1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3 #include <linux/kvm_host.h>
7 #include "kvm_cache_regs.h"
13 #include <linux/module.h>
14 #include <linux/mod_devicetable.h>
15 #include <linux/kernel.h>
16 #include <linux/vmalloc.h>
17 #include <linux/highmem.h>
18 #include <linux/amd-iommu.h>
19 #include <linux/sched.h>
20 #include <linux/trace_events.h>
21 #include <linux/slab.h>
22 #include <linux/hashtable.h>
23 #include <linux/objtool.h>
24 #include <linux/psp-sev.h>
25 #include <linux/file.h>
26 #include <linux/pagemap.h>
27 #include <linux/swap.h>
28 #include <linux/rwsem.h>
29 #include <linux/cc_platform.h>
30 #include <linux/smp.h>
33 #include <asm/perf_event.h>
34 #include <asm/tlbflush.h>
36 #include <asm/debugreg.h>
37 #include <asm/kvm_para.h>
38 #include <asm/irq_remapping.h>
39 #include <asm/spec-ctrl.h>
40 #include <asm/cpu_device_id.h>
41 #include <asm/traps.h>
42 #include <asm/reboot.h>
43 #include <asm/fpu/api.h>
45 #include <trace/events/ipi.h>
52 #include "kvm_onhyperv.h"
53 #include "svm_onhyperv.h"
55 MODULE_AUTHOR("Qumranet");
56 MODULE_LICENSE("GPL");
59 static const struct x86_cpu_id svm_cpu_id[] = {
60 X86_MATCH_FEATURE(X86_FEATURE_SVM, NULL),
63 MODULE_DEVICE_TABLE(x86cpu, svm_cpu_id);
66 #define SEG_TYPE_LDT 2
67 #define SEG_TYPE_BUSY_TSS16 3
69 static bool erratum_383_found __read_mostly;
71 u32 msrpm_offsets[MSRPM_OFFSETS] __read_mostly;
74 * Set osvw_len to higher value when updated Revision Guides
75 * are published and we know what the new status bits are
77 static uint64_t osvw_len = 4, osvw_status;
79 static DEFINE_PER_CPU(u64, current_tsc_ratio);
81 #define X2APIC_MSR(x) (APIC_BASE_MSR + (x >> 4))
83 static const struct svm_direct_access_msrs {
84 u32 index; /* Index of the MSR */
85 bool always; /* True if intercept is initially cleared */
86 } direct_access_msrs[MAX_DIRECT_ACCESS_MSRS] = {
87 { .index = MSR_STAR, .always = true },
88 { .index = MSR_IA32_SYSENTER_CS, .always = true },
89 { .index = MSR_IA32_SYSENTER_EIP, .always = false },
90 { .index = MSR_IA32_SYSENTER_ESP, .always = false },
92 { .index = MSR_GS_BASE, .always = true },
93 { .index = MSR_FS_BASE, .always = true },
94 { .index = MSR_KERNEL_GS_BASE, .always = true },
95 { .index = MSR_LSTAR, .always = true },
96 { .index = MSR_CSTAR, .always = true },
97 { .index = MSR_SYSCALL_MASK, .always = true },
99 { .index = MSR_IA32_SPEC_CTRL, .always = false },
100 { .index = MSR_IA32_PRED_CMD, .always = false },
101 { .index = MSR_IA32_FLUSH_CMD, .always = false },
102 { .index = MSR_IA32_LASTBRANCHFROMIP, .always = false },
103 { .index = MSR_IA32_LASTBRANCHTOIP, .always = false },
104 { .index = MSR_IA32_LASTINTFROMIP, .always = false },
105 { .index = MSR_IA32_LASTINTTOIP, .always = false },
106 { .index = MSR_EFER, .always = false },
107 { .index = MSR_IA32_CR_PAT, .always = false },
108 { .index = MSR_AMD64_SEV_ES_GHCB, .always = true },
109 { .index = MSR_TSC_AUX, .always = false },
110 { .index = X2APIC_MSR(APIC_ID), .always = false },
111 { .index = X2APIC_MSR(APIC_LVR), .always = false },
112 { .index = X2APIC_MSR(APIC_TASKPRI), .always = false },
113 { .index = X2APIC_MSR(APIC_ARBPRI), .always = false },
114 { .index = X2APIC_MSR(APIC_PROCPRI), .always = false },
115 { .index = X2APIC_MSR(APIC_EOI), .always = false },
116 { .index = X2APIC_MSR(APIC_RRR), .always = false },
117 { .index = X2APIC_MSR(APIC_LDR), .always = false },
118 { .index = X2APIC_MSR(APIC_DFR), .always = false },
119 { .index = X2APIC_MSR(APIC_SPIV), .always = false },
120 { .index = X2APIC_MSR(APIC_ISR), .always = false },
121 { .index = X2APIC_MSR(APIC_TMR), .always = false },
122 { .index = X2APIC_MSR(APIC_IRR), .always = false },
123 { .index = X2APIC_MSR(APIC_ESR), .always = false },
124 { .index = X2APIC_MSR(APIC_ICR), .always = false },
125 { .index = X2APIC_MSR(APIC_ICR2), .always = false },
129 * AMD does not virtualize APIC TSC-deadline timer mode, but it is
130 * emulated by KVM. When setting APIC LVTT (0x832) register bit 18,
131 * the AVIC hardware would generate GP fault. Therefore, always
132 * intercept the MSR 0x832, and do not setup direct_access_msr.
134 { .index = X2APIC_MSR(APIC_LVTTHMR), .always = false },
135 { .index = X2APIC_MSR(APIC_LVTPC), .always = false },
136 { .index = X2APIC_MSR(APIC_LVT0), .always = false },
137 { .index = X2APIC_MSR(APIC_LVT1), .always = false },
138 { .index = X2APIC_MSR(APIC_LVTERR), .always = false },
139 { .index = X2APIC_MSR(APIC_TMICT), .always = false },
140 { .index = X2APIC_MSR(APIC_TMCCT), .always = false },
141 { .index = X2APIC_MSR(APIC_TDCR), .always = false },
142 { .index = MSR_INVALID, .always = false },
146 * These 2 parameters are used to config the controls for Pause-Loop Exiting:
147 * pause_filter_count: On processors that support Pause filtering(indicated
148 * by CPUID Fn8000_000A_EDX), the VMCB provides a 16 bit pause filter
149 * count value. On VMRUN this value is loaded into an internal counter.
150 * Each time a pause instruction is executed, this counter is decremented
151 * until it reaches zero at which time a #VMEXIT is generated if pause
152 * intercept is enabled. Refer to AMD APM Vol 2 Section 15.14.4 Pause
153 * Intercept Filtering for more details.
154 * This also indicate if ple logic enabled.
156 * pause_filter_thresh: In addition, some processor families support advanced
157 * pause filtering (indicated by CPUID Fn8000_000A_EDX) upper bound on
158 * the amount of time a guest is allowed to execute in a pause loop.
159 * In this mode, a 16-bit pause filter threshold field is added in the
160 * VMCB. The threshold value is a cycle count that is used to reset the
161 * pause counter. As with simple pause filtering, VMRUN loads the pause
162 * count value from VMCB into an internal counter. Then, on each pause
163 * instruction the hardware checks the elapsed number of cycles since
164 * the most recent pause instruction against the pause filter threshold.
165 * If the elapsed cycle count is greater than the pause filter threshold,
166 * then the internal pause count is reloaded from the VMCB and execution
167 * continues. If the elapsed cycle count is less than the pause filter
168 * threshold, then the internal pause count is decremented. If the count
169 * value is less than zero and PAUSE intercept is enabled, a #VMEXIT is
170 * triggered. If advanced pause filtering is supported and pause filter
171 * threshold field is set to zero, the filter will operate in the simpler,
175 static unsigned short pause_filter_thresh = KVM_DEFAULT_PLE_GAP;
176 module_param(pause_filter_thresh, ushort, 0444);
178 static unsigned short pause_filter_count = KVM_SVM_DEFAULT_PLE_WINDOW;
179 module_param(pause_filter_count, ushort, 0444);
181 /* Default doubles per-vcpu window every exit. */
182 static unsigned short pause_filter_count_grow = KVM_DEFAULT_PLE_WINDOW_GROW;
183 module_param(pause_filter_count_grow, ushort, 0444);
185 /* Default resets per-vcpu window every exit to pause_filter_count. */
186 static unsigned short pause_filter_count_shrink = KVM_DEFAULT_PLE_WINDOW_SHRINK;
187 module_param(pause_filter_count_shrink, ushort, 0444);
189 /* Default is to compute the maximum so we can never overflow. */
190 static unsigned short pause_filter_count_max = KVM_SVM_DEFAULT_PLE_WINDOW_MAX;
191 module_param(pause_filter_count_max, ushort, 0444);
194 * Use nested page tables by default. Note, NPT may get forced off by
195 * svm_hardware_setup() if it's unsupported by hardware or the host kernel.
197 bool npt_enabled = true;
198 module_param_named(npt, npt_enabled, bool, 0444);
200 /* allow nested virtualization in KVM/SVM */
201 static int nested = true;
202 module_param(nested, int, S_IRUGO);
204 /* enable/disable Next RIP Save */
206 module_param(nrips, int, 0444);
208 /* enable/disable Virtual VMLOAD VMSAVE */
209 static int vls = true;
210 module_param(vls, int, 0444);
212 /* enable/disable Virtual GIF */
214 module_param(vgif, int, 0444);
216 /* enable/disable LBR virtualization */
217 static int lbrv = true;
218 module_param(lbrv, int, 0444);
220 static int tsc_scaling = true;
221 module_param(tsc_scaling, int, 0444);
224 * enable / disable AVIC. Because the defaults differ for APICv
225 * support between VMX and SVM we cannot use module_param_named.
228 module_param(avic, bool, 0444);
230 bool __read_mostly dump_invalid_vmcb;
231 module_param(dump_invalid_vmcb, bool, 0644);
234 bool intercept_smi = true;
235 module_param(intercept_smi, bool, 0444);
238 module_param(vnmi, bool, 0444);
240 static bool svm_gp_erratum_intercept = true;
242 static u8 rsm_ins_bytes[] = "\x0f\xaa";
244 static unsigned long iopm_base;
246 DEFINE_PER_CPU(struct svm_cpu_data, svm_data);
249 * Only MSR_TSC_AUX is switched via the user return hook. EFER is switched via
250 * the VMCB, and the SYSCALL/SYSENTER MSRs are handled by VMLOAD/VMSAVE.
252 * RDTSCP and RDPID are not used in the kernel, specifically to allow KVM to
253 * defer the restoration of TSC_AUX until the CPU returns to userspace.
255 static int tsc_aux_uret_slot __read_mostly = -1;
257 static const u32 msrpm_ranges[] = {0, 0xc0000000, 0xc0010000};
259 #define NUM_MSR_MAPS ARRAY_SIZE(msrpm_ranges)
260 #define MSRS_RANGE_SIZE 2048
261 #define MSRS_IN_RANGE (MSRS_RANGE_SIZE * 8 / 2)
263 u32 svm_msrpm_offset(u32 msr)
268 for (i = 0; i < NUM_MSR_MAPS; i++) {
269 if (msr < msrpm_ranges[i] ||
270 msr >= msrpm_ranges[i] + MSRS_IN_RANGE)
273 offset = (msr - msrpm_ranges[i]) / 4; /* 4 msrs per u8 */
274 offset += (i * MSRS_RANGE_SIZE); /* add range offset */
276 /* Now we have the u8 offset - but need the u32 offset */
280 /* MSR not in any range */
284 static void svm_flush_tlb_current(struct kvm_vcpu *vcpu);
286 static int get_npt_level(void)
289 return pgtable_l5_enabled() ? PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
291 return PT32E_ROOT_LEVEL;
295 int svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)
297 struct vcpu_svm *svm = to_svm(vcpu);
298 u64 old_efer = vcpu->arch.efer;
299 vcpu->arch.efer = efer;
302 /* Shadow paging assumes NX to be available. */
305 if (!(efer & EFER_LMA))
309 if ((old_efer & EFER_SVME) != (efer & EFER_SVME)) {
310 if (!(efer & EFER_SVME)) {
311 svm_leave_nested(vcpu);
312 svm_set_gif(svm, true);
313 /* #GP intercept is still needed for vmware backdoor */
314 if (!enable_vmware_backdoor)
315 clr_exception_intercept(svm, GP_VECTOR);
318 * Free the nested guest state, unless we are in SMM.
319 * In this case we will return to the nested guest
320 * as soon as we leave SMM.
323 svm_free_nested(svm);
326 int ret = svm_allocate_nested(svm);
329 vcpu->arch.efer = old_efer;
334 * Never intercept #GP for SEV guests, KVM can't
335 * decrypt guest memory to workaround the erratum.
337 if (svm_gp_erratum_intercept && !sev_guest(vcpu->kvm))
338 set_exception_intercept(svm, GP_VECTOR);
342 svm->vmcb->save.efer = efer | EFER_SVME;
343 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
347 static u32 svm_get_interrupt_shadow(struct kvm_vcpu *vcpu)
349 struct vcpu_svm *svm = to_svm(vcpu);
352 if (svm->vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK)
353 ret = KVM_X86_SHADOW_INT_STI | KVM_X86_SHADOW_INT_MOV_SS;
357 static void svm_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
359 struct vcpu_svm *svm = to_svm(vcpu);
362 svm->vmcb->control.int_state &= ~SVM_INTERRUPT_SHADOW_MASK;
364 svm->vmcb->control.int_state |= SVM_INTERRUPT_SHADOW_MASK;
367 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, int emul_type,
368 void *insn, int insn_len);
370 static int __svm_skip_emulated_instruction(struct kvm_vcpu *vcpu,
371 bool commit_side_effects)
373 struct vcpu_svm *svm = to_svm(vcpu);
374 unsigned long old_rflags;
377 * SEV-ES does not expose the next RIP. The RIP update is controlled by
378 * the type of exit and the #VC handler in the guest.
380 if (sev_es_guest(vcpu->kvm))
383 if (nrips && svm->vmcb->control.next_rip != 0) {
384 WARN_ON_ONCE(!static_cpu_has(X86_FEATURE_NRIPS));
385 svm->next_rip = svm->vmcb->control.next_rip;
388 if (!svm->next_rip) {
390 * FIXME: Drop this when kvm_emulate_instruction() does the
391 * right thing and treats "can't emulate" as outright failure
394 if (!svm_can_emulate_instruction(vcpu, EMULTYPE_SKIP, NULL, 0))
397 if (unlikely(!commit_side_effects))
398 old_rflags = svm->vmcb->save.rflags;
400 if (!kvm_emulate_instruction(vcpu, EMULTYPE_SKIP))
403 if (unlikely(!commit_side_effects))
404 svm->vmcb->save.rflags = old_rflags;
406 kvm_rip_write(vcpu, svm->next_rip);
410 if (likely(commit_side_effects))
411 svm_set_interrupt_shadow(vcpu, 0);
416 static int svm_skip_emulated_instruction(struct kvm_vcpu *vcpu)
418 return __svm_skip_emulated_instruction(vcpu, true);
421 static int svm_update_soft_interrupt_rip(struct kvm_vcpu *vcpu)
423 unsigned long rip, old_rip = kvm_rip_read(vcpu);
424 struct vcpu_svm *svm = to_svm(vcpu);
427 * Due to architectural shortcomings, the CPU doesn't always provide
428 * NextRIP, e.g. if KVM intercepted an exception that occurred while
429 * the CPU was vectoring an INTO/INT3 in the guest. Temporarily skip
430 * the instruction even if NextRIP is supported to acquire the next
431 * RIP so that it can be shoved into the NextRIP field, otherwise
432 * hardware will fail to advance guest RIP during event injection.
433 * Drop the exception/interrupt if emulation fails and effectively
434 * retry the instruction, it's the least awful option. If NRIPS is
435 * in use, the skip must not commit any side effects such as clearing
436 * the interrupt shadow or RFLAGS.RF.
438 if (!__svm_skip_emulated_instruction(vcpu, !nrips))
441 rip = kvm_rip_read(vcpu);
444 * Save the injection information, even when using next_rip, as the
445 * VMCB's next_rip will be lost (cleared on VM-Exit) if the injection
446 * doesn't complete due to a VM-Exit occurring while the CPU is
447 * vectoring the event. Decoding the instruction isn't guaranteed to
448 * work as there may be no backing instruction, e.g. if the event is
449 * being injected by L1 for L2, or if the guest is patching INT3 into
450 * a different instruction.
452 svm->soft_int_injected = true;
453 svm->soft_int_csbase = svm->vmcb->save.cs.base;
454 svm->soft_int_old_rip = old_rip;
455 svm->soft_int_next_rip = rip;
458 kvm_rip_write(vcpu, old_rip);
460 if (static_cpu_has(X86_FEATURE_NRIPS))
461 svm->vmcb->control.next_rip = rip;
466 static void svm_inject_exception(struct kvm_vcpu *vcpu)
468 struct kvm_queued_exception *ex = &vcpu->arch.exception;
469 struct vcpu_svm *svm = to_svm(vcpu);
471 kvm_deliver_exception_payload(vcpu, ex);
473 if (kvm_exception_is_soft(ex->vector) &&
474 svm_update_soft_interrupt_rip(vcpu))
477 svm->vmcb->control.event_inj = ex->vector
479 | (ex->has_error_code ? SVM_EVTINJ_VALID_ERR : 0)
480 | SVM_EVTINJ_TYPE_EXEPT;
481 svm->vmcb->control.event_inj_err = ex->error_code;
484 static void svm_init_erratum_383(void)
490 if (!static_cpu_has_bug(X86_BUG_AMD_TLB_MMATCH))
493 /* Use _safe variants to not break nested virtualization */
494 val = native_read_msr_safe(MSR_AMD64_DC_CFG, &err);
500 low = lower_32_bits(val);
501 high = upper_32_bits(val);
503 native_write_msr_safe(MSR_AMD64_DC_CFG, low, high);
505 erratum_383_found = true;
508 static void svm_init_osvw(struct kvm_vcpu *vcpu)
511 * Guests should see errata 400 and 415 as fixed (assuming that
512 * HLT and IO instructions are intercepted).
514 vcpu->arch.osvw.length = (osvw_len >= 3) ? (osvw_len) : 3;
515 vcpu->arch.osvw.status = osvw_status & ~(6ULL);
518 * By increasing VCPU's osvw.length to 3 we are telling the guest that
519 * all osvw.status bits inside that length, including bit 0 (which is
520 * reserved for erratum 298), are valid. However, if host processor's
521 * osvw_len is 0 then osvw_status[0] carries no information. We need to
522 * be conservative here and therefore we tell the guest that erratum 298
523 * is present (because we really don't know).
525 if (osvw_len == 0 && boot_cpu_data.x86 == 0x10)
526 vcpu->arch.osvw.status |= 1;
529 static bool __kvm_is_svm_supported(void)
531 int cpu = smp_processor_id();
532 struct cpuinfo_x86 *c = &cpu_data(cpu);
536 if (c->x86_vendor != X86_VENDOR_AMD &&
537 c->x86_vendor != X86_VENDOR_HYGON) {
538 pr_err("CPU %d isn't AMD or Hygon\n", cpu);
542 if (!cpu_has(c, X86_FEATURE_SVM)) {
543 pr_err("SVM not supported by CPU %d\n", cpu);
547 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) {
548 pr_info("KVM is unsupported when running as an SEV guest\n");
552 rdmsrl(MSR_VM_CR, vm_cr);
553 if (vm_cr & (1 << SVM_VM_CR_SVM_DISABLE)) {
554 pr_err("SVM disabled (by BIOS) in MSR_VM_CR on CPU %d\n", cpu);
561 static bool kvm_is_svm_supported(void)
566 supported = __kvm_is_svm_supported();
572 static int svm_check_processor_compat(void)
574 if (!__kvm_is_svm_supported())
580 static void __svm_write_tsc_multiplier(u64 multiplier)
582 if (multiplier == __this_cpu_read(current_tsc_ratio))
585 wrmsrl(MSR_AMD64_TSC_RATIO, multiplier);
586 __this_cpu_write(current_tsc_ratio, multiplier);
589 static inline void kvm_cpu_svm_disable(void)
593 wrmsrl(MSR_VM_HSAVE_PA, 0);
594 rdmsrl(MSR_EFER, efer);
595 if (efer & EFER_SVME) {
597 * Force GIF=1 prior to disabling SVM, e.g. to ensure INIT and
598 * NMI aren't blocked.
601 wrmsrl(MSR_EFER, efer & ~EFER_SVME);
605 static void svm_emergency_disable(void)
607 kvm_rebooting = true;
609 kvm_cpu_svm_disable();
612 static void svm_hardware_disable(void)
614 /* Make sure we clean up behind us */
616 __svm_write_tsc_multiplier(SVM_TSC_RATIO_DEFAULT);
618 kvm_cpu_svm_disable();
620 amd_pmu_disable_virt();
623 static int svm_hardware_enable(void)
626 struct svm_cpu_data *sd;
628 int me = raw_smp_processor_id();
630 rdmsrl(MSR_EFER, efer);
631 if (efer & EFER_SVME)
634 sd = per_cpu_ptr(&svm_data, me);
635 sd->asid_generation = 1;
636 sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
637 sd->next_asid = sd->max_asid + 1;
638 sd->min_asid = max_sev_asid + 1;
640 wrmsrl(MSR_EFER, efer | EFER_SVME);
642 wrmsrl(MSR_VM_HSAVE_PA, sd->save_area_pa);
644 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
646 * Set the default value, even if we don't use TSC scaling
647 * to avoid having stale value in the msr
649 __svm_write_tsc_multiplier(SVM_TSC_RATIO_DEFAULT);
656 * Note that it is possible to have a system with mixed processor
657 * revisions and therefore different OSVW bits. If bits are not the same
658 * on different processors then choose the worst case (i.e. if erratum
659 * is present on one processor and not on another then assume that the
660 * erratum is present everywhere).
