xref: /openbmc/linux/arch/x86/kernel/sev.c (revision 3db55767)
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 #include <linux/psp-sev.h>
26 #include <uapi/linux/sev-guest.h>
27 
28 #include <asm/cpu_entry_area.h>
29 #include <asm/stacktrace.h>
30 #include <asm/sev.h>
31 #include <asm/insn-eval.h>
32 #include <asm/fpu/xcr.h>
33 #include <asm/processor.h>
34 #include <asm/realmode.h>
35 #include <asm/setup.h>
36 #include <asm/traps.h>
37 #include <asm/svm.h>
38 #include <asm/smp.h>
39 #include <asm/cpu.h>
40 #include <asm/apic.h>
41 #include <asm/cpuid.h>
42 #include <asm/cmdline.h>
43 
44 #define DR7_RESET_VALUE        0x400
45 
46 /* AP INIT values as documented in the APM2  section "Processor Initialization State" */
47 #define AP_INIT_CS_LIMIT		0xffff
48 #define AP_INIT_DS_LIMIT		0xffff
49 #define AP_INIT_LDTR_LIMIT		0xffff
50 #define AP_INIT_GDTR_LIMIT		0xffff
51 #define AP_INIT_IDTR_LIMIT		0xffff
52 #define AP_INIT_TR_LIMIT		0xffff
53 #define AP_INIT_RFLAGS_DEFAULT		0x2
54 #define AP_INIT_DR6_DEFAULT		0xffff0ff0
55 #define AP_INIT_GPAT_DEFAULT		0x0007040600070406ULL
56 #define AP_INIT_XCR0_DEFAULT		0x1
57 #define AP_INIT_X87_FTW_DEFAULT		0x5555
58 #define AP_INIT_X87_FCW_DEFAULT		0x0040
59 #define AP_INIT_CR0_DEFAULT		0x60000010
60 #define AP_INIT_MXCSR_DEFAULT		0x1f80
61 
62 /* For early boot hypervisor communication in SEV-ES enabled guests */
63 static struct ghcb boot_ghcb_page __bss_decrypted __aligned(PAGE_SIZE);
64 
65 /*
66  * Needs to be in the .data section because we need it NULL before bss is
67  * cleared
68  */
69 static struct ghcb *boot_ghcb __section(".data");
70 
71 /* Bitmap of SEV features supported by the hypervisor */
72 static u64 sev_hv_features __ro_after_init;
73 
74 /* #VC handler runtime per-CPU data */
75 struct sev_es_runtime_data {
76 	struct ghcb ghcb_page;
77 
78 	/*
79 	 * Reserve one page per CPU as backup storage for the unencrypted GHCB.
80 	 * It is needed when an NMI happens while the #VC handler uses the real
81 	 * GHCB, and the NMI handler itself is causing another #VC exception. In
82 	 * that case the GHCB content of the first handler needs to be backed up
83 	 * and restored.
84 	 */
85 	struct ghcb backup_ghcb;
86 
87 	/*
88 	 * Mark the per-cpu GHCBs as in-use to detect nested #VC exceptions.
89 	 * There is no need for it to be atomic, because nothing is written to
90 	 * the GHCB between the read and the write of ghcb_active. So it is safe
91 	 * to use it when a nested #VC exception happens before the write.
92 	 *
93 	 * This is necessary for example in the #VC->NMI->#VC case when the NMI
94 	 * happens while the first #VC handler uses the GHCB. When the NMI code
95 	 * raises a second #VC handler it might overwrite the contents of the
96 	 * GHCB written by the first handler. To avoid this the content of the
97 	 * GHCB is saved and restored when the GHCB is detected to be in use
98 	 * already.
99 	 */
100 	bool ghcb_active;
101 	bool backup_ghcb_active;
102 
103 	/*
104 	 * Cached DR7 value - write it on DR7 writes and return it on reads.
105 	 * That value will never make it to the real hardware DR7 as debugging
106 	 * is currently unsupported in SEV-ES guests.
107 	 */
108 	unsigned long dr7;
109 };
110 
111 struct ghcb_state {
112 	struct ghcb *ghcb;
113 };
114 
115 static DEFINE_PER_CPU(struct sev_es_runtime_data*, runtime_data);
116 DEFINE_STATIC_KEY_FALSE(sev_es_enable_key);
117 
118 static DEFINE_PER_CPU(struct sev_es_save_area *, sev_vmsa);
119 
120 struct sev_config {
121 	__u64 debug		: 1,
122 	      __reserved	: 63;
123 };
124 
125 static struct sev_config sev_cfg __read_mostly;
126 
127 static __always_inline bool on_vc_stack(struct pt_regs *regs)
128 {
129 	unsigned long sp = regs->sp;
130 
131 	/* User-mode RSP is not trusted */
132 	if (user_mode(regs))
133 		return false;
134 
135 	/* SYSCALL gap still has user-mode RSP */
136 	if (ip_within_syscall_gap(regs))
137 		return false;
138 
139 	return ((sp >= __this_cpu_ist_bottom_va(VC)) && (sp < __this_cpu_ist_top_va(VC)));
140 }
141 
142 /*
143  * This function handles the case when an NMI is raised in the #VC
144  * exception handler entry code, before the #VC handler has switched off
145  * its IST stack. In this case, the IST entry for #VC must be adjusted,
146  * so that any nested #VC exception will not overwrite the stack
147  * contents of the interrupted #VC handler.
148  *
149  * The IST entry is adjusted unconditionally so that it can be also be
150  * unconditionally adjusted back in __sev_es_ist_exit(). Otherwise a
151  * nested sev_es_ist_exit() call may adjust back the IST entry too
152  * early.
153  *
154  * The __sev_es_ist_enter() and __sev_es_ist_exit() functions always run
155  * on the NMI IST stack, as they are only called from NMI handling code
156  * right now.
157  */
158 void noinstr __sev_es_ist_enter(struct pt_regs *regs)
159 {
160 	unsigned long old_ist, new_ist;
161 
162 	/* Read old IST entry */
163 	new_ist = old_ist = __this_cpu_read(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC]);
164 
165 	/*
166 	 * If NMI happened while on the #VC IST stack, set the new IST
167 	 * value below regs->sp, so that the interrupted stack frame is
168 	 * not overwritten by subsequent #VC exceptions.
169 	 */
170 	if (on_vc_stack(regs))
171 		new_ist = regs->sp;
172 
173 	/*
174 	 * Reserve additional 8 bytes and store old IST value so this
175 	 * adjustment can be unrolled in __sev_es_ist_exit().
176 	 */
177 	new_ist -= sizeof(old_ist);
178 	*(unsigned long *)new_ist = old_ist;
179 
180 	/* Set new IST entry */
181 	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], new_ist);
182 }
183 
184 void noinstr __sev_es_ist_exit(void)
185 {
186 	unsigned long ist;
187 
188 	/* Read IST entry */
189 	ist = __this_cpu_read(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC]);
190 
191 	if (WARN_ON(ist == __this_cpu_ist_top_va(VC)))
192 		return;
193 
194 	/* Read back old IST entry and write it to the TSS */
195 	this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *(unsigned long *)ist);
196 }
197 
198 /*
199  * Nothing shall interrupt this code path while holding the per-CPU
200  * GHCB. The backup GHCB is only for NMIs interrupting this path.
201  *
202  * Callers must disable local interrupts around it.
203  */
204 static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
205 {
206 	struct sev_es_runtime_data *data;
207 	struct ghcb *ghcb;
208 
209 	WARN_ON(!irqs_disabled());
210 
211 	data = this_cpu_read(runtime_data);
212 	ghcb = &data->ghcb_page;
213 
214 	if (unlikely(data->ghcb_active)) {
215 		/* GHCB is already in use - save its contents */
216 
217 		if (unlikely(data->backup_ghcb_active)) {
218 			/*
219 			 * Backup-GHCB is also already in use. There is no way
220 			 * to continue here so just kill the machine. To make
221 			 * panic() work, mark GHCBs inactive so that messages
222 			 * can be printed out.
223 			 */
224 			data->ghcb_active        = false;
225 			data->backup_ghcb_active = false;
226 
227 			instrumentation_begin();
228 			panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
229 			instrumentation_end();
230 		}
231 
232 		/* Mark backup_ghcb active before writing to it */
233 		data->backup_ghcb_active = true;
234 
235 		state->ghcb = &data->backup_ghcb;
236 
237 		/* Backup GHCB content */
238 		*state->ghcb = *ghcb;
239 	} else {
240 		state->ghcb = NULL;
241 		data->ghcb_active = true;
242 	}
243 
244 	return ghcb;
245 }
246 
247 static inline u64 sev_es_rd_ghcb_msr(void)
248 {
249 	return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
250 }
251 
252 static __always_inline void sev_es_wr_ghcb_msr(u64 val)
253 {
254 	u32 low, high;
255 
256 	low  = (u32)(val);
257 	high = (u32)(val >> 32);
258 
259 	native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
260 }
261 
262 static int vc_fetch_insn_kernel(struct es_em_ctxt *ctxt,
263 				unsigned char *buffer)
264 {
265 	return copy_from_kernel_nofault(buffer, (unsigned char *)ctxt->regs->ip, MAX_INSN_SIZE);
266 }
267 
268 static enum es_result __vc_decode_user_insn(struct es_em_ctxt *ctxt)
269 {
270 	char buffer[MAX_INSN_SIZE];
271 	int insn_bytes;
272 
273 	insn_bytes = insn_fetch_from_user_inatomic(ctxt->regs, buffer);
274 	if (insn_bytes == 0) {
275 		/* Nothing could be copied */
276 		ctxt->fi.vector     = X86_TRAP_PF;
277 		ctxt->fi.error_code = X86_PF_INSTR | X86_PF_USER;
278 		ctxt->fi.cr2        = ctxt->regs->ip;
279 		return ES_EXCEPTION;
280 	} else if (insn_bytes == -EINVAL) {
281 		/* Effective RIP could not be calculated */
282 		ctxt->fi.vector     = X86_TRAP_GP;
283 		ctxt->fi.error_code = 0;
284 		ctxt->fi.cr2        = 0;
285 		return ES_EXCEPTION;
286 	}
287 
288 	if (!insn_decode_from_regs(&ctxt->insn, ctxt->regs, buffer, insn_bytes))
289 		return ES_DECODE_FAILED;
290 
291 	if (ctxt->insn.immediate.got)
292 		return ES_OK;
293 	else
294 		return ES_DECODE_FAILED;
295 }
296 
297 static enum es_result __vc_decode_kern_insn(struct es_em_ctxt *ctxt)
298 {
299 	char buffer[MAX_INSN_SIZE];
300 	int res, ret;
301 
302 	res = vc_fetch_insn_kernel(ctxt, buffer);
303 	if (res) {
304 		ctxt->fi.vector     = X86_TRAP_PF;
305 		ctxt->fi.error_code = X86_PF_INSTR;
306 		ctxt->fi.cr2        = ctxt->regs->ip;
307 		return ES_EXCEPTION;
308 	}
309 
310 	ret = insn_decode(&ctxt->insn, buffer, MAX_INSN_SIZE, INSN_MODE_64);
311 	if (ret < 0)
312 		return ES_DECODE_FAILED;
313 	else
314 		return ES_OK;
315 }
316 
317 static enum es_result vc_decode_insn(struct es_em_ctxt *ctxt)
318 {
319 	if (user_mode(ctxt->regs))
320 		return __vc_decode_user_insn(ctxt);
321 	else
322 		return __vc_decode_kern_insn(ctxt);
323 }
324 
325 static enum es_result vc_write_mem(struct es_em_ctxt *ctxt,
326 				   char *dst, char *buf, size_t size)
327 {
328 	unsigned long error_code = X86_PF_PROT | X86_PF_WRITE;
329 
330 	/*
331 	 * This function uses __put_user() independent of whether kernel or user
332 	 * memory is accessed. This works fine because __put_user() does no
333 	 * sanity checks of the pointer being accessed. All that it does is
334 	 * to report when the access failed.
