xref: /openbmc/linux/arch/x86/mm/kasan_init_64.c (revision 050e9baa)
1 // SPDX-License-Identifier: GPL-2.0
2 #define DISABLE_BRANCH_PROFILING
3 #define pr_fmt(fmt) "kasan: " fmt
4 
5 /* cpu_feature_enabled() cannot be used this early */
6 #define USE_EARLY_PGTABLE_L5
7 
8 #include <linux/bootmem.h>
9 #include <linux/kasan.h>
10 #include <linux/kdebug.h>
11 #include <linux/memblock.h>
12 #include <linux/mm.h>
13 #include <linux/sched.h>
14 #include <linux/sched/task.h>
15 #include <linux/vmalloc.h>
16 
17 #include <asm/e820/types.h>
18 #include <asm/pgalloc.h>
19 #include <asm/tlbflush.h>
20 #include <asm/sections.h>
21 #include <asm/pgtable.h>
22 #include <asm/cpu_entry_area.h>
23 
24 extern struct range pfn_mapped[E820_MAX_ENTRIES];
25 
26 static p4d_t tmp_p4d_table[MAX_PTRS_PER_P4D] __initdata __aligned(PAGE_SIZE);
27 
28 static __init void *early_alloc(size_t size, int nid, bool panic)
29 {
30 	if (panic)
31 		return memblock_virt_alloc_try_nid(size, size,
32 			__pa(MAX_DMA_ADDRESS), BOOTMEM_ALLOC_ACCESSIBLE, nid);
33 	else
34 		return memblock_virt_alloc_try_nid_nopanic(size, size,
35 			__pa(MAX_DMA_ADDRESS), BOOTMEM_ALLOC_ACCESSIBLE, nid);
36 }
37 
38 static void __init kasan_populate_pmd(pmd_t *pmd, unsigned long addr,
39 				      unsigned long end, int nid)
40 {
41 	pte_t *pte;
42 
43 	if (pmd_none(*pmd)) {
44 		void *p;
45 
46 		if (boot_cpu_has(X86_FEATURE_PSE) &&
47 		    ((end - addr) == PMD_SIZE) &&
48 		    IS_ALIGNED(addr, PMD_SIZE)) {
49 			p = early_alloc(PMD_SIZE, nid, false);
50 			if (p && pmd_set_huge(pmd, __pa(p), PAGE_KERNEL))
51 				return;
52 			else if (p)
53 				memblock_free(__pa(p), PMD_SIZE);
54 		}
55 
56 		p = early_alloc(PAGE_SIZE, nid, true);
57 		pmd_populate_kernel(&init_mm, pmd, p);
58 	}
59 
60 	pte = pte_offset_kernel(pmd, addr);
61 	do {
62 		pte_t entry;
63 		void *p;
64 
65 		if (!pte_none(*pte))
66 			continue;
67 
68 		p = early_alloc(PAGE_SIZE, nid, true);
69 		entry = pfn_pte(PFN_DOWN(__pa(p)), PAGE_KERNEL);
70 		set_pte_at(&init_mm, addr, pte, entry);
71 	} while (pte++, addr += PAGE_SIZE, addr != end);
72 }
73 
74 static void __init kasan_populate_pud(pud_t *pud, unsigned long addr,
75 				      unsigned long end, int nid)
76 {
77 	pmd_t *pmd;
78 	unsigned long next;
79 
80 	if (pud_none(*pud)) {
81 		void *p;
82 
83 		if (boot_cpu_has(X86_FEATURE_GBPAGES) &&
84 		    ((end - addr) == PUD_SIZE) &&
85 		    IS_ALIGNED(addr, PUD_SIZE)) {
86 			p = early_alloc(PUD_SIZE, nid, false);
87 			if (p && pud_set_huge(pud, __pa(p), PAGE_KERNEL))