662 if (cpu_has(&boot_cpu_data, X86_FEATURE_OSVW)) {
663 uint64_t len, status = 0;
666 len = native_read_msr_safe(MSR_AMD64_OSVW_ID_LENGTH, &err);
668 status = native_read_msr_safe(MSR_AMD64_OSVW_STATUS,
672 osvw_status = osvw_len = 0;
676 osvw_status |= status;
677 osvw_status &= (1ULL << osvw_len) - 1;
680 osvw_status = osvw_len = 0;
682 svm_init_erratum_383();
684 amd_pmu_enable_virt();
689 static void svm_cpu_uninit(int cpu)
691 struct svm_cpu_data *sd = per_cpu_ptr(&svm_data, cpu);
696 kfree(sd->sev_vmcbs);
697 __free_page(sd->save_area);
698 sd->save_area_pa = 0;
699 sd->save_area = NULL;
702 static int svm_cpu_init(int cpu)
704 struct svm_cpu_data *sd = per_cpu_ptr(&svm_data, cpu);
707 memset(sd, 0, sizeof(struct svm_cpu_data));
708 sd->save_area = alloc_page(GFP_KERNEL | __GFP_ZERO);
712 ret = sev_cpu_init(sd);
716 sd->save_area_pa = __sme_page_pa(sd->save_area);
720 __free_page(sd->save_area);
721 sd->save_area = NULL;
726 static void set_dr_intercepts(struct vcpu_svm *svm)
728 struct vmcb *vmcb = svm->vmcb01.ptr;
730 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR0_READ);
731 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR1_READ);
732 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR2_READ);
733 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR3_READ);
734 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR4_READ);
735 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR5_READ);
736 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR6_READ);
737 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR0_WRITE);
738 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR1_WRITE);
739 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR2_WRITE);
740 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR3_WRITE);
741 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR4_WRITE);
742 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR5_WRITE);
743 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR6_WRITE);
744 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_READ);
745 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_WRITE);
747 recalc_intercepts(svm);
750 static void clr_dr_intercepts(struct vcpu_svm *svm)
752 struct vmcb *vmcb = svm->vmcb01.ptr;
754 vmcb->control.intercepts[INTERCEPT_DR] = 0;
756 recalc_intercepts(svm);
759 static int direct_access_msr_slot(u32 msr)
763 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++)
764 if (direct_access_msrs[i].index == msr)
770 static void set_shadow_msr_intercept(struct kvm_vcpu *vcpu, u32 msr, int read,
773 struct vcpu_svm *svm = to_svm(vcpu);
774 int slot = direct_access_msr_slot(msr);
779 /* Set the shadow bitmaps to the desired intercept states */
781 set_bit(slot, svm->shadow_msr_intercept.read);
783 clear_bit(slot, svm->shadow_msr_intercept.read);
786 set_bit(slot, svm->shadow_msr_intercept.write);
788 clear_bit(slot, svm->shadow_msr_intercept.write);
791 static bool valid_msr_intercept(u32 index)
793 return direct_access_msr_slot(index) != -ENOENT;
796 static bool msr_write_intercepted(struct kvm_vcpu *vcpu, u32 msr)
804 * For non-nested case:
805 * If the L01 MSR bitmap does not intercept the MSR, then we need to
809 * If the L02 MSR bitmap does not intercept the MSR, then we need to
812 msrpm = is_guest_mode(vcpu) ? to_svm(vcpu)->nested.msrpm:
815 offset = svm_msrpm_offset(msr);
816 bit_write = 2 * (msr & 0x0f) + 1;
819 BUG_ON(offset == MSR_INVALID);
821 return test_bit(bit_write, &tmp);
824 static void set_msr_interception_bitmap(struct kvm_vcpu *vcpu, u32 *msrpm,
825 u32 msr, int read, int write)
827 struct vcpu_svm *svm = to_svm(vcpu);
828 u8 bit_read, bit_write;
833 * If this warning triggers extend the direct_access_msrs list at the
834 * beginning of the file
836 WARN_ON(!valid_msr_intercept(msr));
838 /* Enforce non allowed MSRs to trap */
839 if (read && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_READ))
842 if (write && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_WRITE))
845 offset = svm_msrpm_offset(msr);
846 bit_read = 2 * (msr & 0x0f);
847 bit_write = 2 * (msr & 0x0f) + 1;
850 BUG_ON(offset == MSR_INVALID);
852 read ? clear_bit(bit_read, &tmp) : set_bit(bit_read, &tmp);
853 write ? clear_bit(bit_write, &tmp) : set_bit(bit_write, &tmp);
857 svm_hv_vmcb_dirty_nested_enlightenments(vcpu);
858 svm->nested.force_msr_bitmap_recalc = true;
861 void set_msr_interception(struct kvm_vcpu *vcpu, u32 *msrpm, u32 msr,
864 set_shadow_msr_intercept(vcpu, msr, read, write);
865 set_msr_interception_bitmap(vcpu, msrpm, msr, read, write);
868 u32 *svm_vcpu_alloc_msrpm(void)
870 unsigned int order = get_order(MSRPM_SIZE);
871 struct page *pages = alloc_pages(GFP_KERNEL_ACCOUNT, order);
877 msrpm = page_address(pages);
878 memset(msrpm, 0xff, PAGE_SIZE * (1 << order));
883 void svm_vcpu_init_msrpm(struct kvm_vcpu *vcpu, u32 *msrpm)
887 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
888 if (!direct_access_msrs[i].always)
890 set_msr_interception(vcpu, msrpm, direct_access_msrs[i].index, 1, 1);
894 void svm_set_x2apic_msr_interception(struct vcpu_svm *svm, bool intercept)
898 if (intercept == svm->x2avic_msrs_intercepted)
901 if (!x2avic_enabled ||
902 !apic_x2apic_mode(svm->vcpu.arch.apic))
905 for (i = 0; i < MAX_DIRECT_ACCESS_MSRS; i++) {
906 int index = direct_access_msrs[i].index;
908 if ((index < APIC_BASE_MSR) ||
909 (index > APIC_BASE_MSR + 0xff))
911 set_msr_interception(&svm->vcpu, svm->msrpm, index,
912 !intercept, !intercept);
915 svm->x2avic_msrs_intercepted = intercept;
918 void svm_vcpu_free_msrpm(u32 *msrpm)
920 __free_pages(virt_to_page(msrpm), get_order(MSRPM_SIZE));
923 static void svm_msr_filter_changed(struct kvm_vcpu *vcpu)
925 struct vcpu_svm *svm = to_svm(vcpu);
929 * Set intercept permissions for all direct access MSRs again. They
930 * will automatically get filtered through the MSR filter, so we are
931 * back in sync after this.
933 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
934 u32 msr = direct_access_msrs[i].index;
935 u32 read = test_bit(i, svm->shadow_msr_intercept.read);
936 u32 write = test_bit(i, svm->shadow_msr_intercept.write);
938 set_msr_interception_bitmap(vcpu, svm->msrpm, msr, read, write);
942 static void add_msr_offset(u32 offset)
946 for (i = 0; i < MSRPM_OFFSETS; ++i) {
948 /* Offset already in list? */
949 if (msrpm_offsets[i] == offset)
952 /* Slot used by another offset? */
953 if (msrpm_offsets[i] != MSR_INVALID)
956 /* Add offset to list */
957 msrpm_offsets[i] = offset;
963 * If this BUG triggers the msrpm_offsets table has an overflow. Just
964 * increase MSRPM_OFFSETS in this case.
969 static void init_msrpm_offsets(void)
973 memset(msrpm_offsets, 0xff, sizeof(msrpm_offsets));
975 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
978 offset = svm_msrpm_offset(direct_access_msrs[i].index);
979 BUG_ON(offset == MSR_INVALID);
981 add_msr_offset(offset);
985 void svm_copy_lbrs(struct vmcb *to_vmcb, struct vmcb *from_vmcb)
987 to_vmcb->save.dbgctl = from_vmcb->save.dbgctl;
988 to_vmcb->save.br_from = from_vmcb->save.br_from;
989 to_vmcb->save.br_to = from_vmcb->save.br_to;
990 to_vmcb->save.last_excp_from = from_vmcb->save.last_excp_from;
991 to_vmcb->save.last_excp_to = from_vmcb->save.last_excp_to;
993 vmcb_mark_dirty(to_vmcb, VMCB_LBR);
996 static void svm_enable_lbrv(struct kvm_vcpu *vcpu)
998 struct vcpu_svm *svm = to_svm(vcpu);
1000 svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
1001 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
1002 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
1003 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
1004 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
1006 /* Move the LBR msrs to the vmcb02 so that the guest can see them. */
1007 if (is_guest_mode(vcpu))
1008 svm_copy_lbrs(svm->vmcb, svm->vmcb01.ptr);
1011 static void svm_disable_lbrv(struct kvm_vcpu *vcpu)
1013 struct vcpu_svm *svm = to_svm(vcpu);
1015 svm->vmcb->control.virt_ext &= ~LBR_CTL_ENABLE_MASK;
1016 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 0, 0);
1017 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 0, 0);
1018 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 0, 0);
1019 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 0, 0);
1022 * Move the LBR msrs back to the vmcb01 to avoid copying them
1023 * on nested guest entries.
1025 if (is_guest_mode(vcpu))
1026 svm_copy_lbrs(svm->vmcb01.ptr, svm->vmcb);
1029 static struct vmcb *svm_get_lbr_vmcb(struct vcpu_svm *svm)
1032 * If LBR virtualization is disabled, the LBR MSRs are always kept in
1033 * vmcb01. If LBR virtualization is enabled and L1 is running VMs of
1034 * its own, the MSRs are moved between vmcb01 and vmcb02 as needed.
1036 return svm->vmcb->control.virt_ext & LBR_CTL_ENABLE_MASK ? svm->vmcb :
1040 void svm_update_lbrv(struct kvm_vcpu *vcpu)
1042 struct vcpu_svm *svm = to_svm(vcpu);
1043 bool current_enable_lbrv = svm->vmcb->control.virt_ext & LBR_CTL_ENABLE_MASK;
1044 bool enable_lbrv = (svm_get_lbr_vmcb(svm)->save.dbgctl & DEBUGCTLMSR_LBR) ||
1045 (is_guest_mode(vcpu) && guest_can_use(vcpu, X86_FEATURE_LBRV) &&
1046 (svm->nested.ctl.virt_ext & LBR_CTL_ENABLE_MASK));
1048 if (enable_lbrv == current_enable_lbrv)
1052 svm_enable_lbrv(vcpu);
1054 svm_disable_lbrv(vcpu);
1057 void disable_nmi_singlestep(struct vcpu_svm *svm)
1059 svm->nmi_singlestep = false;
1061 if (!(svm->vcpu.guest_debug & KVM_GUESTDBG_SINGLESTEP)) {
1062 /* Clear our flags if they were not set by the guest */
1063 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
1064 svm->vmcb->save.rflags &= ~X86_EFLAGS_TF;
1065 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
1066 svm->vmcb->save.rflags &= ~X86_EFLAGS_RF;
1070 static void grow_ple_window(struct kvm_vcpu *vcpu)
1072 struct vcpu_svm *svm = to_svm(vcpu);
1073 struct vmcb_control_area *control = &svm->vmcb->control;
1074 int old = control->pause_filter_count;
1076 if (kvm_pause_in_guest(vcpu->kvm))
1079 control->pause_filter_count = __grow_ple_window(old,
1081 pause_filter_count_grow,
1082 pause_filter_count_max);
1084 if (control->pause_filter_count != old) {
1085 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1086 trace_kvm_ple_window_update(vcpu->vcpu_id,
1087 control->pause_filter_count, old);
1091 static void shrink_ple_window(struct kvm_vcpu *vcpu)
1093 struct vcpu_svm *svm = to_svm(vcpu);
1094 struct vmcb_control_area *control = &svm->vmcb->control;
1095 int old = control->pause_filter_count;
1097 if (kvm_pause_in_guest(vcpu->kvm))
1100 control->pause_filter_count =
1101 __shrink_ple_window(old,
1103 pause_filter_count_shrink,
1104 pause_filter_count);
1105 if (control->pause_filter_count != old) {
1106 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1107 trace_kvm_ple_window_update(vcpu->vcpu_id,
1108 control->pause_filter_count, old);
1112 static void svm_hardware_unsetup(void)
1116 sev_hardware_unsetup();
1118 for_each_possible_cpu(cpu)
1119 svm_cpu_uninit(cpu);
1121 __free_pages(pfn_to_page(iopm_base >> PAGE_SHIFT),
1122 get_order(IOPM_SIZE));
1126 static void init_seg(struct vmcb_seg *seg)
1129 seg->attrib = SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK |
1130 SVM_SELECTOR_WRITE_MASK; /* Read/Write Data Segment */
1131 seg->limit = 0xffff;
1135 static void init_sys_seg(struct vmcb_seg *seg, uint32_t type)
1138 seg->attrib = SVM_SELECTOR_P_MASK | type;
1139 seg->limit = 0xffff;
1143 static u64 svm_get_l2_tsc_offset(struct kvm_vcpu *vcpu)
1145 struct vcpu_svm *svm = to_svm(vcpu);
1147 return svm->nested.ctl.tsc_offset;
1150 static u64 svm_get_l2_tsc_multiplier(struct kvm_vcpu *vcpu)
1152 struct vcpu_svm *svm = to_svm(vcpu);
1154 return svm->tsc_ratio_msr;
1157 static void svm_write_tsc_offset(struct kvm_vcpu *vcpu)
1159 struct vcpu_svm *svm = to_svm(vcpu);
1161 svm->vmcb01.ptr->control.tsc_offset = vcpu->arch.l1_tsc_offset;
1162 svm->vmcb->control.tsc_offset = vcpu->arch.tsc_offset;
1163 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1166 void svm_write_tsc_multiplier(struct kvm_vcpu *vcpu)
1169 if (to_svm(vcpu)->guest_state_loaded)
1170 __svm_write_tsc_multiplier(vcpu->arch.tsc_scaling_ratio);
1174 /* Evaluate instruction intercepts that depend on guest CPUID features. */
1175 static void svm_recalc_instruction_intercepts(struct kvm_vcpu *vcpu,
1176 struct vcpu_svm *svm)
1179 * Intercept INVPCID if shadow paging is enabled to sync/free shadow
1180 * roots, or if INVPCID is disabled in the guest to inject #UD.
1182 if (kvm_cpu_cap_has(X86_FEATURE_INVPCID)) {
1184 !guest_cpuid_has(&svm->vcpu, X86_FEATURE_INVPCID))
1185 svm_set_intercept(svm, INTERCEPT_INVPCID);
1187 svm_clr_intercept(svm, INTERCEPT_INVPCID);
1190 if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP)) {
1191 if (guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP))
1192 svm_clr_intercept(svm, INTERCEPT_RDTSCP);
1194 svm_set_intercept(svm, INTERCEPT_RDTSCP);
1198 static inline void init_vmcb_after_set_cpuid(struct kvm_vcpu *vcpu)
1200 struct vcpu_svm *svm = to_svm(vcpu);
1202 if (guest_cpuid_is_intel(vcpu)) {
1204 * We must intercept SYSENTER_EIP and SYSENTER_ESP
1205 * accesses because the processor only stores 32 bits.
1206 * For the same reason we cannot use virtual VMLOAD/VMSAVE.
1208 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1209 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1210 svm->vmcb->control.virt_ext &= ~VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
1212 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 0, 0);
1213 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 0, 0);
1216 * If hardware supports Virtual VMLOAD VMSAVE then enable it
1217 * in VMCB and clear intercepts to avoid #VMEXIT.
1220 svm_clr_intercept(svm, INTERCEPT_VMLOAD);
1221 svm_clr_intercept(svm, INTERCEPT_VMSAVE);
1222 svm->vmcb->control.virt_ext |= VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
1224 /* No need to intercept these MSRs */
1225 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 1, 1);
1226 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 1, 1);
1230 static void init_vmcb(struct kvm_vcpu *vcpu)
1232 struct vcpu_svm *svm = to_svm(vcpu);
1233 struct vmcb *vmcb = svm->vmcb01.ptr;
1234 struct vmcb_control_area *control = &vmcb->control;
1235 struct vmcb_save_area *save = &vmcb->save;
1237 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1238 svm_set_intercept(svm, INTERCEPT_CR3_READ);
1239 svm_set_intercept(svm, INTERCEPT_CR4_READ);
1240 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1241 svm_set_intercept(svm, INTERCEPT_CR3_WRITE);
1242 svm_set_intercept(svm, INTERCEPT_CR4_WRITE);
1243 if (!kvm_vcpu_apicv_active(vcpu))
1244 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
1246 set_dr_intercepts(svm);
1248 set_exception_intercept(svm, PF_VECTOR);
1249 set_exception_intercept(svm, UD_VECTOR);
1250 set_exception_intercept(svm, MC_VECTOR);
1251 set_exception_intercept(svm, AC_VECTOR);
1252 set_exception_intercept(svm, DB_VECTOR);
1254 * Guest access to VMware backdoor ports could legitimately
1255 * trigger #GP because of TSS I/O permission bitmap.
1256 * We intercept those #GP and allow access to them anyway
1259 if (enable_vmware_backdoor)
1260 set_exception_intercept(svm, GP_VECTOR);
1262 svm_set_intercept(svm, INTERCEPT_INTR);
1263 svm_set_intercept(svm, INTERCEPT_NMI);
1266 svm_set_intercept(svm, INTERCEPT_SMI);
1268 svm_set_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1269 svm_set_intercept(svm, INTERCEPT_RDPMC);
1270 svm_set_intercept(svm, INTERCEPT_CPUID);
1271 svm_set_intercept(svm, INTERCEPT_INVD);
1272 svm_set_intercept(svm, INTERCEPT_INVLPG);
1273 svm_set_intercept(svm, INTERCEPT_INVLPGA);
1274 svm_set_intercept(svm, INTERCEPT_IOIO_PROT);
1275 svm_set_intercept(svm, INTERCEPT_MSR_PROT);
1276 svm_set_intercept(svm, INTERCEPT_TASK_SWITCH);
1277 svm_set_intercept(svm, INTERCEPT_SHUTDOWN);
1278 svm_set_intercept(svm, INTERCEPT_VMRUN);
1279 svm_set_intercept(svm, INTERCEPT_VMMCALL);
1280 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1281 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1282 svm_set_intercept(svm, INTERCEPT_STGI);
1283 svm_set_intercept(svm, INTERCEPT_CLGI);
1284 svm_set_intercept(svm, INTERCEPT_SKINIT);
1285 svm_set_intercept(svm, INTERCEPT_WBINVD);
1286 svm_set_intercept(svm, INTERCEPT_XSETBV);
1287 svm_set_intercept(svm, INTERCEPT_RDPRU);
1288 svm_set_intercept(svm, INTERCEPT_RSM);
1290 if (!kvm_mwait_in_guest(vcpu->kvm)) {
1291 svm_set_intercept(svm, INTERCEPT_MONITOR);
1292 svm_set_intercept(svm, INTERCEPT_MWAIT);
1295 if (!kvm_hlt_in_guest(vcpu->kvm))
1296 svm_set_intercept(svm, INTERCEPT_HLT);
1298 control->iopm_base_pa = __sme_set(iopm_base);
1299 control->msrpm_base_pa = __sme_set(__pa(svm->msrpm));
1300 control->int_ctl = V_INTR_MASKING_MASK;
1302 init_seg(&save->es);
1303 init_seg(&save->ss);
1304 init_seg(&save->ds);
1305 init_seg(&save->fs);
1306 init_seg(&save->gs);
1308 save->cs.selector = 0xf000;
1309 save->cs.base = 0xffff0000;
1310 /* Executable/Readable Code Segment */
1311 save->cs.attrib = SVM_SELECTOR_READ_MASK | SVM_SELECTOR_P_MASK |
1312 SVM_SELECTOR_S_MASK | SVM_SELECTOR_CODE_MASK;
1313 save->cs.limit = 0xffff;
1315 save->gdtr.base = 0;
1316 save->gdtr.limit = 0xffff;
1317 save->idtr.base = 0;
1318 save->idtr.limit = 0xffff;
1320 init_sys_seg(&save->ldtr, SEG_TYPE_LDT);
1321 init_sys_seg(&save->tr, SEG_TYPE_BUSY_TSS16);
1324 /* Setup VMCB for Nested Paging */
1325 control->nested_ctl |= SVM_NESTED_CTL_NP_ENABLE;
1326 svm_clr_intercept(svm, INTERCEPT_INVLPG);
1327 clr_exception_intercept(svm, PF_VECTOR);
1328 svm_clr_intercept(svm, INTERCEPT_CR3_READ);
1329 svm_clr_intercept(svm, INTERCEPT_CR3_WRITE);
1330 save->g_pat = vcpu->arch.pat;
1333 svm->current_vmcb->asid_generation = 0;
1336 svm->nested.vmcb12_gpa = INVALID_GPA;
1337 svm->nested.last_vmcb12_gpa = INVALID_GPA;
1339 if (!kvm_pause_in_guest(vcpu->kvm)) {
1340 control->pause_filter_count = pause_filter_count;
1341 if (pause_filter_thresh)
1342 control->pause_filter_thresh = pause_filter_thresh;
1343 svm_set_intercept(svm, INTERCEPT_PAUSE);
1345 svm_clr_intercept(svm, INTERCEPT_PAUSE);
1348 svm_recalc_instruction_intercepts(vcpu, svm);
1351 * If the host supports V_SPEC_CTRL then disable the interception
1352 * of MSR_IA32_SPEC_CTRL.
1354 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
1355 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
1357 if (kvm_vcpu_apicv_active(vcpu))
1358 avic_init_vmcb(svm, vmcb);
1361 svm->vmcb->control.int_ctl |= V_NMI_ENABLE_MASK;
1364 svm_clr_intercept(svm, INTERCEPT_STGI);
1365 svm_clr_intercept(svm, INTERCEPT_CLGI);
1366 svm->vmcb->control.int_ctl |= V_GIF_ENABLE_MASK;
1369 if (sev_guest(vcpu->kvm))
1372 svm_hv_init_vmcb(vmcb);
1373 init_vmcb_after_set_cpuid(vcpu);
1375 vmcb_mark_all_dirty(vmcb);
1380 static void __svm_vcpu_reset(struct kvm_vcpu *vcpu)
1382 struct vcpu_svm *svm = to_svm(vcpu);
1384 svm_vcpu_init_msrpm(vcpu, svm->msrpm);
1386 svm_init_osvw(vcpu);
1387 vcpu->arch.microcode_version = 0x01000065;
1388 svm->tsc_ratio_msr = kvm_caps.default_tsc_scaling_ratio;
1390 svm->nmi_masked = false;
1391 svm->awaiting_iret_completion = false;
1393 if (sev_es_guest(vcpu->kvm))
1394 sev_es_vcpu_reset(svm);
1397 static void svm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
1399 struct vcpu_svm *svm = to_svm(vcpu);
1402 svm->virt_spec_ctrl = 0;
1407 __svm_vcpu_reset(vcpu);
1410 void svm_switch_vmcb(struct vcpu_svm *svm, struct kvm_vmcb_info *target_vmcb)
1412 svm->current_vmcb = target_vmcb;
1413 svm->vmcb = target_vmcb->ptr;
1416 static int svm_vcpu_create(struct kvm_vcpu *vcpu)
1418 struct vcpu_svm *svm;
1419 struct page *vmcb01_page;
1420 struct page *vmsa_page = NULL;
1423 BUILD_BUG_ON(offsetof(struct vcpu_svm, vcpu) != 0);
1427 vmcb01_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1431 if (sev_es_guest(vcpu->kvm)) {
1433 * SEV-ES guests require a separate VMSA page used to contain
1434 * the encrypted register state of the guest.
