1 #define pr_fmt(fmt) "SVM: " fmt
3 #include <linux/kvm_host.h>
7 #include "kvm_cache_regs.h"
12 #include <linux/module.h>
13 #include <linux/mod_devicetable.h>
14 #include <linux/kernel.h>
15 #include <linux/vmalloc.h>
16 #include <linux/highmem.h>
17 #include <linux/amd-iommu.h>
18 #include <linux/sched.h>
19 #include <linux/trace_events.h>
20 #include <linux/slab.h>
21 #include <linux/hashtable.h>
22 #include <linux/objtool.h>
23 #include <linux/psp-sev.h>
24 #include <linux/file.h>
25 #include <linux/pagemap.h>
26 #include <linux/swap.h>
27 #include <linux/rwsem.h>
28 #include <linux/cc_platform.h>
31 #include <asm/perf_event.h>
32 #include <asm/tlbflush.h>
34 #include <asm/debugreg.h>
35 #include <asm/kvm_para.h>
36 #include <asm/irq_remapping.h>
37 #include <asm/spec-ctrl.h>
38 #include <asm/cpu_device_id.h>
39 #include <asm/traps.h>
40 #include <asm/fpu/api.h>
42 #include <asm/virtext.h>
48 #include "kvm_onhyperv.h"
49 #include "svm_onhyperv.h"
51 MODULE_AUTHOR("Qumranet");
52 MODULE_LICENSE("GPL");
55 static const struct x86_cpu_id svm_cpu_id[] = {
56 X86_MATCH_FEATURE(X86_FEATURE_SVM, NULL),
59 MODULE_DEVICE_TABLE(x86cpu, svm_cpu_id);
62 #define SEG_TYPE_LDT 2
63 #define SEG_TYPE_BUSY_TSS16 3
65 #define SVM_FEATURE_LBRV (1 << 1)
66 #define SVM_FEATURE_SVML (1 << 2)
67 #define SVM_FEATURE_TSC_RATE (1 << 4)
68 #define SVM_FEATURE_VMCB_CLEAN (1 << 5)
69 #define SVM_FEATURE_FLUSH_ASID (1 << 6)
70 #define SVM_FEATURE_DECODE_ASSIST (1 << 7)
71 #define SVM_FEATURE_PAUSE_FILTER (1 << 10)
73 #define DEBUGCTL_RESERVED_BITS (~(0x3fULL))
75 #define TSC_RATIO_RSVD 0xffffff0000000000ULL
76 #define TSC_RATIO_MIN 0x0000000000000001ULL
77 #define TSC_RATIO_MAX 0x000000ffffffffffULL
79 static bool erratum_383_found __read_mostly;
81 u32 msrpm_offsets[MSRPM_OFFSETS] __read_mostly;
84 * Set osvw_len to higher value when updated Revision Guides
85 * are published and we know what the new status bits are
87 static uint64_t osvw_len = 4, osvw_status;
89 static DEFINE_PER_CPU(u64, current_tsc_ratio);
90 #define TSC_RATIO_DEFAULT 0x0100000000ULL
92 static const struct svm_direct_access_msrs {
93 u32 index; /* Index of the MSR */
94 bool always; /* True if intercept is initially cleared */
95 } direct_access_msrs[MAX_DIRECT_ACCESS_MSRS] = {
96 { .index = MSR_STAR, .always = true },
97 { .index = MSR_IA32_SYSENTER_CS, .always = true },
98 { .index = MSR_IA32_SYSENTER_EIP, .always = false },
99 { .index = MSR_IA32_SYSENTER_ESP, .always = false },
101 { .index = MSR_GS_BASE, .always = true },
102 { .index = MSR_FS_BASE, .always = true },
103 { .index = MSR_KERNEL_GS_BASE, .always = true },
104 { .index = MSR_LSTAR, .always = true },
105 { .index = MSR_CSTAR, .always = true },
106 { .index = MSR_SYSCALL_MASK, .always = true },
108 { .index = MSR_IA32_SPEC_CTRL, .always = false },
109 { .index = MSR_IA32_PRED_CMD, .always = false },
110 { .index = MSR_IA32_LASTBRANCHFROMIP, .always = false },
111 { .index = MSR_IA32_LASTBRANCHTOIP, .always = false },
112 { .index = MSR_IA32_LASTINTFROMIP, .always = false },
113 { .index = MSR_IA32_LASTINTTOIP, .always = false },
114 { .index = MSR_EFER, .always = false },
115 { .index = MSR_IA32_CR_PAT, .always = false },
116 { .index = MSR_AMD64_SEV_ES_GHCB, .always = true },
117 { .index = MSR_INVALID, .always = false },
121 * These 2 parameters are used to config the controls for Pause-Loop Exiting:
122 * pause_filter_count: On processors that support Pause filtering(indicated
123 * by CPUID Fn8000_000A_EDX), the VMCB provides a 16 bit pause filter
124 * count value. On VMRUN this value is loaded into an internal counter.
125 * Each time a pause instruction is executed, this counter is decremented
126 * until it reaches zero at which time a #VMEXIT is generated if pause
127 * intercept is enabled. Refer to AMD APM Vol 2 Section 15.14.4 Pause
128 * Intercept Filtering for more details.
129 * This also indicate if ple logic enabled.
131 * pause_filter_thresh: In addition, some processor families support advanced
132 * pause filtering (indicated by CPUID Fn8000_000A_EDX) upper bound on
133 * the amount of time a guest is allowed to execute in a pause loop.
134 * In this mode, a 16-bit pause filter threshold field is added in the
135 * VMCB. The threshold value is a cycle count that is used to reset the
136 * pause counter. As with simple pause filtering, VMRUN loads the pause
137 * count value from VMCB into an internal counter. Then, on each pause
138 * instruction the hardware checks the elapsed number of cycles since
139 * the most recent pause instruction against the pause filter threshold.
140 * If the elapsed cycle count is greater than the pause filter threshold,
141 * then the internal pause count is reloaded from the VMCB and execution
142 * continues. If the elapsed cycle count is less than the pause filter
143 * threshold, then the internal pause count is decremented. If the count
144 * value is less than zero and PAUSE intercept is enabled, a #VMEXIT is
145 * triggered. If advanced pause filtering is supported and pause filter
146 * threshold field is set to zero, the filter will operate in the simpler,
150 static unsigned short pause_filter_thresh = KVM_DEFAULT_PLE_GAP;
151 module_param(pause_filter_thresh, ushort, 0444);
153 static unsigned short pause_filter_count = KVM_SVM_DEFAULT_PLE_WINDOW;
154 module_param(pause_filter_count, ushort, 0444);
156 /* Default doubles per-vcpu window every exit. */
157 static unsigned short pause_filter_count_grow = KVM_DEFAULT_PLE_WINDOW_GROW;
158 module_param(pause_filter_count_grow, ushort, 0444);
160 /* Default resets per-vcpu window every exit to pause_filter_count. */
161 static unsigned short pause_filter_count_shrink = KVM_DEFAULT_PLE_WINDOW_SHRINK;
162 module_param(pause_filter_count_shrink, ushort, 0444);
164 /* Default is to compute the maximum so we can never overflow. */
165 static unsigned short pause_filter_count_max = KVM_SVM_DEFAULT_PLE_WINDOW_MAX;
166 module_param(pause_filter_count_max, ushort, 0444);
169 * Use nested page tables by default. Note, NPT may get forced off by
170 * svm_hardware_setup() if it's unsupported by hardware or the host kernel.
172 bool npt_enabled = true;
173 module_param_named(npt, npt_enabled, bool, 0444);
175 /* allow nested virtualization in KVM/SVM */
176 static int nested = true;
177 module_param(nested, int, S_IRUGO);
179 /* enable/disable Next RIP Save */
180 static int nrips = true;
181 module_param(nrips, int, 0444);
183 /* enable/disable Virtual VMLOAD VMSAVE */
184 static int vls = true;
185 module_param(vls, int, 0444);
187 /* enable/disable Virtual GIF */
188 static int vgif = true;
189 module_param(vgif, int, 0444);
191 /* enable/disable LBR virtualization */
192 static int lbrv = true;
193 module_param(lbrv, int, 0444);
195 static int tsc_scaling = true;
196 module_param(tsc_scaling, int, 0444);
199 * enable / disable AVIC. Because the defaults differ for APICv
200 * support between VMX and SVM we cannot use module_param_named.
203 module_param(avic, bool, 0444);
205 bool __read_mostly dump_invalid_vmcb;
206 module_param(dump_invalid_vmcb, bool, 0644);
209 bool intercept_smi = true;
210 module_param(intercept_smi, bool, 0444);
213 static bool svm_gp_erratum_intercept = true;
215 static u8 rsm_ins_bytes[] = "\x0f\xaa";
217 static unsigned long iopm_base;
219 struct kvm_ldttss_desc {
222 unsigned base1:8, type:5, dpl:2, p:1;
223 unsigned limit1:4, zero0:3, g:1, base2:8;
226 } __attribute__((packed));
228 DEFINE_PER_CPU(struct svm_cpu_data *, svm_data);
231 * Only MSR_TSC_AUX is switched via the user return hook. EFER is switched via
232 * the VMCB, and the SYSCALL/SYSENTER MSRs are handled by VMLOAD/VMSAVE.
234 * RDTSCP and RDPID are not used in the kernel, specifically to allow KVM to
235 * defer the restoration of TSC_AUX until the CPU returns to userspace.
237 static int tsc_aux_uret_slot __read_mostly = -1;
239 static const u32 msrpm_ranges[] = {0, 0xc0000000, 0xc0010000};
241 #define NUM_MSR_MAPS ARRAY_SIZE(msrpm_ranges)
242 #define MSRS_RANGE_SIZE 2048
243 #define MSRS_IN_RANGE (MSRS_RANGE_SIZE * 8 / 2)
245 u32 svm_msrpm_offset(u32 msr)
250 for (i = 0; i < NUM_MSR_MAPS; i++) {
251 if (msr < msrpm_ranges[i] ||
252 msr >= msrpm_ranges[i] + MSRS_IN_RANGE)
255 offset = (msr - msrpm_ranges[i]) / 4; /* 4 msrs per u8 */
256 offset += (i * MSRS_RANGE_SIZE); /* add range offset */
258 /* Now we have the u8 offset - but need the u32 offset */
262 /* MSR not in any range */
266 static void svm_flush_tlb_current(struct kvm_vcpu *vcpu);
268 static int get_npt_level(void)
271 return pgtable_l5_enabled() ? PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
273 return PT32E_ROOT_LEVEL;
277 int svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)
279 struct vcpu_svm *svm = to_svm(vcpu);
280 u64 old_efer = vcpu->arch.efer;
281 vcpu->arch.efer = efer;
284 /* Shadow paging assumes NX to be available. */
287 if (!(efer & EFER_LMA))
291 if ((old_efer & EFER_SVME) != (efer & EFER_SVME)) {
292 if (!(efer & EFER_SVME)) {
293 svm_leave_nested(vcpu);
294 svm_set_gif(svm, true);
295 /* #GP intercept is still needed for vmware backdoor */
296 if (!enable_vmware_backdoor)
297 clr_exception_intercept(svm, GP_VECTOR);
300 * Free the nested guest state, unless we are in SMM.
301 * In this case we will return to the nested guest
302 * as soon as we leave SMM.
305 svm_free_nested(svm);
308 int ret = svm_allocate_nested(svm);
311 vcpu->arch.efer = old_efer;
316 * Never intercept #GP for SEV guests, KVM can't
317 * decrypt guest memory to workaround the erratum.
319 if (svm_gp_erratum_intercept && !sev_guest(vcpu->kvm))
320 set_exception_intercept(svm, GP_VECTOR);
324 svm->vmcb->save.efer = efer | EFER_SVME;
325 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
329 static int is_external_interrupt(u32 info)
331 info &= SVM_EVTINJ_TYPE_MASK | SVM_EVTINJ_VALID;
332 return info == (SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR);
335 static u32 svm_get_interrupt_shadow(struct kvm_vcpu *vcpu)
337 struct vcpu_svm *svm = to_svm(vcpu);
340 if (svm->vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK)
341 ret = KVM_X86_SHADOW_INT_STI | KVM_X86_SHADOW_INT_MOV_SS;
345 static void svm_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
347 struct vcpu_svm *svm = to_svm(vcpu);
350 svm->vmcb->control.int_state &= ~SVM_INTERRUPT_SHADOW_MASK;
352 svm->vmcb->control.int_state |= SVM_INTERRUPT_SHADOW_MASK;
356 static int svm_skip_emulated_instruction(struct kvm_vcpu *vcpu)
358 struct vcpu_svm *svm = to_svm(vcpu);
361 * SEV-ES does not expose the next RIP. The RIP update is controlled by
362 * the type of exit and the #VC handler in the guest.
