1// SPDX-License-Identifier: GPL-2.0
2/*
3 * This file implements KASLR memory randomization for x86_64. It randomizes
4 * the virtual address space of kernel memory regions (physical memory
5 * mapping, vmalloc & vmemmap) for x86_64. This security feature mitigates
6 * exploits relying on predictable kernel addresses.
7 *
8 * Entropy is generated using the KASLR early boot functions now shared in
9 * the lib directory (originally written by Kees Cook). Randomization is
10 * done on PGD & P4D/PUD page table levels to increase possible addresses.
11 * The physical memory mapping code was adapted to support P4D/PUD level
12 * virtual addresses. This implementation on the best configuration provides
13 * 30,000 possible virtual addresses in average for each memory region.
14 * An additional low memory page is used to ensure each CPU can start with
15 * a PGD aligned virtual address (for realmode).
16 *
17 * The order of each memory region is not changed. The feature looks at
18 * the available space for the regions based on different configuration
19 * options and randomizes the base and space between each. The size of the
20 * physical memory mapping is the available physical memory.
21 */
22
23#include <linux/kernel.h>
24#include <linux/init.h>
25#include <linux/prandom.h>
26#include <linux/memblock.h>
27#include <linux/pgtable.h>
28
29#include <asm/setup.h>
30#include <asm/kaslr.h>
31
32#include "mm_internal.h"
33
34#define TB_SHIFT 40
35
36/*
37 * The end address could depend on more configuration options to make the
38 * highest amount of space for randomization available, but that's too hard
39 * to keep straight and caused issues already.
40 */
41static const unsigned long vaddr_end = CPU_ENTRY_AREA_BASE;
42
43/*
44 * Memory regions randomized by KASLR (except modules that use a separate logic
45 * earlier during boot). The list is ordered based on virtual addresses. This
46 * order is kept after randomization.
47 */
48static __initdata struct kaslr_memory_region {
49 unsigned long *base;
50 unsigned long *end;
51 unsigned long size_tb;
52} kaslr_regions[] = {
53 {
54 .base = &page_offset_base,
55 .end = &direct_map_physmem_end,
56 },
57 {
58 .base = &vmalloc_base,
59 },
60 {
61 .base = &vmemmap_base,
62 },
63};
64
65/*
66 * The end of the physical address space that can be mapped directly by the
67 * kernel. This starts out at (1<<MAX_PHYSMEM_BITS) - 1), but KASLR may reduce
68 * that in order to increase the available entropy for mapping other regions.
69 */
70unsigned long direct_map_physmem_end __ro_after_init;
71
72/* Get size in bytes used by the memory region */
73static inline unsigned long get_padding(struct kaslr_memory_region *region)
74{
75 return (region->size_tb << TB_SHIFT);
76}
77
78/* Initialize base and padding for each memory region randomized with KASLR */
79void __init kernel_randomize_memory(void)
80{
81 size_t i;
82 unsigned long vaddr_start, vaddr;
83 unsigned long rand, memory_tb;
84 struct rnd_state rand_state;
85 unsigned long remain_entropy;
86 unsigned long vmemmap_size;
87
88 vaddr_start = pgtable_l5_enabled() ? __PAGE_OFFSET_BASE_L5 : __PAGE_OFFSET_BASE_L4;
89 vaddr = vaddr_start;
90
91 /*
92 * These BUILD_BUG_ON checks ensure the memory layout is consistent
93 * with the vaddr_start/vaddr_end variables. These checks are very
94 * limited....
95 */
96 BUILD_BUG_ON(vaddr_start >= vaddr_end);
97 BUILD_BUG_ON(vaddr_end != CPU_ENTRY_AREA_BASE);
98 BUILD_BUG_ON(vaddr_end > __START_KERNEL_map);
99
100 /* Preset the end of the possible address space for physical memory */
101 direct_map_physmem_end = ((1ULL << MAX_PHYSMEM_BITS) - 1);
102 if (!kaslr_memory_enabled())
103 return;
104
105 kaslr_regions[0].size_tb = 1 << (MAX_PHYSMEM_BITS - TB_SHIFT);
106 kaslr_regions[1].size_tb = VMALLOC_SIZE_TB;
107
108 /*
109 * Update Physical memory mapping to available and
110 * add padding if needed (especially for memory hotplug support).
