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.
15 #include <generated/compile.h>
16 #include <linux/module.h>
17 #include <linux/uts.h>
18 #include <linux/utsname.h>
19 #include <generated/utsrelease.h>
21 /* Simplified build-specific string for starting entropy. */
22 static const char build_str
[] = UTS_RELEASE
" (" LINUX_COMPILE_BY
"@"
23 LINUX_COMPILE_HOST
") (" LINUX_COMPILER
") " UTS_VERSION
;
25 static unsigned long rotate_xor(unsigned long hash
, const void *area
,
29 unsigned long *ptr
= (unsigned long *)area
;
31 for (i
= 0; i
< size
/ sizeof(hash
); i
++) {
32 /* Rotate by odd number of bits and XOR. */
33 hash
= (hash
<< ((sizeof(hash
) * 8) - 7)) | (hash
>> 7);
40 /* Attempt to create a simple but unpredictable starting entropy. */
41 static unsigned long get_boot_seed(void)
43 unsigned long hash
= 0;
45 hash
= rotate_xor(hash
, build_str
, sizeof(build_str
));
46 hash
= rotate_xor(hash
, boot_params
, sizeof(*boot_params
));
51 #define KASLR_COMPRESSED_BOOT
52 #include "../../lib/kaslr.c"
59 enum mem_avoid_index
{
60 MEM_AVOID_ZO_RANGE
= 0,
67 static struct mem_vector mem_avoid
[MEM_AVOID_MAX
];
69 static bool mem_overlaps(struct mem_vector
*one
, struct mem_vector
*two
)
71 /* Item one is entirely before item two. */
72 if (one
->start
+ one
->size
<= two
->start
)
74 /* Item one is entirely after item two. */
75 if (one
->start
>= two
->start
+ two
->size
)
81 * In theory, KASLR can put the kernel anywhere in the range of [16M, 64T).
82 * The mem_avoid array is used to store the ranges that need to be avoided
83 * when KASLR searches for an appropriate random address. We must avoid any
84 * regions that are unsafe to overlap with during decompression, and other
85 * things like the initrd, cmdline and boot_params. This comment seeks to
86 * explain mem_avoid as clearly as possible since incorrect mem_avoid
87 * memory ranges lead to really hard to debug boot failures.
89 * The initrd, cmdline, and boot_params are trivial to identify for
90 * avoiding. They are MEM_AVOID_INITRD, MEM_AVOID_CMDLINE, and
91 * MEM_AVOID_BOOTPARAMS respectively below.
93 * What is not obvious how to avoid is the range of memory that is used
94 * during decompression (MEM_AVOID_ZO_RANGE below). This range must cover
95 * the compressed kernel (ZO) and its run space, which is used to extract
96 * the uncompressed kernel (VO) and relocs.
98 * ZO's full run size sits against the end of the decompression buffer, so
99 * we can calculate where text, data, bss, etc of ZO are positioned more
102 * For additional background, the decompression calculations can be found
103 * in header.S, and the memory diagram is based on the one found in misc.c.
105 * The following conditions are already enforced by the image layouts and
107 * - input + input_size >= output + output_size
108 * - kernel_total_size <= init_size
109 * - kernel_total_size <= output_size (see Note below)
110 * - output + init_size >= output + output_size
112 * (Note that kernel_total_size and output_size have no fundamental
113 * relationship, but output_size is passed to choose_random_location
114 * as a maximum of the two. The diagram is showing a case where
115 * kernel_total_size is larger than output_size, but this case is
116 * handled by bumping output_size.)
118 * The above conditions can be illustrated by a diagram:
120 * 0 output input input+input_size output+init_size
123 * |-----|--------|--------|--------------|-----------|--|-------------|
126 * output+init_size-ZO_INIT_SIZE output+output_size output+kernel_total_size
128 * [output, output+init_size) is the entire memory range used for
129 * extracting the compressed image.
131 * [output, output+kernel_total_size) is the range needed for the
132 * uncompressed kernel (VO) and its run size (bss, brk, etc).