364 if (sev_es_guest(vcpu->kvm))
367 if (nrips && svm->vmcb->control.next_rip != 0) {
368 WARN_ON_ONCE(!static_cpu_has(X86_FEATURE_NRIPS));
369 svm->next_rip = svm->vmcb->control.next_rip;
372 if (!svm->next_rip) {
373 if (!kvm_emulate_instruction(vcpu, EMULTYPE_SKIP))
376 kvm_rip_write(vcpu, svm->next_rip);
380 svm_set_interrupt_shadow(vcpu, 0);
385 static void svm_queue_exception(struct kvm_vcpu *vcpu)
387 struct vcpu_svm *svm = to_svm(vcpu);
388 unsigned nr = vcpu->arch.exception.nr;
389 bool has_error_code = vcpu->arch.exception.has_error_code;
390 u32 error_code = vcpu->arch.exception.error_code;
392 kvm_deliver_exception_payload(vcpu);
394 if (nr == BP_VECTOR && !nrips) {
395 unsigned long rip, old_rip = kvm_rip_read(vcpu);
398 * For guest debugging where we have to reinject #BP if some
399 * INT3 is guest-owned:
400 * Emulate nRIP by moving RIP forward. Will fail if injection
401 * raises a fault that is not intercepted. Still better than
402 * failing in all cases.
404 (void)svm_skip_emulated_instruction(vcpu);
405 rip = kvm_rip_read(vcpu);
406 svm->int3_rip = rip + svm->vmcb->save.cs.base;
407 svm->int3_injected = rip - old_rip;
410 svm->vmcb->control.event_inj = nr
412 | (has_error_code ? SVM_EVTINJ_VALID_ERR : 0)
413 | SVM_EVTINJ_TYPE_EXEPT;
414 svm->vmcb->control.event_inj_err = error_code;
417 static void svm_init_erratum_383(void)
423 if (!static_cpu_has_bug(X86_BUG_AMD_TLB_MMATCH))
426 /* Use _safe variants to not break nested virtualization */
427 val = native_read_msr_safe(MSR_AMD64_DC_CFG, &err);
433 low = lower_32_bits(val);
434 high = upper_32_bits(val);
436 native_write_msr_safe(MSR_AMD64_DC_CFG, low, high);
438 erratum_383_found = true;
441 static void svm_init_osvw(struct kvm_vcpu *vcpu)
444 * Guests should see errata 400 and 415 as fixed (assuming that
445 * HLT and IO instructions are intercepted).
447 vcpu->arch.osvw.length = (osvw_len >= 3) ? (osvw_len) : 3;
448 vcpu->arch.osvw.status = osvw_status & ~(6ULL);
451 * By increasing VCPU's osvw.length to 3 we are telling the guest that
452 * all osvw.status bits inside that length, including bit 0 (which is
453 * reserved for erratum 298), are valid. However, if host processor's
454 * osvw_len is 0 then osvw_status[0] carries no information. We need to
455 * be conservative here and therefore we tell the guest that erratum 298
456 * is present (because we really don't know).
458 if (osvw_len == 0 && boot_cpu_data.x86 == 0x10)
459 vcpu->arch.osvw.status |= 1;
462 static int has_svm(void)
466 if (!cpu_has_svm(&msg)) {
467 printk(KERN_INFO "has_svm: %s\n", msg);
471 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) {
472 pr_info("KVM is unsupported when running as an SEV guest\n");
479 static void svm_hardware_disable(void)
481 /* Make sure we clean up behind us */
483 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
487 amd_pmu_disable_virt();
490 static int svm_hardware_enable(void)
493 struct svm_cpu_data *sd;
495 struct desc_struct *gdt;
496 int me = raw_smp_processor_id();
498 rdmsrl(MSR_EFER, efer);
499 if (efer & EFER_SVME)
503 pr_err("%s: err EOPNOTSUPP on %d\n", __func__, me);
506 sd = per_cpu(svm_data, me);
508 pr_err("%s: svm_data is NULL on %d\n", __func__, me);
512 sd->asid_generation = 1;
513 sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
514 sd->next_asid = sd->max_asid + 1;
515 sd->min_asid = max_sev_asid + 1;
517 gdt = get_current_gdt_rw();
518 sd->tss_desc = (struct kvm_ldttss_desc *)(gdt + GDT_ENTRY_TSS);
520 wrmsrl(MSR_EFER, efer | EFER_SVME);
522 wrmsrl(MSR_VM_HSAVE_PA, __sme_page_pa(sd->save_area));
524 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
526 * Set the default value, even if we don't use TSC scaling
527 * to avoid having stale value in the msr
529 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
530 __this_cpu_write(current_tsc_ratio, TSC_RATIO_DEFAULT);
537 * Note that it is possible to have a system with mixed processor
538 * revisions and therefore different OSVW bits. If bits are not the same
539 * on different processors then choose the worst case (i.e. if erratum
540 * is present on one processor and not on another then assume that the
541 * erratum is present everywhere).
543 if (cpu_has(&boot_cpu_data, X86_FEATURE_OSVW)) {
544 uint64_t len, status = 0;
547 len = native_read_msr_safe(MSR_AMD64_OSVW_ID_LENGTH, &err);
549 status = native_read_msr_safe(MSR_AMD64_OSVW_STATUS,
553 osvw_status = osvw_len = 0;
557 osvw_status |= status;
558 osvw_status &= (1ULL << osvw_len) - 1;
561 osvw_status = osvw_len = 0;
563 svm_init_erratum_383();
565 amd_pmu_enable_virt();
570 static void svm_cpu_uninit(int cpu)
572 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
577 per_cpu(svm_data, cpu) = NULL;
578 kfree(sd->sev_vmcbs);
579 __free_page(sd->save_area);
583 static int svm_cpu_init(int cpu)
585 struct svm_cpu_data *sd;
588 sd = kzalloc(sizeof(struct svm_cpu_data), GFP_KERNEL);
592 sd->save_area = alloc_page(GFP_KERNEL | __GFP_ZERO);
596 ret = sev_cpu_init(sd);
600 per_cpu(svm_data, cpu) = sd;
605 __free_page(sd->save_area);
612 static int direct_access_msr_slot(u32 msr)
616 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++)
617 if (direct_access_msrs[i].index == msr)
623 static void set_shadow_msr_intercept(struct kvm_vcpu *vcpu, u32 msr, int read,
626 struct vcpu_svm *svm = to_svm(vcpu);
627 int slot = direct_access_msr_slot(msr);
632 /* Set the shadow bitmaps to the desired intercept states */
634 set_bit(slot, svm->shadow_msr_intercept.read);
636 clear_bit(slot, svm->shadow_msr_intercept.read);
639 set_bit(slot, svm->shadow_msr_intercept.write);
641 clear_bit(slot, svm->shadow_msr_intercept.write);
644 static bool valid_msr_intercept(u32 index)
646 return direct_access_msr_slot(index) != -ENOENT;
649 static bool msr_write_intercepted(struct kvm_vcpu *vcpu, u32 msr)
656 msrpm = is_guest_mode(vcpu) ? to_svm(vcpu)->nested.msrpm:
659 offset = svm_msrpm_offset(msr);
660 bit_write = 2 * (msr & 0x0f) + 1;
663 BUG_ON(offset == MSR_INVALID);
665 return !!test_bit(bit_write, &tmp);
668 static void set_msr_interception_bitmap(struct kvm_vcpu *vcpu, u32 *msrpm,
669 u32 msr, int read, int write)
671 struct vcpu_svm *svm = to_svm(vcpu);
672 u8 bit_read, bit_write;
677 * If this warning triggers extend the direct_access_msrs list at the
678 * beginning of the file
680 WARN_ON(!valid_msr_intercept(msr));
682 /* Enforce non allowed MSRs to trap */
683 if (read && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_READ))
686 if (write && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_WRITE))
689 offset = svm_msrpm_offset(msr);
690 bit_read = 2 * (msr & 0x0f);
691 bit_write = 2 * (msr & 0x0f) + 1;
694 BUG_ON(offset == MSR_INVALID);
696 read ? clear_bit(bit_read, &tmp) : set_bit(bit_read, &tmp);
697 write ? clear_bit(bit_write, &tmp) : set_bit(bit_write, &tmp);
701 svm_hv_vmcb_dirty_nested_enlightenments(vcpu);
702 svm->nested.force_msr_bitmap_recalc = true;
705 void set_msr_interception(struct kvm_vcpu *vcpu, u32 *msrpm, u32 msr,
708 set_shadow_msr_intercept(vcpu, msr, read, write);
709 set_msr_interception_bitmap(vcpu, msrpm, msr, read, write);
712 u32 *svm_vcpu_alloc_msrpm(void)
714 unsigned int order = get_order(MSRPM_SIZE);
715 struct page *pages = alloc_pages(GFP_KERNEL_ACCOUNT, order);
721 msrpm = page_address(pages);
722 memset(msrpm, 0xff, PAGE_SIZE * (1 << order));
727 void svm_vcpu_init_msrpm(struct kvm_vcpu *vcpu, u32 *msrpm)
731 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
732 if (!direct_access_msrs[i].always)
734 set_msr_interception(vcpu, msrpm, direct_access_msrs[i].index, 1, 1);
739 void svm_vcpu_free_msrpm(u32 *msrpm)
741 __free_pages(virt_to_page(msrpm), get_order(MSRPM_SIZE));
744 static void svm_msr_filter_changed(struct kvm_vcpu *vcpu)
746 struct vcpu_svm *svm = to_svm(vcpu);
750 * Set intercept permissions for all direct access MSRs again. They
751 * will automatically get filtered through the MSR filter, so we are
752 * back in sync after this.
754 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
755 u32 msr = direct_access_msrs[i].index;
756 u32 read = test_bit(i, svm->shadow_msr_intercept.read);
757 u32 write = test_bit(i, svm->shadow_msr_intercept.write);
759 set_msr_interception_bitmap(vcpu, svm->msrpm, msr, read, write);
763 static void add_msr_offset(u32 offset)
767 for (i = 0; i < MSRPM_OFFSETS; ++i) {
769 /* Offset already in list? */
770 if (msrpm_offsets[i] == offset)
773 /* Slot used by another offset? */
774 if (msrpm_offsets[i] != MSR_INVALID)
777 /* Add offset to list */
778 msrpm_offsets[i] = offset;
784 * If this BUG triggers the msrpm_offsets table has an overflow. Just
785 * increase MSRPM_OFFSETS in this case.
790 static void init_msrpm_offsets(void)
794 memset(msrpm_offsets, 0xff, sizeof(msrpm_offsets));
796 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
799 offset = svm_msrpm_offset(direct_access_msrs[i].index);
800 BUG_ON(offset == MSR_INVALID);
802 add_msr_offset(offset);
806 static void svm_enable_lbrv(struct kvm_vcpu *vcpu)
808 struct vcpu_svm *svm = to_svm(vcpu);
810 svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
811 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
812 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
813 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
814 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
817 static void svm_disable_lbrv(struct kvm_vcpu *vcpu)
819 struct vcpu_svm *svm = to_svm(vcpu);
821 svm->vmcb->control.virt_ext &= ~LBR_CTL_ENABLE_MASK;
822 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 0, 0);
823 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 0, 0);
824 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 0, 0);
825 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 0, 0);
828 void disable_nmi_singlestep(struct vcpu_svm *svm)
830 svm->nmi_singlestep = false;
832 if (!(svm->vcpu.guest_debug & KVM_GUESTDBG_SINGLESTEP)) {
833 /* Clear our flags if they were not set by the guest */
834 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
835 svm->vmcb->save.rflags &= ~X86_EFLAGS_TF;
836 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
837 svm->vmcb->save.rflags &= ~X86_EFLAGS_RF;
841 static void grow_ple_window(struct kvm_vcpu *vcpu)
843 struct vcpu_svm *svm = to_svm(vcpu);
844 struct vmcb_control_area *control = &svm->vmcb->control;
845 int old = control->pause_filter_count;
847 control->pause_filter_count = __grow_ple_window(old,
849 pause_filter_count_grow,
850 pause_filter_count_max);
852 if (control->pause_filter_count != old) {
853 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
854 trace_kvm_ple_window_update(vcpu->vcpu_id,
855 control->pause_filter_count, old);
859 static void shrink_ple_window(struct kvm_vcpu *vcpu)
861 struct vcpu_svm *svm = to_svm(vcpu);
862 struct vmcb_control_area *control = &svm->vmcb->control;
863 int old = control->pause_filter_count;
865 control->pause_filter_count =
866 __shrink_ple_window(old,
868 pause_filter_count_shrink,
870 if (control->pause_filter_count != old) {
871 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
872 trace_kvm_ple_window_update(vcpu->vcpu_id,
873 control->pause_filter_count, old);
877 static void svm_hardware_unsetup(void)
881 sev_hardware_unsetup();
883 for_each_possible_cpu(cpu)
886 __free_pages(pfn_to_page(iopm_base >> PAGE_SHIFT),
887 get_order(IOPM_SIZE));
891 static void init_seg(struct vmcb_seg *seg)
894 seg->attrib = SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK |
895 SVM_SELECTOR_WRITE_MASK; /* Read/Write Data Segment */
900 static void init_sys_seg(struct vmcb_seg *seg, uint32_t type)
903 seg->attrib = SVM_SELECTOR_P_MASK | type;
908 static u64 svm_get_l2_tsc_offset(struct kvm_vcpu *vcpu)
910 struct vcpu_svm *svm = to_svm(vcpu);
912 return svm->nested.ctl.tsc_offset;
915 static u64 svm_get_l2_tsc_multiplier(struct kvm_vcpu *vcpu)
917 struct vcpu_svm *svm = to_svm(vcpu);
919 return svm->tsc_ratio_msr;
922 static void svm_write_tsc_offset(struct kvm_vcpu *vcpu, u64 offset)
924 struct vcpu_svm *svm = to_svm(vcpu);
926 svm->vmcb01.ptr->control.tsc_offset = vcpu->arch.l1_tsc_offset;
927 svm->vmcb->control.tsc_offset = offset;
928 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
931 void svm_write_tsc_multiplier(struct kvm_vcpu *vcpu, u64 multiplier)
933 wrmsrl(MSR_AMD64_TSC_RATIO, multiplier);
936 /* Evaluate instruction intercepts that depend on guest CPUID features. */
937 static void svm_recalc_instruction_intercepts(struct kvm_vcpu *vcpu,
938 struct vcpu_svm *svm)
941 * Intercept INVPCID if shadow paging is enabled to sync/free shadow
942 * roots, or if INVPCID is disabled in the guest to inject #UD.
