Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * prepare to run common code
4 *
5 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
6 */
7
8#define DISABLE_BRANCH_PROFILING
9
10/* cpu_feature_enabled() cannot be used this early */
11#define USE_EARLY_PGTABLE_L5
12
13#include <linux/init.h>
14#include <linux/linkage.h>
15#include <linux/types.h>
16#include <linux/kernel.h>
17#include <linux/string.h>
18#include <linux/percpu.h>
19#include <linux/start_kernel.h>
20#include <linux/io.h>
21#include <linux/memblock.h>
22#include <linux/cc_platform.h>
23#include <linux/pgtable.h>
24
25#include <asm/processor.h>
26#include <asm/proto.h>
27#include <asm/smp.h>
28#include <asm/setup.h>
29#include <asm/desc.h>
30#include <asm/tlbflush.h>
31#include <asm/sections.h>
32#include <asm/kdebug.h>
33#include <asm/e820/api.h>
34#include <asm/bios_ebda.h>
35#include <asm/bootparam_utils.h>
36#include <asm/microcode.h>
37#include <asm/kasan.h>
38#include <asm/fixmap.h>
39#include <asm/realmode.h>
40#include <asm/extable.h>
41#include <asm/trapnr.h>
42#include <asm/sev.h>
43#include <asm/tdx.h>
44
45/*
46 * Manage page tables very early on.
47 */
48extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD];
49static unsigned int __initdata next_early_pgt;
50pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX);
51
52#ifdef CONFIG_X86_5LEVEL
53unsigned int __pgtable_l5_enabled __ro_after_init;
54unsigned int pgdir_shift __ro_after_init = 39;
55EXPORT_SYMBOL(pgdir_shift);
56unsigned int ptrs_per_p4d __ro_after_init = 1;
57EXPORT_SYMBOL(ptrs_per_p4d);
58#endif
59
60#ifdef CONFIG_DYNAMIC_MEMORY_LAYOUT
61unsigned long page_offset_base __ro_after_init = __PAGE_OFFSET_BASE_L4;
62EXPORT_SYMBOL(page_offset_base);
63unsigned long vmalloc_base __ro_after_init = __VMALLOC_BASE_L4;
64EXPORT_SYMBOL(vmalloc_base);
65unsigned long vmemmap_base __ro_after_init = __VMEMMAP_BASE_L4;
66EXPORT_SYMBOL(vmemmap_base);
67#endif
68
69/*
70 * GDT used on the boot CPU before switching to virtual addresses.
71 */
72static struct desc_struct startup_gdt[GDT_ENTRIES] = {
73 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
74 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
75 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
76};
77
78/*
79 * Address needs to be set at runtime because it references the startup_gdt
80 * while the kernel still uses a direct mapping.
81 */
82static struct desc_ptr startup_gdt_descr = {
83 .size = sizeof(startup_gdt),
84 .address = 0,
85};
86
87#define __head __section(".head.text")
88
89static void __head *fixup_pointer(void *ptr, unsigned long physaddr)
90{
91 return ptr - (void *)_text + (void *)physaddr;
92}
93
94static unsigned long __head *fixup_long(void *ptr, unsigned long physaddr)
95{
96 return fixup_pointer(ptr, physaddr);
97}
98
99#ifdef CONFIG_X86_5LEVEL
100static unsigned int __head *fixup_int(void *ptr, unsigned long physaddr)
101{
102 return fixup_pointer(ptr, physaddr);
103}
104
105static bool __head check_la57_support(unsigned long physaddr)
106{
107 /*
108 * 5-level paging is detected and enabled at kernel decompression
109 * stage. Only check if it has been enabled there.
110 */
111 if (!(native_read_cr4() & X86_CR4_LA57))
112 return false;
113
114 *fixup_int(&__pgtable_l5_enabled, physaddr) = 1;
115 *fixup_int(&pgdir_shift, physaddr) = 48;
116 *fixup_int(&ptrs_per_p4d, physaddr) = 512;
117 *fixup_long(&page_offset_base, physaddr) = __PAGE_OFFSET_BASE_L5;
118 *fixup_long(&vmalloc_base, physaddr) = __VMALLOC_BASE_L5;
119 *fixup_long(&vmemmap_base, physaddr) = __VMEMMAP_BASE_L5;
120
121 return true;
122}
123#else
124static bool __head check_la57_support(unsigned long physaddr)
125{
126 return false;
127}
128#endif
129
130static unsigned long __head sme_postprocess_startup(struct boot_params *bp, pmdval_t *pmd)
131{
132 unsigned long vaddr, vaddr_end;
133 int i;
134
135 /* Encrypt the kernel and related (if SME is active) */
136 sme_encrypt_kernel(bp);
137
138 /*
139 * Clear the memory encryption mask from the .bss..decrypted section.
140 * The bss section will be memset to zero later in the initialization so
141 * there is no need to zero it after changing the memory encryption
142 * attribute.
