xref: /openbmc/linux/arch/x86/mm/kasan_init_64.c (revision cfbb9be8)
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 #include <asm/cpu_entry_area.h>
19 
20 extern struct range pfn_mapped[E820_MAX_ENTRIES];
21 
22 static p4d_t tmp_p4d_table[PTRS_PER_P4D] __initdata __aligned(PAGE_SIZE);
23 
24 static __init void *early_alloc(size_t size, int nid, bool panic)
25 {
26 	if (panic)
27 		return memblock_virt_alloc_try_nid(size, size,
28 			__pa(MAX_DMA_ADDRESS), BOOTMEM_ALLOC_ACCESSIBLE, nid);
29 	else
30 		return memblock_virt_alloc_try_nid_nopanic(size, size,
31 			__pa(MAX_DMA_ADDRESS), BOOTMEM_ALLOC_ACCESSIBLE, nid);
32 }
33 
34 static void __init kasan_populate_pmd(pmd_t *pmd, unsigned long addr,
35 				      unsigned long end, int nid)
36 {
37 	pte_t *pte;
38 
39 	if (pmd_none(*pmd)) {
40 		void *p;
41 
42 		if (boot_cpu_has(X86_FEATURE_PSE) &&
43 		    ((end - addr) == PMD_SIZE) &&
44 		    IS_ALIGNED(addr, PMD_SIZE)) {
45 			p = early_alloc(PMD_SIZE, nid, false);
46 			if (p && pmd_set_huge(pmd, __pa(p), PAGE_KERNEL))
47 				return;
48 			else if (p)
49 				memblock_free(__pa(p), PMD_SIZE);
50 		}
51 
52 		p = early_alloc(PAGE_SIZE, nid, true);
53 		pmd_populate_kernel(&init_mm, pmd, p);
54 	}
55 
56 	pte = pte_offset_kernel(pmd, addr);
57 	do {
58 		pte_t entry;
59 		void *p;
60 
61 		if (!pte_none(*pte))
62 			continue;
63 
64 		p = early_alloc(PAGE_SIZE, nid, true);
65 		entry = pfn_pte(PFN_DOWN(__pa(p)), PAGE_KERNEL);
66 		set_pte_at(&init_mm, addr, pte, entry);
67 	} while (pte++, addr += PAGE_SIZE, addr != end);
68 }
69 
70 static void __init kasan_populate_pud(pud_t *pud, unsigned long addr,
71 				      unsigned long end, int nid)
72 {
73 	pmd_t *pmd;
74 	unsigned long next;
75 
76 	if (pud_none(*pud)) {
77 		void *p;
78 
79 		if (boot_cpu_has(X86_FEATURE_GBPAGES) &&
80 		    ((end - addr) == PUD_SIZE) &&
81 		    IS_ALIGNED(addr, PUD_SIZE)) {
82 			p = early_alloc(PUD_SIZE, nid, false);
83 			if (p && pud_set_huge(pud, __pa(p), PAGE_KERNEL))
84 				return;
85 			else if (p)
86 				memblock_free(__pa(p), PUD_SIZE);
87 		}
88 
89 		p = early_alloc(PAGE_SIZE, nid, true);
90 		pud_populate(&init_mm, pud, p);
91 	}
92 
93 	pmd = pmd_offset(pud, addr);
94 	do {
95 		next = pmd_addr_end(addr, end);
96 		if (!pmd_large(*pmd))
97 			kasan_populate_pmd(pmd, addr, next, nid);
98 	} while (pmd++, addr = next, addr != end);
99 }
100 
101 static void __init kasan_populate_p4d(p4d_t *p4d, unsigned long addr,
102 				      unsigned long end, int nid)
103 {
104 	pud_t *pud;
105 	unsigned long next;
106 
107 	if (p4d_none(*p4d)) {
108 		void *p = early_alloc(PAGE_SIZE, nid, true);
109 
110 		p4d_populate(&init_mm, p4d, p);
111 	}
112 
113 	pud = pud_offset(p4d, addr);
114 	do {
115 		next = pud_addr_end(addr, end);