111 */
112 BUG_ON(kaslr_regions[0].base != &page_offset_base);
113 memory_tb = DIV_ROUND_UP(max_pfn << PAGE_SHIFT, 1UL << TB_SHIFT) +
114 CONFIG_RANDOMIZE_MEMORY_PHYSICAL_PADDING;
115
116 /*
117 * Adapt physical memory region size based on available memory,
118 * except when CONFIG_PCI_P2PDMA is enabled. P2PDMA exposes the
119 * device BAR space assuming the direct map space is large enough
120 * for creating a ZONE_DEVICE mapping in the direct map corresponding
121 * to the physical BAR address.
122 */
123 if (!IS_ENABLED(CONFIG_PCI_P2PDMA) && (memory_tb < kaslr_regions[0].size_tb))
124 kaslr_regions[0].size_tb = memory_tb;
125
126 /*
127 * Calculate the vmemmap region size in TBs, aligned to a TB
128 * boundary.
129 */
130 vmemmap_size = (kaslr_regions[0].size_tb << (TB_SHIFT - PAGE_SHIFT)) *
131 sizeof(struct page);
132 kaslr_regions[2].size_tb = DIV_ROUND_UP(vmemmap_size, 1UL << TB_SHIFT);
133
134 /* Calculate entropy available between regions */
135 remain_entropy = vaddr_end - vaddr_start;
136 for (i = 0; i < ARRAY_SIZE(kaslr_regions); i++)
137 remain_entropy -= get_padding(region: &kaslr_regions[i]);
138
139 prandom_seed_state(state: &rand_state, seed: kaslr_get_random_long(purpose: "Memory"));
140
141 for (i = 0; i < ARRAY_SIZE(kaslr_regions); i++) {
142 unsigned long entropy;
143
144 /*
145 * Select a random virtual address using the extra entropy
146 * available.
147 */
148 entropy = remain_entropy / (ARRAY_SIZE(kaslr_regions) - i);
149 prandom_bytes_state(state: &rand_state, buf: &rand, nbytes: sizeof(rand));
150 entropy = (rand % (entropy + 1)) & PUD_MASK;
151 vaddr += entropy;
152 *kaslr_regions[i].base = vaddr;
153
154 /* Calculate the end of the region */
155 vaddr += get_padding(region: &kaslr_regions[i]);
156 /*
157 * KASLR trims the maximum possible size of the
158 * direct-map. Update the direct_map_physmem_end boundary.
159 * No rounding required as the region starts
160 * PUD aligned and size is in units of TB.
161 */
162 if (kaslr_regions[i].end)
163 *kaslr_regions[i].end = __pa_nodebug(vaddr - 1);
164
165 /* Add a minimum padding based on randomization alignment. */
166 vaddr = round_up(vaddr + 1, PUD_SIZE);
167 remain_entropy -= entropy;
168 }
169}
170
171void __meminit init_trampoline_kaslr(void)
172{
173 pud_t *pud_page_tramp, *pud, *pud_tramp;
174 p4d_t *p4d_page_tramp, *p4d, *p4d_tramp;
175 unsigned long paddr, vaddr;
176 pgd_t *pgd;
177
178 pud_page_tramp = alloc_low_page();
179
180 /*
181 * There are two mappings for the low 1MB area, the direct mapping
182 * and the 1:1 mapping for the real mode trampoline:
183 *
184 * Direct mapping: virt_addr = phys_addr + PAGE_OFFSET
185 * 1:1 mapping: virt_addr = phys_addr
186 */
187 paddr = 0;
188 vaddr = (unsigned long)__va(paddr);
189 pgd = pgd_offset_k(vaddr);
190
191 p4d = p4d_offset(pgd, address: vaddr);
192 pud = pud_offset(p4d, address: vaddr);
193
194 pud_tramp = pud_page_tramp + pud_index(address: paddr);
195 *pud_tramp = *pud;
196
197 if (pgtable_l5_enabled()) {
198 p4d_page_tramp = alloc_low_page();
199
200 p4d_tramp = p4d_page_tramp + p4d_index(address: paddr);
201
202 set_p4d(p4d_tramp,
203 __p4d(_KERNPG_TABLE | __pa(pud_page_tramp)));
204
205 trampoline_pgd_entry =
206 __pgd(_KERNPG_TABLE | __pa(p4d_page_tramp));
207 } else {
208 trampoline_pgd_entry =
209 __pgd(_KERNPG_TABLE | __pa(pud_page_tramp));
210 }
211}
212