143 */
144 if (sme_get_me_mask()) {
145 vaddr = (unsigned long)__start_bss_decrypted;
146 vaddr_end = (unsigned long)__end_bss_decrypted;
147
148 for (; vaddr < vaddr_end; vaddr += PMD_SIZE) {
149 /*
150 * On SNP, transition the page to shared in the RMP table so that
151 * it is consistent with the page table attribute change.
152 *
153 * __start_bss_decrypted has a virtual address in the high range
154 * mapping (kernel .text). PVALIDATE, by way of
155 * early_snp_set_memory_shared(), requires a valid virtual
156 * address but the kernel is currently running off of the identity
157 * mapping so use __pa() to get a *currently* valid virtual address.
158 */
159 early_snp_set_memory_shared(__pa(vaddr), __pa(vaddr), PTRS_PER_PMD);
160
161 i = pmd_index(vaddr);
162 pmd[i] -= sme_get_me_mask();
163 }
164 }
165
166 /*
167 * Return the SME encryption mask (if SME is active) to be used as a
168 * modifier for the initial pgdir entry programmed into CR3.
169 */
170 return sme_get_me_mask();
171}
172
173/* Code in __startup_64() can be relocated during execution, but the compiler
174 * doesn't have to generate PC-relative relocations when accessing globals from
175 * that function. Clang actually does not generate them, which leads to
176 * boot-time crashes. To work around this problem, every global pointer must
177 * be adjusted using fixup_pointer().
178 */
179unsigned long __head __startup_64(unsigned long physaddr,
180 struct boot_params *bp)
181{
182 unsigned long load_delta, *p;
183 unsigned long pgtable_flags;
184 pgdval_t *pgd;
185 p4dval_t *p4d;
186 pudval_t *pud;
187 pmdval_t *pmd, pmd_entry;
188 pteval_t *mask_ptr;
189 bool la57;
190 int i;
191 unsigned int *next_pgt_ptr;
192
193 la57 = check_la57_support(physaddr);
194
195 /* Is the address too large? */
196 if (physaddr >> MAX_PHYSMEM_BITS)
197 for (;;);
198
199 /*
200 * Compute the delta between the address I am compiled to run at
201 * and the address I am actually running at.
202 */
203 load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
204
205 /* Is the address not 2M aligned? */
206 if (load_delta & ~PMD_MASK)
207 for (;;);
208
209 /* Include the SME encryption mask in the fixup value */
210 load_delta += sme_get_me_mask();
211
212 /* Fixup the physical addresses in the page table */
213
214 pgd = fixup_pointer(&early_top_pgt, physaddr);
215 p = pgd + pgd_index(__START_KERNEL_map);
216 if (la57)
217 *p = (unsigned long)level4_kernel_pgt;
218 else
219 *p = (unsigned long)level3_kernel_pgt;
220 *p += _PAGE_TABLE_NOENC - __START_KERNEL_map + load_delta;
221
222 if (la57) {
223 p4d = fixup_pointer(&level4_kernel_pgt, physaddr);
224 p4d[511] += load_delta;
225 }
226
227 pud = fixup_pointer(&level3_kernel_pgt, physaddr);
228 pud[510] += load_delta;
229 pud[511] += load_delta;
230
231 pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
232 for (i = FIXMAP_PMD_TOP; i > FIXMAP_PMD_TOP - FIXMAP_PMD_NUM; i--)
233 pmd[i] += load_delta;
234
235 /*
236 * Set up the identity mapping for the switchover. These
237 * entries should *NOT* have the global bit set! This also
238 * creates a bunch of nonsense entries but that is fine --
239 * it avoids problems around wraparound.
240 */
241
242 next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr);
243 pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
244 pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
245
246 pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask();
247
248 if (la57) {
249 p4d = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++],
250 physaddr);
251
252 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
253 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
254 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
255
256 i = physaddr >> P4D_SHIFT;
257 p4d[(i + 0) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags;
258 p4d[(i + 1) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags;
259 } else {
260 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
261 pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
262 pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
263 }
264
265 i = physaddr >> PUD_SHIFT;
266 pud[(i + 0) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags;
267 pud[(i + 1) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags;
268
269 pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
270 /* Filter out unsupported __PAGE_KERNEL_* bits: */
271 mask_ptr = fixup_pointer(&__supported_pte_mask, physaddr);
272 pmd_entry &= *mask_ptr;
273 pmd_entry += sme_get_me_mask();
274 pmd_entry += physaddr;
275
276 for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
277 int idx = i + (physaddr >> PMD_SHIFT);
278
279 pmd[idx % PTRS_PER_PMD] = pmd_entry + i * PMD_SIZE;
280 }
281
282 /*
283 * Fixup the kernel text+data virtual addresses. Note that
284 * we might write invalid pmds, when the kernel is relocated
285 * cleanup_highmap() fixes this up along with the mappings
286 * beyond _end.
