1 // SPDX-License-Identifier: GPL-2.0 1 // SPDX-License-Identifier: GPL-2.0
2 /* 2 /*
3 * KMSAN initialization routines. 3 * KMSAN initialization routines.
4 * 4 *
5 * Copyright (C) 2017-2021 Google LLC 5 * Copyright (C) 2017-2021 Google LLC
6 * Author: Alexander Potapenko <glider@google. 6 * Author: Alexander Potapenko <glider@google.com>
7 * 7 *
8 */ 8 */
9 9
10 #include "kmsan.h" 10 #include "kmsan.h"
11 11
12 #include <asm/sections.h> 12 #include <asm/sections.h>
13 #include <linux/mm.h> 13 #include <linux/mm.h>
14 #include <linux/memblock.h> 14 #include <linux/memblock.h>
15 15
16 #include "../internal.h" 16 #include "../internal.h"
17 17
18 #define NUM_FUTURE_RANGES 128 18 #define NUM_FUTURE_RANGES 128
19 struct start_end_pair { 19 struct start_end_pair {
20 u64 start, end; 20 u64 start, end;
21 }; 21 };
22 22
23 static struct start_end_pair start_end_pairs[N 23 static struct start_end_pair start_end_pairs[NUM_FUTURE_RANGES] __initdata;
24 static int future_index __initdata; 24 static int future_index __initdata;
25 25
26 /* 26 /*
27 * Record a range of memory for which the meta 27 * Record a range of memory for which the metadata pages will be created once
28 * the page allocator becomes available. 28 * the page allocator becomes available.
29 */ 29 */
30 static void __init kmsan_record_future_shadow_ 30 static void __init kmsan_record_future_shadow_range(void *start, void *end)
31 { 31 {
32 u64 nstart = (u64)start, nend = (u64)e 32 u64 nstart = (u64)start, nend = (u64)end, cstart, cend;
33 bool merged = false; 33 bool merged = false;
34 34
35 KMSAN_WARN_ON(future_index == NUM_FUTU 35 KMSAN_WARN_ON(future_index == NUM_FUTURE_RANGES);
36 KMSAN_WARN_ON((nstart >= nend) || !! 36 KMSAN_WARN_ON((nstart >= nend) || !nstart || !nend);
37 /* Virtual address 0 is <<
38 (!IS_ENABLED(CONFIG_S390 <<
39 !nend); <<
40 nstart = ALIGN_DOWN(nstart, PAGE_SIZE) 37 nstart = ALIGN_DOWN(nstart, PAGE_SIZE);
41 nend = ALIGN(nend, PAGE_SIZE); 38 nend = ALIGN(nend, PAGE_SIZE);
42 39
43 /* 40 /*
44 * Scan the existing ranges to see if 41 * Scan the existing ranges to see if any of them overlaps with
45 * [start, end). In that case, merge t 42 * [start, end). In that case, merge the two ranges instead of
46 * creating a new one. 43 * creating a new one.
47 * The number of ranges is less than 2 44 * The number of ranges is less than 20, so there is no need to organize
48 * them into a more intelligent data s 45 * them into a more intelligent data structure.
49 */ 46 */
50 for (int i = 0; i < future_index; i++) 47 for (int i = 0; i < future_index; i++) {
51 cstart = start_end_pairs[i].st 48 cstart = start_end_pairs[i].start;
52 cend = start_end_pairs[i].end; 49 cend = start_end_pairs[i].end;
53 if ((cstart < nstart && cend < 50 if ((cstart < nstart && cend < nstart) ||
54 (cstart > nend && cend > n 51 (cstart > nend && cend > nend))
55 /* ranges are disjoint 52 /* ranges are disjoint - do not merge */
56 continue; 53 continue;
57 start_end_pairs[i].start = min 54 start_end_pairs[i].start = min(nstart, cstart);
58 start_end_pairs[i].end = max(n 55 start_end_pairs[i].end = max(nend, cend);
59 merged = true; 56 merged = true;
60 break; 57 break;
61 } 58 }
62 if (merged) 59 if (merged)
63 return; 60 return;
64 start_end_pairs[future_index].start = 61 start_end_pairs[future_index].start = nstart;
65 start_end_pairs[future_index].end = ne 62 start_end_pairs[future_index].end = nend;
66 future_index++; 63 future_index++;
67 } 64 }
68 65
69 /* 66 /*
70 * Initialize the shadow for existing mappings 67 * Initialize the shadow for existing mappings during kernel initialization.
71 * These include kernel text/data sections, NO 68 * These include kernel text/data sections, NODE_DATA and future ranges
72 * registered while creating other data (e.g. 69 * registered while creating other data (e.g. percpu).
73 * 70 *
74 * Allocations via memblock can be only done b 71 * Allocations via memblock can be only done before slab is initialized.