944 if (kvm_cpu_cap_has(X86_FEATURE_INVPCID)) {
946 !guest_cpuid_has(&svm->vcpu, X86_FEATURE_INVPCID))
947 svm_set_intercept(svm, INTERCEPT_INVPCID);
949 svm_clr_intercept(svm, INTERCEPT_INVPCID);
952 if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP)) {
953 if (guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP))
954 svm_clr_intercept(svm, INTERCEPT_RDTSCP);
956 svm_set_intercept(svm, INTERCEPT_RDTSCP);
960 static inline void init_vmcb_after_set_cpuid(struct kvm_vcpu *vcpu)
962 struct vcpu_svm *svm = to_svm(vcpu);
964 if (guest_cpuid_is_intel(vcpu)) {
966 * We must intercept SYSENTER_EIP and SYSENTER_ESP
967 * accesses because the processor only stores 32 bits.
968 * For the same reason we cannot use virtual VMLOAD/VMSAVE.
970 svm_set_intercept(svm, INTERCEPT_VMLOAD);
971 svm_set_intercept(svm, INTERCEPT_VMSAVE);
972 svm->vmcb->control.virt_ext &= ~VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
974 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 0, 0);
975 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 0, 0);
978 * If hardware supports Virtual VMLOAD VMSAVE then enable it
979 * in VMCB and clear intercepts to avoid #VMEXIT.
982 svm_clr_intercept(svm, INTERCEPT_VMLOAD);
983 svm_clr_intercept(svm, INTERCEPT_VMSAVE);
984 svm->vmcb->control.virt_ext |= VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
986 /* No need to intercept these MSRs */
987 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 1, 1);
988 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 1, 1);
992 static void init_vmcb(struct kvm_vcpu *vcpu)
994 struct vcpu_svm *svm = to_svm(vcpu);
995 struct vmcb_control_area *control = &svm->vmcb->control;
996 struct vmcb_save_area *save = &svm->vmcb->save;
998 svm_set_intercept(svm, INTERCEPT_CR0_READ);
999 svm_set_intercept(svm, INTERCEPT_CR3_READ);
1000 svm_set_intercept(svm, INTERCEPT_CR4_READ);
1001 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1002 svm_set_intercept(svm, INTERCEPT_CR3_WRITE);
1003 svm_set_intercept(svm, INTERCEPT_CR4_WRITE);
1004 if (!kvm_vcpu_apicv_active(vcpu))
1005 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
1007 set_dr_intercepts(svm);
1009 set_exception_intercept(svm, PF_VECTOR);
1010 set_exception_intercept(svm, UD_VECTOR);
1011 set_exception_intercept(svm, MC_VECTOR);
1012 set_exception_intercept(svm, AC_VECTOR);
1013 set_exception_intercept(svm, DB_VECTOR);
1015 * Guest access to VMware backdoor ports could legitimately
1016 * trigger #GP because of TSS I/O permission bitmap.
1017 * We intercept those #GP and allow access to them anyway
1018 * as VMware does. Don't intercept #GP for SEV guests as KVM can't
1019 * decrypt guest memory to decode the faulting instruction.
1021 if (enable_vmware_backdoor && !sev_guest(vcpu->kvm))
1022 set_exception_intercept(svm, GP_VECTOR);
1024 svm_set_intercept(svm, INTERCEPT_INTR);
1025 svm_set_intercept(svm, INTERCEPT_NMI);
1028 svm_set_intercept(svm, INTERCEPT_SMI);
1030 svm_set_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1031 svm_set_intercept(svm, INTERCEPT_RDPMC);
1032 svm_set_intercept(svm, INTERCEPT_CPUID);
1033 svm_set_intercept(svm, INTERCEPT_INVD);
1034 svm_set_intercept(svm, INTERCEPT_INVLPG);
1035 svm_set_intercept(svm, INTERCEPT_INVLPGA);
1036 svm_set_intercept(svm, INTERCEPT_IOIO_PROT);
1037 svm_set_intercept(svm, INTERCEPT_MSR_PROT);
1038 svm_set_intercept(svm, INTERCEPT_TASK_SWITCH);
1039 svm_set_intercept(svm, INTERCEPT_SHUTDOWN);
1040 svm_set_intercept(svm, INTERCEPT_VMRUN);
1041 svm_set_intercept(svm, INTERCEPT_VMMCALL);
1042 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1043 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1044 svm_set_intercept(svm, INTERCEPT_STGI);
1045 svm_set_intercept(svm, INTERCEPT_CLGI);
1046 svm_set_intercept(svm, INTERCEPT_SKINIT);
1047 svm_set_intercept(svm, INTERCEPT_WBINVD);
1048 svm_set_intercept(svm, INTERCEPT_XSETBV);
1049 svm_set_intercept(svm, INTERCEPT_RDPRU);
1050 svm_set_intercept(svm, INTERCEPT_RSM);
1052 if (!kvm_mwait_in_guest(vcpu->kvm)) {
1053 svm_set_intercept(svm, INTERCEPT_MONITOR);
1054 svm_set_intercept(svm, INTERCEPT_MWAIT);
1057 if (!kvm_hlt_in_guest(vcpu->kvm))
1058 svm_set_intercept(svm, INTERCEPT_HLT);
1060 control->iopm_base_pa = __sme_set(iopm_base);
1061 control->msrpm_base_pa = __sme_set(__pa(svm->msrpm));
1062 control->int_ctl = V_INTR_MASKING_MASK;
1064 init_seg(&save->es);
1065 init_seg(&save->ss);
1066 init_seg(&save->ds);
1067 init_seg(&save->fs);
1068 init_seg(&save->gs);
1070 save->cs.selector = 0xf000;
1071 save->cs.base = 0xffff0000;
1072 /* Executable/Readable Code Segment */
1073 save->cs.attrib = SVM_SELECTOR_READ_MASK | SVM_SELECTOR_P_MASK |
1074 SVM_SELECTOR_S_MASK | SVM_SELECTOR_CODE_MASK;
1075 save->cs.limit = 0xffff;
1077 save->gdtr.base = 0;
1078 save->gdtr.limit = 0xffff;
1079 save->idtr.base = 0;
1080 save->idtr.limit = 0xffff;
1082 init_sys_seg(&save->ldtr, SEG_TYPE_LDT);
1083 init_sys_seg(&save->tr, SEG_TYPE_BUSY_TSS16);
1086 /* Setup VMCB for Nested Paging */
1087 control->nested_ctl |= SVM_NESTED_CTL_NP_ENABLE;
1088 svm_clr_intercept(svm, INTERCEPT_INVLPG);
1089 clr_exception_intercept(svm, PF_VECTOR);
1090 svm_clr_intercept(svm, INTERCEPT_CR3_READ);
1091 svm_clr_intercept(svm, INTERCEPT_CR3_WRITE);
1092 save->g_pat = vcpu->arch.pat;
1095 svm->current_vmcb->asid_generation = 0;
1098 svm->nested.vmcb12_gpa = INVALID_GPA;
1099 svm->nested.last_vmcb12_gpa = INVALID_GPA;
1101 if (!kvm_pause_in_guest(vcpu->kvm)) {
1102 control->pause_filter_count = pause_filter_count;
1103 if (pause_filter_thresh)
1104 control->pause_filter_thresh = pause_filter_thresh;
1105 svm_set_intercept(svm, INTERCEPT_PAUSE);
1107 svm_clr_intercept(svm, INTERCEPT_PAUSE);
1110 svm_recalc_instruction_intercepts(vcpu, svm);
1113 * If the host supports V_SPEC_CTRL then disable the interception
1114 * of MSR_IA32_SPEC_CTRL.
1116 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
1117 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
1119 if (kvm_vcpu_apicv_active(vcpu))
1120 avic_init_vmcb(svm);
1123 svm_clr_intercept(svm, INTERCEPT_STGI);
1124 svm_clr_intercept(svm, INTERCEPT_CLGI);
1125 svm->vmcb->control.int_ctl |= V_GIF_ENABLE_MASK;
1128 if (sev_guest(vcpu->kvm)) {
1129 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
1130 clr_exception_intercept(svm, UD_VECTOR);
1132 if (sev_es_guest(vcpu->kvm)) {
1133 /* Perform SEV-ES specific VMCB updates */
1134 sev_es_init_vmcb(svm);
1138 svm_hv_init_vmcb(svm->vmcb);
1139 init_vmcb_after_set_cpuid(vcpu);
1141 vmcb_mark_all_dirty(svm->vmcb);
1146 static void __svm_vcpu_reset(struct kvm_vcpu *vcpu)
1148 struct vcpu_svm *svm = to_svm(vcpu);
1150 svm_vcpu_init_msrpm(vcpu, svm->msrpm);
1152 svm_init_osvw(vcpu);
1153 vcpu->arch.microcode_version = 0x01000065;
1154 svm->tsc_ratio_msr = kvm_default_tsc_scaling_ratio;
1156 if (sev_es_guest(vcpu->kvm))
1157 sev_es_vcpu_reset(svm);
1160 static void svm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
1162 struct vcpu_svm *svm = to_svm(vcpu);
1165 svm->virt_spec_ctrl = 0;
1170 __svm_vcpu_reset(vcpu);
1173 void svm_switch_vmcb(struct vcpu_svm *svm, struct kvm_vmcb_info *target_vmcb)
1175 svm->current_vmcb = target_vmcb;
1176 svm->vmcb = target_vmcb->ptr;
1179 static int svm_vcpu_create(struct kvm_vcpu *vcpu)
1181 struct vcpu_svm *svm;
1182 struct page *vmcb01_page;
1183 struct page *vmsa_page = NULL;
1186 BUILD_BUG_ON(offsetof(struct vcpu_svm, vcpu) != 0);
1190 vmcb01_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1194 if (sev_es_guest(vcpu->kvm)) {
1196 * SEV-ES guests require a separate VMSA page used to contain
1197 * the encrypted register state of the guest.
1199 vmsa_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1201 goto error_free_vmcb_page;
1204 * SEV-ES guests maintain an encrypted version of their FPU
1205 * state which is restored and saved on VMRUN and VMEXIT.
1206 * Mark vcpu->arch.guest_fpu->fpstate as scratch so it won't
1207 * do xsave/xrstor on it.
1209 fpstate_set_confidential(&vcpu->arch.guest_fpu);
1212 err = avic_init_vcpu(svm);
1214 goto error_free_vmsa_page;
1216 svm->msrpm = svm_vcpu_alloc_msrpm();
1219 goto error_free_vmsa_page;
1222 svm->vmcb01.ptr = page_address(vmcb01_page);
1223 svm->vmcb01.pa = __sme_set(page_to_pfn(vmcb01_page) << PAGE_SHIFT);
1224 svm_switch_vmcb(svm, &svm->vmcb01);
1227 svm->sev_es.vmsa = page_address(vmsa_page);
1229 svm->guest_state_loaded = false;
1233 error_free_vmsa_page:
1235 __free_page(vmsa_page);
1236 error_free_vmcb_page:
1237 __free_page(vmcb01_page);
1242 static void svm_clear_current_vmcb(struct vmcb *vmcb)
1246 for_each_online_cpu(i)
1247 cmpxchg(&per_cpu(svm_data, i)->current_vmcb, vmcb, NULL);
1250 static void svm_vcpu_free(struct kvm_vcpu *vcpu)
1252 struct vcpu_svm *svm = to_svm(vcpu);
1255 * The vmcb page can be recycled, causing a false negative in
1256 * svm_vcpu_load(). So, ensure that no logical CPU has this
1257 * vmcb page recorded as its current vmcb.
1259 svm_clear_current_vmcb(svm->vmcb);
1261 svm_free_nested(svm);
1263 sev_free_vcpu(vcpu);
1265 __free_page(pfn_to_page(__sme_clr(svm->vmcb01.pa) >> PAGE_SHIFT));
1266 __free_pages(virt_to_page(svm->msrpm), get_order(MSRPM_SIZE));
1269 static void svm_prepare_switch_to_guest(struct kvm_vcpu *vcpu)
1271 struct vcpu_svm *svm = to_svm(vcpu);
1272 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
1274 if (sev_es_guest(vcpu->kvm))
1275 sev_es_unmap_ghcb(svm);
1277 if (svm->guest_state_loaded)
1281 * Save additional host state that will be restored on VMEXIT (sev-es)
1282 * or subsequent vmload of host save area.
