1 #define pr_fmt(fmt) "SVM: " fmt
3 #include <linux/kvm_host.h>
7 #include "kvm_cache_regs.h"
12 #include <linux/module.h>
13 #include <linux/mod_devicetable.h>
14 #include <linux/kernel.h>
15 #include <linux/vmalloc.h>
16 #include <linux/highmem.h>
17 #include <linux/amd-iommu.h>
18 #include <linux/sched.h>
19 #include <linux/trace_events.h>
20 #include <linux/slab.h>
21 #include <linux/hashtable.h>
22 #include <linux/objtool.h>
23 #include <linux/psp-sev.h>
24 #include <linux/file.h>
25 #include <linux/pagemap.h>
26 #include <linux/swap.h>
27 #include <linux/rwsem.h>
30 #include <asm/perf_event.h>
31 #include <asm/tlbflush.h>
33 #include <asm/debugreg.h>
34 #include <asm/kvm_para.h>
35 #include <asm/irq_remapping.h>
36 #include <asm/spec-ctrl.h>
37 #include <asm/cpu_device_id.h>
38 #include <asm/traps.h>
40 #include <asm/virtext.h>
46 #define __ex(x) __kvm_handle_fault_on_reboot(x)
48 MODULE_AUTHOR("Qumranet");
49 MODULE_LICENSE("GPL");
52 static const struct x86_cpu_id svm_cpu_id[] = {
53 X86_MATCH_FEATURE(X86_FEATURE_SVM, NULL),
56 MODULE_DEVICE_TABLE(x86cpu, svm_cpu_id);
59 #define SEG_TYPE_LDT 2
60 #define SEG_TYPE_BUSY_TSS16 3
62 #define SVM_FEATURE_LBRV (1 << 1)
63 #define SVM_FEATURE_SVML (1 << 2)
64 #define SVM_FEATURE_TSC_RATE (1 << 4)
65 #define SVM_FEATURE_VMCB_CLEAN (1 << 5)
66 #define SVM_FEATURE_FLUSH_ASID (1 << 6)
67 #define SVM_FEATURE_DECODE_ASSIST (1 << 7)
68 #define SVM_FEATURE_PAUSE_FILTER (1 << 10)
70 #define DEBUGCTL_RESERVED_BITS (~(0x3fULL))
72 #define TSC_RATIO_RSVD 0xffffff0000000000ULL
73 #define TSC_RATIO_MIN 0x0000000000000001ULL
74 #define TSC_RATIO_MAX 0x000000ffffffffffULL
76 static bool erratum_383_found __read_mostly;
78 u32 msrpm_offsets[MSRPM_OFFSETS] __read_mostly;
81 * Set osvw_len to higher value when updated Revision Guides
82 * are published and we know what the new status bits are
84 static uint64_t osvw_len = 4, osvw_status;
86 static DEFINE_PER_CPU(u64, current_tsc_ratio);
87 #define TSC_RATIO_DEFAULT 0x0100000000ULL
89 static const struct svm_direct_access_msrs {
90 u32 index; /* Index of the MSR */
91 bool always; /* True if intercept is initially cleared */
92 } direct_access_msrs[MAX_DIRECT_ACCESS_MSRS] = {
93 { .index = MSR_STAR, .always = true },
94 { .index = MSR_IA32_SYSENTER_CS, .always = true },
95 { .index = MSR_IA32_SYSENTER_EIP, .always = false },
96 { .index = MSR_IA32_SYSENTER_ESP, .always = false },
98 { .index = MSR_GS_BASE, .always = true },
99 { .index = MSR_FS_BASE, .always = true },
100 { .index = MSR_KERNEL_GS_BASE, .always = true },
101 { .index = MSR_LSTAR, .always = true },
102 { .index = MSR_CSTAR, .always = true },
103 { .index = MSR_SYSCALL_MASK, .always = true },
105 { .index = MSR_IA32_SPEC_CTRL, .always = false },
106 { .index = MSR_IA32_PRED_CMD, .always = false },
107 { .index = MSR_IA32_LASTBRANCHFROMIP, .always = false },
108 { .index = MSR_IA32_LASTBRANCHTOIP, .always = false },
109 { .index = MSR_IA32_LASTINTFROMIP, .always = false },
110 { .index = MSR_IA32_LASTINTTOIP, .always = false },
111 { .index = MSR_EFER, .always = false },
112 { .index = MSR_IA32_CR_PAT, .always = false },
113 { .index = MSR_AMD64_SEV_ES_GHCB, .always = true },
114 { .index = MSR_INVALID, .always = false },
118 * These 2 parameters are used to config the controls for Pause-Loop Exiting:
119 * pause_filter_count: On processors that support Pause filtering(indicated
120 * by CPUID Fn8000_000A_EDX), the VMCB provides a 16 bit pause filter
121 * count value. On VMRUN this value is loaded into an internal counter.
122 * Each time a pause instruction is executed, this counter is decremented
123 * until it reaches zero at which time a #VMEXIT is generated if pause
124 * intercept is enabled. Refer to AMD APM Vol 2 Section 15.14.4 Pause
125 * Intercept Filtering for more details.
126 * This also indicate if ple logic enabled.
128 * pause_filter_thresh: In addition, some processor families support advanced
129 * pause filtering (indicated by CPUID Fn8000_000A_EDX) upper bound on
130 * the amount of time a guest is allowed to execute in a pause loop.
131 * In this mode, a 16-bit pause filter threshold field is added in the
132 * VMCB. The threshold value is a cycle count that is used to reset the
133 * pause counter. As with simple pause filtering, VMRUN loads the pause
134 * count value from VMCB into an internal counter. Then, on each pause
135 * instruction the hardware checks the elapsed number of cycles since
136 * the most recent pause instruction against the pause filter threshold.
137 * If the elapsed cycle count is greater than the pause filter threshold,
138 * then the internal pause count is reloaded from the VMCB and execution
139 * continues. If the elapsed cycle count is less than the pause filter
140 * threshold, then the internal pause count is decremented. If the count
141 * value is less than zero and PAUSE intercept is enabled, a #VMEXIT is
142 * triggered. If advanced pause filtering is supported and pause filter
143 * threshold field is set to zero, the filter will operate in the simpler,
147 static unsigned short pause_filter_thresh = KVM_DEFAULT_PLE_GAP;
148 module_param(pause_filter_thresh, ushort, 0444);
150 static unsigned short pause_filter_count = KVM_SVM_DEFAULT_PLE_WINDOW;
151 module_param(pause_filter_count, ushort, 0444);
153 /* Default doubles per-vcpu window every exit. */
154 static unsigned short pause_filter_count_grow = KVM_DEFAULT_PLE_WINDOW_GROW;
155 module_param(pause_filter_count_grow, ushort, 0444);
157 /* Default resets per-vcpu window every exit to pause_filter_count. */
158 static unsigned short pause_filter_count_shrink = KVM_DEFAULT_PLE_WINDOW_SHRINK;
159 module_param(pause_filter_count_shrink, ushort, 0444);
161 /* Default is to compute the maximum so we can never overflow. */
162 static unsigned short pause_filter_count_max = KVM_SVM_DEFAULT_PLE_WINDOW_MAX;
163 module_param(pause_filter_count_max, ushort, 0444);
166 * Use nested page tables by default. Note, NPT may get forced off by
167 * svm_hardware_setup() if it's unsupported by hardware or the host kernel.
169 bool npt_enabled = true;
170 module_param_named(npt, npt_enabled, bool, 0444);
172 /* allow nested virtualization in KVM/SVM */
173 static int nested = true;
174 module_param(nested, int, S_IRUGO);
176 /* enable/disable Next RIP Save */
177 static int nrips = true;
178 module_param(nrips, int, 0444);
180 /* enable/disable Virtual VMLOAD VMSAVE */
181 static int vls = true;
182 module_param(vls, int, 0444);
184 /* enable/disable Virtual GIF */
185 static int vgif = true;
186 module_param(vgif, int, 0444);
188 bool __read_mostly dump_invalid_vmcb;
189 module_param(dump_invalid_vmcb, bool, 0644);
191 static bool svm_gp_erratum_intercept = true;
193 static u8 rsm_ins_bytes[] = "\x0f\xaa";
195 static unsigned long iopm_base;
197 struct kvm_ldttss_desc {
200 unsigned base1:8, type:5, dpl:2, p:1;
201 unsigned limit1:4, zero0:3, g:1, base2:8;
204 } __attribute__((packed));
206 DEFINE_PER_CPU(struct svm_cpu_data *, svm_data);
209 * Only MSR_TSC_AUX is switched via the user return hook. EFER is switched via
210 * the VMCB, and the SYSCALL/SYSENTER MSRs are handled by VMLOAD/VMSAVE.
212 * RDTSCP and RDPID are not used in the kernel, specifically to allow KVM to
213 * defer the restoration of TSC_AUX until the CPU returns to userspace.
215 static int tsc_aux_uret_slot __read_mostly = -1;
217 static const u32 msrpm_ranges[] = {0, 0xc0000000, 0xc0010000};
219 #define NUM_MSR_MAPS ARRAY_SIZE(msrpm_ranges)
220 #define MSRS_RANGE_SIZE 2048
221 #define MSRS_IN_RANGE (MSRS_RANGE_SIZE * 8 / 2)
223 u32 svm_msrpm_offset(u32 msr)
228 for (i = 0; i < NUM_MSR_MAPS; i++) {
229 if (msr < msrpm_ranges[i] ||
230 msr >= msrpm_ranges[i] + MSRS_IN_RANGE)
233 offset = (msr - msrpm_ranges[i]) / 4; /* 4 msrs per u8 */
234 offset += (i * MSRS_RANGE_SIZE); /* add range offset */
236 /* Now we have the u8 offset - but need the u32 offset */
240 /* MSR not in any range */
244 #define MAX_INST_SIZE 15
246 static int get_max_npt_level(void)
249 return PT64_ROOT_4LEVEL;
251 return PT32E_ROOT_LEVEL;
255 int svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)
257 struct vcpu_svm *svm = to_svm(vcpu);
258 u64 old_efer = vcpu->arch.efer;
259 vcpu->arch.efer = efer;
262 /* Shadow paging assumes NX to be available. */
265 if (!(efer & EFER_LMA))
269 if ((old_efer & EFER_SVME) != (efer & EFER_SVME)) {
270 if (!(efer & EFER_SVME)) {
271 svm_leave_nested(svm);
272 svm_set_gif(svm, true);
273 /* #GP intercept is still needed for vmware backdoor */
274 if (!enable_vmware_backdoor)
275 clr_exception_intercept(svm, GP_VECTOR);
278 * Free the nested guest state, unless we are in SMM.
279 * In this case we will return to the nested guest
280 * as soon as we leave SMM.
283 svm_free_nested(svm);
286 int ret = svm_allocate_nested(svm);
289 vcpu->arch.efer = old_efer;
293 if (svm_gp_erratum_intercept)
294 set_exception_intercept(svm, GP_VECTOR);
298 svm->vmcb->save.efer = efer | EFER_SVME;
299 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
303 static int is_external_interrupt(u32 info)
305 info &= SVM_EVTINJ_TYPE_MASK | SVM_EVTINJ_VALID;
306 return info == (SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR);
309 static u32 svm_get_interrupt_shadow(struct kvm_vcpu *vcpu)
311 struct vcpu_svm *svm = to_svm(vcpu);
314 if (svm->vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK)
315 ret = KVM_X86_SHADOW_INT_STI | KVM_X86_SHADOW_INT_MOV_SS;
319 static void svm_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
321 struct vcpu_svm *svm = to_svm(vcpu);
324 svm->vmcb->control.int_state &= ~SVM_INTERRUPT_SHADOW_MASK;
326 svm->vmcb->control.int_state |= SVM_INTERRUPT_SHADOW_MASK;
330 static int skip_emulated_instruction(struct kvm_vcpu *vcpu)
332 struct vcpu_svm *svm = to_svm(vcpu);
335 * SEV-ES does not expose the next RIP. The RIP update is controlled by
336 * the type of exit and the #VC handler in the guest.
338 if (sev_es_guest(vcpu->kvm))
341 if (nrips && svm->vmcb->control.next_rip != 0) {
342 WARN_ON_ONCE(!static_cpu_has(X86_FEATURE_NRIPS));
343 svm->next_rip = svm->vmcb->control.next_rip;
346 if (!svm->next_rip) {
347 if (!kvm_emulate_instruction(vcpu, EMULTYPE_SKIP))
350 kvm_rip_write(vcpu, svm->next_rip);
354 svm_set_interrupt_shadow(vcpu, 0);
359 static void svm_queue_exception(struct kvm_vcpu *vcpu)
361 struct vcpu_svm *svm = to_svm(vcpu);
362 unsigned nr = vcpu->arch.exception.nr;
363 bool has_error_code = vcpu->arch.exception.has_error_code;
364 u32 error_code = vcpu->arch.exception.error_code;
366 kvm_deliver_exception_payload(vcpu);
368 if (nr == BP_VECTOR && !nrips) {
369 unsigned long rip, old_rip = kvm_rip_read(vcpu);
372 * For guest debugging where we have to reinject #BP if some
373 * INT3 is guest-owned:
374 * Emulate nRIP by moving RIP forward. Will fail if injection
375 * raises a fault that is not intercepted. Still better than
376 * failing in all cases.
378 (void)skip_emulated_instruction(vcpu);
379 rip = kvm_rip_read(vcpu);
380 svm->int3_rip = rip + svm->vmcb->save.cs.base;
381 svm->int3_injected = rip - old_rip;
384 svm->vmcb->control.event_inj = nr
386 | (has_error_code ? SVM_EVTINJ_VALID_ERR : 0)
387 | SVM_EVTINJ_TYPE_EXEPT;
388 svm->vmcb->control.event_inj_err = error_code;
391 static void svm_init_erratum_383(void)
397 if (!static_cpu_has_bug(X86_BUG_AMD_TLB_MMATCH))
400 /* Use _safe variants to not break nested virtualization */
401 val = native_read_msr_safe(MSR_AMD64_DC_CFG, &err);
407 low = lower_32_bits(val);
408 high = upper_32_bits(val);
410 native_write_msr_safe(MSR_AMD64_DC_CFG, low, high);
412 erratum_383_found = true;
415 static void svm_init_osvw(struct kvm_vcpu *vcpu)
418 * Guests should see errata 400 and 415 as fixed (assuming that
419 * HLT and IO instructions are intercepted).
421 vcpu->arch.osvw.length = (osvw_len >= 3) ? (osvw_len) : 3;
422 vcpu->arch.osvw.status = osvw_status & ~(6ULL);
425 * By increasing VCPU's osvw.length to 3 we are telling the guest that
426 * all osvw.status bits inside that length, including bit 0 (which is
427 * reserved for erratum 298), are valid. However, if host processor's
428 * osvw_len is 0 then osvw_status[0] carries no information. We need to
429 * be conservative here and therefore we tell the guest that erratum 298
430 * is present (because we really don't know).
432 if (osvw_len == 0 && boot_cpu_data.x86 == 0x10)
433 vcpu->arch.osvw.status |= 1;
436 static int has_svm(void)
440 if (!cpu_has_svm(&msg)) {
441 printk(KERN_INFO "has_svm: %s\n", msg);
446 pr_info("KVM is unsupported when running as an SEV guest\n");
453 static void svm_hardware_disable(void)
455 /* Make sure we clean up behind us */
456 if (static_cpu_has(X86_FEATURE_TSCRATEMSR))
457 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
461 amd_pmu_disable_virt();
464 static int svm_hardware_enable(void)
467 struct svm_cpu_data *sd;
469 struct desc_struct *gdt;
470 int me = raw_smp_processor_id();
472 rdmsrl(MSR_EFER, efer);
473 if (efer & EFER_SVME)
477 pr_err("%s: err EOPNOTSUPP on %d\n", __func__, me);
480 sd = per_cpu(svm_data, me);
482 pr_err("%s: svm_data is NULL on %d\n", __func__, me);
486 sd->asid_generation = 1;
487 sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
488 sd->next_asid = sd->max_asid + 1;
489 sd->min_asid = max_sev_asid + 1;
491 gdt = get_current_gdt_rw();
492 sd->tss_desc = (struct kvm_ldttss_desc *)(gdt + GDT_ENTRY_TSS);
494 wrmsrl(MSR_EFER, efer | EFER_SVME);
496 wrmsrl(MSR_VM_HSAVE_PA, __sme_page_pa(sd->save_area));
498 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
499 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
500 __this_cpu_write(current_tsc_ratio, TSC_RATIO_DEFAULT);
507 * Note that it is possible to have a system with mixed processor
508 * revisions and therefore different OSVW bits. If bits are not the same
509 * on different processors then choose the worst case (i.e. if erratum
510 * is present on one processor and not on another then assume that the
511 * erratum is present everywhere).
513 if (cpu_has(&boot_cpu_data, X86_FEATURE_OSVW)) {
514 uint64_t len, status = 0;
517 len = native_read_msr_safe(MSR_AMD64_OSVW_ID_LENGTH, &err);
519 status = native_read_msr_safe(MSR_AMD64_OSVW_STATUS,
523 osvw_status = osvw_len = 0;
527 osvw_status |= status;
528 osvw_status &= (1ULL << osvw_len) - 1;
531 osvw_status = osvw_len = 0;
533 svm_init_erratum_383();
535 amd_pmu_enable_virt();
540 static void svm_cpu_uninit(int cpu)
542 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
547 per_cpu(svm_data, cpu) = NULL;
548 kfree(sd->sev_vmcbs);
549 __free_page(sd->save_area);
553 static int svm_cpu_init(int cpu)
555 struct svm_cpu_data *sd;
558 sd = kzalloc(sizeof(struct svm_cpu_data), GFP_KERNEL);
562 sd->save_area = alloc_page(GFP_KERNEL);
566 clear_page(page_address(sd->save_area));
568 ret = sev_cpu_init(sd);
572 per_cpu(svm_data, cpu) = sd;
577 __free_page(sd->save_area);
584 static int direct_access_msr_slot(u32 msr)
588 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++)
589 if (direct_access_msrs[i].index == msr)
595 static void set_shadow_msr_intercept(struct kvm_vcpu *vcpu, u32 msr, int read,
598 struct vcpu_svm *svm = to_svm(vcpu);
599 int slot = direct_access_msr_slot(msr);
604 /* Set the shadow bitmaps to the desired intercept states */
606 set_bit(slot, svm->shadow_msr_intercept.read);
608 clear_bit(slot, svm->shadow_msr_intercept.read);
611 set_bit(slot, svm->shadow_msr_intercept.write);
613 clear_bit(slot, svm->shadow_msr_intercept.write);
616 static bool valid_msr_intercept(u32 index)
618 return direct_access_msr_slot(index) != -ENOENT;
621 static bool msr_write_intercepted(struct kvm_vcpu *vcpu, u32 msr)
628 msrpm = is_guest_mode(vcpu) ? to_svm(vcpu)->nested.msrpm:
631 offset = svm_msrpm_offset(msr);
632 bit_write = 2 * (msr & 0x0f) + 1;
635 BUG_ON(offset == MSR_INVALID);
637 return !!test_bit(bit_write, &tmp);
640 static void set_msr_interception_bitmap(struct kvm_vcpu *vcpu, u32 *msrpm,
641 u32 msr, int read, int write)
643 u8 bit_read, bit_write;
648 * If this warning triggers extend the direct_access_msrs list at the
649 * beginning of the file
651 WARN_ON(!valid_msr_intercept(msr));
653 /* Enforce non allowed MSRs to trap */
654 if (read && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_READ))
657 if (write && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_WRITE))
660 offset = svm_msrpm_offset(msr);
661 bit_read = 2 * (msr & 0x0f);
662 bit_write = 2 * (msr & 0x0f) + 1;
665 BUG_ON(offset == MSR_INVALID);
667 read ? clear_bit(bit_read, &tmp) : set_bit(bit_read, &tmp);
668 write ? clear_bit(bit_write, &tmp) : set_bit(bit_write, &tmp);
673 void set_msr_interception(struct kvm_vcpu *vcpu, u32 *msrpm, u32 msr,
676 set_shadow_msr_intercept(vcpu, msr, read, write);
677 set_msr_interception_bitmap(vcpu, msrpm, msr, read, write);
680 u32 *svm_vcpu_alloc_msrpm(void)
682 unsigned int order = get_order(MSRPM_SIZE);
683 struct page *pages = alloc_pages(GFP_KERNEL_ACCOUNT, order);
689 msrpm = page_address(pages);
690 memset(msrpm, 0xff, PAGE_SIZE * (1 << order));
695 void svm_vcpu_init_msrpm(struct kvm_vcpu *vcpu, u32 *msrpm)
699 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
700 if (!direct_access_msrs[i].always)
702 set_msr_interception(vcpu, msrpm, direct_access_msrs[i].index, 1, 1);
707 void svm_vcpu_free_msrpm(u32 *msrpm)
709 __free_pages(virt_to_page(msrpm), get_order(MSRPM_SIZE));
712 static void svm_msr_filter_changed(struct kvm_vcpu *vcpu)
714 struct vcpu_svm *svm = to_svm(vcpu);
718 * Set intercept permissions for all direct access MSRs again. They
719 * will automatically get filtered through the MSR filter, so we are
720 * back in sync after this.