134 * [output, output+output_size) is VO plus relocs (i.e. the entire
135 * uncompressed payload contained by ZO). This is the area of the buffer
136 * written to during decompression.
138 * [output+init_size-ZO_INIT_SIZE, output+init_size) is the worst-case
139 * range of the copied ZO and decompression code. (i.e. the range
140 * covered backwards of size ZO_INIT_SIZE, starting from output+init_size.)
142 * [input, input+input_size) is the original copied compressed image (ZO)
143 * (i.e. it does not include its run size). This range must be avoided
144 * because it contains the data used for decompression.
146 * [input+input_size, output+init_size) is [_text, _end) for ZO. This
147 * range includes ZO's heap and stack, and must be avoided since it
148 * performs the decompression.
150 * Since the above two ranges need to be avoided and they are adjacent,
151 * they can be merged, resulting in: [input, output+init_size) which
152 * becomes the MEM_AVOID_ZO_RANGE below.
154 static void mem_avoid_init(unsigned long input
, unsigned long input_size
,
155 unsigned long output
)
157 unsigned long init_size
= boot_params
->hdr
.init_size
;
158 u64 initrd_start
, initrd_size
;
159 u64 cmd_line
, cmd_line_size
;
163 * Avoid the region that is unsafe to overlap during
166 mem_avoid
[MEM_AVOID_ZO_RANGE
].start
= input
;
167 mem_avoid
[MEM_AVOID_ZO_RANGE
].size
= (output
+ init_size
) - input
;
168 add_identity_map(mem_avoid
[MEM_AVOID_ZO_RANGE
].start
,
169 mem_avoid
[MEM_AVOID_ZO_RANGE
].size
);
172 initrd_start
= (u64
)boot_params
->ext_ramdisk_image
<< 32;
173 initrd_start
|= boot_params
->hdr
.ramdisk_image
;
174 initrd_size
= (u64
)boot_params
->ext_ramdisk_size
<< 32;
175 initrd_size
|= boot_params
->hdr
.ramdisk_size
;
176 mem_avoid
[MEM_AVOID_INITRD
].start
= initrd_start
;
177 mem_avoid
[MEM_AVOID_INITRD
].size
= initrd_size
;
178 /* No need to set mapping for initrd, it will be handled in VO. */
180 /* Avoid kernel command line. */
181 cmd_line
= (u64
)boot_params
->ext_cmd_line_ptr
<< 32;
182 cmd_line
|= boot_params
->hdr
.cmd_line_ptr
;
183 /* Calculate size of cmd_line. */
184 ptr
= (char *)(unsigned long)cmd_line
;
185 for (cmd_line_size
= 0; ptr
[cmd_line_size
++]; )
187 mem_avoid
[MEM_AVOID_CMDLINE
].start
= cmd_line
;
188 mem_avoid
[MEM_AVOID_CMDLINE
].size
= cmd_line_size
;
189 add_identity_map(mem_avoid
[MEM_AVOID_CMDLINE
].start
,
190 mem_avoid
[MEM_AVOID_CMDLINE
].size
);
192 /* Avoid boot parameters. */
193 mem_avoid
[MEM_AVOID_BOOTPARAMS
].start
= (unsigned long)boot_params
;
194 mem_avoid
[MEM_AVOID_BOOTPARAMS
].size
= sizeof(*boot_params
);
195 add_identity_map(mem_avoid
[MEM_AVOID_BOOTPARAMS
].start
,
196 mem_avoid
[MEM_AVOID_BOOTPARAMS
].size
);
198 /* We don't need to set a mapping for setup_data. */
200 #ifdef CONFIG_X86_VERBOSE_BOOTUP
201 /* Make sure video RAM can be used. */
202 add_identity_map(0, PMD_SIZE
);
207 * Does this memory vector overlap a known avoided area? If so, record the
208 * overlap region with the lowest address.