287 *
288 * Only the region occupied by the kernel image has so far
289 * been checked against the table of usable memory regions
290 * provided by the firmware, so invalidate pages outside that
291 * region. A page table entry that maps to a reserved area of
292 * memory would allow processor speculation into that area,
293 * and on some hardware (particularly the UV platform) even
294 * speculative access to some reserved areas is caught as an
295 * error, causing the BIOS to halt the system.
296 */
297
298 pmd = fixup_pointer(level2_kernel_pgt, physaddr);
299
300 /* invalidate pages before the kernel image */
301 for (i = 0; i < pmd_index((unsigned long)_text); i++)
302 pmd[i] &= ~_PAGE_PRESENT;
303
304 /* fixup pages that are part of the kernel image */
305 for (; i <= pmd_index((unsigned long)_end); i++)
306 if (pmd[i] & _PAGE_PRESENT)
307 pmd[i] += load_delta;
308
309 /* invalidate pages after the kernel image */
310 for (; i < PTRS_PER_PMD; i++)
311 pmd[i] &= ~_PAGE_PRESENT;
312
313 /*
314 * Fixup phys_base - remove the memory encryption mask to obtain
315 * the true physical address.
316 */
317 *fixup_long(&phys_base, physaddr) += load_delta - sme_get_me_mask();
318
319 return sme_postprocess_startup(bp, pmd);
320}
321
322/* Wipe all early page tables except for the kernel symbol map */
323static void __init reset_early_page_tables(void)
324{
325 memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
326 next_early_pgt = 0;
327 write_cr3(__sme_pa_nodebug(early_top_pgt));
328}
329
330/* Create a new PMD entry */
331bool __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
332{
333 unsigned long physaddr = address - __PAGE_OFFSET;
334 pgdval_t pgd, *pgd_p;
335 p4dval_t p4d, *p4d_p;
336 pudval_t pud, *pud_p;
337 pmdval_t *pmd_p;
338
339 /* Invalid address or early pgt is done ? */
340 if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
341 return false;
342
343again:
344 pgd_p = &early_top_pgt[pgd_index(address)].pgd;
345 pgd = *pgd_p;
346
347 /*
348 * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
349 * critical -- __PAGE_OFFSET would point us back into the dynamic
350 * range and we might end up looping forever...
351 */
352 if (!pgtable_l5_enabled())
353 p4d_p = pgd_p;
354 else if (pgd)
355 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
356 else {
357 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
358 reset_early_page_tables();
359 goto again;
360 }
361
362 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
363 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
364 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
365 }
366 p4d_p += p4d_index(address);
367 p4d = *p4d_p;
368
369 if (p4d)
370 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
371 else {
372 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
373 reset_early_page_tables();
374 goto again;
375 }
376
377 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
378 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
379 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
380 }
381 pud_p += pud_index(address);
382 pud = *pud_p;
383
384 if (pud)
385 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
386 else {
387 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
388 reset_early_page_tables();
389 goto again;
390 }
391
392 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
393 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
394 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
395 }
396 pmd_p[pmd_index(address)] = pmd;
397
398 return true;
399}
400
401static bool __init early_make_pgtable(unsigned long address)
402{
403 unsigned long physaddr = address - __PAGE_OFFSET;
404 pmdval_t pmd;
405
406 pmd = (physaddr & PMD_MASK) + early_pmd_flags;
407
408 return __early_make_pgtable(address, pmd);
409}
410
411void __init do_early_exception(struct pt_regs *regs, int trapnr)
412{
413 if (trapnr == X86_TRAP_PF &&
414 early_make_pgtable(native_read_cr2()))
415 return;
416
417 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT) &&
418 trapnr == X86_TRAP_VC && handle_vc_boot_ghcb(regs))
419 return;
420
421 if (trapnr == X86_TRAP_VE && tdx_early_handle_ve(regs))
422 return;
423
424 early_fixup_exception(regs, trapnr);
425}
426
427/* Don't add a printk in there. printk relies on the PDA which is not initialized
428 yet. */
429void __init clear_bss(void)
430{
431 memset(__bss_start, 0,
432 (unsigned long) __bss_stop - (unsigned long) __bss_start);
433 memset(__brk_base, 0,
434 (unsigned long) __brk_limit - (unsigned long) __brk_base);
435}
436
437static unsigned long get_cmd_line_ptr(void)
438{
439 unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
440
441 cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
442
443 return cmd_line_ptr;
444}
445
446static void __init copy_bootdata(char *real_mode_data)
447{
448 char * command_line;
449 unsigned long cmd_line_ptr;
450
451 /*
452 * If SME is active, this will create decrypted mappings of the
453 * boot data in advance of the copy operations.
454 */
455 sme_map_bootdata(real_mode_data);
456
457 memcpy(&boot_params, real_mode_data, sizeof(boot_params));
458 sanitize_boot_params(&boot_params);
459 cmd_line_ptr = get_cmd_line_ptr();
460 if (cmd_line_ptr) {
461 command_line = __va(cmd_line_ptr);
462 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
463 }
464
465 /*
466 * The old boot data is no longer needed and won't be reserved,
467 * freeing up that memory for use by the system. If SME is active,
468 * we need to remove the mappings that were created so that the
469 * memory doesn't remain mapped as decrypted.