722 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
723 u32 msr = direct_access_msrs[i].index;
724 u32 read = test_bit(i, svm->shadow_msr_intercept.read);
725 u32 write = test_bit(i, svm->shadow_msr_intercept.write);
727 set_msr_interception_bitmap(vcpu, svm->msrpm, msr, read, write);
731 static void add_msr_offset(u32 offset)
735 for (i = 0; i < MSRPM_OFFSETS; ++i) {
737 /* Offset already in list? */
738 if (msrpm_offsets[i] == offset)
741 /* Slot used by another offset? */
742 if (msrpm_offsets[i] != MSR_INVALID)
745 /* Add offset to list */
746 msrpm_offsets[i] = offset;
752 * If this BUG triggers the msrpm_offsets table has an overflow. Just
753 * increase MSRPM_OFFSETS in this case.
758 static void init_msrpm_offsets(void)
762 memset(msrpm_offsets, 0xff, sizeof(msrpm_offsets));
764 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
767 offset = svm_msrpm_offset(direct_access_msrs[i].index);
768 BUG_ON(offset == MSR_INVALID);
770 add_msr_offset(offset);
774 static void svm_enable_lbrv(struct kvm_vcpu *vcpu)
776 struct vcpu_svm *svm = to_svm(vcpu);
778 svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
779 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
780 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
781 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
782 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
785 static void svm_disable_lbrv(struct kvm_vcpu *vcpu)
787 struct vcpu_svm *svm = to_svm(vcpu);
789 svm->vmcb->control.virt_ext &= ~LBR_CTL_ENABLE_MASK;
790 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 0, 0);
791 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 0, 0);
792 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 0, 0);
793 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 0, 0);
796 void disable_nmi_singlestep(struct vcpu_svm *svm)
798 svm->nmi_singlestep = false;
800 if (!(svm->vcpu.guest_debug & KVM_GUESTDBG_SINGLESTEP)) {
801 /* Clear our flags if they were not set by the guest */
802 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
803 svm->vmcb->save.rflags &= ~X86_EFLAGS_TF;
804 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
805 svm->vmcb->save.rflags &= ~X86_EFLAGS_RF;
809 static void grow_ple_window(struct kvm_vcpu *vcpu)
811 struct vcpu_svm *svm = to_svm(vcpu);
812 struct vmcb_control_area *control = &svm->vmcb->control;
813 int old = control->pause_filter_count;
815 control->pause_filter_count = __grow_ple_window(old,
817 pause_filter_count_grow,
818 pause_filter_count_max);
820 if (control->pause_filter_count != old) {
821 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
822 trace_kvm_ple_window_update(vcpu->vcpu_id,
823 control->pause_filter_count, old);
827 static void shrink_ple_window(struct kvm_vcpu *vcpu)
829 struct vcpu_svm *svm = to_svm(vcpu);
830 struct vmcb_control_area *control = &svm->vmcb->control;
831 int old = control->pause_filter_count;
833 control->pause_filter_count =
834 __shrink_ple_window(old,
836 pause_filter_count_shrink,
838 if (control->pause_filter_count != old) {
839 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
840 trace_kvm_ple_window_update(vcpu->vcpu_id,
841 control->pause_filter_count, old);
846 * The default MMIO mask is a single bit (excluding the present bit),
847 * which could conflict with the memory encryption bit. Check for
848 * memory encryption support and override the default MMIO mask if
849 * memory encryption is enabled.
851 static __init void svm_adjust_mmio_mask(void)
853 unsigned int enc_bit, mask_bit;
856 /* If there is no memory encryption support, use existing mask */
857 if (cpuid_eax(0x80000000) < 0x8000001f)
860 /* If memory encryption is not enabled, use existing mask */
861 rdmsrl(MSR_K8_SYSCFG, msr);
862 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
865 enc_bit = cpuid_ebx(0x8000001f) & 0x3f;
866 mask_bit = boot_cpu_data.x86_phys_bits;
868 /* Increment the mask bit if it is the same as the encryption bit */
869 if (enc_bit == mask_bit)
873 * If the mask bit location is below 52, then some bits above the
874 * physical addressing limit will always be reserved, so use the
875 * rsvd_bits() function to generate the mask. This mask, along with
876 * the present bit, will be used to generate a page fault with
879 * If the mask bit location is 52 (or above), then clear the mask.
881 mask = (mask_bit < 52) ? rsvd_bits(mask_bit, 51) | PT_PRESENT_MASK : 0;
883 kvm_mmu_set_mmio_spte_mask(mask, mask, PT_WRITABLE_MASK | PT_USER_MASK);
886 static void svm_hardware_teardown(void)
890 sev_hardware_teardown();
892 for_each_possible_cpu(cpu)
895 __free_pages(pfn_to_page(iopm_base >> PAGE_SHIFT),
896 get_order(IOPM_SIZE));
900 static __init void svm_set_cpu_caps(void)
906 /* CPUID 0x80000001 and 0x8000000A (SVM features) */
908 kvm_cpu_cap_set(X86_FEATURE_SVM);
911 kvm_cpu_cap_set(X86_FEATURE_NRIPS);
914 kvm_cpu_cap_set(X86_FEATURE_NPT);
916 /* Nested VM can receive #VMEXIT instead of triggering #GP */
917 kvm_cpu_cap_set(X86_FEATURE_SVME_ADDR_CHK);
920 /* CPUID 0x80000008 */
921 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) ||
922 boot_cpu_has(X86_FEATURE_AMD_SSBD))
923 kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
925 /* CPUID 0x8000001F (SME/SEV features) */
929 static __init int svm_hardware_setup(void)
932 struct page *iopm_pages;
935 unsigned int order = get_order(IOPM_SIZE);
937 iopm_pages = alloc_pages(GFP_KERNEL, order);
942 iopm_va = page_address(iopm_pages);
943 memset(iopm_va, 0xff, PAGE_SIZE * (1 << order));
944 iopm_base = page_to_pfn(iopm_pages) << PAGE_SHIFT;
946 init_msrpm_offsets();
948 supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
950 if (boot_cpu_has(X86_FEATURE_NX))
951 kvm_enable_efer_bits(EFER_NX);
953 if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
954 kvm_enable_efer_bits(EFER_FFXSR);
956 if (boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
957 kvm_has_tsc_control = true;
958 kvm_max_tsc_scaling_ratio = TSC_RATIO_MAX;
959 kvm_tsc_scaling_ratio_frac_bits = 32;
962 tsc_aux_uret_slot = kvm_add_user_return_msr(MSR_TSC_AUX);
964 /* Check for pause filtering support */
965 if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
966 pause_filter_count = 0;
967 pause_filter_thresh = 0;
968 } else if (!boot_cpu_has(X86_FEATURE_PFTHRESHOLD)) {
969 pause_filter_thresh = 0;
973 printk(KERN_INFO "kvm: Nested Virtualization enabled\n");
974 kvm_enable_efer_bits(EFER_SVME | EFER_LMSLE);
978 * KVM's MMU doesn't support using 2-level paging for itself, and thus
979 * NPT isn't supported if the host is using 2-level paging since host
980 * CR4 is unchanged on VMRUN.
982 if (!IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_X86_PAE))
985 if (!boot_cpu_has(X86_FEATURE_NPT))
988 kvm_configure_mmu(npt_enabled, get_max_npt_level(), PG_LEVEL_1G);
989 pr_info("kvm: Nested Paging %sabled\n", npt_enabled ? "en" : "dis");
991 /* Note, SEV setup consumes npt_enabled. */
992 sev_hardware_setup();
994 svm_adjust_mmio_mask();
996 for_each_possible_cpu(cpu) {
997 r = svm_cpu_init(cpu);
1003 if (!boot_cpu_has(X86_FEATURE_NRIPS))
1009 !boot_cpu_has(X86_FEATURE_AVIC) ||
1010 !IS_ENABLED(CONFIG_X86_LOCAL_APIC)) {
1013 pr_info("AVIC enabled\n");
1015 amd_iommu_register_ga_log_notifier(&avic_ga_log_notifier);
1021 !boot_cpu_has(X86_FEATURE_V_VMSAVE_VMLOAD) ||
1022 !IS_ENABLED(CONFIG_X86_64)) {
1025 pr_info("Virtual VMLOAD VMSAVE supported\n");
1029 if (boot_cpu_has(X86_FEATURE_SVME_ADDR_CHK))
1030 svm_gp_erratum_intercept = false;
1033 if (!boot_cpu_has(X86_FEATURE_VGIF))
1036 pr_info("Virtual GIF supported\n");
1042 * It seems that on AMD processors PTE's accessed bit is
1043 * being set by the CPU hardware before the NPF vmexit.
1044 * This is not expected behaviour and our tests fail because
1046 * A workaround here is to disable support for
1047 * GUEST_MAXPHYADDR < HOST_MAXPHYADDR if NPT is enabled.
1048 * In this case userspace can know if there is support using
1049 * KVM_CAP_SMALLER_MAXPHYADDR extension and decide how to handle
1051 * If future AMD CPU models change the behaviour described above,
1052 * this variable can be changed accordingly
1054 allow_smaller_maxphyaddr = !npt_enabled;
1059 svm_hardware_teardown();
1063 static void init_seg(struct vmcb_seg *seg)
1066 seg->attrib = SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK |
1067 SVM_SELECTOR_WRITE_MASK; /* Read/Write Data Segment */
1068 seg->limit = 0xffff;
1072 static void init_sys_seg(struct vmcb_seg *seg, uint32_t type)
1075 seg->attrib = SVM_SELECTOR_P_MASK | type;
1076 seg->limit = 0xffff;
1080 static u64 svm_write_l1_tsc_offset(struct kvm_vcpu *vcpu, u64 offset)
1082 struct vcpu_svm *svm = to_svm(vcpu);
1083 u64 g_tsc_offset = 0;
1085 if (is_guest_mode(vcpu)) {
1086 /* Write L1's TSC offset. */
1087 g_tsc_offset = svm->vmcb->control.tsc_offset -
1088 svm->vmcb01.ptr->control.tsc_offset;
1089 svm->vmcb01.ptr->control.tsc_offset = offset;
1092 trace_kvm_write_tsc_offset(vcpu->vcpu_id,
1093 svm->vmcb->control.tsc_offset - g_tsc_offset,
1096 svm->vmcb->control.tsc_offset = offset + g_tsc_offset;
1098 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1099 return svm->vmcb->control.tsc_offset;
1102 /* Evaluate instruction intercepts that depend on guest CPUID features. */
1103 static void svm_recalc_instruction_intercepts(struct kvm_vcpu *vcpu,
1104 struct vcpu_svm *svm)
1107 * Intercept INVPCID if shadow paging is enabled to sync/free shadow
1108 * roots, or if INVPCID is disabled in the guest to inject #UD.
1110 if (kvm_cpu_cap_has(X86_FEATURE_INVPCID)) {
1112 !guest_cpuid_has(&svm->vcpu, X86_FEATURE_INVPCID))
1113 svm_set_intercept(svm, INTERCEPT_INVPCID);
1115 svm_clr_intercept(svm, INTERCEPT_INVPCID);
1118 if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP)) {
1119 if (guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP))
1120 svm_clr_intercept(svm, INTERCEPT_RDTSCP);
1122 svm_set_intercept(svm, INTERCEPT_RDTSCP);
1126 static void init_vmcb(struct kvm_vcpu *vcpu)
1128 struct vcpu_svm *svm = to_svm(vcpu);
1129 struct vmcb_control_area *control = &svm->vmcb->control;
1130 struct vmcb_save_area *save = &svm->vmcb->save;
1132 vcpu->arch.hflags = 0;
1134 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1135 svm_set_intercept(svm, INTERCEPT_CR3_READ);
1136 svm_set_intercept(svm, INTERCEPT_CR4_READ);
1137 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1138 svm_set_intercept(svm, INTERCEPT_CR3_WRITE);
1139 svm_set_intercept(svm, INTERCEPT_CR4_WRITE);
1140 if (!kvm_vcpu_apicv_active(vcpu))
1141 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
1143 set_dr_intercepts(svm);
1145 set_exception_intercept(svm, PF_VECTOR);
1146 set_exception_intercept(svm, UD_VECTOR);
1147 set_exception_intercept(svm, MC_VECTOR);
1148 set_exception_intercept(svm, AC_VECTOR);
1149 set_exception_intercept(svm, DB_VECTOR);
1151 * Guest access to VMware backdoor ports could legitimately
1152 * trigger #GP because of TSS I/O permission bitmap.
1153 * We intercept those #GP and allow access to them anyway
1156 if (enable_vmware_backdoor)
1157 set_exception_intercept(svm, GP_VECTOR);
1159 svm_set_intercept(svm, INTERCEPT_INTR);
1160 svm_set_intercept(svm, INTERCEPT_NMI);
1161 svm_set_intercept(svm, INTERCEPT_SMI);
1162 svm_set_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1163 svm_set_intercept(svm, INTERCEPT_RDPMC);
1164 svm_set_intercept(svm, INTERCEPT_CPUID);
1165 svm_set_intercept(svm, INTERCEPT_INVD);
1166 svm_set_intercept(svm, INTERCEPT_INVLPG);
1167 svm_set_intercept(svm, INTERCEPT_INVLPGA);
1168 svm_set_intercept(svm, INTERCEPT_IOIO_PROT);
1169 svm_set_intercept(svm, INTERCEPT_MSR_PROT);
1170 svm_set_intercept(svm, INTERCEPT_TASK_SWITCH);
1171 svm_set_intercept(svm, INTERCEPT_SHUTDOWN);
1172 svm_set_intercept(svm, INTERCEPT_VMRUN);
1173 svm_set_intercept(svm, INTERCEPT_VMMCALL);
1174 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1175 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1176 svm_set_intercept(svm, INTERCEPT_STGI);
1177 svm_set_intercept(svm, INTERCEPT_CLGI);
1178 svm_set_intercept(svm, INTERCEPT_SKINIT);
1179 svm_set_intercept(svm, INTERCEPT_WBINVD);
1180 svm_set_intercept(svm, INTERCEPT_XSETBV);
1181 svm_set_intercept(svm, INTERCEPT_RDPRU);
1182 svm_set_intercept(svm, INTERCEPT_RSM);
1184 if (!kvm_mwait_in_guest(vcpu->kvm)) {
1185 svm_set_intercept(svm, INTERCEPT_MONITOR);
1186 svm_set_intercept(svm, INTERCEPT_MWAIT);
1189 if (!kvm_hlt_in_guest(vcpu->kvm))
1190 svm_set_intercept(svm, INTERCEPT_HLT);
1192 control->iopm_base_pa = __sme_set(iopm_base);
1193 control->msrpm_base_pa = __sme_set(__pa(svm->msrpm));
1194 control->int_ctl = V_INTR_MASKING_MASK;
1196 init_seg(&save->es);
1197 init_seg(&save->ss);
1198 init_seg(&save->ds);
1199 init_seg(&save->fs);
1200 init_seg(&save->gs);
1202 save->cs.selector = 0xf000;
1203 save->cs.base = 0xffff0000;
1204 /* Executable/Readable Code Segment */
1205 save->cs.attrib = SVM_SELECTOR_READ_MASK | SVM_SELECTOR_P_MASK |
1206 SVM_SELECTOR_S_MASK | SVM_SELECTOR_CODE_MASK;
1207 save->cs.limit = 0xffff;
1209 save->gdtr.limit = 0xffff;
1210 save->idtr.limit = 0xffff;
1212 init_sys_seg(&save->ldtr, SEG_TYPE_LDT);
1213 init_sys_seg(&save->tr, SEG_TYPE_BUSY_TSS16);
1215 svm_set_cr4(vcpu, 0);
1216 svm_set_efer(vcpu, 0);
1217 save->dr6 = 0xffff0ff0;
1218 kvm_set_rflags(vcpu, X86_EFLAGS_FIXED);
1219 save->rip = 0x0000fff0;
1220 vcpu->arch.regs[VCPU_REGS_RIP] = save->rip;
1223 * svm_set_cr0() sets PG and WP and clears NW and CD on save->cr0.
1224 * It also updates the guest-visible cr0 value.
1226 svm_set_cr0(vcpu, X86_CR0_NW | X86_CR0_CD | X86_CR0_ET);
1227 kvm_mmu_reset_context(vcpu);
1229 save->cr4 = X86_CR4_PAE;
1233 /* Setup VMCB for Nested Paging */
1234 control->nested_ctl |= SVM_NESTED_CTL_NP_ENABLE;
1235 svm_clr_intercept(svm, INTERCEPT_INVLPG);
1236 clr_exception_intercept(svm, PF_VECTOR);
1237 svm_clr_intercept(svm, INTERCEPT_CR3_READ);
1238 svm_clr_intercept(svm, INTERCEPT_CR3_WRITE);
1239 save->g_pat = vcpu->arch.pat;
1243 svm->current_vmcb->asid_generation = 0;
1246 svm->nested.vmcb12_gpa = INVALID_GPA;
1247 svm->nested.last_vmcb12_gpa = INVALID_GPA;
1248 vcpu->arch.hflags = 0;
1250 if (!kvm_pause_in_guest(vcpu->kvm)) {
1251 control->pause_filter_count = pause_filter_count;
1252 if (pause_filter_thresh)
1253 control->pause_filter_thresh = pause_filter_thresh;
1254 svm_set_intercept(svm, INTERCEPT_PAUSE);
1256 svm_clr_intercept(svm, INTERCEPT_PAUSE);
1259 svm_recalc_instruction_intercepts(vcpu, svm);
1262 * If the host supports V_SPEC_CTRL then disable the interception
1263 * of MSR_IA32_SPEC_CTRL.