1436 vmsa_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1438 goto error_free_vmcb_page;
1441 * SEV-ES guests maintain an encrypted version of their FPU
1442 * state which is restored and saved on VMRUN and VMEXIT.
1443 * Mark vcpu->arch.guest_fpu->fpstate as scratch so it won't
1444 * do xsave/xrstor on it.
1446 fpstate_set_confidential(&vcpu->arch.guest_fpu);
1449 err = avic_init_vcpu(svm);
1451 goto error_free_vmsa_page;
1453 svm->msrpm = svm_vcpu_alloc_msrpm();
1456 goto error_free_vmsa_page;
1459 svm->x2avic_msrs_intercepted = true;
1461 svm->vmcb01.ptr = page_address(vmcb01_page);
1462 svm->vmcb01.pa = __sme_set(page_to_pfn(vmcb01_page) << PAGE_SHIFT);
1463 svm_switch_vmcb(svm, &svm->vmcb01);
1466 svm->sev_es.vmsa = page_address(vmsa_page);
1468 svm->guest_state_loaded = false;
1472 error_free_vmsa_page:
1474 __free_page(vmsa_page);
1475 error_free_vmcb_page:
1476 __free_page(vmcb01_page);
1481 static void svm_clear_current_vmcb(struct vmcb *vmcb)
1485 for_each_online_cpu(i)
1486 cmpxchg(per_cpu_ptr(&svm_data.current_vmcb, i), vmcb, NULL);
1489 static void svm_vcpu_free(struct kvm_vcpu *vcpu)
1491 struct vcpu_svm *svm = to_svm(vcpu);
1494 * The vmcb page can be recycled, causing a false negative in
1495 * svm_vcpu_load(). So, ensure that no logical CPU has this
1496 * vmcb page recorded as its current vmcb.
1498 svm_clear_current_vmcb(svm->vmcb);
1500 svm_leave_nested(vcpu);
1501 svm_free_nested(svm);
1503 sev_free_vcpu(vcpu);
1505 __free_page(pfn_to_page(__sme_clr(svm->vmcb01.pa) >> PAGE_SHIFT));
1506 __free_pages(virt_to_page(svm->msrpm), get_order(MSRPM_SIZE));
1509 static void svm_prepare_switch_to_guest(struct kvm_vcpu *vcpu)
1511 struct vcpu_svm *svm = to_svm(vcpu);
1512 struct svm_cpu_data *sd = per_cpu_ptr(&svm_data, vcpu->cpu);
1514 if (sev_es_guest(vcpu->kvm))
1515 sev_es_unmap_ghcb(svm);
1517 if (svm->guest_state_loaded)
1521 * Save additional host state that will be restored on VMEXIT (sev-es)
1522 * or subsequent vmload of host save area.
1524 vmsave(sd->save_area_pa);
1525 if (sev_es_guest(vcpu->kvm)) {
1526 struct sev_es_save_area *hostsa;
1527 hostsa = (struct sev_es_save_area *)(page_address(sd->save_area) + 0x400);
1529 sev_es_prepare_switch_to_guest(hostsa);
1533 __svm_write_tsc_multiplier(vcpu->arch.tsc_scaling_ratio);
1535 if (likely(tsc_aux_uret_slot >= 0))
1536 kvm_set_user_return_msr(tsc_aux_uret_slot, svm->tsc_aux, -1ull);
1538 svm->guest_state_loaded = true;
1541 static void svm_prepare_host_switch(struct kvm_vcpu *vcpu)
1543 to_svm(vcpu)->guest_state_loaded = false;
1546 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1548 struct vcpu_svm *svm = to_svm(vcpu);
1549 struct svm_cpu_data *sd = per_cpu_ptr(&svm_data, cpu);
1551 if (sd->current_vmcb != svm->vmcb) {
1552 sd->current_vmcb = svm->vmcb;
1554 if (!cpu_feature_enabled(X86_FEATURE_IBPB_ON_VMEXIT))
1555 indirect_branch_prediction_barrier();
1557 if (kvm_vcpu_apicv_active(vcpu))
1558 avic_vcpu_load(vcpu, cpu);
1561 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
1563 if (kvm_vcpu_apicv_active(vcpu))
1564 avic_vcpu_put(vcpu);
1566 svm_prepare_host_switch(vcpu);
1568 ++vcpu->stat.host_state_reload;
1571 static unsigned long svm_get_rflags(struct kvm_vcpu *vcpu)
1573 struct vcpu_svm *svm = to_svm(vcpu);
1574 unsigned long rflags = svm->vmcb->save.rflags;
1576 if (svm->nmi_singlestep) {
1577 /* Hide our flags if they were not set by the guest */
1578 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
1579 rflags &= ~X86_EFLAGS_TF;
1580 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
1581 rflags &= ~X86_EFLAGS_RF;
1586 static void svm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
1588 if (to_svm(vcpu)->nmi_singlestep)
1589 rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
1592 * Any change of EFLAGS.VM is accompanied by a reload of SS
1593 * (caused by either a task switch or an inter-privilege IRET),
1594 * so we do not need to update the CPL here.
1596 to_svm(vcpu)->vmcb->save.rflags = rflags;
1599 static bool svm_get_if_flag(struct kvm_vcpu *vcpu)
1601 struct vmcb *vmcb = to_svm(vcpu)->vmcb;
1603 return sev_es_guest(vcpu->kvm)
1604 ? vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK
1605 : kvm_get_rflags(vcpu) & X86_EFLAGS_IF;
1608 static void svm_cache_reg(struct kvm_vcpu *vcpu, enum kvm_reg reg)
1610 kvm_register_mark_available(vcpu, reg);
1613 case VCPU_EXREG_PDPTR:
1615 * When !npt_enabled, mmu->pdptrs[] is already available since
1616 * it is always updated per SDM when moving to CRs.
1619 load_pdptrs(vcpu, kvm_read_cr3(vcpu));
1622 KVM_BUG_ON(1, vcpu->kvm);
1626 static void svm_set_vintr(struct vcpu_svm *svm)
1628 struct vmcb_control_area *control;
1631 * The following fields are ignored when AVIC is enabled
1633 WARN_ON(kvm_vcpu_apicv_activated(&svm->vcpu));
1635 svm_set_intercept(svm, INTERCEPT_VINTR);
1638 * Recalculating intercepts may have cleared the VINTR intercept. If
1639 * V_INTR_MASKING is enabled in vmcb12, then the effective RFLAGS.IF
1640 * for L1 physical interrupts is L1's RFLAGS.IF at the time of VMRUN.
1641 * Requesting an interrupt window if save.RFLAGS.IF=0 is pointless as
1642 * interrupts will never be unblocked while L2 is running.
1644 if (!svm_is_intercept(svm, INTERCEPT_VINTR))
1648 * This is just a dummy VINTR to actually cause a vmexit to happen.
1649 * Actual injection of virtual interrupts happens through EVENTINJ.
1651 control = &svm->vmcb->control;
1652 control->int_vector = 0x0;
1653 control->int_ctl &= ~V_INTR_PRIO_MASK;
1654 control->int_ctl |= V_IRQ_MASK |
1655 ((/*control->int_vector >> 4*/ 0xf) << V_INTR_PRIO_SHIFT);
1656 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1659 static void svm_clear_vintr(struct vcpu_svm *svm)
1661 svm_clr_intercept(svm, INTERCEPT_VINTR);
1663 /* Drop int_ctl fields related to VINTR injection. */
1664 svm->vmcb->control.int_ctl &= ~V_IRQ_INJECTION_BITS_MASK;
1665 if (is_guest_mode(&svm->vcpu)) {
1666 svm->vmcb01.ptr->control.int_ctl &= ~V_IRQ_INJECTION_BITS_MASK;
1668 WARN_ON((svm->vmcb->control.int_ctl & V_TPR_MASK) !=
1669 (svm->nested.ctl.int_ctl & V_TPR_MASK));
1671 svm->vmcb->control.int_ctl |= svm->nested.ctl.int_ctl &
1672 V_IRQ_INJECTION_BITS_MASK;
1674 svm->vmcb->control.int_vector = svm->nested.ctl.int_vector;
1677 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1680 static struct vmcb_seg *svm_seg(struct kvm_vcpu *vcpu, int seg)
1682 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1683 struct vmcb_save_area *save01 = &to_svm(vcpu)->vmcb01.ptr->save;
1686 case VCPU_SREG_CS: return &save->cs;
1687 case VCPU_SREG_DS: return &save->ds;
1688 case VCPU_SREG_ES: return &save->es;
1689 case VCPU_SREG_FS: return &save01->fs;
1690 case VCPU_SREG_GS: return &save01->gs;
1691 case VCPU_SREG_SS: return &save->ss;
1692 case VCPU_SREG_TR: return &save01->tr;
1693 case VCPU_SREG_LDTR: return &save01->ldtr;
1699 static u64 svm_get_segment_base(struct kvm_vcpu *vcpu, int seg)
1701 struct vmcb_seg *s = svm_seg(vcpu, seg);
1706 static void svm_get_segment(struct kvm_vcpu *vcpu,
1707 struct kvm_segment *var, int seg)
1709 struct vmcb_seg *s = svm_seg(vcpu, seg);
1711 var->base = s->base;
1712 var->limit = s->limit;
1713 var->selector = s->selector;
1714 var->type = s->attrib & SVM_SELECTOR_TYPE_MASK;
1715 var->s = (s->attrib >> SVM_SELECTOR_S_SHIFT) & 1;
1716 var->dpl = (s->attrib >> SVM_SELECTOR_DPL_SHIFT) & 3;
1717 var->present = (s->attrib >> SVM_SELECTOR_P_SHIFT) & 1;
1718 var->avl = (s->attrib >> SVM_SELECTOR_AVL_SHIFT) & 1;
1719 var->l = (s->attrib >> SVM_SELECTOR_L_SHIFT) & 1;
1720 var->db = (s->attrib >> SVM_SELECTOR_DB_SHIFT) & 1;
1723 * AMD CPUs circa 2014 track the G bit for all segments except CS.
1724 * However, the SVM spec states that the G bit is not observed by the
1725 * CPU, and some VMware virtual CPUs drop the G bit for all segments.
1726 * So let's synthesize a legal G bit for all segments, this helps
1727 * running KVM nested. It also helps cross-vendor migration, because
1728 * Intel's vmentry has a check on the 'G' bit.
1730 var->g = s->limit > 0xfffff;
1733 * AMD's VMCB does not have an explicit unusable field, so emulate it
1734 * for cross vendor migration purposes by "not present"
1736 var->unusable = !var->present;
1741 * Work around a bug where the busy flag in the tr selector
1751 * The accessed bit must always be set in the segment
1752 * descriptor cache, although it can be cleared in the
1753 * descriptor, the cached bit always remains at 1. Since
1754 * Intel has a check on this, set it here to support
1755 * cross-vendor migration.
1762 * On AMD CPUs sometimes the DB bit in the segment
1763 * descriptor is left as 1, although the whole segment has
1764 * been made unusable. Clear it here to pass an Intel VMX
1765 * entry check when cross vendor migrating.
1769 /* This is symmetric with svm_set_segment() */
1770 var->dpl = to_svm(vcpu)->vmcb->save.cpl;
1775 static int svm_get_cpl(struct kvm_vcpu *vcpu)
1777 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1782 static void svm_get_cs_db_l_bits(struct kvm_vcpu *vcpu, int *db, int *l)
1784 struct kvm_segment cs;
1786 svm_get_segment(vcpu, &cs, VCPU_SREG_CS);
1791 static void svm_get_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1793 struct vcpu_svm *svm = to_svm(vcpu);
1795 dt->size = svm->vmcb->save.idtr.limit;
1796 dt->address = svm->vmcb->save.idtr.base;
1799 static void svm_set_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1801 struct vcpu_svm *svm = to_svm(vcpu);
1803 svm->vmcb->save.idtr.limit = dt->size;
1804 svm->vmcb->save.idtr.base = dt->address ;
1805 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1808 static void svm_get_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1810 struct vcpu_svm *svm = to_svm(vcpu);
1812 dt->size = svm->vmcb->save.gdtr.limit;
1813 dt->address = svm->vmcb->save.gdtr.base;
1816 static void svm_set_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1818 struct vcpu_svm *svm = to_svm(vcpu);
1820 svm->vmcb->save.gdtr.limit = dt->size;
1821 svm->vmcb->save.gdtr.base = dt->address ;
1822 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1825 static void sev_post_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3)
1827 struct vcpu_svm *svm = to_svm(vcpu);
1830 * For guests that don't set guest_state_protected, the cr3 update is
1831 * handled via kvm_mmu_load() while entering the guest. For guests
1832 * that do (SEV-ES/SEV-SNP), the cr3 update needs to be written to
1833 * VMCB save area now, since the save area will become the initial
1834 * contents of the VMSA, and future VMCB save area updates won't be
1837 if (sev_es_guest(vcpu->kvm)) {
1838 svm->vmcb->save.cr3 = cr3;
1839 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1843 static bool svm_is_valid_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1848 void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1850 struct vcpu_svm *svm = to_svm(vcpu);
1852 bool old_paging = is_paging(vcpu);
1854 #ifdef CONFIG_X86_64
1855 if (vcpu->arch.efer & EFER_LME && !vcpu->arch.guest_state_protected) {
1856 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
1857 vcpu->arch.efer |= EFER_LMA;
1858 svm->vmcb->save.efer |= EFER_LMA | EFER_LME;
1861 if (is_paging(vcpu) && !(cr0 & X86_CR0_PG)) {
1862 vcpu->arch.efer &= ~EFER_LMA;
1863 svm->vmcb->save.efer &= ~(EFER_LMA | EFER_LME);
1867 vcpu->arch.cr0 = cr0;
1870 hcr0 |= X86_CR0_PG | X86_CR0_WP;
1871 if (old_paging != is_paging(vcpu))
1872 svm_set_cr4(vcpu, kvm_read_cr4(vcpu));
1876 * re-enable caching here because the QEMU bios
1877 * does not do it - this results in some delay at
1880 if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
1881 hcr0 &= ~(X86_CR0_CD | X86_CR0_NW);
1883 svm->vmcb->save.cr0 = hcr0;
1884 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1887 * SEV-ES guests must always keep the CR intercepts cleared. CR
1888 * tracking is done using the CR write traps.
1890 if (sev_es_guest(vcpu->kvm))
1894 /* Selective CR0 write remains on. */
1895 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1896 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1898 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1899 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1903 static bool svm_is_valid_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1908 void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1910 unsigned long host_cr4_mce = cr4_read_shadow() & X86_CR4_MCE;
1911 unsigned long old_cr4 = vcpu->arch.cr4;
1913 if (npt_enabled && ((old_cr4 ^ cr4) & X86_CR4_PGE))
1914 svm_flush_tlb_current(vcpu);
1916 vcpu->arch.cr4 = cr4;
1920 if (!is_paging(vcpu))
1921 cr4 &= ~(X86_CR4_SMEP | X86_CR4_SMAP | X86_CR4_PKE);
1923 cr4 |= host_cr4_mce;
1924 to_svm(vcpu)->vmcb->save.cr4 = cr4;
1925 vmcb_mark_dirty(to_svm(vcpu)->vmcb, VMCB_CR);
1927 if ((cr4 ^ old_cr4) & (X86_CR4_OSXSAVE | X86_CR4_PKE))
1928 kvm_update_cpuid_runtime(vcpu);
1931 static void svm_set_segment(struct kvm_vcpu *vcpu,
1932 struct kvm_segment *var, int seg)
1934 struct vcpu_svm *svm = to_svm(vcpu);
1935 struct vmcb_seg *s = svm_seg(vcpu, seg);
1937 s->base = var->base;
1938 s->limit = var->limit;
1939 s->selector = var->selector;
1940 s->attrib = (var->type & SVM_SELECTOR_TYPE_MASK);
1941 s->attrib |= (var->s & 1) << SVM_SELECTOR_S_SHIFT;
1942 s->attrib |= (var->dpl & 3) << SVM_SELECTOR_DPL_SHIFT;
1943 s->attrib |= ((var->present & 1) && !var->unusable) << SVM_SELECTOR_P_SHIFT;
1944 s->attrib |= (var->avl & 1) << SVM_SELECTOR_AVL_SHIFT;
1945 s->attrib |= (var->l & 1) << SVM_SELECTOR_L_SHIFT;
1946 s->attrib |= (var->db & 1) << SVM_SELECTOR_DB_SHIFT;
1947 s->attrib |= (var->g & 1) << SVM_SELECTOR_G_SHIFT;
1950 * This is always accurate, except if SYSRET returned to a segment
1951 * with SS.DPL != 3. Intel does not have this quirk, and always
1952 * forces SS.DPL to 3 on sysret, so we ignore that case; fixing it
1953 * would entail passing the CPL to userspace and back.
1955 if (seg == VCPU_SREG_SS)
1956 /* This is symmetric with svm_get_segment() */
1957 svm->vmcb->save.cpl = (var->dpl & 3);
1959 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
1962 static void svm_update_exception_bitmap(struct kvm_vcpu *vcpu)
1964 struct vcpu_svm *svm = to_svm(vcpu);
1966 clr_exception_intercept(svm, BP_VECTOR);
1968 if (vcpu->guest_debug & KVM_GUESTDBG_ENABLE) {
1969 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP)
1970 set_exception_intercept(svm, BP_VECTOR);
1974 static void new_asid(struct vcpu_svm *svm, struct svm_cpu_data *sd)
1976 if (sd->next_asid > sd->max_asid) {
1977 ++sd->asid_generation;
1978 sd->next_asid = sd->min_asid;
1979 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
1980 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1983 svm->current_vmcb->asid_generation = sd->asid_generation;
1984 svm->asid = sd->next_asid++;
1987 static void svm_set_dr6(struct vcpu_svm *svm, unsigned long value)
1989 struct vmcb *vmcb = svm->vmcb;
1991 if (svm->vcpu.arch.guest_state_protected)
1994 if (unlikely(value != vmcb->save.dr6)) {
1995 vmcb->save.dr6 = value;
1996 vmcb_mark_dirty(vmcb, VMCB_DR);
2000 static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
2002 struct vcpu_svm *svm = to_svm(vcpu);
2004 if (WARN_ON_ONCE(sev_es_guest(vcpu->kvm)))
2007 get_debugreg(vcpu->arch.db[0], 0);
2008 get_debugreg(vcpu->arch.db[1], 1);
2009 get_debugreg(vcpu->arch.db[2], 2);
2010 get_debugreg(vcpu->arch.db[3], 3);
2012 * We cannot reset svm->vmcb->save.dr6 to DR6_ACTIVE_LOW here,
2013 * because db_interception might need it. We can do it before vmentry.
2015 vcpu->arch.dr6 = svm->vmcb->save.dr6;
2016 vcpu->arch.dr7 = svm->vmcb->save.dr7;
2017 vcpu->arch.switch_db_regs &= ~KVM_DEBUGREG_WONT_EXIT;
2018 set_dr_intercepts(svm);
2021 static void svm_set_dr7(struct kvm_vcpu *vcpu, unsigned long value)
2023 struct vcpu_svm *svm = to_svm(vcpu);
2025 if (vcpu->arch.guest_state_protected)
2028 svm->vmcb->save.dr7 = value;
2029 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
2032 static int pf_interception(struct kvm_vcpu *vcpu)
2034 struct vcpu_svm *svm = to_svm(vcpu);
2036 u64 fault_address = svm->vmcb->control.exit_info_2;
2037 u64 error_code = svm->vmcb->control.exit_info_1;
2039 return kvm_handle_page_fault(vcpu, error_code, fault_address,
2040 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
2041 svm->vmcb->control.insn_bytes : NULL,
2042 svm->vmcb->control.insn_len);
2045 static int npf_interception(struct kvm_vcpu *vcpu)
2047 struct vcpu_svm *svm = to_svm(vcpu);
2049 u64 fault_address = svm->vmcb->control.exit_info_2;
2050 u64 error_code = svm->vmcb->control.exit_info_1;
2052 trace_kvm_page_fault(vcpu, fault_address, error_code);
2053 return kvm_mmu_page_fault(vcpu, fault_address, error_code,
2054 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
2055 svm->vmcb->control.insn_bytes : NULL,
2056 svm->vmcb->control.insn_len);
2059 static int db_interception(struct kvm_vcpu *vcpu)
2061 struct kvm_run *kvm_run = vcpu->run;
2062 struct vcpu_svm *svm = to_svm(vcpu);
2064 if (!(vcpu->guest_debug &
2065 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) &&
2066 !svm->nmi_singlestep) {
2067 u32 payload = svm->vmcb->save.dr6 ^ DR6_ACTIVE_LOW;
2068 kvm_queue_exception_p(vcpu, DB_VECTOR, payload);
2072 if (svm->nmi_singlestep) {
2073 disable_nmi_singlestep(svm);
2074 /* Make sure we check for pending NMIs upon entry */
2075 kvm_make_request(KVM_REQ_EVENT, vcpu);
2078 if (vcpu->guest_debug &
2079 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) {
2080 kvm_run->exit_reason = KVM_EXIT_DEBUG;
2081 kvm_run->debug.arch.dr6 = svm->vmcb->save.dr6;
2082 kvm_run->debug.arch.dr7 = svm->vmcb->save.dr7;
2083 kvm_run->debug.arch.pc =
2084 svm->vmcb->save.cs.base + svm->vmcb->save.rip;
2085 kvm_run->debug.arch.exception = DB_VECTOR;
2092 static int bp_interception(struct kvm_vcpu *vcpu)
2094 struct vcpu_svm *svm = to_svm(vcpu);
2095 struct kvm_run *kvm_run = vcpu->run;
2097 kvm_run->exit_reason = KVM_EXIT_DEBUG;
2098 kvm_run->debug.arch.pc = svm->vmcb->save.cs.base + svm->vmcb->save.rip;
2099 kvm_run->debug.arch.exception = BP_VECTOR;
2103 static int ud_interception(struct kvm_vcpu *vcpu)
2105 return handle_ud(vcpu);
2108 static int ac_interception(struct kvm_vcpu *vcpu)
2110 kvm_queue_exception_e(vcpu, AC_VECTOR, 0);
2114 static bool is_erratum_383(void)
2119 if (!erratum_383_found)
2122 value = native_read_msr_safe(MSR_IA32_MC0_STATUS, &err);
2126 /* Bit 62 may or may not be set for this mce */
2127 value &= ~(1ULL << 62);
2129 if (value != 0xb600000000010015ULL)
2132 /* Clear MCi_STATUS registers */
2133 for (i = 0; i < 6; ++i)
2134 native_write_msr_safe(MSR_IA32_MCx_STATUS(i), 0, 0);
2136 value = native_read_msr_safe(MSR_IA32_MCG_STATUS, &err);
2140 value &= ~(1ULL << 2);
2141 low = lower_32_bits(value);
2142 high = upper_32_bits(value);
2144 native_write_msr_safe(MSR_IA32_MCG_STATUS, low, high);
2147 /* Flush tlb to evict multi-match entries */
2153 static void svm_handle_mce(struct kvm_vcpu *vcpu)
2155 if (is_erratum_383()) {
2157 * Erratum 383 triggered. Guest state is corrupt so kill the
2160 pr_err("Guest triggered AMD Erratum 383\n");
2162 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2168 * On an #MC intercept the MCE handler is not called automatically in
2169 * the host. So do it by hand here.