88 				return;
89 			else if (p)
90 				memblock_free(__pa(p), PUD_SIZE);
91 		}
92 
93 		p = early_alloc(PAGE_SIZE, nid, true);
94 		pud_populate(&init_mm, pud, p);
95 	}
96 
97 	pmd = pmd_offset(pud, addr);
98 	do {
99 		next = pmd_addr_end(addr, end);
100 		if (!pmd_large(*pmd))
101 			kasan_populate_pmd(pmd, addr, next, nid);
102 	} while (pmd++, addr = next, addr != end);
103 }
104 
105 static void __init kasan_populate_p4d(p4d_t *p4d, unsigned long addr,
106 				      unsigned long end, int nid)
107 {
108 	pud_t *pud;
109 	unsigned long next;
110 
111 	if (p4d_none(*p4d)) {
112 		void *p = early_alloc(PAGE_SIZE, nid, true);
113 
114 		p4d_populate(&init_mm, p4d, p);
115 	}
116 
117 	pud = pud_offset(p4d, addr);
118 	do {
119 		next = pud_addr_end(addr, end);
120 		if (!pud_large(*pud))
121 			kasan_populate_pud(pud, addr, next, nid);
122 	} while (pud++, addr = next, addr != end);
123 }
124 
125 static void __init kasan_populate_pgd(pgd_t *pgd, unsigned long addr,
126 				      unsigned long end, int nid)
127 {
128 	void *p;
129 	p4d_t *p4d;
130 	unsigned long next;
131 
132 	if (pgd_none(*pgd)) {
133 		p = early_alloc(PAGE_SIZE, nid, true);
134 		pgd_populate(&init_mm, pgd, p);
135 	}
136 
137 	p4d = p4d_offset(pgd, addr);
138 	do {
139 		next = p4d_addr_end(addr, end);
140 		kasan_populate_p4d(p4d, addr, next, nid);
141 	} while (p4d++, addr = next, addr != end);
142 }
143 
144 static void __init kasan_populate_shadow(unsigned long addr, unsigned long end,
145 					 int nid)
146 {
147 	pgd_t *pgd;
148 	unsigned long next;
149 
150 	addr = addr & PAGE_MASK;
151 	end = round_up(end, PAGE_SIZE);
152 	pgd = pgd_offset_k(addr);
153 	do {
154 		next = pgd_addr_end(addr, end);
155 		kasan_populate_pgd(pgd, addr, next, nid);
156 	} while (pgd++, addr = next, addr != end);
157 }
158 
159 static void __init map_range(struct range *range)
160 {
161 	unsigned long start;
162 	unsigned long end;
163 
164 	start = (unsigned long)kasan_mem_to_shadow(pfn_to_kaddr(range->start));
165 	end = (unsigned long)kasan_mem_to_shadow(pfn_to_kaddr(range->end));
166 
167 	kasan_populate_shadow(start, end, early_pfn_to_nid(range->start));
168 }
169 
170 static void __init clear_pgds(unsigned long start,
171 			unsigned long end)
172 {
173 	pgd_t *pgd;
174 	/* See comment in kasan_init() */
175 	unsigned long pgd_end = end & PGDIR_MASK;
176 
177 	for (; start < pgd_end; start += PGDIR_SIZE) {
178 		pgd = pgd_offset_k(start);
179 		/*
180 		 * With folded p4d, pgd_clear() is nop, use p4d_clear()
181 		 * instead.