1265 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
1266 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
1268 if (kvm_vcpu_apicv_active(vcpu))
1269 avic_init_vmcb(svm);
1272 svm_clr_intercept(svm, INTERCEPT_STGI);
1273 svm_clr_intercept(svm, INTERCEPT_CLGI);
1274 svm->vmcb->control.int_ctl |= V_GIF_ENABLE_MASK;
1277 if (sev_guest(vcpu->kvm)) {
1278 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
1279 clr_exception_intercept(svm, UD_VECTOR);
1281 if (sev_es_guest(vcpu->kvm)) {
1282 /* Perform SEV-ES specific VMCB updates */
1283 sev_es_init_vmcb(svm);
1287 vmcb_mark_all_dirty(svm->vmcb);
1293 static void svm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
1295 struct vcpu_svm *svm = to_svm(vcpu);
1300 svm->virt_spec_ctrl = 0;
1303 vcpu->arch.apic_base = APIC_DEFAULT_PHYS_BASE |
1304 MSR_IA32_APICBASE_ENABLE;
1305 if (kvm_vcpu_is_reset_bsp(vcpu))
1306 vcpu->arch.apic_base |= MSR_IA32_APICBASE_BSP;
1310 kvm_cpuid(vcpu, &eax, &dummy, &dummy, &dummy, false);
1311 kvm_rdx_write(vcpu, eax);
1313 if (kvm_vcpu_apicv_active(vcpu) && !init_event)
1314 avic_update_vapic_bar(svm, APIC_DEFAULT_PHYS_BASE);
1317 void svm_switch_vmcb(struct vcpu_svm *svm, struct kvm_vmcb_info *target_vmcb)
1319 svm->current_vmcb = target_vmcb;
1320 svm->vmcb = target_vmcb->ptr;
1323 static int svm_create_vcpu(struct kvm_vcpu *vcpu)
1325 struct vcpu_svm *svm;
1326 struct page *vmcb01_page;
1327 struct page *vmsa_page = NULL;
1330 BUILD_BUG_ON(offsetof(struct vcpu_svm, vcpu) != 0);
1334 vmcb01_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1338 if (sev_es_guest(vcpu->kvm)) {
1340 * SEV-ES guests require a separate VMSA page used to contain
1341 * the encrypted register state of the guest.
1343 vmsa_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1345 goto error_free_vmcb_page;
1348 * SEV-ES guests maintain an encrypted version of their FPU
1349 * state which is restored and saved on VMRUN and VMEXIT.
1350 * Free the fpu structure to prevent KVM from attempting to
1351 * access the FPU state.
1353 kvm_free_guest_fpu(vcpu);
1356 err = avic_init_vcpu(svm);
1358 goto error_free_vmsa_page;
1360 /* We initialize this flag to true to make sure that the is_running
1361 * bit would be set the first time the vcpu is loaded.
1363 if (irqchip_in_kernel(vcpu->kvm) && kvm_apicv_activated(vcpu->kvm))
1364 svm->avic_is_running = true;
1366 svm->msrpm = svm_vcpu_alloc_msrpm();
1369 goto error_free_vmsa_page;
1372 svm_vcpu_init_msrpm(vcpu, svm->msrpm);
1374 svm->vmcb01.ptr = page_address(vmcb01_page);
1375 svm->vmcb01.pa = __sme_set(page_to_pfn(vmcb01_page) << PAGE_SHIFT);
1378 svm->vmsa = page_address(vmsa_page);
1380 svm->guest_state_loaded = false;
1382 svm_switch_vmcb(svm, &svm->vmcb01);
1385 svm_init_osvw(vcpu);
1386 vcpu->arch.microcode_version = 0x01000065;
1388 if (sev_es_guest(vcpu->kvm))
1389 /* Perform SEV-ES specific VMCB creation updates */
1390 sev_es_create_vcpu(svm);
1394 error_free_vmsa_page:
1396 __free_page(vmsa_page);
1397 error_free_vmcb_page:
1398 __free_page(vmcb01_page);
1403 static void svm_clear_current_vmcb(struct vmcb *vmcb)
1407 for_each_online_cpu(i)
1408 cmpxchg(&per_cpu(svm_data, i)->current_vmcb, vmcb, NULL);
1411 static void svm_free_vcpu(struct kvm_vcpu *vcpu)
1413 struct vcpu_svm *svm = to_svm(vcpu);
1416 * The vmcb page can be recycled, causing a false negative in
1417 * svm_vcpu_load(). So, ensure that no logical CPU has this
1418 * vmcb page recorded as its current vmcb.
1420 svm_clear_current_vmcb(svm->vmcb);
1422 svm_free_nested(svm);
1424 sev_free_vcpu(vcpu);
1426 __free_page(pfn_to_page(__sme_clr(svm->vmcb01.pa) >> PAGE_SHIFT));
1427 __free_pages(virt_to_page(svm->msrpm), get_order(MSRPM_SIZE));
1430 static void svm_prepare_guest_switch(struct kvm_vcpu *vcpu)
1432 struct vcpu_svm *svm = to_svm(vcpu);
1433 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
1435 if (svm->guest_state_loaded)
1439 * Save additional host state that will be restored on VMEXIT (sev-es)
1440 * or subsequent vmload of host save area.
1442 if (sev_es_guest(vcpu->kvm)) {
1443 sev_es_prepare_guest_switch(svm, vcpu->cpu);
1445 vmsave(__sme_page_pa(sd->save_area));
1448 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
1449 u64 tsc_ratio = vcpu->arch.tsc_scaling_ratio;
1450 if (tsc_ratio != __this_cpu_read(current_tsc_ratio)) {
1451 __this_cpu_write(current_tsc_ratio, tsc_ratio);
1452 wrmsrl(MSR_AMD64_TSC_RATIO, tsc_ratio);
1456 if (likely(tsc_aux_uret_slot >= 0))
1457 kvm_set_user_return_msr(tsc_aux_uret_slot, svm->tsc_aux, -1ull);
1459 svm->guest_state_loaded = true;
1462 static void svm_prepare_host_switch(struct kvm_vcpu *vcpu)
1464 to_svm(vcpu)->guest_state_loaded = false;
1467 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1469 struct vcpu_svm *svm = to_svm(vcpu);
1470 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1472 if (sd->current_vmcb != svm->vmcb) {
1473 sd->current_vmcb = svm->vmcb;
1474 indirect_branch_prediction_barrier();
1476 avic_vcpu_load(vcpu, cpu);
1479 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
1481 avic_vcpu_put(vcpu);
1482 svm_prepare_host_switch(vcpu);
1484 ++vcpu->stat.host_state_reload;
1487 static unsigned long svm_get_rflags(struct kvm_vcpu *vcpu)
1489 struct vcpu_svm *svm = to_svm(vcpu);
1490 unsigned long rflags = svm->vmcb->save.rflags;
1492 if (svm->nmi_singlestep) {
1493 /* Hide our flags if they were not set by the guest */
1494 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
1495 rflags &= ~X86_EFLAGS_TF;
1496 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
1497 rflags &= ~X86_EFLAGS_RF;
1502 static void svm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
1504 if (to_svm(vcpu)->nmi_singlestep)
1505 rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
1508 * Any change of EFLAGS.VM is accompanied by a reload of SS
1509 * (caused by either a task switch or an inter-privilege IRET),
1510 * so we do not need to update the CPL here.
1512 to_svm(vcpu)->vmcb->save.rflags = rflags;
1515 static void svm_cache_reg(struct kvm_vcpu *vcpu, enum kvm_reg reg)
1518 case VCPU_EXREG_PDPTR:
1519 BUG_ON(!npt_enabled);
1520 load_pdptrs(vcpu, vcpu->arch.walk_mmu, kvm_read_cr3(vcpu));
1527 static void svm_set_vintr(struct vcpu_svm *svm)
1529 struct vmcb_control_area *control;
1531 /* The following fields are ignored when AVIC is enabled */
1532 WARN_ON(kvm_vcpu_apicv_active(&svm->vcpu));
1533 svm_set_intercept(svm, INTERCEPT_VINTR);
1536 * This is just a dummy VINTR to actually cause a vmexit to happen.
1537 * Actual injection of virtual interrupts happens through EVENTINJ.
1539 control = &svm->vmcb->control;
1540 control->int_vector = 0x0;
1541 control->int_ctl &= ~V_INTR_PRIO_MASK;
1542 control->int_ctl |= V_IRQ_MASK |
1543 ((/*control->int_vector >> 4*/ 0xf) << V_INTR_PRIO_SHIFT);
1544 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1547 static void svm_clear_vintr(struct vcpu_svm *svm)
1549 const u32 mask = V_TPR_MASK | V_GIF_ENABLE_MASK | V_GIF_MASK | V_INTR_MASKING_MASK;
1550 svm_clr_intercept(svm, INTERCEPT_VINTR);
1552 /* Drop int_ctl fields related to VINTR injection. */
1553 svm->vmcb->control.int_ctl &= mask;
1554 if (is_guest_mode(&svm->vcpu)) {
1555 svm->vmcb01.ptr->control.int_ctl &= mask;
1557 WARN_ON((svm->vmcb->control.int_ctl & V_TPR_MASK) !=
1558 (svm->nested.ctl.int_ctl & V_TPR_MASK));
1559 svm->vmcb->control.int_ctl |= svm->nested.ctl.int_ctl & ~mask;
1562 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1565 static struct vmcb_seg *svm_seg(struct kvm_vcpu *vcpu, int seg)
1567 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1568 struct vmcb_save_area *save01 = &to_svm(vcpu)->vmcb01.ptr->save;
1571 case VCPU_SREG_CS: return &save->cs;
1572 case VCPU_SREG_DS: return &save->ds;
1573 case VCPU_SREG_ES: return &save->es;
1574 case VCPU_SREG_FS: return &save01->fs;
1575 case VCPU_SREG_GS: return &save01->gs;
1576 case VCPU_SREG_SS: return &save->ss;
1577 case VCPU_SREG_TR: return &save01->tr;
1578 case VCPU_SREG_LDTR: return &save01->ldtr;
1584 static u64 svm_get_segment_base(struct kvm_vcpu *vcpu, int seg)
1586 struct vmcb_seg *s = svm_seg(vcpu, seg);
1591 static void svm_get_segment(struct kvm_vcpu *vcpu,
1592 struct kvm_segment *var, int seg)
1594 struct vmcb_seg *s = svm_seg(vcpu, seg);
1596 var->base = s->base;
1597 var->limit = s->limit;
1598 var->selector = s->selector;
1599 var->type = s->attrib & SVM_SELECTOR_TYPE_MASK;
1600 var->s = (s->attrib >> SVM_SELECTOR_S_SHIFT) & 1;
1601 var->dpl = (s->attrib >> SVM_SELECTOR_DPL_SHIFT) & 3;
1602 var->present = (s->attrib >> SVM_SELECTOR_P_SHIFT) & 1;
1603 var->avl = (s->attrib >> SVM_SELECTOR_AVL_SHIFT) & 1;
1604 var->l = (s->attrib >> SVM_SELECTOR_L_SHIFT) & 1;
1605 var->db = (s->attrib >> SVM_SELECTOR_DB_SHIFT) & 1;
1608 * AMD CPUs circa 2014 track the G bit for all segments except CS.
1609 * However, the SVM spec states that the G bit is not observed by the
1610 * CPU, and some VMware virtual CPUs drop the G bit for all segments.
1611 * So let's synthesize a legal G bit for all segments, this helps
1612 * running KVM nested. It also helps cross-vendor migration, because
1613 * Intel's vmentry has a check on the 'G' bit.
1615 var->g = s->limit > 0xfffff;
1618 * AMD's VMCB does not have an explicit unusable field, so emulate it
1619 * for cross vendor migration purposes by "not present"
1621 var->unusable = !var->present;
1626 * Work around a bug where the busy flag in the tr selector
1636 * The accessed bit must always be set in the segment
1637 * descriptor cache, although it can be cleared in the
1638 * descriptor, the cached bit always remains at 1. Since
1639 * Intel has a check on this, set it here to support
1640 * cross-vendor migration.
1647 * On AMD CPUs sometimes the DB bit in the segment
1648 * descriptor is left as 1, although the whole segment has
1649 * been made unusable. Clear it here to pass an Intel VMX
1650 * entry check when cross vendor migrating.
1654 /* This is symmetric with svm_set_segment() */
1655 var->dpl = to_svm(vcpu)->vmcb->save.cpl;
1660 static int svm_get_cpl(struct kvm_vcpu *vcpu)
1662 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1667 static void svm_get_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1669 struct vcpu_svm *svm = to_svm(vcpu);
1671 dt->size = svm->vmcb->save.idtr.limit;
1672 dt->address = svm->vmcb->save.idtr.base;
1675 static void svm_set_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1677 struct vcpu_svm *svm = to_svm(vcpu);
1679 svm->vmcb->save.idtr.limit = dt->size;
1680 svm->vmcb->save.idtr.base = dt->address ;
1681 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1684 static void svm_get_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1686 struct vcpu_svm *svm = to_svm(vcpu);
1688 dt->size = svm->vmcb->save.gdtr.limit;
1689 dt->address = svm->vmcb->save.gdtr.base;
1692 static void svm_set_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1694 struct vcpu_svm *svm = to_svm(vcpu);
1696 svm->vmcb->save.gdtr.limit = dt->size;
1697 svm->vmcb->save.gdtr.base = dt->address ;
1698 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1701 void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1703 struct vcpu_svm *svm = to_svm(vcpu);
1706 #ifdef CONFIG_X86_64
1707 if (vcpu->arch.efer & EFER_LME && !vcpu->arch.guest_state_protected) {
1708 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
1709 vcpu->arch.efer |= EFER_LMA;
1710 svm->vmcb->save.efer |= EFER_LMA | EFER_LME;
1713 if (is_paging(vcpu) && !(cr0 & X86_CR0_PG)) {
1714 vcpu->arch.efer &= ~EFER_LMA;
1715 svm->vmcb->save.efer &= ~(EFER_LMA | EFER_LME);
1719 vcpu->arch.cr0 = cr0;
1722 hcr0 |= X86_CR0_PG | X86_CR0_WP;
1725 * re-enable caching here because the QEMU bios
1726 * does not do it - this results in some delay at
1729 if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
1730 hcr0 &= ~(X86_CR0_CD | X86_CR0_NW);
1732 svm->vmcb->save.cr0 = hcr0;
1733 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1736 * SEV-ES guests must always keep the CR intercepts cleared. CR
1737 * tracking is done using the CR write traps.
1739 if (sev_es_guest(vcpu->kvm))
1743 /* Selective CR0 write remains on. */
1744 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1745 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1747 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1748 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1752 static bool svm_is_valid_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1757 void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1759 unsigned long host_cr4_mce = cr4_read_shadow() & X86_CR4_MCE;
1760 unsigned long old_cr4 = vcpu->arch.cr4;
1762 if (npt_enabled && ((old_cr4 ^ cr4) & X86_CR4_PGE))
1763 svm_flush_tlb(vcpu);
1765 vcpu->arch.cr4 = cr4;
1768 cr4 |= host_cr4_mce;
1769 to_svm(vcpu)->vmcb->save.cr4 = cr4;
1770 vmcb_mark_dirty(to_svm(vcpu)->vmcb, VMCB_CR);
1772 if ((cr4 ^ old_cr4) & (X86_CR4_OSXSAVE | X86_CR4_PKE))
1773 kvm_update_cpuid_runtime(vcpu);
1776 static void svm_set_segment(struct kvm_vcpu *vcpu,
1777 struct kvm_segment *var, int seg)
1779 struct vcpu_svm *svm = to_svm(vcpu);
1780 struct vmcb_seg *s = svm_seg(vcpu, seg);
1782 s->base = var->base;
1783 s->limit = var->limit;
1784 s->selector = var->selector;
1785 s->attrib = (var->type & SVM_SELECTOR_TYPE_MASK);
1786 s->attrib |= (var->s & 1) << SVM_SELECTOR_S_SHIFT;
1787 s->attrib |= (var->dpl & 3) << SVM_SELECTOR_DPL_SHIFT;
1788 s->attrib |= ((var->present & 1) && !var->unusable) << SVM_SELECTOR_P_SHIFT;
1789 s->attrib |= (var->avl & 1) << SVM_SELECTOR_AVL_SHIFT;
1790 s->attrib |= (var->l & 1) << SVM_SELECTOR_L_SHIFT;
1791 s->attrib |= (var->db & 1) << SVM_SELECTOR_DB_SHIFT;
1792 s->attrib |= (var->g & 1) << SVM_SELECTOR_G_SHIFT;
1795 * This is always accurate, except if SYSRET returned to a segment
1796 * with SS.DPL != 3. Intel does not have this quirk, and always
1797 * forces SS.DPL to 3 on sysret, so we ignore that case; fixing it
1798 * would entail passing the CPL to userspace and back.
1800 if (seg == VCPU_SREG_SS)
1801 /* This is symmetric with svm_get_segment() */
1802 svm->vmcb->save.cpl = (var->dpl & 3);
1804 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
1807 static void svm_update_exception_bitmap(struct kvm_vcpu *vcpu)
1809 struct vcpu_svm *svm = to_svm(vcpu);
1811 clr_exception_intercept(svm, BP_VECTOR);
1813 if (vcpu->guest_debug & KVM_GUESTDBG_ENABLE) {
1814 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP)
1815 set_exception_intercept(svm, BP_VECTOR);
1819 static void new_asid(struct vcpu_svm *svm, struct svm_cpu_data *sd)
1821 if (sd->next_asid > sd->max_asid) {
1822 ++sd->asid_generation;
1823 sd->next_asid = sd->min_asid;
1824 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
1825 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1828 svm->current_vmcb->asid_generation = sd->asid_generation;
1829 svm->asid = sd->next_asid++;
1832 static void svm_set_dr6(struct vcpu_svm *svm, unsigned long value)
1834 struct vmcb *vmcb = svm->vmcb;
1836 if (svm->vcpu.arch.guest_state_protected)
1839 if (unlikely(value != vmcb->save.dr6)) {
1840 vmcb->save.dr6 = value;
1841 vmcb_mark_dirty(vmcb, VMCB_DR);
1845 static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
1847 struct vcpu_svm *svm = to_svm(vcpu);
1849 if (vcpu->arch.guest_state_protected)
1852 get_debugreg(vcpu->arch.db[0], 0);
1853 get_debugreg(vcpu->arch.db[1], 1);
1854 get_debugreg(vcpu->arch.db[2], 2);
1855 get_debugreg(vcpu->arch.db[3], 3);
1857 * We cannot reset svm->vmcb->save.dr6 to DR6_ACTIVE_LOW here,
1858 * because db_interception might need it. We can do it before vmentry.
