xref: /openbmc/linux/arch/x86/kernel/sev.c (revision 65844828)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * AMD Memory Encryption Support
4  *
5  * Copyright (C) 2019 SUSE
6  *
7  * Author: Joerg Roedel <jroedel@suse.de>
8  */
9 
10 #define pr_fmt(fmt)	"SEV: " fmt
11 
12 #include <linux/sched/debug.h>	/* For show_regs() */
13 #include <linux/percpu-defs.h>
14 #include <linux/cc_platform.h>
15 #include <linux/printk.h>
16 #include <linux/mm_types.h>
17 #include <linux/set_memory.h>
18 #include <linux/memblock.h>
19 #include <linux/kernel.h>
20 #include <linux/mm.h>
21 #include <linux/cpumask.h>
22 #include <linux/efi.h>
23 #include <linux/platform_device.h>
24 #include <linux/io.h>
25 
26 #include <asm/cpu_entry_area.h>
27 #include <asm/stacktrace.h>
28 #include <asm/sev.h>
29 #include <asm/insn-eval.h>
30 #include <asm/fpu/xcr.h>
31 #include <asm/processor.h>
32 #include <asm/realmode.h>
33 #include <asm/setup.h>
34 #include <asm/traps.h>
35 #include <asm/svm.h>
36 #include <asm/smp.h>
37 #include <asm/cpu.h>
38 #include <asm/apic.h>
39 #include <asm/cpuid.h>
40 #include <asm/cmdline.h>
41 
42 #define DR7_RESET_VALUE        0x400
43 
44 /* AP INIT values as documented in the APM2  section "Processor Initialization State" */
45 #define AP_INIT_CS_LIMIT		0xffff
46 #define AP_INIT_DS_LIMIT		0xffff
47 #define AP_INIT_LDTR_LIMIT		0xffff
48 #define AP_INIT_GDTR_LIMIT		0xffff
49 #define AP_INIT_IDTR_LIMIT		0xffff
50 #define AP_INIT_TR_LIMIT		0xffff
51 #define AP_INIT_RFLAGS_DEFAULT		0x2
52 #define AP_INIT_DR6_DEFAULT		0xffff0ff0
53 #define AP_INIT_GPAT_DEFAULT		0x0007040600070406ULL
54 #define AP_INIT_XCR0_DEFAULT		0x1
55 #define AP_INIT_X87_FTW_DEFAULT		0x5555
56 #define AP_INIT_X87_FCW_DEFAULT		0x0040
57 #define AP_INIT_CR0_DEFAULT		0x60000010
58 #define AP_INIT_MXCSR_DEFAULT		0x1f80
59 
60 /* For early boot hypervisor communication in SEV-ES enabled guests */
61 static struct ghcb boot_ghcb_page __bss_decrypted __aligned(PAGE_SIZE);
62 
63 /*
64  * Needs to be in the .data section because we need it NULL before bss is
65  * cleared
66  */
67 static struct ghcb *boot_ghcb __section(".data");
68 
69 /* Bitmap of SEV features supported by the hypervisor */
70 static u64 sev_hv_features __ro_after_init;
71 
72 /* #VC handler runtime per-CPU data */
73 struct sev_es_runtime_data {
74 	struct ghcb ghcb_page;
75 
76 	/*
77 	 * Reserve one page per CPU as backup storage for the unencrypted GHCB.
78 	 * It is needed when an NMI happens while the #VC handler uses the real
79 	 * GHCB, and the NMI handler itself is causing another #VC exception. In
80 	 * that case the GHCB content of the first handler needs to be backed up
81 	 * and restored.
82 	 */
83 	struct ghcb backup_ghcb;
84 
85 	/*
86 	 * Mark the per-cpu GHCBs as in-use to detect nested #VC exceptions.
87 	 * There is no need for it to be atomic, because nothing is written to
88 	 * the GHCB between the read and the write of ghcb_active. So it is safe
89 	 * to use it when a nested #VC exception happens before the write.
90 	 *
91 	 * This is necessary for example in the #VC->NMI->#VC case when the NMI
92 	 * happens while the first #VC handler uses the GHCB. When the NMI code
93 	 * raises a second #VC handler it might overwrite the contents of the
94 	 * GHCB written by the first handler. To avoid this the content of the
95 	 * GHCB is saved and restored when the GHCB is detected to be in use
96 	 * already.
97 	 */
98 	bool ghcb_active;
99 	bool backup_ghcb_active;
100 
101 	/*
102 	 * Cached DR7 value - write it on DR7 writes and return it on reads.
103 	 * That value will never make it to the real hardware DR7 as debugging
104 	 * is currently unsupported in SEV-ES guests.
105 	 */
106 	unsigned long dr7;
107 };
108 
109 struct ghcb_state {
110 	struct ghcb *ghcb;
111 };
112 
113 static DEFINE_PER_CPU(struct sev_es_runtime_data*, runtime_data);
114 DEFINE_STATIC_KEY_FALSE(sev_es_enable_key);
115 
116 static DEFINE_PER_CPU(struct sev_es_save_area *, sev_vmsa);
117 
118 struct sev_config {
119 	__u64 debug		: 1,
120 	      __reserved	: 63;
121 };
122 
123 static struct sev_config sev_cfg __read_mostly;
124 
125 static __always_inline bool on_vc_stack(struct pt_regs *regs)
126 {
127 	unsigned long sp = regs->sp;
128 
129 	/* User-mode RSP is not trusted */
130 	if (user_mode(regs))
131 		return false;
132 
133 	/* SYSCALL gap still has user-mode RSP */
134 	if (ip_within_syscall_gap(regs))
135 		return false;
136 
137 	return ((sp >= __this_cpu_ist_bottom_va(VC)) && (sp < __this_cpu_ist_top_va(VC)));
138 }
139 
140 /*
141  * This function handles the case when an NMI is raised in the #VC
142  * exception handler entry code, before the #VC handler has switched off
143  * its IST stack. In this case, the IST entry for #VC must be adjusted,
144  * so that any nested #VC exception will not overwrite the stack
145  * contents of the interrupted #VC handler.
146  *
147  * The IST entry is adjusted unconditionally so that it can be also be
148  * unconditionally adjusted back in __sev_es_ist_exit(). Otherwise a
149  * nested sev_es_ist_exit() call may adjust back the IST entry too
150  * early.
151  *
152  * The __sev_es_ist_enter() and __sev_es_ist_exit() functions always run
153  * on the NMI IST stack, as they are only called from NMI handling code
154  * right now.
155  */
156 void noinstr __sev_es_ist_enter(struct pt_regs *regs)
157 {
158 	unsigned long old_ist, new_ist;
159 
160 	/* Read old IST entry */
161 	new_ist = old_ist = __this_cpu_read(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC]);
162 
163 	/*
164 	 * If NMI happened while on the #VC IST stack, set the new IST
165 	 * value below regs->sp, so that the interrupted stack frame is
166 	 * not overwritten by subsequent #VC exceptions.
167 	 */
168 	if (on_vc_stack(regs))
169 		new_ist = regs->sp;
170 
171 	/*
172 	 * Reserve additional 8 bytes and store old IST value so this
173 	 * adjustment can be unrolled in __sev_es_ist_exit().
174 	 */
175 	new_ist -= sizeof(old_ist);
176 	*(unsigned long *)new_ist = old_ist;
177 
178 	/* Set new IST entry */
179 	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], new_ist);
180 }
181 
182 void noinstr __sev_es_ist_exit(void)
183 {
184 	unsigned long ist;
185 
186 	/* Read IST entry */
187 	ist = __this_cpu_read(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC]);
188 
189 	if (WARN_ON(ist == __this_cpu_ist_top_va(VC)))
190 		return;
191 
192 	/* Read back old IST entry and write it to the TSS */
193 	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *(unsigned long *)ist);
194 }
195 
196 /*
197  * Nothing shall interrupt this code path while holding the per-CPU
198  * GHCB. The backup GHCB is only for NMIs interrupting this path.
199  *
200  * Callers must disable local interrupts around it.
201  */
202 static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
203 {
204 	struct sev_es_runtime_data *data;
205 	struct ghcb *ghcb;
206 
207 	WARN_ON(!irqs_disabled());
208 
209 	data = this_cpu_read(runtime_data);
210 	ghcb = &data->ghcb_page;
211 
212 	if (unlikely(data->ghcb_active)) {
213 		/* GHCB is already in use - save its contents */
214 
215 		if (unlikely(data->backup_ghcb_active)) {
216 			/*
217 			 * Backup-GHCB is also already in use. There is no way
218 			 * to continue here so just kill the machine. To make
219 			 * panic() work, mark GHCBs inactive so that messages
220 			 * can be printed out.
221 			 */
222 			data->ghcb_active        = false;
223 			data->backup_ghcb_active = false;
224 
225 			instrumentation_begin();
226 			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
227 			instrumentation_end();
228 		}
229 
230 		/* Mark backup_ghcb active before writing to it */
231 		data->backup_ghcb_active = true;
232 
233 		state->ghcb = &data->backup_ghcb;
234 
235 		/* Backup GHCB content */
236 		*state->ghcb = *ghcb;
237 	} else {
238 		state->ghcb = NULL;
239 		data->ghcb_active = true;
240 	}
241 
242 	return ghcb;
243 }
244 
245 static inline u64 sev_es_rd_ghcb_msr(void)
246 {
247 	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
248 }
249 
250 static __always_inline void sev_es_wr_ghcb_msr(u64 val)
251 {
252 	u32 low, high;
253 
254 	low  = (u32)(val);
255 	high = (u32)(val >> 32);
256 
257 	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
258 }
259 
260 static int vc_fetch_insn_kernel(struct es_em_ctxt *ctxt,
261 				unsigned char *buffer)
262 {
263 	return copy_from_kernel_nofault(buffer, (unsigned char *)ctxt->regs->ip, MAX_INSN_SIZE);
264 }
265 
266 static enum es_result __vc_decode_user_insn(struct es_em_ctxt *ctxt)
267 {
268 	char buffer[MAX_INSN_SIZE];
269 	int insn_bytes;
270 
271 	insn_bytes = insn_fetch_from_user_inatomic(ctxt->regs, buffer);
272 	if (insn_bytes == 0) {
273 		/* Nothing could be copied */
274 		ctxt->fi.vector     = X86_TRAP_PF;
275 		ctxt->fi.error_code = X86_PF_INSTR | X86_PF_USER;
276 		ctxt->fi.cr2        = ctxt->regs->ip;
277 		return ES_EXCEPTION;
278 	} else if (insn_bytes == -EINVAL) {
279 		/* Effective RIP could not be calculated */
280 		ctxt->fi.vector     = X86_TRAP_GP;
281 		ctxt->fi.error_code = 0;
282 		ctxt->fi.cr2        = 0;
283 		return ES_EXCEPTION;
284 	}
285 
286 	if (!insn_decode_from_regs(&ctxt->insn, ctxt->regs, buffer, insn_bytes))
287 		return ES_DECODE_FAILED;
288 
289 	if (ctxt->insn.immediate.got)
290 		return ES_OK;
291 	else
292 		return ES_DECODE_FAILED;
293 }
294 
295 static enum es_result __vc_decode_kern_insn(struct es_em_ctxt *ctxt)
296 {
297 	char buffer[MAX_INSN_SIZE];
298 	int res, ret;
299 
300 	res = vc_fetch_insn_kernel(ctxt, buffer);
301 	if (res) {
302 		ctxt->fi.vector     = X86_TRAP_PF;
303 		ctxt->fi.error_code = X86_PF_INSTR;
304 		ctxt->fi.cr2        = ctxt->regs->ip;
305 		return ES_EXCEPTION;
306 	}
307 
308 	ret = insn_decode(&ctxt->insn, buffer, MAX_INSN_SIZE, INSN_MODE_64);
309 	if (ret < 0)
310 		return ES_DECODE_FAILED;
311 	else
312 		return ES_OK;
313 }
314 
315 static enum es_result vc_decode_insn(struct es_em_ctxt *ctxt)
316 {
317 	if (user_mode(ctxt->regs))
318 		return __vc_decode_user_insn(ctxt);
319 	else
320 		return __vc_decode_kern_insn(ctxt);
321 }
322 
323 static enum es_result vc_write_mem(struct es_em_ctxt *ctxt,
324 				   char *dst, char *buf, size_t size)
325 {
326 	unsigned long error_code = X86_PF_PROT | X86_PF_WRITE;
327 
328 	/*
329 	 * This function uses __put_user() independent of whether kernel or user
330 	 * memory is accessed. This works fine because __put_user() does no
331 	 * sanity checks of the pointer being accessed. All that it does is
332 	 * to report when the access failed.
