xref: /linux/arch/x86/kernel/head64.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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/asm.h>
26 #include <asm/page_64.h>
27 #include <asm/processor.h>
28 #include <asm/proto.h>
29 #include <asm/smp.h>
30 #include <asm/setup.h>
31 #include <asm/desc.h>
32 #include <asm/tlbflush.h>
33 #include <asm/sections.h>
34 #include <asm/kdebug.h>
35 #include <asm/e820/api.h>
36 #include <asm/bios_ebda.h>
37 #include <asm/bootparam_utils.h>
38 #include <asm/microcode.h>
39 #include <asm/kasan.h>
40 #include <asm/fixmap.h>
41 #include <asm/realmode.h>
42 #include <asm/extable.h>
43 #include <asm/trapnr.h>
44 #include <asm/sev.h>
45 #include <asm/tdx.h>
46 #include <asm/init.h>
47 
48 /*
49  * Manage page tables very early on.
50  */
51 extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD];
52 static unsigned int __initdata next_early_pgt;
53 pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX);
54 
55 #ifdef CONFIG_X86_5LEVEL
56 unsigned int __pgtable_l5_enabled __ro_after_init;
57 unsigned int pgdir_shift __ro_after_init = 39;
58 EXPORT_SYMBOL(pgdir_shift);
59 unsigned int ptrs_per_p4d __ro_after_init = 1;
60 EXPORT_SYMBOL(ptrs_per_p4d);
61 #endif
62 
63 #ifdef CONFIG_DYNAMIC_MEMORY_LAYOUT
64 unsigned long page_offset_base __ro_after_init = __PAGE_OFFSET_BASE_L4;
65 EXPORT_SYMBOL(page_offset_base);
66 unsigned long vmalloc_base __ro_after_init = __VMALLOC_BASE_L4;
67 EXPORT_SYMBOL(vmalloc_base);
68 unsigned long vmemmap_base __ro_after_init = __VMEMMAP_BASE_L4;
69 EXPORT_SYMBOL(vmemmap_base);
70 #endif
71 
72 static inline bool check_la57_support(void)
73 {
74 	if (!IS_ENABLED(CONFIG_X86_5LEVEL))
75 		return false;
76 
77 	/*
78 	 * 5-level paging is detected and enabled at kernel decompression
79 	 * stage. Only check if it has been enabled there.
80 	 */
81 	if (!(native_read_cr4() & X86_CR4_LA57))
82 		return false;
83 
84 	RIP_REL_REF(__pgtable_l5_enabled)	= 1;
85 	RIP_REL_REF(pgdir_shift)		= 48;
86 	RIP_REL_REF(ptrs_per_p4d)		= 512;
87 	RIP_REL_REF(page_offset_base)		= __PAGE_OFFSET_BASE_L5;
88 	RIP_REL_REF(vmalloc_base)		= __VMALLOC_BASE_L5;
89 	RIP_REL_REF(vmemmap_base)		= __VMEMMAP_BASE_L5;
90 
91 	return true;
92 }
93 
94 static unsigned long __head sme_postprocess_startup(struct boot_params *bp, pmdval_t *pmd)
95 {
96 	unsigned long vaddr, vaddr_end;
97 	int i;
98 
99 	/* Encrypt the kernel and related (if SME is active) */
100 	sme_encrypt_kernel(bp);
101 
102 	/*
103 	 * Clear the memory encryption mask from the .bss..decrypted section.
104 	 * The bss section will be memset to zero later in the initialization so
105 	 * there is no need to zero it after changing the memory encryption
106 	 * attribute.
107 	 */
108 	if (sme_get_me_mask()) {
109 		vaddr = (unsigned long)__start_bss_decrypted;
110 		vaddr_end = (unsigned long)__end_bss_decrypted;
111 
112 		for (; vaddr < vaddr_end; vaddr += PMD_SIZE) {
113 			/*
114 			 * On SNP, transition the page to shared in the RMP table so that
115 			 * it is consistent with the page table attribute change.