1860 vcpu->arch.dr6 = svm->vmcb->save.dr6;
1861 vcpu->arch.dr7 = svm->vmcb->save.dr7;
1862 vcpu->arch.switch_db_regs &= ~KVM_DEBUGREG_WONT_EXIT;
1863 set_dr_intercepts(svm);
1866 static void svm_set_dr7(struct kvm_vcpu *vcpu, unsigned long value)
1868 struct vcpu_svm *svm = to_svm(vcpu);
1870 if (vcpu->arch.guest_state_protected)
1873 svm->vmcb->save.dr7 = value;
1874 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
1877 static int pf_interception(struct kvm_vcpu *vcpu)
1879 struct vcpu_svm *svm = to_svm(vcpu);
1881 u64 fault_address = svm->vmcb->control.exit_info_2;
1882 u64 error_code = svm->vmcb->control.exit_info_1;
1884 return kvm_handle_page_fault(vcpu, error_code, fault_address,
1885 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1886 svm->vmcb->control.insn_bytes : NULL,
1887 svm->vmcb->control.insn_len);
1890 static int npf_interception(struct kvm_vcpu *vcpu)
1892 struct vcpu_svm *svm = to_svm(vcpu);
1894 u64 fault_address = __sme_clr(svm->vmcb->control.exit_info_2);
1895 u64 error_code = svm->vmcb->control.exit_info_1;
1897 trace_kvm_page_fault(fault_address, error_code);
1898 return kvm_mmu_page_fault(vcpu, fault_address, error_code,
1899 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1900 svm->vmcb->control.insn_bytes : NULL,
1901 svm->vmcb->control.insn_len);
1904 static int db_interception(struct kvm_vcpu *vcpu)
1906 struct kvm_run *kvm_run = vcpu->run;
1907 struct vcpu_svm *svm = to_svm(vcpu);
1909 if (!(vcpu->guest_debug &
1910 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) &&
1911 !svm->nmi_singlestep) {
1912 u32 payload = svm->vmcb->save.dr6 ^ DR6_ACTIVE_LOW;
1913 kvm_queue_exception_p(vcpu, DB_VECTOR, payload);
1917 if (svm->nmi_singlestep) {
1918 disable_nmi_singlestep(svm);
1919 /* Make sure we check for pending NMIs upon entry */
1920 kvm_make_request(KVM_REQ_EVENT, vcpu);
1923 if (vcpu->guest_debug &
1924 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) {
1925 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1926 kvm_run->debug.arch.dr6 = svm->vmcb->save.dr6;
1927 kvm_run->debug.arch.dr7 = svm->vmcb->save.dr7;
1928 kvm_run->debug.arch.pc =
1929 svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1930 kvm_run->debug.arch.exception = DB_VECTOR;
1937 static int bp_interception(struct kvm_vcpu *vcpu)
1939 struct vcpu_svm *svm = to_svm(vcpu);
1940 struct kvm_run *kvm_run = vcpu->run;
1942 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1943 kvm_run->debug.arch.pc = svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1944 kvm_run->debug.arch.exception = BP_VECTOR;
1948 static int ud_interception(struct kvm_vcpu *vcpu)
1950 return handle_ud(vcpu);
1953 static int ac_interception(struct kvm_vcpu *vcpu)
1955 kvm_queue_exception_e(vcpu, AC_VECTOR, 0);
1959 static bool is_erratum_383(void)
1964 if (!erratum_383_found)
1967 value = native_read_msr_safe(MSR_IA32_MC0_STATUS, &err);
1971 /* Bit 62 may or may not be set for this mce */
1972 value &= ~(1ULL << 62);
1974 if (value != 0xb600000000010015ULL)
1977 /* Clear MCi_STATUS registers */
1978 for (i = 0; i < 6; ++i)
1979 native_write_msr_safe(MSR_IA32_MCx_STATUS(i), 0, 0);
1981 value = native_read_msr_safe(MSR_IA32_MCG_STATUS, &err);
1985 value &= ~(1ULL << 2);
1986 low = lower_32_bits(value);
1987 high = upper_32_bits(value);
1989 native_write_msr_safe(MSR_IA32_MCG_STATUS, low, high);
1992 /* Flush tlb to evict multi-match entries */
1998 static void svm_handle_mce(struct kvm_vcpu *vcpu)
2000 if (is_erratum_383()) {
2002 * Erratum 383 triggered. Guest state is corrupt so kill the
2005 pr_err("KVM: Guest triggered AMD Erratum 383\n");
2007 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2013 * On an #MC intercept the MCE handler is not called automatically in
2014 * the host. So do it by hand here.
2016 kvm_machine_check();
2019 static int mc_interception(struct kvm_vcpu *vcpu)
2024 static int shutdown_interception(struct kvm_vcpu *vcpu)
2026 struct kvm_run *kvm_run = vcpu->run;
2027 struct vcpu_svm *svm = to_svm(vcpu);
2030 * The VM save area has already been encrypted so it
2031 * cannot be reinitialized - just terminate.
2033 if (sev_es_guest(vcpu->kvm))
2037 * VMCB is undefined after a SHUTDOWN intercept
2038 * so reinitialize it.
2040 clear_page(svm->vmcb);
2043 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
2047 static int io_interception(struct kvm_vcpu *vcpu)
2049 struct vcpu_svm *svm = to_svm(vcpu);
2050 u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
2051 int size, in, string;
2054 ++vcpu->stat.io_exits;
2055 string = (io_info & SVM_IOIO_STR_MASK) != 0;
2056 in = (io_info & SVM_IOIO_TYPE_MASK) != 0;
2057 port = io_info >> 16;
2058 size = (io_info & SVM_IOIO_SIZE_MASK) >> SVM_IOIO_SIZE_SHIFT;
2061 if (sev_es_guest(vcpu->kvm))
2062 return sev_es_string_io(svm, size, port, in);
2064 return kvm_emulate_instruction(vcpu, 0);
2067 svm->next_rip = svm->vmcb->control.exit_info_2;
2069 return kvm_fast_pio(vcpu, size, port, in);
2072 static int nmi_interception(struct kvm_vcpu *vcpu)
2077 static int intr_interception(struct kvm_vcpu *vcpu)
2079 ++vcpu->stat.irq_exits;
2083 static int vmload_vmsave_interception(struct kvm_vcpu *vcpu, bool vmload)
2085 struct vcpu_svm *svm = to_svm(vcpu);
2086 struct vmcb *vmcb12;
2087 struct kvm_host_map map;
2090 if (nested_svm_check_permissions(vcpu))
2093 ret = kvm_vcpu_map(vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2096 kvm_inject_gp(vcpu, 0);
2102 ret = kvm_skip_emulated_instruction(vcpu);
2105 nested_svm_vmloadsave(vmcb12, svm->vmcb);
2106 svm->sysenter_eip_hi = 0;
2107 svm->sysenter_esp_hi = 0;
2109 nested_svm_vmloadsave(svm->vmcb, vmcb12);
2111 kvm_vcpu_unmap(vcpu, &map, true);
2116 static int vmload_interception(struct kvm_vcpu *vcpu)
2118 return vmload_vmsave_interception(vcpu, true);
2121 static int vmsave_interception(struct kvm_vcpu *vcpu)
2123 return vmload_vmsave_interception(vcpu, false);
2126 static int vmrun_interception(struct kvm_vcpu *vcpu)
2128 if (nested_svm_check_permissions(vcpu))
2131 return nested_svm_vmrun(vcpu);
2141 /* Return NONE_SVM_INSTR if not SVM instrs, otherwise return decode result */
2142 static int svm_instr_opcode(struct kvm_vcpu *vcpu)
2144 struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt;
2146 if (ctxt->b != 0x1 || ctxt->opcode_len != 2)
2147 return NONE_SVM_INSTR;
2149 switch (ctxt->modrm) {
2150 case 0xd8: /* VMRUN */
2151 return SVM_INSTR_VMRUN;
2152 case 0xda: /* VMLOAD */
2153 return SVM_INSTR_VMLOAD;
2154 case 0xdb: /* VMSAVE */
2155 return SVM_INSTR_VMSAVE;
2160 return NONE_SVM_INSTR;
2163 static int emulate_svm_instr(struct kvm_vcpu *vcpu, int opcode)
2165 const int guest_mode_exit_codes[] = {
2166 [SVM_INSTR_VMRUN] = SVM_EXIT_VMRUN,
2167 [SVM_INSTR_VMLOAD] = SVM_EXIT_VMLOAD,
2168 [SVM_INSTR_VMSAVE] = SVM_EXIT_VMSAVE,
2170 int (*const svm_instr_handlers[])(struct kvm_vcpu *vcpu) = {
2171 [SVM_INSTR_VMRUN] = vmrun_interception,
2172 [SVM_INSTR_VMLOAD] = vmload_interception,
2173 [SVM_INSTR_VMSAVE] = vmsave_interception,
2175 struct vcpu_svm *svm = to_svm(vcpu);
2178 if (is_guest_mode(vcpu)) {
2179 /* Returns '1' or -errno on failure, '0' on success. */
2180 ret = nested_svm_simple_vmexit(svm, guest_mode_exit_codes[opcode]);
2185 return svm_instr_handlers[opcode](vcpu);
2189 * #GP handling code. Note that #GP can be triggered under the following two
2191 * 1) SVM VM-related instructions (VMRUN/VMSAVE/VMLOAD) that trigger #GP on
2192 * some AMD CPUs when EAX of these instructions are in the reserved memory
2193 * regions (e.g. SMM memory on host).
2194 * 2) VMware backdoor
2196 static int gp_interception(struct kvm_vcpu *vcpu)
2198 struct vcpu_svm *svm = to_svm(vcpu);
2199 u32 error_code = svm->vmcb->control.exit_info_1;
2202 /* Both #GP cases have zero error_code */
2206 /* Decode the instruction for usage later */
2207 if (x86_decode_emulated_instruction(vcpu, 0, NULL, 0) != EMULATION_OK)
2210 opcode = svm_instr_opcode(vcpu);
2212 if (opcode == NONE_SVM_INSTR) {
2213 if (!enable_vmware_backdoor)
2217 * VMware backdoor emulation on #GP interception only handles
2218 * IN{S}, OUT{S}, and RDPMC.
2220 if (!is_guest_mode(vcpu))
2221 return kvm_emulate_instruction(vcpu,
2222 EMULTYPE_VMWARE_GP | EMULTYPE_NO_DECODE);
2224 return emulate_svm_instr(vcpu, opcode);
2227 kvm_queue_exception_e(vcpu, GP_VECTOR, error_code);
2231 void svm_set_gif(struct vcpu_svm *svm, bool value)
2235 * If VGIF is enabled, the STGI intercept is only added to
2236 * detect the opening of the SMI/NMI window; remove it now.
2237 * Likewise, clear the VINTR intercept, we will set it
2238 * again while processing KVM_REQ_EVENT if needed.
2240 if (vgif_enabled(svm))
2241 svm_clr_intercept(svm, INTERCEPT_STGI);
2242 if (svm_is_intercept(svm, INTERCEPT_VINTR))
2243 svm_clear_vintr(svm);
2246 if (svm->vcpu.arch.smi_pending ||
2247 svm->vcpu.arch.nmi_pending ||
2248 kvm_cpu_has_injectable_intr(&svm->vcpu))
2249 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2254 * After a CLGI no interrupts should come. But if vGIF is
2255 * in use, we still rely on the VINTR intercept (rather than
2256 * STGI) to detect an open interrupt window.
2258 if (!vgif_enabled(svm))
2259 svm_clear_vintr(svm);
2263 static int stgi_interception(struct kvm_vcpu *vcpu)
2267 if (nested_svm_check_permissions(vcpu))
2270 ret = kvm_skip_emulated_instruction(vcpu);
2271 svm_set_gif(to_svm(vcpu), true);
2275 static int clgi_interception(struct kvm_vcpu *vcpu)
2279 if (nested_svm_check_permissions(vcpu))
2282 ret = kvm_skip_emulated_instruction(vcpu);
2283 svm_set_gif(to_svm(vcpu), false);
2287 static int invlpga_interception(struct kvm_vcpu *vcpu)
2289 gva_t gva = kvm_rax_read(vcpu);
2290 u32 asid = kvm_rcx_read(vcpu);
2292 /* FIXME: Handle an address size prefix. */
2293 if (!is_long_mode(vcpu))
2296 trace_kvm_invlpga(to_svm(vcpu)->vmcb->save.rip, asid, gva);
2298 /* Let's treat INVLPGA the same as INVLPG (can be optimized!) */
2299 kvm_mmu_invlpg(vcpu, gva);
2301 return kvm_skip_emulated_instruction(vcpu);
2304 static int skinit_interception(struct kvm_vcpu *vcpu)
2306 trace_kvm_skinit(to_svm(vcpu)->vmcb->save.rip, kvm_rax_read(vcpu));
2308 kvm_queue_exception(vcpu, UD_VECTOR);
2312 static int task_switch_interception(struct kvm_vcpu *vcpu)
2314 struct vcpu_svm *svm = to_svm(vcpu);
2317 int int_type = svm->vmcb->control.exit_int_info &
2318 SVM_EXITINTINFO_TYPE_MASK;
2319 int int_vec = svm->vmcb->control.exit_int_info & SVM_EVTINJ_VEC_MASK;
2321 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_TYPE_MASK;
2323 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID;
2324 bool has_error_code = false;
2327 tss_selector = (u16)svm->vmcb->control.exit_info_1;
2329 if (svm->vmcb->control.exit_info_2 &
2330 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
2331 reason = TASK_SWITCH_IRET;
2332 else if (svm->vmcb->control.exit_info_2 &
2333 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
2334 reason = TASK_SWITCH_JMP;
2336 reason = TASK_SWITCH_GATE;
2338 reason = TASK_SWITCH_CALL;
2340 if (reason == TASK_SWITCH_GATE) {
2342 case SVM_EXITINTINFO_TYPE_NMI:
2343 vcpu->arch.nmi_injected = false;
2345 case SVM_EXITINTINFO_TYPE_EXEPT:
2346 if (svm->vmcb->control.exit_info_2 &
2347 (1ULL << SVM_EXITINFOSHIFT_TS_HAS_ERROR_CODE)) {
2348 has_error_code = true;
2350 (u32)svm->vmcb->control.exit_info_2;
2352 kvm_clear_exception_queue(vcpu);
2354 case SVM_EXITINTINFO_TYPE_INTR:
2355 kvm_clear_interrupt_queue(vcpu);
2362 if (reason != TASK_SWITCH_GATE ||
2363 int_type == SVM_EXITINTINFO_TYPE_SOFT ||
2364 (int_type == SVM_EXITINTINFO_TYPE_EXEPT &&
2365 (int_vec == OF_VECTOR || int_vec == BP_VECTOR))) {
2366 if (!skip_emulated_instruction(vcpu))
2370 if (int_type != SVM_EXITINTINFO_TYPE_SOFT)
2373 return kvm_task_switch(vcpu, tss_selector, int_vec, reason,
2374 has_error_code, error_code);
2377 static int iret_interception(struct kvm_vcpu *vcpu)
2379 struct vcpu_svm *svm = to_svm(vcpu);
2381 ++vcpu->stat.nmi_window_exits;
2382 vcpu->arch.hflags |= HF_IRET_MASK;
2383 if (!sev_es_guest(vcpu->kvm)) {
2384 svm_clr_intercept(svm, INTERCEPT_IRET);
2385 svm->nmi_iret_rip = kvm_rip_read(vcpu);
2387 kvm_make_request(KVM_REQ_EVENT, vcpu);
2391 static int invlpg_interception(struct kvm_vcpu *vcpu)
2393 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2394 return kvm_emulate_instruction(vcpu, 0);
2396 kvm_mmu_invlpg(vcpu, to_svm(vcpu)->vmcb->control.exit_info_1);
2397 return kvm_skip_emulated_instruction(vcpu);
2400 static int emulate_on_interception(struct kvm_vcpu *vcpu)
2402 return kvm_emulate_instruction(vcpu, 0);
2405 static int rsm_interception(struct kvm_vcpu *vcpu)
2407 return kvm_emulate_instruction_from_buffer(vcpu, rsm_ins_bytes, 2);
2410 static bool check_selective_cr0_intercepted(struct kvm_vcpu *vcpu,
2413 struct vcpu_svm *svm = to_svm(vcpu);
2414 unsigned long cr0 = vcpu->arch.cr0;
2417 if (!is_guest_mode(vcpu) ||
2418 (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_SELECTIVE_CR0))))
2421 cr0 &= ~SVM_CR0_SELECTIVE_MASK;
2422 val &= ~SVM_CR0_SELECTIVE_MASK;
2425 svm->vmcb->control.exit_code = SVM_EXIT_CR0_SEL_WRITE;
2426 ret = (nested_svm_exit_handled(svm) == NESTED_EXIT_DONE);
2432 #define CR_VALID (1ULL << 63)
2434 static int cr_interception(struct kvm_vcpu *vcpu)
2436 struct vcpu_svm *svm = to_svm(vcpu);
2441 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2442 return emulate_on_interception(vcpu);
2444 if (unlikely((svm->vmcb->control.exit_info_1 & CR_VALID) == 0))
2445 return emulate_on_interception(vcpu);
2447 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2448 if (svm->vmcb->control.exit_code == SVM_EXIT_CR0_SEL_WRITE)
2449 cr = SVM_EXIT_WRITE_CR0 - SVM_EXIT_READ_CR0;
2451 cr = svm->vmcb->control.exit_code - SVM_EXIT_READ_CR0;
2454 if (cr >= 16) { /* mov to cr */
2456 val = kvm_register_read(vcpu, reg);
2457 trace_kvm_cr_write(cr, val);
2460 if (!check_selective_cr0_intercepted(vcpu, val))
2461 err = kvm_set_cr0(vcpu, val);
2467 err = kvm_set_cr3(vcpu, val);
2470 err = kvm_set_cr4(vcpu, val);
2473 err = kvm_set_cr8(vcpu, val);
2476 WARN(1, "unhandled write to CR%d", cr);
2477 kvm_queue_exception(vcpu, UD_VECTOR);
2480 } else { /* mov from cr */
2483 val = kvm_read_cr0(vcpu);
2486 val = vcpu->arch.cr2;
2489 val = kvm_read_cr3(vcpu);
2492 val = kvm_read_cr4(vcpu);
2495 val = kvm_get_cr8(vcpu);
2498 WARN(1, "unhandled read from CR%d", cr);
2499 kvm_queue_exception(vcpu, UD_VECTOR);
2502 kvm_register_write(vcpu, reg, val);
2503 trace_kvm_cr_read(cr, val);
2505 return kvm_complete_insn_gp(vcpu, err);
2508 static int cr_trap(struct kvm_vcpu *vcpu)
2510 struct vcpu_svm *svm = to_svm(vcpu);
2511 unsigned long old_value, new_value;
2515 new_value = (unsigned long)svm->vmcb->control.exit_info_1;
2517 cr = svm->vmcb->control.exit_code - SVM_EXIT_CR0_WRITE_TRAP;
2520 old_value = kvm_read_cr0(vcpu);
2521 svm_set_cr0(vcpu, new_value);
2523 kvm_post_set_cr0(vcpu, old_value, new_value);
2526 old_value = kvm_read_cr4(vcpu);
2527 svm_set_cr4(vcpu, new_value);
2529 kvm_post_set_cr4(vcpu, old_value, new_value);
2532 ret = kvm_set_cr8(vcpu, new_value);
2535 WARN(1, "unhandled CR%d write trap", cr);
2536 kvm_queue_exception(vcpu, UD_VECTOR);
2540 return kvm_complete_insn_gp(vcpu, ret);
2543 static int dr_interception(struct kvm_vcpu *vcpu)
2545 struct vcpu_svm *svm = to_svm(vcpu);
2550 if (vcpu->guest_debug == 0) {
2552 * No more DR vmexits; force a reload of the debug registers
2553 * and reenter on this instruction. The next vmexit will
2554 * retrieve the full state of the debug registers.