335 	 *
336 	 * Also, this function runs in atomic context, so __put_user() is not
337 	 * allowed to sleep. The page-fault handler detects that it is running
338 	 * in atomic context and will not try to take mmap_sem and handle the
339 	 * fault, so additional pagefault_enable()/disable() calls are not
340 	 * needed.
341 	 *
342 	 * The access can't be done via copy_to_user() here because
343 	 * vc_write_mem() must not use string instructions to access unsafe
344 	 * memory. The reason is that MOVS is emulated by the #VC handler by
345 	 * splitting the move up into a read and a write and taking a nested #VC
346 	 * exception on whatever of them is the MMIO access. Using string
347 	 * instructions here would cause infinite nesting.
348 	 */
349 	switch (size) {
350 	case 1: {
351 		u8 d1;
352 		u8 __user *target = (u8 __user *)dst;
353 
354 		memcpy(&d1, buf, 1);
355 		if (__put_user(d1, target))
356 			goto fault;
357 		break;
358 	}
359 	case 2: {
360 		u16 d2;
361 		u16 __user *target = (u16 __user *)dst;
362 
363 		memcpy(&d2, buf, 2);
364 		if (__put_user(d2, target))
365 			goto fault;
366 		break;
367 	}
368 	case 4: {
369 		u32 d4;
370 		u32 __user *target = (u32 __user *)dst;
371 
372 		memcpy(&d4, buf, 4);
373 		if (__put_user(d4, target))
374 			goto fault;
375 		break;
376 	}
377 	case 8: {
378 		u64 d8;
379 		u64 __user *target = (u64 __user *)dst;
380 
381 		memcpy(&d8, buf, 8);
382 		if (__put_user(d8, target))
383 			goto fault;
384 		break;
385 	}
386 	default:
387 		WARN_ONCE(1, "%s: Invalid size: %zu\n", __func__, size);
388 		return ES_UNSUPPORTED;
389 	}
390 
391 	return ES_OK;
392 
393 fault:
394 	if (user_mode(ctxt->regs))
395 		error_code |= X86_PF_USER;
396 
397 	ctxt->fi.vector = X86_TRAP_PF;
398 	ctxt->fi.error_code = error_code;
399 	ctxt->fi.cr2 = (unsigned long)dst;
400 
401 	return ES_EXCEPTION;
402 }
403 
404 static enum es_result vc_read_mem(struct es_em_ctxt *ctxt,
405 				  char *src, char *buf, size_t size)
406 {
407 	unsigned long error_code = X86_PF_PROT;
408 
409 	/*
410 	 * This function uses __get_user() independent of whether kernel or user
411 	 * memory is accessed. This works fine because __get_user() does no
412 	 * sanity checks of the pointer being accessed. All that it does is
413 	 * to report when the access failed.
414 	 *
415 	 * Also, this function runs in atomic context, so __get_user() is not
416 	 * allowed to sleep. The page-fault handler detects that it is running
417 	 * in atomic context and will not try to take mmap_sem and handle the
418 	 * fault, so additional pagefault_enable()/disable() calls are not
419 	 * needed.
420 	 *
421 	 * The access can't be done via copy_from_user() here because
422 	 * vc_read_mem() must not use string instructions to access unsafe
423 	 * memory. The reason is that MOVS is emulated by the #VC handler by
424 	 * splitting the move up into a read and a write and taking a nested #VC
425 	 * exception on whatever of them is the MMIO access. Using string
426 	 * instructions here would cause infinite nesting.
427 	 */
428 	switch (size) {
429 	case 1: {
430 		u8 d1;
431 		u8 __user *s = (u8 __user *)src;
432 
433 		if (__get_user(d1, s))
434 			goto fault;
435 		memcpy(buf, &d1, 1);
436 		break;
437 	}
438 	case 2: {
439 		u16 d2;
440 		u16 __user *s = (u16 __user *)src;
441 
442 		if (__get_user(d2, s))
443 			goto fault;
444 		memcpy(buf, &d2, 2);
445 		break;
446 	}
447 	case 4: {
448 		u32 d4;
449 		u32 __user *s = (u32 __user *)src;
450 
451 		if (__get_user(d4, s))
452 			goto fault;
453 		memcpy(buf, &d4, 4);
454 		break;
455 	}
456 	case 8: {
457 		u64 d8;
458 		u64 __user *s = (u64 __user *)src;
459 		if (__get_user(d8, s))
460 			goto fault;
461 		memcpy(buf, &d8, 8);
462 		break;
463 	}
464 	default:
465 		WARN_ONCE(1, "%s: Invalid size: %zu\n", __func__, size);
466 		return ES_UNSUPPORTED;
467 	}
468 
469 	return ES_OK;
470 
471 fault:
472 	if (user_mode(ctxt->regs))
473 		error_code |= X86_PF_USER;
474 
475 	ctxt->fi.vector = X86_TRAP_PF;
476 	ctxt->fi.error_code = error_code;
477 	ctxt->fi.cr2 = (unsigned long)src;
478 
479 	return ES_EXCEPTION;
480 }
481 
482 static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt *ctxt,
483 					   unsigned long vaddr, phys_addr_t *paddr)
484 {
485 	unsigned long va = (unsigned long)vaddr;
486 	unsigned int level;
487 	phys_addr_t pa;
488 	pgd_t *pgd;
489 	pte_t *pte;
490 
491 	pgd = __va(read_cr3_pa());
492 	pgd = &pgd[pgd_index(va)];
493 	pte = lookup_address_in_pgd(pgd, va, &level);
494 	if (!pte) {
495 		ctxt->fi.vector     = X86_TRAP_PF;
496 		ctxt->fi.cr2        = vaddr;
497 		ctxt->fi.error_code = 0;
498 
499 		if (user_mode(ctxt->regs))
500 			ctxt->fi.error_code |= X86_PF_USER;
501 
502 		return ES_EXCEPTION;
503 	}
504 
505 	if (WARN_ON_ONCE(pte_val(*pte) & _PAGE_ENC))
506 		/* Emulated MMIO to/from encrypted memory not supported */
507 		return ES_UNSUPPORTED;
508 
509 	pa = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
510 	pa |= va & ~page_level_mask(level);
511 
512 	*paddr = pa;
513 
514 	return ES_OK;
515 }
516 
517 /* Include code shared with pre-decompression boot stage */
518 #include "sev-shared.c"
519 
520 static noinstr void __sev_put_ghcb(struct ghcb_state *state)
521 {
522 	struct sev_es_runtime_data *data;
523 	struct ghcb *ghcb;
524 
525 	WARN_ON(!irqs_disabled());
526 
527 	data = this_cpu_read(runtime_data);
528 	ghcb = &data->ghcb_page;
529 
530 	if (state->ghcb) {
531 		/* Restore GHCB from Backup */
532 		*ghcb = *state->ghcb;
533 		data->backup_ghcb_active = false;
534 		state->ghcb = NULL;
535 	} else {
536 		/*
537 		 * Invalidate the GHCB so a VMGEXIT instruction issued
538 		 * from userspace won't appear to be valid.
539 		 */
540 		vc_ghcb_invalidate(ghcb);
541 		data->ghcb_active = false;
542 	}
543 }
544 
545 void noinstr __sev_es_nmi_complete(void)
546 {
547 	struct ghcb_state state;
548 	struct ghcb *ghcb;
549 
550 	ghcb = __sev_get_ghcb(&state);
551 
552 	vc_ghcb_invalidate(ghcb);
553 	ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_NMI_COMPLETE);
554 	ghcb_set_sw_exit_info_1(ghcb, 0);
555 	ghcb_set_sw_exit_info_2(ghcb, 0);
556 
557 	sev_es_wr_ghcb_msr(__pa_nodebug(ghcb));
558 	VMGEXIT();
559 
560 	__sev_put_ghcb(&state);
561 }
562 
563 static u64 __init get_secrets_page(void)
564 {
565 	u64 pa_data = boot_params.cc_blob_address;
566 	struct cc_blob_sev_info info;
567 	void *map;
568 
569 	/*
570 	 * The CC blob contains the address of the secrets page, check if the
571 	 * blob is present.