182 		 */
183 		if (pgtable_l5_enabled())
184 			pgd_clear(pgd);
185 		else
186 			p4d_clear(p4d_offset(pgd, start));
187 	}
188 
189 	pgd = pgd_offset_k(start);
190 	for (; start < end; start += P4D_SIZE)
191 		p4d_clear(p4d_offset(pgd, start));
192 }
193 
194 static inline p4d_t *early_p4d_offset(pgd_t *pgd, unsigned long addr)
195 {
196 	unsigned long p4d;
197 
198 	if (!pgtable_l5_enabled())
199 		return (p4d_t *)pgd;
200 
201 	p4d = __pa_nodebug(pgd_val(*pgd)) & PTE_PFN_MASK;
202 	p4d += __START_KERNEL_map - phys_base;
203 	return (p4d_t *)p4d + p4d_index(addr);
204 }
205 
206 static void __init kasan_early_p4d_populate(pgd_t *pgd,
207 		unsigned long addr,
208 		unsigned long end)
209 {
210 	pgd_t pgd_entry;
211 	p4d_t *p4d, p4d_entry;
212 	unsigned long next;
213 
214 	if (pgd_none(*pgd)) {
215 		pgd_entry = __pgd(_KERNPG_TABLE | __pa_nodebug(kasan_zero_p4d));
216 		set_pgd(pgd, pgd_entry);
217 	}
218 
219 	p4d = early_p4d_offset(pgd, addr);
220 	do {
221 		next = p4d_addr_end(addr, end);
222 
223 		if (!p4d_none(*p4d))
224 			continue;
225 
226 		p4d_entry = __p4d(_KERNPG_TABLE | __pa_nodebug(kasan_zero_pud));
227 		set_p4d(p4d, p4d_entry);
228 	} while (p4d++, addr = next, addr != end && p4d_none(*p4d));
229 }
230 
231 static void __init kasan_map_early_shadow(pgd_t *pgd)
232 {
233 	/* See comment in kasan_init() */
234 	unsigned long addr = KASAN_SHADOW_START & PGDIR_MASK;
235 	unsigned long end = KASAN_SHADOW_END;
236 	unsigned long next;
237 
238 	pgd += pgd_index(addr);
239 	do {
240 		next = pgd_addr_end(addr, end);
241 		kasan_early_p4d_populate(pgd, addr, next);
242 	} while (pgd++, addr = next, addr != end);
243 }
244 
245 #ifdef CONFIG_KASAN_INLINE
246 static int kasan_die_handler(struct notifier_block *self,
247 			     unsigned long val,
248 			     void *data)
249 {
250 	if (val == DIE_GPF) {
251 		pr_emerg("CONFIG_KASAN_INLINE enabled\n");
252 		pr_emerg("GPF could be caused by NULL-ptr deref or user memory access\n");
253 	}
254 	return NOTIFY_OK;
255 }
256 
257 static struct notifier_block kasan_die_notifier = {
258 	.notifier_call = kasan_die_handler,
259 };
260 #endif
261 
262 void __init kasan_early_init(void)
263 {
264 	int i;
265 	pteval_t pte_val = __pa_nodebug(kasan_zero_page) | __PAGE_KERNEL | _PAGE_ENC;
266 	pmdval_t pmd_val = __pa_nodebug(kasan_zero_pte) | _KERNPG_TABLE;
267 	pudval_t pud_val = __pa_nodebug(kasan_zero_pmd) | _KERNPG_TABLE;
268 	p4dval_t p4d_val = __pa_nodebug(kasan_zero_pud) | _KERNPG_TABLE;
269 
270 	/* Mask out unsupported __PAGE_KERNEL bits: */
271 	pte_val &= __default_kernel_pte_mask;
272 	pmd_val &= __default_kernel_pte_mask;
273 	pud_val &= __default_kernel_pte_mask;
274 	p4d_val &= __default_kernel_pte_mask;
275 
276 	for (i = 0; i < PTRS_PER_PTE; i++)
277 		kasan_zero_pte[i] = __pte(pte_val);
278 
279 	for (i = 0; i < PTRS_PER_PMD; i++)
280 		kasan_zero_pmd[i] = __pmd(pmd_val);
281 
282 	for (i = 0; i < PTRS_PER_PUD; i++)
283 		kasan_zero_pud[i] = __pud(pud_val);
284 
285 	for (i = 0; pgtable_l5_enabled() && i < PTRS_PER_P4D; i++)
286 		kasan_zero_p4d[i] = __p4d(p4d_val);
287 
288 	kasan_map_early_shadow(early_top_pgt);
289 	kasan_map_early_shadow(init_top_pgt);
290 }
291 
292 void __init kasan_init(void)
293 {
294 	int i;
295 	void *shadow_cpu_entry_begin, *shadow_cpu_entry_end;
296 
297 #ifdef CONFIG_KASAN_INLINE
298 	register_die_notifier(&kasan_die_notifier);
299 #endif
300 
301 	memcpy(early_top_pgt, init_top_pgt, sizeof(early_top_pgt));
302 
303 	/*
304 	 * We use the same shadow offset for 4- and 5-level paging to
305 	 * facilitate boot-time switching between paging modes.
306 	 * As result in 5-level paging mode KASAN_SHADOW_START and
307 	 * KASAN_SHADOW_END are not aligned to PGD boundary.
308 	 *
309 	 * KASAN_SHADOW_START doesn't share PGD with anything else.
310 	 * We claim whole PGD entry to make things easier.