1284 vmsave(__sme_page_pa(sd->save_area));
1285 if (sev_es_guest(vcpu->kvm)) {
1286 struct vmcb_save_area *hostsa;
1287 hostsa = (struct vmcb_save_area *)(page_address(sd->save_area) + 0x400);
1289 sev_es_prepare_switch_to_guest(hostsa);
1293 u64 tsc_ratio = vcpu->arch.tsc_scaling_ratio;
1294 if (tsc_ratio != __this_cpu_read(current_tsc_ratio)) {
1295 __this_cpu_write(current_tsc_ratio, tsc_ratio);
1296 wrmsrl(MSR_AMD64_TSC_RATIO, tsc_ratio);
1300 if (likely(tsc_aux_uret_slot >= 0))
1301 kvm_set_user_return_msr(tsc_aux_uret_slot, svm->tsc_aux, -1ull);
1303 svm->guest_state_loaded = true;
1306 static void svm_prepare_host_switch(struct kvm_vcpu *vcpu)
1308 to_svm(vcpu)->guest_state_loaded = false;
1311 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1313 struct vcpu_svm *svm = to_svm(vcpu);
1314 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1316 if (sd->current_vmcb != svm->vmcb) {
1317 sd->current_vmcb = svm->vmcb;
1318 indirect_branch_prediction_barrier();
1320 if (kvm_vcpu_apicv_active(vcpu))
1321 __avic_vcpu_load(vcpu, cpu);
1324 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
1326 if (kvm_vcpu_apicv_active(vcpu))
1327 __avic_vcpu_put(vcpu);
1329 svm_prepare_host_switch(vcpu);
1331 ++vcpu->stat.host_state_reload;
1334 static unsigned long svm_get_rflags(struct kvm_vcpu *vcpu)
1336 struct vcpu_svm *svm = to_svm(vcpu);
1337 unsigned long rflags = svm->vmcb->save.rflags;
1339 if (svm->nmi_singlestep) {
1340 /* Hide our flags if they were not set by the guest */
1341 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
1342 rflags &= ~X86_EFLAGS_TF;
1343 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
1344 rflags &= ~X86_EFLAGS_RF;
1349 static void svm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
1351 if (to_svm(vcpu)->nmi_singlestep)
1352 rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
1355 * Any change of EFLAGS.VM is accompanied by a reload of SS
1356 * (caused by either a task switch or an inter-privilege IRET),
1357 * so we do not need to update the CPL here.
1359 to_svm(vcpu)->vmcb->save.rflags = rflags;
1362 static bool svm_get_if_flag(struct kvm_vcpu *vcpu)
1364 struct vmcb *vmcb = to_svm(vcpu)->vmcb;
1366 return sev_es_guest(vcpu->kvm)
1367 ? vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK
1368 : kvm_get_rflags(vcpu) & X86_EFLAGS_IF;
1371 static void svm_cache_reg(struct kvm_vcpu *vcpu, enum kvm_reg reg)
1373 kvm_register_mark_available(vcpu, reg);
1376 case VCPU_EXREG_PDPTR:
1378 * When !npt_enabled, mmu->pdptrs[] is already available since
1379 * it is always updated per SDM when moving to CRs.
1382 load_pdptrs(vcpu, kvm_read_cr3(vcpu));
1385 KVM_BUG_ON(1, vcpu->kvm);
1389 static void svm_set_vintr(struct vcpu_svm *svm)
1391 struct vmcb_control_area *control;
1394 * The following fields are ignored when AVIC is enabled
1396 WARN_ON(kvm_apicv_activated(svm->vcpu.kvm));
1398 svm_set_intercept(svm, INTERCEPT_VINTR);
1401 * This is just a dummy VINTR to actually cause a vmexit to happen.
1402 * Actual injection of virtual interrupts happens through EVENTINJ.
1404 control = &svm->vmcb->control;
1405 control->int_vector = 0x0;
1406 control->int_ctl &= ~V_INTR_PRIO_MASK;
1407 control->int_ctl |= V_IRQ_MASK |
1408 ((/*control->int_vector >> 4*/ 0xf) << V_INTR_PRIO_SHIFT);
1409 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1412 static void svm_clear_vintr(struct vcpu_svm *svm)
1414 svm_clr_intercept(svm, INTERCEPT_VINTR);
1416 /* Drop int_ctl fields related to VINTR injection. */
1417 svm->vmcb->control.int_ctl &= ~V_IRQ_INJECTION_BITS_MASK;
1418 if (is_guest_mode(&svm->vcpu)) {
1419 svm->vmcb01.ptr->control.int_ctl &= ~V_IRQ_INJECTION_BITS_MASK;
1421 WARN_ON((svm->vmcb->control.int_ctl & V_TPR_MASK) !=
1422 (svm->nested.ctl.int_ctl & V_TPR_MASK));
1424 svm->vmcb->control.int_ctl |= svm->nested.ctl.int_ctl &
1425 V_IRQ_INJECTION_BITS_MASK;
1427 svm->vmcb->control.int_vector = svm->nested.ctl.int_vector;
1430 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1433 static struct vmcb_seg *svm_seg(struct kvm_vcpu *vcpu, int seg)
1435 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1436 struct vmcb_save_area *save01 = &to_svm(vcpu)->vmcb01.ptr->save;
1439 case VCPU_SREG_CS: return &save->cs;
1440 case VCPU_SREG_DS: return &save->ds;
1441 case VCPU_SREG_ES: return &save->es;
1442 case VCPU_SREG_FS: return &save01->fs;
1443 case VCPU_SREG_GS: return &save01->gs;
1444 case VCPU_SREG_SS: return &save->ss;
1445 case VCPU_SREG_TR: return &save01->tr;
1446 case VCPU_SREG_LDTR: return &save01->ldtr;
1452 static u64 svm_get_segment_base(struct kvm_vcpu *vcpu, int seg)
1454 struct vmcb_seg *s = svm_seg(vcpu, seg);
1459 static void svm_get_segment(struct kvm_vcpu *vcpu,
1460 struct kvm_segment *var, int seg)
1462 struct vmcb_seg *s = svm_seg(vcpu, seg);
1464 var->base = s->base;
1465 var->limit = s->limit;
1466 var->selector = s->selector;
1467 var->type = s->attrib & SVM_SELECTOR_TYPE_MASK;
1468 var->s = (s->attrib >> SVM_SELECTOR_S_SHIFT) & 1;
1469 var->dpl = (s->attrib >> SVM_SELECTOR_DPL_SHIFT) & 3;
1470 var->present = (s->attrib >> SVM_SELECTOR_P_SHIFT) & 1;
1471 var->avl = (s->attrib >> SVM_SELECTOR_AVL_SHIFT) & 1;
1472 var->l = (s->attrib >> SVM_SELECTOR_L_SHIFT) & 1;
1473 var->db = (s->attrib >> SVM_SELECTOR_DB_SHIFT) & 1;
1476 * AMD CPUs circa 2014 track the G bit for all segments except CS.
1477 * However, the SVM spec states that the G bit is not observed by the
1478 * CPU, and some VMware virtual CPUs drop the G bit for all segments.
1479 * So let's synthesize a legal G bit for all segments, this helps
1480 * running KVM nested. It also helps cross-vendor migration, because
1481 * Intel's vmentry has a check on the 'G' bit.
1483 var->g = s->limit > 0xfffff;
1486 * AMD's VMCB does not have an explicit unusable field, so emulate it
1487 * for cross vendor migration purposes by "not present"
1489 var->unusable = !var->present;
1494 * Work around a bug where the busy flag in the tr selector
1504 * The accessed bit must always be set in the segment
1505 * descriptor cache, although it can be cleared in the
1506 * descriptor, the cached bit always remains at 1. Since
1507 * Intel has a check on this, set it here to support
1508 * cross-vendor migration.
1515 * On AMD CPUs sometimes the DB bit in the segment
1516 * descriptor is left as 1, although the whole segment has
1517 * been made unusable. Clear it here to pass an Intel VMX
1518 * entry check when cross vendor migrating.
1522 /* This is symmetric with svm_set_segment() */
1523 var->dpl = to_svm(vcpu)->vmcb->save.cpl;
1528 static int svm_get_cpl(struct kvm_vcpu *vcpu)
1530 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1535 static void svm_get_cs_db_l_bits(struct kvm_vcpu *vcpu, int *db, int *l)
1537 struct kvm_segment cs;
1539 svm_get_segment(vcpu, &cs, VCPU_SREG_CS);
1544 static void svm_get_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1546 struct vcpu_svm *svm = to_svm(vcpu);
1548 dt->size = svm->vmcb->save.idtr.limit;
1549 dt->address = svm->vmcb->save.idtr.base;
1552 static void svm_set_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1554 struct vcpu_svm *svm = to_svm(vcpu);
1556 svm->vmcb->save.idtr.limit = dt->size;
1557 svm->vmcb->save.idtr.base = dt->address ;
1558 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1561 static void svm_get_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1563 struct vcpu_svm *svm = to_svm(vcpu);
1565 dt->size = svm->vmcb->save.gdtr.limit;
1566 dt->address = svm->vmcb->save.gdtr.base;
1569 static void svm_set_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1571 struct vcpu_svm *svm = to_svm(vcpu);
1573 svm->vmcb->save.gdtr.limit = dt->size;
1574 svm->vmcb->save.gdtr.base = dt->address ;
1575 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1578 static void sev_post_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3)
1580 struct vcpu_svm *svm = to_svm(vcpu);
1583 * For guests that don't set guest_state_protected, the cr3 update is
1584 * handled via kvm_mmu_load() while entering the guest. For guests
1585 * that do (SEV-ES/SEV-SNP), the cr3 update needs to be written to
1586 * VMCB save area now, since the save area will become the initial
1587 * contents of the VMSA, and future VMCB save area updates won't be
1590 if (sev_es_guest(vcpu->kvm)) {
1591 svm->vmcb->save.cr3 = cr3;
1592 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1596 void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1598 struct vcpu_svm *svm = to_svm(vcpu);
1600 bool old_paging = is_paging(vcpu);
1602 #ifdef CONFIG_X86_64
1603 if (vcpu->arch.efer & EFER_LME && !vcpu->arch.guest_state_protected) {
1604 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
1605 vcpu->arch.efer |= EFER_LMA;
1606 svm->vmcb->save.efer |= EFER_LMA | EFER_LME;
1609 if (is_paging(vcpu) && !(cr0 & X86_CR0_PG)) {
1610 vcpu->arch.efer &= ~EFER_LMA;
1611 svm->vmcb->save.efer &= ~(EFER_LMA | EFER_LME);
1615 vcpu->arch.cr0 = cr0;
1618 hcr0 |= X86_CR0_PG | X86_CR0_WP;
1619 if (old_paging != is_paging(vcpu))
1620 svm_set_cr4(vcpu, kvm_read_cr4(vcpu));
1624 * re-enable caching here because the QEMU bios
1625 * does not do it - this results in some delay at
1628 if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
1629 hcr0 &= ~(X86_CR0_CD | X86_CR0_NW);
1631 svm->vmcb->save.cr0 = hcr0;
1632 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1635 * SEV-ES guests must always keep the CR intercepts cleared. CR
1636 * tracking is done using the CR write traps.
1638 if (sev_es_guest(vcpu->kvm))
1642 /* Selective CR0 write remains on. */
1643 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1644 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1646 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1647 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1651 static bool svm_is_valid_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1656 void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1658 unsigned long host_cr4_mce = cr4_read_shadow() & X86_CR4_MCE;
1659 unsigned long old_cr4 = vcpu->arch.cr4;
1661 if (npt_enabled && ((old_cr4 ^ cr4) & X86_CR4_PGE))
1662 svm_flush_tlb_current(vcpu);
1664 vcpu->arch.cr4 = cr4;
1668 if (!is_paging(vcpu))
1669 cr4 &= ~(X86_CR4_SMEP | X86_CR4_SMAP | X86_CR4_PKE);
1671 cr4 |= host_cr4_mce;
1672 to_svm(vcpu)->vmcb->save.cr4 = cr4;
1673 vmcb_mark_dirty(to_svm(vcpu)->vmcb, VMCB_CR);
1675 if ((cr4 ^ old_cr4) & (X86_CR4_OSXSAVE | X86_CR4_PKE))
1676 kvm_update_cpuid_runtime(vcpu);
1679 static void svm_set_segment(struct kvm_vcpu *vcpu,
1680 struct kvm_segment *var, int seg)
1682 struct vcpu_svm *svm = to_svm(vcpu);
1683 struct vmcb_seg *s = svm_seg(vcpu, seg);
1685 s->base = var->base;
1686 s->limit = var->limit;
1687 s->selector = var->selector;
1688 s->attrib = (var->type & SVM_SELECTOR_TYPE_MASK);
1689 s->attrib |= (var->s & 1) << SVM_SELECTOR_S_SHIFT;
1690 s->attrib |= (var->dpl & 3) << SVM_SELECTOR_DPL_SHIFT;
1691 s->attrib |= ((var->present & 1) && !var->unusable) << SVM_SELECTOR_P_SHIFT;
1692 s->attrib |= (var->avl & 1) << SVM_SELECTOR_AVL_SHIFT;
1693 s->attrib |= (var->l & 1) << SVM_SELECTOR_L_SHIFT;
1694 s->attrib |= (var->db & 1) << SVM_SELECTOR_DB_SHIFT;
1695 s->attrib |= (var->g & 1) << SVM_SELECTOR_G_SHIFT;
1698 * This is always accurate, except if SYSRET returned to a segment
1699 * with SS.DPL != 3. Intel does not have this quirk, and always
1700 * forces SS.DPL to 3 on sysret, so we ignore that case; fixing it
1701 * would entail passing the CPL to userspace and back.
