1 // SPDX-License-Identifier: GPL-2.0-only
3 * AMD Memory Encryption Support
5 * Copyright (C) 2019 SUSE
7 * Author: Joerg Roedel <jroedel@suse.de>
10 #define pr_fmt(fmt) "SEV: " fmt
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>
21 #include <linux/cpumask.h>
22 #include <linux/efi.h>
23 #include <linux/platform_device.h>
25 #include <linux/psp-sev.h>
26 #include <uapi/linux/sev-guest.h>
28 #include <asm/cpu_entry_area.h>
29 #include <asm/stacktrace.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>
41 #include <asm/cpuid.h>
42 #include <asm/cmdline.h>
44 #define DR7_RESET_VALUE 0x400
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
62 /* For early boot hypervisor communication in SEV-ES enabled guests */
63 static struct ghcb boot_ghcb_page __bss_decrypted __aligned(PAGE_SIZE);
66 * Needs to be in the .data section because we need it NULL before bss is
69 static struct ghcb *boot_ghcb __section(".data");
71 /* Bitmap of SEV features supported by the hypervisor */
72 static u64 sev_hv_features __ro_after_init;
74 /* #VC handler runtime per-CPU data */
75 struct sev_es_runtime_data {
76 struct ghcb ghcb_page;
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
85 struct ghcb backup_ghcb;
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.
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
101 bool backup_ghcb_active;
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.
115 static DEFINE_PER_CPU(struct sev_es_runtime_data*, runtime_data);
116 static DEFINE_PER_CPU(struct sev_es_save_area *, sev_vmsa);
122 * A flag used by __set_pages_state() that indicates when the
123 * per-CPU GHCB has been created and registered and thus can be
124 * used by the BSP instead of the early boot GHCB.
126 * For APs, the per-CPU GHCB is created before they are started
127 * and registered upon startup, so this flag can be used globally
128 * for the BSP and APs.
130 ghcbs_initialized : 1,
135 static struct sev_config sev_cfg __read_mostly;
137 static __always_inline bool on_vc_stack(struct pt_regs *regs)
139 unsigned long sp = regs->sp;
141 /* User-mode RSP is not trusted */
145 /* SYSCALL gap still has user-mode RSP */
146 if (ip_within_syscall_gap(regs))
149 return ((sp >= __this_cpu_ist_bottom_va(VC)) && (sp < __this_cpu_ist_top_va(VC)));
153 * This function handles the case when an NMI is raised in the #VC
154 * exception handler entry code, before the #VC handler has switched off
155 * its IST stack. In this case, the IST entry for #VC must be adjusted,
156 * so that any nested #VC exception will not overwrite the stack
157 * contents of the interrupted #VC handler.
159 * The IST entry is adjusted unconditionally so that it can be also be
160 * unconditionally adjusted back in __sev_es_ist_exit(). Otherwise a
161 * nested sev_es_ist_exit() call may adjust back the IST entry too
164 * The __sev_es_ist_enter() and __sev_es_ist_exit() functions always run
165 * on the NMI IST stack, as they are only called from NMI handling code
168 void noinstr __sev_es_ist_enter(struct pt_regs *regs)
170 unsigned long old_ist, new_ist;
172 /* Read old IST entry */
173 new_ist = old_ist = __this_cpu_read(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC]);
176 * If NMI happened while on the #VC IST stack, set the new IST
177 * value below regs->sp, so that the interrupted stack frame is
178 * not overwritten by subsequent #VC exceptions.
180 if (on_vc_stack(regs))
184 * Reserve additional 8 bytes and store old IST value so this
185 * adjustment can be unrolled in __sev_es_ist_exit().
187 new_ist -= sizeof(old_ist);
188 *(unsigned long *)new_ist = old_ist;
190 /* Set new IST entry */
191 this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], new_ist);
194 void noinstr __sev_es_ist_exit(void)
199 ist = __this_cpu_read(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC]);
201 if (WARN_ON(ist == __this_cpu_ist_top_va(VC)))
204 /* Read back old IST entry and write it to the TSS */
205 this_cpu_write(cpu_tss_rw.x86_tss.ist[IST_INDEX_VC], *(unsigned long *)ist);
209 * Nothing shall interrupt this code path while holding the per-CPU
210 * GHCB. The backup GHCB is only for NMIs interrupting this path.
212 * Callers must disable local interrupts around it.
214 static noinstr struct ghcb *__sev_get_ghcb(struct ghcb_state *state)
216 struct sev_es_runtime_data *data;
219 WARN_ON(!irqs_disabled());
221 data = this_cpu_read(runtime_data);
222 ghcb = &data->ghcb_page;
224 if (unlikely(data->ghcb_active)) {
225 /* GHCB is already in use - save its contents */
227 if (unlikely(data->backup_ghcb_active)) {
229 * Backup-GHCB is also already in use. There is no way
230 * to continue here so just kill the machine. To make
231 * panic() work, mark GHCBs inactive so that messages
232 * can be printed out.
234 data->ghcb_active = false;
235 data->backup_ghcb_active = false;
237 instrumentation_begin();
238 panic("Unable to handle #VC exception! GHCB and Backup GHCB are already in use");
239 instrumentation_end();
242 /* Mark backup_ghcb active before writing to it */
243 data->backup_ghcb_active = true;
245 state->ghcb = &data->backup_ghcb;
247 /* Backup GHCB content */
248 *state->ghcb = *ghcb;
251 data->ghcb_active = true;
257 static inline u64 sev_es_rd_ghcb_msr(void)
259 return __rdmsr(MSR_AMD64_SEV_ES_GHCB);
262 static __always_inline void sev_es_wr_ghcb_msr(u64 val)
267 high = (u32)(val >> 32);
269 native_wrmsr(MSR_AMD64_SEV_ES_GHCB, low, high);
272 static int vc_fetch_insn_kernel(struct es_em_ctxt *ctxt,
273 unsigned char *buffer)
275 return copy_from_kernel_nofault(buffer, (unsigned char *)ctxt->regs->ip, MAX_INSN_SIZE);
278 static enum es_result __vc_decode_user_insn(struct es_em_ctxt *ctxt)
280 char buffer[MAX_INSN_SIZE];
283 insn_bytes = insn_fetch_from_user_inatomic(ctxt->regs, buffer);
284 if (insn_bytes == 0) {
285 /* Nothing could be copied */
286 ctxt->fi.vector = X86_TRAP_PF;
287 ctxt->fi.error_code = X86_PF_INSTR | X86_PF_USER;
288 ctxt->fi.cr2 = ctxt->regs->ip;
290 } else if (insn_bytes == -EINVAL) {
291 /* Effective RIP could not be calculated */
292 ctxt->fi.vector = X86_TRAP_GP;
293 ctxt->fi.error_code = 0;
298 if (!insn_decode_from_regs(&ctxt->insn, ctxt->regs, buffer, insn_bytes))
299 return ES_DECODE_FAILED;
301 if (ctxt->insn.immediate.got)
304 return ES_DECODE_FAILED;
307 static enum es_result __vc_decode_kern_insn(struct es_em_ctxt *ctxt)
309 char buffer[MAX_INSN_SIZE];
312 res = vc_fetch_insn_kernel(ctxt, buffer);
314 ctxt->fi.vector = X86_TRAP_PF;
315 ctxt->fi.error_code = X86_PF_INSTR;
316 ctxt->fi.cr2 = ctxt->regs->ip;
320 ret = insn_decode(&ctxt->insn, buffer, MAX_INSN_SIZE, INSN_MODE_64);
322 return ES_DECODE_FAILED;
327 static enum es_result vc_decode_insn(struct es_em_ctxt *ctxt)
329 if (user_mode(ctxt->regs))
330 return __vc_decode_user_insn(ctxt);
332 return __vc_decode_kern_insn(ctxt);
335 static enum es_result vc_write_mem(struct es_em_ctxt *ctxt,
336 char *dst, char *buf, size_t size)
338 unsigned long error_code = X86_PF_PROT | X86_PF_WRITE;
341 * This function uses __put_user() independent of whether kernel or user
342 * memory is accessed. This works fine because __put_user() does no
343 * sanity checks of the pointer being accessed. All that it does is
344 * to report when the access failed.