75 */ 72 */
76 void __init kmsan_init_shadow(void) 73 void __init kmsan_init_shadow(void)
77 { 74 {
78 const size_t nd_size = sizeof(pg_data_ !! 75 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
79 phys_addr_t p_start, p_end; 76 phys_addr_t p_start, p_end;
80 u64 loop; 77 u64 loop;
81 int nid; 78 int nid;
82 79
83 for_each_reserved_mem_range(loop, &p_s 80 for_each_reserved_mem_range(loop, &p_start, &p_end)
84 kmsan_record_future_shadow_ran 81 kmsan_record_future_shadow_range(phys_to_virt(p_start),
85 82 phys_to_virt(p_end));
86 /* Allocate shadow for .data */ 83 /* Allocate shadow for .data */
87 kmsan_record_future_shadow_range(_sdat 84 kmsan_record_future_shadow_range(_sdata, _edata);
88 85
89 for_each_online_node(nid) 86 for_each_online_node(nid)
90 kmsan_record_future_shadow_ran 87 kmsan_record_future_shadow_range(
91 NODE_DATA(nid), (char 88 NODE_DATA(nid), (char *)NODE_DATA(nid) + nd_size);
92 89
93 for (int i = 0; i < future_index; i++) 90 for (int i = 0; i < future_index; i++)
94 kmsan_init_alloc_meta_for_rang 91 kmsan_init_alloc_meta_for_range(
95 (void *)start_end_pair 92 (void *)start_end_pairs[i].start,
96 (void *)start_end_pair 93 (void *)start_end_pairs[i].end);
97 } 94 }
98 95
99 struct metadata_page_pair { 96 struct metadata_page_pair {
100 struct page *shadow, *origin; 97 struct page *shadow, *origin;
101 }; 98 };
102 static struct metadata_page_pair held_back[NR_ !! 99 static struct metadata_page_pair held_back[MAX_ORDER + 1] __initdata;
103 100
104 /* 101 /*
105 * Eager metadata allocation. When the membloc 102 * Eager metadata allocation. When the memblock allocator is freeing pages to
106 * pagealloc, we use 2/3 of them as metadata f 103 * pagealloc, we use 2/3 of them as metadata for the remaining 1/3.
107 * We store the pointers to the returned block 104 * We store the pointers to the returned blocks of pages in held_back[] grouped
108 * by their order: when kmsan_memblock_free_pa 105 * by their order: when kmsan_memblock_free_pages() is called for the first
109 * time with a certain order, it is reserved a 106 * time with a certain order, it is reserved as a shadow block, for the second
110 * time - as an origin block. On the third tim 107 * time - as an origin block. On the third time the incoming block receives its
111 * shadow and origin ranges from the previousl 108 * shadow and origin ranges from the previously saved shadow and origin blocks,
112 * after which held_back[order] can be used ag 109 * after which held_back[order] can be used again.
113 * 110 *
114 * At the very end there may be leftover block 111 * At the very end there may be leftover blocks in held_back[]. They are
115 * collected later by kmsan_memblock_discard() 112 * collected later by kmsan_memblock_discard().
116 */ 113 */
117 bool kmsan_memblock_free_pages(struct page *pa 114 bool kmsan_memblock_free_pages(struct page *page, unsigned int order)
118 { 115 {
119 struct page *shadow, *origin; 116 struct page *shadow, *origin;
120 117
121 if (!held_back[order].shadow) { 118 if (!held_back[order].shadow) {
122 held_back[order].shadow = page 119 held_back[order].shadow = page;
123 return false; 120 return false;
124 } 121 }
125 if (!held_back[order].origin) { 122 if (!held_back[order].origin) {
126 held_back[order].origin = page 123 held_back[order].origin = page;
127 return false; 124 return false;
128 } 125 }
129 shadow = held_back[order].shadow; 126 shadow = held_back[order].shadow;
130 origin = held_back[order].origin; 127 origin = held_back[order].origin;
131 kmsan_setup_meta(page, shadow, origin, 128 kmsan_setup_meta(page, shadow, origin, order);
132 129
133 held_back[order].shadow = NULL; 130 held_back[order].shadow = NULL;
134 held_back[order].origin = NULL; 131 held_back[order].origin = NULL;
135 return true; 132 return true;
136 } 133 }
137 134
138 #define MAX_BLOCKS 8 135 #define MAX_BLOCKS 8
139 struct smallstack { 136 struct smallstack {
140 struct page *items[MAX_BLOCKS]; 137 struct page *items[MAX_BLOCKS];
141 int index; 138 int index;
142 int order; 139 int order;
143 }; 140 };
144 141
145 static struct smallstack collect = { 142 static struct smallstack collect = {
146 .index = 0, 143 .index = 0,
147 .order = MAX_PAGE_ORDER, !! 144 .order = MAX_ORDER,
148 }; 145 };
149 146
150 static void smallstack_push(struct smallstack 147 static void smallstack_push(struct smallstack *stack, struct page *pages)
151 { 148 {
152 KMSAN_WARN_ON(stack->index == MAX_BLOC 149 KMSAN_WARN_ON(stack->index == MAX_BLOCKS);
