]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 | 2 | /* |
835c34a1 | 3 | * prepare to run common code |
1da177e4 LT |
4 | * |
5 | * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE | |
1da177e4 LT |
6 | */ |
7 | ||
be3606ff | 8 | #define DISABLE_BRANCH_PROFILING |
1da177e4 LT |
9 | #include <linux/init.h> |
10 | #include <linux/linkage.h> | |
11 | #include <linux/types.h> | |
12 | #include <linux/kernel.h> | |
13 | #include <linux/string.h> | |
14 | #include <linux/percpu.h> | |
eaf76e8b | 15 | #include <linux/start_kernel.h> |
8b664aa6 | 16 | #include <linux/io.h> |
72d7c3b3 | 17 | #include <linux/memblock.h> |
5868f365 | 18 | #include <linux/mem_encrypt.h> |
1da177e4 LT |
19 | |
20 | #include <asm/processor.h> | |
21 | #include <asm/proto.h> | |
22 | #include <asm/smp.h> | |
1da177e4 LT |
23 | #include <asm/setup.h> |
24 | #include <asm/desc.h> | |
f6c2e333 | 25 | #include <asm/pgtable.h> |
cfd243d4 | 26 | #include <asm/tlbflush.h> |
2bc0414e | 27 | #include <asm/sections.h> |
718fc13b | 28 | #include <asm/kdebug.h> |
66441bd3 | 29 | #include <asm/e820/api.h> |
47a3d5da | 30 | #include <asm/bios_ebda.h> |
5dcd14ec | 31 | #include <asm/bootparam_utils.h> |
feddc9de | 32 | #include <asm/microcode.h> |
ef7f0d6a | 33 | #include <asm/kasan.h> |
1da177e4 | 34 | |
8170e6be PA |
35 | /* |
36 | * Manage page tables very early on. | |
37 | */ | |
8170e6be | 38 | extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD]; |
c88d7150 | 39 | static unsigned int __initdata next_early_pgt; |
5e427ec2 | 40 | pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX); |
8170e6be | 41 | |
26179670 KS |
42 | #define __head __section(.head.text) |
43 | ||
44 | static void __head *fixup_pointer(void *ptr, unsigned long physaddr) | |
c88d7150 KS |
45 | { |
46 | return ptr - (void *)_text + (void *)physaddr; | |
47 | } | |
48 | ||
aca20d54 TL |
49 | unsigned long __head __startup_64(unsigned long physaddr, |
50 | struct boot_params *bp) | |
c88d7150 KS |
51 | { |
52 | unsigned long load_delta, *p; | |
5868f365 | 53 | unsigned long pgtable_flags; |
c88d7150 | 54 | pgdval_t *pgd; |
032370b9 | 55 | p4dval_t *p4d; |
c88d7150 KS |
56 | pudval_t *pud; |
57 | pmdval_t *pmd, pmd_entry; | |
58 | int i; | |
187e91fe | 59 | unsigned int *next_pgt_ptr; |
c88d7150 KS |
60 | |
61 | /* Is the address too large? */ | |
62 | if (physaddr >> MAX_PHYSMEM_BITS) | |
63 | for (;;); | |
64 | ||
65 | /* | |
66 | * Compute the delta between the address I am compiled to run at | |
67 | * and the address I am actually running at. | |
68 | */ | |
69 | load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map); | |
70 | ||
71 | /* Is the address not 2M aligned? */ | |
72 | if (load_delta & ~PMD_PAGE_MASK) | |
73 | for (;;); | |
74 | ||
5868f365 | 75 | /* Activate Secure Memory Encryption (SME) if supported and enabled */ |
aca20d54 | 76 | sme_enable(bp); |
5868f365 TL |
77 | |
78 | /* Include the SME encryption mask in the fixup value */ | |
79 | load_delta += sme_get_me_mask(); | |
80 | ||
c88d7150 KS |
81 | /* Fixup the physical addresses in the page table */ |
82 | ||
65ade2f8 | 83 | pgd = fixup_pointer(&early_top_pgt, physaddr); |
c88d7150 KS |
84 | pgd[pgd_index(__START_KERNEL_map)] += load_delta; |
85 | ||
032370b9 KS |
86 | if (IS_ENABLED(CONFIG_X86_5LEVEL)) { |
87 | p4d = fixup_pointer(&level4_kernel_pgt, physaddr); | |
88 | p4d[511] += load_delta; | |
89 | } | |
90 | ||
c88d7150 KS |
91 | pud = fixup_pointer(&level3_kernel_pgt, physaddr); |
92 | pud[510] += load_delta; | |
93 | pud[511] += load_delta; | |
94 | ||
95 | pmd = fixup_pointer(level2_fixmap_pgt, physaddr); | |
96 | pmd[506] += load_delta; | |
97 | ||
98 | /* | |
99 | * Set up the identity mapping for the switchover. These | |