346 * Also, this function runs in atomic context, so __put_user() is not
347 * allowed to sleep. The page-fault handler detects that it is running
348 * in atomic context and will not try to take mmap_sem and handle the
349 * fault, so additional pagefault_enable()/disable() calls are not
352 * The access can't be done via copy_to_user() here because
353 * vc_write_mem() must not use string instructions to access unsafe
354 * memory. The reason is that MOVS is emulated by the #VC handler by
355 * splitting the move up into a read and a write and taking a nested #VC
356 * exception on whatever of them is the MMIO access. Using string
357 * instructions here would cause infinite nesting.
362 u8 __user *target = (u8 __user *)dst;
365 if (__put_user(d1, target))
371 u16 __user *target = (u16 __user *)dst;
374 if (__put_user(d2, target))
380 u32 __user *target = (u32 __user *)dst;
383 if (__put_user(d4, target))
389 u64 __user *target = (u64 __user *)dst;
392 if (__put_user(d8, target))
397 WARN_ONCE(1, "%s: Invalid size: %zu\n", __func__, size);
398 return ES_UNSUPPORTED;
404 if (user_mode(ctxt->regs))
405 error_code |= X86_PF_USER;
407 ctxt->fi.vector = X86_TRAP_PF;
408 ctxt->fi.error_code = error_code;
409 ctxt->fi.cr2 = (unsigned long)dst;
414 static enum es_result vc_read_mem(struct es_em_ctxt *ctxt,
415 char *src, char *buf, size_t size)
417 unsigned long error_code = X86_PF_PROT;
420 * This function uses __get_user() independent of whether kernel or user
421 * memory is accessed. This works fine because __get_user() does no
422 * sanity checks of the pointer being accessed. All that it does is
423 * to report when the access failed.
425 * Also, this function runs in atomic context, so __get_user() is not
426 * allowed to sleep. The page-fault handler detects that it is running
427 * in atomic context and will not try to take mmap_sem and handle the
428 * fault, so additional pagefault_enable()/disable() calls are not
431 * The access can't be done via copy_from_user() here because
432 * vc_read_mem() must not use string instructions to access unsafe
433 * memory. The reason is that MOVS is emulated by the #VC handler by
434 * splitting the move up into a read and a write and taking a nested #VC
435 * exception on whatever of them is the MMIO access. Using string
436 * instructions here would cause infinite nesting.
441 u8 __user *s = (u8 __user *)src;
443 if (__get_user(d1, s))
450 u16 __user *s = (u16 __user *)src;
452 if (__get_user(d2, s))
459 u32 __user *s = (u32 __user *)src;
461 if (__get_user(d4, s))
468 u64 __user *s = (u64 __user *)src;
469 if (__get_user(d8, s))
475 WARN_ONCE(1, "%s: Invalid size: %zu\n", __func__, size);
476 return ES_UNSUPPORTED;
482 if (user_mode(ctxt->regs))
483 error_code |= X86_PF_USER;
485 ctxt->fi.vector = X86_TRAP_PF;
486 ctxt->fi.error_code = error_code;
487 ctxt->fi.cr2 = (unsigned long)src;
492 static enum es_result vc_slow_virt_to_phys(struct ghcb *ghcb, struct es_em_ctxt *ctxt,
493 unsigned long vaddr, phys_addr_t *paddr)
495 unsigned long va = (unsigned long)vaddr;
501 pgd = __va(read_cr3_pa());
502 pgd = &pgd[pgd_index(va)];
503 pte = lookup_address_in_pgd(pgd, va, &level);
505 ctxt->fi.vector = X86_TRAP_PF;
506 ctxt->fi.cr2 = vaddr;
507 ctxt->fi.error_code = 0;
509 if (user_mode(ctxt->regs))
510 ctxt->fi.error_code |= X86_PF_USER;
515 if (WARN_ON_ONCE(pte_val(*pte) & _PAGE_ENC))
516 /* Emulated MMIO to/from encrypted memory not supported */
517 return ES_UNSUPPORTED;
519 pa = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
520 pa |= va & ~page_level_mask(level);
527 static enum es_result vc_ioio_check(struct es_em_ctxt *ctxt, u16 port, size_t size)
531 if (user_mode(ctxt->regs)) {
532 struct thread_struct *t = ¤t->thread;
533 struct io_bitmap *iobm = t->io_bitmap;
539 for (idx = port; idx < port + size; ++idx) {
540 if (test_bit(idx, iobm->bitmap))
548 ctxt->fi.vector = X86_TRAP_GP;
549 ctxt->fi.error_code = 0;
554 /* Include code shared with pre-decompression boot stage */
555 #include "sev-shared.c"
557 static noinstr void __sev_put_ghcb(struct ghcb_state *state)
559 struct sev_es_runtime_data *data;
562 WARN_ON(!irqs_disabled());
564 data = this_cpu_read(runtime_data);
565 ghcb = &data->ghcb_page;
568 /* Restore GHCB from Backup */
569 *ghcb = *state->ghcb;
570 data->backup_ghcb_active = false;
574 * Invalidate the GHCB so a VMGEXIT instruction issued
575 * from userspace won't appear to be valid.
577 vc_ghcb_invalidate(ghcb);
578 data->ghcb_active = false;
582 void noinstr __sev_es_nmi_complete(void)
584 struct ghcb_state state;
587 ghcb = __sev_get_ghcb(&state);
589 vc_ghcb_invalidate(ghcb);
590 ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_NMI_COMPLETE);
591 ghcb_set_sw_exit_info_1(ghcb, 0);
592 ghcb_set_sw_exit_info_2(ghcb, 0);
594 sev_es_wr_ghcb_msr(__pa_nodebug(ghcb));
597 __sev_put_ghcb(&state);
600 static u64 __init get_secrets_page(void)
602 u64 pa_data = boot_params.cc_blob_address;
603 struct cc_blob_sev_info info;
607 * The CC blob contains the address of the secrets page, check if the
613 map = early_memremap(pa_data, sizeof(info));
615 pr_err("Unable to locate SNP secrets page: failed to map the Confidential Computing blob.\n");
618 memcpy(&info, map, sizeof(info));
619 early_memunmap(map, sizeof(info));
621 /* smoke-test the secrets page passed */
622 if (!info.secrets_phys || info.secrets_len != PAGE_SIZE)
625 return info.secrets_phys;
628 static u64 __init get_snp_jump_table_addr(void)
630 struct snp_secrets_page_layout *layout;
634 pa = get_secrets_page();
638 mem = ioremap_encrypted(pa, PAGE_SIZE);
640 pr_err("Unable to locate AP jump table address: failed to map the SNP secrets page.\n");
644 layout = (__force struct snp_secrets_page_layout *)mem;
646 addr = layout->os_area.ap_jump_table_pa;
652 static u64 __init get_jump_table_addr(void)
654 struct ghcb_state state;
659 if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
660 return get_snp_jump_table_addr();
662 local_irq_save(flags);
664 ghcb = __sev_get_ghcb(&state);
666 vc_ghcb_invalidate(ghcb);
667 ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_JUMP_TABLE);
668 ghcb_set_sw_exit_info_1(ghcb, SVM_VMGEXIT_GET_AP_JUMP_TABLE);
669 ghcb_set_sw_exit_info_2(ghcb, 0);
671 sev_es_wr_ghcb_msr(__pa(ghcb));
674 if (ghcb_sw_exit_info_1_is_valid(ghcb) &&
675 ghcb_sw_exit_info_2_is_valid(ghcb))
676 ret = ghcb->save.sw_exit_info_2;
678 __sev_put_ghcb(&state);
680 local_irq_restore(flags);
685 static void early_set_pages_state(unsigned long vaddr, unsigned long paddr,
686 unsigned long npages, enum psc_op op)
688 unsigned long paddr_end;
692 vaddr = vaddr & PAGE_MASK;
694 paddr = paddr & PAGE_MASK;
695 paddr_end = paddr + (npages << PAGE_SHIFT);
697 while (paddr < paddr_end) {
698 if (op == SNP_PAGE_STATE_SHARED) {
699 /* Page validation must be rescinded before changing to shared */
700 ret = pvalidate(vaddr, RMP_PG_SIZE_4K, false);
701 if (WARN(ret, "Failed to validate address 0x%lx ret %d", paddr, ret))
706 * Use the MSR protocol because this function can be called before
707 * the GHCB is established.
