLINE, and
272 * MEM_AVOID_BOOTPARAMS respectively below.
274 * What is not obvious how to avoid is the range of memory that is used
275 * during decompression (MEM_AVOID_ZO_RANGE below). This range must cover
276 * the compressed kernel (ZO) and its run space, which is used to extract
277 * the uncompressed kernel (VO) and relocs.
279 * ZO's full run size sits against the end of the decompression buffer, so
280 * we can calculate where text, data, bss, etc of ZO are positioned more
283 * For additional background, the decompression calculations can be found
284 * in header.S, and the memory diagram is based on the one found in misc.c.
286 * The following conditions are already enforced by the image layouts and
288 * - input + input_size >= output + output_size
289 * - kernel_total_size <= init_size
290 * - kernel_total_size <= output_size (see Note below)
291 * - output + init_size >= output + output_size
293 * (Note that kernel_total_size and output_size have no fundamental
294 * relationship, but output_size is passed to choose_random_location
295 * as a maximum of the two. The diagram is showing a case where
296 * kernel_total_size is larger than output_size, but this case is
297 * handled by bumping output_size.)
299 * The above conditions can be illustrated by a diagram:
301 * 0 output input input+input_size output+init_size
304 * |-----|--------|--------|--------------|-----------|--|-------------|
307 * output+init_size-ZO_INIT_SIZE output+output_size output+kernel_total_size
309 * [output, output+init_size) is the entire memory range used for
310 * extracting the compressed image.
312 * [output, output+kernel_total_size) is the range needed for the
313 * uncompressed kernel (VO) and its run size (bss, brk, etc).
315 * [output, output+output_size) is VO plus relocs (i.e. the entire
316 * uncompressed payload contained by ZO). This is the area of the buffer
317 * written to during decompression.
319 * [output+init_size-ZO_INIT_SIZE, output+init_size) is the worst-case
320 * range of the copied ZO and decompression code. (i.e. the range
321 * covered backwards of size ZO_INIT_SIZE, starting from output+init_size.)
323 * [input, input+input_size) is the original copied compressed image (ZO)
324 * (i.e. it does not include its run size). This range must be avoided
325 * because it contains the data used for decompression.
327 * [input+input_size, output+init_size) is [_text, _end) for ZO. This
328 * range includes ZO's heap and stack, and must be avoided since it
329 * performs the decompression.
331 * Since the above two ranges need to be avoided and they are adjacent,
332 * they can be merged, resulting in: [input, output+init_size) which
333 * becomes the MEM_AVOID_ZO_RANGE below.
335 static void mem_avoid_init(unsigned long input, unsigned long input_size,
336 unsigned long output)
338 unsigned long init_size = boot_params->hdr.init_size;
339 u64 initrd_start, initrd_size;
340 u64 cmd_line, cmd_line_size;
344 * Avoid the region that is unsafe to overlap during
347 mem_avoid[MEM_AVOID_ZO_RANGE].start = input;
348 mem_avoid[MEM_AVOID_ZO_RANGE].size = (output + init_size) - input;
349 add_identity_map(mem_avoid[MEM_AVOID_ZO_RANGE].start,
350 mem_avoid[MEM_AVOID_ZO_RANGE].size);
353 initrd_start = (u64)boot_params->ext_ramdisk_image << 32;
354 initrd_start |= boot_params->hdr.ramdisk_image;
355 initrd_size = (u64)boot_params->ext_ramdisk_size << 32;
356 initrd_size |= boot_params->hdr.ramdisk_size;