333 	 *
334 	 * Also, this function runs in atomic context, so __put_user() is not
335 	 * allowed to sleep. The page-fault handler detects that it is running
336 	 * in atomic context and will not try to take mmap_sem and handle the
337 	 * fault, so additional pagefault_enable()/disable() calls are not
338 	 * needed.
339 	 *
340 	 * The access can't be done via copy_to_user() here because
341 	 * vc_write_mem() must not use string instructions to access unsafe
342 	 * memory. The reason is that MOVS is emulated by the #VC handler by
343 	 * splitting the move up into a read and a write and taking a nested #VC
344 	 * exception on whatever of them is the MMIO access. Using string
345 	 * instructions here would cause infinite nesting.
346 	 */
347 	switch (size) {
348 	case 1: {
349 		u8 d1;
350 		u8 __user *target = (u8 __user *)dst;
351 
352 		memcpy(&d1, buf, 1);
353 		if (__put_user(d1, target))
354 			goto fault;
355 		break;
356 	}
357 	case 2: {
358 		u16 d2;
359 		u16 __user *target = (u16 __user *)dst;
360 
361 		memcpy(&d2, buf, 2);
362 		if (__put_user(d2, target))
363 			goto fault;
364 		break;
365 	}
366 	case 4: {
367 		u32 d4;
368 		u32 __user *target = (u32 __user *)dst;
369 
370 		memcpy(&d4, buf, 4);
371 		if (__put_user(d4, target))
372 			goto fault;
373 		break;
374 	}
375 	case 8: {
376 		u64 d8;
377 		u64 __user *target = (u64 __user *)dst;
378 
379 		memcpy(&d8, buf, 8);
380 		if (__put_user(d8, target))
381 			goto fault;
382 		break;
383 	}
384 	default:
385 		WARN_ONCE(1, "%s: Invalid size: %zu\n", __func__, size);
386 		return ES_UNSUPPORTED;
387 	}
388 
389 	return ES_OK;
390 
391 fault:
392 	if (user_mode(ctxt->regs))
393 		error_code |= X86_PF_USER;
394 
395 	ctxt->fi.vector = X86_TRAP_PF;
396 	ctxt->fi.error_code = error_code;
397 	ctxt->fi.cr2 = (unsigned long)dst;
398 
399 	return ES_EXCEPTION;
400 }
401 
402 static enum es_result vc_read_mem(struct es_em_ctxt *ctxt,
403 				  char *src, char *buf, size_t size)
404 {
405 	unsigned long error_code = X86_PF_PROT;
406 
407 	/*
408 	 * This function uses __get_user() independent of whether kernel or user
409 	 * memory is accessed. This works fine because __get_user() does no
410 	 * sanity checks of the pointer being accessed. All that it does is
411 	 * to report when the access failed.
412 	 *
413 	 * Also, this function runs in atomic context, so __get_user() is not
414 	 * allowed to sleep. The page-fault handler detects that it is running
415 	 * in atomic context and will not try to take mmap_sem and handle the
416 	 * fault, so additional pagefault_enable()/disable() calls are not
417 	 * needed.
418 	 *
419 	 * The access can't be done via copy_from_user() here because
420 	 * vc_read_mem() must not use string instructions to access unsafe
421 	 * memory. The reason is that MOVS is emulated by the #VC handler by
422 	 * splitting the move up into a read and a write and taking a nested #VC
423 	 * exception on whatever of them is the MMIO access. Using string
424 	 * instructions here would cause infinite nesting.
425 	 */
426 	switch (size) {
427 	case 1: {
428 		u8 d1;
429 		u8 __user *s = (u8 __user *)src;
430 
431 		if (__get_user(d1, s))
432 			goto fault;
433 		memcpy(buf, &d1, 1);
434 		break;
435 	}
436 	case 2: {
437 		u16 d2;
438 		u16 __user *s = (u16 __user *)src;
439 
440 		if (__get_user(d2, s))
441 			goto fault;
442 		memcpy(buf, &d2, 2);
443 		break;
444 	}
445 	case 4: {
446 		u32 d4;
447 		u32 __user *s = (u32 __user *)src;
448 
449 		if (__get_user(d4, s))
450 			goto fault;
451 		memcpy(buf, &d4, 4);
452 		break;
453 	}
454 	case 8: {
455 		u64 d8;
456 		u64 __user *s = (u64 __user *)src;
457 		if (__get_user(d8, s))
458 			goto fault;
459 		memcpy(buf, &d8, 8);
460 		break;
461 	}
462 	default:
463 		WARN_ONCE(1, "%s: Invalid size: %zu\n", __func__, size);
464 		return ES_UNSUPPORTED;
465 	}
466 
467 	return ES_OK;
468 
469 fault:
470 	if (user_mode(ctxt->regs))
471 		error_code |= X86_PF_USER;
472 
473 	ctxt->fi.vector = X86_TRAP_PF;
474 	ctxt->fi.error_code = error_code;
475 	ctxt->fi.cr2 = (unsigned long)src;
476 
477 	return ES_EXCEPTION;
478 }
479 
480 static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt *ctxt,
481 					   unsigned long vaddr, phys_addr_t *paddr)
482 {
483 	unsigned long va = (unsigned long)vaddr;
484 	unsigned int level;
485 	phys_addr_t pa;
486 	pgd_t *pgd;
487 	pte_t *pte;
488 
489 	pgd = __va(read_cr3_pa());
490 	pgd = &pgd[pgd_index(va)];
491 	pte = lookup_address_in_pgd(pgd, va, &level);
492 	if (!pte) {
493 		ctxt->fi.vector     = X86_TRAP_PF;
494 		ctxt->fi.cr2        = vaddr;
495 		ctxt->fi.error_code = 0;
496 
497 		if (user_mode(ctxt->regs))
498 			ctxt->fi.error_code |= X86_PF_USER;
499 
500 		return ES_EXCEPTION;
501 	}
502 
503 	if (WARN_ON_ONCE(pte_val(*pte) & _PAGE_ENC))
504 		/* Emulated MMIO to/from encrypted memory not supported */
505 		return ES_UNSUPPORTED;
506 
507 	pa = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
508 	pa |= va & ~page_level_mask(level);
509 
510 	*paddr = pa;
511 
512 	return ES_OK;
513 }
514 
515 /* Include code shared with pre-decompression boot stage */
516 #include "sev-shared.c"
517 
518 static noinstr void __sev_put_ghcb(struct ghcb_state *state)
519 {
520 	struct sev_es_runtime_data *data;
521 	struct ghcb *ghcb;
522 
523 	WARN_ON(!irqs_disabled());
524 
525 	data = this_cpu_read(runtime_data);
526 	ghcb = &data->ghcb_page;
527 
528 	if (state->ghcb) {
529 		/* Restore GHCB from Backup */
530 		*ghcb = *state->ghcb;
531 		data->backup_ghcb_active = false;
532 		state->ghcb = NULL;
533 	} else {
534 		/*
535 		 * Invalidate the GHCB so a VMGEXIT instruction issued
536 		 * from userspace won't appear to be valid.
537 		 */
538 		vc_ghcb_invalidate(ghcb);
539 		data->ghcb_active = false;
540 	}
541 }
542 
543 void noinstr __sev_es_nmi_complete(void)
544 {
545 	struct ghcb_state state;
546 	struct ghcb *ghcb;
547 
548 	ghcb = __sev_get_ghcb(&state);
549 
550 	vc_ghcb_invalidate(ghcb);
551 	ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_NMI_COMPLETE);
552 	ghcb_set_sw_exit_info_1(ghcb, 0);
553 	ghcb_set_sw_exit_info_2(ghcb, 0);
554 
555 	sev_es_wr_ghcb_msr(__pa_nodebug(ghcb));
556 	VMGEXIT();
557 
558 	__sev_put_ghcb(&state);
559 }
560 
561 static u64 __init get_secrets_page(void)
562 {
563 	u64 pa_data = boot_params.cc_blob_address;
564 	struct cc_blob_sev_info info;
565 	void *map;
566 
567 	/*
568 	 * The CC blob contains the address of the secrets page, check if the
569 	 * blob is present.