116 			 *
117 			 * __start_bss_decrypted has a virtual address in the high range
118 			 * mapping (kernel .text). PVALIDATE, by way of
119 			 * early_snp_set_memory_shared(), requires a valid virtual
120 			 * address but the kernel is currently running off of the identity
121 			 * mapping so use __pa() to get a *currently* valid virtual address.
122 			 */
123 			early_snp_set_memory_shared(__pa(vaddr), __pa(vaddr), PTRS_PER_PMD);
124 
125 			i = pmd_index(vaddr);
126 			pmd[i] -= sme_get_me_mask();
127 		}
128 	}
129 
130 	/*
131 	 * Return the SME encryption mask (if SME is active) to be used as a
132 	 * modifier for the initial pgdir entry programmed into CR3.
133 	 */
134 	return sme_get_me_mask();
135 }
136 
137 /* Code in __startup_64() can be relocated during execution, but the compiler
138  * doesn't have to generate PC-relative relocations when accessing globals from
139  * that function. Clang actually does not generate them, which leads to
140  * boot-time crashes. To work around this problem, every global pointer must
141  * be accessed using RIP_REL_REF().
142  */
143 unsigned long __head __startup_64(unsigned long physaddr,
144 				  struct boot_params *bp)
145 {
146 	pmd_t (*early_pgts)[PTRS_PER_PMD] = RIP_REL_REF(early_dynamic_pgts);
147 	unsigned long pgtable_flags;
148 	unsigned long load_delta;
149 	pgdval_t *pgd;
150 	p4dval_t *p4d;
151 	pudval_t *pud;
152 	pmdval_t *pmd, pmd_entry;
153 	bool la57;
154 	int i;
155 
156 	la57 = check_la57_support();
157 
158 	/* Is the address too large? */
159 	if (physaddr >> MAX_PHYSMEM_BITS)
160 		for (;;);
161 
162 	/*
163 	 * Compute the delta between the address I am compiled to run at
164 	 * and the address I am actually running at.
165 	 */
166 	load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
167 	RIP_REL_REF(phys_base) = load_delta;
168 
169 	/* Is the address not 2M aligned? */
170 	if (load_delta & ~PMD_MASK)
171 		for (;;);
172 
173 	/* Include the SME encryption mask in the fixup value */
174 	load_delta += sme_get_me_mask();
175 
176 	/* Fixup the physical addresses in the page table */
177 
178 	pgd = &RIP_REL_REF(early_top_pgt)->pgd;
179 	pgd[pgd_index(__START_KERNEL_map)] += load_delta;
180 
181 	if (la57) {
182 		p4d = (p4dval_t *)&RIP_REL_REF(level4_kernel_pgt);
183 		p4d[MAX_PTRS_PER_P4D - 1] += load_delta;
184 
185 		pgd[pgd_index(__START_KERNEL_map)] = (pgdval_t)p4d | _PAGE_TABLE;
186 	}
187 
188 	RIP_REL_REF(level3_kernel_pgt)[PTRS_PER_PUD - 2].pud += load_delta;
189 	RIP_REL_REF(level3_kernel_pgt)[PTRS_PER_PUD - 1].pud += load_delta;
190 
191 	for (i = FIXMAP_PMD_TOP; i > FIXMAP_PMD_TOP - FIXMAP_PMD_NUM; i--)
192 		RIP_REL_REF(level2_fixmap_pgt)[i].pmd += load_delta;
193 
194 	/*
195 	 * Set up the identity mapping for the switchover.  These
196 	 * entries should *NOT* have the global bit set!  This also
197 	 * creates a bunch of nonsense entries but that is fine --
198 	 * it avoids problems around wraparound.