572 	 */
573 	if (!pa_data)
574 		return 0;
575 
576 	map = early_memremap(pa_data, sizeof(info));
577 	if (!map) {
578 		pr_err("Unable to locate SNP secrets page: failed to map the Confidential Computing blob.\n");
579 		return 0;
580 	}
581 	memcpy(&info, map, sizeof(info));
582 	early_memunmap(map, sizeof(info));
583 
584 	/* smoke-test the secrets page passed */
585 	if (!info.secrets_phys || info.secrets_len != PAGE_SIZE)
586 		return 0;
587 
588 	return info.secrets_phys;
589 }
590 
591 static u64 __init get_snp_jump_table_addr(void)
592 {
593 	struct snp_secrets_page_layout *layout;
594 	void __iomem *mem;
595 	u64 pa, addr;
596 
597 	pa = get_secrets_page();
598 	if (!pa)
599 		return 0;
600 
601 	mem = ioremap_encrypted(pa, PAGE_SIZE);
602 	if (!mem) {
603 		pr_err("Unable to locate AP jump table address: failed to map the SNP secrets page.\n");
604 		return 0;
605 	}
606 
607 	layout = (__force struct snp_secrets_page_layout *)mem;
608 
609 	addr = layout->os_area.ap_jump_table_pa;
610 	iounmap(mem);
611 
612 	return addr;
613 }
614 
615 static u64 __init get_jump_table_addr(void)
616 {
617 	struct ghcb_state state;
618 	unsigned long flags;
619 	struct ghcb *ghcb;
620 	u64 ret = 0;
621 
622 	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
623 		return get_snp_jump_table_addr();
624 
625 	local_irq_save(flags);
626 
627 	ghcb = __sev_get_ghcb(&state);
628 
629 	vc_ghcb_invalidate(ghcb);
630 	ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_JUMP_TABLE);
631 	ghcb_set_sw_exit_info_1(ghcb, SVM_VMGEXIT_GET_AP_JUMP_TABLE);
632 	ghcb_set_sw_exit_info_2(ghcb, 0);
633 
634 	sev_es_wr_ghcb_msr(__pa(ghcb));
635 	VMGEXIT();
636 
637 	if (ghcb_sw_exit_info_1_is_valid(ghcb) &&
638 	    ghcb_sw_exit_info_2_is_valid(ghcb))
639 		ret = ghcb->save.sw_exit_info_2;
640 
641 	__sev_put_ghcb(&state);
642 
643 	local_irq_restore(flags);
644 
645 	return ret;
646 }
647 
648 static void pvalidate_pages(unsigned long vaddr, unsigned int npages, bool validate)
649 {
650 	unsigned long vaddr_end;
651 	int rc;
652 
653 	vaddr = vaddr & PAGE_MASK;
654 	vaddr_end = vaddr + (npages << PAGE_SHIFT);
655 
656 	while (vaddr < vaddr_end) {
657 		rc = pvalidate(vaddr, RMP_PG_SIZE_4K, validate);
658 		if (WARN(rc, "Failed to validate address 0x%lx ret %d", vaddr, rc))
659 			sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PVALIDATE);
660 
661 		vaddr = vaddr + PAGE_SIZE;
662 	}
663 }
664 
665 static void __init early_set_pages_state(unsigned long paddr, unsigned int npages, enum psc_op op)
666 {
667 	unsigned long paddr_end;
668 	u64 val;
669 
670 	paddr = paddr & PAGE_MASK;
671 	paddr_end = paddr + (npages << PAGE_SHIFT);
672 
673 	while (paddr < paddr_end) {
674 		/*
675 		 * Use the MSR protocol because this function can be called before
676 		 * the GHCB is established.
677 		 */
678 		sev_es_wr_ghcb_msr(GHCB_MSR_PSC_REQ_GFN(paddr >> PAGE_SHIFT, op));
679 		VMGEXIT();
680 
681 		val = sev_es_rd_ghcb_msr();
682 
683 		if (WARN(GHCB_RESP_CODE(val) != GHCB_MSR_PSC_RESP,
684 			 "Wrong PSC response code: 0x%x\n",
685 			 (unsigned int)GHCB_RESP_CODE(val)))
686 			goto e_term;
687 
688 		if (WARN(GHCB_MSR_PSC_RESP_VAL(val),
689 			 "Failed to change page state to '%s' paddr 0x%lx error 0x%llx\n",
690 			 op == SNP_PAGE_STATE_PRIVATE ? "private" : "shared",
691 			 paddr, GHCB_MSR_PSC_RESP_VAL(val)))
692 			goto e_term;
693 
694 		paddr = paddr + PAGE_SIZE;
695 	}
696 
697 	return;
698 
699 e_term:
700 	sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PSC);
701 }
702 
703 void __init early_snp_set_memory_private(unsigned long vaddr, unsigned long paddr,
704 					 unsigned int npages)
705 {
706 	/*
707 	 * This can be invoked in early boot while running identity mapped, so
708 	 * use an open coded check for SNP instead of using cc_platform_has().
709 	 * This eliminates worries about jump tables or checking boot_cpu_data
710 	 * in the cc_platform_has() function.
711 	 */
712 	if (!(sev_status & MSR_AMD64_SEV_SNP_ENABLED))
713 		return;
714 
715 	 /*
716 	  * Ask the hypervisor to mark the memory pages as private in the RMP
717 	  * table.
718 	  */
719 	early_set_pages_state(paddr, npages, SNP_PAGE_STATE_PRIVATE);
720 
721 	/* Validate the memory pages after they've been added in the RMP table. */
722 	pvalidate_pages(vaddr, npages, true);
723 }
724 
725 void __init early_snp_set_memory_shared(unsigned long vaddr, unsigned long paddr,
726 					unsigned int npages)
727 {
728 	/*
729 	 * This can be invoked in early boot while running identity mapped, so
730 	 * use an open coded check for SNP instead of using cc_platform_has().
731 	 * This eliminates worries about jump tables or checking boot_cpu_data
732 	 * in the cc_platform_has() function.
733 	 */
734 	if (!(sev_status & MSR_AMD64_SEV_SNP_ENABLED))
735 		return;
736 
737 	/* Invalidate the memory pages before they are marked shared in the RMP table. */
738 	pvalidate_pages(vaddr, npages, false);
739 
740 	 /* Ask hypervisor to mark the memory pages shared in the RMP table. */
741 	early_set_pages_state(paddr, npages, SNP_PAGE_STATE_SHARED);
742 }
743 
744 void __init snp_prep_memory(unsigned long paddr, unsigned int sz, enum psc_op op)
745 {
746 	unsigned long vaddr, npages;
747 
748 	vaddr = (unsigned long)__va(paddr);
749 	npages = PAGE_ALIGN(sz) >> PAGE_SHIFT;
750 
751 	if (op == SNP_PAGE_STATE_PRIVATE)
752 		early_snp_set_memory_private(vaddr, paddr, npages);
753 	else if (op == SNP_PAGE_STATE_SHARED)
754 		early_snp_set_memory_shared(vaddr, paddr, npages);
755 	else
756 		WARN(1, "invalid memory op %d\n", op);
757 }
758 
759 static int vmgexit_psc(struct snp_psc_desc *desc)
760 {
761 	int cur_entry, end_entry, ret = 0;
762 	struct snp_psc_desc *data;
763 	struct ghcb_state state;
764 	struct es_em_ctxt ctxt;
765 	unsigned long flags;
766 	struct ghcb *ghcb;
767 
768 	/*
769 	 * __sev_get_ghcb() needs to run with IRQs disabled because it is using
770 	 * a per-CPU GHCB.
771 	 */
772 	local_irq_save(flags);
773 
774 	ghcb = __sev_get_ghcb(&state);
775 	if (!ghcb) {
776 		ret = 1;
777 		goto out_unlock;
778 	}
779 
780 	/* Copy the input desc into GHCB shared buffer */
781 	data = (struct snp_psc_desc *)ghcb->shared_buffer;
782 	memcpy(ghcb->shared_buffer, desc, min_t(int, GHCB_SHARED_BUF_SIZE, sizeof(*desc)));
783 
784 	/*
785 	 * As per the GHCB specification, the hypervisor can resume the guest
786 	 * before processing all the entries. Check whether all the entries
787 	 * are processed. If not, then keep retrying. Note, the hypervisor
788 	 * will update the data memory directly to indicate the status, so
789 	 * reference the data->hdr everywhere.
790 	 *
791 	 * The strategy here is to wait for the hypervisor to change the page
792 	 * state in the RMP table before guest accesses the memory pages. If the
793 	 * page state change was not successful, then later memory access will
794 	 * result in a crash.
795 	 */
796 	cur_entry = data->hdr.cur_entry;
797 	end_entry = data->hdr.end_entry;
798 
799 	while (data->hdr.cur_entry <= data->hdr.end_entry) {
800 		ghcb_set_sw_scratch(ghcb, (u64)__pa(data));
801 
802 		/* This will advance the shared buffer data points to. */
803 		ret = sev_es_ghcb_hv_call(ghcb, &ctxt, SVM_VMGEXIT_PSC, 0, 0);
804 
805 		/*
806 		 * Page State Change VMGEXIT can pass error code through
807 		 * exit_info_2.
808 		 */
809 		if (WARN(ret || ghcb->save.sw_exit_info_2,
810 			 "SNP: PSC failed ret=%d exit_info_2=%llx\n",
811 			 ret, ghcb->save.sw_exit_info_2)) {
812 			ret = 1;
813 			goto out;
814 		}
815 
816 		/* Verify that reserved bit is not set */
817 		if (WARN(data->hdr.reserved, "Reserved bit is set in the PSC header\n")) {
818 			ret = 1;
819 			goto out;
820 		}
821 
822 		/*
823 		 * Sanity check that entry processing is not going backwards.
824 		 * This will happen only if hypervisor is tricking us.
825 		 */
826 		if (WARN(data->hdr.end_entry > end_entry || cur_entry > data->hdr.cur_entry,
827 "SNP: PSC processing going backward, end_entry %d (got %d) cur_entry %d (got %d)\n",
828 			 end_entry, data->hdr.end_entry, cur_entry, data->hdr.cur_entry)) {
829 			ret = 1;
830 			goto out;
831 		}
832 	}
833 
834 out:
835 	__sev_put_ghcb(&state);
836 
837 out_unlock:
838 	local_irq_restore(flags);
839 
840 	return ret;
841 }
842 
843 static void __set_pages_state(struct snp_psc_desc *data, unsigned long vaddr,
844 			      unsigned long vaddr_end, int op)
845 {
846 	struct psc_hdr *hdr;
847 	struct psc_entry *e;
848 	unsigned long pfn;
849 	int i;
850 
851 	hdr = &data->hdr;
852 	e = data->entries;
853 
854 	memset(data, 0, sizeof(*data));
855 	i = 0;
856 
857 	while (vaddr < vaddr_end) {
858 		if (is_vmalloc_addr((void *)vaddr))
859 			pfn = vmalloc_to_pfn((void *)vaddr);
860 		else
861 			pfn = __pa(vaddr) >> PAGE_SHIFT;
862 
863 		e->gfn = pfn;
864 		e->operation = op;
865 		hdr->end_entry = i;
866 
867 		/*
868 		 * Current SNP implementation doesn't keep track of the RMP page
869 		 * size so use 4K for simplicity.