570 	 */
571 	if (!pa_data)
572 		return 0;
573 
574 	map = early_memremap(pa_data, sizeof(info));
575 	if (!map) {
576 		pr_err("Unable to locate SNP secrets page: failed to map the Confidential Computing blob.\n");
577 		return 0;
578 	}
579 	memcpy(&info, map, sizeof(info));
580 	early_memunmap(map, sizeof(info));
581 
582 	/* smoke-test the secrets page passed */
583 	if (!info.secrets_phys || info.secrets_len != PAGE_SIZE)
584 		return 0;
585 
586 	return info.secrets_phys;
587 }
588 
589 static u64 __init get_snp_jump_table_addr(void)
590 {
591 	struct snp_secrets_page_layout *layout;
592 	void __iomem *mem;
593 	u64 pa, addr;
594 
595 	pa = get_secrets_page();
596 	if (!pa)
597 		return 0;
598 
599 	mem = ioremap_encrypted(pa, PAGE_SIZE);
600 	if (!mem) {
601 		pr_err("Unable to locate AP jump table address: failed to map the SNP secrets page.\n");
602 		return 0;
603 	}
604 
605 	layout = (__force struct snp_secrets_page_layout *)mem;
606 
607 	addr = layout->os_area.ap_jump_table_pa;
608 	iounmap(mem);
609 
610 	return addr;
611 }
612 
613 static u64 __init get_jump_table_addr(void)
614 {
615 	struct ghcb_state state;
616 	unsigned long flags;
617 	struct ghcb *ghcb;
618 	u64 ret = 0;
619 
620 	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
621 		return get_snp_jump_table_addr();
622 
623 	local_irq_save(flags);
624 
625 	ghcb = __sev_get_ghcb(&state);
626 
627 	vc_ghcb_invalidate(ghcb);
628 	ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_JUMP_TABLE);
629 	ghcb_set_sw_exit_info_1(ghcb, SVM_VMGEXIT_GET_AP_JUMP_TABLE);
630 	ghcb_set_sw_exit_info_2(ghcb, 0);
631 
632 	sev_es_wr_ghcb_msr(__pa(ghcb));
633 	VMGEXIT();
634 
635 	if (ghcb_sw_exit_info_1_is_valid(ghcb) &&
636 	    ghcb_sw_exit_info_2_is_valid(ghcb))
637 		ret = ghcb->save.sw_exit_info_2;
638 
639 	__sev_put_ghcb(&state);
640 
641 	local_irq_restore(flags);
642 
643 	return ret;
644 }
645 
646 static void pvalidate_pages(unsigned long vaddr, unsigned int npages, bool validate)
647 {
648 	unsigned long vaddr_end;
649 	int rc;
650 
651 	vaddr = vaddr & PAGE_MASK;
652 	vaddr_end = vaddr + (npages << PAGE_SHIFT);
653 
654 	while (vaddr < vaddr_end) {
655 		rc = pvalidate(vaddr, RMP_PG_SIZE_4K, validate);
656 		if (WARN(rc, "Failed to validate address 0x%lx ret %d", vaddr, rc))
657 			sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PVALIDATE);
658 
659 		vaddr = vaddr + PAGE_SIZE;
660 	}
661 }
662 
663 static void __init early_set_pages_state(unsigned long paddr, unsigned int npages, enum psc_op op)
664 {
665 	unsigned long paddr_end;
666 	u64 val;
667 
668 	paddr = paddr & PAGE_MASK;
669 	paddr_end = paddr + (npages << PAGE_SHIFT);
670 
671 	while (paddr < paddr_end) {
672 		/*
673 		 * Use the MSR protocol because this function can be called before
674 		 * the GHCB is established.
675 		 */
676 		sev_es_wr_ghcb_msr(GHCB_MSR_PSC_REQ_GFN(paddr >> PAGE_SHIFT, op));
677 		VMGEXIT();
678 
679 		val = sev_es_rd_ghcb_msr();
680 
681 		if (WARN(GHCB_RESP_CODE(val) != GHCB_MSR_PSC_RESP,
682 			 "Wrong PSC response code: 0x%x\n",
683 			 (unsigned int)GHCB_RESP_CODE(val)))
684 			goto e_term;
685 
686 		if (WARN(GHCB_MSR_PSC_RESP_VAL(val),
687 			 "Failed to change page state to '%s' paddr 0x%lx error 0x%llx\n",
688 			 op == SNP_PAGE_STATE_PRIVATE ? "private" : "shared",
689 			 paddr, GHCB_MSR_PSC_RESP_VAL(val)))
690 			goto e_term;
691 
692 		paddr = paddr + PAGE_SIZE;
693 	}
694 
695 	return;
696 
697 e_term:
698 	sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PSC);
699 }
700 
701 void __init early_snp_set_memory_private(unsigned long vaddr, unsigned long paddr,
702 					 unsigned int npages)
703 {
704 	/*
705 	 * This can be invoked in early boot while running identity mapped, so
706 	 * use an open coded check for SNP instead of using cc_platform_has().
707 	 * This eliminates worries about jump tables or checking boot_cpu_data
708 	 * in the cc_platform_has() function.
709 	 */
710 	if (!(sev_status & MSR_AMD64_SEV_SNP_ENABLED))
711 		return;
712 
713 	 /*
714 	  * Ask the hypervisor to mark the memory pages as private in the RMP
715 	  * table.
716 	  */
717 	early_set_pages_state(paddr, npages, SNP_PAGE_STATE_PRIVATE);
718 
719 	/* Validate the memory pages after they've been added in the RMP table. */
720 	pvalidate_pages(vaddr, npages, true);
721 }
722 
723 void __init early_snp_set_memory_shared(unsigned long vaddr, unsigned long paddr,
724 					unsigned int npages)
725 {
726 	/*
727 	 * This can be invoked in early boot while running identity mapped, so
728 	 * use an open coded check for SNP instead of using cc_platform_has().
729 	 * This eliminates worries about jump tables or checking boot_cpu_data
730 	 * in the cc_platform_has() function.
731 	 */
732 	if (!(sev_status & MSR_AMD64_SEV_SNP_ENABLED))
733 		return;
734 
735 	/* Invalidate the memory pages before they are marked shared in the RMP table. */
736 	pvalidate_pages(vaddr, npages, false);
737 
738 	 /* Ask hypervisor to mark the memory pages shared in the RMP table. */
739 	early_set_pages_state(paddr, npages, SNP_PAGE_STATE_SHARED);
740 }
741 
742 void __init snp_prep_memory(unsigned long paddr, unsigned int sz, enum psc_op op)
743 {
744 	unsigned long vaddr, npages;
745 
746 	vaddr = (unsigned long)__va(paddr);
747 	npages = PAGE_ALIGN(sz) >> PAGE_SHIFT;
748 
749 	if (op == SNP_PAGE_STATE_PRIVATE)
750 		early_snp_set_memory_private(vaddr, paddr, npages);
751 	else if (op == SNP_PAGE_STATE_SHARED)
752 		early_snp_set_memory_shared(vaddr, paddr, npages);
753 	else
754 		WARN(1, "invalid memory op %d\n", op);
755 }
756 
757 static int vmgexit_psc(struct snp_psc_desc *desc)
758 {
759 	int cur_entry, end_entry, ret = 0;
760 	struct snp_psc_desc *data;
761 	struct ghcb_state state;
762 	struct es_em_ctxt ctxt;
763 	unsigned long flags;
764 	struct ghcb *ghcb;
765 
766 	/*
767 	 * __sev_get_ghcb() needs to run with IRQs disabled because it is using
768 	 * a per-CPU GHCB.
769 	 */
770 	local_irq_save(flags);
771 
772 	ghcb = __sev_get_ghcb(&state);
773 	if (!ghcb) {
774 		ret = 1;
775 		goto out_unlock;
776 	}
777 
778 	/* Copy the input desc into GHCB shared buffer */
779 	data = (struct snp_psc_desc *)ghcb->shared_buffer;
780 	memcpy(ghcb->shared_buffer, desc, min_t(int, GHCB_SHARED_BUF_SIZE, sizeof(*desc)));
781 
782 	/*
783 	 * As per the GHCB specification, the hypervisor can resume the guest
784 	 * before processing all the entries. Check whether all the entries
785 	 * are processed. If not, then keep retrying. Note, the hypervisor
786 	 * will update the data memory directly to indicate the status, so
787 	 * reference the data->hdr everywhere.
788 	 *
789 	 * The strategy here is to wait for the hypervisor to change the page
790 	 * state in the RMP table before guest accesses the memory pages. If the
791 	 * page state change was not successful, then later memory access will
792 	 * result in a crash.
793 	 */
794 	cur_entry = data->hdr.cur_entry;
795 	end_entry = data->hdr.end_entry;
796 
797 	while (data->hdr.cur_entry <= data->hdr.end_entry) {
798 		ghcb_set_sw_scratch(ghcb, (u64)__pa(data));
799 
800 		/* This will advance the shared buffer data points to. */
801 		ret = sev_es_ghcb_hv_call(ghcb, &ctxt, SVM_VMGEXIT_PSC, 0, 0);
802 
803 		/*
804 		 * Page State Change VMGEXIT can pass error code through
805 		 * exit_info_2.
806 		 */
807 		if (WARN(ret || ghcb->save.sw_exit_info_2,
808 			 "SNP: PSC failed ret=%d exit_info_2=%llx\n",
809 			 ret, ghcb->save.sw_exit_info_2)) {
810 			ret = 1;
811 			goto out;
812 		}
813 
814 		/* Verify that reserved bit is not set */
815 		if (WARN(data->hdr.reserved, "Reserved bit is set in the PSC header\n")) {
816 			ret = 1;
817 			goto out;
818 		}
819 
820 		/*
821 		 * Sanity check that entry processing is not going backwards.
822 		 * This will happen only if hypervisor is tricking us.
823 		 */
824 		if (WARN(data->hdr.end_entry > end_entry || cur_entry > data->hdr.cur_entry,
825 "SNP: PSC processing going backward, end_entry %d (got %d) cur_entry %d (got %d)\n",
826 			 end_entry, data->hdr.end_entry, cur_entry, data->hdr.cur_entry)) {
827 			ret = 1;
828 			goto out;
829 		}
830 	}
831 
832 out:
833 	__sev_put_ghcb(&state);
834 
835 out_unlock:
836 	local_irq_restore(flags);
837 
838 	return ret;
839 }
840 
841 static void __set_pages_state(struct snp_psc_desc *data, unsigned long vaddr,
842 			      unsigned long vaddr_end, int op)
843 {
844 	struct psc_hdr *hdr;
845 	struct psc_entry *e;
846 	unsigned long pfn;
847 	int i;
848 
849 	hdr = &data->hdr;
850 	e = data->entries;
851 
852 	memset(data, 0, sizeof(*data));
853 	i = 0;
854 
855 	while (vaddr < vaddr_end) {
856 		if (is_vmalloc_addr((void *)vaddr))
857 			pfn = vmalloc_to_pfn((void *)vaddr);
858 		else
859 			pfn = __pa(vaddr) >> PAGE_SHIFT;
860 
861 		e->gfn = pfn;
862 		e->operation = op;
863 		hdr->end_entry = i;
864 
865 		/*
866 		 * Current SNP implementation doesn't keep track of the RMP page
867 		 * size so use 4K for simplicity.