199 	 */
200 
201 	pud = &early_pgts[0]->pmd;
202 	pmd = &early_pgts[1]->pmd;
203 	RIP_REL_REF(next_early_pgt) = 2;
204 
205 	pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask();
206 
207 	if (la57) {
208 		p4d = &early_pgts[RIP_REL_REF(next_early_pgt)++]->pmd;
209 
210 		i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
211 		pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
212 		pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
213 
214 		i = physaddr >> P4D_SHIFT;
215 		p4d[(i + 0) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags;
216 		p4d[(i + 1) % PTRS_PER_P4D] = (pgdval_t)pud + pgtable_flags;
217 	} else {
218 		i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
219 		pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
220 		pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
221 	}
222 
223 	i = physaddr >> PUD_SHIFT;
224 	pud[(i + 0) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags;
225 	pud[(i + 1) % PTRS_PER_PUD] = (pudval_t)pmd + pgtable_flags;
226 
227 	pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
228 	/* Filter out unsupported __PAGE_KERNEL_* bits: */
229 	pmd_entry &= RIP_REL_REF(__supported_pte_mask);
230 	pmd_entry += sme_get_me_mask();
231 	pmd_entry +=  physaddr;
232 
233 	for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
234 		int idx = i + (physaddr >> PMD_SHIFT);
235 
236 		pmd[idx % PTRS_PER_PMD] = pmd_entry + i * PMD_SIZE;
237 	}
238 
239 	/*
240 	 * Fixup the kernel text+data virtual addresses. Note that
241 	 * we might write invalid pmds, when the kernel is relocated
242 	 * cleanup_highmap() fixes this up along with the mappings
243 	 * beyond _end.
244 	 *
245 	 * Only the region occupied by the kernel image has so far
246 	 * been checked against the table of usable memory regions
247 	 * provided by the firmware, so invalidate pages outside that
248 	 * region. A page table entry that maps to a reserved area of
249 	 * memory would allow processor speculation into that area,
250 	 * and on some hardware (particularly the UV platform) even
251 	 * speculative access to some reserved areas is caught as an
252 	 * error, causing the BIOS to halt the system.
253 	 */
254 
255 	pmd = &RIP_REL_REF(level2_kernel_pgt)->pmd;
256 
257 	/* invalidate pages before the kernel image */
258 	for (i = 0; i < pmd_index((unsigned long)_text); i++)
259 		pmd[i] &= ~_PAGE_PRESENT;
260 
261 	/* fixup pages that are part of the kernel image */
262 	for (; i <= pmd_index((unsigned long)_end); i++)
263 		if (pmd[i] & _PAGE_PRESENT)
264 			pmd[i] += load_delta;
265 
266 	/* invalidate pages after the kernel image */
267 	for (; i < PTRS_PER_PMD; i++)
268 		pmd[i] &= ~_PAGE_PRESENT;
269 
270 	return sme_postprocess_startup(bp, pmd);
271 }
272 
273 /* Wipe all early page tables except for the kernel symbol map */
274 static void __init reset_early_page_tables(void)
275 {
276 	memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
277 	next_early_pgt = 0;
278 	write_cr3(__sme_pa_nodebug(early_top_pgt));
279 }
280 
281 /* Create a new PMD entry */
282 bool __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
283 {
284 	unsigned long physaddr = address - __PAGE_OFFSET;
285 	pgdval_t pgd, *pgd_p;
286 	p4dval_t p4d, *p4d_p;
287 	pudval_t pud, *pud_p;
288 	pmdval_t *pmd_p;
289 
290 	/* Invalid address or early pgt is done ?  */
291 	if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
292 		return false;
293 
294 again:
295 	pgd_p = &early_top_pgt[pgd_index(address)].pgd;
296 	pgd = *pgd_p;
297 
298 	/*
299 	 * The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
300 	 * critical -- __PAGE_OFFSET would point us back into the dynamic
301 	 * range and we might end up looping forever...