116 		if (!pud_large(*pud))
117 			kasan_populate_pud(pud, addr, next, nid);
118 	} while (pud++, addr = next, addr != end);
119 }
120 
121 static void __init kasan_populate_pgd(pgd_t *pgd, unsigned long addr,
122 				      unsigned long end, int nid)
123 {
124 	void *p;
125 	p4d_t *p4d;
126 	unsigned long next;
127 
128 	if (pgd_none(*pgd)) {
129 		p = early_alloc(PAGE_SIZE, nid, true);
130 		pgd_populate(&init_mm, pgd, p);
131 	}
132 
133 	p4d = p4d_offset(pgd, addr);
134 	do {
135 		next = p4d_addr_end(addr, end);
136 		kasan_populate_p4d(p4d, addr, next, nid);
137 	} while (p4d++, addr = next, addr != end);
138 }
139 
140 static void __init kasan_populate_shadow(unsigned long addr, unsigned long end,
141 					 int nid)
142 {
143 	pgd_t *pgd;
144 	unsigned long next;
145 
146 	addr = addr & PAGE_MASK;
147 	end = round_up(end, PAGE_SIZE);
148 	pgd = pgd_offset_k(addr);
149 	do {
150 		next = pgd_addr_end(addr, end);
151 		kasan_populate_pgd(pgd, addr, next, nid);
152 	} while (pgd++, addr = next, addr != end);
153 }
154 
155 static void __init map_range(struct range *range)
156 {
157 	unsigned long start;
158 	unsigned long end;
159 
160 	start = (unsigned long)kasan_mem_to_shadow(pfn_to_kaddr(range->start));
161 	end = (unsigned long)kasan_mem_to_shadow(pfn_to_kaddr(range->end));
162 
163 	kasan_populate_shadow(start, end, early_pfn_to_nid(range->start));
164 }
165 
166 static void __init clear_pgds(unsigned long start,
167 			unsigned long end)
168 {
169 	pgd_t *pgd;
170 	/* See comment in kasan_init() */
171 	unsigned long pgd_end = end & PGDIR_MASK;
172 
173 	for (; start < pgd_end; start += PGDIR_SIZE) {
174 		pgd = pgd_offset_k(start);
175 		/*
176 		 * With folded p4d, pgd_clear() is nop, use p4d_clear()
177 		 * instead.
178 		 */
179 		if (CONFIG_PGTABLE_LEVELS < 5)
180 			p4d_clear(p4d_offset(pgd, start));
181 		else
182 			pgd_clear(pgd);
183 	}
184 
185 	pgd = pgd_offset_k(start);
186 	for (; start < end; start += P4D_SIZE)
187 		p4d_clear(p4d_offset(pgd, start));
188 }
189 
190 static inline p4d_t *early_p4d_offset(pgd_t *pgd, unsigned long addr)
191 {
192 	unsigned long p4d;
193 
194 	if (!IS_ENABLED(CONFIG_X86_5LEVEL))
195 		return (p4d_t *)pgd;
196 
197 	p4d = __pa_nodebug(pgd_val(*pgd)) & PTE_PFN_MASK;
198 	p4d += __START_KERNEL_map - phys_base;
199 	return (p4d_t *)p4d + p4d_index(addr);
200 }
201 
202 static void __init kasan_early_p4d_populate(pgd_t *pgd,
203 		unsigned long addr,
204 		unsigned long end)
205 {
206 	pgd_t pgd_entry;
207 	p4d_t *p4d, p4d_entry;
208 	unsigned long next;
209 
210 	if (pgd_none(*pgd)) {
211 		pgd_entry = __pgd(_KERNPG_TABLE | __pa_nodebug(kasan_zero_p4d));
212 		set_pgd(pgd, pgd_entry);
213 	}
214 
215 	p4d = early_p4d_offset(pgd, addr);
216 	do {
217 		next = p4d_addr_end(addr, end);