868 		 */
869 		e->pagesize = RMP_PG_SIZE_4K;
870 
871 		vaddr = vaddr + PAGE_SIZE;
872 		e++;
873 		i++;
874 	}
875 
876 	if (vmgexit_psc(data))
877 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PSC);
878 }
879 
880 static void set_pages_state(unsigned long vaddr, unsigned int npages, int op)
881 {
882 	unsigned long vaddr_end, next_vaddr;
883 	struct snp_psc_desc *desc;
884 
885 	desc = kmalloc(sizeof(*desc), GFP_KERNEL_ACCOUNT);
886 	if (!desc)
887 		panic("SNP: failed to allocate memory for PSC descriptor\n");
888 
889 	vaddr = vaddr & PAGE_MASK;
890 	vaddr_end = vaddr + (npages << PAGE_SHIFT);
891 
892 	while (vaddr < vaddr_end) {
893 		/* Calculate the last vaddr that fits in one struct snp_psc_desc. */
894 		next_vaddr = min_t(unsigned long, vaddr_end,
895 				   (VMGEXIT_PSC_MAX_ENTRY * PAGE_SIZE) + vaddr);
896 
897 		__set_pages_state(desc, vaddr, next_vaddr, op);
898 
899 		vaddr = next_vaddr;
900 	}
901 
902 	kfree(desc);
903 }
904 
905 void snp_set_memory_shared(unsigned long vaddr, unsigned int npages)
906 {
907 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
908 		return;
909 
910 	pvalidate_pages(vaddr, npages, false);
911 
912 	set_pages_state(vaddr, npages, SNP_PAGE_STATE_SHARED);
913 }
914 
915 void snp_set_memory_private(unsigned long vaddr, unsigned int npages)
916 {
917 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
918 		return;
919 
920 	set_pages_state(vaddr, npages, SNP_PAGE_STATE_PRIVATE);
921 
922 	pvalidate_pages(vaddr, npages, true);
923 }
924 
925 static int snp_set_vmsa(void *va, bool vmsa)
926 {
927 	u64 attrs;
928 
929 	/*
930 	 * Running at VMPL0 allows the kernel to change the VMSA bit for a page
931 	 * using the RMPADJUST instruction. However, for the instruction to
932 	 * succeed it must target the permissions of a lesser privileged
933 	 * (higher numbered) VMPL level, so use VMPL1 (refer to the RMPADJUST
934 	 * instruction in the AMD64 APM Volume 3).
935 	 */
936 	attrs = 1;
937 	if (vmsa)
938 		attrs |= RMPADJUST_VMSA_PAGE_BIT;
939 
940 	return rmpadjust((unsigned long)va, RMP_PG_SIZE_4K, attrs);
941 }
942 
943 #define __ATTR_BASE		(SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK)
944 #define INIT_CS_ATTRIBS		(__ATTR_BASE | SVM_SELECTOR_READ_MASK | SVM_SELECTOR_CODE_MASK)
945 #define INIT_DS_ATTRIBS		(__ATTR_BASE | SVM_SELECTOR_WRITE_MASK)
946 
947 #define INIT_LDTR_ATTRIBS	(SVM_SELECTOR_P_MASK | 2)
948 #define INIT_TR_ATTRIBS		(SVM_SELECTOR_P_MASK | 3)
949 
950 static void *snp_alloc_vmsa_page(void)
951 {
952 	struct page *p;
953 
954 	/*
955 	 * Allocate VMSA page to work around the SNP erratum where the CPU will
956 	 * incorrectly signal an RMP violation #PF if a large page (2MB or 1GB)
957 	 * collides with the RMP entry of VMSA page. The recommended workaround
958 	 * is to not use a large page.
959 	 *
960 	 * Allocate an 8k page which is also 8k-aligned.
961 	 */
962 	p = alloc_pages(GFP_KERNEL_ACCOUNT | __GFP_ZERO, 1);
963 	if (!p)
964 		return NULL;
965 
966 	split_page(p, 1);
967 
968 	/* Free the first 4k. This page may be 2M/1G aligned and cannot be used. */
969 	__free_page(p);
970 
971 	return page_address(p + 1);
972 }
973 
974 static void snp_cleanup_vmsa(struct sev_es_save_area *vmsa)
975 {
976 	int err;
977 
978 	err = snp_set_vmsa(vmsa, false);
979 	if (err)
980 		pr_err("clear VMSA page failed (%u), leaking page\n", err);
981 	else
982 		free_page((unsigned long)vmsa);
983 }
984 
985 static int wakeup_cpu_via_vmgexit(int apic_id, unsigned long start_ip)
986 {
987 	struct sev_es_save_area *cur_vmsa, *vmsa;
988 	struct ghcb_state state;
989 	unsigned long flags;
990 	struct ghcb *ghcb;
991 	u8 sipi_vector;
992 	int cpu, ret;
993 	u64 cr4;
994 
995 	/*
996 	 * The hypervisor SNP feature support check has happened earlier, just check
997 	 * the AP_CREATION one here.
998 	 */
999 	if (!(sev_hv_features & GHCB_HV_FT_SNP_AP_CREATION))
1000 		return -EOPNOTSUPP;
1001 
1002 	/*
1003 	 * Verify the desired start IP against the known trampoline start IP
1004 	 * to catch any future new trampolines that may be introduced that
1005 	 * would require a new protected guest entry point.
1006 	 */
1007 	if (WARN_ONCE(start_ip != real_mode_header->trampoline_start,
1008 		      "Unsupported SNP start_ip: %lx\n", start_ip))
1009 		return -EINVAL;
1010 
1011 	/* Override start_ip with known protected guest start IP */
1012 	start_ip = real_mode_header->sev_es_trampoline_start;
1013 
1014 	/* Find the logical CPU for the APIC ID */
1015 	for_each_present_cpu(cpu) {
1016 		if (arch_match_cpu_phys_id(cpu, apic_id))
1017 			break;
1018 	}
1019 	if (cpu >= nr_cpu_ids)
1020 		return -EINVAL;
1021 
1022 	cur_vmsa = per_cpu(sev_vmsa, cpu);
1023 
1024 	/*
1025 	 * A new VMSA is created each time because there is no guarantee that
1026 	 * the current VMSA is the kernels or that the vCPU is not running. If
1027 	 * an attempt was done to use the current VMSA with a running vCPU, a
1028 	 * #VMEXIT of that vCPU would wipe out all of the settings being done
1029 	 * here.
1030 	 */
1031 	vmsa = (struct sev_es_save_area *)snp_alloc_vmsa_page();
1032 	if (!vmsa)
1033 		return -ENOMEM;
1034 
1035 	/* CR4 should maintain the MCE value */
1036 	cr4 = native_read_cr4() & X86_CR4_MCE;
1037 
1038 	/* Set the CS value based on the start_ip converted to a SIPI vector */
1039 	sipi_vector		= (start_ip >> 12);
1040 	vmsa->cs.base		= sipi_vector << 12;
1041 	vmsa->cs.limit		= AP_INIT_CS_LIMIT;
1042 	vmsa->cs.attrib		= INIT_CS_ATTRIBS;
1043 	vmsa->cs.selector	= sipi_vector << 8;
1044 
1045 	/* Set the RIP value based on start_ip */
1046 	vmsa->rip		= start_ip & 0xfff;
1047 
1048 	/* Set AP INIT defaults as documented in the APM */
1049 	vmsa->ds.limit		= AP_INIT_DS_LIMIT;
1050 	vmsa->ds.attrib		= INIT_DS_ATTRIBS;
1051 	vmsa->es		= vmsa->ds;
1052 	vmsa->fs		= vmsa->ds;
1053 	vmsa->gs		= vmsa->ds;
1054 	vmsa->ss		= vmsa->ds;
1055 
1056 	vmsa->gdtr.limit	= AP_INIT_GDTR_LIMIT;
1057 	vmsa->ldtr.limit	= AP_INIT_LDTR_LIMIT;
1058 	vmsa->ldtr.attrib	= INIT_LDTR_ATTRIBS;
1059 	vmsa->idtr.limit	= AP_INIT_IDTR_LIMIT;
1060 	vmsa->tr.limit		= AP_INIT_TR_LIMIT;
1061 	vmsa->tr.attrib		= INIT_TR_ATTRIBS;
1062 
1063 	vmsa->cr4		= cr4;
1064 	vmsa->cr0		= AP_INIT_CR0_DEFAULT;
1065 	vmsa->dr7		= DR7_RESET_VALUE;
1066 	vmsa->dr6		= AP_INIT_DR6_DEFAULT;
1067 	vmsa->rflags		= AP_INIT_RFLAGS_DEFAULT;
1068 	vmsa->g_pat		= AP_INIT_GPAT_DEFAULT;
1069 	vmsa->xcr0		= AP_INIT_XCR0_DEFAULT;
1070 	vmsa->mxcsr		= AP_INIT_MXCSR_DEFAULT;
1071 	vmsa->x87_ftw		= AP_INIT_X87_FTW_DEFAULT;
1072 	vmsa->x87_fcw		= AP_INIT_X87_FCW_DEFAULT;
1073 
1074 	/* SVME must be set. */
1075 	vmsa->efer		= EFER_SVME;
1076 
1077 	/*
1078 	 * Set the SNP-specific fields for this VMSA:
1079 	 *   VMPL level
1080 	 *   SEV_FEATURES (matches the SEV STATUS MSR right shifted 2 bits)
1081 	 */
1082 	vmsa->vmpl		= 0;
1083 	vmsa->sev_features	= sev_status >> 2;
1084 
1085 	/* Switch the page over to a VMSA page now that it is initialized */
1086 	ret = snp_set_vmsa(vmsa, true);
1087 	if (ret) {
1088 		pr_err("set VMSA page failed (%u)\n", ret);
1089 		free_page((unsigned long)vmsa);
1090 
1091 		return -EINVAL;
1092 	}
1093 
1094 	/* Issue VMGEXIT AP Creation NAE event */
1095 	local_irq_save(flags);
1096 
1097 	ghcb = __sev_get_ghcb(&state);
1098 
1099 	vc_ghcb_invalidate(ghcb);
1100 	ghcb_set_rax(ghcb, vmsa->sev_features);
1101 	ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_CREATION);
1102 	ghcb_set_sw_exit_info_1(ghcb, ((u64)apic_id << 32) | SVM_VMGEXIT_AP_CREATE);
1103 	ghcb_set_sw_exit_info_2(ghcb, __pa(vmsa));
1104 
1105 	sev_es_wr_ghcb_msr(__pa(ghcb));
1106 	VMGEXIT();
1107 
1108 	if (!ghcb_sw_exit_info_1_is_valid(ghcb) ||
1109 	    lower_32_bits(ghcb->save.sw_exit_info_1)) {
1110 		pr_err("SNP AP Creation error\n");
1111 		ret = -EINVAL;
1112 	}
1113 
1114 	__sev_put_ghcb(&state);
1115 
1116 	local_irq_restore(flags);
1117 
1118 	/* Perform cleanup if there was an error */
1119 	if (ret) {
1120 		snp_cleanup_vmsa(vmsa);
1121 		vmsa = NULL;
1122 	}
1123 
1124 	/* Free up any previous VMSA page */
1125 	if (cur_vmsa)
1126 		snp_cleanup_vmsa(cur_vmsa);
1127 
1128 	/* Record the current VMSA page */
1129 	per_cpu(sev_vmsa, cpu) = vmsa;
1130 
1131 	return ret;
1132 }
1133 
1134 void snp_set_wakeup_secondary_cpu(void)
1135 {
1136 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
1137 		return;
1138 
1139 	/*
1140 	 * Always set this override if SNP is enabled. This makes it the
1141 	 * required method to start APs under SNP. If the hypervisor does
1142 	 * not support AP creation, then no APs will be started.