709 sev_es_wr_ghcb_msr(GHCB_MSR_PSC_REQ_GFN(paddr >> PAGE_SHIFT, op));
712 val = sev_es_rd_ghcb_msr();
714 if (WARN(GHCB_RESP_CODE(val) != GHCB_MSR_PSC_RESP,
715 "Wrong PSC response code: 0x%x\n",
716 (unsigned int)GHCB_RESP_CODE(val)))
719 if (WARN(GHCB_MSR_PSC_RESP_VAL(val),
720 "Failed to change page state to '%s' paddr 0x%lx error 0x%llx\n",
721 op == SNP_PAGE_STATE_PRIVATE ? "private" : "shared",
722 paddr, GHCB_MSR_PSC_RESP_VAL(val)))
725 if (op == SNP_PAGE_STATE_PRIVATE) {
726 /* Page validation must be performed after changing to private */
727 ret = pvalidate(vaddr, RMP_PG_SIZE_4K, true);
728 if (WARN(ret, "Failed to validate address 0x%lx ret %d", paddr, ret))
739 sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PSC);
742 void __init early_snp_set_memory_private(unsigned long vaddr, unsigned long paddr,
743 unsigned long npages)
746 * This can be invoked in early boot while running identity mapped, so
747 * use an open coded check for SNP instead of using cc_platform_has().
748 * This eliminates worries about jump tables or checking boot_cpu_data
749 * in the cc_platform_has() function.
751 if (!(sev_status & MSR_AMD64_SEV_SNP_ENABLED))
755 * Ask the hypervisor to mark the memory pages as private in the RMP
758 early_set_pages_state(vaddr, paddr, npages, SNP_PAGE_STATE_PRIVATE);
761 void __init early_snp_set_memory_shared(unsigned long vaddr, unsigned long paddr,
762 unsigned long npages)
765 * This can be invoked in early boot while running identity mapped, so
766 * use an open coded check for SNP instead of using cc_platform_has().
767 * This eliminates worries about jump tables or checking boot_cpu_data
768 * in the cc_platform_has() function.
770 if (!(sev_status & MSR_AMD64_SEV_SNP_ENABLED))
773 /* Ask hypervisor to mark the memory pages shared in the RMP table. */
774 early_set_pages_state(vaddr, paddr, npages, SNP_PAGE_STATE_SHARED);
777 void __init snp_prep_memory(unsigned long paddr, unsigned int sz, enum psc_op op)
779 unsigned long vaddr, npages;
781 vaddr = (unsigned long)__va(paddr);
782 npages = PAGE_ALIGN(sz) >> PAGE_SHIFT;
784 if (op == SNP_PAGE_STATE_PRIVATE)
785 early_snp_set_memory_private(vaddr, paddr, npages);
786 else if (op == SNP_PAGE_STATE_SHARED)
787 early_snp_set_memory_shared(vaddr, paddr, npages);
789 WARN(1, "invalid memory op %d\n", op);
792 static unsigned long __set_pages_state(struct snp_psc_desc *data, unsigned long vaddr,
793 unsigned long vaddr_end, int op)
795 struct ghcb_state state;
796 bool use_large_entry;
807 memset(data, 0, sizeof(*data));
810 while (vaddr < vaddr_end && i < ARRAY_SIZE(data->entries)) {
813 if (is_vmalloc_addr((void *)vaddr)) {
814 pfn = vmalloc_to_pfn((void *)vaddr);
815 use_large_entry = false;
817 pfn = __pa(vaddr) >> PAGE_SHIFT;
818 use_large_entry = true;
824 if (use_large_entry && IS_ALIGNED(vaddr, PMD_SIZE) &&
825 (vaddr_end - vaddr) >= PMD_SIZE) {
826 e->pagesize = RMP_PG_SIZE_2M;
829 e->pagesize = RMP_PG_SIZE_4K;
837 /* Page validation must be rescinded before changing to shared */
838 if (op == SNP_PAGE_STATE_SHARED)
839 pvalidate_pages(data);
841 local_irq_save(flags);
843 if (sev_cfg.ghcbs_initialized)
844 ghcb = __sev_get_ghcb(&state);
848 /* Invoke the hypervisor to perform the page state changes */
849 if (!ghcb || vmgexit_psc(ghcb, data))
850 sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PSC);
852 if (sev_cfg.ghcbs_initialized)
853 __sev_put_ghcb(&state);
855 local_irq_restore(flags);
857 /* Page validation must be performed after changing to private */
858 if (op == SNP_PAGE_STATE_PRIVATE)
859 pvalidate_pages(data);
864 static void set_pages_state(unsigned long vaddr, unsigned long npages, int op)
866 struct snp_psc_desc desc;
867 unsigned long vaddr_end;
869 /* Use the MSR protocol when a GHCB is not available. */
871 return early_set_pages_state(vaddr, __pa(vaddr), npages, op);
873 vaddr = vaddr & PAGE_MASK;
874 vaddr_end = vaddr + (npages << PAGE_SHIFT);
876 while (vaddr < vaddr_end)
877 vaddr = __set_pages_state(&desc, vaddr, vaddr_end, op);
880 void snp_set_memory_shared(unsigned long vaddr, unsigned long npages)
882 if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
885 set_pages_state(vaddr, npages, SNP_PAGE_STATE_SHARED);
888 void snp_set_memory_private(unsigned long vaddr, unsigned long npages)
890 if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
893 set_pages_state(vaddr, npages, SNP_PAGE_STATE_PRIVATE);
896 void snp_accept_memory(phys_addr_t start, phys_addr_t end)
898 unsigned long vaddr, npages;
900 if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
903 vaddr = (unsigned long)__va(start);
904 npages = (end - start) >> PAGE_SHIFT;
906 set_pages_state(vaddr, npages, SNP_PAGE_STATE_PRIVATE);
909 static int snp_set_vmsa(void *va, bool vmsa)
914 * Running at VMPL0 allows the kernel to change the VMSA bit for a page
915 * using the RMPADJUST instruction. However, for the instruction to
916 * succeed it must target the permissions of a lesser privileged
917 * (higher numbered) VMPL level, so use VMPL1 (refer to the RMPADJUST
918 * instruction in the AMD64 APM Volume 3).