210 static bool mem_avoid_overlap(struct mem_vector
*img
,
211 struct mem_vector
*overlap
)
214 struct setup_data
*ptr
;
215 unsigned long earliest
= img
->start
+ img
->size
;
216 bool is_overlapping
= false;
218 for (i
= 0; i
< MEM_AVOID_MAX
; i
++) {
219 if (mem_overlaps(img
, &mem_avoid
[i
]) &&
220 mem_avoid
[i
].start
< earliest
) {
221 *overlap
= mem_avoid
[i
];
222 earliest
= overlap
->start
;
223 is_overlapping
= true;
227 /* Avoid all entries in the setup_data linked list. */
228 ptr
= (struct setup_data
*)(unsigned long)boot_params
->hdr
.setup_data
;
230 struct mem_vector avoid
;
232 avoid
.start
= (unsigned long)ptr
;
233 avoid
.size
= sizeof(*ptr
) + ptr
->len
;
235 if (mem_overlaps(img
, &avoid
) && (avoid
.start
< earliest
)) {
237 earliest
= overlap
->start
;
238 is_overlapping
= true;
241 ptr
= (struct setup_data
*)(unsigned long)ptr
->next
;
244 return is_overlapping
;
252 #define MAX_SLOT_AREA 100
254 static struct slot_area slot_areas
[MAX_SLOT_AREA
];
256 static unsigned long slot_max
;
258 static unsigned long slot_area_index
;
260 static void store_slot_info(struct mem_vector
*region
, unsigned long image_size
)
262 struct slot_area slot_area
;
264 if (slot_area_index
== MAX_SLOT_AREA
)
267 slot_area
.addr
= region
->start
;
268 slot_area
.num
= (region
->size
- image_size
) /
269 CONFIG_PHYSICAL_ALIGN
+ 1;
271 if (slot_area
.num
> 0) {
272 slot_areas
[slot_area_index
++] = slot_area
;
273 slot_max
+= slot_area
.num
;
277 static unsigned long slots_fetch_random(void)
282 /* Handle case of no slots stored. */
286 slot
= kaslr_get_random_long("Physical") % slot_max
;
288 for (i
= 0; i
< slot_area_index
; i
++) {
289 if (slot
>= slot_areas
[i
].num
) {
290 slot
-= slot_areas
[i
].num
;
293 return slot_areas
[i
].addr
+ slot
* CONFIG_PHYSICAL_ALIGN
;
296 if (i
== slot_area_index
)
297 debug_putstr("slots_fetch_random() failed!?\n");
301 static void process_e820_entry(struct e820entry
*entry
,
302 unsigned long minimum
,
303 unsigned long image_size
)
305 struct mem_vector region
, overlap
;
306 struct slot_area slot_area
;
307 unsigned long start_orig
;
309 /* Skip non-RAM entries. */
310 if (entry
->type
!= E820_RAM
)
313 /* On 32-bit, ignore entries entirely above our maximum. */
314 if (IS_ENABLED(CONFIG_X86_32
) && entry
->addr
>= KERNEL_IMAGE_SIZE
)
317 /* Ignore entries entirely below our minimum. */
318 if (entry
->addr
+ entry
->size
< minimum
)
321 region
.start
= entry
->addr
;
322 region
.size
= entry
->size
;
324 /* Give up if slot area array is full. */
325 while (slot_area_index
< MAX_SLOT_AREA
) {
326 start_orig
= region
.start
;
328 /* Potentially raise address to minimum location. */
329 if (region
.start
< minimum
)
330 region
.start
= minimum
;
332 /* Potentially raise address to meet alignment needs. */
333 region
.start
= ALIGN(region
.start
, CONFIG_PHYSICAL_ALIGN
);
335 /* Did we raise the address above this e820 region? */
336 if (region
.start
> entry
->addr
+ entry
->size
)
339 /* Reduce size by any delta from the original address. */
340 region
.size
-= region
.start
- start_orig
;
342 /* On 32-bit, reduce region size to fit within max size. */
343 if (IS_ENABLED(CONFIG_X86_32
) &&
344 region
.start
+ region
.size
> KERNEL_IMAGE_SIZE
)
345 region
.size
= KERNEL_IMAGE_SIZE
- region
.start
;
347 /* Return if region can't contain decompressed kernel */
348 if (region
.size
< image_size
)
351 /* If nothing overlaps, store the region and return. */
352 if (!mem_avoid_overlap(®ion
, &overlap
)) {