870 		 */
871 		e->pagesize = RMP_PG_SIZE_4K;
872 
873 		vaddr = vaddr + PAGE_SIZE;
874 		e++;
875 		i++;
876 	}
877 
878 	if (vmgexit_psc(data))
879 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PSC);
880 }
881 
882 static void set_pages_state(unsigned long vaddr, unsigned int npages, int op)
883 {
884 	unsigned long vaddr_end, next_vaddr;
885 	struct snp_psc_desc *desc;
886 
887 	desc = kmalloc(sizeof(*desc), GFP_KERNEL_ACCOUNT);
888 	if (!desc)
889 		panic("SNP: failed to allocate memory for PSC descriptor\n");
890 
891 	vaddr = vaddr & PAGE_MASK;
892 	vaddr_end = vaddr + (npages << PAGE_SHIFT);
893 
894 	while (vaddr < vaddr_end) {
895 		/* Calculate the last vaddr that fits in one struct snp_psc_desc. */
896 		next_vaddr = min_t(unsigned long, vaddr_end,
897 				   (VMGEXIT_PSC_MAX_ENTRY * PAGE_SIZE) + vaddr);
898 
899 		__set_pages_state(desc, vaddr, next_vaddr, op);
900 
901 		vaddr = next_vaddr;
902 	}
903 
904 	kfree(desc);
905 }
906 
907 void snp_set_memory_shared(unsigned long vaddr, unsigned int npages)
908 {
909 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
910 		return;
911 
912 	pvalidate_pages(vaddr, npages, false);
913 
914 	set_pages_state(vaddr, npages, SNP_PAGE_STATE_SHARED);
915 }
916 
917 void snp_set_memory_private(unsigned long vaddr, unsigned int npages)
918 {
919 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
920 		return;
921 
922 	set_pages_state(vaddr, npages, SNP_PAGE_STATE_PRIVATE);
923 
924 	pvalidate_pages(vaddr, npages, true);
925 }
926 
927 static int snp_set_vmsa(void *va, bool vmsa)
928 {
929 	u64 attrs;
930 
931 	/*
932 	 * Running at VMPL0 allows the kernel to change the VMSA bit for a page
933 	 * using the RMPADJUST instruction. However, for the instruction to
934 	 * succeed it must target the permissions of a lesser privileged
935 	 * (higher numbered) VMPL level, so use VMPL1 (refer to the RMPADJUST
936 	 * instruction in the AMD64 APM Volume 3).
937 	 */
938 	attrs = 1;
939 	if (vmsa)
940 		attrs |= RMPADJUST_VMSA_PAGE_BIT;
941 
942 	return rmpadjust((unsigned long)va, RMP_PG_SIZE_4K, attrs);
943 }
944 
945 #define __ATTR_BASE		(SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK)
946 #define INIT_CS_ATTRIBS		(__ATTR_BASE | SVM_SELECTOR_READ_MASK | SVM_SELECTOR_CODE_MASK)
947 #define INIT_DS_ATTRIBS		(__ATTR_BASE | SVM_SELECTOR_WRITE_MASK)
948 
949 #define INIT_LDTR_ATTRIBS	(SVM_SELECTOR_P_MASK | 2)
950 #define INIT_TR_ATTRIBS		(SVM_SELECTOR_P_MASK | 3)
951 
952 static void *snp_alloc_vmsa_page(void)
953 {
954 	struct page *p;
955 
956 	/*
957 	 * Allocate VMSA page to work around the SNP erratum where the CPU will
958 	 * incorrectly signal an RMP violation #PF if a large page (2MB or 1GB)
959 	 * collides with the RMP entry of VMSA page. The recommended workaround
960 	 * is to not use a large page.
961 	 *
962 	 * Allocate an 8k page which is also 8k-aligned.
963 	 */
964 	p = alloc_pages(GFP_KERNEL_ACCOUNT | __GFP_ZERO, 1);
965 	if (!p)
966 		return NULL;
967 
968 	split_page(p, 1);
969 
970 	/* Free the first 4k. This page may be 2M/1G aligned and cannot be used. */
971 	__free_page(p);
972 
973 	return page_address(p + 1);
974 }
975 
976 static void snp_cleanup_vmsa(struct sev_es_save_area *vmsa)
977 {
978 	int err;
979 
980 	err = snp_set_vmsa(vmsa, false);
981 	if (err)
982 		pr_err("clear VMSA page failed (%u), leaking page\n", err);
983 	else
984 		free_page((unsigned long)vmsa);
985 }
986 
987 static int wakeup_cpu_via_vmgexit(int apic_id, unsigned long start_ip)
988 {
989 	struct sev_es_save_area *cur_vmsa, *vmsa;
990 	struct ghcb_state state;
991 	unsigned long flags;
992 	struct ghcb *ghcb;
993 	u8 sipi_vector;
994 	int cpu, ret;
995 	u64 cr4;
996 
997 	/*
998 	 * The hypervisor SNP feature support check has happened earlier, just check
999 	 * the AP_CREATION one here.
1000 	 */
1001 	if (!(sev_hv_features & GHCB_HV_FT_SNP_AP_CREATION))
1002 		return -EOPNOTSUPP;
1003 
1004 	/*
1005 	 * Verify the desired start IP against the known trampoline start IP
1006 	 * to catch any future new trampolines that may be introduced that
1007 	 * would require a new protected guest entry point.
1008 	 */
1009 	if (WARN_ONCE(start_ip != real_mode_header->trampoline_start,
1010 		      "Unsupported SNP start_ip: %lx\n", start_ip))
1011 		return -EINVAL;
1012 
1013 	/* Override start_ip with known protected guest start IP */
1014 	start_ip = real_mode_header->sev_es_trampoline_start;
1015 
1016 	/* Find the logical CPU for the APIC ID */
1017 	for_each_present_cpu(cpu) {
1018 		if (arch_match_cpu_phys_id(cpu, apic_id))
1019 			break;
1020 	}
1021 	if (cpu >= nr_cpu_ids)
1022 		return -EINVAL;
1023 
1024 	cur_vmsa = per_cpu(sev_vmsa, cpu);
1025 
1026 	/*
1027 	 * A new VMSA is created each time because there is no guarantee that
1028 	 * the current VMSA is the kernels or that the vCPU is not running. If
1029 	 * an attempt was done to use the current VMSA with a running vCPU, a
1030 	 * #VMEXIT of that vCPU would wipe out all of the settings being done
1031 	 * here.
1032 	 */
1033 	vmsa = (struct sev_es_save_area *)snp_alloc_vmsa_page();
1034 	if (!vmsa)
1035 		return -ENOMEM;
1036 
1037 	/* CR4 should maintain the MCE value */
1038 	cr4 = native_read_cr4() & X86_CR4_MCE;
1039 
1040 	/* Set the CS value based on the start_ip converted to a SIPI vector */
1041 	sipi_vector		= (start_ip >> 12);
1042 	vmsa->cs.base		= sipi_vector << 12;
1043 	vmsa->cs.limit		= AP_INIT_CS_LIMIT;
1044 	vmsa->cs.attrib		= INIT_CS_ATTRIBS;
1045 	vmsa->cs.selector	= sipi_vector << 8;
1046 
1047 	/* Set the RIP value based on start_ip */
1048 	vmsa->rip		= start_ip & 0xfff;
1049 
1050 	/* Set AP INIT defaults as documented in the APM */
1051 	vmsa->ds.limit		= AP_INIT_DS_LIMIT;
1052 	vmsa->ds.attrib		= INIT_DS_ATTRIBS;
1053 	vmsa->es		= vmsa->ds;
1054 	vmsa->fs		= vmsa->ds;
1055 	vmsa->gs		= vmsa->ds;
1056 	vmsa->ss		= vmsa->ds;
1057 
1058 	vmsa->gdtr.limit	= AP_INIT_GDTR_LIMIT;
1059 	vmsa->ldtr.limit	= AP_INIT_LDTR_LIMIT;
1060 	vmsa->ldtr.attrib	= INIT_LDTR_ATTRIBS;
1061 	vmsa->idtr.limit	= AP_INIT_IDTR_LIMIT;
1062 	vmsa->tr.limit		= AP_INIT_TR_LIMIT;
1063 	vmsa->tr.attrib		= INIT_TR_ATTRIBS;
1064 
1065 	vmsa->cr4		= cr4;
1066 	vmsa->cr0		= AP_INIT_CR0_DEFAULT;
1067 	vmsa->dr7		= DR7_RESET_VALUE;
1068 	vmsa->dr6		= AP_INIT_DR6_DEFAULT;
1069 	vmsa->rflags		= AP_INIT_RFLAGS_DEFAULT;
1070 	vmsa->g_pat		= AP_INIT_GPAT_DEFAULT;
1071 	vmsa->xcr0		= AP_INIT_XCR0_DEFAULT;
1072 	vmsa->mxcsr		= AP_INIT_MXCSR_DEFAULT;
1073 	vmsa->x87_ftw		= AP_INIT_X87_FTW_DEFAULT;
1074 	vmsa->x87_fcw		= AP_INIT_X87_FCW_DEFAULT;
1075 
1076 	/* SVME must be set. */
1077 	vmsa->efer		= EFER_SVME;
1078 
1079 	/*
1080 	 * Set the SNP-specific fields for this VMSA:
1081 	 *   VMPL level
1082 	 *   SEV_FEATURES (matches the SEV STATUS MSR right shifted 2 bits)
1083 	 */
1084 	vmsa->vmpl		= 0;
1085 	vmsa->sev_features	= sev_status >> 2;
1086 
1087 	/* Switch the page over to a VMSA page now that it is initialized */
1088 	ret = snp_set_vmsa(vmsa, true);
1089 	if (ret) {
1090 		pr_err("set VMSA page failed (%u)\n", ret);
1091 		free_page((unsigned long)vmsa);
1092 
1093 		return -EINVAL;
1094 	}
1095 
1096 	/* Issue VMGEXIT AP Creation NAE event */
1097 	local_irq_save(flags);
1098 
1099 	ghcb = __sev_get_ghcb(&state);
1100 
1101 	vc_ghcb_invalidate(ghcb);
1102 	ghcb_set_rax(ghcb, vmsa->sev_features);
1103 	ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_CREATION);
1104 	ghcb_set_sw_exit_info_1(ghcb, ((u64)apic_id << 32) | SVM_VMGEXIT_AP_CREATE);
1105 	ghcb_set_sw_exit_info_2(ghcb, __pa(vmsa));
1106 
1107 	sev_es_wr_ghcb_msr(__pa(ghcb));
1108 	VMGEXIT();
1109 
1110 	if (!ghcb_sw_exit_info_1_is_valid(ghcb) ||
1111 	    lower_32_bits(ghcb->save.sw_exit_info_1)) {
1112 		pr_err("SNP AP Creation error\n");
1113 		ret = -EINVAL;
1114 	}
1115 
1116 	__sev_put_ghcb(&state);
1117 
1118 	local_irq_restore(flags);
1119 
1120 	/* Perform cleanup if there was an error */
1121 	if (ret) {
1122 		snp_cleanup_vmsa(vmsa);
1123 		vmsa = NULL;
1124 	}
1125 
1126 	/* Free up any previous VMSA page */
1127 	if (cur_vmsa)
1128 		snp_cleanup_vmsa(cur_vmsa);
1129 
1130 	/* Record the current VMSA page */
1131 	per_cpu(sev_vmsa, cpu) = vmsa;
1132 
1133 	return ret;
1134 }
1135 
1136 void snp_set_wakeup_secondary_cpu(void)
1137 {
1138 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
1139 		return;
1140 
1141 	/*
1142 	 * Always set this override if SNP is enabled. This makes it the
1143 	 * required method to start APs under SNP. If the hypervisor does
1144 	 * not support AP creation, then no APs will be started.