218 
219 		if (!p4d_none(*p4d))
220 			continue;
221 
222 		p4d_entry = __p4d(_KERNPG_TABLE | __pa_nodebug(kasan_zero_pud));
223 		set_p4d(p4d, p4d_entry);
224 	} while (p4d++, addr = next, addr != end && p4d_none(*p4d));
225 }
226 
227 static void __init kasan_map_early_shadow(pgd_t *pgd)
228 {
229 	/* See comment in kasan_init() */
230 	unsigned long addr = KASAN_SHADOW_START & PGDIR_MASK;
231 	unsigned long end = KASAN_SHADOW_END;
232 	unsigned long next;
233 
234 	pgd += pgd_index(addr);
235 	do {
236 		next = pgd_addr_end(addr, end);
237 		kasan_early_p4d_populate(pgd, addr, next);
238 	} while (pgd++, addr = next, addr != end);
239 }
240 
241 #ifdef CONFIG_KASAN_INLINE
242 static int kasan_die_handler(struct notifier_block *self,
243 			     unsigned long val,
244 			     void *data)
245 {
246 	if (val == DIE_GPF) {
247 		pr_emerg("CONFIG_KASAN_INLINE enabled\n");
248 		pr_emerg("GPF could be caused by NULL-ptr deref or user memory access\n");
249 	}
250 	return NOTIFY_OK;
251 }
252 
253 static struct notifier_block kasan_die_notifier = {
254 	.notifier_call = kasan_die_handler,
255 };
256 #endif
257 
258 void __init kasan_early_init(void)
259 {
260 	int i;
261 	pteval_t pte_val = __pa_nodebug(kasan_zero_page) | __PAGE_KERNEL | _PAGE_ENC;
262 	pmdval_t pmd_val = __pa_nodebug(kasan_zero_pte) | _KERNPG_TABLE;
263 	pudval_t pud_val = __pa_nodebug(kasan_zero_pmd) | _KERNPG_TABLE;
264 	p4dval_t p4d_val = __pa_nodebug(kasan_zero_pud) | _KERNPG_TABLE;
265 
266 	for (i = 0; i < PTRS_PER_PTE; i++)
267 		kasan_zero_pte[i] = __pte(pte_val);
268 
269 	for (i = 0; i < PTRS_PER_PMD; i++)
270 		kasan_zero_pmd[i] = __pmd(pmd_val);
271 
272 	for (i = 0; i < PTRS_PER_PUD; i++)
273 		kasan_zero_pud[i] = __pud(pud_val);
274 
275 	for (i = 0; IS_ENABLED(CONFIG_X86_5LEVEL) && i < PTRS_PER_P4D; i++)
276 		kasan_zero_p4d[i] = __p4d(p4d_val);
277 
278 	kasan_map_early_shadow(early_top_pgt);
279 	kasan_map_early_shadow(init_top_pgt);
280 }
281 
282 void __init kasan_init(void)
283 {
284 	int i;
285 	void *shadow_cpu_entry_begin, *shadow_cpu_entry_end;
286 
287 #ifdef CONFIG_KASAN_INLINE
288 	register_die_notifier(&kasan_die_notifier);
289 #endif
290 
291 	memcpy(early_top_pgt, init_top_pgt, sizeof(early_top_pgt));
292 
293 	/*
294 	 * We use the same shadow offset for 4- and 5-level paging to
295 	 * facilitate boot-time switching between paging modes.
296 	 * As result in 5-level paging mode KASAN_SHADOW_START and
297 	 * KASAN_SHADOW_END are not aligned to PGD boundary.
298 	 *
299 	 * KASAN_SHADOW_START doesn't share PGD with anything else.
300 	 * We claim whole PGD entry to make things easier.
301 	 *
302 	 * KASAN_SHADOW_END lands in the last PGD entry and it collides with
303 	 * bunch of things like kernel code, modules, EFI mapping, etc.
