xref: /openbmc/linux/arch/x86/mm/kasan_init_64.c (revision 65417d9f)
1 // SPDX-License-Identifier: GPL-2.0
2 #define DISABLE_BRANCH_PROFILING
3 #define pr_fmt(fmt) "kasan: " fmt
4 #include <linux/bootmem.h>
5 #include <linux/kasan.h>
6 #include <linux/kdebug.h>
7 #include <linux/memblock.h>
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/sched/task.h>
11 #include <linux/vmalloc.h>
12 
13 #include <asm/e820/types.h>
14 #include <asm/pgalloc.h>
15 #include <asm/tlbflush.h>
16 #include <asm/sections.h>
17 #include <asm/pgtable.h>
18 
19 extern struct range pfn_mapped[E820_MAX_ENTRIES];
20 
21 static p4d_t tmp_p4d_table[PTRS_PER_P4D] __initdata __aligned(PAGE_SIZE);
22 
23 static __init void *early_alloc(size_t size, int nid)
24 {
25 	return memblock_virt_alloc_try_nid_nopanic(size, size,
26 		__pa(MAX_DMA_ADDRESS), BOOTMEM_ALLOC_ACCESSIBLE, nid);
27 }
28 
29 static void __init kasan_populate_pmd(pmd_t *pmd, unsigned long addr,
30 				      unsigned long end, int nid)
31 {
32 	pte_t *pte;
33 
34 	if (pmd_none(*pmd)) {
35 		void *p;
36 
37 		if (boot_cpu_has(X86_FEATURE_PSE) &&
38 		    ((end - addr) == PMD_SIZE) &&
39 		    IS_ALIGNED(addr, PMD_SIZE)) {
40 			p = early_alloc(PMD_SIZE, nid);
41 			if (p && pmd_set_huge(pmd, __pa(p), PAGE_KERNEL))
42 				return;
43 			else if (p)
44 				memblock_free(__pa(p), PMD_SIZE);
45 		}
46 
47 		p = early_alloc(PAGE_SIZE, nid);
48 		pmd_populate_kernel(&init_mm, pmd, p);
49 	}
50 
51 	pte = pte_offset_kernel(pmd, addr);
52 	do {
53 		pte_t entry;
54 		void *p;
55 
56 		if (!pte_none(*pte))
57 			continue;
58 
59 		p = early_alloc(PAGE_SIZE, nid);
60 		entry = pfn_pte(PFN_DOWN(__pa(p)), PAGE_KERNEL);
61 		set_pte_at(&init_mm, addr, pte, entry);
62 	} while (pte++, addr += PAGE_SIZE, addr != end);
63 }
64 
65 static void __init kasan_populate_pud(pud_t *pud, unsigned long addr,
66 				      unsigned long end, int nid)
67 {
68 	pmd_t *pmd;
69 	unsigned long next;
70 
71 	if (pud_none(*pud)) {
72 		void *p;
73 
74 		if (boot_cpu_has(X86_FEATURE_GBPAGES) &&
75 		    ((end - addr) == PUD_SIZE) &&
76 		    IS_ALIGNED(addr, PUD_SIZE)) {
77 			p = early_alloc(PUD_SIZE, nid);
78 			if (p && pud_set_huge(pud, __pa(p), PAGE_KERNEL))
79 				return;
80 			else if (p)
81 				memblock_free(__pa(p), PUD_SIZE);
82 		}
83 
84 		p = early_alloc(PAGE_SIZE, nid);
85 		pud_populate(&init_mm, pud, p);
86 	}
87 
88 	pmd = pmd_offset(pud, addr);
89 	do {
90 		next = pmd_addr_end(addr, end);
91 		if (!pmd_large(*pmd))
92 			kasan_populate_pmd(pmd, addr, next, nid);
93 	} while (pmd++, addr = next, addr != end);
94 }
95 
96 static void __init kasan_populate_p4d(p4d_t *p4d, unsigned long addr,
97 				      unsigned long end, int nid)
98 {
99 	pud_t *pud;
100 	unsigned long next;
101 
102 	if (p4d_none(*p4d)) {
103 		void *p = early_alloc(PAGE_SIZE, nid);
104 
105 		p4d_populate(&init_mm, p4d, p);
106 	}
107 
108 	pud = pud_offset(p4d, addr);
109 	do {
110 		next = pud_addr_end(addr, end);
111 		if (!pud_large(*pud))
112 			kasan_populate_pud(pud, addr, next, nid);
113 	} while (pud++, addr = next, addr != end);
114 }
115 
116 static void __init kasan_populate_pgd(pgd_t *pgd, unsigned long addr,
117 				      unsigned long end, int nid)
118 {
119 	void *p;
120 	p4d_t *p4d;
121 	unsigned long next;
122 
123 	if (pgd_none(*pgd)) {
124 		p = early_alloc(PAGE_SIZE, nid);
125 		pgd_populate(&init_mm, pgd, p);
126 	}
127 
128 	p4d = p4d_offset(pgd, addr);
129 	do {
130 		next = p4d_addr_end(addr, end);