1703 if (seg == VCPU_SREG_SS)
1704 /* This is symmetric with svm_get_segment() */
1705 svm->vmcb->save.cpl = (var->dpl & 3);
1707 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
1710 static void svm_update_exception_bitmap(struct kvm_vcpu *vcpu)
1712 struct vcpu_svm *svm = to_svm(vcpu);
1714 clr_exception_intercept(svm, BP_VECTOR);
1716 if (vcpu->guest_debug & KVM_GUESTDBG_ENABLE) {
1717 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP)
1718 set_exception_intercept(svm, BP_VECTOR);
1722 static void new_asid(struct vcpu_svm *svm, struct svm_cpu_data *sd)
1724 if (sd->next_asid > sd->max_asid) {
1725 ++sd->asid_generation;
1726 sd->next_asid = sd->min_asid;
1727 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
1728 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1731 svm->current_vmcb->asid_generation = sd->asid_generation;
1732 svm->asid = sd->next_asid++;
1735 static void svm_set_dr6(struct vcpu_svm *svm, unsigned long value)
1737 struct vmcb *vmcb = svm->vmcb;
1739 if (svm->vcpu.arch.guest_state_protected)
1742 if (unlikely(value != vmcb->save.dr6)) {
1743 vmcb->save.dr6 = value;
1744 vmcb_mark_dirty(vmcb, VMCB_DR);
1748 static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
1750 struct vcpu_svm *svm = to_svm(vcpu);
1752 if (vcpu->arch.guest_state_protected)
1755 get_debugreg(vcpu->arch.db[0], 0);
1756 get_debugreg(vcpu->arch.db[1], 1);
1757 get_debugreg(vcpu->arch.db[2], 2);
1758 get_debugreg(vcpu->arch.db[3], 3);
1760 * We cannot reset svm->vmcb->save.dr6 to DR6_ACTIVE_LOW here,
1761 * because db_interception might need it. We can do it before vmentry.
1763 vcpu->arch.dr6 = svm->vmcb->save.dr6;
1764 vcpu->arch.dr7 = svm->vmcb->save.dr7;
1765 vcpu->arch.switch_db_regs &= ~KVM_DEBUGREG_WONT_EXIT;
1766 set_dr_intercepts(svm);
1769 static void svm_set_dr7(struct kvm_vcpu *vcpu, unsigned long value)
1771 struct vcpu_svm *svm = to_svm(vcpu);
1773 if (vcpu->arch.guest_state_protected)
1776 svm->vmcb->save.dr7 = value;
1777 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
1780 static int pf_interception(struct kvm_vcpu *vcpu)
1782 struct vcpu_svm *svm = to_svm(vcpu);
1784 u64 fault_address = svm->vmcb->control.exit_info_2;
1785 u64 error_code = svm->vmcb->control.exit_info_1;
1787 return kvm_handle_page_fault(vcpu, error_code, fault_address,
1788 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1789 svm->vmcb->control.insn_bytes : NULL,
1790 svm->vmcb->control.insn_len);
1793 static int npf_interception(struct kvm_vcpu *vcpu)
1795 struct vcpu_svm *svm = to_svm(vcpu);
1797 u64 fault_address = svm->vmcb->control.exit_info_2;
1798 u64 error_code = svm->vmcb->control.exit_info_1;
1800 trace_kvm_page_fault(fault_address, error_code);
1801 return kvm_mmu_page_fault(vcpu, fault_address, error_code,
1802 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1803 svm->vmcb->control.insn_bytes : NULL,
1804 svm->vmcb->control.insn_len);
1807 static int db_interception(struct kvm_vcpu *vcpu)
1809 struct kvm_run *kvm_run = vcpu->run;
1810 struct vcpu_svm *svm = to_svm(vcpu);
1812 if (!(vcpu->guest_debug &
1813 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) &&
1814 !svm->nmi_singlestep) {
1815 u32 payload = svm->vmcb->save.dr6 ^ DR6_ACTIVE_LOW;
1816 kvm_queue_exception_p(vcpu, DB_VECTOR, payload);
1820 if (svm->nmi_singlestep) {
1821 disable_nmi_singlestep(svm);
1822 /* Make sure we check for pending NMIs upon entry */
1823 kvm_make_request(KVM_REQ_EVENT, vcpu);
1826 if (vcpu->guest_debug &
1827 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) {
1828 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1829 kvm_run->debug.arch.dr6 = svm->vmcb->save.dr6;
1830 kvm_run->debug.arch.dr7 = svm->vmcb->save.dr7;
1831 kvm_run->debug.arch.pc =
1832 svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1833 kvm_run->debug.arch.exception = DB_VECTOR;
1840 static int bp_interception(struct kvm_vcpu *vcpu)
1842 struct vcpu_svm *svm = to_svm(vcpu);
1843 struct kvm_run *kvm_run = vcpu->run;
1845 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1846 kvm_run->debug.arch.pc = svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1847 kvm_run->debug.arch.exception = BP_VECTOR;
1851 static int ud_interception(struct kvm_vcpu *vcpu)
1853 return handle_ud(vcpu);
1856 static int ac_interception(struct kvm_vcpu *vcpu)
1858 kvm_queue_exception_e(vcpu, AC_VECTOR, 0);
1862 static bool is_erratum_383(void)
1867 if (!erratum_383_found)
1870 value = native_read_msr_safe(MSR_IA32_MC0_STATUS, &err);
1874 /* Bit 62 may or may not be set for this mce */
1875 value &= ~(1ULL << 62);
1877 if (value != 0xb600000000010015ULL)
1880 /* Clear MCi_STATUS registers */
1881 for (i = 0; i < 6; ++i)
1882 native_write_msr_safe(MSR_IA32_MCx_STATUS(i), 0, 0);
1884 value = native_read_msr_safe(MSR_IA32_MCG_STATUS, &err);
1888 value &= ~(1ULL << 2);
1889 low = lower_32_bits(value);
1890 high = upper_32_bits(value);
1892 native_write_msr_safe(MSR_IA32_MCG_STATUS, low, high);
1895 /* Flush tlb to evict multi-match entries */
1901 static void svm_handle_mce(struct kvm_vcpu *vcpu)
1903 if (is_erratum_383()) {
1905 * Erratum 383 triggered. Guest state is corrupt so kill the
1908 pr_err("KVM: Guest triggered AMD Erratum 383\n");
1910 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
1916 * On an #MC intercept the MCE handler is not called automatically in
1917 * the host. So do it by hand here.
1919 kvm_machine_check();
1922 static int mc_interception(struct kvm_vcpu *vcpu)
1927 static int shutdown_interception(struct kvm_vcpu *vcpu)
1929 struct kvm_run *kvm_run = vcpu->run;
1930 struct vcpu_svm *svm = to_svm(vcpu);
1933 * The VM save area has already been encrypted so it
1934 * cannot be reinitialized - just terminate.
1936 if (sev_es_guest(vcpu->kvm))
1940 * VMCB is undefined after a SHUTDOWN intercept. INIT the vCPU to put
1941 * the VMCB in a known good state. Unfortuately, KVM doesn't have
1942 * KVM_MP_STATE_SHUTDOWN and can't add it without potentially breaking
1943 * userspace. At a platform view, INIT is acceptable behavior as
1944 * there exist bare metal platforms that automatically INIT the CPU
1945 * in response to shutdown.
1947 clear_page(svm->vmcb);
1948 kvm_vcpu_reset(vcpu, true);
1950 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
1954 static int io_interception(struct kvm_vcpu *vcpu)
1956 struct vcpu_svm *svm = to_svm(vcpu);
1957 u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
1958 int size, in, string;
1961 ++vcpu->stat.io_exits;
1962 string = (io_info & SVM_IOIO_STR_MASK) != 0;
1963 in = (io_info & SVM_IOIO_TYPE_MASK) != 0;
1964 port = io_info >> 16;
1965 size = (io_info & SVM_IOIO_SIZE_MASK) >> SVM_IOIO_SIZE_SHIFT;
1968 if (sev_es_guest(vcpu->kvm))
1969 return sev_es_string_io(svm, size, port, in);
1971 return kvm_emulate_instruction(vcpu, 0);
1974 svm->next_rip = svm->vmcb->control.exit_info_2;
1976 return kvm_fast_pio(vcpu, size, port, in);
1979 static int nmi_interception(struct kvm_vcpu *vcpu)
1984 static int smi_interception(struct kvm_vcpu *vcpu)
1989 static int intr_interception(struct kvm_vcpu *vcpu)
1991 ++vcpu->stat.irq_exits;
1995 static int vmload_vmsave_interception(struct kvm_vcpu *vcpu, bool vmload)
1997 struct vcpu_svm *svm = to_svm(vcpu);
1998 struct vmcb *vmcb12;
1999 struct kvm_host_map map;
2002 if (nested_svm_check_permissions(vcpu))
2005 ret = kvm_vcpu_map(vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2008 kvm_inject_gp(vcpu, 0);
2014 ret = kvm_skip_emulated_instruction(vcpu);
2017 svm_copy_vmloadsave_state(svm->vmcb, vmcb12);
2018 svm->sysenter_eip_hi = 0;
2019 svm->sysenter_esp_hi = 0;
2021 svm_copy_vmloadsave_state(vmcb12, svm->vmcb);
2024 kvm_vcpu_unmap(vcpu, &map, true);
2029 static int vmload_interception(struct kvm_vcpu *vcpu)
2031 return vmload_vmsave_interception(vcpu, true);
2034 static int vmsave_interception(struct kvm_vcpu *vcpu)
2036 return vmload_vmsave_interception(vcpu, false);
2039 static int vmrun_interception(struct kvm_vcpu *vcpu)
2041 if (nested_svm_check_permissions(vcpu))
2044 return nested_svm_vmrun(vcpu);
2054 /* Return NONE_SVM_INSTR if not SVM instrs, otherwise return decode result */
2055 static int svm_instr_opcode(struct kvm_vcpu *vcpu)
2057 struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt;
2059 if (ctxt->b != 0x1 || ctxt->opcode_len != 2)
2060 return NONE_SVM_INSTR;
2062 switch (ctxt->modrm) {
2063 case 0xd8: /* VMRUN */
2064 return SVM_INSTR_VMRUN;
2065 case 0xda: /* VMLOAD */
2066 return SVM_INSTR_VMLOAD;
2067 case 0xdb: /* VMSAVE */
2068 return SVM_INSTR_VMSAVE;
2073 return NONE_SVM_INSTR;
2076 static int emulate_svm_instr(struct kvm_vcpu *vcpu, int opcode)
2078 const int guest_mode_exit_codes[] = {
2079 [SVM_INSTR_VMRUN] = SVM_EXIT_VMRUN,
2080 [SVM_INSTR_VMLOAD] = SVM_EXIT_VMLOAD,
2081 [SVM_INSTR_VMSAVE] = SVM_EXIT_VMSAVE,
2083 int (*const svm_instr_handlers[])(struct kvm_vcpu *vcpu) = {
2084 [SVM_INSTR_VMRUN] = vmrun_interception,
2085 [SVM_INSTR_VMLOAD] = vmload_interception,
2086 [SVM_INSTR_VMSAVE] = vmsave_interception,
2088 struct vcpu_svm *svm = to_svm(vcpu);
2091 if (is_guest_mode(vcpu)) {
2092 /* Returns '1' or -errno on failure, '0' on success. */
2093 ret = nested_svm_simple_vmexit(svm, guest_mode_exit_codes[opcode]);
2098 return svm_instr_handlers[opcode](vcpu);
2102 * #GP handling code. Note that #GP can be triggered under the following two
2104 * 1) SVM VM-related instructions (VMRUN/VMSAVE/VMLOAD) that trigger #GP on
2105 * some AMD CPUs when EAX of these instructions are in the reserved memory
2106 * regions (e.g. SMM memory on host).
2107 * 2) VMware backdoor
2109 static int gp_interception(struct kvm_vcpu *vcpu)
2111 struct vcpu_svm *svm = to_svm(vcpu);
2112 u32 error_code = svm->vmcb->control.exit_info_1;
2115 /* Both #GP cases have zero error_code */
2119 /* Decode the instruction for usage later */
2120 if (x86_decode_emulated_instruction(vcpu, 0, NULL, 0) != EMULATION_OK)
2123 opcode = svm_instr_opcode(vcpu);
2125 if (opcode == NONE_SVM_INSTR) {
2126 if (!enable_vmware_backdoor)
2130 * VMware backdoor emulation on #GP interception only handles
2131 * IN{S}, OUT{S}, and RDPMC.
2133 if (!is_guest_mode(vcpu))
2134 return kvm_emulate_instruction(vcpu,
2135 EMULTYPE_VMWARE_GP | EMULTYPE_NO_DECODE);
2137 /* All SVM instructions expect page aligned RAX */
2138 if (svm->vmcb->save.rax & ~PAGE_MASK)
2141 return emulate_svm_instr(vcpu, opcode);
2145 kvm_queue_exception_e(vcpu, GP_VECTOR, error_code);
2149 void svm_set_gif(struct vcpu_svm *svm, bool value)
2153 * If VGIF is enabled, the STGI intercept is only added to
2154 * detect the opening of the SMI/NMI window; remove it now.
2155 * Likewise, clear the VINTR intercept, we will set it
2156 * again while processing KVM_REQ_EVENT if needed.