2171 kvm_machine_check();
2174 static int mc_interception(struct kvm_vcpu *vcpu)
2179 static int shutdown_interception(struct kvm_vcpu *vcpu)
2181 struct kvm_run *kvm_run = vcpu->run;
2182 struct vcpu_svm *svm = to_svm(vcpu);
2185 * The VM save area has already been encrypted so it
2186 * cannot be reinitialized - just terminate.
2188 if (sev_es_guest(vcpu->kvm))
2192 * VMCB is undefined after a SHUTDOWN intercept. INIT the vCPU to put
2193 * the VMCB in a known good state. Unfortuately, KVM doesn't have
2194 * KVM_MP_STATE_SHUTDOWN and can't add it without potentially breaking
2195 * userspace. At a platform view, INIT is acceptable behavior as
2196 * there exist bare metal platforms that automatically INIT the CPU
2197 * in response to shutdown.
2199 clear_page(svm->vmcb);
2200 kvm_vcpu_reset(vcpu, true);
2202 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
2206 static int io_interception(struct kvm_vcpu *vcpu)
2208 struct vcpu_svm *svm = to_svm(vcpu);
2209 u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
2210 int size, in, string;
2213 ++vcpu->stat.io_exits;
2214 string = (io_info & SVM_IOIO_STR_MASK) != 0;
2215 in = (io_info & SVM_IOIO_TYPE_MASK) != 0;
2216 port = io_info >> 16;
2217 size = (io_info & SVM_IOIO_SIZE_MASK) >> SVM_IOIO_SIZE_SHIFT;
2220 if (sev_es_guest(vcpu->kvm))
2221 return sev_es_string_io(svm, size, port, in);
2223 return kvm_emulate_instruction(vcpu, 0);
2226 svm->next_rip = svm->vmcb->control.exit_info_2;
2228 return kvm_fast_pio(vcpu, size, port, in);
2231 static int nmi_interception(struct kvm_vcpu *vcpu)
2236 static int smi_interception(struct kvm_vcpu *vcpu)
2241 static int intr_interception(struct kvm_vcpu *vcpu)
2243 ++vcpu->stat.irq_exits;
2247 static int vmload_vmsave_interception(struct kvm_vcpu *vcpu, bool vmload)
2249 struct vcpu_svm *svm = to_svm(vcpu);
2250 struct vmcb *vmcb12;
2251 struct kvm_host_map map;
2254 if (nested_svm_check_permissions(vcpu))
2257 ret = kvm_vcpu_map(vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2260 kvm_inject_gp(vcpu, 0);
2266 ret = kvm_skip_emulated_instruction(vcpu);
2269 svm_copy_vmloadsave_state(svm->vmcb, vmcb12);
2270 svm->sysenter_eip_hi = 0;
2271 svm->sysenter_esp_hi = 0;
2273 svm_copy_vmloadsave_state(vmcb12, svm->vmcb);
2276 kvm_vcpu_unmap(vcpu, &map, true);
2281 static int vmload_interception(struct kvm_vcpu *vcpu)
2283 return vmload_vmsave_interception(vcpu, true);
2286 static int vmsave_interception(struct kvm_vcpu *vcpu)
2288 return vmload_vmsave_interception(vcpu, false);
2291 static int vmrun_interception(struct kvm_vcpu *vcpu)
2293 if (nested_svm_check_permissions(vcpu))
2296 return nested_svm_vmrun(vcpu);
2306 /* Return NONE_SVM_INSTR if not SVM instrs, otherwise return decode result */
2307 static int svm_instr_opcode(struct kvm_vcpu *vcpu)
2309 struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt;
2311 if (ctxt->b != 0x1 || ctxt->opcode_len != 2)
2312 return NONE_SVM_INSTR;
2314 switch (ctxt->modrm) {
2315 case 0xd8: /* VMRUN */
2316 return SVM_INSTR_VMRUN;
2317 case 0xda: /* VMLOAD */
2318 return SVM_INSTR_VMLOAD;
2319 case 0xdb: /* VMSAVE */
2320 return SVM_INSTR_VMSAVE;
2325 return NONE_SVM_INSTR;
2328 static int emulate_svm_instr(struct kvm_vcpu *vcpu, int opcode)
2330 const int guest_mode_exit_codes[] = {
2331 [SVM_INSTR_VMRUN] = SVM_EXIT_VMRUN,
2332 [SVM_INSTR_VMLOAD] = SVM_EXIT_VMLOAD,
2333 [SVM_INSTR_VMSAVE] = SVM_EXIT_VMSAVE,
2335 int (*const svm_instr_handlers[])(struct kvm_vcpu *vcpu) = {
2336 [SVM_INSTR_VMRUN] = vmrun_interception,
2337 [SVM_INSTR_VMLOAD] = vmload_interception,
2338 [SVM_INSTR_VMSAVE] = vmsave_interception,
2340 struct vcpu_svm *svm = to_svm(vcpu);
2343 if (is_guest_mode(vcpu)) {
2344 /* Returns '1' or -errno on failure, '0' on success. */
2345 ret = nested_svm_simple_vmexit(svm, guest_mode_exit_codes[opcode]);
2350 return svm_instr_handlers[opcode](vcpu);
2354 * #GP handling code. Note that #GP can be triggered under the following two
2356 * 1) SVM VM-related instructions (VMRUN/VMSAVE/VMLOAD) that trigger #GP on
2357 * some AMD CPUs when EAX of these instructions are in the reserved memory
2358 * regions (e.g. SMM memory on host).
2359 * 2) VMware backdoor
2361 static int gp_interception(struct kvm_vcpu *vcpu)
2363 struct vcpu_svm *svm = to_svm(vcpu);
2364 u32 error_code = svm->vmcb->control.exit_info_1;
2367 /* Both #GP cases have zero error_code */
2371 /* Decode the instruction for usage later */
2372 if (x86_decode_emulated_instruction(vcpu, 0, NULL, 0) != EMULATION_OK)
2375 opcode = svm_instr_opcode(vcpu);
2377 if (opcode == NONE_SVM_INSTR) {
2378 if (!enable_vmware_backdoor)
2382 * VMware backdoor emulation on #GP interception only handles
2383 * IN{S}, OUT{S}, and RDPMC.
2385 if (!is_guest_mode(vcpu))
2386 return kvm_emulate_instruction(vcpu,
2387 EMULTYPE_VMWARE_GP | EMULTYPE_NO_DECODE);
2389 /* All SVM instructions expect page aligned RAX */
2390 if (svm->vmcb->save.rax & ~PAGE_MASK)
2393 return emulate_svm_instr(vcpu, opcode);
2397 kvm_queue_exception_e(vcpu, GP_VECTOR, error_code);
2401 void svm_set_gif(struct vcpu_svm *svm, bool value)
2405 * If VGIF is enabled, the STGI intercept is only added to
2406 * detect the opening of the SMI/NMI window; remove it now.
2407 * Likewise, clear the VINTR intercept, we will set it
2408 * again while processing KVM_REQ_EVENT if needed.
2411 svm_clr_intercept(svm, INTERCEPT_STGI);
2412 if (svm_is_intercept(svm, INTERCEPT_VINTR))
2413 svm_clear_vintr(svm);
2416 if (svm->vcpu.arch.smi_pending ||
2417 svm->vcpu.arch.nmi_pending ||
2418 kvm_cpu_has_injectable_intr(&svm->vcpu) ||
2419 kvm_apic_has_pending_init_or_sipi(&svm->vcpu))
2420 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2425 * After a CLGI no interrupts should come. But if vGIF is
2426 * in use, we still rely on the VINTR intercept (rather than
2427 * STGI) to detect an open interrupt window.
2430 svm_clear_vintr(svm);
2434 static int stgi_interception(struct kvm_vcpu *vcpu)
2438 if (nested_svm_check_permissions(vcpu))
2441 ret = kvm_skip_emulated_instruction(vcpu);
2442 svm_set_gif(to_svm(vcpu), true);
2446 static int clgi_interception(struct kvm_vcpu *vcpu)
2450 if (nested_svm_check_permissions(vcpu))
2453 ret = kvm_skip_emulated_instruction(vcpu);
2454 svm_set_gif(to_svm(vcpu), false);
2458 static int invlpga_interception(struct kvm_vcpu *vcpu)
2460 gva_t gva = kvm_rax_read(vcpu);
2461 u32 asid = kvm_rcx_read(vcpu);
2463 /* FIXME: Handle an address size prefix. */
2464 if (!is_long_mode(vcpu))
2467 trace_kvm_invlpga(to_svm(vcpu)->vmcb->save.rip, asid, gva);
2469 /* Let's treat INVLPGA the same as INVLPG (can be optimized!) */
2470 kvm_mmu_invlpg(vcpu, gva);
2472 return kvm_skip_emulated_instruction(vcpu);
2475 static int skinit_interception(struct kvm_vcpu *vcpu)
2477 trace_kvm_skinit(to_svm(vcpu)->vmcb->save.rip, kvm_rax_read(vcpu));
2479 kvm_queue_exception(vcpu, UD_VECTOR);
2483 static int task_switch_interception(struct kvm_vcpu *vcpu)
2485 struct vcpu_svm *svm = to_svm(vcpu);
2488 int int_type = svm->vmcb->control.exit_int_info &
2489 SVM_EXITINTINFO_TYPE_MASK;
2490 int int_vec = svm->vmcb->control.exit_int_info & SVM_EVTINJ_VEC_MASK;
2492 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_TYPE_MASK;
2494 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID;
2495 bool has_error_code = false;
2498 tss_selector = (u16)svm->vmcb->control.exit_info_1;
2500 if (svm->vmcb->control.exit_info_2 &
2501 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
2502 reason = TASK_SWITCH_IRET;
2503 else if (svm->vmcb->control.exit_info_2 &
2504 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
2505 reason = TASK_SWITCH_JMP;
2507 reason = TASK_SWITCH_GATE;
2509 reason = TASK_SWITCH_CALL;
2511 if (reason == TASK_SWITCH_GATE) {
2513 case SVM_EXITINTINFO_TYPE_NMI:
2514 vcpu->arch.nmi_injected = false;
2516 case SVM_EXITINTINFO_TYPE_EXEPT:
2517 if (svm->vmcb->control.exit_info_2 &
2518 (1ULL << SVM_EXITINFOSHIFT_TS_HAS_ERROR_CODE)) {
2519 has_error_code = true;
2521 (u32)svm->vmcb->control.exit_info_2;
2523 kvm_clear_exception_queue(vcpu);
2525 case SVM_EXITINTINFO_TYPE_INTR:
2526 case SVM_EXITINTINFO_TYPE_SOFT:
2527 kvm_clear_interrupt_queue(vcpu);
2534 if (reason != TASK_SWITCH_GATE ||
2535 int_type == SVM_EXITINTINFO_TYPE_SOFT ||
2536 (int_type == SVM_EXITINTINFO_TYPE_EXEPT &&
2537 (int_vec == OF_VECTOR || int_vec == BP_VECTOR))) {
2538 if (!svm_skip_emulated_instruction(vcpu))
2542 if (int_type != SVM_EXITINTINFO_TYPE_SOFT)
2545 return kvm_task_switch(vcpu, tss_selector, int_vec, reason,
2546 has_error_code, error_code);
2549 static void svm_clr_iret_intercept(struct vcpu_svm *svm)
2551 if (!sev_es_guest(svm->vcpu.kvm))
2552 svm_clr_intercept(svm, INTERCEPT_IRET);
2555 static void svm_set_iret_intercept(struct vcpu_svm *svm)
2557 if (!sev_es_guest(svm->vcpu.kvm))
2558 svm_set_intercept(svm, INTERCEPT_IRET);
2561 static int iret_interception(struct kvm_vcpu *vcpu)
2563 struct vcpu_svm *svm = to_svm(vcpu);
2565 WARN_ON_ONCE(sev_es_guest(vcpu->kvm));
2567 ++vcpu->stat.nmi_window_exits;
2568 svm->awaiting_iret_completion = true;
2570 svm_clr_iret_intercept(svm);
2571 svm->nmi_iret_rip = kvm_rip_read(vcpu);
2573 kvm_make_request(KVM_REQ_EVENT, vcpu);
2577 static int invlpg_interception(struct kvm_vcpu *vcpu)
2579 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2580 return kvm_emulate_instruction(vcpu, 0);
2582 kvm_mmu_invlpg(vcpu, to_svm(vcpu)->vmcb->control.exit_info_1);
2583 return kvm_skip_emulated_instruction(vcpu);
2586 static int emulate_on_interception(struct kvm_vcpu *vcpu)
2588 return kvm_emulate_instruction(vcpu, 0);
2591 static int rsm_interception(struct kvm_vcpu *vcpu)
2593 return kvm_emulate_instruction_from_buffer(vcpu, rsm_ins_bytes, 2);
2596 static bool check_selective_cr0_intercepted(struct kvm_vcpu *vcpu,
2599 struct vcpu_svm *svm = to_svm(vcpu);
2600 unsigned long cr0 = vcpu->arch.cr0;
2603 if (!is_guest_mode(vcpu) ||
2604 (!(vmcb12_is_intercept(&svm->nested.ctl, INTERCEPT_SELECTIVE_CR0))))
2607 cr0 &= ~SVM_CR0_SELECTIVE_MASK;
2608 val &= ~SVM_CR0_SELECTIVE_MASK;
2611 svm->vmcb->control.exit_code = SVM_EXIT_CR0_SEL_WRITE;
2612 ret = (nested_svm_exit_handled(svm) == NESTED_EXIT_DONE);
2618 #define CR_VALID (1ULL << 63)
2620 static int cr_interception(struct kvm_vcpu *vcpu)
2622 struct vcpu_svm *svm = to_svm(vcpu);
2627 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2628 return emulate_on_interception(vcpu);
2630 if (unlikely((svm->vmcb->control.exit_info_1 & CR_VALID) == 0))
2631 return emulate_on_interception(vcpu);
2633 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2634 if (svm->vmcb->control.exit_code == SVM_EXIT_CR0_SEL_WRITE)
2635 cr = SVM_EXIT_WRITE_CR0 - SVM_EXIT_READ_CR0;
2637 cr = svm->vmcb->control.exit_code - SVM_EXIT_READ_CR0;
2640 if (cr >= 16) { /* mov to cr */
2642 val = kvm_register_read(vcpu, reg);
2643 trace_kvm_cr_write(cr, val);
2646 if (!check_selective_cr0_intercepted(vcpu, val))
2647 err = kvm_set_cr0(vcpu, val);
2653 err = kvm_set_cr3(vcpu, val);
2656 err = kvm_set_cr4(vcpu, val);
2659 err = kvm_set_cr8(vcpu, val);
2662 WARN(1, "unhandled write to CR%d", cr);
2663 kvm_queue_exception(vcpu, UD_VECTOR);
2666 } else { /* mov from cr */
2669 val = kvm_read_cr0(vcpu);
2672 val = vcpu->arch.cr2;
2675 val = kvm_read_cr3(vcpu);
2678 val = kvm_read_cr4(vcpu);
2681 val = kvm_get_cr8(vcpu);
2684 WARN(1, "unhandled read from CR%d", cr);
2685 kvm_queue_exception(vcpu, UD_VECTOR);
2688 kvm_register_write(vcpu, reg, val);
2689 trace_kvm_cr_read(cr, val);
2691 return kvm_complete_insn_gp(vcpu, err);
2694 static int cr_trap(struct kvm_vcpu *vcpu)
2696 struct vcpu_svm *svm = to_svm(vcpu);
2697 unsigned long old_value, new_value;
2701 new_value = (unsigned long)svm->vmcb->control.exit_info_1;
2703 cr = svm->vmcb->control.exit_code - SVM_EXIT_CR0_WRITE_TRAP;
2706 old_value = kvm_read_cr0(vcpu);
2707 svm_set_cr0(vcpu, new_value);
2709 kvm_post_set_cr0(vcpu, old_value, new_value);
2712 old_value = kvm_read_cr4(vcpu);
2713 svm_set_cr4(vcpu, new_value);
2715 kvm_post_set_cr4(vcpu, old_value, new_value);
2718 ret = kvm_set_cr8(vcpu, new_value);
2721 WARN(1, "unhandled CR%d write trap", cr);
2722 kvm_queue_exception(vcpu, UD_VECTOR);
2726 return kvm_complete_insn_gp(vcpu, ret);
2729 static int dr_interception(struct kvm_vcpu *vcpu)
2731 struct vcpu_svm *svm = to_svm(vcpu);
2737 * SEV-ES intercepts DR7 only to disable guest debugging and the guest issues a VMGEXIT
2738 * for DR7 write only. KVM cannot change DR7 (always swapped as type 'A') so return early.
2740 if (sev_es_guest(vcpu->kvm))
2743 if (vcpu->guest_debug == 0) {
2745 * No more DR vmexits; force a reload of the debug registers
2746 * and reenter on this instruction. The next vmexit will
2747 * retrieve the full state of the debug registers.
2749 clr_dr_intercepts(svm);
2750 vcpu->arch.switch_db_regs |= KVM_DEBUGREG_WONT_EXIT;
2754 if (!boot_cpu_has(X86_FEATURE_DECODEASSISTS))
2755 return emulate_on_interception(vcpu);
2757 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2758 dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
2759 if (dr >= 16) { /* mov to DRn */
2761 val = kvm_register_read(vcpu, reg);
2762 err = kvm_set_dr(vcpu, dr, val);
2764 kvm_get_dr(vcpu, dr, &val);
2765 kvm_register_write(vcpu, reg, val);
2768 return kvm_complete_insn_gp(vcpu, err);
2771 static int cr8_write_interception(struct kvm_vcpu *vcpu)
2775 u8 cr8_prev = kvm_get_cr8(vcpu);
2776 /* instruction emulation calls kvm_set_cr8() */
2777 r = cr_interception(vcpu);
2778 if (lapic_in_kernel(vcpu))
2780 if (cr8_prev <= kvm_get_cr8(vcpu))
2782 vcpu->run->exit_reason = KVM_EXIT_SET_TPR;
2786 static int efer_trap(struct kvm_vcpu *vcpu)
2788 struct msr_data msr_info;
2792 * Clear the EFER_SVME bit from EFER. The SVM code always sets this
2793 * bit in svm_set_efer(), but __kvm_valid_efer() checks it against
2794 * whether the guest has X86_FEATURE_SVM - this avoids a failure if
2795 * the guest doesn't have X86_FEATURE_SVM.