302 	 */
303 	if (!pgtable_l5_enabled())
304 		p4d_p = pgd_p;
305 	else if (pgd)
306 		p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
307 	else {
308 		if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
309 			reset_early_page_tables();
310 			goto again;
311 		}
312 
313 		p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
314 		memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
315 		*pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
316 	}
317 	p4d_p += p4d_index(address);
318 	p4d = *p4d_p;
319 
320 	if (p4d)
321 		pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
322 	else {
323 		if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
324 			reset_early_page_tables();
325 			goto again;
326 		}
327 
328 		pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
329 		memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
330 		*p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
331 	}
332 	pud_p += pud_index(address);
333 	pud = *pud_p;
334 
335 	if (pud)
336 		pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
337 	else {
338 		if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
339 			reset_early_page_tables();
340 			goto again;
341 		}
342 
343 		pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
344 		memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
345 		*pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
346 	}
347 	pmd_p[pmd_index(address)] = pmd;
348 
349 	return true;
350 }
351 
352 static bool __init early_make_pgtable(unsigned long address)
353 {
354 	unsigned long physaddr = address - __PAGE_OFFSET;
355 	pmdval_t pmd;
356 
357 	pmd = (physaddr & PMD_MASK) + early_pmd_flags;
358 
359 	return __early_make_pgtable(address, pmd);
360 }
361 
362 void __init do_early_exception(struct pt_regs *regs, int trapnr)
363 {
364 	if (trapnr == X86_TRAP_PF &&
365 	    early_make_pgtable(native_read_cr2()))
366 		return;
367 
368 	if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT) &&
369 	    trapnr == X86_TRAP_VC && handle_vc_boot_ghcb(regs))
370 		return;
371 
372 	if (trapnr == X86_TRAP_VE && tdx_early_handle_ve(regs))
373 		return;
374 
375 	early_fixup_exception(regs, trapnr);
376 }
377 
378 /* Don't add a printk in there. printk relies on the PDA which is not initialized
379    yet. */
380 void __init clear_bss(void)
381 {
382 	memset(__bss_start, 0,
383 	       (unsigned long) __bss_stop - (unsigned long) __bss_start);
384 	memset(__brk_base, 0,
385 	       (unsigned long) __brk_limit - (unsigned long) __brk_base);
386 }
387 
388 static unsigned long get_cmd_line_ptr(void)
389 {
390 	unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
391 
392 	cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
393 
394 	return cmd_line_ptr;
395 }
396 
397 static void __init copy_bootdata(char *real_mode_data)
398 {
399 	char * command_line;
400 	unsigned long cmd_line_ptr;
401 
402 	/*
403 	 * If SME is active, this will create decrypted mappings of the
404 	 * boot data in advance of the copy operations.
405 	 */
406 	sme_map_bootdata(real_mode_data);
407 
408 	memcpy(&boot_params, real_mode_data, sizeof(boot_params));
409 	sanitize_boot_params(&boot_params);
410 	cmd_line_ptr = get_cmd_line_ptr();
411 	if (cmd_line_ptr) {
412 		command_line = __va(cmd_line_ptr);
413 		memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
414 	}
415 
416 	/*
417 	 * The old boot data is no longer needed and won't be reserved,
418 	 * freeing up that memory for use by the system. If SME is active,
419 	 * we need to remove the mappings that were created so that the
420 	 * memory doesn't remain mapped as decrypted.
421 	 */
422 	sme_unmap_bootdata(real_mode_data);
423 }
424 
425 asmlinkage __visible void __init __noreturn x86_64_start_kernel(char * real_mode_data)
426 {
427 	/*
428 	 * Build-time sanity checks on the kernel image and module
429 	 * area mappings. (these are purely build-time and produce no code)
430 	 */
431 	BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
432 	BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
433 	BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
434 	BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
435 	BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
436 	BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
437 	MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
438 				(__START_KERNEL & PGDIR_MASK)));
439 	BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
440 
441 	cr4_init_shadow();
442 
443 	/* Kill off the identity-map trampoline */
444 	reset_early_page_tables();
445 
446 	clear_bss();
447 
448 	/*
449 	 * This needs to happen *before* kasan_early_init() because latter maps stuff
450 	 * into that page.
451 	 */
452 	clear_page(init_top_pgt);
453 
454 	/*
455 	 * SME support may update early_pmd_flags to include the memory
456 	 * encryption mask, so it needs to be called before anything
457 	 * that may generate a page fault.
458 	 */
459 	sme_early_init();
460 
461 	kasan_early_init();
462 
463 	/*
464 	 * Flush global TLB entries which could be left over from the trampoline page
465 	 * table.
466 	 *
467 	 * This needs to happen *after* kasan_early_init() as KASAN-enabled .configs
468 	 * instrument native_write_cr4() so KASAN must be initialized for that
469 	 * instrumentation to work.