922 attrs |= RMPADJUST_VMSA_PAGE_BIT;
924 return rmpadjust((unsigned long)va, RMP_PG_SIZE_4K, attrs);
927 #define __ATTR_BASE (SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK)
928 #define INIT_CS_ATTRIBS (__ATTR_BASE | SVM_SELECTOR_READ_MASK | SVM_SELECTOR_CODE_MASK)
929 #define INIT_DS_ATTRIBS (__ATTR_BASE | SVM_SELECTOR_WRITE_MASK)
931 #define INIT_LDTR_ATTRIBS (SVM_SELECTOR_P_MASK | 2)
932 #define INIT_TR_ATTRIBS (SVM_SELECTOR_P_MASK | 3)
934 static void *snp_alloc_vmsa_page(void)
939 * Allocate VMSA page to work around the SNP erratum where the CPU will
940 * incorrectly signal an RMP violation #PF if a large page (2MB or 1GB)
941 * collides with the RMP entry of VMSA page. The recommended workaround
942 * is to not use a large page.
944 * Allocate an 8k page which is also 8k-aligned.
946 p = alloc_pages(GFP_KERNEL_ACCOUNT | __GFP_ZERO, 1);
952 /* Free the first 4k. This page may be 2M/1G aligned and cannot be used. */
955 return page_address(p + 1);
958 static void snp_cleanup_vmsa(struct sev_es_save_area *vmsa)
962 err = snp_set_vmsa(vmsa, false);
964 pr_err("clear VMSA page failed (%u), leaking page\n", err);
966 free_page((unsigned long)vmsa);
969 static int wakeup_cpu_via_vmgexit(u32 apic_id, unsigned long start_ip)
971 struct sev_es_save_area *cur_vmsa, *vmsa;
972 struct ghcb_state state;
980 * The hypervisor SNP feature support check has happened earlier, just check
981 * the AP_CREATION one here.
983 if (!(sev_hv_features & GHCB_HV_FT_SNP_AP_CREATION))
987 * Verify the desired start IP against the known trampoline start IP
988 * to catch any future new trampolines that may be introduced that
989 * would require a new protected guest entry point.
991 if (WARN_ONCE(start_ip != real_mode_header->trampoline_start,
992 "Unsupported SNP start_ip: %lx\n", start_ip))
995 /* Override start_ip with known protected guest start IP */
996 start_ip = real_mode_header->sev_es_trampoline_start;
998 /* Find the logical CPU for the APIC ID */
999 for_each_present_cpu(cpu) {
1000 if (arch_match_cpu_phys_id(cpu, apic_id))
1003 if (cpu >= nr_cpu_ids)
1006 cur_vmsa = per_cpu(sev_vmsa, cpu);
1009 * A new VMSA is created each time because there is no guarantee that
1010 * the current VMSA is the kernels or that the vCPU is not running. If
1011 * an attempt was done to use the current VMSA with a running vCPU, a
1012 * #VMEXIT of that vCPU would wipe out all of the settings being done
1015 vmsa = (struct sev_es_save_area *)snp_alloc_vmsa_page();
1019 /* CR4 should maintain the MCE value */
1020 cr4 = native_read_cr4() & X86_CR4_MCE;
1022 /* Set the CS value based on the start_ip converted to a SIPI vector */
1023 sipi_vector = (start_ip >> 12);
1024 vmsa->cs.base = sipi_vector << 12;
1025 vmsa->cs.limit = AP_INIT_CS_LIMIT;
1026 vmsa->cs.attrib = INIT_CS_ATTRIBS;
1027 vmsa->cs.selector = sipi_vector << 8;
1029 /* Set the RIP value based on start_ip */
1030 vmsa->rip = start_ip & 0xfff;
1032 /* Set AP INIT defaults as documented in the APM */
1033 vmsa->ds.limit = AP_INIT_DS_LIMIT;
1034 vmsa->ds.attrib = INIT_DS_ATTRIBS;
1035 vmsa->es = vmsa->ds;
1036 vmsa->fs = vmsa->ds;
1037 vmsa->gs = vmsa->ds;
1038 vmsa->ss = vmsa->ds;
1040 vmsa->gdtr.limit = AP_INIT_GDTR_LIMIT;
1041 vmsa->ldtr.limit = AP_INIT_LDTR_LIMIT;
1042 vmsa->ldtr.attrib = INIT_LDTR_ATTRIBS;
1043 vmsa->idtr.limit = AP_INIT_IDTR_LIMIT;
1044 vmsa->tr.limit = AP_INIT_TR_LIMIT;
1045 vmsa->tr.attrib = INIT_TR_ATTRIBS;
1048 vmsa->cr0 = AP_INIT_CR0_DEFAULT;
1049 vmsa->dr7 = DR7_RESET_VALUE;
1050 vmsa->dr6 = AP_INIT_DR6_DEFAULT;
1051 vmsa->rflags = AP_INIT_RFLAGS_DEFAULT;
1052 vmsa->g_pat = AP_INIT_GPAT_DEFAULT;
1053 vmsa->xcr0 = AP_INIT_XCR0_DEFAULT;
1054 vmsa->mxcsr = AP_INIT_MXCSR_DEFAULT;
1055 vmsa->x87_ftw = AP_INIT_X87_FTW_DEFAULT;
1056 vmsa->x87_fcw = AP_INIT_X87_FCW_DEFAULT;
1058 /* SVME must be set. */
1059 vmsa->efer = EFER_SVME;
1062 * Set the SNP-specific fields for this VMSA:
1064 * SEV_FEATURES (matches the SEV STATUS MSR right shifted 2 bits)
1067 vmsa->sev_features = sev_status >> 2;
1069 /* Switch the page over to a VMSA page now that it is initialized */
1070 ret = snp_set_vmsa(vmsa, true);
1072 pr_err("set VMSA page failed (%u)\n", ret);
1073 free_page((unsigned long)vmsa);
1078 /* Issue VMGEXIT AP Creation NAE event */
1079 local_irq_save(flags);
1081 ghcb = __sev_get_ghcb(&state);
1083 vc_ghcb_invalidate(ghcb);
1084 ghcb_set_rax(ghcb, vmsa->sev_features);
1085 ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_CREATION);
1086 ghcb_set_sw_exit_info_1(ghcb, ((u64)apic_id << 32) | SVM_VMGEXIT_AP_CREATE);
1087 ghcb_set_sw_exit_info_2(ghcb, __pa(vmsa));
1089 sev_es_wr_ghcb_msr(__pa(ghcb));
1092 if (!ghcb_sw_exit_info_1_is_valid(ghcb) ||
1093 lower_32_bits(ghcb->save.sw_exit_info_1)) {
1094 pr_err("SNP AP Creation error\n");
1098 __sev_put_ghcb(&state);
1100 local_irq_restore(flags);
1102 /* Perform cleanup if there was an error */
1104 snp_cleanup_vmsa(vmsa);
1108 /* Free up any previous VMSA page */
1110 snp_cleanup_vmsa(cur_vmsa);
1112 /* Record the current VMSA page */
1113 per_cpu(sev_vmsa, cpu) = vmsa;
1118 void __init snp_set_wakeup_secondary_cpu(void)
1120 if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
1124 * Always set this override if SNP is enabled. This makes it the
1125 * required method to start APs under SNP. If the hypervisor does
1126 * not support AP creation, then no APs will be started.
1128 apic_update_callback(wakeup_secondary_cpu, wakeup_cpu_via_vmgexit);
1131 int __init sev_es_setup_ap_jump_table(struct real_mode_header *rmh)
1133 u16 startup_cs, startup_ip;
1134 phys_addr_t jump_table_pa;
1135 u64 jump_table_addr;
1136 u16 __iomem *jump_table;
1138 jump_table_addr = get_jump_table_addr();
1140 /* On UP guests there is no jump table so this is not a failure */
1141 if (!jump_table_addr)
1144 /* Check if AP Jump Table is page-aligned */
1145 if (jump_table_addr & ~PAGE_MASK)
1148 jump_table_pa = jump_table_addr & PAGE_MASK;
1150 startup_cs = (u16)(rmh->trampoline_start >> 4);
1151 startup_ip = (u16)(rmh->sev_es_trampoline_start -
1152 rmh->trampoline_start);
1154 jump_table = ioremap_encrypted(jump_table_pa, PAGE_SIZE);
1158 writew(startup_ip, &jump_table[0]);
1159 writew(startup_cs, &jump_table[1]);
1161 iounmap(jump_table);
1167 * This is needed by the OVMF UEFI firmware which will use whatever it finds in
1168 * the GHCB MSR as its GHCB to talk to the hypervisor. So make sure the per-cpu
1169 * runtime GHCBs used by the kernel are also mapped in the EFI page-table.