357 mem_avoid[MEM_AVOID_INITRD].start = initrd_start;
358 mem_avoid[MEM_AVOID_INITRD].size = initrd_size;
359 /* No need to set mapping for initrd, it will be handled in VO. */
361 /* Avoid kernel command line. */
362 cmd_line = (u64)boot_params->ext_cmd_line_ptr << 32;
363 cmd_line |= boot_params->hdr.cmd_line_ptr;
364 /* Calculate size of cmd_line. */
365 ptr = (char *)(unsigned long)cmd_line;
366 for (cmd_line_size = 0; ptr[cmd_line_size++]; )
368 mem_avoid[MEM_AVOID_CMDLINE].start = cmd_line;
369 mem_avoid[MEM_AVOID_CMDLINE].size = cmd_line_size;
370 add_identity_map(mem_avoid[MEM_AVOID_CMDLINE].start,
371 mem_avoid[MEM_AVOID_CMDLINE].size);
373 /* Avoid boot parameters. */
374 mem_avoid[MEM_AVOID_BOOTPARAMS].start = (unsigned long)boot_params;
375 mem_avoid[MEM_AVOID_BOOTPARAMS].size = sizeof(*boot_params);
376 add_identity_map(mem_avoid[MEM_AVOID_BOOTPARAMS].start,
377 mem_avoid[MEM_AVOID_BOOTPARAMS].size);
379 /* We don't need to set a mapping for setup_data. */
381 /* Mark the memmap regions we need to avoid */
384 #ifdef CONFIG_X86_VERBOSE_BOOTUP
385 /* Make sure video RAM can be used. */
386 add_identity_map(0, PMD_SIZE);
391 * Does this memory vector overlap a known avoided area? If so, record the
392 * overlap region with the lowest address.
394 static bool mem_avoid_overlap(struct mem_vector *img,
395 struct mem_vector *overlap)
398 struct setup_data *ptr;
399 unsigned long earliest = img->start + img->size;
400 bool is_overlapping = false;
402 for (i = 0; i < MEM_AVOID_MAX; i++) {
403 if (mem_overlaps(img, &mem_avoid[i]) &&
404 mem_avoid[i].start < earliest) {
405 *overlap = mem_avoid[i];
406 earliest = overlap->start;
407 is_overlapping = true;
411 /* Avoid all entries in the setup_data linked list. */
412 ptr = (struct setup_data *)(unsigned long)boot_params->hdr.setup_data;
414 struct mem_vector avoid;
416 avoid.start = (unsigned long)ptr;
417 avoid.size = sizeof(*ptr) + ptr->len;
419 if (mem_overlaps(img, &avoid) && (avoid.start < earliest)) {
421 earliest = overlap->start;
422 is_overlapping = true;
425 ptr = (struct setup_data *)(unsigned long)ptr->next;
428 return is_overlapping;
436 #define MAX_SLOT_AREA 100
438 static struct slot_area slot_areas[MAX_SLOT_AREA];
440 static unsigned long slot_max;
442 static unsigned long slot_area_index;
444 static void store_slot_info(struct mem_vector *region, unsigned long image_size)
446 struct slot_area slot_area;
448 if (slot_area_index == MAX_SLOT_AREA)
451 slot_area.addr = region->start;
452 slot_area.num = (region->size - image_size) /
453 CONFIG_PHYSICAL_ALIGN + 1;
455 if (slot_area.num > 0) {
456 slot_areas[slot_area_index++] = slot_area;
457 slot_max += slot_area.num;
461 static unsigned long slots_fetch_random(void)
466 /* Handle case of no slots stored. */
470 slot = kaslr_get_random_long("Physical") % slot_max;
472 for (i = 0; i < slot_area_index; i++) {
473 if (slot >= slot_areas[i].num) {
474 slot -= slot_areas[i].num;
477 return slot_areas[i].addr + slot * CONFIG_PHYSICAL_ALIGN;
480 if (i == slot_area_index)
481 debug_putstr("slots_fetch_random() failed!?\n");
485 static void process_mem_region(struct mem_vector *entry,
486 unsigned long minimum,
487 unsigned long image_size)
489 struct mem_vector region, overlap;
490 struct slot_area slot_area;
491 unsigned long start_orig, end;
492 struct mem_vector cur_entry;
494 /* On 32-bit, ignore entries entirely above our maximum. */