131 		kasan_populate_p4d(p4d, addr, next, nid);
132 	} while (p4d++, addr = next, addr != end);
133 }
134 
135 static void __init kasan_populate_shadow(unsigned long addr, unsigned long end,
136 					 int nid)
137 {
138 	pgd_t *pgd;
139 	unsigned long next;
140 
141 	addr = addr & PAGE_MASK;
142 	end = round_up(end, PAGE_SIZE);
143 	pgd = pgd_offset_k(addr);
144 	do {
145 		next = pgd_addr_end(addr, end);
146 		kasan_populate_pgd(pgd, addr, next, nid);
147 	} while (pgd++, addr = next, addr != end);
148 }
149 
150 static void __init map_range(struct range *range)
151 {
152 	unsigned long start;
153 	unsigned long end;
154 
155 	start = (unsigned long)kasan_mem_to_shadow(pfn_to_kaddr(range->start));
156 	end = (unsigned long)kasan_mem_to_shadow(pfn_to_kaddr(range->end));
157 
158 	kasan_populate_shadow(start, end, early_pfn_to_nid(range->start));
159 }
160 
161 static void __init clear_pgds(unsigned long start,
162 			unsigned long end)
163 {
164 	pgd_t *pgd;
165 	/* See comment in kasan_init() */
166 	unsigned long pgd_end = end & PGDIR_MASK;
167 
168 	for (; start < pgd_end; start += PGDIR_SIZE) {
169 		pgd = pgd_offset_k(start);
170 		/*
171 		 * With folded p4d, pgd_clear() is nop, use p4d_clear()
172 		 * instead.
173 		 */
174 		if (CONFIG_PGTABLE_LEVELS < 5)
175 			p4d_clear(p4d_offset(pgd, start));
176 		else
177 			pgd_clear(pgd);
178 	}
179 
180 	pgd = pgd_offset_k(start);
181 	for (; start < end; start += P4D_SIZE)
182 		p4d_clear(p4d_offset(pgd, start));
183 }
184 
185 static inline p4d_t *early_p4d_offset(pgd_t *pgd, unsigned long addr)
186 {
187 	unsigned long p4d;
188 
189 	if (!IS_ENABLED(CONFIG_X86_5LEVEL))
190 		return (p4d_t *)pgd;
191 
192 	p4d = __pa_nodebug(pgd_val(*pgd)) & PTE_PFN_MASK;
193 	p4d += __START_KERNEL_map - phys_base;
194 	return (p4d_t *)p4d + p4d_index(addr);
195 }
196 
197 static void __init kasan_early_p4d_populate(pgd_t *pgd,
198 		unsigned long addr,
199 		unsigned long end)
200 {
201 	pgd_t pgd_entry;
202 	p4d_t *p4d, p4d_entry;
203 	unsigned long next;
204 
205 	if (pgd_none(*pgd)) {
206 		pgd_entry = __pgd(_KERNPG_TABLE | __pa_nodebug(kasan_zero_p4d));
207 		set_pgd(pgd, pgd_entry);
208 	}
209 
210 	p4d = early_p4d_offset(pgd, addr);
211 	do {
212 		next = p4d_addr_end(addr, end);
213 
214 		if (!p4d_none(*p4d))
215 			continue;
216 
217 		p4d_entry = __p4d(_KERNPG_TABLE | __pa_nodebug(kasan_zero_pud));
218 		set_p4d(p4d, p4d_entry);
219 	} while (p4d++, addr = next, addr != end && p4d_none(*p4d));
220 }
221 
222 static void __init kasan_map_early_shadow(pgd_t *pgd)
223 {
224 	/* See comment in kasan_init() */
225 	unsigned long addr = KASAN_SHADOW_START & PGDIR_MASK;
226 	unsigned long end = KASAN_SHADOW_END;
227 	unsigned long next;
228 
229 	pgd += pgd_index(addr);
230 	do {
231 		next = pgd_addr_end(addr, end);
232 		kasan_early_p4d_populate(pgd, addr, next);
233 	} while (pgd++, addr = next, addr != end);
234 }
235 
236 #ifdef CONFIG_KASAN_INLINE
237 static int kasan_die_handler(struct notifier_block *self,
238 			     unsigned long val,
239 			     void *data)
240 {
241 	if (val == DIE_GPF) {
242 		pr_emerg("CONFIG_KASAN_INLINE enabled\n");
243 		pr_emerg("GPF could be caused by NULL-ptr deref or user memory access\n");
244 	}
245 	return NOTIFY_OK;
246 }
247 
248 static struct notifier_block kasan_die_notifier = {
249 	.notifier_call = kasan_die_handler,
250 };
251 #endif
252 
253 void __init kasan_early_init(void)
254 {
255 	int i;