1143 	 */
1144 	apic->wakeup_secondary_cpu = wakeup_cpu_via_vmgexit;
1145 }
1146 
1147 int __init sev_es_setup_ap_jump_table(struct real_mode_header *rmh)
1148 {
1149 	u16 startup_cs, startup_ip;
1150 	phys_addr_t jump_table_pa;
1151 	u64 jump_table_addr;
1152 	u16 __iomem *jump_table;
1153 
1154 	jump_table_addr = get_jump_table_addr();
1155 
1156 	/* On UP guests there is no jump table so this is not a failure */
1157 	if (!jump_table_addr)
1158 		return 0;
1159 
1160 	/* Check if AP Jump Table is page-aligned */
1161 	if (jump_table_addr & ~PAGE_MASK)
1162 		return -EINVAL;
1163 
1164 	jump_table_pa = jump_table_addr & PAGE_MASK;
1165 
1166 	startup_cs = (u16)(rmh->trampoline_start >> 4);
1167 	startup_ip = (u16)(rmh->sev_es_trampoline_start -
1168 			   rmh->trampoline_start);
1169 
1170 	jump_table = ioremap_encrypted(jump_table_pa, PAGE_SIZE);
1171 	if (!jump_table)
1172 		return -EIO;
1173 
1174 	writew(startup_ip, &jump_table[0]);
1175 	writew(startup_cs, &jump_table[1]);
1176 
1177 	iounmap(jump_table);
1178 
1179 	return 0;
1180 }
1181 
1182 /*
1183  * This is needed by the OVMF UEFI firmware which will use whatever it finds in
1184  * the GHCB MSR as its GHCB to talk to the hypervisor. So make sure the per-cpu
1185  * runtime GHCBs used by the kernel are also mapped in the EFI page-table.
1186  */
1187 int __init sev_es_efi_map_ghcbs(pgd_t *pgd)
1188 {
1189 	struct sev_es_runtime_data *data;
1190 	unsigned long address, pflags;
1191 	int cpu;
1192 	u64 pfn;
1193 
1194 	if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
1195 		return 0;
1196 
1197 	pflags = _PAGE_NX | _PAGE_RW;
1198 
1199 	for_each_possible_cpu(cpu) {
1200 		data = per_cpu(runtime_data, cpu);
1201 
1202 		address = __pa(&data->ghcb_page);
1203 		pfn = address >> PAGE_SHIFT;
1204 
1205 		if (kernel_map_pages_in_pgd(pgd, pfn, address, 1, pflags))
1206 			return 1;
1207 	}
1208 
1209 	return 0;
1210 }
1211 
1212 static enum es_result vc_handle_msr(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
1213 {
1214 	struct pt_regs *regs = ctxt->regs;
1215 	enum es_result ret;
1216 	u64 exit_info_1;
1217 
1218 	/* Is it a WRMSR? */
1219 	exit_info_1 = (ctxt->insn.opcode.bytes[1] == 0x30) ? 1 : 0;
1220 
1221 	ghcb_set_rcx(ghcb, regs->cx);
1222 	if (exit_info_1) {
1223 		ghcb_set_rax(ghcb, regs->ax);
1224 		ghcb_set_rdx(ghcb, regs->dx);
1225 	}
1226 
1227 	ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_MSR, exit_info_1, 0);
1228 
1229 	if ((ret == ES_OK) && (!exit_info_1)) {
1230 		regs->ax = ghcb->save.rax;
1231 		regs->dx = ghcb->save.rdx;
1232 	}
1233 
1234 	return ret;
1235 }
1236 
1237 static void snp_register_per_cpu_ghcb(void)
1238 {
1239 	struct sev_es_runtime_data *data;
1240 	struct ghcb *ghcb;
1241 
1242 	data = this_cpu_read(runtime_data);
1243 	ghcb = &data->ghcb_page;
1244 
1245 	snp_register_ghcb_early(__pa(ghcb));
1246 }
1247 
1248 void setup_ghcb(void)
1249 {
1250 	if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
1251 		return;
1252 
1253 	/* First make sure the hypervisor talks a supported protocol. */
1254 	if (!sev_es_negotiate_protocol())
1255 		sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
1256 
1257 	/*
1258 	 * Check whether the runtime #VC exception handler is active. It uses
1259 	 * the per-CPU GHCB page which is set up by sev_es_init_vc_handling().
1260 	 *
1261 	 * If SNP is active, register the per-CPU GHCB page so that the runtime
1262 	 * exception handler can use it.
1263 	 */
1264 	if (initial_vc_handler == (unsigned long)kernel_exc_vmm_communication) {
1265 		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
1266 			snp_register_per_cpu_ghcb();
1267 
1268 		return;
1269 	}
1270 
1271 	/*
1272 	 * Clear the boot_ghcb. The first exception comes in before the bss
1273 	 * section is cleared.
1274 	 */
1275 	memset(&boot_ghcb_page, 0, PAGE_SIZE);
1276 
1277 	/* Alright - Make the boot-ghcb public */
1278 	boot_ghcb = &boot_ghcb_page;
1279 
1280 	/* SNP guest requires that GHCB GPA must be registered. */
1281 	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
1282 		snp_register_ghcb_early(__pa(&boot_ghcb_page));
1283 }
1284 
1285 #ifdef CONFIG_HOTPLUG_CPU
1286 static void sev_es_ap_hlt_loop(void)
1287 {
1288 	struct ghcb_state state;
1289 	struct ghcb *ghcb;
1290 
1291 	ghcb = __sev_get_ghcb(&state);
1292 
1293 	while (true) {
1294 		vc_ghcb_invalidate(ghcb);
1295 		ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_HLT_LOOP);
1296 		ghcb_set_sw_exit_info_1(ghcb, 0);
1297 		ghcb_set_sw_exit_info_2(ghcb, 0);
1298 
1299 		sev_es_wr_ghcb_msr(__pa(ghcb));
1300 		VMGEXIT();
1301 
1302 		/* Wakeup signal? */
1303 		if (ghcb_sw_exit_info_2_is_valid(ghcb) &&
1304 		    ghcb->save.sw_exit_info_2)
1305 			break;
1306 	}
1307 
1308 	__sev_put_ghcb(&state);
1309 }
1310 
1311 /*
1312  * Play_dead handler when running under SEV-ES. This is needed because
1313  * the hypervisor can't deliver an SIPI request to restart the AP.
1314  * Instead the kernel has to issue a VMGEXIT to halt the VCPU until the
1315  * hypervisor wakes it up again.
1316  */
1317 static void sev_es_play_dead(void)
1318 {
1319 	play_dead_common();
1320 
1321 	/* IRQs now disabled */
1322 
1323 	sev_es_ap_hlt_loop();
1324 
1325 	/*
1326 	 * If we get here, the VCPU was woken up again. Jump to CPU
1327 	 * startup code to get it back online.
1328 	 */
1329 	start_cpu0();
1330 }
1331 #else  /* CONFIG_HOTPLUG_CPU */
1332 #define sev_es_play_dead	native_play_dead
1333 #endif /* CONFIG_HOTPLUG_CPU */
1334 
1335 #ifdef CONFIG_SMP
1336 static void __init sev_es_setup_play_dead(void)
1337 {
1338 	smp_ops.play_dead = sev_es_play_dead;
1339 }
1340 #else
1341 static inline void sev_es_setup_play_dead(void) { }
1342 #endif
1343 
1344 static void __init alloc_runtime_data(int cpu)
1345 {
1346 	struct sev_es_runtime_data *data;
1347 
1348 	data = memblock_alloc(sizeof(*data), PAGE_SIZE);
1349 	if (!data)
1350 		panic("Can't allocate SEV-ES runtime data");
1351 
1352 	per_cpu(runtime_data, cpu) = data;
1353 }
1354 
1355 static void __init init_ghcb(int cpu)
1356 {
1357 	struct sev_es_runtime_data *data;
1358 	int err;
1359 
1360 	data = per_cpu(runtime_data, cpu);
1361 
1362 	err = early_set_memory_decrypted((unsigned long)&data->ghcb_page,
1363 					 sizeof(data->ghcb_page));
1364 	if (err)
1365 		panic("Can't map GHCBs unencrypted");
1366 
1367 	memset(&data->ghcb_page, 0, sizeof(data->ghcb_page));
1368 
1369 	data->ghcb_active = false;
1370 	data->backup_ghcb_active = false;
1371 }
1372 
1373 void __init sev_es_init_vc_handling(void)
1374 {
1375 	int cpu;
1376 
1377 	BUILD_BUG_ON(offsetof(struct sev_es_runtime_data, ghcb_page) % PAGE_SIZE);
1378 
1379 	if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
1380 		return;
1381 
1382 	if (!sev_es_check_cpu_features())
1383 		panic("SEV-ES CPU Features missing");
1384 
1385 	/*
1386 	 * SNP is supported in v2 of the GHCB spec which mandates support for HV
1387 	 * features.