495 if (IS_ENABLED(CONFIG_X86_32) && entry->start >= KERNEL_IMAGE_SIZE)
498 /* Ignore entries entirely below our minimum. */
499 if (entry->start + entry->size < minimum)
502 /* Ignore entries above memory limit */
503 end = min(entry->size + entry->start, mem_limit);
504 if (entry->start >= end)
506 cur_entry.start = entry->start;
507 cur_entry.size = end - entry->start;
509 region.start = cur_entry.start;
510 region.size = cur_entry.size;
512 /* Give up if slot area array is full. */
513 while (slot_area_index < MAX_SLOT_AREA) {
514 start_orig = region.start;
516 /* Potentially raise address to minimum location. */
517 if (region.start < minimum)
518 region.start = minimum;
520 /* Potentially raise address to meet alignment needs. */
521 region.start = ALIGN(region.start, CONFIG_PHYSICAL_ALIGN);
523 /* Did we raise the address above the passed in memory entry? */
524 if (region.start > cur_entry.start + cur_entry.size)
527 /* Reduce size by any delta from the original address. */
528 region.size -= region.start - start_orig;
530 /* On 32-bit, reduce region size to fit within max size. */
531 if (IS_ENABLED(CONFIG_X86_32) &&
532 region.start + region.size > KERNEL_IMAGE_SIZE)
533 region.size = KERNEL_IMAGE_SIZE - region.start;
535 /* Return if region can't contain decompressed kernel */
536 if (region.size < image_size)
539 /* If nothing overlaps, store the region and return. */
540 if (!mem_avoid_overlap(®ion, &overlap)) {
541 store_slot_info(®ion, image_size);
545 /* Store beginning of region if holds at least image_size. */
546 if (overlap.start > region.start + image_size) {
547 struct mem_vector beginning;
549 beginning.start = region.start;
550 beginning.size = overlap.start - region.start;
551 store_slot_info(&beginning, image_size);
554 /* Return if overlap extends to or past end of region. */
555 if (overlap.start + overlap.size >= region.start + region.size)
558 /* Clip off the overlapping region and start over. */
559 region.size -= overlap.start - region.start + overlap.size;
560 region.start = overlap.start + overlap.size;
566 * Returns true if mirror region found (and must have been processed
570 process_efi_entries(unsigned long minimum, unsigned long image_size)
572 struct efi_info *e = &boot_params->efi_info;
573 bool efi_mirror_found = false;
574 struct mem_vector region;
575 efi_memory_desc_t *md;
581 signature = (char *)&e->efi_loader_signature;
582 if (strncmp(signature, EFI32_LOADER_SIGNATURE, 4) &&
583 strncmp(signature, EFI64_LOADER_SIGNATURE, 4))
587 /* Can't handle data above 4GB at this time */
588 if (e->efi_memmap_hi) {
589 warn("EFI memmap is above 4GB, can't be handled now on x86_32. EFI should be disabled.\n");
592 pmap = e->efi_memmap;
594 pmap = (e->efi_memmap | ((__u64)e->efi_memmap_hi << 32));
597 nr_desc = e->efi_memmap_size / e->efi_memdesc_size;
598 for (i = 0; i < nr_desc; i++) {
599 md = efi_early_memdesc_ptr(pmap, e->efi_memdesc_size, i);
600 if (md->attribute & EFI_MEMORY_MORE_RELIABLE) {
601 efi_mirror_found = true;
606 for (i = 0; i < nr_desc; i++) {
607 md = efi_early_memdesc_ptr(pmap, e->efi_memdesc_size, i);
610 * Here we are more conservative in picking free memory than
611 * the EFI spec allows:
613 * According to the spec, EFI_BOOT_SERVICES_{CODE|DATA} are also
614 * free memory and thus available to place the kernel image into,
615 * but in practice there's firmware where using that memory leads
618 * Only EFI_CONVENTIONAL_MEMORY is guaranteed to be free.