2556 clr_dr_intercepts(svm);
2557 vcpu->arch.switch_db_regs |= KVM_DEBUGREG_WONT_EXIT;
2561 if (!boot_cpu_has(X86_FEATURE_DECODEASSISTS))
2562 return emulate_on_interception(vcpu);
2564 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2565 dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
2566 if (dr >= 16) { /* mov to DRn */
2568 val = kvm_register_read(vcpu, reg);
2569 err = kvm_set_dr(vcpu, dr, val);
2571 kvm_get_dr(vcpu, dr, &val);
2572 kvm_register_write(vcpu, reg, val);
2575 return kvm_complete_insn_gp(vcpu, err);
2578 static int cr8_write_interception(struct kvm_vcpu *vcpu)
2582 u8 cr8_prev = kvm_get_cr8(vcpu);
2583 /* instruction emulation calls kvm_set_cr8() */
2584 r = cr_interception(vcpu);
2585 if (lapic_in_kernel(vcpu))
2587 if (cr8_prev <= kvm_get_cr8(vcpu))
2589 vcpu->run->exit_reason = KVM_EXIT_SET_TPR;
2593 static int efer_trap(struct kvm_vcpu *vcpu)
2595 struct msr_data msr_info;
2599 * Clear the EFER_SVME bit from EFER. The SVM code always sets this
2600 * bit in svm_set_efer(), but __kvm_valid_efer() checks it against
2601 * whether the guest has X86_FEATURE_SVM - this avoids a failure if
2602 * the guest doesn't have X86_FEATURE_SVM.
2604 msr_info.host_initiated = false;
2605 msr_info.index = MSR_EFER;
2606 msr_info.data = to_svm(vcpu)->vmcb->control.exit_info_1 & ~EFER_SVME;
2607 ret = kvm_set_msr_common(vcpu, &msr_info);
2609 return kvm_complete_insn_gp(vcpu, ret);
2612 static int svm_get_msr_feature(struct kvm_msr_entry *msr)
2616 switch (msr->index) {
2617 case MSR_F10H_DECFG:
2618 if (boot_cpu_has(X86_FEATURE_LFENCE_RDTSC))
2619 msr->data |= MSR_F10H_DECFG_LFENCE_SERIALIZE;
2621 case MSR_IA32_PERF_CAPABILITIES:
2624 return KVM_MSR_RET_INVALID;
2630 static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
2632 struct vcpu_svm *svm = to_svm(vcpu);
2634 switch (msr_info->index) {
2636 msr_info->data = svm->vmcb01.ptr->save.star;
2638 #ifdef CONFIG_X86_64
2640 msr_info->data = svm->vmcb01.ptr->save.lstar;
2643 msr_info->data = svm->vmcb01.ptr->save.cstar;
2645 case MSR_KERNEL_GS_BASE:
2646 msr_info->data = svm->vmcb01.ptr->save.kernel_gs_base;
2648 case MSR_SYSCALL_MASK:
2649 msr_info->data = svm->vmcb01.ptr->save.sfmask;
2652 case MSR_IA32_SYSENTER_CS:
2653 msr_info->data = svm->vmcb01.ptr->save.sysenter_cs;
2655 case MSR_IA32_SYSENTER_EIP:
2656 msr_info->data = (u32)svm->vmcb01.ptr->save.sysenter_eip;
2657 if (guest_cpuid_is_intel(vcpu))
2658 msr_info->data |= (u64)svm->sysenter_eip_hi << 32;
2660 case MSR_IA32_SYSENTER_ESP:
2661 msr_info->data = svm->vmcb01.ptr->save.sysenter_esp;
2662 if (guest_cpuid_is_intel(vcpu))
2663 msr_info->data |= (u64)svm->sysenter_esp_hi << 32;
2666 if (tsc_aux_uret_slot < 0)
2668 if (!msr_info->host_initiated &&
2669 !guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP) &&
2670 !guest_cpuid_has(vcpu, X86_FEATURE_RDPID))
2672 msr_info->data = svm->tsc_aux;
2675 * Nobody will change the following 5 values in the VMCB so we can
2676 * safely return them on rdmsr. They will always be 0 until LBRV is
2679 case MSR_IA32_DEBUGCTLMSR:
2680 msr_info->data = svm->vmcb->save.dbgctl;
2682 case MSR_IA32_LASTBRANCHFROMIP:
2683 msr_info->data = svm->vmcb->save.br_from;
2685 case MSR_IA32_LASTBRANCHTOIP:
2686 msr_info->data = svm->vmcb->save.br_to;
2688 case MSR_IA32_LASTINTFROMIP:
2689 msr_info->data = svm->vmcb->save.last_excp_from;
2691 case MSR_IA32_LASTINTTOIP:
2692 msr_info->data = svm->vmcb->save.last_excp_to;
2694 case MSR_VM_HSAVE_PA:
2695 msr_info->data = svm->nested.hsave_msr;
2698 msr_info->data = svm->nested.vm_cr_msr;
2700 case MSR_IA32_SPEC_CTRL:
2701 if (!msr_info->host_initiated &&
2702 !guest_has_spec_ctrl_msr(vcpu))
2705 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2706 msr_info->data = svm->vmcb->save.spec_ctrl;
2708 msr_info->data = svm->spec_ctrl;
2710 case MSR_AMD64_VIRT_SPEC_CTRL:
2711 if (!msr_info->host_initiated &&
2712 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2715 msr_info->data = svm->virt_spec_ctrl;
2717 case MSR_F15H_IC_CFG: {
2721 family = guest_cpuid_family(vcpu);
2722 model = guest_cpuid_model(vcpu);
2724 if (family < 0 || model < 0)
2725 return kvm_get_msr_common(vcpu, msr_info);
2729 if (family == 0x15 &&
2730 (model >= 0x2 && model < 0x20))
2731 msr_info->data = 0x1E;
2734 case MSR_F10H_DECFG:
2735 msr_info->data = svm->msr_decfg;
2738 return kvm_get_msr_common(vcpu, msr_info);
2743 static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
2745 struct vcpu_svm *svm = to_svm(vcpu);
2746 if (!err || !sev_es_guest(vcpu->kvm) || WARN_ON_ONCE(!svm->ghcb))
2747 return kvm_complete_insn_gp(vcpu, err);
2749 ghcb_set_sw_exit_info_1(svm->ghcb, 1);
2750 ghcb_set_sw_exit_info_2(svm->ghcb,
2752 SVM_EVTINJ_TYPE_EXEPT |
2757 static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
2759 struct vcpu_svm *svm = to_svm(vcpu);
2760 int svm_dis, chg_mask;
2762 if (data & ~SVM_VM_CR_VALID_MASK)
2765 chg_mask = SVM_VM_CR_VALID_MASK;
2767 if (svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK)
2768 chg_mask &= ~(SVM_VM_CR_SVM_LOCK_MASK | SVM_VM_CR_SVM_DIS_MASK);
2770 svm->nested.vm_cr_msr &= ~chg_mask;
2771 svm->nested.vm_cr_msr |= (data & chg_mask);
2773 svm_dis = svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK;
2775 /* check for svm_disable while efer.svme is set */
2776 if (svm_dis && (vcpu->arch.efer & EFER_SVME))
2782 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
2784 struct vcpu_svm *svm = to_svm(vcpu);
2787 u32 ecx = msr->index;
2788 u64 data = msr->data;
2790 case MSR_IA32_CR_PAT:
2791 if (!kvm_mtrr_valid(vcpu, MSR_IA32_CR_PAT, data))
2793 vcpu->arch.pat = data;
2794 svm->vmcb01.ptr->save.g_pat = data;
2795 if (is_guest_mode(vcpu))
2796 nested_vmcb02_compute_g_pat(svm);
2797 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
2799 case MSR_IA32_SPEC_CTRL:
2800 if (!msr->host_initiated &&
2801 !guest_has_spec_ctrl_msr(vcpu))
2804 if (kvm_spec_ctrl_test_value(data))
2807 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2808 svm->vmcb->save.spec_ctrl = data;
2810 svm->spec_ctrl = data;
2816 * When it's written (to non-zero) for the first time, pass
2820 * The handling of the MSR bitmap for L2 guests is done in
2821 * nested_svm_vmrun_msrpm.
2822 * We update the L1 MSR bit as well since it will end up
2823 * touching the MSR anyway now.
2825 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
2827 case MSR_IA32_PRED_CMD:
2828 if (!msr->host_initiated &&
2829 !guest_has_pred_cmd_msr(vcpu))
2832 if (data & ~PRED_CMD_IBPB)
2834 if (!boot_cpu_has(X86_FEATURE_IBPB))
2839 wrmsrl(MSR_IA32_PRED_CMD, PRED_CMD_IBPB);
2840 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_PRED_CMD, 0, 1);
2842 case MSR_AMD64_VIRT_SPEC_CTRL:
2843 if (!msr->host_initiated &&
2844 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2847 if (data & ~SPEC_CTRL_SSBD)
2850 svm->virt_spec_ctrl = data;
2853 svm->vmcb01.ptr->save.star = data;
2855 #ifdef CONFIG_X86_64
2857 svm->vmcb01.ptr->save.lstar = data;
2860 svm->vmcb01.ptr->save.cstar = data;
2862 case MSR_KERNEL_GS_BASE:
2863 svm->vmcb01.ptr->save.kernel_gs_base = data;
2865 case MSR_SYSCALL_MASK:
2866 svm->vmcb01.ptr->save.sfmask = data;
2869 case MSR_IA32_SYSENTER_CS:
2870 svm->vmcb01.ptr->save.sysenter_cs = data;
2872 case MSR_IA32_SYSENTER_EIP:
2873 svm->vmcb01.ptr->save.sysenter_eip = (u32)data;
2875 * We only intercept the MSR_IA32_SYSENTER_{EIP|ESP} msrs
2876 * when we spoof an Intel vendor ID (for cross vendor migration).
2877 * In this case we use this intercept to track the high
2878 * 32 bit part of these msrs to support Intel's
2879 * implementation of SYSENTER/SYSEXIT.
2881 svm->sysenter_eip_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2883 case MSR_IA32_SYSENTER_ESP:
2884 svm->vmcb01.ptr->save.sysenter_esp = (u32)data;
2885 svm->sysenter_esp_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2888 if (tsc_aux_uret_slot < 0)
2891 if (!msr->host_initiated &&
2892 !guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP) &&
2893 !guest_cpuid_has(vcpu, X86_FEATURE_RDPID))
2897 * Per Intel's SDM, bits 63:32 are reserved, but AMD's APM has
2898 * incomplete and conflicting architectural behavior. Current
2899 * AMD CPUs completely ignore bits 63:32, i.e. they aren't
2900 * reserved and always read as zeros. Emulate AMD CPU behavior
2901 * to avoid explosions if the vCPU is migrated from an AMD host
2907 * TSC_AUX is usually changed only during boot and never read
2908 * directly. Intercept TSC_AUX instead of exposing it to the
2909 * guest via direct_access_msrs, and switch it via user return.
2912 r = kvm_set_user_return_msr(tsc_aux_uret_slot, data, -1ull);
2917 svm->tsc_aux = data;
2919 case MSR_IA32_DEBUGCTLMSR:
2920 if (!boot_cpu_has(X86_FEATURE_LBRV)) {
2921 vcpu_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTL 0x%llx, nop\n",
2925 if (data & DEBUGCTL_RESERVED_BITS)
2928 svm->vmcb->save.dbgctl = data;
2929 vmcb_mark_dirty(svm->vmcb, VMCB_LBR);
2930 if (data & (1ULL<<0))
2931 svm_enable_lbrv(vcpu);
2933 svm_disable_lbrv(vcpu);
2935 case MSR_VM_HSAVE_PA:
2936 svm->nested.hsave_msr = data;
2939 return svm_set_vm_cr(vcpu, data);
2941 vcpu_unimpl(vcpu, "unimplemented wrmsr: 0x%x data 0x%llx\n", ecx, data);
2943 case MSR_F10H_DECFG: {
2944 struct kvm_msr_entry msr_entry;
2946 msr_entry.index = msr->index;
2947 if (svm_get_msr_feature(&msr_entry))
2950 /* Check the supported bits */
2951 if (data & ~msr_entry.data)
2954 /* Don't allow the guest to change a bit, #GP */
2955 if (!msr->host_initiated && (data ^ msr_entry.data))
2958 svm->msr_decfg = data;
2961 case MSR_IA32_APICBASE:
2962 if (kvm_vcpu_apicv_active(vcpu))
2963 avic_update_vapic_bar(to_svm(vcpu), data);
2966 return kvm_set_msr_common(vcpu, msr);
2971 static int msr_interception(struct kvm_vcpu *vcpu)
2973 if (to_svm(vcpu)->vmcb->control.exit_info_1)
2974 return kvm_emulate_wrmsr(vcpu);
2976 return kvm_emulate_rdmsr(vcpu);
2979 static int interrupt_window_interception(struct kvm_vcpu *vcpu)
2981 kvm_make_request(KVM_REQ_EVENT, vcpu);
2982 svm_clear_vintr(to_svm(vcpu));
2985 * For AVIC, the only reason to end up here is ExtINTs.
2986 * In this case AVIC was temporarily disabled for
2987 * requesting the IRQ window and we have to re-enable it.
2989 svm_toggle_avic_for_irq_window(vcpu, true);
2991 ++vcpu->stat.irq_window_exits;
2995 static int pause_interception(struct kvm_vcpu *vcpu)
3000 * CPL is not made available for an SEV-ES guest, therefore
3001 * vcpu->arch.preempted_in_kernel can never be true. Just
3002 * set in_kernel to false as well.
3004 in_kernel = !sev_es_guest(vcpu->kvm) && svm_get_cpl(vcpu) == 0;
3006 if (!kvm_pause_in_guest(vcpu->kvm))
3007 grow_ple_window(vcpu);
3009 kvm_vcpu_on_spin(vcpu, in_kernel);
3010 return kvm_skip_emulated_instruction(vcpu);
3013 static int invpcid_interception(struct kvm_vcpu *vcpu)
3015 struct vcpu_svm *svm = to_svm(vcpu);
3019 if (!guest_cpuid_has(vcpu, X86_FEATURE_INVPCID)) {
3020 kvm_queue_exception(vcpu, UD_VECTOR);
3025 * For an INVPCID intercept:
3026 * EXITINFO1 provides the linear address of the memory operand.
3027 * EXITINFO2 provides the contents of the register operand.