2158 if (vgif_enabled(svm))
2159 svm_clr_intercept(svm, INTERCEPT_STGI);
2160 if (svm_is_intercept(svm, INTERCEPT_VINTR))
2161 svm_clear_vintr(svm);
2164 if (svm->vcpu.arch.smi_pending ||
2165 svm->vcpu.arch.nmi_pending ||
2166 kvm_cpu_has_injectable_intr(&svm->vcpu))
2167 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2172 * After a CLGI no interrupts should come. But if vGIF is
2173 * in use, we still rely on the VINTR intercept (rather than
2174 * STGI) to detect an open interrupt window.
2176 if (!vgif_enabled(svm))
2177 svm_clear_vintr(svm);
2181 static int stgi_interception(struct kvm_vcpu *vcpu)
2185 if (nested_svm_check_permissions(vcpu))
2188 ret = kvm_skip_emulated_instruction(vcpu);
2189 svm_set_gif(to_svm(vcpu), true);
2193 static int clgi_interception(struct kvm_vcpu *vcpu)
2197 if (nested_svm_check_permissions(vcpu))
2200 ret = kvm_skip_emulated_instruction(vcpu);
2201 svm_set_gif(to_svm(vcpu), false);
2205 static int invlpga_interception(struct kvm_vcpu *vcpu)
2207 gva_t gva = kvm_rax_read(vcpu);
2208 u32 asid = kvm_rcx_read(vcpu);
2210 /* FIXME: Handle an address size prefix. */
2211 if (!is_long_mode(vcpu))
2214 trace_kvm_invlpga(to_svm(vcpu)->vmcb->save.rip, asid, gva);
2216 /* Let's treat INVLPGA the same as INVLPG (can be optimized!) */
2217 kvm_mmu_invlpg(vcpu, gva);
2219 return kvm_skip_emulated_instruction(vcpu);
2222 static int skinit_interception(struct kvm_vcpu *vcpu)
2224 trace_kvm_skinit(to_svm(vcpu)->vmcb->save.rip, kvm_rax_read(vcpu));
2226 kvm_queue_exception(vcpu, UD_VECTOR);
2230 static int task_switch_interception(struct kvm_vcpu *vcpu)
2232 struct vcpu_svm *svm = to_svm(vcpu);
2235 int int_type = svm->vmcb->control.exit_int_info &
2236 SVM_EXITINTINFO_TYPE_MASK;
2237 int int_vec = svm->vmcb->control.exit_int_info & SVM_EVTINJ_VEC_MASK;
2239 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_TYPE_MASK;
2241 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID;
2242 bool has_error_code = false;
2245 tss_selector = (u16)svm->vmcb->control.exit_info_1;
2247 if (svm->vmcb->control.exit_info_2 &
2248 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
2249 reason = TASK_SWITCH_IRET;
2250 else if (svm->vmcb->control.exit_info_2 &
2251 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
2252 reason = TASK_SWITCH_JMP;
2254 reason = TASK_SWITCH_GATE;
2256 reason = TASK_SWITCH_CALL;
2258 if (reason == TASK_SWITCH_GATE) {
2260 case SVM_EXITINTINFO_TYPE_NMI:
2261 vcpu->arch.nmi_injected = false;
2263 case SVM_EXITINTINFO_TYPE_EXEPT:
2264 if (svm->vmcb->control.exit_info_2 &
2265 (1ULL << SVM_EXITINFOSHIFT_TS_HAS_ERROR_CODE)) {
2266 has_error_code = true;
2268 (u32)svm->vmcb->control.exit_info_2;
2270 kvm_clear_exception_queue(vcpu);
2272 case SVM_EXITINTINFO_TYPE_INTR:
2273 kvm_clear_interrupt_queue(vcpu);
2280 if (reason != TASK_SWITCH_GATE ||
2281 int_type == SVM_EXITINTINFO_TYPE_SOFT ||
2282 (int_type == SVM_EXITINTINFO_TYPE_EXEPT &&
2283 (int_vec == OF_VECTOR || int_vec == BP_VECTOR))) {
2284 if (!svm_skip_emulated_instruction(vcpu))
2288 if (int_type != SVM_EXITINTINFO_TYPE_SOFT)
2291 return kvm_task_switch(vcpu, tss_selector, int_vec, reason,
2292 has_error_code, error_code);
2295 static int iret_interception(struct kvm_vcpu *vcpu)
2297 struct vcpu_svm *svm = to_svm(vcpu);
2299 ++vcpu->stat.nmi_window_exits;
2300 vcpu->arch.hflags |= HF_IRET_MASK;
2301 if (!sev_es_guest(vcpu->kvm)) {
2302 svm_clr_intercept(svm, INTERCEPT_IRET);
2303 svm->nmi_iret_rip = kvm_rip_read(vcpu);
2305 kvm_make_request(KVM_REQ_EVENT, vcpu);
2309 static int invlpg_interception(struct kvm_vcpu *vcpu)
2311 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2312 return kvm_emulate_instruction(vcpu, 0);
2314 kvm_mmu_invlpg(vcpu, to_svm(vcpu)->vmcb->control.exit_info_1);
2315 return kvm_skip_emulated_instruction(vcpu);
2318 static int emulate_on_interception(struct kvm_vcpu *vcpu)
2320 return kvm_emulate_instruction(vcpu, 0);
2323 static int rsm_interception(struct kvm_vcpu *vcpu)
2325 return kvm_emulate_instruction_from_buffer(vcpu, rsm_ins_bytes, 2);
2328 static bool check_selective_cr0_intercepted(struct kvm_vcpu *vcpu,
2331 struct vcpu_svm *svm = to_svm(vcpu);
2332 unsigned long cr0 = vcpu->arch.cr0;
2335 if (!is_guest_mode(vcpu) ||
2336 (!(vmcb12_is_intercept(&svm->nested.ctl, INTERCEPT_SELECTIVE_CR0))))
2339 cr0 &= ~SVM_CR0_SELECTIVE_MASK;
2340 val &= ~SVM_CR0_SELECTIVE_MASK;
2343 svm->vmcb->control.exit_code = SVM_EXIT_CR0_SEL_WRITE;
2344 ret = (nested_svm_exit_handled(svm) == NESTED_EXIT_DONE);
2350 #define CR_VALID (1ULL << 63)
2352 static int cr_interception(struct kvm_vcpu *vcpu)
2354 struct vcpu_svm *svm = to_svm(vcpu);
2359 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2360 return emulate_on_interception(vcpu);
2362 if (unlikely((svm->vmcb->control.exit_info_1 & CR_VALID) == 0))
2363 return emulate_on_interception(vcpu);
2365 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2366 if (svm->vmcb->control.exit_code == SVM_EXIT_CR0_SEL_WRITE)
2367 cr = SVM_EXIT_WRITE_CR0 - SVM_EXIT_READ_CR0;
2369 cr = svm->vmcb->control.exit_code - SVM_EXIT_READ_CR0;
2372 if (cr >= 16) { /* mov to cr */
2374 val = kvm_register_read(vcpu, reg);
2375 trace_kvm_cr_write(cr, val);
2378 if (!check_selective_cr0_intercepted(vcpu, val))
2379 err = kvm_set_cr0(vcpu, val);
2385 err = kvm_set_cr3(vcpu, val);
2388 err = kvm_set_cr4(vcpu, val);
2391 err = kvm_set_cr8(vcpu, val);
2394 WARN(1, "unhandled write to CR%d", cr);
2395 kvm_queue_exception(vcpu, UD_VECTOR);
2398 } else { /* mov from cr */
2401 val = kvm_read_cr0(vcpu);
2404 val = vcpu->arch.cr2;
2407 val = kvm_read_cr3(vcpu);
2410 val = kvm_read_cr4(vcpu);
2413 val = kvm_get_cr8(vcpu);
2416 WARN(1, "unhandled read from CR%d", cr);
2417 kvm_queue_exception(vcpu, UD_VECTOR);
2420 kvm_register_write(vcpu, reg, val);
2421 trace_kvm_cr_read(cr, val);
2423 return kvm_complete_insn_gp(vcpu, err);
2426 static int cr_trap(struct kvm_vcpu *vcpu)
2428 struct vcpu_svm *svm = to_svm(vcpu);
2429 unsigned long old_value, new_value;
2433 new_value = (unsigned long)svm->vmcb->control.exit_info_1;
2435 cr = svm->vmcb->control.exit_code - SVM_EXIT_CR0_WRITE_TRAP;
2438 old_value = kvm_read_cr0(vcpu);
2439 svm_set_cr0(vcpu, new_value);
2441 kvm_post_set_cr0(vcpu, old_value, new_value);
2444 old_value = kvm_read_cr4(vcpu);
2445 svm_set_cr4(vcpu, new_value);
2447 kvm_post_set_cr4(vcpu, old_value, new_value);
2450 ret = kvm_set_cr8(vcpu, new_value);
2453 WARN(1, "unhandled CR%d write trap", cr);
2454 kvm_queue_exception(vcpu, UD_VECTOR);
2458 return kvm_complete_insn_gp(vcpu, ret);
2461 static int dr_interception(struct kvm_vcpu *vcpu)
2463 struct vcpu_svm *svm = to_svm(vcpu);
2468 if (vcpu->guest_debug == 0) {
2470 * No more DR vmexits; force a reload of the debug registers
2471 * and reenter on this instruction. The next vmexit will
2472 * retrieve the full state of the debug registers.
2474 clr_dr_intercepts(svm);
2475 vcpu->arch.switch_db_regs |= KVM_DEBUGREG_WONT_EXIT;
2479 if (!boot_cpu_has(X86_FEATURE_DECODEASSISTS))
2480 return emulate_on_interception(vcpu);
2482 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2483 dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
2484 if (dr >= 16) { /* mov to DRn */
2486 val = kvm_register_read(vcpu, reg);
2487 err = kvm_set_dr(vcpu, dr, val);
2489 kvm_get_dr(vcpu, dr, &val);
2490 kvm_register_write(vcpu, reg, val);
2493 return kvm_complete_insn_gp(vcpu, err);
2496 static int cr8_write_interception(struct kvm_vcpu *vcpu)
2500 u8 cr8_prev = kvm_get_cr8(vcpu);
2501 /* instruction emulation calls kvm_set_cr8() */
2502 r = cr_interception(vcpu);
2503 if (lapic_in_kernel(vcpu))
2505 if (cr8_prev <= kvm_get_cr8(vcpu))
2507 vcpu->run->exit_reason = KVM_EXIT_SET_TPR;
2511 static int efer_trap(struct kvm_vcpu *vcpu)
2513 struct msr_data msr_info;
2517 * Clear the EFER_SVME bit from EFER. The SVM code always sets this
2518 * bit in svm_set_efer(), but __kvm_valid_efer() checks it against
2519 * whether the guest has X86_FEATURE_SVM - this avoids a failure if
2520 * the guest doesn't have X86_FEATURE_SVM.