470 */
471 sme_unmap_bootdata(real_mode_data);
472}
473
474asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
475{
476 /*
477 * Build-time sanity checks on the kernel image and module
478 * area mappings. (these are purely build-time and produce no code)
479 */
480 BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
481 BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
482 BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
483 BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
484 BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
485 BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
486 MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
487 (__START_KERNEL & PGDIR_MASK)));
488 BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
489
490 cr4_init_shadow();
491
492 /* Kill off the identity-map trampoline */
493 reset_early_page_tables();
494
495 clear_bss();
496
497 /*
498 * This needs to happen *before* kasan_early_init() because latter maps stuff
499 * into that page.
500 */
501 clear_page(init_top_pgt);
502
503 /*
504 * SME support may update early_pmd_flags to include the memory
505 * encryption mask, so it needs to be called before anything
506 * that may generate a page fault.
507 */
508 sme_early_init();
509
510 kasan_early_init();
511
512 /*
513 * Flush global TLB entries which could be left over from the trampoline page
514 * table.
515 *
516 * This needs to happen *after* kasan_early_init() as KASAN-enabled .configs
517 * instrument native_write_cr4() so KASAN must be initialized for that
518 * instrumentation to work.
519 */
520 __native_tlb_flush_global(this_cpu_read(cpu_tlbstate.cr4));
521
522 idt_setup_early_handler();
523
524 /* Needed before cc_platform_has() can be used for TDX */
525 tdx_early_init();
526
527 copy_bootdata(__va(real_mode_data));
528
529 /*
530 * Load microcode early on BSP.
531 */
532 load_ucode_bsp();
533
534 /* set init_top_pgt kernel high mapping*/
535 init_top_pgt[511] = early_top_pgt[511];
536
537 x86_64_start_reservations(real_mode_data);
538}
539
540void __init x86_64_start_reservations(char *real_mode_data)
541{
542 /* version is always not zero if it is copied */
543 if (!boot_params.hdr.version)
544 copy_bootdata(__va(real_mode_data));
545
546 x86_early_init_platform_quirks();
547
548 switch (boot_params.hdr.hardware_subarch) {
549 case X86_SUBARCH_INTEL_MID:
550 x86_intel_mid_early_setup();
551 break;
552 default:
553 break;
554 }
555
556 start_kernel();
557}
558
559/*
560 * Data structures and code used for IDT setup in head_64.S. The bringup-IDT is
561 * used until the idt_table takes over. On the boot CPU this happens in
562 * x86_64_start_kernel(), on secondary CPUs in start_secondary(). In both cases
563 * this happens in the functions called from head_64.S.
564 *
565 * The idt_table can't be used that early because all the code modifying it is
566 * in idt.c and can be instrumented by tracing or KASAN, which both don't work
567 * during early CPU bringup. Also the idt_table has the runtime vectors
568 * configured which require certain CPU state to be setup already (like TSS),
569 * which also hasn't happened yet in early CPU bringup.
570 */
571static gate_desc bringup_idt_table[NUM_EXCEPTION_VECTORS] __page_aligned_data;
572
573static struct desc_ptr bringup_idt_descr = {
574 .size = (NUM_EXCEPTION_VECTORS * sizeof(gate_desc)) - 1,
575 .address = 0, /* Set at runtime */
576};
577
578static void set_bringup_idt_handler(gate_desc *idt, int n, void *handler)
579{
580#ifdef CONFIG_AMD_MEM_ENCRYPT
581 struct idt_data data;
582 gate_desc desc;
583
584 init_idt_data(&data, n, handler);
585 idt_init_desc(&desc, &data);
586 native_write_idt_entry(idt, n, &desc);
587#endif
588}
589
590/* This runs while still in the direct mapping */
591static void startup_64_load_idt(unsigned long physbase)
592{
593 struct desc_ptr *desc = fixup_pointer(&bringup_idt_descr, physbase);
594 gate_desc *idt = fixup_pointer(bringup_idt_table, physbase);
595
596
597 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT)) {
598 void *handler;
599
600 /* VMM Communication Exception */
601 handler = fixup_pointer(vc_no_ghcb, physbase);
602 set_bringup_idt_handler(idt, X86_TRAP_VC, handler);
603 }
604
605 desc->address = (unsigned long)idt;
606 native_load_idt(desc);
607}
608
609/* This is used when running on kernel addresses */
610void early_setup_idt(void)
611{
612 /* VMM Communication Exception */
613 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT)) {
614 setup_ghcb();
615 set_bringup_idt_handler(bringup_idt_table, X86_TRAP_VC, vc_boot_ghcb);
616 }
617
618 bringup_idt_descr.address = (unsigned long)bringup_idt_table;
619 native_load_idt(&bringup_idt_descr);
620}
621
622/*
623 * Setup boot CPU state needed before kernel switches to virtual addresses.