1171 int __init sev_es_efi_map_ghcbs(pgd_t *pgd)
1173 struct sev_es_runtime_data *data;
1174 unsigned long address, pflags;
1178 if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
1181 pflags = _PAGE_NX | _PAGE_RW;
1183 for_each_possible_cpu(cpu) {
1184 data = per_cpu(runtime_data, cpu);
1186 address = __pa(&data->ghcb_page);
1187 pfn = address >> PAGE_SHIFT;
1189 if (kernel_map_pages_in_pgd(pgd, pfn, address, 1, pflags))
1196 static enum es_result vc_handle_msr(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
1198 struct pt_regs *regs = ctxt->regs;
1202 /* Is it a WRMSR? */
1203 exit_info_1 = (ctxt->insn.opcode.bytes[1] == 0x30) ? 1 : 0;
1205 ghcb_set_rcx(ghcb, regs->cx);
1207 ghcb_set_rax(ghcb, regs->ax);
1208 ghcb_set_rdx(ghcb, regs->dx);
1211 ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_MSR, exit_info_1, 0);
1213 if ((ret == ES_OK) && (!exit_info_1)) {
1214 regs->ax = ghcb->save.rax;
1215 regs->dx = ghcb->save.rdx;
1221 static void snp_register_per_cpu_ghcb(void)
1223 struct sev_es_runtime_data *data;
1226 data = this_cpu_read(runtime_data);
1227 ghcb = &data->ghcb_page;
1229 snp_register_ghcb_early(__pa(ghcb));
1232 void setup_ghcb(void)
1234 if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
1237 /* First make sure the hypervisor talks a supported protocol. */
1238 if (!sev_es_negotiate_protocol())
1239 sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
1242 * Check whether the runtime #VC exception handler is active. It uses
1243 * the per-CPU GHCB page which is set up by sev_es_init_vc_handling().
1245 * If SNP is active, register the per-CPU GHCB page so that the runtime
1246 * exception handler can use it.
1248 if (initial_vc_handler == (unsigned long)kernel_exc_vmm_communication) {
1249 if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
1250 snp_register_per_cpu_ghcb();
1252 sev_cfg.ghcbs_initialized = true;
1258 * Clear the boot_ghcb. The first exception comes in before the bss
1259 * section is cleared.
1261 memset(&boot_ghcb_page, 0, PAGE_SIZE);
1263 /* Alright - Make the boot-ghcb public */
1264 boot_ghcb = &boot_ghcb_page;
1266 /* SNP guest requires that GHCB GPA must be registered. */
1267 if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
1268 snp_register_ghcb_early(__pa(&boot_ghcb_page));
1271 #ifdef CONFIG_HOTPLUG_CPU
1272 static void sev_es_ap_hlt_loop(void)
1274 struct ghcb_state state;
1277 ghcb = __sev_get_ghcb(&state);
1280 vc_ghcb_invalidate(ghcb);
1281 ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_AP_HLT_LOOP);
1282 ghcb_set_sw_exit_info_1(ghcb, 0);
1283 ghcb_set_sw_exit_info_2(ghcb, 0);
1285 sev_es_wr_ghcb_msr(__pa(ghcb));
1288 /* Wakeup signal? */
1289 if (ghcb_sw_exit_info_2_is_valid(ghcb) &&
1290 ghcb->save.sw_exit_info_2)
1294 __sev_put_ghcb(&state);
1298 * Play_dead handler when running under SEV-ES. This is needed because
1299 * the hypervisor can't deliver an SIPI request to restart the AP.
1300 * Instead the kernel has to issue a VMGEXIT to halt the VCPU until the
1301 * hypervisor wakes it up again.
1303 static void sev_es_play_dead(void)
1307 /* IRQs now disabled */
1309 sev_es_ap_hlt_loop();
1312 * If we get here, the VCPU was woken up again. Jump to CPU
1313 * startup code to get it back online.
1317 #else /* CONFIG_HOTPLUG_CPU */
1318 #define sev_es_play_dead native_play_dead
1319 #endif /* CONFIG_HOTPLUG_CPU */
1322 static void __init sev_es_setup_play_dead(void)
1324 smp_ops.play_dead = sev_es_play_dead;
1327 static inline void sev_es_setup_play_dead(void) { }
1330 static void __init alloc_runtime_data(int cpu)
1332 struct sev_es_runtime_data *data;
1334 data = memblock_alloc(sizeof(*data), PAGE_SIZE);
1336 panic("Can't allocate SEV-ES runtime data");
1338 per_cpu(runtime_data, cpu) = data;
1341 static void __init init_ghcb(int cpu)
1343 struct sev_es_runtime_data *data;
1346 data = per_cpu(runtime_data, cpu);
1348 err = early_set_memory_decrypted((unsigned long)&data->ghcb_page,
1349 sizeof(data->ghcb_page));
1351 panic("Can't map GHCBs unencrypted");
1353 memset(&data->ghcb_page, 0, sizeof(data->ghcb_page));
1355 data->ghcb_active = false;
1356 data->backup_ghcb_active = false;
1359 void __init sev_es_init_vc_handling(void)
1363 BUILD_BUG_ON(offsetof(struct sev_es_runtime_data, ghcb_page) % PAGE_SIZE);
1365 if (!cc_platform_has(CC_ATTR_GUEST_STATE_ENCRYPT))
1368 if (!sev_es_check_cpu_features())
1369 panic("SEV-ES CPU Features missing");
1372 * SNP is supported in v2 of the GHCB spec which mandates support for HV
1375 if (cc_platform_has(CC_ATTR_GUEST_SEV_SNP)) {
1376 sev_hv_features = get_hv_features();
1378 if (!(sev_hv_features & GHCB_HV_FT_SNP))
1379 sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SNP_UNSUPPORTED);
1382 /* Initialize per-cpu GHCB pages */
1383 for_each_possible_cpu(cpu) {
1384 alloc_runtime_data(cpu);
1388 sev_es_setup_play_dead();
1390 /* Secondary CPUs use the runtime #VC handler */
1391 initial_vc_handler = (unsigned long)kernel_exc_vmm_communication;
1394 static void __init vc_early_forward_exception(struct es_em_ctxt *ctxt)
1396 int trapnr = ctxt->fi.vector;
1398 if (trapnr == X86_TRAP_PF)
1399 native_write_cr2(ctxt->fi.cr2);
1401 ctxt->regs->orig_ax = ctxt->fi.error_code;
1402 do_early_exception(ctxt->regs, trapnr);
1405 static long *vc_insn_get_rm(struct es_em_ctxt *ctxt)
1410 reg_array = (long *)ctxt->regs;
1411 offset = insn_get_modrm_rm_off(&ctxt->insn, ctxt->regs);
1416 offset /= sizeof(long);
1418 return reg_array + offset;
1420 static enum es_result vc_do_mmio(struct ghcb *ghcb, struct es_em_ctxt *ctxt,
1421 unsigned int bytes, bool read)
1423 u64 exit_code, exit_info_1, exit_info_2;
1424 unsigned long ghcb_pa = __pa(ghcb);
1429 ref = insn_get_addr_ref(&ctxt->insn, ctxt->regs);
1430 if (ref == (void __user *)-1L)
1431 return ES_UNSUPPORTED;
1433 exit_code = read ? SVM_VMGEXIT_MMIO_READ : SVM_VMGEXIT_MMIO_WRITE;
1435 res = vc_slow_virt_to_phys(ghcb, ctxt, (unsigned long)ref, &paddr);
1437 if (res == ES_EXCEPTION && !read)
1438 ctxt->fi.error_code |= X86_PF_WRITE;
1443 exit_info_1 = paddr;
1444 /* Can never be greater than 8 */
1445 exit_info_2 = bytes;
1447 ghcb_set_sw_scratch(ghcb, ghcb_pa + offsetof(struct ghcb, shared_buffer));
1449 return sev_es_ghcb_hv_call(ghcb, ctxt, exit_code, exit_info_1, exit_info_2);
1453 * The MOVS instruction has two memory operands, which raises the
1454 * problem that it is not known whether the access to the source or the
1455 * destination caused the #VC exception (and hence whether an MMIO read
1456 * or write operation needs to be emulated).