1145 	 */
1146 	apic->wakeup_secondary_cpu = wakeup_cpu_via_vmgexit;
1147 }
1148 
1149 int __init sev_es_setup_ap_jump_table(struct real_mode_header *rmh)
1150 {
1151 	u16 startup_cs, startup_ip;
1152 	phys_addr_t jump_table_pa;
1153 	u64 jump_table_addr;
1154 	u16 __iomem *jump_table;
1155 
1156 	jump_table_addr = get_jump_table_addr();
1157 
1158 	/* On UP guests there is no jump table so this is not a failure */
1159 	if (!jump_table_addr)
1160 		return 0;
1161 
1162 	/* Check if AP Jump Table is page-aligned */
1163 	if (jump_table_addr & ~PAGE_MASK)
1164 		return -EINVAL;
1165 
1166 	jump_table_pa = jump_table_addr & PAGE_MASK;
1167 
1168 	startup_cs = (u16)(rmh->trampoline_start >> 4);
1169 	startup_ip = (u16)(rmh->sev_es_trampoline_start -
1170 			   rmh->trampoline_start);
1171 
1172 	jump_table = ioremap_encrypted(jump_table_pa, PAGE_SIZE);
1173 	if (!jump_table)
1174 		return -EIO;
1175 
1176 	writew(startup_ip, &jump_table[0]);
1177 	writew(startup_cs, &jump_table[1]);
1178 
1179 	iounmap(jump_table);
1180 
1181 	return 0;
1182 }
1183 
1184 /*
1185  * This is needed by the OVMF UEFI firmware which will use whatever it finds in
1186  * the GHCB MSR as its GHCB to talk to the hypervisor. So make sure the per-cpu
1187  * runtime GHCBs used by the kernel are also mapped in the EFI page-table.
1188  */
1189 int __init sev_es_efi_map_ghcbs(pgd_t *pgd)
1190 {
1191 	struct sev_es_runtime_data *data;
1192 	unsigned long address, pflags;
1193 	int cpu;
1194 	u64 pfn;
1195 
1196 	if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
1197 		return 0;
1198 
1199 	pflags = _PAGE_NX | _PAGE_RW;
1200 
1201 	for_each_possible_cpu(cpu) {
1202 		data = per_cpu(runtime_data, cpu);
1203 
1204 		address = __pa(&data->ghcb_page);
1205 		pfn = address >> PAGE_SHIFT;
1206 
1207 		if (kernel_map_pages_in_pgd(pgd, pfn, address, 1, pflags))
1208 			return 1;
1209 	}
1210 
1211 	return 0;
1212 }
1213 
1214 static enum es_result vc_handle_msr(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
1215 {
1216 	struct pt_regs *regs = ctxt->regs;
1217 	enum es_result ret;
1218 	u64 exit_info_1;
1219 
1220 	/* Is it a WRMSR? */
1221 	exit_info_1 = (ctxt->insn.opcode.bytes[1] == 0x30) ? 1 : 0;
1222 
1223 	ghcb_set_rcx(ghcb, regs->cx);
1224 	if (exit_info_1) {
1225 		ghcb_set_rax(ghcb, regs->ax);
1226 		ghcb_set_rdx(ghcb, regs->dx);
1227 	}
1228 
1229 	ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_MSR, exit_info_1, 0);
1230 
1231 	if ((ret == ES_OK) && (!exit_info_1)) {
1232 		regs->ax = ghcb->save.rax;
1233 		regs->dx = ghcb->save.rdx;
1234 	}
1235 
1236 	return ret;
1237 }
1238 
1239 static void snp_register_per_cpu_ghcb(void)
1240 {
1241 	struct sev_es_runtime_data *data;
1242 	struct ghcb *ghcb;
1243 
1244 	data = this_cpu_read(runtime_data);
1245 	ghcb = &data->ghcb_page;
1246 
1247 	snp_register_ghcb_early(__pa(ghcb));
1248 }
1249 
1250 void setup_ghcb(void)
1251 {
1252 	if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
1253 		return;
1254 
1255 	/* First make sure the hypervisor talks a supported protocol. */
1256 	if (!sev_es_negotiate_protocol())
1257 		sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
1258 
1259 	/*
1260 	 * Check whether the runtime #VC exception handler is active. It uses
1261 	 * the per-CPU GHCB page which is set up by sev_es_init_vc_handling().
1262 	 *
1263 	 * If SNP is active, register the per-CPU GHCB page so that the runtime
1264 	 * exception handler can use it.
1265 	 */
1266 	if (initial_vc_handler == (unsigned long)kernel_exc_vmm_communication) {
1267 		if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
1268 			snp_register_per_cpu_ghcb();
1269 
1270 		return;
1271 	}
1272 
1273 	/*
1274 	 * Clear the boot_ghcb. The first exception comes in before the bss
1275 	 * section is cleared.
1276 	 */
1277 	memset(&boot_ghcb_page, 0, PAGE_SIZE);
1278 
1279 	/* Alright - Make the boot-ghcb public */
1280 	boot_ghcb = &boot_ghcb_page;
1281 
1282 	/* SNP guest requires that GHCB GPA must be registered. */
1283 	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
1284 		snp_register_ghcb_early(__pa(&boot_ghcb_page));
1285 }
1286 
1287 #ifdef CONFIG_HOTPLUG_CPU
1288 static void sev_es_ap_hlt_loop(void)
1289 {
1290 	struct ghcb_state state;
1291 	struct ghcb *ghcb;
1292 
1293 	ghcb = __sev_get_ghcb(&state);
1294 
1295 	while (true) {
1296 		vc_ghcb_invalidate(ghcb);
1297 		ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_HLT_LOOP);
1298 		ghcb_set_sw_exit_info_1(ghcb, 0);
1299 		ghcb_set_sw_exit_info_2(ghcb, 0);
1300 
1301 		sev_es_wr_ghcb_msr(__pa(ghcb));
1302 		VMGEXIT();
1303 
1304 		/* Wakeup signal? */
1305 		if (ghcb_sw_exit_info_2_is_valid(ghcb) &&
1306 		    ghcb->save.sw_exit_info_2)
1307 			break;
1308 	}
1309 
1310 	__sev_put_ghcb(&state);
1311 }
1312 
1313 /*
1314  * Play_dead handler when running under SEV-ES. This is needed because
1315  * the hypervisor can't deliver an SIPI request to restart the AP.
1316  * Instead the kernel has to issue a VMGEXIT to halt the VCPU until the
1317  * hypervisor wakes it up again.
1318  */
1319 static void sev_es_play_dead(void)
1320 {
1321 	play_dead_common();
1322 
1323 	/* IRQs now disabled */
1324 
1325 	sev_es_ap_hlt_loop();
1326 
1327 	/*
1328 	 * If we get here, the VCPU was woken up again. Jump to CPU
1329 	 * startup code to get it back online.
1330 	 */
1331 	start_cpu0();
1332 }
1333 #else  /* CONFIG_HOTPLUG_CPU */
1334 #define sev_es_play_dead	native_play_dead
1335 #endif /* CONFIG_HOTPLUG_CPU */
1336 
1337 #ifdef CONFIG_SMP
1338 static void __init sev_es_setup_play_dead(void)
1339 {
1340 	smp_ops.play_dead = sev_es_play_dead;
1341 }
1342 #else
1343 static inline void sev_es_setup_play_dead(void) { }
1344 #endif
1345 
1346 static void __init alloc_runtime_data(int cpu)
1347 {
1348 	struct sev_es_runtime_data *data;
1349 
1350 	data = memblock_alloc(sizeof(*data), PAGE_SIZE);
1351 	if (!data)
1352 		panic("Can't allocate SEV-ES runtime data");
1353 
1354 	per_cpu(runtime_data, cpu) = data;
1355 }
1356 
1357 static void __init init_ghcb(int cpu)
1358 {
1359 	struct sev_es_runtime_data *data;
1360 	int err;
1361 
1362 	data = per_cpu(runtime_data, cpu);
1363 
1364 	err = early_set_memory_decrypted((unsigned long)&data->ghcb_page,
1365 					 sizeof(data->ghcb_page));
1366 	if (err)
1367 		panic("Can't map GHCBs unencrypted");
1368 
1369 	memset(&data->ghcb_page, 0, sizeof(data->ghcb_page));
1370 
1371 	data->ghcb_active = false;
1372 	data->backup_ghcb_active = false;
1373 }
1374 
1375 void __init sev_es_init_vc_handling(void)
1376 {
1377 	int cpu;
1378 
1379 	BUILD_BUG_ON(offsetof(struct sev_es_runtime_data, ghcb_page) % PAGE_SIZE);
1380 
1381 	if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
1382 		return;
1383 
1384 	if (!sev_es_check_cpu_features())
1385 		panic("SEV-ES CPU Features missing");
1386 
1387 	/*
1388 	 * SNP is supported in v2 of the GHCB spec which mandates support for HV
1389 	 * features.
1390 	 */
1391 	if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
1392 		sev_hv_features = get_hv_features();
1393 
1394 		if (!(sev_hv_features & GHCB_HV_FT_SNP))
1395 			sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SNP_UNSUPPORTED);
1396 	}
1397 
1398 	/* Enable SEV-ES special handling */
1399 	static_branch_enable(&sev_es_enable_key);
1400 
1401 	/* Initialize per-cpu GHCB pages */
1402 	for_each_possible_cpu(cpu) {
1403 		alloc_runtime_data(cpu);
1404 		init_ghcb(cpu);
1405 	}
1406 
1407 	sev_es_setup_play_dead();
1408 
1409 	/* Secondary CPUs use the runtime #VC handler */
1410 	initial_vc_handler = (unsigned long)kernel_exc_vmm_communication;
1411 }
1412 
1413 static void __init vc_early_forward_exception(struct es_em_ctxt *ctxt)
1414 {
1415 	int trapnr = ctxt->fi.vector;
1416 
1417 	if (trapnr == X86_TRAP_PF)
1418 		native_write_cr2(ctxt->fi.cr2);
1419 
1420 	ctxt->regs->orig_ax = ctxt->fi.error_code;
1421 	do_early_exception(ctxt->regs, trapnr);
1422 }
1423 
1424 static long *vc_insn_get_rm(struct es_em_ctxt *ctxt)
1425 {
1426 	long *reg_array;
1427 	int offset;
1428 
1429 	reg_array = (long *)ctxt->regs;
1430 	offset    = insn_get_modrm_rm_off(&ctxt->insn, ctxt->regs);
1431 
1432 	if (offset < 0)
1433 		return NULL;
1434 
1435 	offset /= sizeof(long);
1436 
1437 	return reg_array + offset;
1438 }
1439 static enum es_result vc_do_mmio(struct ghcb *ghcb, struct es_em_ctxt *ctxt,
1440 				 unsigned int bytes, bool read)
1441 {
1442 	u64 exit_code, exit_info_1, exit_info_2;
1443 	unsigned long ghcb_pa = __pa(ghcb);
1444 	enum es_result res;
1445 	phys_addr_t paddr;
1446 	void __user *ref;
1447 
1448 	ref = insn_get_addr_ref(&ctxt->insn, ctxt->regs);
1449 	if (ref == (void __user *)-1L)
1450 		return ES_UNSUPPORTED;
1451 
1452 	exit_code = read ? SVM_VMGEXIT_MMIO_READ : SVM_VMGEXIT_MMIO_WRITE;
1453 
1454 	res = vc_slow_virt_to_phys(ghcb, ctxt, (unsigned long)ref, &paddr);
1455 	if (res != ES_OK) {
1456 		if (res == ES_EXCEPTION && !read)
1457 			ctxt->fi.error_code |= X86_PF_WRITE;
1458 
1459 		return res;
1460 	}
1461 
1462 	exit_info_1 = paddr;
1463 	/* Can never be greater than 8 */
1464 	exit_info_2 = bytes;
1465 
1466 	ghcb_set_sw_scratch(ghcb, ghcb_pa + offsetof(struct ghcb, shared_buffer));
1467 
1468 	return sev_es_ghcb_hv_call(ghcb, ctxt, exit_code, exit_info_1, exit_info_2);
1469 }
1470 
1471 /*
1472  * The MOVS instruction has two memory operands, which raises the
1473  * problem that it is not known whether the access to the source or the
1474  * destination caused the #VC exception (and hence whether an MMIO read
1475  * or write operation needs to be emulated).