620 if (md->type != EFI_CONVENTIONAL_MEMORY)
623 if (efi_mirror_found &&
624 !(md->attribute & EFI_MEMORY_MORE_RELIABLE))
627 region.start = md->phys_addr;
628 region.size = md->num_pages << EFI_PAGE_SHIFT;
629 process_mem_region(®ion, minimum, image_size);
630 if (slot_area_index == MAX_SLOT_AREA) {
631 debug_putstr("Aborted EFI scan (slot_areas full)!\n");
639 process_efi_entries(unsigned long minimum, unsigned long image_size)
645 static void process_e820_entries(unsigned long minimum,
646 unsigned long image_size)
649 struct mem_vector region;
650 struct boot_e820_entry *entry;
652 /* Verify potential e820 positions, appending to slots list. */
653 for (i = 0; i < boot_params->e820_entries; i++) {
654 entry = &boot_params->e820_table[i];
655 /* Skip non-RAM entries. */
656 if (entry->type != E820_TYPE_RAM)
658 region.start = entry->addr;
659 region.size = entry->size;
660 process_mem_region(®ion, minimum, image_size);
661 if (slot_area_index == MAX_SLOT_AREA) {
662 debug_putstr("Aborted e820 scan (slot_areas full)!\n");
668 static unsigned long find_random_phys_addr(unsigned long minimum,
669 unsigned long image_size)
671 /* Check if we had too many memmaps. */
672 if (memmap_too_large) {
673 debug_putstr("Aborted memory entries scan (more than 4 memmap= args)!\n");
677 /* Make sure minimum is aligned. */
678 minimum = ALIGN(minimum, CONFIG_PHYSICAL_ALIGN);
680 if (process_efi_entries(minimum, image_size))
681 return slots_fetch_random();
683 process_e820_entries(minimum, image_size);
684 return slots_fetch_random();
687 static unsigned long find_random_virt_addr(unsigned long minimum,
688 unsigned long image_size)
690 unsigned long slots, random_addr;
692 /* Make sure minimum is aligned. */
693 minimum = ALIGN(minimum, CONFIG_PHYSICAL_ALIGN);
694 /* Align image_size for easy slot calculations. */
695 image_size = ALIGN(image_size, CONFIG_PHYSICAL_ALIGN);
698 * There are how many CONFIG_PHYSICAL_ALIGN-sized slots
699 * that can hold image_size within the range of minimum to
702 slots = (KERNEL_IMAGE_SIZE - minimum - image_size) /
703 CONFIG_PHYSICAL_ALIGN + 1;
705 random_addr = kaslr_get_random_long("Virtual") % slots;
707 return random_addr * CONFIG_PHYSICAL_ALIGN + minimum;
711 * Since this function examines addresses much more numerically,
712 * it takes the input and output pointers as 'unsigned long'.
714 void choose_random_location(unsigned long input,
715 unsigned long input_size,
716 unsigned long *output,
717 unsigned long output_size,
718 unsigned long *virt_addr)
720 unsigned long random_addr, min_addr;
722 if (cmdline_find_option_bool("nokaslr")) {
723 warn("KASLR disabled: 'nokaslr' on cmdline.");
727 boot_params->hdr.loadflags |= KASLR_FLAG;
729 /* Prepare to add new identity pagetables on demand. */
730 initialize_identity_maps();
732 /* Record the various known unsafe memory ranges. */
733 mem_avoid_init(input, input_size, *output);
736 * Low end of the randomization range should be the
737 * smaller of 512M or the initial kernel image
740 min_addr = min(*output, 512UL << 20);
742 /* Walk available memory entries to find a random address. */
743 random_addr = find_random_phys_addr(min_addr, output_size);
745 warn("Physical KASLR disabled: no suitable memory region!");
747 /* Update the new physical address location. */
748 if (*output != random_addr) {
749 add_identity_map(random_addr, output_size);
750 *output = random_addr;
754 * This loads the identity mapping page table.
755 * This should only be done if a new physical address
756 * is found for the kernel, otherwise we should keep
757 * the old page table to make it be like the "nokaslr"
760 finalize_identity_maps();
764 /* Pick random virtual address starting from LOAD_PHYSICAL_ADDR. */
765 if (IS_ENABLED(CONFIG_X86_64))
766 random_addr = find_random_virt_addr(LOAD_PHYSICAL_ADDR, output_size);
767 *virt_addr = random_addr;