353 store_slot_info(®ion
, image_size
);
357 /* Store beginning of region if holds at least image_size. */
358 if (overlap
.start
> region
.start
+ image_size
) {
359 struct mem_vector beginning
;
361 beginning
.start
= region
.start
;
362 beginning
.size
= overlap
.start
- region
.start
;
363 store_slot_info(&beginning
, image_size
);
366 /* Return if overlap extends to or past end of region. */
367 if (overlap
.start
+ overlap
.size
>= region
.start
+ region
.size
)
370 /* Clip off the overlapping region and start over. */
371 region
.size
-= overlap
.start
- region
.start
+ overlap
.size
;
372 region
.start
= overlap
.start
+ overlap
.size
;
376 static unsigned long find_random_phys_addr(unsigned long minimum
,
377 unsigned long image_size
)
382 /* Make sure minimum is aligned. */
383 minimum
= ALIGN(minimum
, CONFIG_PHYSICAL_ALIGN
);
385 /* Verify potential e820 positions, appending to slots list. */
386 for (i
= 0; i
< boot_params
->e820_entries
; i
++) {
387 process_e820_entry(&boot_params
->e820_map
[i
], minimum
,
389 if (slot_area_index
== MAX_SLOT_AREA
) {
390 debug_putstr("Aborted e820 scan (slot_areas full)!\n");
395 return slots_fetch_random();
398 static unsigned long find_random_virt_addr(unsigned long minimum
,
399 unsigned long image_size
)
401 unsigned long slots
, random_addr
;
403 /* Make sure minimum is aligned. */
404 minimum
= ALIGN(minimum
, CONFIG_PHYSICAL_ALIGN
);
405 /* Align image_size for easy slot calculations. */
406 image_size
= ALIGN(image_size
, CONFIG_PHYSICAL_ALIGN
);
409 * There are how many CONFIG_PHYSICAL_ALIGN-sized slots
410 * that can hold image_size within the range of minimum to
413 slots
= (KERNEL_IMAGE_SIZE
- minimum
- image_size
) /
414 CONFIG_PHYSICAL_ALIGN
+ 1;
416 random_addr
= kaslr_get_random_long("Virtual") % slots
;
418 return random_addr
* CONFIG_PHYSICAL_ALIGN
+ minimum
;
422 * Since this function examines addresses much more numerically,
423 * it takes the input and output pointers as 'unsigned long'.
425 void choose_random_location(unsigned long input
,
426 unsigned long input_size
,
427 unsigned long *output
,
428 unsigned long output_size
,
429 unsigned long *virt_addr
)
431 unsigned long random_addr
, min_addr
;
433 if (cmdline_find_option_bool("nokaslr")) {
434 warn("KASLR disabled: 'nokaslr' on cmdline.");
438 boot_params
->hdr
.loadflags
|= KASLR_FLAG
;
440 /* Prepare to add new identity pagetables on demand. */
441 initialize_identity_maps();
443 /* Record the various known unsafe memory ranges. */
444 mem_avoid_init(input
, input_size
, *output
);
447 * Low end of the randomization range should be the
448 * smaller of 512M or the initial kernel image
451 min_addr
= min(*output
, 512UL << 20);
453 /* Walk e820 and find a random address. */
454 random_addr
= find_random_phys_addr(min_addr
, output_size
);
456 warn("KASLR disabled: could not find suitable E820 region!");
458 /* Update the new physical address location. */
459 if (*output
!= random_addr
) {
460 add_identity_map(random_addr
, output_size
);
461 *output
= random_addr
;
465 * This loads the identity mapping page table.
466 * This should only be done if a new physical address
467 * is found for the kernel, otherwise we should keep
468 * the old page table to make it be like the "nokaslr"
471 finalize_identity_maps();
475 /* Pick random virtual address starting from LOAD_PHYSICAL_ADDR. */
476 if (IS_ENABLED(CONFIG_X86_64
))
477 random_addr
= find_random_virt_addr(LOAD_PHYSICAL_ADDR
, output_size
);
478 *virt_addr
= random_addr
;