2797 msr_info.host_initiated = false;
2798 msr_info.index = MSR_EFER;
2799 msr_info.data = to_svm(vcpu)->vmcb->control.exit_info_1 & ~EFER_SVME;
2800 ret = kvm_set_msr_common(vcpu, &msr_info);
2802 return kvm_complete_insn_gp(vcpu, ret);
2805 static int svm_get_msr_feature(struct kvm_msr_entry *msr)
2809 switch (msr->index) {
2810 case MSR_AMD64_DE_CFG:
2811 if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
2812 msr->data |= MSR_AMD64_DE_CFG_LFENCE_SERIALIZE;
2815 return KVM_MSR_RET_INVALID;
2821 static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
2823 struct vcpu_svm *svm = to_svm(vcpu);
2825 switch (msr_info->index) {
2826 case MSR_AMD64_TSC_RATIO:
2827 if (!msr_info->host_initiated &&
2828 !guest_can_use(vcpu, X86_FEATURE_TSCRATEMSR))
2830 msr_info->data = svm->tsc_ratio_msr;
2833 msr_info->data = svm->vmcb01.ptr->save.star;
2835 #ifdef CONFIG_X86_64
2837 msr_info->data = svm->vmcb01.ptr->save.lstar;
2840 msr_info->data = svm->vmcb01.ptr->save.cstar;
2842 case MSR_KERNEL_GS_BASE:
2843 msr_info->data = svm->vmcb01.ptr->save.kernel_gs_base;
2845 case MSR_SYSCALL_MASK:
2846 msr_info->data = svm->vmcb01.ptr->save.sfmask;
2849 case MSR_IA32_SYSENTER_CS:
2850 msr_info->data = svm->vmcb01.ptr->save.sysenter_cs;
2852 case MSR_IA32_SYSENTER_EIP:
2853 msr_info->data = (u32)svm->vmcb01.ptr->save.sysenter_eip;
2854 if (guest_cpuid_is_intel(vcpu))
2855 msr_info->data |= (u64)svm->sysenter_eip_hi << 32;
2857 case MSR_IA32_SYSENTER_ESP:
2858 msr_info->data = svm->vmcb01.ptr->save.sysenter_esp;
2859 if (guest_cpuid_is_intel(vcpu))
2860 msr_info->data |= (u64)svm->sysenter_esp_hi << 32;
2863 msr_info->data = svm->tsc_aux;
2865 case MSR_IA32_DEBUGCTLMSR:
2866 msr_info->data = svm_get_lbr_vmcb(svm)->save.dbgctl;
2868 case MSR_IA32_LASTBRANCHFROMIP:
2869 msr_info->data = svm_get_lbr_vmcb(svm)->save.br_from;
2871 case MSR_IA32_LASTBRANCHTOIP:
2872 msr_info->data = svm_get_lbr_vmcb(svm)->save.br_to;
2874 case MSR_IA32_LASTINTFROMIP:
2875 msr_info->data = svm_get_lbr_vmcb(svm)->save.last_excp_from;
2877 case MSR_IA32_LASTINTTOIP:
2878 msr_info->data = svm_get_lbr_vmcb(svm)->save.last_excp_to;
2880 case MSR_VM_HSAVE_PA:
2881 msr_info->data = svm->nested.hsave_msr;
2884 msr_info->data = svm->nested.vm_cr_msr;
2886 case MSR_IA32_SPEC_CTRL:
2887 if (!msr_info->host_initiated &&
2888 !guest_has_spec_ctrl_msr(vcpu))
2891 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2892 msr_info->data = svm->vmcb->save.spec_ctrl;
2894 msr_info->data = svm->spec_ctrl;
2896 case MSR_AMD64_VIRT_SPEC_CTRL:
2897 if (!msr_info->host_initiated &&
2898 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2901 msr_info->data = svm->virt_spec_ctrl;
2903 case MSR_F15H_IC_CFG: {
2907 family = guest_cpuid_family(vcpu);
2908 model = guest_cpuid_model(vcpu);
2910 if (family < 0 || model < 0)
2911 return kvm_get_msr_common(vcpu, msr_info);
2915 if (family == 0x15 &&
2916 (model >= 0x2 && model < 0x20))
2917 msr_info->data = 0x1E;
2920 case MSR_AMD64_DE_CFG:
2921 msr_info->data = svm->msr_decfg;
2924 return kvm_get_msr_common(vcpu, msr_info);
2929 static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
2931 struct vcpu_svm *svm = to_svm(vcpu);
2932 if (!err || !sev_es_guest(vcpu->kvm) || WARN_ON_ONCE(!svm->sev_es.ghcb))
2933 return kvm_complete_insn_gp(vcpu, err);
2935 ghcb_set_sw_exit_info_1(svm->sev_es.ghcb, 1);
2936 ghcb_set_sw_exit_info_2(svm->sev_es.ghcb,
2938 SVM_EVTINJ_TYPE_EXEPT |
2943 static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
2945 struct vcpu_svm *svm = to_svm(vcpu);
2946 int svm_dis, chg_mask;
2948 if (data & ~SVM_VM_CR_VALID_MASK)
2951 chg_mask = SVM_VM_CR_VALID_MASK;
2953 if (svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK)
2954 chg_mask &= ~(SVM_VM_CR_SVM_LOCK_MASK | SVM_VM_CR_SVM_DIS_MASK);
2956 svm->nested.vm_cr_msr &= ~chg_mask;
2957 svm->nested.vm_cr_msr |= (data & chg_mask);
2959 svm_dis = svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK;
2961 /* check for svm_disable while efer.svme is set */
2962 if (svm_dis && (vcpu->arch.efer & EFER_SVME))
2968 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
2970 struct vcpu_svm *svm = to_svm(vcpu);
2973 u32 ecx = msr->index;
2974 u64 data = msr->data;
2976 case MSR_AMD64_TSC_RATIO:
2978 if (!guest_can_use(vcpu, X86_FEATURE_TSCRATEMSR)) {
2980 if (!msr->host_initiated)
2983 * In case TSC scaling is not enabled, always
2984 * leave this MSR at the default value.
2986 * Due to bug in qemu 6.2.0, it would try to set
2987 * this msr to 0 if tsc scaling is not enabled.
2988 * Ignore this value as well.
2990 if (data != 0 && data != svm->tsc_ratio_msr)
2995 if (data & SVM_TSC_RATIO_RSVD)
2998 svm->tsc_ratio_msr = data;
3000 if (guest_can_use(vcpu, X86_FEATURE_TSCRATEMSR) &&
3001 is_guest_mode(vcpu))
3002 nested_svm_update_tsc_ratio_msr(vcpu);
3005 case MSR_IA32_CR_PAT:
3006 ret = kvm_set_msr_common(vcpu, msr);
3010 svm->vmcb01.ptr->save.g_pat = data;
3011 if (is_guest_mode(vcpu))
3012 nested_vmcb02_compute_g_pat(svm);
3013 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
3015 case MSR_IA32_SPEC_CTRL:
3016 if (!msr->host_initiated &&
3017 !guest_has_spec_ctrl_msr(vcpu))
3020 if (kvm_spec_ctrl_test_value(data))
3023 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3024 svm->vmcb->save.spec_ctrl = data;
3026 svm->spec_ctrl = data;
3032 * When it's written (to non-zero) for the first time, pass
3036 * The handling of the MSR bitmap for L2 guests is done in
3037 * nested_svm_vmrun_msrpm.
3038 * We update the L1 MSR bit as well since it will end up
3039 * touching the MSR anyway now.
3041 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
3043 case MSR_AMD64_VIRT_SPEC_CTRL:
3044 if (!msr->host_initiated &&
3045 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
3048 if (data & ~SPEC_CTRL_SSBD)
3051 svm->virt_spec_ctrl = data;
3054 svm->vmcb01.ptr->save.star = data;
3056 #ifdef CONFIG_X86_64
3058 svm->vmcb01.ptr->save.lstar = data;
3061 svm->vmcb01.ptr->save.cstar = data;
3063 case MSR_KERNEL_GS_BASE:
3064 svm->vmcb01.ptr->save.kernel_gs_base = data;
3066 case MSR_SYSCALL_MASK:
3067 svm->vmcb01.ptr->save.sfmask = data;
3070 case MSR_IA32_SYSENTER_CS:
3071 svm->vmcb01.ptr->save.sysenter_cs = data;
3073 case MSR_IA32_SYSENTER_EIP:
3074 svm->vmcb01.ptr->save.sysenter_eip = (u32)data;
3076 * We only intercept the MSR_IA32_SYSENTER_{EIP|ESP} msrs
3077 * when we spoof an Intel vendor ID (for cross vendor migration).
3078 * In this case we use this intercept to track the high
3079 * 32 bit part of these msrs to support Intel's
3080 * implementation of SYSENTER/SYSEXIT.
3082 svm->sysenter_eip_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
3084 case MSR_IA32_SYSENTER_ESP:
3085 svm->vmcb01.ptr->save.sysenter_esp = (u32)data;
3086 svm->sysenter_esp_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
3090 * TSC_AUX is usually changed only during boot and never read
3091 * directly. Intercept TSC_AUX instead of exposing it to the
3092 * guest via direct_access_msrs, and switch it via user return.
3095 ret = kvm_set_user_return_msr(tsc_aux_uret_slot, data, -1ull);
3100 svm->tsc_aux = data;
3102 case MSR_IA32_DEBUGCTLMSR:
3104 kvm_pr_unimpl_wrmsr(vcpu, ecx, data);
3107 if (data & DEBUGCTL_RESERVED_BITS)
3110 svm_get_lbr_vmcb(svm)->save.dbgctl = data;
3111 svm_update_lbrv(vcpu);
3113 case MSR_VM_HSAVE_PA:
3115 * Old kernels did not validate the value written to
3116 * MSR_VM_HSAVE_PA. Allow KVM_SET_MSR to set an invalid
3117 * value to allow live migrating buggy or malicious guests
3118 * originating from those kernels.
3120 if (!msr->host_initiated && !page_address_valid(vcpu, data))
3123 svm->nested.hsave_msr = data & PAGE_MASK;
3126 return svm_set_vm_cr(vcpu, data);
3128 kvm_pr_unimpl_wrmsr(vcpu, ecx, data);
3130 case MSR_AMD64_DE_CFG: {
3131 struct kvm_msr_entry msr_entry;
3133 msr_entry.index = msr->index;
3134 if (svm_get_msr_feature(&msr_entry))
3137 /* Check the supported bits */
3138 if (data & ~msr_entry.data)
3141 /* Don't allow the guest to change a bit, #GP */
3142 if (!msr->host_initiated && (data ^ msr_entry.data))
3145 svm->msr_decfg = data;
3149 return kvm_set_msr_common(vcpu, msr);
3154 static int msr_interception(struct kvm_vcpu *vcpu)
3156 if (to_svm(vcpu)->vmcb->control.exit_info_1)
3157 return kvm_emulate_wrmsr(vcpu);
3159 return kvm_emulate_rdmsr(vcpu);
3162 static int interrupt_window_interception(struct kvm_vcpu *vcpu)
3164 kvm_make_request(KVM_REQ_EVENT, vcpu);
3165 svm_clear_vintr(to_svm(vcpu));
3168 * If not running nested, for AVIC, the only reason to end up here is ExtINTs.
3169 * In this case AVIC was temporarily disabled for
3170 * requesting the IRQ window and we have to re-enable it.
3172 * If running nested, still remove the VM wide AVIC inhibit to
3173 * support case in which the interrupt window was requested when the
3174 * vCPU was not running nested.
3176 * All vCPUs which run still run nested, will remain to have their
3177 * AVIC still inhibited due to per-cpu AVIC inhibition.
3179 kvm_clear_apicv_inhibit(vcpu->kvm, APICV_INHIBIT_REASON_IRQWIN);
3181 ++vcpu->stat.irq_window_exits;
3185 static int pause_interception(struct kvm_vcpu *vcpu)
3189 * CPL is not made available for an SEV-ES guest, therefore
3190 * vcpu->arch.preempted_in_kernel can never be true. Just
3191 * set in_kernel to false as well.
3193 in_kernel = !sev_es_guest(vcpu->kvm) && svm_get_cpl(vcpu) == 0;
3195 grow_ple_window(vcpu);
3197 kvm_vcpu_on_spin(vcpu, in_kernel);
3198 return kvm_skip_emulated_instruction(vcpu);
3201 static int invpcid_interception(struct kvm_vcpu *vcpu)
3203 struct vcpu_svm *svm = to_svm(vcpu);
3207 if (!guest_cpuid_has(vcpu, X86_FEATURE_INVPCID)) {
3208 kvm_queue_exception(vcpu, UD_VECTOR);
3213 * For an INVPCID intercept:
3214 * EXITINFO1 provides the linear address of the memory operand.
3215 * EXITINFO2 provides the contents of the register operand.
3217 type = svm->vmcb->control.exit_info_2;
3218 gva = svm->vmcb->control.exit_info_1;
3220 return kvm_handle_invpcid(vcpu, type, gva);
3223 static int (*const svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {
3224 [SVM_EXIT_READ_CR0] = cr_interception,
3225 [SVM_EXIT_READ_CR3] = cr_interception,
3226 [SVM_EXIT_READ_CR4] = cr_interception,
3227 [SVM_EXIT_READ_CR8] = cr_interception,
3228 [SVM_EXIT_CR0_SEL_WRITE] = cr_interception,
3229 [SVM_EXIT_WRITE_CR0] = cr_interception,
3230 [SVM_EXIT_WRITE_CR3] = cr_interception,
3231 [SVM_EXIT_WRITE_CR4] = cr_interception,
3232 [SVM_EXIT_WRITE_CR8] = cr8_write_interception,
3233 [SVM_EXIT_READ_DR0] = dr_interception,
3234 [SVM_EXIT_READ_DR1] = dr_interception,
3235 [SVM_EXIT_READ_DR2] = dr_interception,
3236 [SVM_EXIT_READ_DR3] = dr_interception,
3237 [SVM_EXIT_READ_DR4] = dr_interception,
3238 [SVM_EXIT_READ_DR5] = dr_interception,
3239 [SVM_EXIT_READ_DR6] = dr_interception,
3240 [SVM_EXIT_READ_DR7] = dr_interception,
3241 [SVM_EXIT_WRITE_DR0] = dr_interception,
3242 [SVM_EXIT_WRITE_DR1] = dr_interception,
3243 [SVM_EXIT_WRITE_DR2] = dr_interception,
3244 [SVM_EXIT_WRITE_DR3] = dr_interception,
3245 [SVM_EXIT_WRITE_DR4] = dr_interception,
3246 [SVM_EXIT_WRITE_DR5] = dr_interception,
3247 [SVM_EXIT_WRITE_DR6] = dr_interception,
3248 [SVM_EXIT_WRITE_DR7] = dr_interception,
3249 [SVM_EXIT_EXCP_BASE + DB_VECTOR] = db_interception,
3250 [SVM_EXIT_EXCP_BASE + BP_VECTOR] = bp_interception,
3251 [SVM_EXIT_EXCP_BASE + UD_VECTOR] = ud_interception,
3252 [SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
3253 [SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
3254 [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
3255 [SVM_EXIT_EXCP_BASE + GP_VECTOR] = gp_interception,
3256 [SVM_EXIT_INTR] = intr_interception,
3257 [SVM_EXIT_NMI] = nmi_interception,
3258 [SVM_EXIT_SMI] = smi_interception,
3259 [SVM_EXIT_VINTR] = interrupt_window_interception,
3260 [SVM_EXIT_RDPMC] = kvm_emulate_rdpmc,
3261 [SVM_EXIT_CPUID] = kvm_emulate_cpuid,
3262 [SVM_EXIT_IRET] = iret_interception,
3263 [SVM_EXIT_INVD] = kvm_emulate_invd,
3264 [SVM_EXIT_PAUSE] = pause_interception,
3265 [SVM_EXIT_HLT] = kvm_emulate_halt,
3266 [SVM_EXIT_INVLPG] = invlpg_interception,
3267 [SVM_EXIT_INVLPGA] = invlpga_interception,
3268 [SVM_EXIT_IOIO] = io_interception,
3269 [SVM_EXIT_MSR] = msr_interception,
3270 [SVM_EXIT_TASK_SWITCH] = task_switch_interception,
3271 [SVM_EXIT_SHUTDOWN] = shutdown_interception,
3272 [SVM_EXIT_VMRUN] = vmrun_interception,
3273 [SVM_EXIT_VMMCALL] = kvm_emulate_hypercall,
3274 [SVM_EXIT_VMLOAD] = vmload_interception,
3275 [SVM_EXIT_VMSAVE] = vmsave_interception,
3276 [SVM_EXIT_STGI] = stgi_interception,
3277 [SVM_EXIT_CLGI] = clgi_interception,
3278 [SVM_EXIT_SKINIT] = skinit_interception,
3279 [SVM_EXIT_RDTSCP] = kvm_handle_invalid_op,
3280 [SVM_EXIT_WBINVD] = kvm_emulate_wbinvd,
3281 [SVM_EXIT_MONITOR] = kvm_emulate_monitor,
3282 [SVM_EXIT_MWAIT] = kvm_emulate_mwait,
3283 [SVM_EXIT_XSETBV] = kvm_emulate_xsetbv,
3284 [SVM_EXIT_RDPRU] = kvm_handle_invalid_op,
3285 [SVM_EXIT_EFER_WRITE_TRAP] = efer_trap,
3286 [SVM_EXIT_CR0_WRITE_TRAP] = cr_trap,
3287 [SVM_EXIT_CR4_WRITE_TRAP] = cr_trap,
3288 [SVM_EXIT_CR8_WRITE_TRAP] = cr_trap,
3289 [SVM_EXIT_INVPCID] = invpcid_interception,
3290 [SVM_EXIT_NPF] = npf_interception,
3291 [SVM_EXIT_RSM] = rsm_interception,
3292 [SVM_EXIT_AVIC_INCOMPLETE_IPI] = avic_incomplete_ipi_interception,
3293 [SVM_EXIT_AVIC_UNACCELERATED_ACCESS] = avic_unaccelerated_access_interception,
3294 [SVM_EXIT_VMGEXIT] = sev_handle_vmgexit,
3297 static void dump_vmcb(struct kvm_vcpu *vcpu)
3299 struct vcpu_svm *svm = to_svm(vcpu);
3300 struct vmcb_control_area *control = &svm->vmcb->control;
3301 struct vmcb_save_area *save = &svm->vmcb->save;
3302 struct vmcb_save_area *save01 = &svm->vmcb01.ptr->save;
3304 if (!dump_invalid_vmcb) {
3305 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3309 pr_err("VMCB %p, last attempted VMRUN on CPU %d\n",
3310 svm->current_vmcb->ptr, vcpu->arch.last_vmentry_cpu);
3311 pr_err("VMCB Control Area:\n");
3312 pr_err("%-20s%04x\n", "cr_read:", control->intercepts[INTERCEPT_CR] & 0xffff);
3313 pr_err("%-20s%04x\n", "cr_write:", control->intercepts[INTERCEPT_CR] >> 16);
3314 pr_err("%-20s%04x\n", "dr_read:", control->intercepts[INTERCEPT_DR] & 0xffff);
3315 pr_err("%-20s%04x\n", "dr_write:", control->intercepts[INTERCEPT_DR] >> 16);
3316 pr_err("%-20s%08x\n", "exceptions:", control->intercepts[INTERCEPT_EXCEPTION]);
3317 pr_err("%-20s%08x %08x\n", "intercepts:",
3318 control->intercepts[INTERCEPT_WORD3],
3319 control->intercepts[INTERCEPT_WORD4]);
3320 pr_err("%-20s%d\n", "pause filter count:", control->pause_filter_count);
3321 pr_err("%-20s%d\n", "pause filter threshold:",
3322 control->pause_filter_thresh);
3323 pr_err("%-20s%016llx\n", "iopm_base_pa:", control->iopm_base_pa);
3324 pr_err("%-20s%016llx\n", "msrpm_base_pa:", control->msrpm_base_pa);
3325 pr_err("%-20s%016llx\n", "tsc_offset:", control->tsc_offset);
3326 pr_err("%-20s%d\n", "asid:", control->asid);
3327 pr_err("%-20s%d\n", "tlb_ctl:", control->tlb_ctl);
3328 pr_err("%-20s%08x\n", "int_ctl:", control->int_ctl);
3329 pr_err("%-20s%08x\n", "int_vector:", control->int_vector);
3330 pr_err("%-20s%08x\n", "int_state:", control->int_state);
3331 pr_err("%-20s%08x\n", "exit_code:", control->exit_code);
3332 pr_err("%-20s%016llx\n", "exit_info1:", control->exit_info_1);
3333 pr_err("%-20s%016llx\n", "exit_info2:", control->exit_info_2);
3334 pr_err("%-20s%08x\n", "exit_int_info:", control->exit_int_info);
3335 pr_err("%-20s%08x\n", "exit_int_info_err:", control->exit_int_info_err);
3336 pr_err("%-20s%lld\n", "nested_ctl:", control->nested_ctl);
3337 pr_err("%-20s%016llx\n", "nested_cr3:", control->nested_cr3);
3338 pr_err("%-20s%016llx\n", "avic_vapic_bar:", control->avic_vapic_bar);
3339 pr_err("%-20s%016llx\n", "ghcb:", control->ghcb_gpa);
3340 pr_err("%-20s%08x\n", "event_inj:", control->event_inj);
3341 pr_err("%-20s%08x\n", "event_inj_err:", control->event_inj_err);
3342 pr_err("%-20s%lld\n", "virt_ext:", control->virt_ext);
3343 pr_err("%-20s%016llx\n", "next_rip:", control->next_rip);
3344 pr_err("%-20s%016llx\n", "avic_backing_page:", control->avic_backing_page);
3345 pr_err("%-20s%016llx\n", "avic_logical_id:", control->avic_logical_id);
3346 pr_err("%-20s%016llx\n", "avic_physical_id:", control->avic_physical_id);
3347 pr_err("%-20s%016llx\n", "vmsa_pa:", control->vmsa_pa);
3348 pr_err("VMCB State Save Area:\n");
3349 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3351 save->es.selector, save->es.attrib,
3352 save->es.limit, save->es.base);
3353 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3355 save->cs.selector, save->cs.attrib,
3356 save->cs.limit, save->cs.base);
3357 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3359 save->ss.selector, save->ss.attrib,
3360 save->ss.limit, save->ss.base);
3361 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3363 save->ds.selector, save->ds.attrib,
3364 save->ds.limit, save->ds.base);
3365 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3367 save01->fs.selector, save01->fs.attrib,
3368 save01->fs.limit, save01->fs.base);
3369 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3371 save01->gs.selector, save01->gs.attrib,
3372 save01->gs.limit, save01->gs.base);
3373 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3375 save->gdtr.selector, save->gdtr.attrib,
3376 save->gdtr.limit, save->gdtr.base);
3377 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3379 save01->ldtr.selector, save01->ldtr.attrib,
3380 save01->ldtr.limit, save01->ldtr.base);
3381 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3383 save->idtr.selector, save->idtr.attrib,
3384 save->idtr.limit, save->idtr.base);
3385 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3387 save01->tr.selector, save01->tr.attrib,
3388 save01->tr.limit, save01->tr.base);
3389 pr_err("vmpl: %d cpl: %d efer: %016llx\n",
3390 save->vmpl, save->cpl, save->efer);
3391 pr_err("%-15s %016llx %-13s %016llx\n",
3392 "cr0:", save->cr0, "cr2:", save->cr2);
3393 pr_err("%-15s %016llx %-13s %016llx\n",
3394 "cr3:", save->cr3, "cr4:", save->cr4);
3395 pr_err("%-15s %016llx %-13s %016llx\n",
3396 "dr6:", save->dr6, "dr7:", save->dr7);
3397 pr_err("%-15s %016llx %-13s %016llx\n",
3398 "rip:", save->rip, "rflags:", save->rflags);
3399 pr_err("%-15s %016llx %-13s %016llx\n",
3400 "rsp:", save->rsp, "rax:", save->rax);
3401 pr_err("%-15s %016llx %-13s %016llx\n",
3402 "star:", save01->star, "lstar:", save01->lstar);
3403 pr_err("%-15s %016llx %-13s %016llx\n",
3404 "cstar:", save01->cstar, "sfmask:", save01->sfmask);
3405 pr_err("%-15s %016llx %-13s %016llx\n",
3406 "kernel_gs_base:", save01->kernel_gs_base,
3407 "sysenter_cs:", save01->sysenter_cs);
3408 pr_err("%-15s %016llx %-13s %016llx\n",
3409 "sysenter_esp:", save01->sysenter_esp,
3410 "sysenter_eip:", save01->sysenter_eip);
3411 pr_err("%-15s %016llx %-13s %016llx\n",
3412 "gpat:", save->g_pat, "dbgctl:", save->dbgctl);
3413 pr_err("%-15s %016llx %-13s %016llx\n",
3414 "br_from:", save->br_from, "br_to:", save->br_to);
3415 pr_err("%-15s %016llx %-13s %016llx\n",
3416 "excp_from:", save->last_excp_from,
3417 "excp_to:", save->last_excp_to);
3420 static bool svm_check_exit_valid(u64 exit_code)
3422 return (exit_code < ARRAY_SIZE(svm_exit_handlers) &&
3423 svm_exit_handlers[exit_code]);
3426 static int svm_handle_invalid_exit(struct kvm_vcpu *vcpu, u64 exit_code)
3428 vcpu_unimpl(vcpu, "svm: unexpected exit reason 0x%llx\n", exit_code);
3430 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3431 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
3432 vcpu->run->internal.ndata = 2;
3433 vcpu->run->internal.data[0] = exit_code;
3434 vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
3438 int svm_invoke_exit_handler(struct kvm_vcpu *vcpu, u64 exit_code)
3440 if (!svm_check_exit_valid(exit_code))
3441 return svm_handle_invalid_exit(vcpu, exit_code);
3443 #ifdef CONFIG_RETPOLINE
3444 if (exit_code == SVM_EXIT_MSR)
3445 return msr_interception(vcpu);
3446 else if (exit_code == SVM_EXIT_VINTR)
3447 return interrupt_window_interception(vcpu);
3448 else if (exit_code == SVM_EXIT_INTR)
3449 return intr_interception(vcpu);
3450 else if (exit_code == SVM_EXIT_HLT)
3451 return kvm_emulate_halt(vcpu);
3452 else if (exit_code == SVM_EXIT_NPF)
3453 return npf_interception(vcpu);
3455 return svm_exit_handlers[exit_code](vcpu);
3458 static void svm_get_exit_info(struct kvm_vcpu *vcpu, u32 *reason,
3459 u64 *info1, u64 *info2,
3460 u32 *intr_info, u32 *error_code)
3462 struct vmcb_control_area *control = &to_svm(vcpu)->vmcb->control;
3464 *reason = control->exit_code;
3465 *info1 = control->exit_info_1;
3466 *info2 = control->exit_info_2;
3467 *intr_info = control->exit_int_info;
3468 if ((*intr_info & SVM_EXITINTINFO_VALID) &&
3469 (*intr_info & SVM_EXITINTINFO_VALID_ERR))
3470 *error_code = control->exit_int_info_err;
3475 static int svm_handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
3477 struct vcpu_svm *svm = to_svm(vcpu);
3478 struct kvm_run *kvm_run = vcpu->run;
3479 u32 exit_code = svm->vmcb->control.exit_code;
3481 /* SEV-ES guests must use the CR write traps to track CR registers. */
3482 if (!sev_es_guest(vcpu->kvm)) {
3483 if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
3484 vcpu->arch.cr0 = svm->vmcb->save.cr0;
3486 vcpu->arch.cr3 = svm->vmcb->save.cr3;
3489 if (is_guest_mode(vcpu)) {
3492 trace_kvm_nested_vmexit(vcpu, KVM_ISA_SVM);
3494 vmexit = nested_svm_exit_special(svm);
3496 if (vmexit == NESTED_EXIT_CONTINUE)
3497 vmexit = nested_svm_exit_handled(svm);
3499 if (vmexit == NESTED_EXIT_DONE)
3503 if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
3504 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3505 kvm_run->fail_entry.hardware_entry_failure_reason
3506 = svm->vmcb->control.exit_code;
3507 kvm_run->fail_entry.cpu = vcpu->arch.last_vmentry_cpu;
3512 if (exit_fastpath != EXIT_FASTPATH_NONE)
3515 return svm_invoke_exit_handler(vcpu, exit_code);
3518 static void pre_svm_run(struct kvm_vcpu *vcpu)
3520 struct svm_cpu_data *sd = per_cpu_ptr(&svm_data, vcpu->cpu);
3521 struct vcpu_svm *svm = to_svm(vcpu);
3524 * If the previous vmrun of the vmcb occurred on a different physical
3525 * cpu, then mark the vmcb dirty and assign a new asid. Hardware's
3526 * vmcb clean bits are per logical CPU, as are KVM's asid assignments.