256 	pteval_t pte_val = __pa_nodebug(kasan_zero_page) | __PAGE_KERNEL | _PAGE_ENC;
257 	pmdval_t pmd_val = __pa_nodebug(kasan_zero_pte) | _KERNPG_TABLE;
258 	pudval_t pud_val = __pa_nodebug(kasan_zero_pmd) | _KERNPG_TABLE;
259 	p4dval_t p4d_val = __pa_nodebug(kasan_zero_pud) | _KERNPG_TABLE;
260 
261 	for (i = 0; i < PTRS_PER_PTE; i++)
262 		kasan_zero_pte[i] = __pte(pte_val);
263 
264 	for (i = 0; i < PTRS_PER_PMD; i++)
265 		kasan_zero_pmd[i] = __pmd(pmd_val);
266 
267 	for (i = 0; i < PTRS_PER_PUD; i++)
268 		kasan_zero_pud[i] = __pud(pud_val);
269 
270 	for (i = 0; IS_ENABLED(CONFIG_X86_5LEVEL) && i < PTRS_PER_P4D; i++)
271 		kasan_zero_p4d[i] = __p4d(p4d_val);
272 
273 	kasan_map_early_shadow(early_top_pgt);
274 	kasan_map_early_shadow(init_top_pgt);
275 }
276 
277 void __init kasan_init(void)
278 {
279 	int i;
280 
281 #ifdef CONFIG_KASAN_INLINE
282 	register_die_notifier(&kasan_die_notifier);
283 #endif
284 
285 	memcpy(early_top_pgt, init_top_pgt, sizeof(early_top_pgt));
286 
287 	/*
288 	 * We use the same shadow offset for 4- and 5-level paging to
289 	 * facilitate boot-time switching between paging modes.
290 	 * As result in 5-level paging mode KASAN_SHADOW_START and
291 	 * KASAN_SHADOW_END are not aligned to PGD boundary.
292 	 *
293 	 * KASAN_SHADOW_START doesn't share PGD with anything else.
294 	 * We claim whole PGD entry to make things easier.
295 	 *
296 	 * KASAN_SHADOW_END lands in the last PGD entry and it collides with
297 	 * bunch of things like kernel code, modules, EFI mapping, etc.
298 	 * We need to take extra steps to not overwrite them.
299 	 */
300 	if (IS_ENABLED(CONFIG_X86_5LEVEL)) {
301 		void *ptr;
302 
303 		ptr = (void *)pgd_page_vaddr(*pgd_offset_k(KASAN_SHADOW_END));
304 		memcpy(tmp_p4d_table, (void *)ptr, sizeof(tmp_p4d_table));
305 		set_pgd(&early_top_pgt[pgd_index(KASAN_SHADOW_END)],
306 				__pgd(__pa(tmp_p4d_table) | _KERNPG_TABLE));
307 	}
308 
309 	load_cr3(early_top_pgt);
310 	__flush_tlb_all();
311 
312 	clear_pgds(KASAN_SHADOW_START & PGDIR_MASK, KASAN_SHADOW_END);
313 
314 	kasan_populate_zero_shadow((void *)(KASAN_SHADOW_START & PGDIR_MASK),
315 			kasan_mem_to_shadow((void *)PAGE_OFFSET));
316 
317 	for (i = 0; i < E820_MAX_ENTRIES; i++) {
318 		if (pfn_mapped[i].end == 0)
319 			break;
320 
321 		map_range(&pfn_mapped[i]);
322 	}
323 
324 	kasan_populate_zero_shadow(
325 		kasan_mem_to_shadow((void *)PAGE_OFFSET + MAXMEM),
326 		kasan_mem_to_shadow((void *)__START_KERNEL_map));
327 
328 	kasan_populate_shadow((unsigned long)kasan_mem_to_shadow(_stext),
329 			      (unsigned long)kasan_mem_to_shadow(_end),
330 			      early_pfn_to_nid(__pa(_stext)));
331 
332 	kasan_populate_zero_shadow(kasan_mem_to_shadow((void *)MODULES_END),
333 			(void *)KASAN_SHADOW_END);
334 
335 	load_cr3(init_top_pgt);
336 	__flush_tlb_all();
337 
338 	/*
339 	 * kasan_zero_page has been used as early shadow memory, thus it may
340 	 * contain some garbage. Now we can clear and write protect it, since
341 	 * after the TLB flush no one should write to it.
342 	 */
343 	memset(kasan_zero_page, 0, PAGE_SIZE);
344 	for (i = 0; i < PTRS_PER_PTE; i++) {
345 		pte_t pte = __pte(__pa(kasan_zero_page) | __PAGE_KERNEL_RO | _PAGE_ENC);
346 		set_pte(&kasan_zero_pte[i], pte);
347 	}
348 	/* Flush TLBs again to be sure that write protection applied. */
349 	__flush_tlb_all();
350 
351 	init_task.kasan_depth = 0;
352 	pr_info("KernelAddressSanitizer initialized\n");
353 }
354