100 | * entries should *NOT* have the global bit set! This also | |
101 | * creates a bunch of nonsense entries but that is fine -- | |
102 | * it avoids problems around wraparound. | |
103 | */ | |
104 | ||
187e91fe AP |
105 | next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr); |
106 | pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr); | |
107 | pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr); | |
c88d7150 | 108 | |
21729f81 | 109 | pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask(); |
c88d7150 | 110 | |
032370b9 KS |
111 | if (IS_ENABLED(CONFIG_X86_5LEVEL)) { |
112 | p4d = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr); | |
113 | ||
114 | i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD; | |
5868f365 TL |
115 | pgd[i + 0] = (pgdval_t)p4d + pgtable_flags; |
116 | pgd[i + 1] = (pgdval_t)p4d + pgtable_flags; | |
032370b9 KS |
117 | |
118 | i = (physaddr >> P4D_SHIFT) % PTRS_PER_P4D; | |
5868f365 TL |
119 | p4d[i + 0] = (pgdval_t)pud + pgtable_flags; |
120 | p4d[i + 1] = (pgdval_t)pud + pgtable_flags; | |
032370b9 KS |
121 | } else { |
122 | i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD; | |
5868f365 TL |
123 | pgd[i + 0] = (pgdval_t)pud + pgtable_flags; |
124 | pgd[i + 1] = (pgdval_t)pud + pgtable_flags; | |
032370b9 | 125 | } |
c88d7150 KS |
126 | |
127 | i = (physaddr >> PUD_SHIFT) % PTRS_PER_PUD; | |
5868f365 TL |
128 | pud[i + 0] = (pudval_t)pmd + pgtable_flags; |
129 | pud[i + 1] = (pudval_t)pmd + pgtable_flags; | |
c88d7150 KS |
130 | |
131 | pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL; | |
5868f365 | 132 | pmd_entry += sme_get_me_mask(); |
c88d7150 KS |
133 | pmd_entry += physaddr; |
134 | ||
135 | for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) { | |
136 | int idx = i + (physaddr >> PMD_SHIFT) % PTRS_PER_PMD; | |
137 | pmd[idx] = pmd_entry + i * PMD_SIZE; | |
138 | } | |
139 | ||
140 | /* | |
141 | * Fixup the kernel text+data virtual addresses. Note that | |
142 | * we might write invalid pmds, when the kernel is relocated | |
143 | * cleanup_highmap() fixes this up along with the mappings | |
144 | * beyond _end. | |
145 | */ | |
146 | ||
147 | pmd = fixup_pointer(level2_kernel_pgt, physaddr); | |
148 | for (i = 0; i < PTRS_PER_PMD; i++) { | |
149 | if (pmd[i] & _PAGE_PRESENT) | |
150 | pmd[i] += load_delta; | |
151 | } | |
152 | ||
5868f365 TL |
153 | /* |
154 | * Fixup phys_base - remove the memory encryption mask to obtain | |
155 | * the true physical address. | |
156 | */ | |
c88d7150 | 157 | p = fixup_pointer(&phys_base, physaddr); |
5868f365 TL |
158 | *p += load_delta - sme_get_me_mask(); |
159 | ||
160 | /* Encrypt the kernel (if SME is active) */ | |
161 | sme_encrypt_kernel(); | |
162 | ||
163 | /* | |
164 | * Return the SME encryption mask (if SME is active) to be used as a | |
165 | * modifier for the initial pgdir entry programmed into CR3. | |
166 | */ | |
167 | return sme_get_me_mask(); | |
168 | } | |
169 | ||
170 | unsigned long __startup_secondary_64(void) | |
171 | { | |
172 | /* | |
173 | * Return the SME encryption mask (if SME is active) to be used as a | |
174 | * modifier for the initial pgdir entry programmed into CR3. | |
175 | */ | |
176 | return sme_get_me_mask(); | |
c88d7150 KS |
177 | } |
178 | ||
8170e6be PA |
179 | /* Wipe all early page tables except for the kernel symbol map */ |
180 | static void __init reset_early_page_tables(void) | |
cfd243d4 | 181 | { |
65ade2f8 | 182 | memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1)); |
8170e6be | 183 | next_early_pgt = 0; |
21729f81 | 184 | write_cr3(__sme_pa_nodebug(early_top_pgt)); |
8170e6be PA |
185 | } |
186 | ||
187 | /* Create a new PMD entry */ | |
b9d05200 | 188 | int __init __early_make_pgtable(unsigned long address, pmdval_t pmd) |