3528 if (unlikely(svm->current_vmcb->cpu != vcpu->cpu)) {
3529 svm->current_vmcb->asid_generation = 0;
3530 vmcb_mark_all_dirty(svm->vmcb);
3531 svm->current_vmcb->cpu = vcpu->cpu;
3534 if (sev_guest(vcpu->kvm))
3535 return pre_sev_run(svm, vcpu->cpu);
3537 /* FIXME: handle wraparound of asid_generation */
3538 if (svm->current_vmcb->asid_generation != sd->asid_generation)
3542 static void svm_inject_nmi(struct kvm_vcpu *vcpu)
3544 struct vcpu_svm *svm = to_svm(vcpu);
3546 svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
3548 if (svm->nmi_l1_to_l2)
3551 svm->nmi_masked = true;
3552 svm_set_iret_intercept(svm);
3553 ++vcpu->stat.nmi_injections;
3556 static bool svm_is_vnmi_pending(struct kvm_vcpu *vcpu)
3558 struct vcpu_svm *svm = to_svm(vcpu);
3560 if (!is_vnmi_enabled(svm))
3563 return !!(svm->vmcb->control.int_ctl & V_NMI_PENDING_MASK);
3566 static bool svm_set_vnmi_pending(struct kvm_vcpu *vcpu)
3568 struct vcpu_svm *svm = to_svm(vcpu);
3570 if (!is_vnmi_enabled(svm))
3573 if (svm->vmcb->control.int_ctl & V_NMI_PENDING_MASK)
3576 svm->vmcb->control.int_ctl |= V_NMI_PENDING_MASK;
3577 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
3580 * Because the pending NMI is serviced by hardware, KVM can't know when
3581 * the NMI is "injected", but for all intents and purposes, passing the
3582 * NMI off to hardware counts as injection.
3584 ++vcpu->stat.nmi_injections;
3589 static void svm_inject_irq(struct kvm_vcpu *vcpu, bool reinjected)
3591 struct vcpu_svm *svm = to_svm(vcpu);
3594 if (vcpu->arch.interrupt.soft) {
3595 if (svm_update_soft_interrupt_rip(vcpu))
3598 type = SVM_EVTINJ_TYPE_SOFT;
3600 type = SVM_EVTINJ_TYPE_INTR;
3603 trace_kvm_inj_virq(vcpu->arch.interrupt.nr,
3604 vcpu->arch.interrupt.soft, reinjected);
3605 ++vcpu->stat.irq_injections;
3607 svm->vmcb->control.event_inj = vcpu->arch.interrupt.nr |
3608 SVM_EVTINJ_VALID | type;
3611 void svm_complete_interrupt_delivery(struct kvm_vcpu *vcpu, int delivery_mode,
3612 int trig_mode, int vector)
3615 * apic->apicv_active must be read after vcpu->mode.
3616 * Pairs with smp_store_release in vcpu_enter_guest.
3618 bool in_guest_mode = (smp_load_acquire(&vcpu->mode) == IN_GUEST_MODE);
3620 /* Note, this is called iff the local APIC is in-kernel. */
3621 if (!READ_ONCE(vcpu->arch.apic->apicv_active)) {
3622 /* Process the interrupt via kvm_check_and_inject_events(). */
3623 kvm_make_request(KVM_REQ_EVENT, vcpu);
3624 kvm_vcpu_kick(vcpu);
3628 trace_kvm_apicv_accept_irq(vcpu->vcpu_id, delivery_mode, trig_mode, vector);
3629 if (in_guest_mode) {
3631 * Signal the doorbell to tell hardware to inject the IRQ. If
3632 * the vCPU exits the guest before the doorbell chimes, hardware
3633 * will automatically process AVIC interrupts at the next VMRUN.
3635 avic_ring_doorbell(vcpu);
3638 * Wake the vCPU if it was blocking. KVM will then detect the
3639 * pending IRQ when checking if the vCPU has a wake event.
3641 kvm_vcpu_wake_up(vcpu);
3645 static void svm_deliver_interrupt(struct kvm_lapic *apic, int delivery_mode,
3646 int trig_mode, int vector)
3648 kvm_lapic_set_irr(vector, apic);
3651 * Pairs with the smp_mb_*() after setting vcpu->guest_mode in
3652 * vcpu_enter_guest() to ensure the write to the vIRR is ordered before
3653 * the read of guest_mode. This guarantees that either VMRUN will see
3654 * and process the new vIRR entry, or that svm_complete_interrupt_delivery
3655 * will signal the doorbell if the CPU has already entered the guest.
3657 smp_mb__after_atomic();
3658 svm_complete_interrupt_delivery(apic->vcpu, delivery_mode, trig_mode, vector);
3661 static void svm_update_cr8_intercept(struct kvm_vcpu *vcpu, int tpr, int irr)
3663 struct vcpu_svm *svm = to_svm(vcpu);
3666 * SEV-ES guests must always keep the CR intercepts cleared. CR
3667 * tracking is done using the CR write traps.
3669 if (sev_es_guest(vcpu->kvm))
3672 if (nested_svm_virtualize_tpr(vcpu))
3675 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3681 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
3684 static bool svm_get_nmi_mask(struct kvm_vcpu *vcpu)
3686 struct vcpu_svm *svm = to_svm(vcpu);
3688 if (is_vnmi_enabled(svm))
3689 return svm->vmcb->control.int_ctl & V_NMI_BLOCKING_MASK;
3691 return svm->nmi_masked;
3694 static void svm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
3696 struct vcpu_svm *svm = to_svm(vcpu);
3698 if (is_vnmi_enabled(svm)) {
3700 svm->vmcb->control.int_ctl |= V_NMI_BLOCKING_MASK;
3702 svm->vmcb->control.int_ctl &= ~V_NMI_BLOCKING_MASK;
3705 svm->nmi_masked = masked;
3707 svm_set_iret_intercept(svm);
3709 svm_clr_iret_intercept(svm);
3713 bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
3715 struct vcpu_svm *svm = to_svm(vcpu);
3716 struct vmcb *vmcb = svm->vmcb;
3721 if (is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3724 if (svm_get_nmi_mask(vcpu))
3727 return vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK;
3730 static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3732 struct vcpu_svm *svm = to_svm(vcpu);
3733 if (svm->nested.nested_run_pending)
3736 if (svm_nmi_blocked(vcpu))
3739 /* An NMI must not be injected into L2 if it's supposed to VM-Exit. */
3740 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3745 bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
3747 struct vcpu_svm *svm = to_svm(vcpu);
3748 struct vmcb *vmcb = svm->vmcb;
3753 if (is_guest_mode(vcpu)) {
3754 /* As long as interrupts are being delivered... */
3755 if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
3756 ? !(svm->vmcb01.ptr->save.rflags & X86_EFLAGS_IF)
3757 : !(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3760 /* ... vmexits aren't blocked by the interrupt shadow */
3761 if (nested_exit_on_intr(svm))
3764 if (!svm_get_if_flag(vcpu))
3768 return (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK);
3771 static int svm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3773 struct vcpu_svm *svm = to_svm(vcpu);
3775 if (svm->nested.nested_run_pending)
3778 if (svm_interrupt_blocked(vcpu))
3782 * An IRQ must not be injected into L2 if it's supposed to VM-Exit,
3783 * e.g. if the IRQ arrived asynchronously after checking nested events.
3785 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_intr(svm))
3791 static void svm_enable_irq_window(struct kvm_vcpu *vcpu)
3793 struct vcpu_svm *svm = to_svm(vcpu);
3796 * In case GIF=0 we can't rely on the CPU to tell us when GIF becomes
3797 * 1, because that's a separate STGI/VMRUN intercept. The next time we
3798 * get that intercept, this function will be called again though and
3799 * we'll get the vintr intercept. However, if the vGIF feature is
3800 * enabled, the STGI interception will not occur. Enable the irq
3801 * window under the assumption that the hardware will set the GIF.
3803 if (vgif || gif_set(svm)) {
3805 * IRQ window is not needed when AVIC is enabled,
3806 * unless we have pending ExtINT since it cannot be injected
3807 * via AVIC. In such case, KVM needs to temporarily disable AVIC,
3808 * and fallback to injecting IRQ via V_IRQ.
3810 * If running nested, AVIC is already locally inhibited
3811 * on this vCPU, therefore there is no need to request
3812 * the VM wide AVIC inhibition.
3814 if (!is_guest_mode(vcpu))
3815 kvm_set_apicv_inhibit(vcpu->kvm, APICV_INHIBIT_REASON_IRQWIN);
3821 static void svm_enable_nmi_window(struct kvm_vcpu *vcpu)
3823 struct vcpu_svm *svm = to_svm(vcpu);
3826 * KVM should never request an NMI window when vNMI is enabled, as KVM
3827 * allows at most one to-be-injected NMI and one pending NMI, i.e. if
3828 * two NMIs arrive simultaneously, KVM will inject one and set
3829 * V_NMI_PENDING for the other. WARN, but continue with the standard
3830 * single-step approach to try and salvage the pending NMI.
3832 WARN_ON_ONCE(is_vnmi_enabled(svm));
3834 if (svm_get_nmi_mask(vcpu) && !svm->awaiting_iret_completion)
3835 return; /* IRET will cause a vm exit */
3838 * SEV-ES guests are responsible for signaling when a vCPU is ready to
3839 * receive a new NMI, as SEV-ES guests can't be single-stepped, i.e.
3840 * KVM can't intercept and single-step IRET to detect when NMIs are
3841 * unblocked (architecturally speaking). See SVM_VMGEXIT_NMI_COMPLETE.
3843 * Note, GIF is guaranteed to be '1' for SEV-ES guests as hardware
3844 * ignores SEV-ES guest writes to EFER.SVME *and* CLGI/STGI are not
3845 * supported NAEs in the GHCB protocol.
3847 if (sev_es_guest(vcpu->kvm))
3850 if (!gif_set(svm)) {
3852 svm_set_intercept(svm, INTERCEPT_STGI);
3853 return; /* STGI will cause a vm exit */
3857 * Something prevents NMI from been injected. Single step over possible
3858 * problem (IRET or exception injection or interrupt shadow)
3860 svm->nmi_singlestep_guest_rflags = svm_get_rflags(vcpu);
3861 svm->nmi_singlestep = true;
3862 svm->vmcb->save.rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
3865 static void svm_flush_tlb_asid(struct kvm_vcpu *vcpu)
3867 struct vcpu_svm *svm = to_svm(vcpu);
3870 * Unlike VMX, SVM doesn't provide a way to flush only NPT TLB entries.
3871 * A TLB flush for the current ASID flushes both "host" and "guest" TLB
3872 * entries, and thus is a superset of Hyper-V's fine grained flushing.
3874 kvm_hv_vcpu_purge_flush_tlb(vcpu);
3877 * Flush only the current ASID even if the TLB flush was invoked via
3878 * kvm_flush_remote_tlbs(). Although flushing remote TLBs requires all
3879 * ASIDs to be flushed, KVM uses a single ASID for L1 and L2, and
3880 * unconditionally does a TLB flush on both nested VM-Enter and nested
3881 * VM-Exit (via kvm_mmu_reset_context()).
3883 if (static_cpu_has(X86_FEATURE_FLUSHBYASID))
3884 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3886 svm->current_vmcb->asid_generation--;
3889 static void svm_flush_tlb_current(struct kvm_vcpu *vcpu)
3891 hpa_t root_tdp = vcpu->arch.mmu->root.hpa;
3894 * When running on Hyper-V with EnlightenedNptTlb enabled, explicitly
3895 * flush the NPT mappings via hypercall as flushing the ASID only
3896 * affects virtual to physical mappings, it does not invalidate guest
3897 * physical to host physical mappings.
3899 if (svm_hv_is_enlightened_tlb_enabled(vcpu) && VALID_PAGE(root_tdp))
3900 hyperv_flush_guest_mapping(root_tdp);
3902 svm_flush_tlb_asid(vcpu);
3905 static void svm_flush_tlb_all(struct kvm_vcpu *vcpu)
3908 * When running on Hyper-V with EnlightenedNptTlb enabled, remote TLB
3909 * flushes should be routed to hv_flush_remote_tlbs() without requesting
3910 * a "regular" remote flush. Reaching this point means either there's
3911 * a KVM bug or a prior hv_flush_remote_tlbs() call failed, both of
3912 * which might be fatal to the guest. Yell, but try to recover.
3914 if (WARN_ON_ONCE(svm_hv_is_enlightened_tlb_enabled(vcpu)))
3915 hv_flush_remote_tlbs(vcpu->kvm);
3917 svm_flush_tlb_asid(vcpu);
3920 static void svm_flush_tlb_gva(struct kvm_vcpu *vcpu, gva_t gva)
3922 struct vcpu_svm *svm = to_svm(vcpu);
3924 invlpga(gva, svm->vmcb->control.asid);
3927 static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)
3929 struct vcpu_svm *svm = to_svm(vcpu);
3931 if (nested_svm_virtualize_tpr(vcpu))
3934 if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
3935 int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
3936 kvm_set_cr8(vcpu, cr8);
3940 static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
3942 struct vcpu_svm *svm = to_svm(vcpu);
3945 if (nested_svm_virtualize_tpr(vcpu) ||
3946 kvm_vcpu_apicv_active(vcpu))
3949 cr8 = kvm_get_cr8(vcpu);
3950 svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
3951 svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
3954 static void svm_complete_soft_interrupt(struct kvm_vcpu *vcpu, u8 vector,
3957 bool is_exception = (type == SVM_EXITINTINFO_TYPE_EXEPT);
3958 bool is_soft = (type == SVM_EXITINTINFO_TYPE_SOFT);
3959 struct vcpu_svm *svm = to_svm(vcpu);
3962 * If NRIPS is enabled, KVM must snapshot the pre-VMRUN next_rip that's
3963 * associated with the original soft exception/interrupt. next_rip is
3964 * cleared on all exits that can occur while vectoring an event, so KVM
3965 * needs to manually set next_rip for re-injection. Unlike the !nrips
3966 * case below, this needs to be done if and only if KVM is re-injecting
3967 * the same event, i.e. if the event is a soft exception/interrupt,
3968 * otherwise next_rip is unused on VMRUN.
3970 if (nrips && (is_soft || (is_exception && kvm_exception_is_soft(vector))) &&
3971 kvm_is_linear_rip(vcpu, svm->soft_int_old_rip + svm->soft_int_csbase))
3972 svm->vmcb->control.next_rip = svm->soft_int_next_rip;
3974 * If NRIPS isn't enabled, KVM must manually advance RIP prior to
3975 * injecting the soft exception/interrupt. That advancement needs to
3976 * be unwound if vectoring didn't complete. Note, the new event may
3977 * not be the injected event, e.g. if KVM injected an INTn, the INTn
3978 * hit a #NP in the guest, and the #NP encountered a #PF, the #NP will
3979 * be the reported vectored event, but RIP still needs to be unwound.
3981 else if (!nrips && (is_soft || is_exception) &&
3982 kvm_is_linear_rip(vcpu, svm->soft_int_next_rip + svm->soft_int_csbase))
3983 kvm_rip_write(vcpu, svm->soft_int_old_rip);
3986 static void svm_complete_interrupts(struct kvm_vcpu *vcpu)
3988 struct vcpu_svm *svm = to_svm(vcpu);
3991 u32 exitintinfo = svm->vmcb->control.exit_int_info;
3992 bool nmi_l1_to_l2 = svm->nmi_l1_to_l2;
3993 bool soft_int_injected = svm->soft_int_injected;
3995 svm->nmi_l1_to_l2 = false;
3996 svm->soft_int_injected = false;
3999 * If we've made progress since setting awaiting_iret_completion, we've
4000 * executed an IRET and can allow NMI injection.
4002 if (svm->awaiting_iret_completion &&
4003 kvm_rip_read(vcpu) != svm->nmi_iret_rip) {
4004 svm->awaiting_iret_completion = false;
4005 svm->nmi_masked = false;
4006 kvm_make_request(KVM_REQ_EVENT, vcpu);
4009 vcpu->arch.nmi_injected = false;
4010 kvm_clear_exception_queue(vcpu);
4011 kvm_clear_interrupt_queue(vcpu);
4013 if (!(exitintinfo & SVM_EXITINTINFO_VALID))
4016 kvm_make_request(KVM_REQ_EVENT, vcpu);
4018 vector = exitintinfo & SVM_EXITINTINFO_VEC_MASK;
4019 type = exitintinfo & SVM_EXITINTINFO_TYPE_MASK;
4021 if (soft_int_injected)
4022 svm_complete_soft_interrupt(vcpu, vector, type);
4025 case SVM_EXITINTINFO_TYPE_NMI:
4026 vcpu->arch.nmi_injected = true;
4027 svm->nmi_l1_to_l2 = nmi_l1_to_l2;
4029 case SVM_EXITINTINFO_TYPE_EXEPT:
4031 * Never re-inject a #VC exception.
4033 if (vector == X86_TRAP_VC)
4036 if (exitintinfo & SVM_EXITINTINFO_VALID_ERR) {
4037 u32 err = svm->vmcb->control.exit_int_info_err;
4038 kvm_requeue_exception_e(vcpu, vector, err);
4041 kvm_requeue_exception(vcpu, vector);
4043 case SVM_EXITINTINFO_TYPE_INTR:
4044 kvm_queue_interrupt(vcpu, vector, false);
4046 case SVM_EXITINTINFO_TYPE_SOFT:
4047 kvm_queue_interrupt(vcpu, vector, true);
4055 static void svm_cancel_injection(struct kvm_vcpu *vcpu)
4057 struct vcpu_svm *svm = to_svm(vcpu);
4058 struct vmcb_control_area *control = &svm->vmcb->control;
4060 control->exit_int_info = control->event_inj;
4061 control->exit_int_info_err = control->event_inj_err;
4062 control->event_inj = 0;
4063 svm_complete_interrupts(vcpu);
4066 static int svm_vcpu_pre_run(struct kvm_vcpu *vcpu)
4071 static fastpath_t svm_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
4073 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_MSR &&
4074 to_svm(vcpu)->vmcb->control.exit_info_1)
4075 return handle_fastpath_set_msr_irqoff(vcpu);
4077 return EXIT_FASTPATH_NONE;
4080 static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu, bool spec_ctrl_intercepted)
4082 struct vcpu_svm *svm = to_svm(vcpu);
4084 guest_state_enter_irqoff();
4086 amd_clear_divider();
4088 if (sev_es_guest(vcpu->kvm))
4089 __svm_sev_es_vcpu_run(svm, spec_ctrl_intercepted);
4091 __svm_vcpu_run(svm, spec_ctrl_intercepted);
4093 guest_state_exit_irqoff();
4096 static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
4098 struct vcpu_svm *svm = to_svm(vcpu);
4099 bool spec_ctrl_intercepted = msr_write_intercepted(vcpu, MSR_IA32_SPEC_CTRL);
4101 trace_kvm_entry(vcpu);
4103 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4104 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4105 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4108 * Disable singlestep if we're injecting an interrupt/exception.