1458 * Instead of playing games with walking page-tables and trying to guess
1459 * whether the source or destination is an MMIO range, split the move
1460 * into two operations, a read and a write with only one memory operand.
1461 * This will cause a nested #VC exception on the MMIO address which can
1464 * This implementation has the benefit that it also supports MOVS where
1465 * source _and_ destination are MMIO regions.
1467 * It will slow MOVS on MMIO down a lot, but in SEV-ES guests it is a
1468 * rare operation. If it turns out to be a performance problem the split
1469 * operations can be moved to memcpy_fromio() and memcpy_toio().
1471 static enum es_result vc_handle_mmio_movs(struct es_em_ctxt *ctxt,
1474 unsigned long ds_base, es_base;
1475 unsigned char *src, *dst;
1476 unsigned char buffer[8];
1481 ds_base = insn_get_seg_base(ctxt->regs, INAT_SEG_REG_DS);
1482 es_base = insn_get_seg_base(ctxt->regs, INAT_SEG_REG_ES);
1484 if (ds_base == -1L || es_base == -1L) {
1485 ctxt->fi.vector = X86_TRAP_GP;
1486 ctxt->fi.error_code = 0;
1487 return ES_EXCEPTION;
1490 src = ds_base + (unsigned char *)ctxt->regs->si;
1491 dst = es_base + (unsigned char *)ctxt->regs->di;
1493 ret = vc_read_mem(ctxt, src, buffer, bytes);
1497 ret = vc_write_mem(ctxt, dst, buffer, bytes);
1501 if (ctxt->regs->flags & X86_EFLAGS_DF)
1506 ctxt->regs->si += off;
1507 ctxt->regs->di += off;
1509 rep = insn_has_rep_prefix(&ctxt->insn);
1511 ctxt->regs->cx -= 1;
1513 if (!rep || ctxt->regs->cx == 0)
1519 static enum es_result vc_handle_mmio(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
1521 struct insn *insn = &ctxt->insn;
1522 enum insn_mmio_type mmio;
1523 unsigned int bytes = 0;
1528 mmio = insn_decode_mmio(insn, &bytes);
1529 if (mmio == INSN_MMIO_DECODE_FAILED)
1530 return ES_DECODE_FAILED;
1532 if (mmio != INSN_MMIO_WRITE_IMM && mmio != INSN_MMIO_MOVS) {
1533 reg_data = insn_get_modrm_reg_ptr(insn, ctxt->regs);
1535 return ES_DECODE_FAILED;
1538 if (user_mode(ctxt->regs))
1539 return ES_UNSUPPORTED;
1542 case INSN_MMIO_WRITE:
1543 memcpy(ghcb->shared_buffer, reg_data, bytes);
1544 ret = vc_do_mmio(ghcb, ctxt, bytes, false);
1546 case INSN_MMIO_WRITE_IMM:
1547 memcpy(ghcb->shared_buffer, insn->immediate1.bytes, bytes);
1548 ret = vc_do_mmio(ghcb, ctxt, bytes, false);
1550 case INSN_MMIO_READ:
1551 ret = vc_do_mmio(ghcb, ctxt, bytes, true);
1555 /* Zero-extend for 32-bit operation */
1559 memcpy(reg_data, ghcb->shared_buffer, bytes);
1561 case INSN_MMIO_READ_ZERO_EXTEND:
1562 ret = vc_do_mmio(ghcb, ctxt, bytes, true);
1566 /* Zero extend based on operand size */
1567 memset(reg_data, 0, insn->opnd_bytes);
1568 memcpy(reg_data, ghcb->shared_buffer, bytes);
1570 case INSN_MMIO_READ_SIGN_EXTEND:
1571 ret = vc_do_mmio(ghcb, ctxt, bytes, true);
1576 u8 *val = (u8 *)ghcb->shared_buffer;
1578 sign_byte = (*val & 0x80) ? 0xff : 0x00;
1580 u16 *val = (u16 *)ghcb->shared_buffer;
1582 sign_byte = (*val & 0x8000) ? 0xff : 0x00;
1585 /* Sign extend based on operand size */
1586 memset(reg_data, sign_byte, insn->opnd_bytes);
1587 memcpy(reg_data, ghcb->shared_buffer, bytes);
1589 case INSN_MMIO_MOVS:
1590 ret = vc_handle_mmio_movs(ctxt, bytes);
1593 ret = ES_UNSUPPORTED;
1600 static enum es_result vc_handle_dr7_write(struct ghcb *ghcb,
1601 struct es_em_ctxt *ctxt)
1603 struct sev_es_runtime_data *data = this_cpu_read(runtime_data);
1604 long val, *reg = vc_insn_get_rm(ctxt);
1607 if (sev_status & MSR_AMD64_SNP_DEBUG_SWAP)
1608 return ES_VMM_ERROR;
1611 return ES_DECODE_FAILED;
1615 /* Upper 32 bits must be written as zeroes */
1617 ctxt->fi.vector = X86_TRAP_GP;
1618 ctxt->fi.error_code = 0;
1619 return ES_EXCEPTION;
1622 /* Clear out other reserved bits and set bit 10 */
1623 val = (val & 0xffff23ffL) | BIT(10);
1625 /* Early non-zero writes to DR7 are not supported */
1626 if (!data && (val & ~DR7_RESET_VALUE))
1627 return ES_UNSUPPORTED;
1629 /* Using a value of 0 for ExitInfo1 means RAX holds the value */
1630 ghcb_set_rax(ghcb, val);
1631 ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_WRITE_DR7, 0, 0);
1641 static enum es_result vc_handle_dr7_read(struct ghcb *ghcb,
1642 struct es_em_ctxt *ctxt)
1644 struct sev_es_runtime_data *data = this_cpu_read(runtime_data);
1645 long *reg = vc_insn_get_rm(ctxt);
1647 if (sev_status & MSR_AMD64_SNP_DEBUG_SWAP)
1648 return ES_VMM_ERROR;
1651 return ES_DECODE_FAILED;
1656 *reg = DR7_RESET_VALUE;
1661 static enum es_result vc_handle_wbinvd(struct ghcb *ghcb,
1662 struct es_em_ctxt *ctxt)
1664 return sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_WBINVD, 0, 0);
1667 static enum es_result vc_handle_rdpmc(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
1671 ghcb_set_rcx(ghcb, ctxt->regs->cx);
1673 ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_RDPMC, 0, 0);
1677 if (!(ghcb_rax_is_valid(ghcb) && ghcb_rdx_is_valid(ghcb)))
1678 return ES_VMM_ERROR;
1680 ctxt->regs->ax = ghcb->save.rax;
1681 ctxt->regs->dx = ghcb->save.rdx;
1686 static enum es_result vc_handle_monitor(struct ghcb *ghcb,
1687 struct es_em_ctxt *ctxt)
1690 * Treat it as a NOP and do not leak a physical address to the
1696 static enum es_result vc_handle_mwait(struct ghcb *ghcb,
1697 struct es_em_ctxt *ctxt)
1699 /* Treat the same as MONITOR/MONITORX */
1703 static enum es_result vc_handle_vmmcall(struct ghcb *ghcb,
1704 struct es_em_ctxt *ctxt)
1708 ghcb_set_rax(ghcb, ctxt->regs->ax);
1709 ghcb_set_cpl(ghcb, user_mode(ctxt->regs) ? 3 : 0);
1711 if (x86_platform.hyper.sev_es_hcall_prepare)
1712 x86_platform.hyper.sev_es_hcall_prepare(ghcb, ctxt->regs);
1714 ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_VMMCALL, 0, 0);
1718 if (!ghcb_rax_is_valid(ghcb))
1719 return ES_VMM_ERROR;
1721 ctxt->regs->ax = ghcb->save.rax;
1724 * Call sev_es_hcall_finish() after regs->ax is already set.