8170e6be PA |
189 | { |
190 | unsigned long physaddr = address - __PAGE_OFFSET; | |
8170e6be | 191 | pgdval_t pgd, *pgd_p; |
032370b9 | 192 | p4dval_t p4d, *p4d_p; |
6b9c75ac | 193 | pudval_t pud, *pud_p; |
b9d05200 | 194 | pmdval_t *pmd_p; |
8170e6be PA |
195 | |
196 | /* Invalid address or early pgt is done ? */ | |
65ade2f8 | 197 | if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt)) |
8170e6be PA |
198 | return -1; |
199 | ||
6b9c75ac | 200 | again: |
65ade2f8 | 201 | pgd_p = &early_top_pgt[pgd_index(address)].pgd; |
8170e6be PA |
202 | pgd = *pgd_p; |
203 | ||
204 | /* | |
205 | * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is | |
206 | * critical -- __PAGE_OFFSET would point us back into the dynamic | |
207 | * range and we might end up looping forever... | |
208 | */ | |
032370b9 KS |
209 | if (!IS_ENABLED(CONFIG_X86_5LEVEL)) |
210 | p4d_p = pgd_p; | |
211 | else if (pgd) | |
212 | p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); | |
213 | else { | |
214 | if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { | |
215 | reset_early_page_tables(); | |
216 | goto again; | |
217 | } | |
218 | ||
219 | p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++]; | |
220 | memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D); | |
221 | *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; | |
222 | } | |
223 | p4d_p += p4d_index(address); | |
224 | p4d = *p4d_p; | |
225 | ||
226 | if (p4d) | |
227 | pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); | |
6b9c75ac YL |
228 | else { |
229 | if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { | |
8170e6be | 230 | reset_early_page_tables(); |
6b9c75ac YL |
231 | goto again; |
232 | } | |
8170e6be PA |
233 | |
234 | pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++]; | |
a91bbe01 | 235 | memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD); |
032370b9 | 236 | *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; |
8170e6be | 237 | } |
6b9c75ac YL |
238 | pud_p += pud_index(address); |
239 | pud = *pud_p; | |
8170e6be | 240 | |
6b9c75ac YL |
241 | if (pud) |
242 | pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base); | |
243 | else { | |
244 | if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) { | |
245 | reset_early_page_tables(); | |
246 | goto again; | |
247 | } | |
248 | ||
249 | pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++]; | |
a91bbe01 | 250 | memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD); |
6b9c75ac YL |
251 | *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE; |
252 | } | |
6b9c75ac | 253 | pmd_p[pmd_index(address)] = pmd; |
8170e6be PA |
254 | |
255 | return 0; | |
cfd243d4 VG |
256 | } |
257 | ||
b9d05200 TL |
258 | int __init early_make_pgtable(unsigned long address) |
259 | { | |
260 | unsigned long physaddr = address - __PAGE_OFFSET; | |
261 | pmdval_t pmd; | |
262 | ||
263 | pmd = (physaddr & PMD_MASK) + early_pmd_flags; | |
264 | ||
265 | return __early_make_pgtable(address, pmd); | |
266 | } | |
267 | ||
1da177e4 LT |
268 | /* Don't add a printk in there. printk relies on the PDA which is not initialized |
269 | yet. */ | |
270 | static void __init clear_bss(void) | |
271 | { | |
1da177e4 | 272 | memset(__bss_start, 0, |
2bc0414e | 273 | (unsigned long) __bss_stop - (unsigned long) __bss_start); |
1da177e4 LT |
274 | } |
275 | ||
f1da834c YL |
276 | static unsigned long get_cmd_line_ptr(void) |
277 | { | |
278 | unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr; | |
279 | ||
ee92d815 YL |
280 | cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32; |
281 | ||
f1da834c YL |
282 | return cmd_line_ptr; |
283 | } | |
284 | ||
1da177e4 LT |