1476  *
1477  * Instead of playing games with walking page-tables and trying to guess
1478  * whether the source or destination is an MMIO range, split the move
1479  * into two operations, a read and a write with only one memory operand.
1480  * This will cause a nested #VC exception on the MMIO address which can
1481  * then be handled.
1482  *
1483  * This implementation has the benefit that it also supports MOVS where
1484  * source _and_ destination are MMIO regions.
1485  *
1486  * It will slow MOVS on MMIO down a lot, but in SEV-ES guests it is a
1487  * rare operation. If it turns out to be a performance problem the split
1488  * operations can be moved to memcpy_fromio() and memcpy_toio().
1489  */
1490 static enum es_result vc_handle_mmio_movs(struct es_em_ctxt *ctxt,
1491 					  unsigned int bytes)
1492 {
1493 	unsigned long ds_base, es_base;
1494 	unsigned char *src, *dst;
1495 	unsigned char buffer[8];
1496 	enum es_result ret;
1497 	bool rep;
1498 	int off;
1499 
1500 	ds_base = insn_get_seg_base(ctxt->regs, INAT_SEG_REG_DS);
1501 	es_base = insn_get_seg_base(ctxt->regs, INAT_SEG_REG_ES);
1502 
1503 	if (ds_base == -1L || es_base == -1L) {
1504 		ctxt->fi.vector = X86_TRAP_GP;
1505 		ctxt->fi.error_code = 0;
1506 		return ES_EXCEPTION;
1507 	}
1508 
1509 	src = ds_base + (unsigned char *)ctxt->regs->si;
1510 	dst = es_base + (unsigned char *)ctxt->regs->di;
1511 
1512 	ret = vc_read_mem(ctxt, src, buffer, bytes);
1513 	if (ret != ES_OK)
1514 		return ret;
1515 
1516 	ret = vc_write_mem(ctxt, dst, buffer, bytes);
1517 	if (ret != ES_OK)
1518 		return ret;
1519 
1520 	if (ctxt->regs->flags & X86_EFLAGS_DF)
1521 		off = -bytes;
1522 	else
1523 		off =  bytes;
1524 
1525 	ctxt->regs->si += off;
1526 	ctxt->regs->di += off;
1527 
1528 	rep = insn_has_rep_prefix(&ctxt->insn);
1529 	if (rep)
1530 		ctxt->regs->cx -= 1;
1531 
1532 	if (!rep || ctxt->regs->cx == 0)
1533 		return ES_OK;
1534 	else
1535 		return ES_RETRY;
1536 }
1537 
1538 static enum es_result vc_handle_mmio(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
1539 {
1540 	struct insn *insn = &ctxt->insn;
1541 	enum insn_mmio_type mmio;
1542 	unsigned int bytes = 0;
1543 	enum es_result ret;
1544 	u8 sign_byte;
1545 	long *reg_data;
1546 
1547 	mmio = insn_decode_mmio(insn, &bytes);
1548 	if (mmio == INSN_MMIO_DECODE_FAILED)
1549 		return ES_DECODE_FAILED;
1550 
1551 	if (mmio != INSN_MMIO_WRITE_IMM && mmio != INSN_MMIO_MOVS) {
1552 		reg_data = insn_get_modrm_reg_ptr(insn, ctxt->regs);
1553 		if (!reg_data)
1554 			return ES_DECODE_FAILED;
1555 	}
1556 
1557 	switch (mmio) {
1558 	case INSN_MMIO_WRITE:
1559 		memcpy(ghcb->shared_buffer, reg_data, bytes);
1560 		ret = vc_do_mmio(ghcb, ctxt, bytes, false);
1561 		break;
1562 	case INSN_MMIO_WRITE_IMM:
1563 		memcpy(ghcb->shared_buffer, insn->immediate1.bytes, bytes);
1564 		ret = vc_do_mmio(ghcb, ctxt, bytes, false);
1565 		break;
1566 	case INSN_MMIO_READ:
1567 		ret = vc_do_mmio(ghcb, ctxt, bytes, true);
1568 		if (ret)
1569 			break;
1570 
1571 		/* Zero-extend for 32-bit operation */
1572 		if (bytes == 4)
1573 			*reg_data = 0;
1574 
1575 		memcpy(reg_data, ghcb->shared_buffer, bytes);
1576 		break;
1577 	case INSN_MMIO_READ_ZERO_EXTEND:
1578 		ret = vc_do_mmio(ghcb, ctxt, bytes, true);
1579 		if (ret)
1580 			break;
1581 
1582 		/* Zero extend based on operand size */
1583 		memset(reg_data, 0, insn->opnd_bytes);
1584 		memcpy(reg_data, ghcb->shared_buffer, bytes);
1585 		break;
1586 	case INSN_MMIO_READ_SIGN_EXTEND:
1587 		ret = vc_do_mmio(ghcb, ctxt, bytes, true);
1588 		if (ret)
1589 			break;
1590 
1591 		if (bytes == 1) {
1592 			u8 *val = (u8 *)ghcb->shared_buffer;
1593 
1594 			sign_byte = (*val & 0x80) ? 0xff : 0x00;
1595 		} else {
1596 			u16 *val = (u16 *)ghcb->shared_buffer;
1597 
1598 			sign_byte = (*val & 0x8000) ? 0xff : 0x00;
1599 		}
1600 
1601 		/* Sign extend based on operand size */
1602 		memset(reg_data, sign_byte, insn->opnd_bytes);
1603 		memcpy(reg_data, ghcb->shared_buffer, bytes);
1604 		break;
1605 	case INSN_MMIO_MOVS:
1606 		ret = vc_handle_mmio_movs(ctxt, bytes);
1607 		break;
1608 	default:
1609 		ret = ES_UNSUPPORTED;
1610 		break;
1611 	}
1612 
1613 	return ret;
1614 }
1615 
1616 static enum es_result vc_handle_dr7_write(struct ghcb *ghcb,
1617 					  struct es_em_ctxt *ctxt)
1618 {
1619 	struct sev_es_runtime_data *data = this_cpu_read(runtime_data);
1620 	long val, *reg = vc_insn_get_rm(ctxt);
1621 	enum es_result ret;
1622 
1623 	if (!reg)
1624 		return ES_DECODE_FAILED;
1625 
1626 	val = *reg;
1627 
1628 	/* Upper 32 bits must be written as zeroes */
1629 	if (val >> 32) {
1630 		ctxt->fi.vector = X86_TRAP_GP;
1631 		ctxt->fi.error_code = 0;
1632 		return ES_EXCEPTION;
1633 	}
1634 
1635 	/* Clear out other reserved bits and set bit 10 */
1636 	val = (val & 0xffff23ffL) | BIT(10);
1637 
1638 	/* Early non-zero writes to DR7 are not supported */
1639 	if (!data && (val & ~DR7_RESET_VALUE))
1640 		return ES_UNSUPPORTED;
1641 
1642 	/* Using a value of 0 for ExitInfo1 means RAX holds the value */
1643 	ghcb_set_rax(ghcb, val);
1644 	ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_WRITE_DR7, 0, 0);
1645 	if (ret != ES_OK)
1646 		return ret;
1647 
1648 	if (data)
1649 		data->dr7 = val;
1650 
1651 	return ES_OK;
1652 }
1653 
1654 static enum es_result vc_handle_dr7_read(struct ghcb *ghcb,
1655 					 struct es_em_ctxt *ctxt)
1656 {
1657 	struct sev_es_runtime_data *data = this_cpu_read(runtime_data);
1658 	long *reg = vc_insn_get_rm(ctxt);
1659 
1660 	if (!reg)
1661 		return ES_DECODE_FAILED;
1662 
1663 	if (data)
1664 		*reg = data->dr7;
1665 	else
1666 		*reg = DR7_RESET_VALUE;
1667 
1668 	return ES_OK;
1669 }
1670 
1671 static enum es_result vc_handle_wbinvd(struct ghcb *ghcb,
1672 				       struct es_em_ctxt *ctxt)
1673 {
1674 	return sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_WBINVD, 0, 0);
1675 }
1676 
1677 static enum es_result vc_handle_rdpmc(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
1678 {
1679 	enum es_result ret;
1680 
1681 	ghcb_set_rcx(ghcb, ctxt->regs->cx);
1682 
1683 	ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_RDPMC, 0, 0);
1684 	if (ret != ES_OK)
1685 		return ret;
1686 
1687 	if (!(ghcb_rax_is_valid(ghcb) && ghcb_rdx_is_valid(ghcb)))
1688 		return ES_VMM_ERROR;
1689 
1690 	ctxt->regs->ax = ghcb->save.rax;
1691 	ctxt->regs->dx = ghcb->save.rdx;
1692 
1693 	return ES_OK;
1694 }
1695 
1696 static enum es_result vc_handle_monitor(struct ghcb *ghcb,
1697 					struct es_em_ctxt *ctxt)
1698 {
1699 	/*
1700 	 * Treat it as a NOP and do not leak a physical address to the
1701 	 * hypervisor.