624 */
625void __head startup_64_setup_env(unsigned long physbase)
626{
627 /* Load GDT */
628 startup_gdt_descr.address = (unsigned long)fixup_pointer(startup_gdt, physbase);
629 native_load_gdt(&startup_gdt_descr);
630
631 /* New GDT is live - reload data segment registers */
632 asm volatile("movl %%eax, %%ds\n"
633 "movl %%eax, %%ss\n"
634 "movl %%eax, %%es\n" : : "a"(__KERNEL_DS) : "memory");
635
636 startup_64_load_idt(physbase);
637}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * prepare to run common code
4 *
5 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
6 */
7
8#define DISABLE_BRANCH_PROFILING
9
10/* cpu_feature_enabled() cannot be used this early */
11#define USE_EARLY_PGTABLE_L5
12
13#include <linux/init.h>
14#include <linux/linkage.h>
15#include <linux/types.h>
16#include <linux/kernel.h>
17#include <linux/string.h>
18#include <linux/percpu.h>
19#include <linux/start_kernel.h>
20#include <linux/io.h>
21#include <linux/memblock.h>
22#include <linux/mem_encrypt.h>
23#include <linux/pgtable.h>
24
25#include <asm/processor.h>
26#include <asm/proto.h>
27#include <asm/smp.h>
28#include <asm/setup.h>
29#include <asm/desc.h>
30#include <asm/tlbflush.h>
31#include <asm/sections.h>
32#include <asm/kdebug.h>
33#include <asm/e820/api.h>
34#include <asm/bios_ebda.h>
35#include <asm/bootparam_utils.h>
36#include <asm/microcode.h>
37#include <asm/kasan.h>
38#include <asm/fixmap.h>
39#include <asm/realmode.h>
40#include <asm/extable.h>
41#include <asm/trapnr.h>
42#include <asm/sev.h>
43
44/*
45 * Manage page tables very early on.
46 */
47extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD];
48static unsigned int __initdata next_early_pgt;
49pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX);
50
51#ifdef CONFIG_X86_5LEVEL
52unsigned int __pgtable_l5_enabled __ro_after_init;
53unsigned int pgdir_shift __ro_after_init = 39;
54EXPORT_SYMBOL(pgdir_shift);
55unsigned int ptrs_per_p4d __ro_after_init = 1;
56EXPORT_SYMBOL(ptrs_per_p4d);
57#endif
58
59#ifdef CONFIG_DYNAMIC_MEMORY_LAYOUT
60unsigned long page_offset_base __ro_after_init = __PAGE_OFFSET_BASE_L4;
61EXPORT_SYMBOL(page_offset_base);
62unsigned long vmalloc_base __ro_after_init = __VMALLOC_BASE_L4;
63EXPORT_SYMBOL(vmalloc_base);
64unsigned long vmemmap_base __ro_after_init = __VMEMMAP_BASE_L4;
65EXPORT_SYMBOL(vmemmap_base);
66#endif
67
68/*
69 * GDT used on the boot CPU before switching to virtual addresses.
70 */
71static struct desc_struct startup_gdt[GDT_ENTRIES] = {
72 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
73 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
74 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
75};
76
77/*
78 * Address needs to be set at runtime because it references the startup_gdt
79 * while the kernel still uses a direct mapping.
80 */
81static struct desc_ptr startup_gdt_descr = {
82 .size = sizeof(startup_gdt),
83 .address = 0,
84};
85
86#define __head __section(".head.text")
87
88static void __head *fixup_pointer(void *ptr, unsigned long physaddr)
89{
90 return ptr - (void *)_text + (void *)physaddr;
91}
92
93static unsigned long __head *fixup_long(void *ptr, unsigned long physaddr)
94{
95 return fixup_pointer(ptr, physaddr);
96}
97
98#ifdef CONFIG_X86_5LEVEL
99static unsigned int __head *fixup_int(void *ptr, unsigned long physaddr)
100{
101 return fixup_pointer(ptr, physaddr);
102}
103
104static bool __head check_la57_support(unsigned long physaddr)
105{
106 /*
107 * 5-level paging is detected and enabled at kernel decompression
108 * stage. Only check if it has been enabled there.
109 */
110 if (!(native_read_cr4() & X86_CR4_LA57))
111 return false;
112
113 *fixup_int(&__pgtable_l5_enabled, physaddr) = 1;
114 *fixup_int(&pgdir_shift, physaddr) = 48;
115 *fixup_int(&ptrs_per_p4d, physaddr) = 512;
116 *fixup_long(&page_offset_base, physaddr) = __PAGE_OFFSET_BASE_L5;
117 *fixup_long(&vmalloc_base, physaddr) = __VMALLOC_BASE_L5;
118 *fixup_long(&vmemmap_base, physaddr) = __VMEMMAP_BASE_L5;
119
120 return true;
121}
122#else
123static bool __head check_la57_support(unsigned long physaddr)
124{
125 return false;
126}
127#endif
128
129/* Code in __startup_64() can be relocated during execution, but the compiler
130 * doesn't have to generate PC-relative relocations when accessing globals from
131 * that function. Clang actually does not generate them, which leads to
132 * boot-time crashes. To work around this problem, every global pointer must
133 * be adjusted using fixup_pointer().