285 | static void __init copy_bootdata(char *real_mode_data) |
286 | { | |
1da177e4 | 287 | char * command_line; |
f1da834c | 288 | unsigned long cmd_line_ptr; |
1da177e4 | 289 | |
b9d05200 TL |
290 | /* |
291 | * If SME is active, this will create decrypted mappings of the | |
292 | * boot data in advance of the copy operations. | |
293 | */ | |
294 | sme_map_bootdata(real_mode_data); | |
295 | ||
30c82645 | 296 | memcpy(&boot_params, real_mode_data, sizeof boot_params); |
5dcd14ec | 297 | sanitize_boot_params(&boot_params); |
f1da834c YL |
298 | cmd_line_ptr = get_cmd_line_ptr(); |
299 | if (cmd_line_ptr) { | |
300 | command_line = __va(cmd_line_ptr); | |
30c82645 | 301 | memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE); |
1da177e4 | 302 | } |
b9d05200 TL |
303 | |
304 | /* | |
305 | * The old boot data is no longer needed and won't be reserved, | |
306 | * freeing up that memory for use by the system. If SME is active, | |
307 | * we need to remove the mappings that were created so that the | |
308 | * memory doesn't remain mapped as decrypted. | |
309 | */ | |
310 | sme_unmap_bootdata(real_mode_data); | |
1da177e4 LT |
311 | } |
312 | ||
2605fc21 | 313 | asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data) |
1da177e4 | 314 | { |
b4e0409a IM |
315 | /* |
316 | * Build-time sanity checks on the kernel image and module | |
317 | * area mappings. (these are purely build-time and produce no code) | |
318 | */ | |
8e3c2a8c BP |
319 | BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map); |
320 | BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE); | |
b4e0409a | 321 | BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE); |
8e3c2a8c | 322 | BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0); |
b4e0409a IM |
323 | BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0); |
324 | BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL)); | |
325 | BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) == | |
326 | (__START_KERNEL & PGDIR_MASK))); | |
66d4bdf2 | 327 | BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END); |
b4e0409a | 328 | |
1e02ce4c AL |
329 | cr4_init_shadow(); |
330 | ||
8170e6be PA |
331 | /* Kill off the identity-map trampoline */ |
332 | reset_early_page_tables(); | |
333 | ||
3df0af0e YL |
334 | clear_bss(); |
335 | ||
65ade2f8 | 336 | clear_page(init_top_pgt); |
d0f77d4d | 337 | |
21729f81 TL |
338 | /* |
339 | * SME support may update early_pmd_flags to include the memory | |
340 | * encryption mask, so it needs to be called before anything | |
341 | * that may generate a page fault. | |
342 | */ | |
343 | sme_early_init(); | |
344 | ||
5d5aa3cf AP |
345 | kasan_early_init(); |
346 | ||
588787fd | 347 | idt_setup_early_handler(); |
f6c2e333 | 348 | |
fa2bbce9 YL |
349 | copy_bootdata(__va(real_mode_data)); |
350 | ||
feddc9de FY |
351 | /* |
352 | * Load microcode early on BSP. | |
353 | */ | |
354 | load_ucode_bsp(); | |
355 | ||
65ade2f8 KS |
356 | /* set init_top_pgt kernel high mapping*/ |
357 | init_top_pgt[511] = early_top_pgt[511]; | |
8170e6be | 358 | |
f97013fd JF |
359 | x86_64_start_reservations(real_mode_data); |
360 | } | |
361 | ||
362 | void __init x86_64_start_reservations(char *real_mode_data) | |
363 | { | |
fa2bbce9 YL |
364 | /* version is always not zero if it is copied */ |
365 | if (!boot_params.hdr.version) | |
366 | copy_bootdata(__va(real_mode_data)); | |
9de819fe | 367 | |
8d152e7a | 368 | x86_early_init_platform_quirks(); |
75175278 | 369 | |
3fda5bb4 AS |
370 | switch (boot_params.hdr.hardware_subarch) { |
371 | case X86_SUBARCH_INTEL_MID: | |
372 | x86_intel_mid_early_setup(); | |
373 | break; | |
374 | default: | |
375 | break; | |
376 | } | |
377 | ||
1da177e4 LT |
378 | start_kernel(); |
379 | } |