1725 * This allows the hypervisor handler to overwrite it again if
1728 if (x86_platform.hyper.sev_es_hcall_finish &&
1729 !x86_platform.hyper.sev_es_hcall_finish(ghcb, ctxt->regs))
1730 return ES_VMM_ERROR;
1735 static enum es_result vc_handle_trap_ac(struct ghcb *ghcb,
1736 struct es_em_ctxt *ctxt)
1739 * Calling ecx_alignment_check() directly does not work, because it
1740 * enables IRQs and the GHCB is active. Forward the exception and call
1741 * it later from vc_forward_exception().
1743 ctxt->fi.vector = X86_TRAP_AC;
1744 ctxt->fi.error_code = 0;
1745 return ES_EXCEPTION;
1748 static enum es_result vc_handle_exitcode(struct es_em_ctxt *ctxt,
1750 unsigned long exit_code)
1752 enum es_result result;
1754 switch (exit_code) {
1755 case SVM_EXIT_READ_DR7:
1756 result = vc_handle_dr7_read(ghcb, ctxt);
1758 case SVM_EXIT_WRITE_DR7:
1759 result = vc_handle_dr7_write(ghcb, ctxt);
1761 case SVM_EXIT_EXCP_BASE + X86_TRAP_AC:
1762 result = vc_handle_trap_ac(ghcb, ctxt);
1764 case SVM_EXIT_RDTSC:
1765 case SVM_EXIT_RDTSCP:
1766 result = vc_handle_rdtsc(ghcb, ctxt, exit_code);
1768 case SVM_EXIT_RDPMC:
1769 result = vc_handle_rdpmc(ghcb, ctxt);
1772 pr_err_ratelimited("#VC exception for INVD??? Seriously???\n");
1773 result = ES_UNSUPPORTED;
1775 case SVM_EXIT_CPUID:
1776 result = vc_handle_cpuid(ghcb, ctxt);
1779 result = vc_handle_ioio(ghcb, ctxt);
1782 result = vc_handle_msr(ghcb, ctxt);
1784 case SVM_EXIT_VMMCALL:
1785 result = vc_handle_vmmcall(ghcb, ctxt);
1787 case SVM_EXIT_WBINVD:
1788 result = vc_handle_wbinvd(ghcb, ctxt);
1790 case SVM_EXIT_MONITOR:
1791 result = vc_handle_monitor(ghcb, ctxt);
1793 case SVM_EXIT_MWAIT:
1794 result = vc_handle_mwait(ghcb, ctxt);
1797 result = vc_handle_mmio(ghcb, ctxt);
1801 * Unexpected #VC exception
1803 result = ES_UNSUPPORTED;
1809 static __always_inline void vc_forward_exception(struct es_em_ctxt *ctxt)
1811 long error_code = ctxt->fi.error_code;
1812 int trapnr = ctxt->fi.vector;
1814 ctxt->regs->orig_ax = ctxt->fi.error_code;
1818 exc_general_protection(ctxt->regs, error_code);
1821 exc_invalid_op(ctxt->regs);
1824 write_cr2(ctxt->fi.cr2);
1825 exc_page_fault(ctxt->regs, error_code);
1828 exc_alignment_check(ctxt->regs, error_code);
1831 pr_emerg("Unsupported exception in #VC instruction emulation - can't continue\n");
1836 static __always_inline bool is_vc2_stack(unsigned long sp)
1838 return (sp >= __this_cpu_ist_bottom_va(VC2) && sp < __this_cpu_ist_top_va(VC2));
1841 static __always_inline bool vc_from_invalid_context(struct pt_regs *regs)
1843 unsigned long sp, prev_sp;
1845 sp = (unsigned long)regs;
1849 * If the code was already executing on the VC2 stack when the #VC
1850 * happened, let it proceed to the normal handling routine. This way the
1851 * code executing on the VC2 stack can cause #VC exceptions to get handled.
1853 return is_vc2_stack(sp) && !is_vc2_stack(prev_sp);
1856 static bool vc_raw_handle_exception(struct pt_regs *regs, unsigned long error_code)
1858 struct ghcb_state state;
1859 struct es_em_ctxt ctxt;
1860 enum es_result result;
1864 ghcb = __sev_get_ghcb(&state);
1866 vc_ghcb_invalidate(ghcb);
1867 result = vc_init_em_ctxt(&ctxt, regs, error_code);
1869 if (result == ES_OK)
1870 result = vc_handle_exitcode(&ctxt, ghcb, error_code);
1872 __sev_put_ghcb(&state);
1874 /* Done - now check the result */
1877 vc_finish_insn(&ctxt);
1879 case ES_UNSUPPORTED:
1880 pr_err_ratelimited("Unsupported exit-code 0x%02lx in #VC exception (IP: 0x%lx)\n",
1881 error_code, regs->ip);
1885 pr_err_ratelimited("Failure in communication with VMM (exit-code 0x%02lx IP: 0x%lx)\n",
1886 error_code, regs->ip);
1889 case ES_DECODE_FAILED:
1890 pr_err_ratelimited("Failed to decode instruction (exit-code 0x%02lx IP: 0x%lx)\n",
1891 error_code, regs->ip);
1895 vc_forward_exception(&ctxt);
1901 pr_emerg("Unknown result in %s():%d\n", __func__, result);
1903 * Emulating the instruction which caused the #VC exception
1904 * failed - can't continue so print debug information
1912 static __always_inline bool vc_is_db(unsigned long error_code)
1914 return error_code == SVM_EXIT_EXCP_BASE + X86_TRAP_DB;
1918 * Runtime #VC exception handler when raised from kernel mode. Runs in NMI mode
1919 * and will panic when an error happens.
1921 DEFINE_IDTENTRY_VC_KERNEL(exc_vmm_communication)
1923 irqentry_state_t irq_state;
1926 * With the current implementation it is always possible to switch to a
1927 * safe stack because #VC exceptions only happen at known places, like
1928 * intercepted instructions or accesses to MMIO areas/IO ports. They can
1929 * also happen with code instrumentation when the hypervisor intercepts
1930 * #DB, but the critical paths are forbidden to be instrumented, so #DB
1931 * exceptions currently also only happen in safe places.
1933 * But keep this here in case the noinstr annotations are violated due
1936 if (unlikely(vc_from_invalid_context(regs))) {
1937 instrumentation_begin();
1938 panic("Can't handle #VC exception from unsupported context\n");
1939 instrumentation_end();
1943 * Handle #DB before calling into !noinstr code to avoid recursive #DB.