2522 msr_info.host_initiated = false;
2523 msr_info.index = MSR_EFER;
2524 msr_info.data = to_svm(vcpu)->vmcb->control.exit_info_1 & ~EFER_SVME;
2525 ret = kvm_set_msr_common(vcpu, &msr_info);
2527 return kvm_complete_insn_gp(vcpu, ret);
2530 static int svm_get_msr_feature(struct kvm_msr_entry *msr)
2534 switch (msr->index) {
2535 case MSR_F10H_DECFG:
2536 if (boot_cpu_has(X86_FEATURE_LFENCE_RDTSC))
2537 msr->data |= MSR_F10H_DECFG_LFENCE_SERIALIZE;
2539 case MSR_IA32_PERF_CAPABILITIES:
2542 return KVM_MSR_RET_INVALID;
2548 static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
2550 struct vcpu_svm *svm = to_svm(vcpu);
2552 switch (msr_info->index) {
2553 case MSR_AMD64_TSC_RATIO:
2554 if (!msr_info->host_initiated && !svm->tsc_scaling_enabled)
2556 msr_info->data = svm->tsc_ratio_msr;
2559 msr_info->data = svm->vmcb01.ptr->save.star;
2561 #ifdef CONFIG_X86_64
2563 msr_info->data = svm->vmcb01.ptr->save.lstar;
2566 msr_info->data = svm->vmcb01.ptr->save.cstar;
2568 case MSR_KERNEL_GS_BASE:
2569 msr_info->data = svm->vmcb01.ptr->save.kernel_gs_base;
2571 case MSR_SYSCALL_MASK:
2572 msr_info->data = svm->vmcb01.ptr->save.sfmask;
2575 case MSR_IA32_SYSENTER_CS:
2576 msr_info->data = svm->vmcb01.ptr->save.sysenter_cs;
2578 case MSR_IA32_SYSENTER_EIP:
2579 msr_info->data = (u32)svm->vmcb01.ptr->save.sysenter_eip;
2580 if (guest_cpuid_is_intel(vcpu))
2581 msr_info->data |= (u64)svm->sysenter_eip_hi << 32;
2583 case MSR_IA32_SYSENTER_ESP:
2584 msr_info->data = svm->vmcb01.ptr->save.sysenter_esp;
2585 if (guest_cpuid_is_intel(vcpu))
2586 msr_info->data |= (u64)svm->sysenter_esp_hi << 32;
2589 msr_info->data = svm->tsc_aux;
2592 * Nobody will change the following 5 values in the VMCB so we can
2593 * safely return them on rdmsr. They will always be 0 until LBRV is
2596 case MSR_IA32_DEBUGCTLMSR:
2597 msr_info->data = svm->vmcb->save.dbgctl;
2599 case MSR_IA32_LASTBRANCHFROMIP:
2600 msr_info->data = svm->vmcb->save.br_from;
2602 case MSR_IA32_LASTBRANCHTOIP:
2603 msr_info->data = svm->vmcb->save.br_to;
2605 case MSR_IA32_LASTINTFROMIP:
2606 msr_info->data = svm->vmcb->save.last_excp_from;
2608 case MSR_IA32_LASTINTTOIP:
2609 msr_info->data = svm->vmcb->save.last_excp_to;
2611 case MSR_VM_HSAVE_PA:
2612 msr_info->data = svm->nested.hsave_msr;
2615 msr_info->data = svm->nested.vm_cr_msr;
2617 case MSR_IA32_SPEC_CTRL:
2618 if (!msr_info->host_initiated &&
2619 !guest_has_spec_ctrl_msr(vcpu))
2622 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2623 msr_info->data = svm->vmcb->save.spec_ctrl;
2625 msr_info->data = svm->spec_ctrl;
2627 case MSR_AMD64_VIRT_SPEC_CTRL:
2628 if (!msr_info->host_initiated &&
2629 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2632 msr_info->data = svm->virt_spec_ctrl;
2634 case MSR_F15H_IC_CFG: {
2638 family = guest_cpuid_family(vcpu);
2639 model = guest_cpuid_model(vcpu);
2641 if (family < 0 || model < 0)
2642 return kvm_get_msr_common(vcpu, msr_info);
2646 if (family == 0x15 &&
2647 (model >= 0x2 && model < 0x20))
2648 msr_info->data = 0x1E;
2651 case MSR_F10H_DECFG:
2652 msr_info->data = svm->msr_decfg;
2655 return kvm_get_msr_common(vcpu, msr_info);
2660 static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
2662 struct vcpu_svm *svm = to_svm(vcpu);
2663 if (!err || !sev_es_guest(vcpu->kvm) || WARN_ON_ONCE(!svm->sev_es.ghcb))
2664 return kvm_complete_insn_gp(vcpu, err);
2666 ghcb_set_sw_exit_info_1(svm->sev_es.ghcb, 1);
2667 ghcb_set_sw_exit_info_2(svm->sev_es.ghcb,
2669 SVM_EVTINJ_TYPE_EXEPT |
2674 static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
2676 struct vcpu_svm *svm = to_svm(vcpu);
2677 int svm_dis, chg_mask;
2679 if (data & ~SVM_VM_CR_VALID_MASK)
2682 chg_mask = SVM_VM_CR_VALID_MASK;
2684 if (svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK)
2685 chg_mask &= ~(SVM_VM_CR_SVM_LOCK_MASK | SVM_VM_CR_SVM_DIS_MASK);
2687 svm->nested.vm_cr_msr &= ~chg_mask;
2688 svm->nested.vm_cr_msr |= (data & chg_mask);
2690 svm_dis = svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK;
2692 /* check for svm_disable while efer.svme is set */
2693 if (svm_dis && (vcpu->arch.efer & EFER_SVME))
2699 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
2701 struct vcpu_svm *svm = to_svm(vcpu);
2704 u32 ecx = msr->index;
2705 u64 data = msr->data;
2707 case MSR_AMD64_TSC_RATIO:
2708 if (!msr->host_initiated && !svm->tsc_scaling_enabled)
2711 if (data & TSC_RATIO_RSVD)
2714 svm->tsc_ratio_msr = data;
2716 if (svm->tsc_scaling_enabled && is_guest_mode(vcpu))
2717 nested_svm_update_tsc_ratio_msr(vcpu);
2720 case MSR_IA32_CR_PAT:
2721 if (!kvm_mtrr_valid(vcpu, MSR_IA32_CR_PAT, data))
2723 vcpu->arch.pat = data;
2724 svm->vmcb01.ptr->save.g_pat = data;
2725 if (is_guest_mode(vcpu))
2726 nested_vmcb02_compute_g_pat(svm);
2727 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
2729 case MSR_IA32_SPEC_CTRL:
2730 if (!msr->host_initiated &&
2731 !guest_has_spec_ctrl_msr(vcpu))
2734 if (kvm_spec_ctrl_test_value(data))
2737 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2738 svm->vmcb->save.spec_ctrl = data;
2740 svm->spec_ctrl = data;
2746 * When it's written (to non-zero) for the first time, pass
2750 * The handling of the MSR bitmap for L2 guests is done in
2751 * nested_svm_vmrun_msrpm.
2752 * We update the L1 MSR bit as well since it will end up
2753 * touching the MSR anyway now.
2755 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
2757 case MSR_IA32_PRED_CMD:
2758 if (!msr->host_initiated &&
2759 !guest_has_pred_cmd_msr(vcpu))
2762 if (data & ~PRED_CMD_IBPB)
2764 if (!boot_cpu_has(X86_FEATURE_IBPB))
2769 wrmsrl(MSR_IA32_PRED_CMD, PRED_CMD_IBPB);
2770 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_PRED_CMD, 0, 1);
2772 case MSR_AMD64_VIRT_SPEC_CTRL:
2773 if (!msr->host_initiated &&
2774 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2777 if (data & ~SPEC_CTRL_SSBD)
2780 svm->virt_spec_ctrl = data;
2783 svm->vmcb01.ptr->save.star = data;
2785 #ifdef CONFIG_X86_64
2787 svm->vmcb01.ptr->save.lstar = data;
2790 svm->vmcb01.ptr->save.cstar = data;
2792 case MSR_KERNEL_GS_BASE:
2793 svm->vmcb01.ptr->save.kernel_gs_base = data;
2795 case MSR_SYSCALL_MASK:
2796 svm->vmcb01.ptr->save.sfmask = data;
2799 case MSR_IA32_SYSENTER_CS:
2800 svm->vmcb01.ptr->save.sysenter_cs = data;
2802 case MSR_IA32_SYSENTER_EIP:
2803 svm->vmcb01.ptr->save.sysenter_eip = (u32)data;
2805 * We only intercept the MSR_IA32_SYSENTER_{EIP|ESP} msrs
2806 * when we spoof an Intel vendor ID (for cross vendor migration).
2807 * In this case we use this intercept to track the high
2808 * 32 bit part of these msrs to support Intel's
2809 * implementation of SYSENTER/SYSEXIT.
2811 svm->sysenter_eip_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2813 case MSR_IA32_SYSENTER_ESP:
2814 svm->vmcb01.ptr->save.sysenter_esp = (u32)data;
2815 svm->sysenter_esp_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2819 * TSC_AUX is usually changed only during boot and never read
2820 * directly. Intercept TSC_AUX instead of exposing it to the
2821 * guest via direct_access_msrs, and switch it via user return.
2824 r = kvm_set_user_return_msr(tsc_aux_uret_slot, data, -1ull);
2829 svm->tsc_aux = data;
2831 case MSR_IA32_DEBUGCTLMSR:
2833 vcpu_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTL 0x%llx, nop\n",
2837 if (data & DEBUGCTL_RESERVED_BITS)
2840 svm->vmcb->save.dbgctl = data;
2841 vmcb_mark_dirty(svm->vmcb, VMCB_LBR);
2842 if (data & (1ULL<<0))
2843 svm_enable_lbrv(vcpu);
2845 svm_disable_lbrv(vcpu);
2847 case MSR_VM_HSAVE_PA:
2849 * Old kernels did not validate the value written to
2850 * MSR_VM_HSAVE_PA. Allow KVM_SET_MSR to set an invalid
2851 * value to allow live migrating buggy or malicious guests
2852 * originating from those kernels.
2854 if (!msr->host_initiated && !page_address_valid(vcpu, data))
2857 svm->nested.hsave_msr = data & PAGE_MASK;
2860 return svm_set_vm_cr(vcpu, data);
2862 vcpu_unimpl(vcpu, "unimplemented wrmsr: 0x%x data 0x%llx\n", ecx, data);
2864 case MSR_F10H_DECFG: {
2865 struct kvm_msr_entry msr_entry;
2867 msr_entry.index = msr->index;
2868 if (svm_get_msr_feature(&msr_entry))
2871 /* Check the supported bits */
2872 if (data & ~msr_entry.data)
2875 /* Don't allow the guest to change a bit, #GP */
2876 if (!msr->host_initiated && (data ^ msr_entry.data))
2879 svm->msr_decfg = data;
2883 return kvm_set_msr_common(vcpu, msr);
2888 static int msr_interception(struct kvm_vcpu *vcpu)
2890 if (to_svm(vcpu)->vmcb->control.exit_info_1)
2891 return kvm_emulate_wrmsr(vcpu);
2893 return kvm_emulate_rdmsr(vcpu);
2896 static int interrupt_window_interception(struct kvm_vcpu *vcpu)
2898 kvm_make_request(KVM_REQ_EVENT, vcpu);
2899 svm_clear_vintr(to_svm(vcpu));
2902 * For AVIC, the only reason to end up here is ExtINTs.
2903 * In this case AVIC was temporarily disabled for
2904 * requesting the IRQ window and we have to re-enable it.
2906 kvm_request_apicv_update(vcpu->kvm, true, APICV_INHIBIT_REASON_IRQWIN);
2908 ++vcpu->stat.irq_window_exits;
2912 static int pause_interception(struct kvm_vcpu *vcpu)
2917 * CPL is not made available for an SEV-ES guest, therefore
2918 * vcpu->arch.preempted_in_kernel can never be true. Just
2919 * set in_kernel to false as well.
2921 in_kernel = !sev_es_guest(vcpu->kvm) && svm_get_cpl(vcpu) == 0;
2923 if (!kvm_pause_in_guest(vcpu->kvm))
2924 grow_ple_window(vcpu);
2926 kvm_vcpu_on_spin(vcpu, in_kernel);
2927 return kvm_skip_emulated_instruction(vcpu);
2930 static int invpcid_interception(struct kvm_vcpu *vcpu)
2932 struct vcpu_svm *svm = to_svm(vcpu);
2936 if (!guest_cpuid_has(vcpu, X86_FEATURE_INVPCID)) {
2937 kvm_queue_exception(vcpu, UD_VECTOR);
2942 * For an INVPCID intercept:
2943 * EXITINFO1 provides the linear address of the memory operand.
2944 * EXITINFO2 provides the contents of the register operand.