134 */
135unsigned long __head __startup_64(unsigned long physaddr,
136 struct boot_params *bp)
137{
138 unsigned long vaddr, vaddr_end;
139 unsigned long load_delta, *p;
140 unsigned long pgtable_flags;
141 pgdval_t *pgd;
142 p4dval_t *p4d;
143 pudval_t *pud;
144 pmdval_t *pmd, pmd_entry;
145 pteval_t *mask_ptr;
146 bool la57;
147 int i;
148 unsigned int *next_pgt_ptr;
149
150 la57 = check_la57_support(physaddr);
151
152 /* Is the address too large? */
153 if (physaddr >> MAX_PHYSMEM_BITS)
154 for (;;);
155
156 /*
157 * Compute the delta between the address I am compiled to run at
158 * and the address I am actually running at.
159 */
160 load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
161
162 /* Is the address not 2M aligned? */
163 if (load_delta & ~PMD_PAGE_MASK)
164 for (;;);
165
166 /* Activate Secure Memory Encryption (SME) if supported and enabled */
167 sme_enable(bp);
168
169 /* Include the SME encryption mask in the fixup value */
170 load_delta += sme_get_me_mask();
171
172 /* Fixup the physical addresses in the page table */
173
174 pgd = fixup_pointer(&early_top_pgt, physaddr);
175 p = pgd + pgd_index(__START_KERNEL_map);
176 if (la57)
177 *p = (unsigned long)level4_kernel_pgt;
178 else
179 *p = (unsigned long)level3_kernel_pgt;
180 *p += _PAGE_TABLE_NOENC - __START_KERNEL_map + load_delta;
181
182 if (la57) {
183 p4d = fixup_pointer(&level4_kernel_pgt, physaddr);
184 p4d[511] += load_delta;
185 }
186
187 pud = fixup_pointer(&level3_kernel_pgt, physaddr);
188 pud[510] += load_delta;
189 pud[511] += load_delta;
190
191 pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
192 for (i = FIXMAP_PMD_TOP; i > FIXMAP_PMD_TOP - FIXMAP_PMD_NUM; i--)
193 pmd[i] += load_delta;
194
195 /*
196 * Set up the identity mapping for the switchover. These
197 * entries should *NOT* have the global bit set! This also
198 * creates a bunch of nonsense entries but that is fine --
199 * it avoids problems around wraparound.
200 */
201
202 next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr);
203 pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
204 pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
205
206 pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask();
207
208 if (la57) {
209 p4d = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++],
210 physaddr);
211
212 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
213 pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
214 pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
215
216 i = physaddr >> P4D_SHIFT;
217 p4d[(i + 0) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags;
218 p4d[(i + 1) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags;
219 } else {
220 i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
221 pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
222 pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
223 }
224
225 i = physaddr >> PUD_SHIFT;
226 pud[(i + 0) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags;
227 pud[(i + 1) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags;
228
229 pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
230 /* Filter out unsupported __PAGE_KERNEL_* bits: */
231 mask_ptr = fixup_pointer(&__supported_pte_mask, physaddr);
232 pmd_entry &= *mask_ptr;
233 pmd_entry += sme_get_me_mask();
234 pmd_entry += physaddr;
235
236 for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
237 int idx = i + (physaddr >> PMD_SHIFT);
238
239 pmd[idx % PTRS_PER_PMD] = pmd_entry + i * PMD_SIZE;
240 }
241
242 /*
243 * Fixup the kernel text+data virtual addresses. Note that
244 * we might write invalid pmds, when the kernel is relocated
245 * cleanup_highmap() fixes this up along with the mappings
246 * beyond _end.
247 *
248 * Only the region occupied by the kernel image has so far
249 * been checked against the table of usable memory regions
250 * provided by the firmware, so invalidate pages outside that
251 * region. A page table entry that maps to a reserved area of
252 * memory would allow processor speculation into that area,
253 * and on some hardware (particularly the UV platform) even
254 * speculative access to some reserved areas is caught as an
255 * error, causing the BIOS to halt the system.
256 */
257
258 pmd = fixup_pointer(level2_kernel_pgt, physaddr);
259
260 /* invalidate pages before the kernel image */
261 for (i = 0; i < pmd_index((unsigned long)_text); i++)
262 pmd[i] &= ~_PAGE_PRESENT;
263
264 /* fixup pages that are part of the kernel image */
265 for (; i <= pmd_index((unsigned long)_end); i++)
266 if (pmd[i] & _PAGE_PRESENT)
267 pmd[i] += load_delta;
268
269 /* invalidate pages after the kernel image */
270 for (; i < PTRS_PER_PMD; i++)
271 pmd[i] &= ~_PAGE_PRESENT;
272
273 /*
274 * Fixup phys_base - remove the memory encryption mask to obtain
275 * the true physical address.