1388 	 */
1389 	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
1390 		sev_hv_features = get_hv_features();
1391 
1392 		if (!(sev_hv_features & GHCB_HV_FT_SNP))
1393 			sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SNP_UNSUPPORTED);
1394 	}
1395 
1396 	/* Enable SEV-ES special handling */
1397 	static_branch_enable(&sev_es_enable_key);
1398 
1399 	/* Initialize per-cpu GHCB pages */
1400 	for_each_possible_cpu(cpu) {
1401 		alloc_runtime_data(cpu);
1402 		init_ghcb(cpu);
1403 	}
1404 
1405 	sev_es_setup_play_dead();
1406 
1407 	/* Secondary CPUs use the runtime #VC handler */
1408 	initial_vc_handler = (unsigned long)kernel_exc_vmm_communication;
1409 }
1410 
1411 static void __init vc_early_forward_exception(struct es_em_ctxt *ctxt)
1412 {
1413 	int trapnr = ctxt->fi.vector;
1414 
1415 	if (trapnr == X86_TRAP_PF)
1416 		native_write_cr2(ctxt->fi.cr2);
1417 
1418 	ctxt->regs->orig_ax = ctxt->fi.error_code;
1419 	do_early_exception(ctxt->regs, trapnr);
1420 }
1421 
1422 static long *vc_insn_get_rm(struct es_em_ctxt *ctxt)
1423 {
1424 	long *reg_array;
1425 	int offset;
1426 
1427 	reg_array = (long *)ctxt->regs;
1428 	offset    = insn_get_modrm_rm_off(&ctxt->insn, ctxt->regs);
1429 
1430 	if (offset < 0)
1431 		return NULL;
1432 
1433 	offset /= sizeof(long);
1434 
1435 	return reg_array + offset;
1436 }
1437 static enum es_result vc_do_mmio(struct ghcb *ghcb, struct es_em_ctxt *ctxt,
1438 				 unsigned int bytes, bool read)
1439 {
1440 	u64 exit_code, exit_info_1, exit_info_2;
1441 	unsigned long ghcb_pa = __pa(ghcb);
1442 	enum es_result res;
1443 	phys_addr_t paddr;
1444 	void __user *ref;
1445 
1446 	ref = insn_get_addr_ref(&ctxt->insn, ctxt->regs);
1447 	if (ref == (void __user *)-1L)
1448 		return ES_UNSUPPORTED;
1449 
1450 	exit_code = read ? SVM_VMGEXIT_MMIO_READ : SVM_VMGEXIT_MMIO_WRITE;
1451 
1452 	res = vc_slow_virt_to_phys(ghcb, ctxt, (unsigned long)ref, &paddr);
1453 	if (res != ES_OK) {
1454 		if (res == ES_EXCEPTION && !read)
1455 			ctxt->fi.error_code |= X86_PF_WRITE;
1456 
1457 		return res;
1458 	}
1459 
1460 	exit_info_1 = paddr;
1461 	/* Can never be greater than 8 */
1462 	exit_info_2 = bytes;
1463 
1464 	ghcb_set_sw_scratch(ghcb, ghcb_pa + offsetof(struct ghcb, shared_buffer));
1465 
1466 	return sev_es_ghcb_hv_call(ghcb, ctxt, exit_code, exit_info_1, exit_info_2);
1467 }
1468 
1469 /*
1470  * The MOVS instruction has two memory operands, which raises the
1471  * problem that it is not known whether the access to the source or the
1472  * destination caused the #VC exception (and hence whether an MMIO read
1473  * or write operation needs to be emulated).
1474  *
1475  * Instead of playing games with walking page-tables and trying to guess
1476  * whether the source or destination is an MMIO range, split the move
1477  * into two operations, a read and a write with only one memory operand.
1478  * This will cause a nested #VC exception on the MMIO address which can
1479  * then be handled.
1480  *
1481  * This implementation has the benefit that it also supports MOVS where
1482  * source _and_ destination are MMIO regions.
1483  *
1484  * It will slow MOVS on MMIO down a lot, but in SEV-ES guests it is a
1485  * rare operation. If it turns out to be a performance problem the split
1486  * operations can be moved to memcpy_fromio() and memcpy_toio().
1487  */
1488 static enum es_result vc_handle_mmio_movs(struct es_em_ctxt *ctxt,
1489 					  unsigned int bytes)
1490 {
1491 	unsigned long ds_base, es_base;
1492 	unsigned char *src, *dst;
1493 	unsigned char buffer[8];
1494 	enum es_result ret;
1495 	bool rep;
1496 	int off;
1497 
1498 	ds_base = insn_get_seg_base(ctxt->regs, INAT_SEG_REG_DS);
1499 	es_base = insn_get_seg_base(ctxt->regs, INAT_SEG_REG_ES);
1500 
1501 	if (ds_base == -1L || es_base == -1L) {
1502 		ctxt->fi.vector = X86_TRAP_GP;
1503 		ctxt->fi.error_code = 0;
1504 		return ES_EXCEPTION;
1505 	}
1506 
1507 	src = ds_base + (unsigned char *)ctxt->regs->si;
1508 	dst = es_base + (unsigned char *)ctxt->regs->di;
1509 
1510 	ret = vc_read_mem(ctxt, src, buffer, bytes);
1511 	if (ret != ES_OK)
1512 		return ret;
1513 
1514 	ret = vc_write_mem(ctxt, dst, buffer, bytes);
1515 	if (ret != ES_OK)
1516 		return ret;
1517 
1518 	if (ctxt->regs->flags & X86_EFLAGS_DF)
1519 		off = -bytes;
1520 	else
1521 		off =  bytes;
1522 
1523 	ctxt->regs->si += off;
1524 	ctxt->regs->di += off;
1525 
1526 	rep = insn_has_rep_prefix(&ctxt->insn);
1527 	if (rep)
1528 		ctxt->regs->cx -= 1;
1529 
1530 	if (!rep || ctxt->regs->cx == 0)
1531 		return ES_OK;
1532 	else
1533 		return ES_RETRY;
1534 }
1535 
1536 static enum es_result vc_handle_mmio(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
1537 {
1538 	struct insn *insn = &ctxt->insn;
1539 	unsigned int bytes = 0;
1540 	enum mmio_type mmio;
1541 	enum es_result ret;
1542 	u8 sign_byte;
1543 	long *reg_data;
1544 
1545 	mmio = insn_decode_mmio(insn, &bytes);
1546 	if (mmio == MMIO_DECODE_FAILED)
1547 		return ES_DECODE_FAILED;
1548 
1549 	if (mmio != MMIO_WRITE_IMM && mmio != MMIO_MOVS) {
1550 		reg_data = insn_get_modrm_reg_ptr(insn, ctxt->regs);
1551 		if (!reg_data)
1552 			return ES_DECODE_FAILED;
1553 	}
1554 
1555 	switch (mmio) {
1556 	case MMIO_WRITE:
1557 		memcpy(ghcb->shared_buffer, reg_data, bytes);
1558 		ret = vc_do_mmio(ghcb, ctxt, bytes, false);
1559 		break;
1560 	case MMIO_WRITE_IMM:
1561 		memcpy(ghcb->shared_buffer, insn->immediate1.bytes, bytes);
1562 		ret = vc_do_mmio(ghcb, ctxt, bytes, false);
1563 		break;
1564 	case MMIO_READ:
1565 		ret = vc_do_mmio(ghcb, ctxt, bytes, true);
1566 		if (ret)
1567 			break;
1568 
1569 		/* Zero-extend for 32-bit operation */
1570 		if (bytes == 4)
1571 			*reg_data = 0;
1572 
1573 		memcpy(reg_data, ghcb->shared_buffer, bytes);
1574 		break;
1575 	case MMIO_READ_ZERO_EXTEND:
1576 		ret = vc_do_mmio(ghcb, ctxt, bytes, true);
1577 		if (ret)
1578 			break;
1579 
1580 		/* Zero extend based on operand size */
1581 		memset(reg_data, 0, insn->opnd_bytes);
1582 		memcpy(reg_data, ghcb->shared_buffer, bytes);
1583 		break;
1584 	case MMIO_READ_SIGN_EXTEND:
1585 		ret = vc_do_mmio(ghcb, ctxt, bytes, true);
1586 		if (ret)
1587 			break;
1588 
1589 		if (bytes == 1) {
1590 			u8 *val = (u8 *)ghcb->shared_buffer;
1591 
1592 			sign_byte = (*val & 0x80) ? 0xff : 0x00;
1593 		} else {
1594 			u16 *val = (u16 *)ghcb->shared_buffer;
1595 
1596 			sign_byte = (*val & 0x8000) ? 0xff : 0x00;
1597 		}
1598 
1599 		/* Sign extend based on operand size */
1600 		memset(reg_data, sign_byte, insn->opnd_bytes);
1601 		memcpy(reg_data, ghcb->shared_buffer, bytes);
1602 		break;
1603 	case MMIO_MOVS:
1604 		ret = vc_handle_mmio_movs(ctxt, bytes);
1605 		break;
1606 	default:
1607 		ret = ES_UNSUPPORTED;
1608 		break;
1609 	}
1610 
1611 	return ret;
1612 }
1613 
1614 static enum es_result vc_handle_dr7_write(struct ghcb *ghcb,
1615 					  struct es_em_ctxt *ctxt)
1616 {
1617 	struct sev_es_runtime_data *data = this_cpu_read(runtime_data);
1618 	long val, *reg = vc_insn_get_rm(ctxt);
1619 	enum es_result ret;
1620 
1621 	if (!reg)
1622 		return ES_DECODE_FAILED;
1623 
1624 	val = *reg;
1625 
1626 	/* Upper 32 bits must be written as zeroes */
1627 	if (val >> 32) {
1628 		ctxt->fi.vector = X86_TRAP_GP;
1629 		ctxt->fi.error_code = 0;
1630 		return ES_EXCEPTION;
1631 	}
1632 
1633 	/* Clear out other reserved bits and set bit 10 */
1634 	val = (val & 0xffff23ffL) | BIT(10);
1635 
1636 	/* Early non-zero writes to DR7 are not supported */
1637 	if (!data && (val & ~DR7_RESET_VALUE))
1638 		return ES_UNSUPPORTED;
1639 
1640 	/* Using a value of 0 for ExitInfo1 means RAX holds the value */
1641 	ghcb_set_rax(ghcb, val);
1642 	ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_WRITE_DR7, 0, 0);
1643 	if (ret != ES_OK)
1644 		return ret;
1645 
1646 	if (data)
1647 		data->dr7 = val;
1648 
1649 	return ES_OK;
1650 }
1651 
1652 static enum es_result vc_handle_dr7_read(struct ghcb *ghcb,
1653 					 struct es_em_ctxt *ctxt)
1654 {
1655 	struct sev_es_runtime_data *data = this_cpu_read(runtime_data);
1656 	long *reg = vc_insn_get_rm(ctxt);
1657 
1658 	if (!reg)
1659 		return ES_DECODE_FAILED;
1660 
1661 	if (data)
1662 		*reg = data->dr7;
1663 	else
1664 		*reg = DR7_RESET_VALUE;
1665 
1666 	return ES_OK;
1667 }
1668 
1669 static enum es_result vc_handle_wbinvd(struct ghcb *ghcb,
1670 				       struct es_em_ctxt *ctxt)
1671 {
1672 	return sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_WBINVD, 0, 0);
1673 }
1674 
1675 static enum es_result vc_handle_rdpmc(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
1676 {
1677 	enum es_result ret;
1678 
1679 	ghcb_set_rcx(ghcb, ctxt->regs->cx);
1680 
1681 	ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_RDPMC, 0, 0);
1682 	if (ret != ES_OK)
1683 		return ret;
1684 
1685 	if (!(ghcb_rax_is_valid(ghcb) && ghcb_rdx_is_valid(ghcb)))
1686 		return ES_VMM_ERROR;
1687 
1688 	ctxt->regs->ax = ghcb->save.rax;
1689 	ctxt->regs->dx = ghcb->save.rdx;
1690 
1691 	return ES_OK;
1692 }
1693 
1694 static enum es_result vc_handle_monitor(struct ghcb *ghcb,
1695 					struct es_em_ctxt *ctxt)
1696 {
1697 	/*
1698 	 * Treat it as a NOP and do not leak a physical address to the
1699 	 * hypervisor.