3029 type = svm->vmcb->control.exit_info_2;
3030 gva = svm->vmcb->control.exit_info_1;
3033 kvm_inject_gp(vcpu, 0);
3037 return kvm_handle_invpcid(vcpu, type, gva);
3040 static int (*const svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {
3041 [SVM_EXIT_READ_CR0] = cr_interception,
3042 [SVM_EXIT_READ_CR3] = cr_interception,
3043 [SVM_EXIT_READ_CR4] = cr_interception,
3044 [SVM_EXIT_READ_CR8] = cr_interception,
3045 [SVM_EXIT_CR0_SEL_WRITE] = cr_interception,
3046 [SVM_EXIT_WRITE_CR0] = cr_interception,
3047 [SVM_EXIT_WRITE_CR3] = cr_interception,
3048 [SVM_EXIT_WRITE_CR4] = cr_interception,
3049 [SVM_EXIT_WRITE_CR8] = cr8_write_interception,
3050 [SVM_EXIT_READ_DR0] = dr_interception,
3051 [SVM_EXIT_READ_DR1] = dr_interception,
3052 [SVM_EXIT_READ_DR2] = dr_interception,
3053 [SVM_EXIT_READ_DR3] = dr_interception,
3054 [SVM_EXIT_READ_DR4] = dr_interception,
3055 [SVM_EXIT_READ_DR5] = dr_interception,
3056 [SVM_EXIT_READ_DR6] = dr_interception,
3057 [SVM_EXIT_READ_DR7] = dr_interception,
3058 [SVM_EXIT_WRITE_DR0] = dr_interception,
3059 [SVM_EXIT_WRITE_DR1] = dr_interception,
3060 [SVM_EXIT_WRITE_DR2] = dr_interception,
3061 [SVM_EXIT_WRITE_DR3] = dr_interception,
3062 [SVM_EXIT_WRITE_DR4] = dr_interception,
3063 [SVM_EXIT_WRITE_DR5] = dr_interception,
3064 [SVM_EXIT_WRITE_DR6] = dr_interception,
3065 [SVM_EXIT_WRITE_DR7] = dr_interception,
3066 [SVM_EXIT_EXCP_BASE + DB_VECTOR] = db_interception,
3067 [SVM_EXIT_EXCP_BASE + BP_VECTOR] = bp_interception,
3068 [SVM_EXIT_EXCP_BASE + UD_VECTOR] = ud_interception,
3069 [SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
3070 [SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
3071 [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
3072 [SVM_EXIT_EXCP_BASE + GP_VECTOR] = gp_interception,
3073 [SVM_EXIT_INTR] = intr_interception,
3074 [SVM_EXIT_NMI] = nmi_interception,
3075 [SVM_EXIT_SMI] = kvm_emulate_as_nop,
3076 [SVM_EXIT_INIT] = kvm_emulate_as_nop,
3077 [SVM_EXIT_VINTR] = interrupt_window_interception,
3078 [SVM_EXIT_RDPMC] = kvm_emulate_rdpmc,
3079 [SVM_EXIT_CPUID] = kvm_emulate_cpuid,
3080 [SVM_EXIT_IRET] = iret_interception,
3081 [SVM_EXIT_INVD] = kvm_emulate_invd,
3082 [SVM_EXIT_PAUSE] = pause_interception,
3083 [SVM_EXIT_HLT] = kvm_emulate_halt,
3084 [SVM_EXIT_INVLPG] = invlpg_interception,
3085 [SVM_EXIT_INVLPGA] = invlpga_interception,
3086 [SVM_EXIT_IOIO] = io_interception,
3087 [SVM_EXIT_MSR] = msr_interception,
3088 [SVM_EXIT_TASK_SWITCH] = task_switch_interception,
3089 [SVM_EXIT_SHUTDOWN] = shutdown_interception,
3090 [SVM_EXIT_VMRUN] = vmrun_interception,
3091 [SVM_EXIT_VMMCALL] = kvm_emulate_hypercall,
3092 [SVM_EXIT_VMLOAD] = vmload_interception,
3093 [SVM_EXIT_VMSAVE] = vmsave_interception,
3094 [SVM_EXIT_STGI] = stgi_interception,
3095 [SVM_EXIT_CLGI] = clgi_interception,
3096 [SVM_EXIT_SKINIT] = skinit_interception,
3097 [SVM_EXIT_RDTSCP] = kvm_handle_invalid_op,
3098 [SVM_EXIT_WBINVD] = kvm_emulate_wbinvd,
3099 [SVM_EXIT_MONITOR] = kvm_emulate_monitor,
3100 [SVM_EXIT_MWAIT] = kvm_emulate_mwait,
3101 [SVM_EXIT_XSETBV] = kvm_emulate_xsetbv,
3102 [SVM_EXIT_RDPRU] = kvm_handle_invalid_op,
3103 [SVM_EXIT_EFER_WRITE_TRAP] = efer_trap,
3104 [SVM_EXIT_CR0_WRITE_TRAP] = cr_trap,
3105 [SVM_EXIT_CR4_WRITE_TRAP] = cr_trap,
3106 [SVM_EXIT_CR8_WRITE_TRAP] = cr_trap,
3107 [SVM_EXIT_INVPCID] = invpcid_interception,
3108 [SVM_EXIT_NPF] = npf_interception,
3109 [SVM_EXIT_RSM] = rsm_interception,
3110 [SVM_EXIT_AVIC_INCOMPLETE_IPI] = avic_incomplete_ipi_interception,
3111 [SVM_EXIT_AVIC_UNACCELERATED_ACCESS] = avic_unaccelerated_access_interception,
3112 [SVM_EXIT_VMGEXIT] = sev_handle_vmgexit,
3115 static void dump_vmcb(struct kvm_vcpu *vcpu)
3117 struct vcpu_svm *svm = to_svm(vcpu);
3118 struct vmcb_control_area *control = &svm->vmcb->control;
3119 struct vmcb_save_area *save = &svm->vmcb->save;
3120 struct vmcb_save_area *save01 = &svm->vmcb01.ptr->save;
3122 if (!dump_invalid_vmcb) {
3123 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3127 pr_err("VMCB Control Area:\n");
3128 pr_err("%-20s%04x\n", "cr_read:", control->intercepts[INTERCEPT_CR] & 0xffff);
3129 pr_err("%-20s%04x\n", "cr_write:", control->intercepts[INTERCEPT_CR] >> 16);
3130 pr_err("%-20s%04x\n", "dr_read:", control->intercepts[INTERCEPT_DR] & 0xffff);
3131 pr_err("%-20s%04x\n", "dr_write:", control->intercepts[INTERCEPT_DR] >> 16);
3132 pr_err("%-20s%08x\n", "exceptions:", control->intercepts[INTERCEPT_EXCEPTION]);
3133 pr_err("%-20s%08x %08x\n", "intercepts:",
3134 control->intercepts[INTERCEPT_WORD3],
3135 control->intercepts[INTERCEPT_WORD4]);
3136 pr_err("%-20s%d\n", "pause filter count:", control->pause_filter_count);
3137 pr_err("%-20s%d\n", "pause filter threshold:",
3138 control->pause_filter_thresh);
3139 pr_err("%-20s%016llx\n", "iopm_base_pa:", control->iopm_base_pa);
3140 pr_err("%-20s%016llx\n", "msrpm_base_pa:", control->msrpm_base_pa);
3141 pr_err("%-20s%016llx\n", "tsc_offset:", control->tsc_offset);
3142 pr_err("%-20s%d\n", "asid:", control->asid);
3143 pr_err("%-20s%d\n", "tlb_ctl:", control->tlb_ctl);
3144 pr_err("%-20s%08x\n", "int_ctl:", control->int_ctl);
3145 pr_err("%-20s%08x\n", "int_vector:", control->int_vector);
3146 pr_err("%-20s%08x\n", "int_state:", control->int_state);
3147 pr_err("%-20s%08x\n", "exit_code:", control->exit_code);
3148 pr_err("%-20s%016llx\n", "exit_info1:", control->exit_info_1);
3149 pr_err("%-20s%016llx\n", "exit_info2:", control->exit_info_2);
3150 pr_err("%-20s%08x\n", "exit_int_info:", control->exit_int_info);
3151 pr_err("%-20s%08x\n", "exit_int_info_err:", control->exit_int_info_err);
3152 pr_err("%-20s%lld\n", "nested_ctl:", control->nested_ctl);
3153 pr_err("%-20s%016llx\n", "nested_cr3:", control->nested_cr3);
3154 pr_err("%-20s%016llx\n", "avic_vapic_bar:", control->avic_vapic_bar);
3155 pr_err("%-20s%016llx\n", "ghcb:", control->ghcb_gpa);
3156 pr_err("%-20s%08x\n", "event_inj:", control->event_inj);
3157 pr_err("%-20s%08x\n", "event_inj_err:", control->event_inj_err);
3158 pr_err("%-20s%lld\n", "virt_ext:", control->virt_ext);
3159 pr_err("%-20s%016llx\n", "next_rip:", control->next_rip);
3160 pr_err("%-20s%016llx\n", "avic_backing_page:", control->avic_backing_page);
3161 pr_err("%-20s%016llx\n", "avic_logical_id:", control->avic_logical_id);
3162 pr_err("%-20s%016llx\n", "avic_physical_id:", control->avic_physical_id);
3163 pr_err("%-20s%016llx\n", "vmsa_pa:", control->vmsa_pa);
3164 pr_err("VMCB State Save Area:\n");
3165 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3167 save->es.selector, save->es.attrib,
3168 save->es.limit, save->es.base);
3169 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3171 save->cs.selector, save->cs.attrib,
3172 save->cs.limit, save->cs.base);
3173 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3175 save->ss.selector, save->ss.attrib,
3176 save->ss.limit, save->ss.base);
3177 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3179 save->ds.selector, save->ds.attrib,
3180 save->ds.limit, save->ds.base);
3181 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3183 save01->fs.selector, save01->fs.attrib,
3184 save01->fs.limit, save01->fs.base);
3185 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3187 save01->gs.selector, save01->gs.attrib,
3188 save01->gs.limit, save01->gs.base);
3189 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3191 save->gdtr.selector, save->gdtr.attrib,
3192 save->gdtr.limit, save->gdtr.base);
3193 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3195 save01->ldtr.selector, save01->ldtr.attrib,
3196 save01->ldtr.limit, save01->ldtr.base);
3197 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3199 save->idtr.selector, save->idtr.attrib,
3200 save->idtr.limit, save->idtr.base);
3201 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3203 save01->tr.selector, save01->tr.attrib,
3204 save01->tr.limit, save01->tr.base);
3205 pr_err("cpl: %d efer: %016llx\n",
3206 save->cpl, save->efer);
3207 pr_err("%-15s %016llx %-13s %016llx\n",
3208 "cr0:", save->cr0, "cr2:", save->cr2);
3209 pr_err("%-15s %016llx %-13s %016llx\n",
3210 "cr3:", save->cr3, "cr4:", save->cr4);
3211 pr_err("%-15s %016llx %-13s %016llx\n",
3212 "dr6:", save->dr6, "dr7:", save->dr7);
3213 pr_err("%-15s %016llx %-13s %016llx\n",
3214 "rip:", save->rip, "rflags:", save->rflags);
3215 pr_err("%-15s %016llx %-13s %016llx\n",
3216 "rsp:", save->rsp, "rax:", save->rax);
3217 pr_err("%-15s %016llx %-13s %016llx\n",
3218 "star:", save01->star, "lstar:", save01->lstar);
3219 pr_err("%-15s %016llx %-13s %016llx\n",
3220 "cstar:", save01->cstar, "sfmask:", save01->sfmask);
3221 pr_err("%-15s %016llx %-13s %016llx\n",
3222 "kernel_gs_base:", save01->kernel_gs_base,
3223 "sysenter_cs:", save01->sysenter_cs);
3224 pr_err("%-15s %016llx %-13s %016llx\n",
3225 "sysenter_esp:", save01->sysenter_esp,
3226 "sysenter_eip:", save01->sysenter_eip);
3227 pr_err("%-15s %016llx %-13s %016llx\n",
3228 "gpat:", save->g_pat, "dbgctl:", save->dbgctl);
3229 pr_err("%-15s %016llx %-13s %016llx\n",
3230 "br_from:", save->br_from, "br_to:", save->br_to);
3231 pr_err("%-15s %016llx %-13s %016llx\n",
3232 "excp_from:", save->last_excp_from,
3233 "excp_to:", save->last_excp_to);
3236 static int svm_handle_invalid_exit(struct kvm_vcpu *vcpu, u64 exit_code)
3238 if (exit_code < ARRAY_SIZE(svm_exit_handlers) &&
3239 svm_exit_handlers[exit_code])
3242 vcpu_unimpl(vcpu, "svm: unexpected exit reason 0x%llx\n", exit_code);
3244 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3245 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
3246 vcpu->run->internal.ndata = 2;
3247 vcpu->run->internal.data[0] = exit_code;
3248 vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
3253 int svm_invoke_exit_handler(struct kvm_vcpu *vcpu, u64 exit_code)
3255 if (svm_handle_invalid_exit(vcpu, exit_code))
3258 #ifdef CONFIG_RETPOLINE
3259 if (exit_code == SVM_EXIT_MSR)
3260 return msr_interception(vcpu);
3261 else if (exit_code == SVM_EXIT_VINTR)
3262 return interrupt_window_interception(vcpu);
3263 else if (exit_code == SVM_EXIT_INTR)
3264 return intr_interception(vcpu);
3265 else if (exit_code == SVM_EXIT_HLT)
3266 return kvm_emulate_halt(vcpu);
3267 else if (exit_code == SVM_EXIT_NPF)
3268 return npf_interception(vcpu);
3270 return svm_exit_handlers[exit_code](vcpu);
3273 static void svm_get_exit_info(struct kvm_vcpu *vcpu, u64 *info1, u64 *info2,
3274 u32 *intr_info, u32 *error_code)
3276 struct vmcb_control_area *control = &to_svm(vcpu)->vmcb->control;
3278 *info1 = control->exit_info_1;
3279 *info2 = control->exit_info_2;
3280 *intr_info = control->exit_int_info;
3281 if ((*intr_info & SVM_EXITINTINFO_VALID) &&
3282 (*intr_info & SVM_EXITINTINFO_VALID_ERR))
3283 *error_code = control->exit_int_info_err;
3288 static int handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
3290 struct vcpu_svm *svm = to_svm(vcpu);
3291 struct kvm_run *kvm_run = vcpu->run;
3292 u32 exit_code = svm->vmcb->control.exit_code;
3294 trace_kvm_exit(exit_code, vcpu, KVM_ISA_SVM);
3296 /* SEV-ES guests must use the CR write traps to track CR registers. */
3297 if (!sev_es_guest(vcpu->kvm)) {
3298 if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
3299 vcpu->arch.cr0 = svm->vmcb->save.cr0;
3301 vcpu->arch.cr3 = svm->vmcb->save.cr3;
3304 if (is_guest_mode(vcpu)) {
3307 trace_kvm_nested_vmexit(exit_code, vcpu, KVM_ISA_SVM);
3309 vmexit = nested_svm_exit_special(svm);
3311 if (vmexit == NESTED_EXIT_CONTINUE)
3312 vmexit = nested_svm_exit_handled(svm);
3314 if (vmexit == NESTED_EXIT_DONE)
3318 if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
3319 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3320 kvm_run->fail_entry.hardware_entry_failure_reason
3321 = svm->vmcb->control.exit_code;
3322 kvm_run->fail_entry.cpu = vcpu->arch.last_vmentry_cpu;
3327 if (is_external_interrupt(svm->vmcb->control.exit_int_info) &&
3328 exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR &&
3329 exit_code != SVM_EXIT_NPF && exit_code != SVM_EXIT_TASK_SWITCH &&
3330 exit_code != SVM_EXIT_INTR && exit_code != SVM_EXIT_NMI)
3331 printk(KERN_ERR "%s: unexpected exit_int_info 0x%x "
3333 __func__, svm->vmcb->control.exit_int_info,
3336 if (exit_fastpath != EXIT_FASTPATH_NONE)
3339 return svm_invoke_exit_handler(vcpu, exit_code);
3342 static void reload_tss(struct kvm_vcpu *vcpu)
3344 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3346 sd->tss_desc->type = 9; /* available 32/64-bit TSS */
3350 static void pre_svm_run(struct kvm_vcpu *vcpu)
3352 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3353 struct vcpu_svm *svm = to_svm(vcpu);
3356 * If the previous vmrun of the vmcb occurred on a different physical
3357 * cpu, then mark the vmcb dirty and assign a new asid. Hardware's
3358 * vmcb clean bits are per logical CPU, as are KVM's asid assignments.
3360 if (unlikely(svm->current_vmcb->cpu != vcpu->cpu)) {
3361 svm->current_vmcb->asid_generation = 0;
3362 vmcb_mark_all_dirty(svm->vmcb);
3363 svm->current_vmcb->cpu = vcpu->cpu;
3366 if (sev_guest(vcpu->kvm))
3367 return pre_sev_run(svm, vcpu->cpu);
3369 /* FIXME: handle wraparound of asid_generation */
3370 if (svm->current_vmcb->asid_generation != sd->asid_generation)
3374 static void svm_inject_nmi(struct kvm_vcpu *vcpu)
3376 struct vcpu_svm *svm = to_svm(vcpu);
3378 svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
3379 vcpu->arch.hflags |= HF_NMI_MASK;
3380 if (!sev_es_guest(vcpu->kvm))
3381 svm_set_intercept(svm, INTERCEPT_IRET);
3382 ++vcpu->stat.nmi_injections;
3385 static void svm_set_irq(struct kvm_vcpu *vcpu)
3387 struct vcpu_svm *svm = to_svm(vcpu);
3389 BUG_ON(!(gif_set(svm)));
3391 trace_kvm_inj_virq(vcpu->arch.interrupt.nr);
3392 ++vcpu->stat.irq_injections;
3394 svm->vmcb->control.event_inj = vcpu->arch.interrupt.nr |
3395 SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR;
3398 static void svm_update_cr8_intercept(struct kvm_vcpu *vcpu, int tpr, int irr)
3400 struct vcpu_svm *svm = to_svm(vcpu);
3403 * SEV-ES guests must always keep the CR intercepts cleared. CR
3404 * tracking is done using the CR write traps.
3406 if (sev_es_guest(vcpu->kvm))
3409 if (nested_svm_virtualize_tpr(vcpu))
3412 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3418 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
3421 bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
3423 struct vcpu_svm *svm = to_svm(vcpu);
3424 struct vmcb *vmcb = svm->vmcb;
3430 if (is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3433 ret = (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK) ||
3434 (vcpu->arch.hflags & HF_NMI_MASK);
3439 static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3441 struct vcpu_svm *svm = to_svm(vcpu);
3442 if (svm->nested.nested_run_pending)
3445 /* An NMI must not be injected into L2 if it's supposed to VM-Exit. */
3446 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3449 return !svm_nmi_blocked(vcpu);
3452 static bool svm_get_nmi_mask(struct kvm_vcpu *vcpu)
3454 return !!(vcpu->arch.hflags & HF_NMI_MASK);
3457 static void svm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
3459 struct vcpu_svm *svm = to_svm(vcpu);
3462 vcpu->arch.hflags |= HF_NMI_MASK;
3463 if (!sev_es_guest(vcpu->kvm))
3464 svm_set_intercept(svm, INTERCEPT_IRET);
3466 vcpu->arch.hflags &= ~HF_NMI_MASK;
3467 if (!sev_es_guest(vcpu->kvm))
3468 svm_clr_intercept(svm, INTERCEPT_IRET);
3472 bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
3474 struct vcpu_svm *svm = to_svm(vcpu);
3475 struct vmcb *vmcb = svm->vmcb;
3480 if (sev_es_guest(vcpu->kvm)) {
3482 * SEV-ES guests to not expose RFLAGS. Use the VMCB interrupt mask
3483 * bit to determine the state of the IF flag.
3485 if (!(vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK))
3487 } else if (is_guest_mode(vcpu)) {
3488 /* As long as interrupts are being delivered... */
3489 if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
3490 ? !(svm->vmcb01.ptr->save.rflags & X86_EFLAGS_IF)
3491 : !(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3494 /* ... vmexits aren't blocked by the interrupt shadow */
3495 if (nested_exit_on_intr(svm))
3498 if (!(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3502 return (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK);
3505 static int svm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3507 struct vcpu_svm *svm = to_svm(vcpu);
3508 if (svm->nested.nested_run_pending)
3512 * An IRQ must not be injected into L2 if it's supposed to VM-Exit,
3513 * e.g. if the IRQ arrived asynchronously after checking nested events.
3515 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_intr(svm))
3518 return !svm_interrupt_blocked(vcpu);
3521 static void svm_enable_irq_window(struct kvm_vcpu *vcpu)
3523 struct vcpu_svm *svm = to_svm(vcpu);
3526 * In case GIF=0 we can't rely on the CPU to tell us when GIF becomes
3527 * 1, because that's a separate STGI/VMRUN intercept. The next time we
3528 * get that intercept, this function will be called again though and
3529 * we'll get the vintr intercept. However, if the vGIF feature is
3530 * enabled, the STGI interception will not occur. Enable the irq
3531 * window under the assumption that the hardware will set the GIF.
3533 if (vgif_enabled(svm) || gif_set(svm)) {
3535 * IRQ window is not needed when AVIC is enabled,
3536 * unless we have pending ExtINT since it cannot be injected
3537 * via AVIC. In such case, we need to temporarily disable AVIC,
3538 * and fallback to injecting IRQ via V_IRQ.
3540 svm_toggle_avic_for_irq_window(vcpu, false);
3545 static void svm_enable_nmi_window(struct kvm_vcpu *vcpu)
3547 struct vcpu_svm *svm = to_svm(vcpu);
3549 if ((vcpu->arch.hflags & (HF_NMI_MASK | HF_IRET_MASK)) == HF_NMI_MASK)
3550 return; /* IRET will cause a vm exit */
3552 if (!gif_set(svm)) {
3553 if (vgif_enabled(svm))
3554 svm_set_intercept(svm, INTERCEPT_STGI);
3555 return; /* STGI will cause a vm exit */
3559 * Something prevents NMI from been injected. Single step over possible
3560 * problem (IRET or exception injection or interrupt shadow)
3562 svm->nmi_singlestep_guest_rflags = svm_get_rflags(vcpu);
3563 svm->nmi_singlestep = true;
3564 svm->vmcb->save.rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
3567 static int svm_set_tss_addr(struct kvm *kvm, unsigned int addr)
3572 static int svm_set_identity_map_addr(struct kvm *kvm, u64 ident_addr)
3577 void svm_flush_tlb(struct kvm_vcpu *vcpu)
3579 struct vcpu_svm *svm = to_svm(vcpu);
3582 * Flush only the current ASID even if the TLB flush was invoked via
3583 * kvm_flush_remote_tlbs(). Although flushing remote TLBs requires all
3584 * ASIDs to be flushed, KVM uses a single ASID for L1 and L2, and
3585 * unconditionally does a TLB flush on both nested VM-Enter and nested
3586 * VM-Exit (via kvm_mmu_reset_context()).