311 	 *
312 	 * KASAN_SHADOW_END lands in the last PGD entry and it collides with
313 	 * bunch of things like kernel code, modules, EFI mapping, etc.
314 	 * We need to take extra steps to not overwrite them.
315 	 */
316 	if (pgtable_l5_enabled()) {
317 		void *ptr;
318 
319 		ptr = (void *)pgd_page_vaddr(*pgd_offset_k(KASAN_SHADOW_END));
320 		memcpy(tmp_p4d_table, (void *)ptr, sizeof(tmp_p4d_table));
321 		set_pgd(&early_top_pgt[pgd_index(KASAN_SHADOW_END)],
322 				__pgd(__pa(tmp_p4d_table) | _KERNPG_TABLE));
323 	}
324 
325 	load_cr3(early_top_pgt);
326 	__flush_tlb_all();
327 
328 	clear_pgds(KASAN_SHADOW_START & PGDIR_MASK, KASAN_SHADOW_END);
329 
330 	kasan_populate_zero_shadow((void *)(KASAN_SHADOW_START & PGDIR_MASK),
331 			kasan_mem_to_shadow((void *)PAGE_OFFSET));
332 
333 	for (i = 0; i < E820_MAX_ENTRIES; i++) {
334 		if (pfn_mapped[i].end == 0)
335 			break;
336 
337 		map_range(&pfn_mapped[i]);
338 	}
339 
340 	shadow_cpu_entry_begin = (void *)CPU_ENTRY_AREA_BASE;
341 	shadow_cpu_entry_begin = kasan_mem_to_shadow(shadow_cpu_entry_begin);
342 	shadow_cpu_entry_begin = (void *)round_down((unsigned long)shadow_cpu_entry_begin,
343 						PAGE_SIZE);
344 
345 	shadow_cpu_entry_end = (void *)(CPU_ENTRY_AREA_BASE +
346 					CPU_ENTRY_AREA_MAP_SIZE);
347 	shadow_cpu_entry_end = kasan_mem_to_shadow(shadow_cpu_entry_end);
348 	shadow_cpu_entry_end = (void *)round_up((unsigned long)shadow_cpu_entry_end,
349 					PAGE_SIZE);
350 
351 	kasan_populate_zero_shadow(
352 		kasan_mem_to_shadow((void *)PAGE_OFFSET + MAXMEM),
353 		shadow_cpu_entry_begin);
354 
355 	kasan_populate_shadow((unsigned long)shadow_cpu_entry_begin,
356 			      (unsigned long)shadow_cpu_entry_end, 0);
357 
358 	kasan_populate_zero_shadow(shadow_cpu_entry_end,
359 				kasan_mem_to_shadow((void *)__START_KERNEL_map));
360 
361 	kasan_populate_shadow((unsigned long)kasan_mem_to_shadow(_stext),
362 			      (unsigned long)kasan_mem_to_shadow(_end),
363 			      early_pfn_to_nid(__pa(_stext)));
364 
365 	kasan_populate_zero_shadow(kasan_mem_to_shadow((void *)MODULES_END),
366 				(void *)KASAN_SHADOW_END);
367 
368 	load_cr3(init_top_pgt);
369 	__flush_tlb_all();
370 
371 	/*
372 	 * kasan_zero_page has been used as early shadow memory, thus it may
373 	 * contain some garbage. Now we can clear and write protect it, since
374 	 * after the TLB flush no one should write to it.
375 	 */
376 	memset(kasan_zero_page, 0, PAGE_SIZE);
377 	for (i = 0; i < PTRS_PER_PTE; i++) {
378 		pte_t pte;
379 		pgprot_t prot;
380 
381 		prot = __pgprot(__PAGE_KERNEL_RO | _PAGE_ENC);
382 		pgprot_val(prot) &= __default_kernel_pte_mask;
383 
384 		pte = __pte(__pa(kasan_zero_page) | pgprot_val(prot));
385 		set_pte(&kasan_zero_pte[i], pte);
386 	}
387 	/* Flush TLBs again to be sure that write protection applied. */
388 	__flush_tlb_all();
389 
390 	init_task.kasan_depth = 0;
391 	pr_info("KernelAddressSanitizer initialized\n");
392 }
393