1700 	 */
1701 	return ES_OK;
1702 }
1703 
1704 static enum es_result vc_handle_mwait(struct ghcb *ghcb,
1705 				      struct es_em_ctxt *ctxt)
1706 {
1707 	/* Treat the same as MONITOR/MONITORX */
1708 	return ES_OK;
1709 }
1710 
1711 static enum es_result vc_handle_vmmcall(struct ghcb *ghcb,
1712 					struct es_em_ctxt *ctxt)
1713 {
1714 	enum es_result ret;
1715 
1716 	ghcb_set_rax(ghcb, ctxt->regs->ax);
1717 	ghcb_set_cpl(ghcb, user_mode(ctxt->regs) ? 3 : 0);
1718 
1719 	if (x86_platform.hyper.sev_es_hcall_prepare)
1720 		x86_platform.hyper.sev_es_hcall_prepare(ghcb, ctxt->regs);
1721 
1722 	ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_VMMCALL, 0, 0);
1723 	if (ret != ES_OK)
1724 		return ret;
1725 
1726 	if (!ghcb_rax_is_valid(ghcb))
1727 		return ES_VMM_ERROR;
1728 
1729 	ctxt->regs->ax = ghcb->save.rax;
1730 
1731 	/*
1732 	 * Call sev_es_hcall_finish() after regs->ax is already set.
1733 	 * This allows the hypervisor handler to overwrite it again if
1734 	 * necessary.
1735 	 */
1736 	if (x86_platform.hyper.sev_es_hcall_finish &&
1737 	    !x86_platform.hyper.sev_es_hcall_finish(ghcb, ctxt->regs))
1738 		return ES_VMM_ERROR;
1739 
1740 	return ES_OK;
1741 }
1742 
1743 static enum es_result vc_handle_trap_ac(struct ghcb *ghcb,
1744 					struct es_em_ctxt *ctxt)
1745 {
1746 	/*
1747 	 * Calling ecx_alignment_check() directly does not work, because it
1748 	 * enables IRQs and the GHCB is active. Forward the exception and call
1749 	 * it later from vc_forward_exception().
1750 	 */
1751 	ctxt->fi.vector = X86_TRAP_AC;
1752 	ctxt->fi.error_code = 0;
1753 	return ES_EXCEPTION;
1754 }
1755 
1756 static enum es_result vc_handle_exitcode(struct es_em_ctxt *ctxt,
1757 					 struct ghcb *ghcb,
1758 					 unsigned long exit_code)
1759 {
1760 	enum es_result result;
1761 
1762 	switch (exit_code) {
1763 	case SVM_EXIT_READ_DR7:
1764 		result = vc_handle_dr7_read(ghcb, ctxt);
1765 		break;
1766 	case SVM_EXIT_WRITE_DR7:
1767 		result = vc_handle_dr7_write(ghcb, ctxt);
1768 		break;
1769 	case SVM_EXIT_EXCP_BASE + X86_TRAP_AC:
1770 		result = vc_handle_trap_ac(ghcb, ctxt);
1771 		break;
1772 	case SVM_EXIT_RDTSC:
1773 	case SVM_EXIT_RDTSCP:
1774 		result = vc_handle_rdtsc(ghcb, ctxt, exit_code);
1775 		break;
1776 	case SVM_EXIT_RDPMC:
1777 		result = vc_handle_rdpmc(ghcb, ctxt);
1778 		break;
1779 	case SVM_EXIT_INVD:
1780 		pr_err_ratelimited("#VC exception for INVD??? Seriously???\n");
1781 		result = ES_UNSUPPORTED;
1782 		break;
1783 	case SVM_EXIT_CPUID:
1784 		result = vc_handle_cpuid(ghcb, ctxt);
1785 		break;
1786 	case SVM_EXIT_IOIO:
1787 		result = vc_handle_ioio(ghcb, ctxt);
1788 		break;
1789 	case SVM_EXIT_MSR:
1790 		result = vc_handle_msr(ghcb, ctxt);
1791 		break;
1792 	case SVM_EXIT_VMMCALL:
1793 		result = vc_handle_vmmcall(ghcb, ctxt);
1794 		break;
1795 	case SVM_EXIT_WBINVD:
1796 		result = vc_handle_wbinvd(ghcb, ctxt);
1797 		break;
1798 	case SVM_EXIT_MONITOR:
1799 		result = vc_handle_monitor(ghcb, ctxt);
1800 		break;
1801 	case SVM_EXIT_MWAIT:
1802 		result = vc_handle_mwait(ghcb, ctxt);
1803 		break;
1804 	case SVM_EXIT_NPF:
1805 		result = vc_handle_mmio(ghcb, ctxt);
1806 		break;
1807 	default:
1808 		/*
1809 		 * Unexpected #VC exception
1810 		 */
1811 		result = ES_UNSUPPORTED;
1812 	}
1813 
1814 	return result;
1815 }
1816 
1817 static __always_inline void vc_forward_exception(struct es_em_ctxt *ctxt)
1818 {
1819 	long error_code = ctxt->fi.error_code;
1820 	int trapnr = ctxt->fi.vector;
1821 
1822 	ctxt->regs->orig_ax = ctxt->fi.error_code;
1823 
1824 	switch (trapnr) {
1825 	case X86_TRAP_GP:
1826 		exc_general_protection(ctxt->regs, error_code);
1827 		break;
1828 	case X86_TRAP_UD:
1829 		exc_invalid_op(ctxt->regs);
1830 		break;
1831 	case X86_TRAP_PF:
1832 		write_cr2(ctxt->fi.cr2);
1833 		exc_page_fault(ctxt->regs, error_code);
1834 		break;
1835 	case X86_TRAP_AC:
1836 		exc_alignment_check(ctxt->regs, error_code);
1837 		break;
1838 	default:
1839 		pr_emerg("Unsupported exception in #VC instruction emulation - can't continue\n");
1840 		BUG();
1841 	}
1842 }
1843 
1844 static __always_inline bool is_vc2_stack(unsigned long sp)
1845 {
1846 	return (sp >= __this_cpu_ist_bottom_va(VC2) && sp < __this_cpu_ist_top_va(VC2));
1847 }
1848 
1849 static __always_inline bool vc_from_invalid_context(struct pt_regs *regs)
1850 {
1851 	unsigned long sp, prev_sp;
1852 
1853 	sp      = (unsigned long)regs;
1854 	prev_sp = regs->sp;
1855 
1856 	/*
1857 	 * If the code was already executing on the VC2 stack when the #VC
1858 	 * happened, let it proceed to the normal handling routine. This way the
1859 	 * code executing on the VC2 stack can cause #VC exceptions to get handled.
1860 	 */
1861 	return is_vc2_stack(sp) && !is_vc2_stack(prev_sp);
1862 }
1863 
1864 static bool vc_raw_handle_exception(struct pt_regs *regs, unsigned long error_code)
1865 {
1866 	struct ghcb_state state;
1867 	struct es_em_ctxt ctxt;
1868 	enum es_result result;
1869 	struct ghcb *ghcb;
1870 	bool ret = true;
1871 
1872 	ghcb = __sev_get_ghcb(&state);
1873 
1874 	vc_ghcb_invalidate(ghcb);
1875 	result = vc_init_em_ctxt(&ctxt, regs, error_code);
1876 
1877 	if (result == ES_OK)
1878 		result = vc_handle_exitcode(&ctxt, ghcb, error_code);
1879 
1880 	__sev_put_ghcb(&state);
1881 
1882 	/* Done - now check the result */
1883 	switch (result) {
1884 	case ES_OK:
1885 		vc_finish_insn(&ctxt);
1886 		break;
1887 	case ES_UNSUPPORTED:
1888 		pr_err_ratelimited("Unsupported exit-code 0x%02lx in #VC exception (IP: 0x%lx)\n",
1889 				   error_code, regs->ip);
1890 		ret = false;
1891 		break;
1892 	case ES_VMM_ERROR:
1893 		pr_err_ratelimited("Failure in communication with VMM (exit-code 0x%02lx IP: 0x%lx)\n",
1894 				   error_code, regs->ip);
1895 		ret = false;
1896 		break;
1897 	case ES_DECODE_FAILED:
1898 		pr_err_ratelimited("Failed to decode instruction (exit-code 0x%02lx IP: 0x%lx)\n",
1899 				   error_code, regs->ip);
1900 		ret = false;
1901 		break;
1902 	case ES_EXCEPTION:
1903 		vc_forward_exception(&ctxt);
1904 		break;
1905 	case ES_RETRY:
1906 		/* Nothing to do */
1907 		break;
1908 	default:
1909 		pr_emerg("Unknown result in %s():%d\n", __func__, result);
1910 		/*
1911 		 * Emulating the instruction which caused the #VC exception
1912 		 * failed - can't continue so print debug information
1913 		 */
1914 		BUG();
1915 	}
1916 
1917 	return ret;
1918 }
1919 
1920 static __always_inline bool vc_is_db(unsigned long error_code)
1921 {
1922 	return error_code == SVM_EXIT_EXCP_BASE + X86_TRAP_DB;
1923 }
1924 
1925 /*
1926  * Runtime #VC exception handler when raised from kernel mode. Runs in NMI mode
1927  * and will panic when an error happens.
1928  */
1929 DEFINE_IDTENTRY_VC_KERNEL(exc_vmm_communication)
1930 {
1931 	irqentry_state_t irq_state;
1932 
1933 	/*
1934 	 * With the current implementation it is always possible to switch to a
1935 	 * safe stack because #VC exceptions only happen at known places, like
1936 	 * intercepted instructions or accesses to MMIO areas/IO ports. They can
1937 	 * also happen with code instrumentation when the hypervisor intercepts
1938 	 * #DB, but the critical paths are forbidden to be instrumented, so #DB
1939 	 * exceptions currently also only happen in safe places.
1940 	 *
1941 	 * But keep this here in case the noinstr annotations are violated due
1942 	 * to bug elsewhere.
1943 	 */
1944 	if (unlikely(vc_from_invalid_context(regs))) {
1945 		instrumentation_begin();
1946 		panic("Can't handle #VC exception from unsupported context\n");
1947 		instrumentation_end();
1948 	}
1949 
1950 	/*
1951 	 * Handle #DB before calling into !noinstr code to avoid recursive #DB.
1952 	 */
1953 	if (vc_is_db(error_code)) {
1954 		exc_debug(regs);
1955 		return;
1956 	}
1957 
1958 	irq_state = irqentry_nmi_enter(regs);
1959 
1960 	instrumentation_begin();
1961 
1962 	if (!vc_raw_handle_exception(regs, error_code)) {
1963 		/* Show some debug info */
1964 		show_regs(regs);
1965 
1966 		/* Ask hypervisor to sev_es_terminate */
1967 		sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
1968 
1969 		/* If that fails and we get here - just panic */
1970 		panic("Returned from Terminate-Request to Hypervisor\n");
1971 	}
1972 
1973 	instrumentation_end();
1974 	irqentry_nmi_exit(regs, irq_state);
1975 }
1976 
1977 /*
1978  * Runtime #VC exception handler when raised from user mode. Runs in IRQ mode
1979  * and will kill the current task with SIGBUS when an error happens.