1945 if (vc_is_db(error_code)) {
1950 irq_state = irqentry_nmi_enter(regs);
1952 instrumentation_begin();
1954 if (!vc_raw_handle_exception(regs, error_code)) {
1955 /* Show some debug info */
1958 /* Ask hypervisor to sev_es_terminate */
1959 sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
1961 /* If that fails and we get here - just panic */
1962 panic("Returned from Terminate-Request to Hypervisor\n");
1965 instrumentation_end();
1966 irqentry_nmi_exit(regs, irq_state);
1970 * Runtime #VC exception handler when raised from user mode. Runs in IRQ mode
1971 * and will kill the current task with SIGBUS when an error happens.
1973 DEFINE_IDTENTRY_VC_USER(exc_vmm_communication)
1976 * Handle #DB before calling into !noinstr code to avoid recursive #DB.
1978 if (vc_is_db(error_code)) {
1979 noist_exc_debug(regs);
1983 irqentry_enter_from_user_mode(regs);
1984 instrumentation_begin();
1986 if (!vc_raw_handle_exception(regs, error_code)) {
1988 * Do not kill the machine if user-space triggered the
1989 * exception. Send SIGBUS instead and let user-space deal with
1992 force_sig_fault(SIGBUS, BUS_OBJERR, (void __user *)0);
1995 instrumentation_end();
1996 irqentry_exit_to_user_mode(regs);
1999 bool __init handle_vc_boot_ghcb(struct pt_regs *regs)
2001 unsigned long exit_code = regs->orig_ax;
2002 struct es_em_ctxt ctxt;
2003 enum es_result result;
2005 vc_ghcb_invalidate(boot_ghcb);
2007 result = vc_init_em_ctxt(&ctxt, regs, exit_code);
2008 if (result == ES_OK)
2009 result = vc_handle_exitcode(&ctxt, boot_ghcb, exit_code);
2011 /* Done - now check the result */
2014 vc_finish_insn(&ctxt);
2016 case ES_UNSUPPORTED:
2017 early_printk("PANIC: Unsupported exit-code 0x%02lx in early #VC exception (IP: 0x%lx)\n",
2018 exit_code, regs->ip);
2021 early_printk("PANIC: Failure in communication with VMM (exit-code 0x%02lx IP: 0x%lx)\n",
2022 exit_code, regs->ip);
2024 case ES_DECODE_FAILED:
2025 early_printk("PANIC: Failed to decode instruction (exit-code 0x%02lx IP: 0x%lx)\n",
2026 exit_code, regs->ip);
2029 vc_early_forward_exception(&ctxt);
2043 sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
2047 * Initial set up of SNP relies on information provided by the
2048 * Confidential Computing blob, which can be passed to the kernel
2049 * in the following ways, depending on how it is booted:
2051 * - when booted via the boot/decompress kernel:
2054 * - when booted directly by firmware/bootloader (e.g. CONFIG_PVH):
2055 * - via a setup_data entry, as defined by the Linux Boot Protocol
2057 * Scan for the blob in that order.
2059 static __init struct cc_blob_sev_info *find_cc_blob(struct boot_params *bp)
2061 struct cc_blob_sev_info *cc_info;
2063 /* Boot kernel would have passed the CC blob via boot_params. */
2064 if (bp->cc_blob_address) {
2065 cc_info = (struct cc_blob_sev_info *)(unsigned long)bp->cc_blob_address;
2070 * If kernel was booted directly, without the use of the
2071 * boot/decompression kernel, the CC blob may have been passed via
2072 * setup_data instead.
2074 cc_info = find_cc_blob_setup_data(bp);
2079 if (cc_info->magic != CC_BLOB_SEV_HDR_MAGIC)
2085 bool __init snp_init(struct boot_params *bp)
2087 struct cc_blob_sev_info *cc_info;
2092 cc_info = find_cc_blob(bp);
2096 setup_cpuid_table(cc_info);
2099 * The CC blob will be used later to access the secrets page. Cache
2100 * it here like the boot kernel does.
2102 bp->cc_blob_address = (u32)(unsigned long)cc_info;
2107 void __init __noreturn snp_abort(void)
2109 sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SNP_UNSUPPORTED);
2112 static void dump_cpuid_table(void)
2114 const struct snp_cpuid_table *cpuid_table = snp_cpuid_get_table();
2117 pr_info("count=%d reserved=0x%x reserved2=0x%llx\n",
2118 cpuid_table->count, cpuid_table->__reserved1, cpuid_table->__reserved2);
2120 for (i = 0; i < SNP_CPUID_COUNT_MAX; i++) {
2121 const struct snp_cpuid_fn *fn = &cpuid_table->fn[i];
2123 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",
2124 i, fn->eax_in, fn->ecx_in, fn->eax, fn->ebx, fn->ecx,
2125 fn->edx, fn->xcr0_in, fn->xss_in, fn->__reserved);
2130 * It is useful from an auditing/testing perspective to provide an easy way
2131 * for the guest owner to know that the CPUID table has been initialized as
2132 * expected, but that initialization happens too early in boot to print any
2133 * sort of indicator, and there's not really any other good place to do it,
2136 static int __init report_cpuid_table(void)
2138 const struct snp_cpuid_table *cpuid_table = snp_cpuid_get_table();
2140 if (!cpuid_table->count)
2143 pr_info("Using SNP CPUID table, %d entries present.\n",
2144 cpuid_table->count);
2151 arch_initcall(report_cpuid_table);
2153 static int __init init_sev_config(char *str)
2157 while ((s = strsep(&str, ","))) {
2158 if (!strcmp(s, "debug")) {
2159 sev_cfg.debug = true;
2163 pr_info("SEV command-line option '%s' was not recognized\n", s);
2168 __setup("sev=", init_sev_config);
2170 int snp_issue_guest_request(u64 exit_code, struct snp_req_data *input, struct snp_guest_request_ioctl *rio)
2172 struct ghcb_state state;
2173 struct es_em_ctxt ctxt;
2174 unsigned long flags;
2178 rio->exitinfo2 = SEV_RET_NO_FW_CALL;
2181 * __sev_get_ghcb() needs to run with IRQs disabled because it is using
2184 local_irq_save(flags);
2186 ghcb = __sev_get_ghcb(&state);
2192 vc_ghcb_invalidate(ghcb);
2194 if (exit_code == SVM_VMGEXIT_EXT_GUEST_REQUEST) {
2195 ghcb_set_rax(ghcb, input->data_gpa);
2196 ghcb_set_rbx(ghcb, input->data_npages);
2199 ret = sev_es_ghcb_hv_call(ghcb, &ctxt, exit_code, input->req_gpa, input->resp_gpa);
2203 rio->exitinfo2 = ghcb->save.sw_exit_info_2;
2204 switch (rio->exitinfo2) {
2208 case SNP_GUEST_VMM_ERR(SNP_GUEST_VMM_ERR_BUSY):
2212 case SNP_GUEST_VMM_ERR(SNP_GUEST_VMM_ERR_INVALID_LEN):
2213 /* Number of expected pages are returned in RBX */
2214 if (exit_code == SVM_VMGEXIT_EXT_GUEST_REQUEST) {
2215 input->data_npages = ghcb_get_rbx(ghcb);
2226 __sev_put_ghcb(&state);
2228 local_irq_restore(flags);
2232 EXPORT_SYMBOL_GPL(snp_issue_guest_request);
2234 static struct platform_device sev_guest_device = {
2235 .name = "sev-guest",
2239 static int __init snp_init_platform_device(void)
2241 struct sev_guest_platform_data data;
2244 if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP))
2247 gpa = get_secrets_page();
2251 data.secrets_gpa = gpa;
2252 if (platform_device_add_data(&sev_guest_device, &data, sizeof(data)))
2255 if (platform_device_register(&sev_guest_device))
2258 pr_info("SNP guest platform device initialized.\n");
2261 device_initcall(snp_init_platform_device);