276 */
277 *fixup_long(&phys_base, physaddr) += load_delta - sme_get_me_mask();
278
279 /* Encrypt the kernel and related (if SME is active) */
280 sme_encrypt_kernel(bp);
281
282 /*
283 * Clear the memory encryption mask from the .bss..decrypted section.
284 * The bss section will be memset to zero later in the initialization so
285 * there is no need to zero it after changing the memory encryption
286 * attribute.
287 */
288 if (mem_encrypt_active()) {
289 vaddr = (unsigned long)__start_bss_decrypted;
290 vaddr_end = (unsigned long)__end_bss_decrypted;
291 for (; vaddr < vaddr_end; vaddr += PMD_SIZE) {
292 i = pmd_index(vaddr);
293 pmd[i] -= sme_get_me_mask();
294 }
295 }
296
297 /*
298 * Return the SME encryption mask (if SME is active) to be used as a
299 * modifier for the initial pgdir entry programmed into CR3.
300 */
301 return sme_get_me_mask();
302}
303
304unsigned long __startup_secondary_64(void)
305{
306 /*
307 * Return the SME encryption mask (if SME is active) to be used as a
308 * modifier for the initial pgdir entry programmed into CR3.
309 */
310 return sme_get_me_mask();
311}
312
313/* Wipe all early page tables except for the kernel symbol map */
314static void __init reset_early_page_tables(void)
315{
316 memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
317 next_early_pgt = 0;
318 write_cr3(__sme_pa_nodebug(early_top_pgt));
319}
320
321/* Create a new PMD entry */
322bool __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
323{
324 unsigned long physaddr = address - __PAGE_OFFSET;
325 pgdval_t pgd, *pgd_p;
326 p4dval_t p4d, *p4d_p;
327 pudval_t pud, *pud_p;
328 pmdval_t *pmd_p;
329
330 /* Invalid address or early pgt is done ? */
331 if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
332 return false;
333
334again:
335 pgd_p = &early_top_pgt[pgd_index(address)].pgd;
336 pgd = *pgd_p;
337
338 /*
339 * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
340 * critical -- __PAGE_OFFSET would point us back into the dynamic
341 * range and we might end up looping forever...
342 */
343 if (!pgtable_l5_enabled())
344 p4d_p = pgd_p;
345 else if (pgd)
346 p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
347 else {
348 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
349 reset_early_page_tables();
350 goto again;
351 }
352
353 p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
354 memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
355 *pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
356 }
357 p4d_p += p4d_index(address);
358 p4d = *p4d_p;
359
360 if (p4d)
361 pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
362 else {
363 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
364 reset_early_page_tables();
365 goto again;
366 }
367
368 pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
369 memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
370 *p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
371 }
372 pud_p += pud_index(address);
373 pud = *pud_p;
374
375 if (pud)
376 pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
377 else {
378 if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
379 reset_early_page_tables();
380 goto again;
381 }
382
383 pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
384 memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
385 *pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
386 }
387 pmd_p[pmd_index(address)] = pmd;
388
389 return true;
390}
391
392static bool __init early_make_pgtable(unsigned long address)
393{
394 unsigned long physaddr = address - __PAGE_OFFSET;
395 pmdval_t pmd;
396
397 pmd = (physaddr & PMD_MASK) + early_pmd_flags;
398
399 return __early_make_pgtable(address, pmd);
400}
401
402void __init do_early_exception(struct pt_regs *regs, int trapnr)
403{
404 if (trapnr == X86_TRAP_PF &&
405 early_make_pgtable(native_read_cr2()))
406 return;
407
408 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT) &&
409 trapnr == X86_TRAP_VC && handle_vc_boot_ghcb(regs))
410 return;
411
412 early_fixup_exception(regs, trapnr);
413}
414
415/* Don't add a printk in there. printk relies on the PDA which is not initialized
416 yet. */
417static void __init clear_bss(void)
418{
419 memset(__bss_start, 0,
420 (unsigned long) __bss_stop - (unsigned long) __bss_start);
421}
422
423static unsigned long get_cmd_line_ptr(void)
424{
425 unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
426
427 cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
428
429 return cmd_line_ptr;
430}
431
432static void __init copy_bootdata(char *real_mode_data)
433{
434 char * command_line;
435 unsigned long cmd_line_ptr;
436
437 /*
438 * If SME is active, this will create decrypted mappings of the
439 * boot data in advance of the copy operations.
440 */
441 sme_map_bootdata(real_mode_data);
442
443 memcpy(&boot_params, real_mode_data, sizeof(boot_params));
444 sanitize_boot_params(&boot_params);
445 cmd_line_ptr = get_cmd_line_ptr();
446 if (cmd_line_ptr) {
447 command_line = __va(cmd_line_ptr);
448 memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
449 }
450
451 /*
452 * The old boot data is no longer needed and won't be reserved,
453 * freeing up that memory for use by the system. If SME is active,
454 * we need to remove the mappings that were created so that the
455 * memory doesn't remain mapped as decrypted.
