1 // SPDX-License-Identifier: GPL-2.0
3 * KMSAN initialization routines.
5 * Copyright (C) 2017-2021 Google LLC
6 * Author: Alexander Potapenko <glider@google.com>
12 #include <asm/sections.h>
14 #include <linux/memblock.h>
16 #include "../internal.h"
18 #define NUM_FUTURE_RANGES 128
19 struct start_end_pair {
23 static struct start_end_pair start_end_pairs[NUM_FUTURE_RANGES] __initdata;
24 static int future_index __initdata;
27 * Record a range of memory for which the metadata pages will be created once
28 * the page allocator becomes available.
30 static void __init kmsan_record_future_shadow_range(void *start, void *end)
32 u64 nstart = (u64)start, nend = (u64)end, cstart, cend;
35 KMSAN_WARN_ON(future_index == NUM_FUTURE_RANGES);
36 KMSAN_WARN_ON((nstart >= nend) || !nstart || !nend);
37 nstart = ALIGN_DOWN(nstart, PAGE_SIZE);
38 nend = ALIGN(nend, PAGE_SIZE);
41 * Scan the existing ranges to see if any of them overlaps with
42 * [start, end). In that case, merge the two ranges instead of
44 * The number of ranges is less than 20, so there is no need to organize
45 * them into a more intelligent data structure.
47 for (int i = 0; i < future_index; i++) {
48 cstart = start_end_pairs[i].start;
49 cend = start_end_pairs[i].end;
50 if ((cstart < nstart && cend < nstart) ||
51 (cstart > nend && cend > nend))
52 /* ranges are disjoint - do not merge */
54 start_end_pairs[i].start = min(nstart, cstart);
55 start_end_pairs[i].end = max(nend, cend);
61 start_end_pairs[future_index].start = nstart;
62 start_end_pairs[future_index].end = nend;
67 * Initialize the shadow for existing mappings during kernel initialization.
68 * These include kernel text/data sections, NODE_DATA and future ranges
69 * registered while creating other data (e.g. percpu).
71 * Allocations via memblock can be only done before slab is initialized.
73 void __init kmsan_init_shadow(void)
75 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
76 phys_addr_t p_start, p_end;
80 for_each_reserved_mem_range(loop, &p_start, &p_end)
81 kmsan_record_future_shadow_range(phys_to_virt(p_start),
83 /* Allocate shadow for .data */
84 kmsan_record_future_shadow_range(_sdata, _edata);
86 for_each_online_node(nid)
87 kmsan_record_future_shadow_range(
88 NODE_DATA(nid), (char *)NODE_DATA(nid) + nd_size);
90 for (int i = 0; i < future_index; i++)
91 kmsan_init_alloc_meta_for_range(
92 (void *)start_end_pairs[i].start,
93 (void *)start_end_pairs[i].end);
96 struct metadata_page_pair {
97 struct page *shadow, *origin;
99 static struct metadata_page_pair held_back[MAX_ORDER] __initdata;
102 * Eager metadata allocation. When the memblock allocator is freeing pages to
103 * pagealloc, we use 2/3 of them as metadata for the remaining 1/3.
104 * We store the pointers to the returned blocks of pages in held_back[] grouped
105 * by their order: when kmsan_memblock_free_pages() is called for the first
106 * time with a certain order, it is reserved as a shadow block, for the second
107 * time - as an origin block. On the third time the incoming block receives its
108 * shadow and origin ranges from the previously saved shadow and origin blocks,
109 * after which held_back[order] can be used again.
111 * At the very end there may be leftover blocks in held_back[]. They are
112 * collected later by kmsan_memblock_discard().
114 bool kmsan_memblock_free_pages(struct page *page, unsigned int order)
116 struct page *shadow, *origin;
118 if (!held_back[order].shadow) {
119 held_back[order].shadow = page;
122 if (!held_back[order].origin) {
123 held_back[order].origin = page;
126 shadow = held_back[order].shadow;
127 origin = held_back[order].origin;
128 kmsan_setup_meta(page, shadow, origin, order);
130 held_back[order].shadow = NULL;
131 held_back[order].origin = NULL;
137 struct page *items[MAX_BLOCKS];
142 static struct smallstack collect = {
147 static void smallstack_push(struct smallstack *stack, struct page *pages)
149 KMSAN_WARN_ON(stack->index == MAX_BLOCKS);
150 stack->items[stack->index] = pages;
155 static struct page *smallstack_pop(struct smallstack *stack)
159 KMSAN_WARN_ON(stack->index == 0);
161 ret = stack->items[stack->index];
162 stack->items[stack->index] = NULL;
166 static void do_collection(void)
168 struct page *page, *shadow, *origin;
170 while (collect.index >= 3) {
171 page = smallstack_pop(&collect);
172 shadow = smallstack_pop(&collect);
173 origin = smallstack_pop(&collect);
174 kmsan_setup_meta(page, shadow, origin, collect.order);
175 __free_pages_core(page, collect.order);
179 static void collect_split(void)
181 struct smallstack tmp = {
182 .order = collect.order - 1,
189 while (collect.index) {
190 page = smallstack_pop(&collect);
191 smallstack_push(&tmp, &page[0]);
192 smallstack_push(&tmp, &page[1 << tmp.order]);
194 __memcpy(&collect, &tmp, sizeof(tmp));
198 * Memblock is about to go away. Split the page blocks left over in held_back[]
199 * and return 1/3 of that memory to the system.
201 static void kmsan_memblock_discard(void)
205 * - push held_back[N].shadow and .origin to @collect;
206 * - while there are >= 3 elements in @collect, do garbage collection:
207 * - pop 3 ranges from @collect;
208 * - use two of them as shadow and origin for the third one;
210 * - split each remaining element from @collect into 2 ranges of
214 collect.order = MAX_ORDER - 1;
215 for (int i = MAX_ORDER - 1; i >= 0; i--) {
216 if (held_back[i].shadow)
217 smallstack_push(&collect, held_back[i].shadow);
218 if (held_back[i].origin)
219 smallstack_push(&collect, held_back[i].origin);
220 held_back[i].shadow = NULL;
221 held_back[i].origin = NULL;
227 void __init kmsan_init_runtime(void)
229 /* Assuming current is init_task */
230 kmsan_internal_task_create(current);
231 kmsan_memblock_discard();
232 pr_info("Starting KernelMemorySanitizer\n");
233 pr_info("ATTENTION: KMSAN is a debugging tool! Do not use it on production machines!\n");
234 kmsan_enabled = true;