1702 	 */
1703 	return ES_OK;
1704 }
1705 
1706 static enum es_result vc_handle_mwait(struct ghcb *ghcb,
1707 				      struct es_em_ctxt *ctxt)
1708 {
1709 	/* Treat the same as MONITOR/MONITORX */
1710 	return ES_OK;
1711 }
1712 
1713 static enum es_result vc_handle_vmmcall(struct ghcb *ghcb,
1714 					struct es_em_ctxt *ctxt)
1715 {
1716 	enum es_result ret;
1717 
1718 	ghcb_set_rax(ghcb, ctxt->regs->ax);
1719 	ghcb_set_cpl(ghcb, user_mode(ctxt->regs) ? 3 : 0);
1720 
1721 	if (x86_platform.hyper.sev_es_hcall_prepare)
1722 		x86_platform.hyper.sev_es_hcall_prepare(ghcb, ctxt->regs);
1723 
1724 	ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_VMMCALL, 0, 0);
1725 	if (ret != ES_OK)
1726 		return ret;
1727 
1728 	if (!ghcb_rax_is_valid(ghcb))
1729 		return ES_VMM_ERROR;
1730 
1731 	ctxt->regs->ax = ghcb->save.rax;
1732 
1733 	/*
1734 	 * Call sev_es_hcall_finish() after regs->ax is already set.
1735 	 * This allows the hypervisor handler to overwrite it again if
1736 	 * necessary.
1737 	 */
1738 	if (x86_platform.hyper.sev_es_hcall_finish &&
1739 	    !x86_platform.hyper.sev_es_hcall_finish(ghcb, ctxt->regs))
1740 		return ES_VMM_ERROR;
1741 
1742 	return ES_OK;
1743 }
1744 
1745 static enum es_result vc_handle_trap_ac(struct ghcb *ghcb,
1746 					struct es_em_ctxt *ctxt)
1747 {
1748 	/*
1749 	 * Calling ecx_alignment_check() directly does not work, because it
1750 	 * enables IRQs and the GHCB is active. Forward the exception and call
1751 	 * it later from vc_forward_exception().
1752 	 */
1753 	ctxt->fi.vector = X86_TRAP_AC;
1754 	ctxt->fi.error_code = 0;
1755 	return ES_EXCEPTION;
1756 }
1757 
1758 static enum es_result vc_handle_exitcode(struct es_em_ctxt *ctxt,
1759 					 struct ghcb *ghcb,
1760 					 unsigned long exit_code)
1761 {
1762 	enum es_result result;
1763 
1764 	switch (exit_code) {
1765 	case SVM_EXIT_READ_DR7:
1766 		result = vc_handle_dr7_read(ghcb, ctxt);
1767 		break;
1768 	case SVM_EXIT_WRITE_DR7:
1769 		result = vc_handle_dr7_write(ghcb, ctxt);
1770 		break;
1771 	case SVM_EXIT_EXCP_BASE + X86_TRAP_AC:
1772 		result = vc_handle_trap_ac(ghcb, ctxt);
1773 		break;
1774 	case SVM_EXIT_RDTSC:
1775 	case SVM_EXIT_RDTSCP:
1776 		result = vc_handle_rdtsc(ghcb, ctxt, exit_code);
1777 		break;
1778 	case SVM_EXIT_RDPMC:
1779 		result = vc_handle_rdpmc(ghcb, ctxt);
1780 		break;
1781 	case SVM_EXIT_INVD:
1782 		pr_err_ratelimited("#VC exception for INVD??? Seriously???\n");
1783 		result = ES_UNSUPPORTED;
1784 		break;
1785 	case SVM_EXIT_CPUID:
1786 		result = vc_handle_cpuid(ghcb, ctxt);
1787 		break;
1788 	case SVM_EXIT_IOIO:
1789 		result = vc_handle_ioio(ghcb, ctxt);
1790 		break;
1791 	case SVM_EXIT_MSR:
1792 		result = vc_handle_msr(ghcb, ctxt);
1793 		break;
1794 	case SVM_EXIT_VMMCALL:
1795 		result = vc_handle_vmmcall(ghcb, ctxt);
1796 		break;
1797 	case SVM_EXIT_WBINVD:
1798 		result = vc_handle_wbinvd(ghcb, ctxt);
1799 		break;
1800 	case SVM_EXIT_MONITOR:
1801 		result = vc_handle_monitor(ghcb, ctxt);
1802 		break;
1803 	case SVM_EXIT_MWAIT:
1804 		result = vc_handle_mwait(ghcb, ctxt);
1805 		break;
1806 	case SVM_EXIT_NPF:
1807 		result = vc_handle_mmio(ghcb, ctxt);
1808 		break;
1809 	default:
1810 		/*
1811 		 * Unexpected #VC exception
1812 		 */
1813 		result = ES_UNSUPPORTED;
1814 	}
1815 
1816 	return result;
1817 }
1818 
1819 static __always_inline void vc_forward_exception(struct es_em_ctxt *ctxt)
1820 {
1821 	long error_code = ctxt->fi.error_code;
1822 	int trapnr = ctxt->fi.vector;
1823 
1824 	ctxt->regs->orig_ax = ctxt->fi.error_code;
1825 
1826 	switch (trapnr) {
1827 	case X86_TRAP_GP:
1828 		exc_general_protection(ctxt->regs, error_code);
1829 		break;
1830 	case X86_TRAP_UD:
1831 		exc_invalid_op(ctxt->regs);
1832 		break;
1833 	case X86_TRAP_PF:
1834 		write_cr2(ctxt->fi.cr2);
1835 		exc_page_fault(ctxt->regs, error_code);
1836 		break;
1837 	case X86_TRAP_AC:
1838 		exc_alignment_check(ctxt->regs, error_code);
1839 		break;
1840 	default:
1841 		pr_emerg("Unsupported exception in #VC instruction emulation - can't continue\n");
1842 		BUG();
1843 	}
1844 }
1845 
1846 static __always_inline bool is_vc2_stack(unsigned long sp)
1847 {
1848 	return (sp >= __this_cpu_ist_bottom_va(VC2) && sp < __this_cpu_ist_top_va(VC2));
1849 }
1850 
1851 static __always_inline bool vc_from_invalid_context(struct pt_regs *regs)
1852 {
1853 	unsigned long sp, prev_sp;
1854 
1855 	sp      = (unsigned long)regs;
1856 	prev_sp = regs->sp;
1857 
1858 	/*
1859 	 * If the code was already executing on the VC2 stack when the #VC
1860 	 * happened, let it proceed to the normal handling routine. This way the
1861 	 * code executing on the VC2 stack can cause #VC exceptions to get handled.
1862 	 */
1863 	return is_vc2_stack(sp) && !is_vc2_stack(prev_sp);
1864 }
1865 
1866 static bool vc_raw_handle_exception(struct pt_regs *regs, unsigned long error_code)
1867 {
1868 	struct ghcb_state state;
1869 	struct es_em_ctxt ctxt;
1870 	enum es_result result;
1871 	struct ghcb *ghcb;
1872 	bool ret = true;
1873 
1874 	ghcb = __sev_get_ghcb(&state);
1875 
1876 	vc_ghcb_invalidate(ghcb);
1877 	result = vc_init_em_ctxt(&ctxt, regs, error_code);
1878 
1879 	if (result == ES_OK)
1880 		result = vc_handle_exitcode(&ctxt, ghcb, error_code);
1881 
1882 	__sev_put_ghcb(&state);
1883 
1884 	/* Done - now check the result */
1885 	switch (result) {
1886 	case ES_OK:
1887 		vc_finish_insn(&ctxt);
1888 		break;
1889 	case ES_UNSUPPORTED:
1890 		pr_err_ratelimited("Unsupported exit-code 0x%02lx in #VC exception (IP: 0x%lx)\n",
1891 				   error_code, regs->ip);
1892 		ret = false;
1893 		break;
1894 	case ES_VMM_ERROR:
1895 		pr_err_ratelimited("Failure in communication with VMM (exit-code 0x%02lx IP: 0x%lx)\n",
1896 				   error_code, regs->ip);
1897 		ret = false;
1898 		break;
1899 	case ES_DECODE_FAILED:
1900 		pr_err_ratelimited("Failed to decode instruction (exit-code 0x%02lx IP: 0x%lx)\n",
1901 				   error_code, regs->ip);
1902 		ret = false;
1903 		break;
1904 	case ES_EXCEPTION:
1905 		vc_forward_exception(&ctxt);
1906 		break;
1907 	case ES_RETRY:
1908 		/* Nothing to do */
1909 		break;
1910 	default:
1911 		pr_emerg("Unknown result in %s():%d\n", __func__, result);
1912 		/*
1913 		 * Emulating the instruction which caused the #VC exception
1914 		 * failed - can't continue so print debug information
1915 		 */
1916 		BUG();
1917 	}
1918 
1919 	return ret;
1920 }
1921 
1922 static __always_inline bool vc_is_db(unsigned long error_code)
1923 {
1924 	return error_code == SVM_EXIT_EXCP_BASE + X86_TRAP_DB;
1925 }
1926 
1927 /*
1928  * Runtime #VC exception handler when raised from kernel mode. Runs in NMI mode
1929  * and will panic when an error happens.
1930  */
1931 DEFINE_IDTENTRY_VC_KERNEL(exc_vmm_communication)
1932 {
1933 	irqentry_state_t irq_state;
1934 
1935 	/*
1936 	 * With the current implementation it is always possible to switch to a
1937 	 * safe stack because #VC exceptions only happen at known places, like
1938 	 * intercepted instructions or accesses to MMIO areas/IO ports. They can
1939 	 * also happen with code instrumentation when the hypervisor intercepts
1940 	 * #DB, but the critical paths are forbidden to be instrumented, so #DB
1941 	 * exceptions currently also only happen in safe places.
1942 	 *
1943 	 * But keep this here in case the noinstr annotations are violated due
1944 	 * to bug elsewhere.
1945 	 */
1946 	if (unlikely(vc_from_invalid_context(regs))) {
1947 		instrumentation_begin();
1948 		panic("Can't handle #VC exception from unsupported context\n");
1949 		instrumentation_end();
1950 	}
1951 
1952 	/*
1953 	 * Handle #DB before calling into !noinstr code to avoid recursive #DB.
1954 	 */
1955 	if (vc_is_db(error_code)) {
1956 		exc_debug(regs);
1957 		return;
1958 	}
1959 
1960 	irq_state = irqentry_nmi_enter(regs);
1961 
1962 	instrumentation_begin();
1963 
1964 	if (!vc_raw_handle_exception(regs, error_code)) {
1965 		/* Show some debug info */
1966 		show_regs(regs);
1967 
1968 		/* Ask hypervisor to sev_es_terminate */
1969 		sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
1970 
1971 		/* If that fails and we get here - just panic */
1972 		panic("Returned from Terminate-Request to Hypervisor\n");
1973 	}
1974 
1975 	instrumentation_end();
1976 	irqentry_nmi_exit(regs, irq_state);
1977 }
1978 
1979 /*
1980  * Runtime #VC exception handler when raised from user mode. Runs in IRQ mode
1981  * and will kill the current task with SIGBUS when an error happens.