456 */
457 sme_unmap_bootdata(real_mode_data);
458}
459
460asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
461{
462 /*
463 * Build-time sanity checks on the kernel image and module
464 * area mappings. (these are purely build-time and produce no code)
465 */
466 BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
467 BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
468 BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
469 BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
470 BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
471 BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
472 MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
473 (__START_KERNEL & PGDIR_MASK)));
474 BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
475
476 cr4_init_shadow();
477
478 /* Kill off the identity-map trampoline */
479 reset_early_page_tables();
480
481 clear_bss();
482
483 clear_page(init_top_pgt);
484
485 /*
486 * SME support may update early_pmd_flags to include the memory
487 * encryption mask, so it needs to be called before anything
488 * that may generate a page fault.
489 */
490 sme_early_init();
491
492 kasan_early_init();
493
494 idt_setup_early_handler();
495
496 copy_bootdata(__va(real_mode_data));
497
498 /*
499 * Load microcode early on BSP.
500 */
501 load_ucode_bsp();
502
503 /* set init_top_pgt kernel high mapping*/
504 init_top_pgt[511] = early_top_pgt[511];
505
506 x86_64_start_reservations(real_mode_data);
507}
508
509void __init x86_64_start_reservations(char *real_mode_data)
510{
511 /* version is always not zero if it is copied */
512 if (!boot_params.hdr.version)
513 copy_bootdata(__va(real_mode_data));
514
515 x86_early_init_platform_quirks();
516
517 switch (boot_params.hdr.hardware_subarch) {
518 case X86_SUBARCH_INTEL_MID:
519 x86_intel_mid_early_setup();
520 break;
521 default:
522 break;
523 }
524
525 start_kernel();
526}
527
528/*
529 * Data structures and code used for IDT setup in head_64.S. The bringup-IDT is
530 * used until the idt_table takes over. On the boot CPU this happens in
531 * x86_64_start_kernel(), on secondary CPUs in start_secondary(). In both cases
532 * this happens in the functions called from head_64.S.
533 *
534 * The idt_table can't be used that early because all the code modifying it is
535 * in idt.c and can be instrumented by tracing or KASAN, which both don't work
536 * during early CPU bringup. Also the idt_table has the runtime vectors
537 * configured which require certain CPU state to be setup already (like TSS),
538 * which also hasn't happened yet in early CPU bringup.
539 */
540static gate_desc bringup_idt_table[NUM_EXCEPTION_VECTORS] __page_aligned_data;
541
542static struct desc_ptr bringup_idt_descr = {
543 .size = (NUM_EXCEPTION_VECTORS * sizeof(gate_desc)) - 1,
544 .address = 0, /* Set at runtime */
545};
546
547static void set_bringup_idt_handler(gate_desc *idt, int n, void *handler)
548{
549#ifdef CONFIG_AMD_MEM_ENCRYPT
550 struct idt_data data;
551 gate_desc desc;
552
553 init_idt_data(&data, n, handler);
554 idt_init_desc(&desc, &data);
555 native_write_idt_entry(idt, n, &desc);
556#endif
557}
558
559/* This runs while still in the direct mapping */
560static void startup_64_load_idt(unsigned long physbase)
561{
562 struct desc_ptr *desc = fixup_pointer(&bringup_idt_descr, physbase);
563 gate_desc *idt = fixup_pointer(bringup_idt_table, physbase);
564
565
566 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT)) {
567 void *handler;
568
569 /* VMM Communication Exception */
570 handler = fixup_pointer(vc_no_ghcb, physbase);
571 set_bringup_idt_handler(idt, X86_TRAP_VC, handler);
572 }
573
574 desc->address = (unsigned long)idt;
575 native_load_idt(desc);
576}
577
578/* This is used when running on kernel addresses */
579void early_setup_idt(void)
580{
581 /* VMM Communication Exception */
582 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT))
583 set_bringup_idt_handler(bringup_idt_table, X86_TRAP_VC, vc_boot_ghcb);
584
585 bringup_idt_descr.address = (unsigned long)bringup_idt_table;
586 native_load_idt(&bringup_idt_descr);
587}
588
589/*
590 * Setup boot CPU state needed before kernel switches to virtual addresses.
591 */
592void __head startup_64_setup_env(unsigned long physbase)
593{
594 /* Load GDT */
595 startup_gdt_descr.address = (unsigned long)fixup_pointer(startup_gdt, physbase);
596 native_load_gdt(&startup_gdt_descr);
597
598 /* New GDT is live - reload data segment registers */
599 asm volatile("movl %%eax, %%ds\n"
600 "movl %%eax, %%ss\n"
601 "movl %%eax, %%es\n" : : "a"(__KERNEL_DS) : "memory");
602
603 startup_64_load_idt(physbase);
604}