3588 if (static_cpu_has(X86_FEATURE_FLUSHBYASID))
3589 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3591 svm->current_vmcb->asid_generation--;
3594 static void svm_flush_tlb_gva(struct kvm_vcpu *vcpu, gva_t gva)
3596 struct vcpu_svm *svm = to_svm(vcpu);
3598 invlpga(gva, svm->vmcb->control.asid);
3601 static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)
3603 struct vcpu_svm *svm = to_svm(vcpu);
3605 if (nested_svm_virtualize_tpr(vcpu))
3608 if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
3609 int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
3610 kvm_set_cr8(vcpu, cr8);
3614 static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
3616 struct vcpu_svm *svm = to_svm(vcpu);
3619 if (nested_svm_virtualize_tpr(vcpu) ||
3620 kvm_vcpu_apicv_active(vcpu))
3623 cr8 = kvm_get_cr8(vcpu);
3624 svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
3625 svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
3628 static void svm_complete_interrupts(struct kvm_vcpu *vcpu)
3630 struct vcpu_svm *svm = to_svm(vcpu);
3633 u32 exitintinfo = svm->vmcb->control.exit_int_info;
3634 unsigned int3_injected = svm->int3_injected;
3636 svm->int3_injected = 0;
3639 * If we've made progress since setting HF_IRET_MASK, we've
3640 * executed an IRET and can allow NMI injection.
3642 if ((vcpu->arch.hflags & HF_IRET_MASK) &&
3643 (sev_es_guest(vcpu->kvm) ||
3644 kvm_rip_read(vcpu) != svm->nmi_iret_rip)) {
3645 vcpu->arch.hflags &= ~(HF_NMI_MASK | HF_IRET_MASK);
3646 kvm_make_request(KVM_REQ_EVENT, vcpu);
3649 vcpu->arch.nmi_injected = false;
3650 kvm_clear_exception_queue(vcpu);
3651 kvm_clear_interrupt_queue(vcpu);
3653 if (!(exitintinfo & SVM_EXITINTINFO_VALID))
3656 kvm_make_request(KVM_REQ_EVENT, vcpu);
3658 vector = exitintinfo & SVM_EXITINTINFO_VEC_MASK;
3659 type = exitintinfo & SVM_EXITINTINFO_TYPE_MASK;
3662 case SVM_EXITINTINFO_TYPE_NMI:
3663 vcpu->arch.nmi_injected = true;
3665 case SVM_EXITINTINFO_TYPE_EXEPT:
3667 * Never re-inject a #VC exception.
3669 if (vector == X86_TRAP_VC)
3673 * In case of software exceptions, do not reinject the vector,
3674 * but re-execute the instruction instead. Rewind RIP first
3675 * if we emulated INT3 before.
3677 if (kvm_exception_is_soft(vector)) {
3678 if (vector == BP_VECTOR && int3_injected &&
3679 kvm_is_linear_rip(vcpu, svm->int3_rip))
3681 kvm_rip_read(vcpu) - int3_injected);
3684 if (exitintinfo & SVM_EXITINTINFO_VALID_ERR) {
3685 u32 err = svm->vmcb->control.exit_int_info_err;
3686 kvm_requeue_exception_e(vcpu, vector, err);
3689 kvm_requeue_exception(vcpu, vector);
3691 case SVM_EXITINTINFO_TYPE_INTR:
3692 kvm_queue_interrupt(vcpu, vector, false);
3699 static void svm_cancel_injection(struct kvm_vcpu *vcpu)
3701 struct vcpu_svm *svm = to_svm(vcpu);
3702 struct vmcb_control_area *control = &svm->vmcb->control;
3704 control->exit_int_info = control->event_inj;
3705 control->exit_int_info_err = control->event_inj_err;
3706 control->event_inj = 0;
3707 svm_complete_interrupts(vcpu);
3710 static fastpath_t svm_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
3712 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_MSR &&
3713 to_svm(vcpu)->vmcb->control.exit_info_1)
3714 return handle_fastpath_set_msr_irqoff(vcpu);
3716 return EXIT_FASTPATH_NONE;
3719 static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
3721 struct vcpu_svm *svm = to_svm(vcpu);
3722 unsigned long vmcb_pa = svm->current_vmcb->pa;
3725 * VMENTER enables interrupts (host state), but the kernel state is
3726 * interrupts disabled when this is invoked. Also tell RCU about
3727 * it. This is the same logic as for exit_to_user_mode().
3729 * This ensures that e.g. latency analysis on the host observes
3730 * guest mode as interrupt enabled.
3732 * guest_enter_irqoff() informs context tracking about the
3733 * transition to guest mode and if enabled adjusts RCU state
3736 instrumentation_begin();
3737 trace_hardirqs_on_prepare();
3738 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
3739 instrumentation_end();
3741 guest_enter_irqoff();
3742 lockdep_hardirqs_on(CALLER_ADDR0);
3744 if (sev_es_guest(vcpu->kvm)) {
3745 __svm_sev_es_vcpu_run(vmcb_pa);
3747 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3750 * Use a single vmcb (vmcb01 because it's always valid) for
3751 * context switching guest state via VMLOAD/VMSAVE, that way
3752 * the state doesn't need to be copied between vmcb01 and
3753 * vmcb02 when switching vmcbs for nested virtualization.
3755 vmload(svm->vmcb01.pa);
3756 __svm_vcpu_run(vmcb_pa, (unsigned long *)&vcpu->arch.regs);
3757 vmsave(svm->vmcb01.pa);
3759 vmload(__sme_page_pa(sd->save_area));
3763 * VMEXIT disables interrupts (host state), but tracing and lockdep
3764 * have them in state 'on' as recorded before entering guest mode.
3765 * Same as enter_from_user_mode().
3767 * guest_exit_irqoff() restores host context and reinstates RCU if
3768 * enabled and required.
3770 * This needs to be done before the below as native_read_msr()
3771 * contains a tracepoint and x86_spec_ctrl_restore_host() calls
3772 * into world and some more.
3774 lockdep_hardirqs_off(CALLER_ADDR0);
3775 guest_exit_irqoff();
3777 instrumentation_begin();
3778 trace_hardirqs_off_finish();
3779 instrumentation_end();
3782 static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
3784 struct vcpu_svm *svm = to_svm(vcpu);
3786 trace_kvm_entry(vcpu);
3788 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
3789 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
3790 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
3793 * Disable singlestep if we're injecting an interrupt/exception.
3794 * We don't want our modified rflags to be pushed on the stack where
3795 * we might not be able to easily reset them if we disabled NMI
3798 if (svm->nmi_singlestep && svm->vmcb->control.event_inj) {
3800 * Event injection happens before external interrupts cause a
3801 * vmexit and interrupts are disabled here, so smp_send_reschedule
3802 * is enough to force an immediate vmexit.
3804 disable_nmi_singlestep(svm);
3805 smp_send_reschedule(vcpu->cpu);
3810 sync_lapic_to_cr8(vcpu);
3812 if (unlikely(svm->asid != svm->vmcb->control.asid)) {
3813 svm->vmcb->control.asid = svm->asid;
3814 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
3816 svm->vmcb->save.cr2 = vcpu->arch.cr2;
3819 * Run with all-zero DR6 unless needed, so that we can get the exact cause
3822 if (unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT))
3823 svm_set_dr6(svm, vcpu->arch.dr6);
3825 svm_set_dr6(svm, DR6_ACTIVE_LOW);
3828 kvm_load_guest_xsave_state(vcpu);
3830 kvm_wait_lapic_expire(vcpu);
3833 * If this vCPU has touched SPEC_CTRL, restore the guest's value if
3834 * it's non-zero. Since vmentry is serialising on affected CPUs, there
3835 * is no need to worry about the conditional branch over the wrmsr
3836 * being speculatively taken.
3838 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3839 x86_spec_ctrl_set_guest(svm->spec_ctrl, svm->virt_spec_ctrl);
3841 svm_vcpu_enter_exit(vcpu);
3844 * We do not use IBRS in the kernel. If this vCPU has used the
3845 * SPEC_CTRL MSR it may have left it on; save the value and
3846 * turn it off. This is much more efficient than blindly adding
3847 * it to the atomic save/restore list. Especially as the former
3848 * (Saving guest MSRs on vmexit) doesn't even exist in KVM.
3850 * For non-nested case:
3851 * If the L01 MSR bitmap does not intercept the MSR, then we need to
3855 * If the L02 MSR bitmap does not intercept the MSR, then we need to
3858 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL) &&
3859 unlikely(!msr_write_intercepted(vcpu, MSR_IA32_SPEC_CTRL)))
3860 svm->spec_ctrl = native_read_msr(MSR_IA32_SPEC_CTRL);
3862 if (!sev_es_guest(vcpu->kvm))
3865 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3866 x86_spec_ctrl_restore_host(svm->spec_ctrl, svm->virt_spec_ctrl);
3868 if (!sev_es_guest(vcpu->kvm)) {
3869 vcpu->arch.cr2 = svm->vmcb->save.cr2;
3870 vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
3871 vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
3872 vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
3875 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3876 kvm_before_interrupt(vcpu);
3878 kvm_load_host_xsave_state(vcpu);
3881 /* Any pending NMI will happen here */
3883 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3884 kvm_after_interrupt(vcpu);
3886 sync_cr8_to_lapic(vcpu);
3889 if (is_guest_mode(vcpu)) {
3890 nested_sync_control_from_vmcb02(svm);
3891 svm->nested.nested_run_pending = 0;
3894 svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
3895 vmcb_mark_all_clean(svm->vmcb);
3897 /* if exit due to PF check for async PF */
3898 if (svm->vmcb->control.exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR)
3899 vcpu->arch.apf.host_apf_flags =
3900 kvm_read_and_reset_apf_flags();
3903 vcpu->arch.regs_avail &= ~(1 << VCPU_EXREG_PDPTR);
3904 vcpu->arch.regs_dirty &= ~(1 << VCPU_EXREG_PDPTR);
3908 * We need to handle MC intercepts here before the vcpu has a chance to
3909 * change the physical cpu
3911 if (unlikely(svm->vmcb->control.exit_code ==
3912 SVM_EXIT_EXCP_BASE + MC_VECTOR))
3913 svm_handle_mce(vcpu);
3915 svm_complete_interrupts(vcpu);
3917 if (is_guest_mode(vcpu))
3918 return EXIT_FASTPATH_NONE;
3920 return svm_exit_handlers_fastpath(vcpu);
3923 static void svm_load_mmu_pgd(struct kvm_vcpu *vcpu, hpa_t root_hpa,
3926 struct vcpu_svm *svm = to_svm(vcpu);
3930 svm->vmcb->control.nested_cr3 = __sme_set(root_hpa);
3931 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
3933 /* Loading L2's CR3 is handled by enter_svm_guest_mode. */
3934 if (!test_bit(VCPU_EXREG_CR3, (ulong *)&vcpu->arch.regs_avail))
3936 cr3 = vcpu->arch.cr3;
3937 } else if (vcpu->arch.mmu->shadow_root_level >= PT64_ROOT_4LEVEL) {
3938 cr3 = __sme_set(root_hpa) | kvm_get_active_pcid(vcpu);
3940 /* PCID in the guest should be impossible with a 32-bit MMU. */
3941 WARN_ON_ONCE(kvm_get_active_pcid(vcpu));
3945 svm->vmcb->save.cr3 = cr3;
3946 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
3949 static int is_disabled(void)
3953 rdmsrl(MSR_VM_CR, vm_cr);
3954 if (vm_cr & (1 << SVM_VM_CR_SVM_DISABLE))
3961 svm_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
3964 * Patch in the VMMCALL instruction:
3966 hypercall[0] = 0x0f;
3967 hypercall[1] = 0x01;
3968 hypercall[2] = 0xd9;
3971 static int __init svm_check_processor_compat(void)
3976 static bool svm_cpu_has_accelerated_tpr(void)
3982 * The kvm parameter can be NULL (module initialization, or invocation before
3983 * VM creation). Be sure to check the kvm parameter before using it.
3985 static bool svm_has_emulated_msr(struct kvm *kvm, u32 index)
3988 case MSR_IA32_MCG_EXT_CTL:
3989 case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:
3991 case MSR_IA32_SMBASE:
3992 /* SEV-ES guests do not support SMM, so report false */
3993 if (kvm && sev_es_guest(kvm))
4003 static u64 svm_get_mt_mask(struct kvm_vcpu *vcpu, gfn_t gfn, bool is_mmio)
4008 static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
4010 struct vcpu_svm *svm = to_svm(vcpu);
4011 struct kvm_cpuid_entry2 *best;
4013 vcpu->arch.xsaves_enabled = guest_cpuid_has(vcpu, X86_FEATURE_XSAVE) &&
4014 boot_cpu_has(X86_FEATURE_XSAVE) &&
4015 boot_cpu_has(X86_FEATURE_XSAVES);
4017 /* Update nrips enabled cache */
4018 svm->nrips_enabled = kvm_cpu_cap_has(X86_FEATURE_NRIPS) &&
4019 guest_cpuid_has(vcpu, X86_FEATURE_NRIPS);
4021 svm_recalc_instruction_intercepts(vcpu, svm);
4023 /* For sev guests, the memory encryption bit is not reserved in CR3. */
4024 if (sev_guest(vcpu->kvm)) {
4025 best = kvm_find_cpuid_entry(vcpu, 0x8000001F, 0);
4027 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
4030 if (kvm_vcpu_apicv_active(vcpu)) {
4032 * AVIC does not work with an x2APIC mode guest. If the X2APIC feature
4033 * is exposed to the guest, disable AVIC.
4035 if (guest_cpuid_has(vcpu, X86_FEATURE_X2APIC))
4036 kvm_request_apicv_update(vcpu->kvm, false,
4037 APICV_INHIBIT_REASON_X2APIC);
4040 * Currently, AVIC does not work with nested virtualization.
4041 * So, we disable AVIC when cpuid for SVM is set in the L1 guest.
4043 if (nested && guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4044 kvm_request_apicv_update(vcpu->kvm, false,
4045 APICV_INHIBIT_REASON_NESTED);
4048 if (guest_cpuid_is_intel(vcpu)) {
4050 * We must intercept SYSENTER_EIP and SYSENTER_ESP
4051 * accesses because the processor only stores 32 bits.
4052 * For the same reason we cannot use virtual VMLOAD/VMSAVE.
4054 svm_set_intercept(svm, INTERCEPT_VMLOAD);
4055 svm_set_intercept(svm, INTERCEPT_VMSAVE);
4056 svm->vmcb->control.virt_ext &= ~VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
4058 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 0, 0);
4059 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 0, 0);
4062 * If hardware supports Virtual VMLOAD VMSAVE then enable it
4063 * in VMCB and clear intercepts to avoid #VMEXIT.