4109 * We don't want our modified rflags to be pushed on the stack where
4110 * we might not be able to easily reset them if we disabled NMI
4113 if (svm->nmi_singlestep && svm->vmcb->control.event_inj) {
4115 * Event injection happens before external interrupts cause a
4116 * vmexit and interrupts are disabled here, so smp_send_reschedule
4117 * is enough to force an immediate vmexit.
4119 disable_nmi_singlestep(svm);
4120 smp_send_reschedule(vcpu->cpu);
4125 sync_lapic_to_cr8(vcpu);
4127 if (unlikely(svm->asid != svm->vmcb->control.asid)) {
4128 svm->vmcb->control.asid = svm->asid;
4129 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
4131 svm->vmcb->save.cr2 = vcpu->arch.cr2;
4133 svm_hv_update_vp_id(svm->vmcb, vcpu);
4136 * Run with all-zero DR6 unless needed, so that we can get the exact cause
4139 if (unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT))
4140 svm_set_dr6(svm, vcpu->arch.dr6);
4142 svm_set_dr6(svm, DR6_ACTIVE_LOW);
4145 kvm_load_guest_xsave_state(vcpu);
4147 kvm_wait_lapic_expire(vcpu);
4150 * If this vCPU has touched SPEC_CTRL, restore the guest's value if
4151 * it's non-zero. Since vmentry is serialising on affected CPUs, there
4152 * is no need to worry about the conditional branch over the wrmsr
4153 * being speculatively taken.
4155 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
4156 x86_spec_ctrl_set_guest(svm->virt_spec_ctrl);
4158 svm_vcpu_enter_exit(vcpu, spec_ctrl_intercepted);
4160 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
4161 x86_spec_ctrl_restore_host(svm->virt_spec_ctrl);
4163 if (!sev_es_guest(vcpu->kvm)) {
4164 vcpu->arch.cr2 = svm->vmcb->save.cr2;
4165 vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
4166 vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
4167 vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
4169 vcpu->arch.regs_dirty = 0;
4171 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
4172 kvm_before_interrupt(vcpu, KVM_HANDLING_NMI);
4174 kvm_load_host_xsave_state(vcpu);
4177 /* Any pending NMI will happen here */
4179 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
4180 kvm_after_interrupt(vcpu);
4182 sync_cr8_to_lapic(vcpu);
4185 if (is_guest_mode(vcpu)) {
4186 nested_sync_control_from_vmcb02(svm);
4188 /* Track VMRUNs that have made past consistency checking */
4189 if (svm->nested.nested_run_pending &&
4190 svm->vmcb->control.exit_code != SVM_EXIT_ERR)
4191 ++vcpu->stat.nested_run;
4193 svm->nested.nested_run_pending = 0;
4196 svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
4197 vmcb_mark_all_clean(svm->vmcb);
4199 /* if exit due to PF check for async PF */
4200 if (svm->vmcb->control.exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR)
4201 vcpu->arch.apf.host_apf_flags =
4202 kvm_read_and_reset_apf_flags();
4204 vcpu->arch.regs_avail &= ~SVM_REGS_LAZY_LOAD_SET;
4207 * We need to handle MC intercepts here before the vcpu has a chance to
4208 * change the physical cpu
4210 if (unlikely(svm->vmcb->control.exit_code ==
4211 SVM_EXIT_EXCP_BASE + MC_VECTOR))
4212 svm_handle_mce(vcpu);
4214 trace_kvm_exit(vcpu, KVM_ISA_SVM);
4216 svm_complete_interrupts(vcpu);
4218 if (is_guest_mode(vcpu))
4219 return EXIT_FASTPATH_NONE;
4221 return svm_exit_handlers_fastpath(vcpu);
4224 static void svm_load_mmu_pgd(struct kvm_vcpu *vcpu, hpa_t root_hpa,
4227 struct vcpu_svm *svm = to_svm(vcpu);
4231 svm->vmcb->control.nested_cr3 = __sme_set(root_hpa);
4232 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
4234 hv_track_root_tdp(vcpu, root_hpa);
4236 cr3 = vcpu->arch.cr3;
4237 } else if (root_level >= PT64_ROOT_4LEVEL) {
4238 cr3 = __sme_set(root_hpa) | kvm_get_active_pcid(vcpu);
4240 /* PCID in the guest should be impossible with a 32-bit MMU. */
4241 WARN_ON_ONCE(kvm_get_active_pcid(vcpu));
4245 svm->vmcb->save.cr3 = cr3;
4246 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
4250 svm_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
4253 * Patch in the VMMCALL instruction:
4255 hypercall[0] = 0x0f;
4256 hypercall[1] = 0x01;
4257 hypercall[2] = 0xd9;
4261 * The kvm parameter can be NULL (module initialization, or invocation before
4262 * VM creation). Be sure to check the kvm parameter before using it.
4264 static bool svm_has_emulated_msr(struct kvm *kvm, u32 index)
4267 case MSR_IA32_MCG_EXT_CTL:
4268 case KVM_FIRST_EMULATED_VMX_MSR ... KVM_LAST_EMULATED_VMX_MSR:
4270 case MSR_IA32_SMBASE:
4271 if (!IS_ENABLED(CONFIG_KVM_SMM))
4273 /* SEV-ES guests do not support SMM, so report false */
4274 if (kvm && sev_es_guest(kvm))
4284 static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
4286 struct vcpu_svm *svm = to_svm(vcpu);
4287 struct kvm_cpuid_entry2 *best;
4290 * SVM doesn't provide a way to disable just XSAVES in the guest, KVM
4291 * can only disable all variants of by disallowing CR4.OSXSAVE from
4292 * being set. As a result, if the host has XSAVE and XSAVES, and the
4293 * guest has XSAVE enabled, the guest can execute XSAVES without
4294 * faulting. Treat XSAVES as enabled in this case regardless of
4295 * whether it's advertised to the guest so that KVM context switches
4296 * XSS on VM-Enter/VM-Exit. Failure to do so would effectively give
4297 * the guest read/write access to the host's XSS.
4299 if (boot_cpu_has(X86_FEATURE_XSAVE) &&
4300 boot_cpu_has(X86_FEATURE_XSAVES) &&
4301 guest_cpuid_has(vcpu, X86_FEATURE_XSAVE))
4302 kvm_governed_feature_set(vcpu, X86_FEATURE_XSAVES);
4304 kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_NRIPS);
4305 kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_TSCRATEMSR);
4306 kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_LBRV);
4309 * Intercept VMLOAD if the vCPU mode is Intel in order to emulate that
4310 * VMLOAD drops bits 63:32 of SYSENTER (ignoring the fact that exposing
4311 * SVM on Intel is bonkers and extremely unlikely to work).
4313 if (!guest_cpuid_is_intel(vcpu))
4314 kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_V_VMSAVE_VMLOAD);
4316 kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_PAUSEFILTER);
4317 kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_PFTHRESHOLD);
4318 kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_VGIF);
4319 kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_VNMI);
4321 svm_recalc_instruction_intercepts(vcpu, svm);
4323 if (boot_cpu_has(X86_FEATURE_IBPB))
4324 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_PRED_CMD, 0,
4325 !!guest_has_pred_cmd_msr(vcpu));
4327 if (boot_cpu_has(X86_FEATURE_FLUSH_L1D))
4328 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_FLUSH_CMD, 0,
4329 !!guest_cpuid_has(vcpu, X86_FEATURE_FLUSH_L1D));
4331 /* For sev guests, the memory encryption bit is not reserved in CR3. */
4332 if (sev_guest(vcpu->kvm)) {
4333 best = kvm_find_cpuid_entry(vcpu, 0x8000001F);
4335 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
4338 init_vmcb_after_set_cpuid(vcpu);
4341 static bool svm_has_wbinvd_exit(void)
4346 #define PRE_EX(exit) { .exit_code = (exit), \
4347 .stage = X86_ICPT_PRE_EXCEPT, }
4348 #define POST_EX(exit) { .exit_code = (exit), \
4349 .stage = X86_ICPT_POST_EXCEPT, }
4350 #define POST_MEM(exit) { .exit_code = (exit), \
4351 .stage = X86_ICPT_POST_MEMACCESS, }
4353 static const struct __x86_intercept {
4355 enum x86_intercept_stage stage;
4356 } x86_intercept_map[] = {
4357 [x86_intercept_cr_read] = POST_EX(SVM_EXIT_READ_CR0),
4358 [x86_intercept_cr_write] = POST_EX(SVM_EXIT_WRITE_CR0),
4359 [x86_intercept_clts] = POST_EX(SVM_EXIT_WRITE_CR0),
4360 [x86_intercept_lmsw] = POST_EX(SVM_EXIT_WRITE_CR0),
4361 [x86_intercept_smsw] = POST_EX(SVM_EXIT_READ_CR0),
4362 [x86_intercept_dr_read] = POST_EX(SVM_EXIT_READ_DR0),
4363 [x86_intercept_dr_write] = POST_EX(SVM_EXIT_WRITE_DR0),
4364 [x86_intercept_sldt] = POST_EX(SVM_EXIT_LDTR_READ),
4365 [x86_intercept_str] = POST_EX(SVM_EXIT_TR_READ),
4366 [x86_intercept_lldt] = POST_EX(SVM_EXIT_LDTR_WRITE),
4367 [x86_intercept_ltr] = POST_EX(SVM_EXIT_TR_WRITE),
4368 [x86_intercept_sgdt] = POST_EX(SVM_EXIT_GDTR_READ),
4369 [x86_intercept_sidt] = POST_EX(SVM_EXIT_IDTR_READ),
4370 [x86_intercept_lgdt] = POST_EX(SVM_EXIT_GDTR_WRITE),
4371 [x86_intercept_lidt] = POST_EX(SVM_EXIT_IDTR_WRITE),
4372 [x86_intercept_vmrun] = POST_EX(SVM_EXIT_VMRUN),
4373 [x86_intercept_vmmcall] = POST_EX(SVM_EXIT_VMMCALL),
4374 [x86_intercept_vmload] = POST_EX(SVM_EXIT_VMLOAD),
4375 [x86_intercept_vmsave] = POST_EX(SVM_EXIT_VMSAVE),
4376 [x86_intercept_stgi] = POST_EX(SVM_EXIT_STGI),
4377 [x86_intercept_clgi] = POST_EX(SVM_EXIT_CLGI),
4378 [x86_intercept_skinit] = POST_EX(SVM_EXIT_SKINIT),
4379 [x86_intercept_invlpga] = POST_EX(SVM_EXIT_INVLPGA),
4380 [x86_intercept_rdtscp] = POST_EX(SVM_EXIT_RDTSCP),
4381 [x86_intercept_monitor] = POST_MEM(SVM_EXIT_MONITOR),
4382 [x86_intercept_mwait] = POST_EX(SVM_EXIT_MWAIT),
4383 [x86_intercept_invlpg] = POST_EX(SVM_EXIT_INVLPG),
4384 [x86_intercept_invd] = POST_EX(SVM_EXIT_INVD),
4385 [x86_intercept_wbinvd] = POST_EX(SVM_EXIT_WBINVD),
4386 [x86_intercept_wrmsr] = POST_EX(SVM_EXIT_MSR),
4387 [x86_intercept_rdtsc] = POST_EX(SVM_EXIT_RDTSC),
4388 [x86_intercept_rdmsr] = POST_EX(SVM_EXIT_MSR),
4389 [x86_intercept_rdpmc] = POST_EX(SVM_EXIT_RDPMC),
4390 [x86_intercept_cpuid] = PRE_EX(SVM_EXIT_CPUID),
4391 [x86_intercept_rsm] = PRE_EX(SVM_EXIT_RSM),
4392 [x86_intercept_pause] = PRE_EX(SVM_EXIT_PAUSE),
4393 [x86_intercept_pushf] = PRE_EX(SVM_EXIT_PUSHF),
4394 [x86_intercept_popf] = PRE_EX(SVM_EXIT_POPF),
4395 [x86_intercept_intn] = PRE_EX(SVM_EXIT_SWINT),
4396 [x86_intercept_iret] = PRE_EX(SVM_EXIT_IRET),
4397 [x86_intercept_icebp] = PRE_EX(SVM_EXIT_ICEBP),
4398 [x86_intercept_hlt] = POST_EX(SVM_EXIT_HLT),
4399 [x86_intercept_in] = POST_EX(SVM_EXIT_IOIO),
4400 [x86_intercept_ins] = POST_EX(SVM_EXIT_IOIO),
4401 [x86_intercept_out] = POST_EX(SVM_EXIT_IOIO),
4402 [x86_intercept_outs] = POST_EX(SVM_EXIT_IOIO),
4403 [x86_intercept_xsetbv] = PRE_EX(SVM_EXIT_XSETBV),
4410 static int svm_check_intercept(struct kvm_vcpu *vcpu,
4411 struct x86_instruction_info *info,
4412 enum x86_intercept_stage stage,
4413 struct x86_exception *exception)
4415 struct vcpu_svm *svm = to_svm(vcpu);
4416 int vmexit, ret = X86EMUL_CONTINUE;
4417 struct __x86_intercept icpt_info;
4418 struct vmcb *vmcb = svm->vmcb;
4420 if (info->intercept >= ARRAY_SIZE(x86_intercept_map))
4423 icpt_info = x86_intercept_map[info->intercept];
4425 if (stage != icpt_info.stage)
4428 switch (icpt_info.exit_code) {
4429 case SVM_EXIT_READ_CR0:
4430 if (info->intercept == x86_intercept_cr_read)
4431 icpt_info.exit_code += info->modrm_reg;
4433 case SVM_EXIT_WRITE_CR0: {
4434 unsigned long cr0, val;
4436 if (info->intercept == x86_intercept_cr_write)
4437 icpt_info.exit_code += info->modrm_reg;
4439 if (icpt_info.exit_code != SVM_EXIT_WRITE_CR0 ||
4440 info->intercept == x86_intercept_clts)
4443 if (!(vmcb12_is_intercept(&svm->nested.ctl,
4444 INTERCEPT_SELECTIVE_CR0)))
4447 cr0 = vcpu->arch.cr0 & ~SVM_CR0_SELECTIVE_MASK;
4448 val = info->src_val & ~SVM_CR0_SELECTIVE_MASK;
4450 if (info->intercept == x86_intercept_lmsw) {
4453 /* lmsw can't clear PE - catch this here */
4454 if (cr0 & X86_CR0_PE)
4459 icpt_info.exit_code = SVM_EXIT_CR0_SEL_WRITE;
4463 case SVM_EXIT_READ_DR0:
4464 case SVM_EXIT_WRITE_DR0:
4465 icpt_info.exit_code += info->modrm_reg;
4468 if (info->intercept == x86_intercept_wrmsr)
4469 vmcb->control.exit_info_1 = 1;
4471 vmcb->control.exit_info_1 = 0;
4473 case SVM_EXIT_PAUSE:
4475 * We get this for NOP only, but pause
4476 * is rep not, check this here
4478 if (info->rep_prefix != REPE_PREFIX)
4481 case SVM_EXIT_IOIO: {
4485 if (info->intercept == x86_intercept_in ||
4486 info->intercept == x86_intercept_ins) {
4487 exit_info = ((info->src_val & 0xffff) << 16) |
4489 bytes = info->dst_bytes;
4491 exit_info = (info->dst_val & 0xffff) << 16;
4492 bytes = info->src_bytes;
4495 if (info->intercept == x86_intercept_outs ||
4496 info->intercept == x86_intercept_ins)
4497 exit_info |= SVM_IOIO_STR_MASK;
4499 if (info->rep_prefix)
4500 exit_info |= SVM_IOIO_REP_MASK;
4502 bytes = min(bytes, 4u);
4504 exit_info |= bytes << SVM_IOIO_SIZE_SHIFT;
4506 exit_info |= (u32)info->ad_bytes << (SVM_IOIO_ASIZE_SHIFT - 1);
4508 vmcb->control.exit_info_1 = exit_info;
4509 vmcb->control.exit_info_2 = info->next_rip;
4517 /* TODO: Advertise NRIPS to guest hypervisor unconditionally */
4518 if (static_cpu_has(X86_FEATURE_NRIPS))
4519 vmcb->control.next_rip = info->next_rip;
4520 vmcb->control.exit_code = icpt_info.exit_code;
4521 vmexit = nested_svm_exit_handled(svm);
4523 ret = (vmexit == NESTED_EXIT_DONE) ? X86EMUL_INTERCEPTED
4530 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
4532 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_INTR)
4533 vcpu->arch.at_instruction_boundary = true;
4536 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
4538 if (!kvm_pause_in_guest(vcpu->kvm))
4539 shrink_ple_window(vcpu);
4542 static void svm_setup_mce(struct kvm_vcpu *vcpu)
4544 /* [63:9] are reserved. */
4545 vcpu->arch.mcg_cap &= 0x1ff;
4548 #ifdef CONFIG_KVM_SMM
4549 bool svm_smi_blocked(struct kvm_vcpu *vcpu)
4551 struct vcpu_svm *svm = to_svm(vcpu);
4553 /* Per APM Vol.2 15.22.2 "Response to SMI" */
4557 return is_smm(vcpu);
4560 static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
4562 struct vcpu_svm *svm = to_svm(vcpu);
4563 if (svm->nested.nested_run_pending)
4566 if (svm_smi_blocked(vcpu))
4569 /* An SMI must not be injected into L2 if it's supposed to VM-Exit. */
4570 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_smi(svm))
4576 static int svm_enter_smm(struct kvm_vcpu *vcpu, union kvm_smram *smram)
4578 struct vcpu_svm *svm = to_svm(vcpu);
4579 struct kvm_host_map map_save;
4582 if (!is_guest_mode(vcpu))
4586 * 32-bit SMRAM format doesn't preserve EFER and SVM state. Userspace is
4587 * responsible for ensuring nested SVM and SMIs are mutually exclusive.
4590 if (!guest_cpuid_has(vcpu, X86_FEATURE_LM))
4593 smram->smram64.svm_guest_flag = 1;
4594 smram->smram64.svm_guest_vmcb_gpa = svm->nested.vmcb12_gpa;
4596 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4597 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4598 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4600 ret = nested_svm_simple_vmexit(svm, SVM_EXIT_SW);
4605 * KVM uses VMCB01 to store L1 host state while L2 runs but
4606 * VMCB01 is going to be used during SMM and thus the state will
4607 * be lost. Temporary save non-VMLOAD/VMSAVE state to the host save
4608 * area pointed to by MSR_VM_HSAVE_PA. APM guarantees that the
4609 * format of the area is identical to guest save area offsetted
4610 * by 0x400 (matches the offset of 'struct vmcb_save_area'
4611 * within 'struct vmcb'). Note: HSAVE area may also be used by
4612 * L1 hypervisor to save additional host context (e.g. KVM does
4613 * that, see svm_prepare_switch_to_guest()) which must be
4616 if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr), &map_save))
4619 BUILD_BUG_ON(offsetof(struct vmcb, save) != 0x400);
4621 svm_copy_vmrun_state(map_save.hva + 0x400,
4622 &svm->vmcb01.ptr->save);
4624 kvm_vcpu_unmap(vcpu, &map_save, true);
4628 static int svm_leave_smm(struct kvm_vcpu *vcpu, const union kvm_smram *smram)
4630 struct vcpu_svm *svm = to_svm(vcpu);
4631 struct kvm_host_map map, map_save;
4632 struct vmcb *vmcb12;
4635 const struct kvm_smram_state_64 *smram64 = &smram->smram64;
4637 if (!guest_cpuid_has(vcpu, X86_FEATURE_LM))
4640 /* Non-zero if SMI arrived while vCPU was in guest mode. */
4641 if (!smram64->svm_guest_flag)
4644 if (!guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4647 if (!(smram64->efer & EFER_SVME))
4650 if (kvm_vcpu_map(vcpu, gpa_to_gfn(smram64->svm_guest_vmcb_gpa), &map))
4654 if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr), &map_save))
4657 if (svm_allocate_nested(svm))
4661 * Restore L1 host state from L1 HSAVE area as VMCB01 was
4662 * used during SMM (see svm_enter_smm())
4665 svm_copy_vmrun_state(&svm->vmcb01.ptr->save, map_save.hva + 0x400);
4668 * Enter the nested guest now
4671 vmcb_mark_all_dirty(svm->vmcb01.ptr);
4674 nested_copy_vmcb_control_to_cache(svm, &vmcb12->control);
4675 nested_copy_vmcb_save_to_cache(svm, &vmcb12->save);
4676 ret = enter_svm_guest_mode(vcpu, smram64->svm_guest_vmcb_gpa, vmcb12, false);
4681 svm->nested.nested_run_pending = 1;
4684 kvm_vcpu_unmap(vcpu, &map_save, true);
4686 kvm_vcpu_unmap(vcpu, &map, true);
4690 static void svm_enable_smi_window(struct kvm_vcpu *vcpu)
4692 struct vcpu_svm *svm = to_svm(vcpu);
4694 if (!gif_set(svm)) {
4696 svm_set_intercept(svm, INTERCEPT_STGI);
4697 /* STGI will cause a vm exit */
4699 /* We must be in SMM; RSM will cause a vmexit anyway. */
4704 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, int emul_type,
4705 void *insn, int insn_len)
4707 bool smep, smap, is_user;
4710 /* Emulation is always possible when KVM has access to all guest state. */
4711 if (!sev_guest(vcpu->kvm))
4714 /* #UD and #GP should never be intercepted for SEV guests. */
4715 WARN_ON_ONCE(emul_type & (EMULTYPE_TRAP_UD |
4716 EMULTYPE_TRAP_UD_FORCED |
4717 EMULTYPE_VMWARE_GP));
4720 * Emulation is impossible for SEV-ES guests as KVM doesn't have access
4721 * to guest register state.
4723 if (sev_es_guest(vcpu->kvm))
4727 * Emulation is possible if the instruction is already decoded, e.g.
4728 * when completing I/O after returning from userspace.