470 	 */
471 	__native_tlb_flush_global(this_cpu_read(cpu_tlbstate.cr4));
472 
473 	idt_setup_early_handler();
474 
475 	/* Needed before cc_platform_has() can be used for TDX */
476 	tdx_early_init();
477 
478 	copy_bootdata(__va(real_mode_data));
479 
480 	/*
481 	 * Load microcode early on BSP.
482 	 */
483 	load_ucode_bsp();
484 
485 	/* set init_top_pgt kernel high mapping*/
486 	init_top_pgt[511] = early_top_pgt[511];
487 
488 	x86_64_start_reservations(real_mode_data);
489 }
490 
491 void __init __noreturn x86_64_start_reservations(char *real_mode_data)
492 {
493 	/* version is always not zero if it is copied */
494 	if (!boot_params.hdr.version)
495 		copy_bootdata(__va(real_mode_data));
496 
497 	x86_early_init_platform_quirks();
498 
499 	switch (boot_params.hdr.hardware_subarch) {
500 	case X86_SUBARCH_INTEL_MID:
501 		x86_intel_mid_early_setup();
502 		break;
503 	default:
504 		break;
505 	}
506 
507 	start_kernel();
508 }
509 
510 /*
511  * Data structures and code used for IDT setup in head_64.S. The bringup-IDT is
512  * used until the idt_table takes over. On the boot CPU this happens in
513  * x86_64_start_kernel(), on secondary CPUs in start_secondary(). In both cases
514  * this happens in the functions called from head_64.S.
515  *
516  * The idt_table can't be used that early because all the code modifying it is
517  * in idt.c and can be instrumented by tracing or KASAN, which both don't work
518  * during early CPU bringup. Also the idt_table has the runtime vectors
519  * configured which require certain CPU state to be setup already (like TSS),
520  * which also hasn't happened yet in early CPU bringup.
521  */
522 static gate_desc bringup_idt_table[NUM_EXCEPTION_VECTORS] __page_aligned_data;
523 
524 /* This may run while still in the direct mapping */
525 static void __head startup_64_load_idt(void *vc_handler)
526 {
527 	struct desc_ptr desc = {
528 		.address = (unsigned long)&RIP_REL_REF(bringup_idt_table),
529 		.size    = sizeof(bringup_idt_table) - 1,
530 	};
531 	struct idt_data data;
532 	gate_desc idt_desc;
533 
534 	/* @vc_handler is set only for a VMM Communication Exception */
535 	if (vc_handler) {
536 		init_idt_data(&data, X86_TRAP_VC, vc_handler);
537 		idt_init_desc(&idt_desc, &data);
538 		native_write_idt_entry((gate_desc *)desc.address, X86_TRAP_VC, &idt_desc);
539 	}
540 
541 	native_load_idt(&desc);
542 }
543 
544 /* This is used when running on kernel addresses */
545 void early_setup_idt(void)
546 {
547 	void *handler = NULL;
548 
549 	if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT)) {
550 		setup_ghcb();
551 		handler = vc_boot_ghcb;
552 	}
553 
554 	startup_64_load_idt(handler);
555 }
556 
557 /*
558  * Setup boot CPU state needed before kernel switches to virtual addresses.
559  */
560 void __head startup_64_setup_gdt_idt(void)
561 {
562 	struct desc_struct *gdt = (void *)(__force unsigned long)init_per_cpu_var(gdt_page.gdt);
563 	void *handler = NULL;
564 
565 	struct desc_ptr startup_gdt_descr = {
566 		.address = (unsigned long)&RIP_REL_REF(*gdt),
567 		.size    = GDT_SIZE - 1,
568 	};
569 
570 	/* Load GDT */
571 	native_load_gdt(&startup_gdt_descr);
572 
573 	/* New GDT is live - reload data segment registers */
574 	asm volatile("movl %%eax, %%ds\n"
575 		     "movl %%eax, %%ss\n"
576 		     "movl %%eax, %%es\n" : : "a"(__KERNEL_DS) : "memory");
577 
578 	if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT))
579 		handler = &RIP_REL_REF(vc_no_ghcb);
580 
581 	startup_64_load_idt(handler);
582 }
583