4066 svm_clr_intercept(svm, INTERCEPT_VMLOAD);
4067 svm_clr_intercept(svm, INTERCEPT_VMSAVE);
4068 svm->vmcb->control.virt_ext |= VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
4070 /* No need to intercept these MSRs */
4071 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 1, 1);
4072 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 1, 1);
4076 static bool svm_has_wbinvd_exit(void)
4081 #define PRE_EX(exit) { .exit_code = (exit), \
4082 .stage = X86_ICPT_PRE_EXCEPT, }
4083 #define POST_EX(exit) { .exit_code = (exit), \
4084 .stage = X86_ICPT_POST_EXCEPT, }
4085 #define POST_MEM(exit) { .exit_code = (exit), \
4086 .stage = X86_ICPT_POST_MEMACCESS, }
4088 static const struct __x86_intercept {
4090 enum x86_intercept_stage stage;
4091 } x86_intercept_map[] = {
4092 [x86_intercept_cr_read] = POST_EX(SVM_EXIT_READ_CR0),
4093 [x86_intercept_cr_write] = POST_EX(SVM_EXIT_WRITE_CR0),
4094 [x86_intercept_clts] = POST_EX(SVM_EXIT_WRITE_CR0),
4095 [x86_intercept_lmsw] = POST_EX(SVM_EXIT_WRITE_CR0),
4096 [x86_intercept_smsw] = POST_EX(SVM_EXIT_READ_CR0),
4097 [x86_intercept_dr_read] = POST_EX(SVM_EXIT_READ_DR0),
4098 [x86_intercept_dr_write] = POST_EX(SVM_EXIT_WRITE_DR0),
4099 [x86_intercept_sldt] = POST_EX(SVM_EXIT_LDTR_READ),
4100 [x86_intercept_str] = POST_EX(SVM_EXIT_TR_READ),
4101 [x86_intercept_lldt] = POST_EX(SVM_EXIT_LDTR_WRITE),
4102 [x86_intercept_ltr] = POST_EX(SVM_EXIT_TR_WRITE),
4103 [x86_intercept_sgdt] = POST_EX(SVM_EXIT_GDTR_READ),
4104 [x86_intercept_sidt] = POST_EX(SVM_EXIT_IDTR_READ),
4105 [x86_intercept_lgdt] = POST_EX(SVM_EXIT_GDTR_WRITE),
4106 [x86_intercept_lidt] = POST_EX(SVM_EXIT_IDTR_WRITE),
4107 [x86_intercept_vmrun] = POST_EX(SVM_EXIT_VMRUN),
4108 [x86_intercept_vmmcall] = POST_EX(SVM_EXIT_VMMCALL),
4109 [x86_intercept_vmload] = POST_EX(SVM_EXIT_VMLOAD),
4110 [x86_intercept_vmsave] = POST_EX(SVM_EXIT_VMSAVE),
4111 [x86_intercept_stgi] = POST_EX(SVM_EXIT_STGI),
4112 [x86_intercept_clgi] = POST_EX(SVM_EXIT_CLGI),
4113 [x86_intercept_skinit] = POST_EX(SVM_EXIT_SKINIT),
4114 [x86_intercept_invlpga] = POST_EX(SVM_EXIT_INVLPGA),
4115 [x86_intercept_rdtscp] = POST_EX(SVM_EXIT_RDTSCP),
4116 [x86_intercept_monitor] = POST_MEM(SVM_EXIT_MONITOR),
4117 [x86_intercept_mwait] = POST_EX(SVM_EXIT_MWAIT),
4118 [x86_intercept_invlpg] = POST_EX(SVM_EXIT_INVLPG),
4119 [x86_intercept_invd] = POST_EX(SVM_EXIT_INVD),
4120 [x86_intercept_wbinvd] = POST_EX(SVM_EXIT_WBINVD),
4121 [x86_intercept_wrmsr] = POST_EX(SVM_EXIT_MSR),
4122 [x86_intercept_rdtsc] = POST_EX(SVM_EXIT_RDTSC),
4123 [x86_intercept_rdmsr] = POST_EX(SVM_EXIT_MSR),
4124 [x86_intercept_rdpmc] = POST_EX(SVM_EXIT_RDPMC),
4125 [x86_intercept_cpuid] = PRE_EX(SVM_EXIT_CPUID),
4126 [x86_intercept_rsm] = PRE_EX(SVM_EXIT_RSM),
4127 [x86_intercept_pause] = PRE_EX(SVM_EXIT_PAUSE),
4128 [x86_intercept_pushf] = PRE_EX(SVM_EXIT_PUSHF),
4129 [x86_intercept_popf] = PRE_EX(SVM_EXIT_POPF),
4130 [x86_intercept_intn] = PRE_EX(SVM_EXIT_SWINT),
4131 [x86_intercept_iret] = PRE_EX(SVM_EXIT_IRET),
4132 [x86_intercept_icebp] = PRE_EX(SVM_EXIT_ICEBP),
4133 [x86_intercept_hlt] = POST_EX(SVM_EXIT_HLT),
4134 [x86_intercept_in] = POST_EX(SVM_EXIT_IOIO),
4135 [x86_intercept_ins] = POST_EX(SVM_EXIT_IOIO),
4136 [x86_intercept_out] = POST_EX(SVM_EXIT_IOIO),
4137 [x86_intercept_outs] = POST_EX(SVM_EXIT_IOIO),
4138 [x86_intercept_xsetbv] = PRE_EX(SVM_EXIT_XSETBV),
4145 static int svm_check_intercept(struct kvm_vcpu *vcpu,
4146 struct x86_instruction_info *info,
4147 enum x86_intercept_stage stage,
4148 struct x86_exception *exception)
4150 struct vcpu_svm *svm = to_svm(vcpu);
4151 int vmexit, ret = X86EMUL_CONTINUE;
4152 struct __x86_intercept icpt_info;
4153 struct vmcb *vmcb = svm->vmcb;
4155 if (info->intercept >= ARRAY_SIZE(x86_intercept_map))
4158 icpt_info = x86_intercept_map[info->intercept];
4160 if (stage != icpt_info.stage)
4163 switch (icpt_info.exit_code) {
4164 case SVM_EXIT_READ_CR0:
4165 if (info->intercept == x86_intercept_cr_read)
4166 icpt_info.exit_code += info->modrm_reg;
4168 case SVM_EXIT_WRITE_CR0: {
4169 unsigned long cr0, val;
4171 if (info->intercept == x86_intercept_cr_write)
4172 icpt_info.exit_code += info->modrm_reg;
4174 if (icpt_info.exit_code != SVM_EXIT_WRITE_CR0 ||
4175 info->intercept == x86_intercept_clts)
4178 if (!(vmcb_is_intercept(&svm->nested.ctl,
4179 INTERCEPT_SELECTIVE_CR0)))
4182 cr0 = vcpu->arch.cr0 & ~SVM_CR0_SELECTIVE_MASK;
4183 val = info->src_val & ~SVM_CR0_SELECTIVE_MASK;
4185 if (info->intercept == x86_intercept_lmsw) {
4188 /* lmsw can't clear PE - catch this here */
4189 if (cr0 & X86_CR0_PE)
4194 icpt_info.exit_code = SVM_EXIT_CR0_SEL_WRITE;
4198 case SVM_EXIT_READ_DR0:
4199 case SVM_EXIT_WRITE_DR0:
4200 icpt_info.exit_code += info->modrm_reg;
4203 if (info->intercept == x86_intercept_wrmsr)
4204 vmcb->control.exit_info_1 = 1;
4206 vmcb->control.exit_info_1 = 0;
4208 case SVM_EXIT_PAUSE:
4210 * We get this for NOP only, but pause
4211 * is rep not, check this here
4213 if (info->rep_prefix != REPE_PREFIX)
4216 case SVM_EXIT_IOIO: {
4220 if (info->intercept == x86_intercept_in ||
4221 info->intercept == x86_intercept_ins) {
4222 exit_info = ((info->src_val & 0xffff) << 16) |
4224 bytes = info->dst_bytes;
4226 exit_info = (info->dst_val & 0xffff) << 16;
4227 bytes = info->src_bytes;
4230 if (info->intercept == x86_intercept_outs ||
4231 info->intercept == x86_intercept_ins)
4232 exit_info |= SVM_IOIO_STR_MASK;
4234 if (info->rep_prefix)
4235 exit_info |= SVM_IOIO_REP_MASK;
4237 bytes = min(bytes, 4u);
4239 exit_info |= bytes << SVM_IOIO_SIZE_SHIFT;
4241 exit_info |= (u32)info->ad_bytes << (SVM_IOIO_ASIZE_SHIFT - 1);
4243 vmcb->control.exit_info_1 = exit_info;
4244 vmcb->control.exit_info_2 = info->next_rip;
4252 /* TODO: Advertise NRIPS to guest hypervisor unconditionally */
4253 if (static_cpu_has(X86_FEATURE_NRIPS))
4254 vmcb->control.next_rip = info->next_rip;
4255 vmcb->control.exit_code = icpt_info.exit_code;
4256 vmexit = nested_svm_exit_handled(svm);
4258 ret = (vmexit == NESTED_EXIT_DONE) ? X86EMUL_INTERCEPTED
4265 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
4269 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
4271 if (!kvm_pause_in_guest(vcpu->kvm))
4272 shrink_ple_window(vcpu);
4275 static void svm_setup_mce(struct kvm_vcpu *vcpu)
4277 /* [63:9] are reserved. */
4278 vcpu->arch.mcg_cap &= 0x1ff;
4281 bool svm_smi_blocked(struct kvm_vcpu *vcpu)
4283 struct vcpu_svm *svm = to_svm(vcpu);
4285 /* Per APM Vol.2 15.22.2 "Response to SMI" */
4289 return is_smm(vcpu);
4292 static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
4294 struct vcpu_svm *svm = to_svm(vcpu);
4295 if (svm->nested.nested_run_pending)
4298 /* An SMI must not be injected into L2 if it's supposed to VM-Exit. */
4299 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_smi(svm))
4302 return !svm_smi_blocked(vcpu);
4305 static int svm_pre_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
4307 struct vcpu_svm *svm = to_svm(vcpu);
4310 if (is_guest_mode(vcpu)) {
4311 /* FED8h - SVM Guest */
4312 put_smstate(u64, smstate, 0x7ed8, 1);
4313 /* FEE0h - SVM Guest VMCB Physical Address */
4314 put_smstate(u64, smstate, 0x7ee0, svm->nested.vmcb12_gpa);
4316 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4317 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4318 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4320 ret = nested_svm_vmexit(svm);
4327 static int svm_pre_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
4329 struct vcpu_svm *svm = to_svm(vcpu);
4330 struct kvm_host_map map;
4333 if (guest_cpuid_has(vcpu, X86_FEATURE_LM)) {
4334 u64 saved_efer = GET_SMSTATE(u64, smstate, 0x7ed0);
4335 u64 guest = GET_SMSTATE(u64, smstate, 0x7ed8);
4336 u64 vmcb12_gpa = GET_SMSTATE(u64, smstate, 0x7ee0);
4339 if (!guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4342 if (!(saved_efer & EFER_SVME))
4345 if (kvm_vcpu_map(vcpu,
4346 gpa_to_gfn(vmcb12_gpa), &map) == -EINVAL)
4349 if (svm_allocate_nested(svm))
4352 ret = enter_svm_guest_mode(vcpu, vmcb12_gpa, map.hva);
4353 kvm_vcpu_unmap(vcpu, &map, true);
4360 static void svm_enable_smi_window(struct kvm_vcpu *vcpu)
4362 struct vcpu_svm *svm = to_svm(vcpu);
4364 if (!gif_set(svm)) {
4365 if (vgif_enabled(svm))
4366 svm_set_intercept(svm, INTERCEPT_STGI);
4367 /* STGI will cause a vm exit */
4369 /* We must be in SMM; RSM will cause a vmexit anyway. */
4373 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, void *insn, int insn_len)
4375 bool smep, smap, is_user;
4379 * When the guest is an SEV-ES guest, emulation is not possible.
4381 if (sev_es_guest(vcpu->kvm))
4385 * Detect and workaround Errata 1096 Fam_17h_00_0Fh.
4388 * When CPU raise #NPF on guest data access and vCPU CR4.SMAP=1, it is
4389 * possible that CPU microcode implementing DecodeAssist will fail
4390 * to read bytes of instruction which caused #NPF. In this case,
4391 * GuestIntrBytes field of the VMCB on a VMEXIT will incorrectly
4392 * return 0 instead of the correct guest instruction bytes.
4394 * This happens because CPU microcode reading instruction bytes
4395 * uses a special opcode which attempts to read data using CPL=0
4396 * privileges. The microcode reads CS:RIP and if it hits a SMAP
4397 * fault, it gives up and returns no instruction bytes.
4400 * We reach here in case CPU supports DecodeAssist, raised #NPF and
4401 * returned 0 in GuestIntrBytes field of the VMCB.
4402 * First, errata can only be triggered in case vCPU CR4.SMAP=1.
4403 * Second, if vCPU CR4.SMEP=1, errata could only be triggered
4404 * in case vCPU CPL==3 (Because otherwise guest would have triggered
4405 * a SMEP fault instead of #NPF).
4406 * Otherwise, vCPU CR4.SMEP=0, errata could be triggered by any vCPU CPL.
4407 * As most guests enable SMAP if they have also enabled SMEP, use above
4408 * logic in order to attempt minimize false-positive of detecting errata
4409 * while still preserving all cases semantic correctness.
4412 * To determine what instruction the guest was executing, the hypervisor
4413 * will have to decode the instruction at the instruction pointer.
4415 * In non SEV guest, hypervisor will be able to read the guest
4416 * memory to decode the instruction pointer when insn_len is zero
4417 * so we return true to indicate that decoding is possible.
4419 * But in the SEV guest, the guest memory is encrypted with the
4420 * guest specific key and hypervisor will not be able to decode the
4421 * instruction pointer so we will not able to workaround it. Lets
4422 * print the error and request to kill the guest.
4424 if (likely(!insn || insn_len))
4428 * If RIP is invalid, go ahead with emulation which will cause an
4429 * internal error exit.
4431 if (!kvm_vcpu_gfn_to_memslot(vcpu, kvm_rip_read(vcpu) >> PAGE_SHIFT))
4434 cr4 = kvm_read_cr4(vcpu);
4435 smep = cr4 & X86_CR4_SMEP;
4436 smap = cr4 & X86_CR4_SMAP;
4437 is_user = svm_get_cpl(vcpu) == 3;
4438 if (smap && (!smep || is_user)) {
4439 if (!sev_guest(vcpu->kvm))
4442 pr_err_ratelimited("KVM: SEV Guest triggered AMD Erratum 1096\n");
4443 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
4449 static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
4451 struct vcpu_svm *svm = to_svm(vcpu);
4454 * TODO: Last condition latch INIT signals on vCPU when
4455 * vCPU is in guest-mode and vmcb12 defines intercept on INIT.
4456 * To properly emulate the INIT intercept,
4457 * svm_check_nested_events() should call nested_svm_vmexit()
4458 * if an INIT signal is pending.
4460 return !gif_set(svm) ||
4461 (vmcb_is_intercept(&svm->vmcb->control, INTERCEPT_INIT));
4464 static void svm_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
4466 if (!sev_es_guest(vcpu->kvm))
4467 return kvm_vcpu_deliver_sipi_vector(vcpu, vector);
4469 sev_vcpu_deliver_sipi_vector(vcpu, vector);
4472 static void svm_vm_destroy(struct kvm *kvm)
4474 avic_vm_destroy(kvm);
4475 sev_vm_destroy(kvm);
4478 static int svm_vm_init(struct kvm *kvm)
4480 if (!pause_filter_count || !pause_filter_thresh)
4481 kvm->arch.pause_in_guest = true;
4484 int ret = avic_vm_init(kvm);
4489 kvm_apicv_init(kvm, avic);
4493 static struct kvm_x86_ops svm_x86_ops __initdata = {
4494 .hardware_unsetup = svm_hardware_teardown,
4495 .hardware_enable = svm_hardware_enable,
4496 .hardware_disable = svm_hardware_disable,
4497 .cpu_has_accelerated_tpr = svm_cpu_has_accelerated_tpr,
4498 .has_emulated_msr = svm_has_emulated_msr,
4500 .vcpu_create = svm_create_vcpu,
4501 .vcpu_free = svm_free_vcpu,
4502 .vcpu_reset = svm_vcpu_reset,
4504 .vm_size = sizeof(struct kvm_svm),
4505 .vm_init = svm_vm_init,
4506 .vm_destroy = svm_vm_destroy,
4508 .prepare_guest_switch = svm_prepare_guest_switch,
4509 .vcpu_load = svm_vcpu_load,
4510 .vcpu_put = svm_vcpu_put,
4511 .vcpu_blocking = svm_vcpu_blocking,
4512 .vcpu_unblocking = svm_vcpu_unblocking,
4514 .update_exception_bitmap = svm_update_exception_bitmap,
4515 .get_msr_feature = svm_get_msr_feature,
4516 .get_msr = svm_get_msr,
4517 .set_msr = svm_set_msr,
4518 .get_segment_base = svm_get_segment_base,
4519 .get_segment = svm_get_segment,
4520 .set_segment = svm_set_segment,
4521 .get_cpl = svm_get_cpl,
4522 .get_cs_db_l_bits = kvm_get_cs_db_l_bits,
4523 .set_cr0 = svm_set_cr0,
4524 .is_valid_cr4 = svm_is_valid_cr4,
4525 .set_cr4 = svm_set_cr4,
4526 .set_efer = svm_set_efer,
4527 .get_idt = svm_get_idt,
4528 .set_idt = svm_set_idt,
4529 .get_gdt = svm_get_gdt,
4530 .set_gdt = svm_set_gdt,
4531 .set_dr7 = svm_set_dr7,
4532 .sync_dirty_debug_regs = svm_sync_dirty_debug_regs,
4533 .cache_reg = svm_cache_reg,
4534 .get_rflags = svm_get_rflags,
4535 .set_rflags = svm_set_rflags,
4537 .tlb_flush_all = svm_flush_tlb,
4538 .tlb_flush_current = svm_flush_tlb,
4539 .tlb_flush_gva = svm_flush_tlb_gva,
4540 .tlb_flush_guest = svm_flush_tlb,
4542 .run = svm_vcpu_run,
4543 .handle_exit = handle_exit,
4544 .skip_emulated_instruction = skip_emulated_instruction,
4545 .update_emulated_instruction = NULL,
4546 .set_interrupt_shadow = svm_set_interrupt_shadow,
4547 .get_interrupt_shadow = svm_get_interrupt_shadow,
4548 .patch_hypercall = svm_patch_hypercall,
4549 .set_irq = svm_set_irq,
4550 .set_nmi = svm_inject_nmi,
4551 .queue_exception = svm_queue_exception,
4552 .cancel_injection = svm_cancel_injection,
4553 .interrupt_allowed = svm_interrupt_allowed,
4554 .nmi_allowed = svm_nmi_allowed,
4555 .get_nmi_mask = svm_get_nmi_mask,
4556 .set_nmi_mask = svm_set_nmi_mask,
4557 .enable_nmi_window = svm_enable_nmi_window,
4558 .enable_irq_window = svm_enable_irq_window,
4559 .update_cr8_intercept = svm_update_cr8_intercept,
4560 .set_virtual_apic_mode = svm_set_virtual_apic_mode,
4561 .refresh_apicv_exec_ctrl = svm_refresh_apicv_exec_ctrl,
4562 .check_apicv_inhibit_reasons = svm_check_apicv_inhibit_reasons,
4563 .pre_update_apicv_exec_ctrl = svm_pre_update_apicv_exec_ctrl,
4564 .load_eoi_exitmap = svm_load_eoi_exitmap,
4565 .hwapic_irr_update = svm_hwapic_irr_update,
4566 .hwapic_isr_update = svm_hwapic_isr_update,
4567 .sync_pir_to_irr = kvm_lapic_find_highest_irr,
4568 .apicv_post_state_restore = avic_post_state_restore,
4570 .set_tss_addr = svm_set_tss_addr,
4571 .set_identity_map_addr = svm_set_identity_map_addr,
4572 .get_mt_mask = svm_get_mt_mask,
4574 .get_exit_info = svm_get_exit_info,
4576 .vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
4578 .has_wbinvd_exit = svm_has_wbinvd_exit,
4580 .write_l1_tsc_offset = svm_write_l1_tsc_offset,
4582 .load_mmu_pgd = svm_load_mmu_pgd,
4584 .check_intercept = svm_check_intercept,
4585 .handle_exit_irqoff = svm_handle_exit_irqoff,
4587 .request_immediate_exit = __kvm_request_immediate_exit,
4589 .sched_in = svm_sched_in,
4591 .pmu_ops = &amd_pmu_ops,
4592 .nested_ops = &svm_nested_ops,
4594 .deliver_posted_interrupt = svm_deliver_avic_intr,
4595 .dy_apicv_has_pending_interrupt = svm_dy_apicv_has_pending_interrupt,
4596 .update_pi_irte = svm_update_pi_irte,
4597 .setup_mce = svm_setup_mce,
4599 .smi_allowed = svm_smi_allowed,
4600 .pre_enter_smm = svm_pre_enter_smm,
4601 .pre_leave_smm = svm_pre_leave_smm,
4602 .enable_smi_window = svm_enable_smi_window,
4604 .mem_enc_op = svm_mem_enc_op,
4605 .mem_enc_reg_region = svm_register_enc_region,
4606 .mem_enc_unreg_region = svm_unregister_enc_region,
4608 .vm_copy_enc_context_from = svm_vm_copy_asid_from,
4610 .can_emulate_instruction = svm_can_emulate_instruction,
4612 .apic_init_signal_blocked = svm_apic_init_signal_blocked,
4614 .msr_filter_changed = svm_msr_filter_changed,
4615 .complete_emulated_msr = svm_complete_emulated_msr,
4617 .vcpu_deliver_sipi_vector = svm_vcpu_deliver_sipi_vector,
4620 static struct kvm_x86_init_ops svm_init_ops __initdata = {
4621 .cpu_has_kvm_support = has_svm,
4622 .disabled_by_bios = is_disabled,
4623 .hardware_setup = svm_hardware_setup,
4624 .check_processor_compatibility = svm_check_processor_compat,
4626 .runtime_ops = &svm_x86_ops,
4629 static int __init svm_init(void)
4631 __unused_size_checks();
4633 return kvm_init(&svm_init_ops, sizeof(struct vcpu_svm),
4634 __alignof__(struct vcpu_svm), THIS_MODULE);
4637 static void __exit svm_exit(void)
4642 module_init(svm_init)
4643 module_exit(svm_exit)