304 	 * We need to take extra steps to not overwrite them.
305 	 */
306 	if (IS_ENABLED(CONFIG_X86_5LEVEL)) {
307 		void *ptr;
308 
309 		ptr = (void *)pgd_page_vaddr(*pgd_offset_k(KASAN_SHADOW_END));
310 		memcpy(tmp_p4d_table, (void *)ptr, sizeof(tmp_p4d_table));
311 		set_pgd(&early_top_pgt[pgd_index(KASAN_SHADOW_END)],
312 				__pgd(__pa(tmp_p4d_table) | _KERNPG_TABLE));
313 	}
314 
315 	load_cr3(early_top_pgt);
316 	__flush_tlb_all();
317 
318 	clear_pgds(KASAN_SHADOW_START & PGDIR_MASK, KASAN_SHADOW_END);
319 
320 	kasan_populate_zero_shadow((void *)(KASAN_SHADOW_START & PGDIR_MASK),
321 			kasan_mem_to_shadow((void *)PAGE_OFFSET));
322 
323 	for (i = 0; i < E820_MAX_ENTRIES; i++) {
324 		if (pfn_mapped[i].end == 0)
325 			break;
326 
327 		map_range(&pfn_mapped[i]);
328 	}
329 
330 	shadow_cpu_entry_begin = (void *)CPU_ENTRY_AREA_BASE;
331 	shadow_cpu_entry_begin = kasan_mem_to_shadow(shadow_cpu_entry_begin);
332 	shadow_cpu_entry_begin = (void *)round_down((unsigned long)shadow_cpu_entry_begin,
333 						PAGE_SIZE);
334 
335 	shadow_cpu_entry_end = (void *)(CPU_ENTRY_AREA_BASE +
336 					CPU_ENTRY_AREA_MAP_SIZE);
337 	shadow_cpu_entry_end = kasan_mem_to_shadow(shadow_cpu_entry_end);
338 	shadow_cpu_entry_end = (void *)round_up((unsigned long)shadow_cpu_entry_end,
339 					PAGE_SIZE);
340 
341 	kasan_populate_zero_shadow(
342 		kasan_mem_to_shadow((void *)PAGE_OFFSET + MAXMEM),
343 		shadow_cpu_entry_begin);
344 
345 	kasan_populate_shadow((unsigned long)shadow_cpu_entry_begin,
346 			      (unsigned long)shadow_cpu_entry_end, 0);
347 
348 	kasan_populate_zero_shadow(shadow_cpu_entry_end,
349 				kasan_mem_to_shadow((void *)__START_KERNEL_map));
350 
351 	kasan_populate_shadow((unsigned long)kasan_mem_to_shadow(_stext),
352 			      (unsigned long)kasan_mem_to_shadow(_end),
353 			      early_pfn_to_nid(__pa(_stext)));
354 
355 	kasan_populate_zero_shadow(kasan_mem_to_shadow((void *)MODULES_END),
356 				(void *)KASAN_SHADOW_END);
357 
358 	load_cr3(init_top_pgt);
359 	__flush_tlb_all();
360 
361 	/*
362 	 * kasan_zero_page has been used as early shadow memory, thus it may
363 	 * contain some garbage. Now we can clear and write protect it, since
364 	 * after the TLB flush no one should write to it.
365 	 */
366 	memset(kasan_zero_page, 0, PAGE_SIZE);
367 	for (i = 0; i < PTRS_PER_PTE; i++) {
368 		pte_t pte = __pte(__pa(kasan_zero_page) | __PAGE_KERNEL_RO | _PAGE_ENC);
369 		set_pte(&kasan_zero_pte[i], pte);
370 	}
371 	/* Flush TLBs again to be sure that write protection applied. */
372 	__flush_tlb_all();
373 
374 	init_task.kasan_depth = 0;
375 	pr_info("KernelAddressSanitizer initialized\n");
376 }
377