4730 if (emul_type & EMULTYPE_NO_DECODE)
4734 * Emulation is possible for SEV guests if and only if a prefilled
4735 * buffer containing the bytes of the intercepted instruction is
4736 * available. SEV guest memory is encrypted with a guest specific key
4737 * and cannot be decrypted by KVM, i.e. KVM would read cyphertext and
4740 * If KVM is NOT trying to simply skip an instruction, inject #UD if
4741 * KVM reached this point without an instruction buffer. In practice,
4742 * this path should never be hit by a well-behaved guest, e.g. KVM
4743 * doesn't intercept #UD or #GP for SEV guests, but this path is still
4744 * theoretically reachable, e.g. via unaccelerated fault-like AVIC
4745 * access, and needs to be handled by KVM to avoid putting the guest
4746 * into an infinite loop. Injecting #UD is somewhat arbitrary, but
4747 * its the least awful option given lack of insight into the guest.
4749 * If KVM is trying to skip an instruction, simply resume the guest.
4750 * If a #NPF occurs while the guest is vectoring an INT3/INTO, then KVM
4751 * will attempt to re-inject the INT3/INTO and skip the instruction.
4752 * In that scenario, retrying the INT3/INTO and hoping the guest will
4753 * make forward progress is the only option that has a chance of
4754 * success (and in practice it will work the vast majority of the time).
4756 if (unlikely(!insn)) {
4757 if (!(emul_type & EMULTYPE_SKIP))
4758 kvm_queue_exception(vcpu, UD_VECTOR);
4763 * Emulate for SEV guests if the insn buffer is not empty. The buffer
4764 * will be empty if the DecodeAssist microcode cannot fetch bytes for
4765 * the faulting instruction because the code fetch itself faulted, e.g.
4766 * the guest attempted to fetch from emulated MMIO or a guest page
4767 * table used to translate CS:RIP resides in emulated MMIO.
4769 if (likely(insn_len))
4773 * Detect and workaround Errata 1096 Fam_17h_00_0Fh.
4776 * When CPU raises #NPF on guest data access and vCPU CR4.SMAP=1, it is
4777 * possible that CPU microcode implementing DecodeAssist will fail to
4778 * read guest memory at CS:RIP and vmcb.GuestIntrBytes will incorrectly
4779 * be '0'. This happens because microcode reads CS:RIP using a _data_
4780 * loap uop with CPL=0 privileges. If the load hits a SMAP #PF, ucode
4781 * gives up and does not fill the instruction bytes buffer.
4783 * As above, KVM reaches this point iff the VM is an SEV guest, the CPU
4784 * supports DecodeAssist, a #NPF was raised, KVM's page fault handler
4785 * triggered emulation (e.g. for MMIO), and the CPU returned 0 in the
4786 * GuestIntrBytes field of the VMCB.
4788 * This does _not_ mean that the erratum has been encountered, as the
4789 * DecodeAssist will also fail if the load for CS:RIP hits a legitimate
4790 * #PF, e.g. if the guest attempt to execute from emulated MMIO and
4791 * encountered a reserved/not-present #PF.
4793 * To hit the erratum, the following conditions must be true:
4794 * 1. CR4.SMAP=1 (obviously).
4795 * 2. CR4.SMEP=0 || CPL=3. If SMEP=1 and CPL<3, the erratum cannot
4796 * have been hit as the guest would have encountered a SMEP
4797 * violation #PF, not a #NPF.
4798 * 3. The #NPF is not due to a code fetch, in which case failure to
4799 * retrieve the instruction bytes is legitimate (see abvoe).
4801 * In addition, don't apply the erratum workaround if the #NPF occurred
4802 * while translating guest page tables (see below).
4804 error_code = to_svm(vcpu)->vmcb->control.exit_info_1;
4805 if (error_code & (PFERR_GUEST_PAGE_MASK | PFERR_FETCH_MASK))
4808 smep = kvm_is_cr4_bit_set(vcpu, X86_CR4_SMEP);
4809 smap = kvm_is_cr4_bit_set(vcpu, X86_CR4_SMAP);
4810 is_user = svm_get_cpl(vcpu) == 3;
4811 if (smap && (!smep || is_user)) {
4812 pr_err_ratelimited("SEV Guest triggered AMD Erratum 1096\n");
4815 * If the fault occurred in userspace, arbitrarily inject #GP
4816 * to avoid killing the guest and to hopefully avoid confusing
4817 * the guest kernel too much, e.g. injecting #PF would not be
4818 * coherent with respect to the guest's page tables. Request
4819 * triple fault if the fault occurred in the kernel as there's
4820 * no fault that KVM can inject without confusing the guest.
4821 * In practice, the triple fault is moot as no sane SEV kernel
4822 * will execute from user memory while also running with SMAP=1.
4825 kvm_inject_gp(vcpu, 0);
4827 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
4832 * If the erratum was not hit, simply resume the guest and let it fault
4833 * again. While awful, e.g. the vCPU may get stuck in an infinite loop
4834 * if the fault is at CPL=0, it's the lesser of all evils. Exiting to
4835 * userspace will kill the guest, and letting the emulator read garbage
4836 * will yield random behavior and potentially corrupt the guest.
4838 * Simply resuming the guest is technically not a violation of the SEV
4839 * architecture. AMD's APM states that all code fetches and page table
4840 * accesses for SEV guest are encrypted, regardless of the C-Bit. The
4841 * APM also states that encrypted accesses to MMIO are "ignored", but
4842 * doesn't explicitly define "ignored", i.e. doing nothing and letting
4843 * the guest spin is technically "ignoring" the access.
4848 static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
4850 struct vcpu_svm *svm = to_svm(vcpu);
4852 return !gif_set(svm);
4855 static void svm_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
4857 if (!sev_es_guest(vcpu->kvm))
4858 return kvm_vcpu_deliver_sipi_vector(vcpu, vector);
4860 sev_vcpu_deliver_sipi_vector(vcpu, vector);
4863 static void svm_vm_destroy(struct kvm *kvm)
4865 avic_vm_destroy(kvm);
4866 sev_vm_destroy(kvm);
4869 static int svm_vm_init(struct kvm *kvm)
4871 if (!pause_filter_count || !pause_filter_thresh)
4872 kvm->arch.pause_in_guest = true;
4875 int ret = avic_vm_init(kvm);
4883 static struct kvm_x86_ops svm_x86_ops __initdata = {
4884 .name = KBUILD_MODNAME,
4886 .check_processor_compatibility = svm_check_processor_compat,
4888 .hardware_unsetup = svm_hardware_unsetup,
4889 .hardware_enable = svm_hardware_enable,
4890 .hardware_disable = svm_hardware_disable,
4891 .has_emulated_msr = svm_has_emulated_msr,
4893 .vcpu_create = svm_vcpu_create,
4894 .vcpu_free = svm_vcpu_free,
4895 .vcpu_reset = svm_vcpu_reset,
4897 .vm_size = sizeof(struct kvm_svm),
4898 .vm_init = svm_vm_init,
4899 .vm_destroy = svm_vm_destroy,
4901 .prepare_switch_to_guest = svm_prepare_switch_to_guest,
4902 .vcpu_load = svm_vcpu_load,
4903 .vcpu_put = svm_vcpu_put,
4904 .vcpu_blocking = avic_vcpu_blocking,
4905 .vcpu_unblocking = avic_vcpu_unblocking,
4907 .update_exception_bitmap = svm_update_exception_bitmap,
4908 .get_msr_feature = svm_get_msr_feature,
4909 .get_msr = svm_get_msr,
4910 .set_msr = svm_set_msr,
4911 .get_segment_base = svm_get_segment_base,
4912 .get_segment = svm_get_segment,
4913 .set_segment = svm_set_segment,
4914 .get_cpl = svm_get_cpl,
4915 .get_cs_db_l_bits = svm_get_cs_db_l_bits,
4916 .is_valid_cr0 = svm_is_valid_cr0,
4917 .set_cr0 = svm_set_cr0,
4918 .post_set_cr3 = sev_post_set_cr3,
4919 .is_valid_cr4 = svm_is_valid_cr4,
4920 .set_cr4 = svm_set_cr4,
4921 .set_efer = svm_set_efer,
4922 .get_idt = svm_get_idt,
4923 .set_idt = svm_set_idt,
4924 .get_gdt = svm_get_gdt,
4925 .set_gdt = svm_set_gdt,
4926 .set_dr7 = svm_set_dr7,
4927 .sync_dirty_debug_regs = svm_sync_dirty_debug_regs,
4928 .cache_reg = svm_cache_reg,
4929 .get_rflags = svm_get_rflags,
4930 .set_rflags = svm_set_rflags,
4931 .get_if_flag = svm_get_if_flag,
4933 .flush_tlb_all = svm_flush_tlb_all,
4934 .flush_tlb_current = svm_flush_tlb_current,
4935 .flush_tlb_gva = svm_flush_tlb_gva,
4936 .flush_tlb_guest = svm_flush_tlb_asid,
4938 .vcpu_pre_run = svm_vcpu_pre_run,
4939 .vcpu_run = svm_vcpu_run,
4940 .handle_exit = svm_handle_exit,
4941 .skip_emulated_instruction = svm_skip_emulated_instruction,
4942 .update_emulated_instruction = NULL,
4943 .set_interrupt_shadow = svm_set_interrupt_shadow,
4944 .get_interrupt_shadow = svm_get_interrupt_shadow,
4945 .patch_hypercall = svm_patch_hypercall,
4946 .inject_irq = svm_inject_irq,
4947 .inject_nmi = svm_inject_nmi,
4948 .is_vnmi_pending = svm_is_vnmi_pending,
4949 .set_vnmi_pending = svm_set_vnmi_pending,
4950 .inject_exception = svm_inject_exception,
4951 .cancel_injection = svm_cancel_injection,
4952 .interrupt_allowed = svm_interrupt_allowed,
4953 .nmi_allowed = svm_nmi_allowed,
4954 .get_nmi_mask = svm_get_nmi_mask,
4955 .set_nmi_mask = svm_set_nmi_mask,
4956 .enable_nmi_window = svm_enable_nmi_window,
4957 .enable_irq_window = svm_enable_irq_window,
4958 .update_cr8_intercept = svm_update_cr8_intercept,
4959 .set_virtual_apic_mode = avic_refresh_virtual_apic_mode,
4960 .refresh_apicv_exec_ctrl = avic_refresh_apicv_exec_ctrl,
4961 .apicv_post_state_restore = avic_apicv_post_state_restore,
4962 .required_apicv_inhibits = AVIC_REQUIRED_APICV_INHIBITS,
4964 .get_exit_info = svm_get_exit_info,
4966 .vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
4968 .has_wbinvd_exit = svm_has_wbinvd_exit,
4970 .get_l2_tsc_offset = svm_get_l2_tsc_offset,
4971 .get_l2_tsc_multiplier = svm_get_l2_tsc_multiplier,
4972 .write_tsc_offset = svm_write_tsc_offset,
4973 .write_tsc_multiplier = svm_write_tsc_multiplier,
4975 .load_mmu_pgd = svm_load_mmu_pgd,
4977 .check_intercept = svm_check_intercept,
4978 .handle_exit_irqoff = svm_handle_exit_irqoff,
4980 .request_immediate_exit = __kvm_request_immediate_exit,
4982 .sched_in = svm_sched_in,
4984 .nested_ops = &svm_nested_ops,
4986 .deliver_interrupt = svm_deliver_interrupt,
4987 .pi_update_irte = avic_pi_update_irte,
4988 .setup_mce = svm_setup_mce,
4990 #ifdef CONFIG_KVM_SMM
4991 .smi_allowed = svm_smi_allowed,
4992 .enter_smm = svm_enter_smm,
4993 .leave_smm = svm_leave_smm,
4994 .enable_smi_window = svm_enable_smi_window,
4997 .mem_enc_ioctl = sev_mem_enc_ioctl,
4998 .mem_enc_register_region = sev_mem_enc_register_region,
4999 .mem_enc_unregister_region = sev_mem_enc_unregister_region,
5000 .guest_memory_reclaimed = sev_guest_memory_reclaimed,
5002 .vm_copy_enc_context_from = sev_vm_copy_enc_context_from,
5003 .vm_move_enc_context_from = sev_vm_move_enc_context_from,
5005 .can_emulate_instruction = svm_can_emulate_instruction,
5007 .apic_init_signal_blocked = svm_apic_init_signal_blocked,
5009 .msr_filter_changed = svm_msr_filter_changed,
5010 .complete_emulated_msr = svm_complete_emulated_msr,
5012 .vcpu_deliver_sipi_vector = svm_vcpu_deliver_sipi_vector,
5013 .vcpu_get_apicv_inhibit_reasons = avic_vcpu_get_apicv_inhibit_reasons,
5017 * The default MMIO mask is a single bit (excluding the present bit),
5018 * which could conflict with the memory encryption bit. Check for
5019 * memory encryption support and override the default MMIO mask if
5020 * memory encryption is enabled.
5022 static __init void svm_adjust_mmio_mask(void)
5024 unsigned int enc_bit, mask_bit;
5027 /* If there is no memory encryption support, use existing mask */
5028 if (cpuid_eax(0x80000000) < 0x8000001f)
5031 /* If memory encryption is not enabled, use existing mask */
5032 rdmsrl(MSR_AMD64_SYSCFG, msr);
5033 if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
5036 enc_bit = cpuid_ebx(0x8000001f) & 0x3f;
5037 mask_bit = boot_cpu_data.x86_phys_bits;
5039 /* Increment the mask bit if it is the same as the encryption bit */
5040 if (enc_bit == mask_bit)
5044 * If the mask bit location is below 52, then some bits above the
5045 * physical addressing limit will always be reserved, so use the
5046 * rsvd_bits() function to generate the mask. This mask, along with
5047 * the present bit, will be used to generate a page fault with
5050 * If the mask bit location is 52 (or above), then clear the mask.
5052 mask = (mask_bit < 52) ? rsvd_bits(mask_bit, 51) | PT_PRESENT_MASK : 0;
5054 kvm_mmu_set_mmio_spte_mask(mask, mask, PT_WRITABLE_MASK | PT_USER_MASK);
5057 static __init void svm_set_cpu_caps(void)
5061 kvm_caps.supported_perf_cap = 0;
5062 kvm_caps.supported_xss = 0;
5064 /* CPUID 0x80000001 and 0x8000000A (SVM features) */
5066 kvm_cpu_cap_set(X86_FEATURE_SVM);
5067 kvm_cpu_cap_set(X86_FEATURE_VMCBCLEAN);
5070 kvm_cpu_cap_set(X86_FEATURE_NRIPS);
5073 kvm_cpu_cap_set(X86_FEATURE_NPT);
5076 kvm_cpu_cap_set(X86_FEATURE_TSCRATEMSR);
5079 kvm_cpu_cap_set(X86_FEATURE_V_VMSAVE_VMLOAD);
5081 kvm_cpu_cap_set(X86_FEATURE_LBRV);
5083 if (boot_cpu_has(X86_FEATURE_PAUSEFILTER))
5084 kvm_cpu_cap_set(X86_FEATURE_PAUSEFILTER);
5086 if (boot_cpu_has(X86_FEATURE_PFTHRESHOLD))
5087 kvm_cpu_cap_set(X86_FEATURE_PFTHRESHOLD);
5090 kvm_cpu_cap_set(X86_FEATURE_VGIF);
5093 kvm_cpu_cap_set(X86_FEATURE_VNMI);
5095 /* Nested VM can receive #VMEXIT instead of triggering #GP */
5096 kvm_cpu_cap_set(X86_FEATURE_SVME_ADDR_CHK);
5099 /* CPUID 0x80000008 */
5100 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) ||
5101 boot_cpu_has(X86_FEATURE_AMD_SSBD))
5102 kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
5106 * Enumerate support for PERFCTR_CORE if and only if KVM has
5107 * access to enough counters to virtualize "core" support,
5108 * otherwise limit vPMU support to the legacy number of counters.
5110 if (kvm_pmu_cap.num_counters_gp < AMD64_NUM_COUNTERS_CORE)
5111 kvm_pmu_cap.num_counters_gp = min(AMD64_NUM_COUNTERS,
5112 kvm_pmu_cap.num_counters_gp);
5114 kvm_cpu_cap_check_and_set(X86_FEATURE_PERFCTR_CORE);
5116 if (kvm_pmu_cap.version != 2 ||
5117 !kvm_cpu_cap_has(X86_FEATURE_PERFCTR_CORE))
5118 kvm_cpu_cap_clear(X86_FEATURE_PERFMON_V2);
5121 /* CPUID 0x8000001F (SME/SEV features) */
5125 static __init int svm_hardware_setup(void)
5128 struct page *iopm_pages;
5131 unsigned int order = get_order(IOPM_SIZE);
5134 * NX is required for shadow paging and for NPT if the NX huge pages
5135 * mitigation is enabled.
5137 if (!boot_cpu_has(X86_FEATURE_NX)) {
5138 pr_err_ratelimited("NX (Execute Disable) not supported\n");
5141 kvm_enable_efer_bits(EFER_NX);
5143 iopm_pages = alloc_pages(GFP_KERNEL, order);
5148 iopm_va = page_address(iopm_pages);
5149 memset(iopm_va, 0xff, PAGE_SIZE * (1 << order));
5150 iopm_base = page_to_pfn(iopm_pages) << PAGE_SHIFT;
5152 init_msrpm_offsets();
5154 kvm_caps.supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS |
5155 XFEATURE_MASK_BNDCSR);
5157 if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
5158 kvm_enable_efer_bits(EFER_FFXSR);
5161 if (!boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
5162 tsc_scaling = false;
5164 pr_info("TSC scaling supported\n");
5165 kvm_caps.has_tsc_control = true;
5168 kvm_caps.max_tsc_scaling_ratio = SVM_TSC_RATIO_MAX;
5169 kvm_caps.tsc_scaling_ratio_frac_bits = 32;
5171 tsc_aux_uret_slot = kvm_add_user_return_msr(MSR_TSC_AUX);
5173 if (boot_cpu_has(X86_FEATURE_AUTOIBRS))
5174 kvm_enable_efer_bits(EFER_AUTOIBRS);
5176 /* Check for pause filtering support */
5177 if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
5178 pause_filter_count = 0;
5179 pause_filter_thresh = 0;
5180 } else if (!boot_cpu_has(X86_FEATURE_PFTHRESHOLD)) {
5181 pause_filter_thresh = 0;
5185 pr_info("Nested Virtualization enabled\n");
5186 kvm_enable_efer_bits(EFER_SVME | EFER_LMSLE);
5190 * KVM's MMU doesn't support using 2-level paging for itself, and thus
5191 * NPT isn't supported if the host is using 2-level paging since host
5192 * CR4 is unchanged on VMRUN.
5194 if (!IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_X86_PAE))
5195 npt_enabled = false;
5197 if (!boot_cpu_has(X86_FEATURE_NPT))
5198 npt_enabled = false;
5200 /* Force VM NPT level equal to the host's paging level */
5201 kvm_configure_mmu(npt_enabled, get_npt_level(),
5202 get_npt_level(), PG_LEVEL_1G);
5203 pr_info("Nested Paging %sabled\n", npt_enabled ? "en" : "dis");
5205 /* Setup shadow_me_value and shadow_me_mask */
5206 kvm_mmu_set_me_spte_mask(sme_me_mask, sme_me_mask);
5208 svm_adjust_mmio_mask();
5210 nrips = nrips && boot_cpu_has(X86_FEATURE_NRIPS);
5213 * Note, SEV setup consumes npt_enabled and enable_mmio_caching (which
5214 * may be modified by svm_adjust_mmio_mask()), as well as nrips.
5216 sev_hardware_setup();
5218 svm_hv_hardware_setup();
5220 for_each_possible_cpu(cpu) {
5221 r = svm_cpu_init(cpu);
5226 enable_apicv = avic = avic && avic_hardware_setup();
5228 if (!enable_apicv) {
5229 svm_x86_ops.vcpu_blocking = NULL;
5230 svm_x86_ops.vcpu_unblocking = NULL;
5231 svm_x86_ops.vcpu_get_apicv_inhibit_reasons = NULL;
5232 } else if (!x2avic_enabled) {
5233 svm_x86_ops.allow_apicv_in_x2apic_without_x2apic_virtualization = true;
5238 !boot_cpu_has(X86_FEATURE_V_VMSAVE_VMLOAD) ||
5239 !IS_ENABLED(CONFIG_X86_64)) {
5242 pr_info("Virtual VMLOAD VMSAVE supported\n");
5246 if (boot_cpu_has(X86_FEATURE_SVME_ADDR_CHK))
5247 svm_gp_erratum_intercept = false;
5250 if (!boot_cpu_has(X86_FEATURE_VGIF))
5253 pr_info("Virtual GIF supported\n");
5256 vnmi = vgif && vnmi && boot_cpu_has(X86_FEATURE_VNMI);
5258 pr_info("Virtual NMI enabled\n");
5261 svm_x86_ops.is_vnmi_pending = NULL;
5262 svm_x86_ops.set_vnmi_pending = NULL;
5267 if (!boot_cpu_has(X86_FEATURE_LBRV))
5270 pr_info("LBR virtualization supported\n");
5274 pr_info("PMU virtualization is disabled\n");
5279 * It seems that on AMD processors PTE's accessed bit is
5280 * being set by the CPU hardware before the NPF vmexit.
5281 * This is not expected behaviour and our tests fail because
5283 * A workaround here is to disable support for
5284 * GUEST_MAXPHYADDR < HOST_MAXPHYADDR if NPT is enabled.
5285 * In this case userspace can know if there is support using
5286 * KVM_CAP_SMALLER_MAXPHYADDR extension and decide how to handle
5288 * If future AMD CPU models change the behaviour described above,
5289 * this variable can be changed accordingly
5291 allow_smaller_maxphyaddr = !npt_enabled;
5296 svm_hardware_unsetup();
5301 static struct kvm_x86_init_ops svm_init_ops __initdata = {
5302 .hardware_setup = svm_hardware_setup,
5304 .runtime_ops = &svm_x86_ops,
5305 .pmu_ops = &amd_pmu_ops,
5308 static void __svm_exit(void)
5310 kvm_x86_vendor_exit();
5312 cpu_emergency_unregister_virt_callback(svm_emergency_disable);
5315 static int __init svm_init(void)
5319 __unused_size_checks();
5321 if (!kvm_is_svm_supported())
5324 r = kvm_x86_vendor_init(&svm_init_ops);
5328 cpu_emergency_register_virt_callback(svm_emergency_disable);
5331 * Common KVM initialization _must_ come last, after this, /dev/kvm is
5332 * exposed to userspace!
5334 r = kvm_init(sizeof(struct vcpu_svm), __alignof__(struct vcpu_svm),
5346 static void __exit svm_exit(void)
5352 module_init(svm_init)
5353 module_exit(svm_exit)