1982  */
1983 DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
1984 {
1985 	/*
1986 	 * Handle #DB before calling into !noinstr code to avoid recursive #DB.
1987 	 */
1988 	if (vc_is_db(error_code)) {
1989 		noist_exc_debug(regs);
1990 		return;
1991 	}
1992 
1993 	irqentry_enter_from_user_mode(regs);
1994 	instrumentation_begin();
1995 
1996 	if (!vc_raw_handle_exception(regs, error_code)) {
1997 		/*
1998 		 * Do not kill the machine if user-space triggered the
1999 		 * exception. Send SIGBUS instead and let user-space deal with
2000 		 * it.
2001 		 */
2002 		force_sig_fault(SIGBUS, BUS_OBJERR, (void __user *)0);
2003 	}
2004 
2005 	instrumentation_end();
2006 	irqentry_exit_to_user_mode(regs);
2007 }
2008 
2009 bool __init handle_vc_boot_ghcb(struct pt_regs *regs)
2010 {
2011 	unsigned long exit_code = regs->orig_ax;
2012 	struct es_em_ctxt ctxt;
2013 	enum es_result result;
2014 
2015 	vc_ghcb_invalidate(boot_ghcb);
2016 
2017 	result = vc_init_em_ctxt(&ctxt, regs, exit_code);
2018 	if (result == ES_OK)
2019 		result = vc_handle_exitcode(&ctxt, boot_ghcb, exit_code);
2020 
2021 	/* Done - now check the result */
2022 	switch (result) {
2023 	case ES_OK:
2024 		vc_finish_insn(&ctxt);
2025 		break;
2026 	case ES_UNSUPPORTED:
2027 		early_printk("PANIC: Unsupported exit-code 0x%02lx in early #VC exception (IP: 0x%lx)\n",
2028 				exit_code, regs->ip);
2029 		goto fail;
2030 	case ES_VMM_ERROR:
2031 		early_printk("PANIC: Failure in communication with VMM (exit-code 0x%02lx IP: 0x%lx)\n",
2032 				exit_code, regs->ip);
2033 		goto fail;
2034 	case ES_DECODE_FAILED:
2035 		early_printk("PANIC: Failed to decode instruction (exit-code 0x%02lx IP: 0x%lx)\n",
2036 				exit_code, regs->ip);
2037 		goto fail;
2038 	case ES_EXCEPTION:
2039 		vc_early_forward_exception(&ctxt);
2040 		break;
2041 	case ES_RETRY:
2042 		/* Nothing to do */
2043 		break;
2044 	default:
2045 		BUG();
2046 	}
2047 
2048 	return true;
2049 
2050 fail:
2051 	show_regs(regs);
2052 
2053 	sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
2054 }
2055 
2056 /*
2057  * Initial set up of SNP relies on information provided by the
2058  * Confidential Computing blob, which can be passed to the kernel
2059  * in the following ways, depending on how it is booted:
2060  *
2061  * - when booted via the boot/decompress kernel:
2062  *   - via boot_params
2063  *
2064  * - when booted directly by firmware/bootloader (e.g. CONFIG_PVH):
2065  *   - via a setup_data entry, as defined by the Linux Boot Protocol
2066  *
2067  * Scan for the blob in that order.
2068  */
2069 static __init struct cc_blob_sev_info *find_cc_blob(struct boot_params *bp)
2070 {
2071 	struct cc_blob_sev_info *cc_info;
2072 
2073 	/* Boot kernel would have passed the CC blob via boot_params. */
2074 	if (bp->cc_blob_address) {
2075 		cc_info = (struct cc_blob_sev_info *)(unsigned long)bp->cc_blob_address;
2076 		goto found_cc_info;
2077 	}
2078 
2079 	/*
2080 	 * If kernel was booted directly, without the use of the
2081 	 * boot/decompression kernel, the CC blob may have been passed via
2082 	 * setup_data instead.
2083 	 */
2084 	cc_info = find_cc_blob_setup_data(bp);
2085 	if (!cc_info)
2086 		return NULL;
2087 
2088 found_cc_info:
2089 	if (cc_info->magic != CC_BLOB_SEV_HDR_MAGIC)
2090 		snp_abort();
2091 
2092 	return cc_info;
2093 }
2094 
2095 bool __init snp_init(struct boot_params *bp)
2096 {
2097 	struct cc_blob_sev_info *cc_info;
2098 
2099 	if (!bp)
2100 		return false;
2101 
2102 	cc_info = find_cc_blob(bp);
2103 	if (!cc_info)
2104 		return false;
2105 
2106 	setup_cpuid_table(cc_info);
2107 
2108 	/*
2109 	 * The CC blob will be used later to access the secrets page. Cache
2110 	 * it here like the boot kernel does.
2111 	 */
2112 	bp->cc_blob_address = (u32)(unsigned long)cc_info;
2113 
2114 	return true;
2115 }
2116 
2117 void __init __noreturn snp_abort(void)
2118 {
2119 	sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SNP_UNSUPPORTED);
2120 }
2121 
2122 static void dump_cpuid_table(void)
2123 {
2124 	const struct snp_cpuid_table *cpuid_table = snp_cpuid_get_table();
2125 	int i = 0;
2126 
2127 	pr_info("count=%d reserved=0x%x reserved2=0x%llx\n",
2128 		cpuid_table->count, cpuid_table->__reserved1, cpuid_table->__reserved2);
2129 
2130 	for (i = 0; i < SNP_CPUID_COUNT_MAX; i++) {
2131 		const struct snp_cpuid_fn *fn = &cpuid_table->fn[i];
2132 
2133 		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",
2134 			i, fn->eax_in, fn->ecx_in, fn->eax, fn->ebx, fn->ecx,
2135 			fn->edx, fn->xcr0_in, fn->xss_in, fn->__reserved);
2136 	}
2137 }
2138 
2139 /*
2140  * It is useful from an auditing/testing perspective to provide an easy way
2141  * for the guest owner to know that the CPUID table has been initialized as
2142  * expected, but that initialization happens too early in boot to print any
2143  * sort of indicator, and there's not really any other good place to do it,
2144  * so do it here.
2145  */
2146 static int __init report_cpuid_table(void)
2147 {
2148 	const struct snp_cpuid_table *cpuid_table = snp_cpuid_get_table();
2149 
2150 	if (!cpuid_table->count)
2151 		return 0;
2152 
2153 	pr_info("Using SNP CPUID table, %d entries present.\n",
2154 		cpuid_table->count);
2155 
2156 	if (sev_cfg.debug)
2157 		dump_cpuid_table();
2158 
2159 	return 0;
2160 }
2161 arch_initcall(report_cpuid_table);
2162 
2163 static int __init init_sev_config(char *str)
2164 {
2165 	char *s;
2166 
2167 	while ((s = strsep(&str, ","))) {
2168 		if (!strcmp(s, "debug")) {
2169 			sev_cfg.debug = true;
2170 			continue;
2171 		}
2172 
2173 		pr_info("SEV command-line option '%s' was not recognized\n", s);
2174 	}
2175 
2176 	return 1;
2177 }
2178 __setup("sev=", init_sev_config);
2179 
2180 int snp_issue_guest_request(u64 exit_code, struct snp_req_data *input, struct snp_guest_request_ioctl *rio)
2181 {
2182 	struct ghcb_state state;
2183 	struct es_em_ctxt ctxt;
2184 	unsigned long flags;
2185 	struct ghcb *ghcb;
2186 	int ret;
2187 
2188 	rio->exitinfo2 = SEV_RET_NO_FW_CALL;
2189 
2190 	/*
2191 	 * __sev_get_ghcb() needs to run with IRQs disabled because it is using
2192 	 * a per-CPU GHCB.
2193 	 */
2194 	local_irq_save(flags);
2195 
2196 	ghcb = __sev_get_ghcb(&state);
2197 	if (!ghcb) {
2198 		ret = -EIO;
2199 		goto e_restore_irq;
2200 	}
2201 
2202 	vc_ghcb_invalidate(ghcb);
2203 
2204 	if (exit_code == SVM_VMGEXIT_EXT_GUEST_REQUEST) {
2205 		ghcb_set_rax(ghcb, input->data_gpa);
2206 		ghcb_set_rbx(ghcb, input->data_npages);
2207 	}
2208 
2209 	ret = sev_es_ghcb_hv_call(ghcb, &ctxt, exit_code, input->req_gpa, input->resp_gpa);
2210 	if (ret)
2211 		goto e_put;
2212 
2213 	rio->exitinfo2 = ghcb->save.sw_exit_info_2;
2214 	switch (rio->exitinfo2) {
2215 	case 0:
2216 		break;
2217 
2218 	case SNP_GUEST_VMM_ERR(SNP_GUEST_VMM_ERR_BUSY):
2219 		ret = -EAGAIN;
2220 		break;
2221 
2222 	case SNP_GUEST_VMM_ERR(SNP_GUEST_VMM_ERR_INVALID_LEN):
2223 		/* Number of expected pages are returned in RBX */
2224 		if (exit_code == SVM_VMGEXIT_EXT_GUEST_REQUEST) {
2225 			input->data_npages = ghcb_get_rbx(ghcb);
2226 			ret = -ENOSPC;
2227 			break;
2228 		}
2229 		fallthrough;
2230 	default:
2231 		ret = -EIO;
2232 		break;
2233 	}
2234 
2235 e_put:
2236 	__sev_put_ghcb(&state);
2237 e_restore_irq:
2238 	local_irq_restore(flags);
2239 
2240 	return ret;
2241 }
2242 EXPORT_SYMBOL_GPL(snp_issue_guest_request);
2243 
2244 static struct platform_device sev_guest_device = {
2245 	.name		= "sev-guest",
2246 	.id		= -1,
2247 };
2248 
2249 static int __init snp_init_platform_device(void)
2250 {
2251 	struct sev_guest_platform_data data;
2252 	u64 gpa;
2253 
2254 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
2255 		return -ENODEV;
2256 
2257 	gpa = get_secrets_page();
2258 	if (!gpa)
2259 		return -ENODEV;
2260 
2261 	data.secrets_gpa = gpa;
2262 	if (platform_device_add_data(&sev_guest_device, &data, sizeof(data)))
2263 		return -ENODEV;
2264 
2265 	if (platform_device_register(&sev_guest_device))
2266 		return -ENODEV;
2267 
2268 	pr_info("SNP guest platform device initialized.\n");
2269 	return 0;
2270 }
2271 device_initcall(snp_init_platform_device);
2272