153 stack->items[stack->index] = pages; 150 stack->items[stack->index] = pages;
154 stack->index++; 151 stack->index++;
155 } 152 }
156 #undef MAX_BLOCKS 153 #undef MAX_BLOCKS
157 154
158 static struct page *smallstack_pop(struct smal 155 static struct page *smallstack_pop(struct smallstack *stack)
159 { 156 {
160 struct page *ret; 157 struct page *ret;
161 158
162 KMSAN_WARN_ON(stack->index == 0); 159 KMSAN_WARN_ON(stack->index == 0);
163 stack->index--; 160 stack->index--;
164 ret = stack->items[stack->index]; 161 ret = stack->items[stack->index];
165 stack->items[stack->index] = NULL; 162 stack->items[stack->index] = NULL;
166 return ret; 163 return ret;
167 } 164 }
168 165
169 static void do_collection(void) 166 static void do_collection(void)
170 { 167 {
171 struct page *page, *shadow, *origin; 168 struct page *page, *shadow, *origin;
172 169
173 while (collect.index >= 3) { 170 while (collect.index >= 3) {
174 page = smallstack_pop(&collect 171 page = smallstack_pop(&collect);
175 shadow = smallstack_pop(&colle 172 shadow = smallstack_pop(&collect);
176 origin = smallstack_pop(&colle 173 origin = smallstack_pop(&collect);
177 kmsan_setup_meta(page, shadow, 174 kmsan_setup_meta(page, shadow, origin, collect.order);
178 __free_pages_core(page, collec !! 175 __free_pages_core(page, collect.order);
179 } 176 }
180 } 177 }
181 178
182 static void collect_split(void) 179 static void collect_split(void)
183 { 180 {
184 struct smallstack tmp = { 181 struct smallstack tmp = {
185 .order = collect.order - 1, 182 .order = collect.order - 1,
186 .index = 0, 183 .index = 0,
187 }; 184 };
188 struct page *page; 185 struct page *page;
189 186
190 if (!collect.order) 187 if (!collect.order)
191 return; 188 return;
192 while (collect.index) { 189 while (collect.index) {
193 page = smallstack_pop(&collect 190 page = smallstack_pop(&collect);
194 smallstack_push(&tmp, &page[0] 191 smallstack_push(&tmp, &page[0]);
195 smallstack_push(&tmp, &page[1 192 smallstack_push(&tmp, &page[1 << tmp.order]);
196 } 193 }
197 __memcpy(&collect, &tmp, sizeof(tmp)); 194 __memcpy(&collect, &tmp, sizeof(tmp));
198 } 195 }
199 196
200 /* 197 /*
201 * Memblock is about to go away. Split the pag 198 * Memblock is about to go away. Split the page blocks left over in held_back[]
202 * and return 1/3 of that memory to the system 199 * and return 1/3 of that memory to the system.
203 */ 200 */
204 static void kmsan_memblock_discard(void) 201 static void kmsan_memblock_discard(void)
205 { 202 {
206 /* 203 /*
207 * For each order=N: 204 * For each order=N:
208 * - push held_back[N].shadow and .or 205 * - push held_back[N].shadow and .origin to @collect;
209 * - while there are >= 3 elements in 206 * - while there are >= 3 elements in @collect, do garbage collection:
210 * - pop 3 ranges from @collect; 207 * - pop 3 ranges from @collect;
211 * - use two of them as shadow and 208 * - use two of them as shadow and origin for the third one;
212 * - repeat; 209 * - repeat;
213 * - split each remaining element fro 210 * - split each remaining element from @collect into 2 ranges of
214 * order=N-1, 211 * order=N-1,
215 * - repeat. 212 * - repeat.
216 */ 213 */
217 collect.order = MAX_PAGE_ORDER; !! 214 collect.order = MAX_ORDER;
218 for (int i = MAX_PAGE_ORDER; i >= 0; i !! 215 for (int i = MAX_ORDER; i >= 0; i--) {
219 if (held_back[i].shadow) 216 if (held_back[i].shadow)
220 smallstack_push(&colle 217 smallstack_push(&collect, held_back[i].shadow);
221 if (held_back[i].origin) 218 if (held_back[i].origin)
222 smallstack_push(&colle 219 smallstack_push(&collect, held_back[i].origin);
223 held_back[i].shadow = NULL; 220 held_back[i].shadow = NULL;
224 held_back[i].origin = NULL; 221 held_back[i].origin = NULL;
225 do_collection(); 222 do_collection();
226 collect_split(); 223 collect_split();
227 } 224 }
228 } 225 }
229 226
230 void __init kmsan_init_runtime(void) 227 void __init kmsan_init_runtime(void)
231 { 228 {
232 /* Assuming current is init_task */ 229 /* Assuming current is init_task */
233 kmsan_internal_task_create(current); 230 kmsan_internal_task_create(current);
234 kmsan_memblock_discard(); 231 kmsan_memblock_discard();
235 pr_info("Starting KernelMemorySanitize 232 pr_info("Starting KernelMemorySanitizer\n");
236 pr_info("ATTENTION: KMSAN is a debuggi 233 pr_info("ATTENTION: KMSAN is a debugging tool! Do not use it on production machines!\n");
237 kmsan_enabled = true; 234 kmsan_enabled = true;
238 } 235 }
239 236
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.