1980  */
1981 DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
1982 {
1983 	/*
1984 	 * Handle #DB before calling into !noinstr code to avoid recursive #DB.
1985 	 */
1986 	if (vc_is_db(error_code)) {
1987 		noist_exc_debug(regs);
1988 		return;
1989 	}
1990 
1991 	irqentry_enter_from_user_mode(regs);
1992 	instrumentation_begin();
1993 
1994 	if (!vc_raw_handle_exception(regs, error_code)) {
1995 		/*
1996 		 * Do not kill the machine if user-space triggered the
1997 		 * exception. Send SIGBUS instead and let user-space deal with
1998 		 * it.
1999 		 */
2000 		force_sig_fault(SIGBUS, BUS_OBJERR, (void __user *)0);
2001 	}
2002 
2003 	instrumentation_end();
2004 	irqentry_exit_to_user_mode(regs);
2005 }
2006 
2007 bool __init handle_vc_boot_ghcb(struct pt_regs *regs)
2008 {
2009 	unsigned long exit_code = regs->orig_ax;
2010 	struct es_em_ctxt ctxt;
2011 	enum es_result result;
2012 
2013 	vc_ghcb_invalidate(boot_ghcb);
2014 
2015 	result = vc_init_em_ctxt(&ctxt, regs, exit_code);
2016 	if (result == ES_OK)
2017 		result = vc_handle_exitcode(&ctxt, boot_ghcb, exit_code);
2018 
2019 	/* Done - now check the result */
2020 	switch (result) {
2021 	case ES_OK:
2022 		vc_finish_insn(&ctxt);
2023 		break;
2024 	case ES_UNSUPPORTED:
2025 		early_printk("PANIC: Unsupported exit-code 0x%02lx in early #VC exception (IP: 0x%lx)\n",
2026 				exit_code, regs->ip);
2027 		goto fail;
2028 	case ES_VMM_ERROR:
2029 		early_printk("PANIC: Failure in communication with VMM (exit-code 0x%02lx IP: 0x%lx)\n",
2030 				exit_code, regs->ip);
2031 		goto fail;
2032 	case ES_DECODE_FAILED:
2033 		early_printk("PANIC: Failed to decode instruction (exit-code 0x%02lx IP: 0x%lx)\n",
2034 				exit_code, regs->ip);
2035 		goto fail;
2036 	case ES_EXCEPTION:
2037 		vc_early_forward_exception(&ctxt);
2038 		break;
2039 	case ES_RETRY:
2040 		/* Nothing to do */
2041 		break;
2042 	default:
2043 		BUG();
2044 	}
2045 
2046 	return true;
2047 
2048 fail:
2049 	show_regs(regs);
2050 
2051 	sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
2052 }
2053 
2054 /*
2055  * Initial set up of SNP relies on information provided by the
2056  * Confidential Computing blob, which can be passed to the kernel
2057  * in the following ways, depending on how it is booted:
2058  *
2059  * - when booted via the boot/decompress kernel:
2060  *   - via boot_params
2061  *
2062  * - when booted directly by firmware/bootloader (e.g. CONFIG_PVH):
2063  *   - via a setup_data entry, as defined by the Linux Boot Protocol
2064  *
2065  * Scan for the blob in that order.
2066  */
2067 static __init struct cc_blob_sev_info *find_cc_blob(struct boot_params *bp)
2068 {
2069 	struct cc_blob_sev_info *cc_info;
2070 
2071 	/* Boot kernel would have passed the CC blob via boot_params. */
2072 	if (bp->cc_blob_address) {
2073 		cc_info = (struct cc_blob_sev_info *)(unsigned long)bp->cc_blob_address;
2074 		goto found_cc_info;
2075 	}
2076 
2077 	/*
2078 	 * If kernel was booted directly, without the use of the
2079 	 * boot/decompression kernel, the CC blob may have been passed via
2080 	 * setup_data instead.
2081 	 */
2082 	cc_info = find_cc_blob_setup_data(bp);
2083 	if (!cc_info)
2084 		return NULL;
2085 
2086 found_cc_info:
2087 	if (cc_info->magic != CC_BLOB_SEV_HDR_MAGIC)
2088 		snp_abort();
2089 
2090 	return cc_info;
2091 }
2092 
2093 bool __init snp_init(struct boot_params *bp)
2094 {
2095 	struct cc_blob_sev_info *cc_info;
2096 
2097 	if (!bp)
2098 		return false;
2099 
2100 	cc_info = find_cc_blob(bp);
2101 	if (!cc_info)
2102 		return false;
2103 
2104 	setup_cpuid_table(cc_info);
2105 
2106 	/*
2107 	 * The CC blob will be used later to access the secrets page. Cache
2108 	 * it here like the boot kernel does.
2109 	 */
2110 	bp->cc_blob_address = (u32)(unsigned long)cc_info;
2111 
2112 	return true;
2113 }
2114 
2115 void __init __noreturn snp_abort(void)
2116 {
2117 	sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SNP_UNSUPPORTED);
2118 }
2119 
2120 static void dump_cpuid_table(void)
2121 {
2122 	const struct snp_cpuid_table *cpuid_table = snp_cpuid_get_table();
2123 	int i = 0;
2124 
2125 	pr_info("count=%d reserved=0x%x reserved2=0x%llx\n",
2126 		cpuid_table->count, cpuid_table->__reserved1, cpuid_table->__reserved2);
2127 
2128 	for (i = 0; i < SNP_CPUID_COUNT_MAX; i++) {
2129 		const struct snp_cpuid_fn *fn = &cpuid_table->fn[i];
2130 
2131 		pr_info("index=%3d fn=0x%08x subfn=0x%08x: eax=0x%08x ebx=0x%08x ecx=0x%08x edx=0x%08x xcr0_in=0x%016llx xss_in=0x%016llx reserved=0x%016llx\n",
2132 			i, fn->eax_in, fn->ecx_in, fn->eax, fn->ebx, fn->ecx,
2133 			fn->edx, fn->xcr0_in, fn->xss_in, fn->__reserved);
2134 	}
2135 }
2136 
2137 /*
2138  * It is useful from an auditing/testing perspective to provide an easy way
2139  * for the guest owner to know that the CPUID table has been initialized as
2140  * expected, but that initialization happens too early in boot to print any
2141  * sort of indicator, and there's not really any other good place to do it,
2142  * so do it here.
2143  */
2144 static int __init report_cpuid_table(void)
2145 {
2146 	const struct snp_cpuid_table *cpuid_table = snp_cpuid_get_table();
2147 
2148 	if (!cpuid_table->count)
2149 		return 0;
2150 
2151 	pr_info("Using SNP CPUID table, %d entries present.\n",
2152 		cpuid_table->count);
2153 
2154 	if (sev_cfg.debug)
2155 		dump_cpuid_table();
2156 
2157 	return 0;
2158 }
2159 arch_initcall(report_cpuid_table);
2160 
2161 static int __init init_sev_config(char *str)
2162 {
2163 	char *s;
2164 
2165 	while ((s = strsep(&str, ","))) {
2166 		if (!strcmp(s, "debug")) {
2167 			sev_cfg.debug = true;
2168 			continue;
2169 		}
2170 
2171 		pr_info("SEV command-line option '%s' was not recognized\n", s);
2172 	}
2173 
2174 	return 1;
2175 }
2176 __setup("sev=", init_sev_config);
2177 
2178 int snp_issue_guest_request(u64 exit_code, struct snp_req_data *input, unsigned long *fw_err)
2179 {
2180 	struct ghcb_state state;
2181 	struct es_em_ctxt ctxt;
2182 	unsigned long flags;
2183 	struct ghcb *ghcb;
2184 	int ret;
2185 
2186 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
2187 		return -ENODEV;
2188 
2189 	if (!fw_err)
2190 		return -EINVAL;
2191 
2192 	/*
2193 	 * __sev_get_ghcb() needs to run with IRQs disabled because it is using
2194 	 * a per-CPU GHCB.
2195 	 */
2196 	local_irq_save(flags);
2197 
2198 	ghcb = __sev_get_ghcb(&state);
2199 	if (!ghcb) {
2200 		ret = -EIO;
2201 		goto e_restore_irq;
2202 	}
2203 
2204 	vc_ghcb_invalidate(ghcb);
2205 
2206 	if (exit_code == SVM_VMGEXIT_EXT_GUEST_REQUEST) {
2207 		ghcb_set_rax(ghcb, input->data_gpa);
2208 		ghcb_set_rbx(ghcb, input->data_npages);
2209 	}
2210 
2211 	ret = sev_es_ghcb_hv_call(ghcb, &ctxt, exit_code, input->req_gpa, input->resp_gpa);
2212 	if (ret)
2213 		goto e_put;
2214 
2215 	if (ghcb->save.sw_exit_info_2) {
2216 		/* Number of expected pages are returned in RBX */
2217 		if (exit_code == SVM_VMGEXIT_EXT_GUEST_REQUEST &&
2218 		    ghcb->save.sw_exit_info_2 == SNP_GUEST_REQ_INVALID_LEN)
2219 			input->data_npages = ghcb_get_rbx(ghcb);
2220 
2221 		*fw_err = ghcb->save.sw_exit_info_2;
2222 
2223 		ret = -EIO;
2224 	}
2225 
2226 e_put:
2227 	__sev_put_ghcb(&state);
2228 e_restore_irq:
2229 	local_irq_restore(flags);
2230 
2231 	return ret;
2232 }
2233 EXPORT_SYMBOL_GPL(snp_issue_guest_request);
2234 
2235 static struct platform_device sev_guest_device = {
2236 	.name		= "sev-guest",
2237 	.id		= -1,
2238 };
2239 
2240 static int __init snp_init_platform_device(void)
2241 {
2242 	struct sev_guest_platform_data data;
2243 	u64 gpa;
2244 
2245 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
2246 		return -ENODEV;
2247 
2248 	gpa = get_secrets_page();
2249 	if (!gpa)
2250 		return -ENODEV;
2251 
2252 	data.secrets_gpa = gpa;
2253 	if (platform_device_add_data(&sev_guest_device, &data, sizeof(data)))
2254 		return -ENODEV;
2255 
2256 	if (platform_device_register(&sev_guest_device))
2257 		return -ENODEV;
2258 
2259 	pr_info("SNP guest platform device initialized.\n");
2260 	return 0;
2261 }
2262 device_initcall(snp_init_platform_device);
2263