2946 type = svm->vmcb->control.exit_info_2;
2947 gva = svm->vmcb->control.exit_info_1;
2949 return kvm_handle_invpcid(vcpu, type, gva);
2952 static int (*const svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {
2953 [SVM_EXIT_READ_CR0] = cr_interception,
2954 [SVM_EXIT_READ_CR3] = cr_interception,
2955 [SVM_EXIT_READ_CR4] = cr_interception,
2956 [SVM_EXIT_READ_CR8] = cr_interception,
2957 [SVM_EXIT_CR0_SEL_WRITE] = cr_interception,
2958 [SVM_EXIT_WRITE_CR0] = cr_interception,
2959 [SVM_EXIT_WRITE_CR3] = cr_interception,
2960 [SVM_EXIT_WRITE_CR4] = cr_interception,
2961 [SVM_EXIT_WRITE_CR8] = cr8_write_interception,
2962 [SVM_EXIT_READ_DR0] = dr_interception,
2963 [SVM_EXIT_READ_DR1] = dr_interception,
2964 [SVM_EXIT_READ_DR2] = dr_interception,
2965 [SVM_EXIT_READ_DR3] = dr_interception,
2966 [SVM_EXIT_READ_DR4] = dr_interception,
2967 [SVM_EXIT_READ_DR5] = dr_interception,
2968 [SVM_EXIT_READ_DR6] = dr_interception,
2969 [SVM_EXIT_READ_DR7] = dr_interception,
2970 [SVM_EXIT_WRITE_DR0] = dr_interception,
2971 [SVM_EXIT_WRITE_DR1] = dr_interception,
2972 [SVM_EXIT_WRITE_DR2] = dr_interception,
2973 [SVM_EXIT_WRITE_DR3] = dr_interception,
2974 [SVM_EXIT_WRITE_DR4] = dr_interception,
2975 [SVM_EXIT_WRITE_DR5] = dr_interception,
2976 [SVM_EXIT_WRITE_DR6] = dr_interception,
2977 [SVM_EXIT_WRITE_DR7] = dr_interception,
2978 [SVM_EXIT_EXCP_BASE + DB_VECTOR] = db_interception,
2979 [SVM_EXIT_EXCP_BASE + BP_VECTOR] = bp_interception,
2980 [SVM_EXIT_EXCP_BASE + UD_VECTOR] = ud_interception,
2981 [SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
2982 [SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
2983 [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
2984 [SVM_EXIT_EXCP_BASE + GP_VECTOR] = gp_interception,
2985 [SVM_EXIT_INTR] = intr_interception,
2986 [SVM_EXIT_NMI] = nmi_interception,
2987 [SVM_EXIT_SMI] = smi_interception,
2988 [SVM_EXIT_VINTR] = interrupt_window_interception,
2989 [SVM_EXIT_RDPMC] = kvm_emulate_rdpmc,
2990 [SVM_EXIT_CPUID] = kvm_emulate_cpuid,
2991 [SVM_EXIT_IRET] = iret_interception,
2992 [SVM_EXIT_INVD] = kvm_emulate_invd,
2993 [SVM_EXIT_PAUSE] = pause_interception,
2994 [SVM_EXIT_HLT] = kvm_emulate_halt,
2995 [SVM_EXIT_INVLPG] = invlpg_interception,
2996 [SVM_EXIT_INVLPGA] = invlpga_interception,
2997 [SVM_EXIT_IOIO] = io_interception,
2998 [SVM_EXIT_MSR] = msr_interception,
2999 [SVM_EXIT_TASK_SWITCH] = task_switch_interception,
3000 [SVM_EXIT_SHUTDOWN] = shutdown_interception,
3001 [SVM_EXIT_VMRUN] = vmrun_interception,
3002 [SVM_EXIT_VMMCALL] = kvm_emulate_hypercall,
3003 [SVM_EXIT_VMLOAD] = vmload_interception,
3004 [SVM_EXIT_VMSAVE] = vmsave_interception,
3005 [SVM_EXIT_STGI] = stgi_interception,
3006 [SVM_EXIT_CLGI] = clgi_interception,
3007 [SVM_EXIT_SKINIT] = skinit_interception,
3008 [SVM_EXIT_RDTSCP] = kvm_handle_invalid_op,
3009 [SVM_EXIT_WBINVD] = kvm_emulate_wbinvd,
3010 [SVM_EXIT_MONITOR] = kvm_emulate_monitor,
3011 [SVM_EXIT_MWAIT] = kvm_emulate_mwait,
3012 [SVM_EXIT_XSETBV] = kvm_emulate_xsetbv,
3013 [SVM_EXIT_RDPRU] = kvm_handle_invalid_op,
3014 [SVM_EXIT_EFER_WRITE_TRAP] = efer_trap,
3015 [SVM_EXIT_CR0_WRITE_TRAP] = cr_trap,
3016 [SVM_EXIT_CR4_WRITE_TRAP] = cr_trap,
3017 [SVM_EXIT_CR8_WRITE_TRAP] = cr_trap,
3018 [SVM_EXIT_INVPCID] = invpcid_interception,
3019 [SVM_EXIT_NPF] = npf_interception,
3020 [SVM_EXIT_RSM] = rsm_interception,
3021 [SVM_EXIT_AVIC_INCOMPLETE_IPI] = avic_incomplete_ipi_interception,
3022 [SVM_EXIT_AVIC_UNACCELERATED_ACCESS] = avic_unaccelerated_access_interception,
3023 [SVM_EXIT_VMGEXIT] = sev_handle_vmgexit,
3026 static void dump_vmcb(struct kvm_vcpu *vcpu)
3028 struct vcpu_svm *svm = to_svm(vcpu);
3029 struct vmcb_control_area *control = &svm->vmcb->control;
3030 struct vmcb_save_area *save = &svm->vmcb->save;
3031 struct vmcb_save_area *save01 = &svm->vmcb01.ptr->save;
3033 if (!dump_invalid_vmcb) {
3034 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3038 pr_err("VMCB %p, last attempted VMRUN on CPU %d\n",
3039 svm->current_vmcb->ptr, vcpu->arch.last_vmentry_cpu);
3040 pr_err("VMCB Control Area:\n");
3041 pr_err("%-20s%04x\n", "cr_read:", control->intercepts[INTERCEPT_CR] & 0xffff);
3042 pr_err("%-20s%04x\n", "cr_write:", control->intercepts[INTERCEPT_CR] >> 16);
3043 pr_err("%-20s%04x\n", "dr_read:", control->intercepts[INTERCEPT_DR] & 0xffff);
3044 pr_err("%-20s%04x\n", "dr_write:", control->intercepts[INTERCEPT_DR] >> 16);
3045 pr_err("%-20s%08x\n", "exceptions:", control->intercepts[INTERCEPT_EXCEPTION]);
3046 pr_err("%-20s%08x %08x\n", "intercepts:",
3047 control->intercepts[INTERCEPT_WORD3],
3048 control->intercepts[INTERCEPT_WORD4]);
3049 pr_err("%-20s%d\n", "pause filter count:", control->pause_filter_count);
3050 pr_err("%-20s%d\n", "pause filter threshold:",
3051 control->pause_filter_thresh);
3052 pr_err("%-20s%016llx\n", "iopm_base_pa:", control->iopm_base_pa);
3053 pr_err("%-20s%016llx\n", "msrpm_base_pa:", control->msrpm_base_pa);
3054 pr_err("%-20s%016llx\n", "tsc_offset:", control->tsc_offset);
3055 pr_err("%-20s%d\n", "asid:", control->asid);
3056 pr_err("%-20s%d\n", "tlb_ctl:", control->tlb_ctl);
3057 pr_err("%-20s%08x\n", "int_ctl:", control->int_ctl);
3058 pr_err("%-20s%08x\n", "int_vector:", control->int_vector);
3059 pr_err("%-20s%08x\n", "int_state:", control->int_state);
3060 pr_err("%-20s%08x\n", "exit_code:", control->exit_code);
3061 pr_err("%-20s%016llx\n", "exit_info1:", control->exit_info_1);
3062 pr_err("%-20s%016llx\n", "exit_info2:", control->exit_info_2);
3063 pr_err("%-20s%08x\n", "exit_int_info:", control->exit_int_info);
3064 pr_err("%-20s%08x\n", "exit_int_info_err:", control->exit_int_info_err);
3065 pr_err("%-20s%lld\n", "nested_ctl:", control->nested_ctl);
3066 pr_err("%-20s%016llx\n", "nested_cr3:", control->nested_cr3);
3067 pr_err("%-20s%016llx\n", "avic_vapic_bar:", control->avic_vapic_bar);
3068 pr_err("%-20s%016llx\n", "ghcb:", control->ghcb_gpa);
3069 pr_err("%-20s%08x\n", "event_inj:", control->event_inj);
3070 pr_err("%-20s%08x\n", "event_inj_err:", control->event_inj_err);
3071 pr_err("%-20s%lld\n", "virt_ext:", control->virt_ext);
3072 pr_err("%-20s%016llx\n", "next_rip:", control->next_rip);
3073 pr_err("%-20s%016llx\n", "avic_backing_page:", control->avic_backing_page);
3074 pr_err("%-20s%016llx\n", "avic_logical_id:", control->avic_logical_id);
3075 pr_err("%-20s%016llx\n", "avic_physical_id:", control->avic_physical_id);
3076 pr_err("%-20s%016llx\n", "vmsa_pa:", control->vmsa_pa);
3077 pr_err("VMCB State Save Area:\n");
3078 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3080 save->es.selector, save->es.attrib,
3081 save->es.limit, save->es.base);
3082 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3084 save->cs.selector, save->cs.attrib,
3085 save->cs.limit, save->cs.base);
3086 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3088 save->ss.selector, save->ss.attrib,
3089 save->ss.limit, save->ss.base);
3090 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3092 save->ds.selector, save->ds.attrib,
3093 save->ds.limit, save->ds.base);
3094 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3096 save01->fs.selector, save01->fs.attrib,
3097 save01->fs.limit, save01->fs.base);
3098 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3100 save01->gs.selector, save01->gs.attrib,
3101 save01->gs.limit, save01->gs.base);
3102 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3104 save->gdtr.selector, save->gdtr.attrib,
3105 save->gdtr.limit, save->gdtr.base);
3106 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3108 save01->ldtr.selector, save01->ldtr.attrib,
3109 save01->ldtr.limit, save01->ldtr.base);
3110 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3112 save->idtr.selector, save->idtr.attrib,
3113 save->idtr.limit, save->idtr.base);
3114 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3116 save01->tr.selector, save01->tr.attrib,
3117 save01->tr.limit, save01->tr.base);
3118 pr_err("cpl: %d efer: %016llx\n",
3119 save->cpl, save->efer);
3120 pr_err("%-15s %016llx %-13s %016llx\n",
3121 "cr0:", save->cr0, "cr2:", save->cr2);
3122 pr_err("%-15s %016llx %-13s %016llx\n",
3123 "cr3:", save->cr3, "cr4:", save->cr4);
3124 pr_err("%-15s %016llx %-13s %016llx\n",
3125 "dr6:", save->dr6, "dr7:", save->dr7);
3126 pr_err("%-15s %016llx %-13s %016llx\n",
3127 "rip:", save->rip, "rflags:", save->rflags);
3128 pr_err("%-15s %016llx %-13s %016llx\n",
3129 "rsp:", save->rsp, "rax:", save->rax);
3130 pr_err("%-15s %016llx %-13s %016llx\n",
3131 "star:", save01->star, "lstar:", save01->lstar);
3132 pr_err("%-15s %016llx %-13s %016llx\n",
3133 "cstar:", save01->cstar, "sfmask:", save01->sfmask);
3134 pr_err("%-15s %016llx %-13s %016llx\n",
3135 "kernel_gs_base:", save01->kernel_gs_base,
3136 "sysenter_cs:", save01->sysenter_cs);
3137 pr_err("%-15s %016llx %-13s %016llx\n",
3138 "sysenter_esp:", save01->sysenter_esp,
3139 "sysenter_eip:", save01->sysenter_eip);
3140 pr_err("%-15s %016llx %-13s %016llx\n",
3141 "gpat:", save->g_pat, "dbgctl:", save->dbgctl);
3142 pr_err("%-15s %016llx %-13s %016llx\n",
3143 "br_from:", save->br_from, "br_to:", save->br_to);
3144 pr_err("%-15s %016llx %-13s %016llx\n",
3145 "excp_from:", save->last_excp_from,
3146 "excp_to:", save->last_excp_to);
3149 static bool svm_check_exit_valid(u64 exit_code)
3151 return (exit_code < ARRAY_SIZE(svm_exit_handlers) &&
3152 svm_exit_handlers[exit_code]);
3155 static int svm_handle_invalid_exit(struct kvm_vcpu *vcpu, u64 exit_code)
3157 vcpu_unimpl(vcpu, "svm: unexpected exit reason 0x%llx\n", exit_code);
3159 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3160 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
3161 vcpu->run->internal.ndata = 2;
3162 vcpu->run->internal.data[0] = exit_code;
3163 vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
3167 int svm_invoke_exit_handler(struct kvm_vcpu *vcpu, u64 exit_code)
3169 if (!svm_check_exit_valid(exit_code))
3170 return svm_handle_invalid_exit(vcpu, exit_code);
3172 #ifdef CONFIG_RETPOLINE
3173 if (exit_code == SVM_EXIT_MSR)
3174 return msr_interception(vcpu);
3175 else if (exit_code == SVM_EXIT_VINTR)
3176 return interrupt_window_interception(vcpu);
3177 else if (exit_code == SVM_EXIT_INTR)
3178 return intr_interception(vcpu);
3179 else if (exit_code == SVM_EXIT_HLT)
3180 return kvm_emulate_halt(vcpu);
3181 else if (exit_code == SVM_EXIT_NPF)
3182 return npf_interception(vcpu);
3184 return svm_exit_handlers[exit_code](vcpu);
3187 static void svm_get_exit_info(struct kvm_vcpu *vcpu, u32 *reason,
3188 u64 *info1, u64 *info2,
3189 u32 *intr_info, u32 *error_code)
3191 struct vmcb_control_area *control = &to_svm(vcpu)->vmcb->control;
3193 *reason = control->exit_code;
3194 *info1 = control->exit_info_1;
3195 *info2 = control->exit_info_2;
3196 *intr_info = control->exit_int_info;
3197 if ((*intr_info & SVM_EXITINTINFO_VALID) &&
3198 (*intr_info & SVM_EXITINTINFO_VALID_ERR))
3199 *error_code = control->exit_int_info_err;
3204 static int svm_handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
3206 struct vcpu_svm *svm = to_svm(vcpu);
3207 struct kvm_run *kvm_run = vcpu->run;
3208 u32 exit_code = svm->vmcb->control.exit_code;
3210 trace_kvm_exit(vcpu, KVM_ISA_SVM);
3212 /* SEV-ES guests must use the CR write traps to track CR registers. */
3213 if (!sev_es_guest(vcpu->kvm)) {
3214 if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
3215 vcpu->arch.cr0 = svm->vmcb->save.cr0;
3217 vcpu->arch.cr3 = svm->vmcb->save.cr3;
3220 if (is_guest_mode(vcpu)) {
3223 trace_kvm_nested_vmexit(vcpu, KVM_ISA_SVM);
3225 vmexit = nested_svm_exit_special(svm);
3227 if (vmexit == NESTED_EXIT_CONTINUE)
3228 vmexit = nested_svm_exit_handled(svm);
3230 if (vmexit == NESTED_EXIT_DONE)
3234 if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
3235 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3236 kvm_run->fail_entry.hardware_entry_failure_reason
3237 = svm->vmcb->control.exit_code;
3238 kvm_run->fail_entry.cpu = vcpu->arch.last_vmentry_cpu;
3243 if (is_external_interrupt(svm->vmcb->control.exit_int_info) &&
3244 exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR &&
3245 exit_code != SVM_EXIT_NPF && exit_code != SVM_EXIT_TASK_SWITCH &&
3246 exit_code != SVM_EXIT_INTR && exit_code != SVM_EXIT_NMI)
3247 printk(KERN_ERR "%s: unexpected exit_int_info 0x%x "
3249 __func__, svm->vmcb->control.exit_int_info,
3252 if (exit_fastpath != EXIT_FASTPATH_NONE)
3255 return svm_invoke_exit_handler(vcpu, exit_code);
3258 static void reload_tss(struct kvm_vcpu *vcpu)
3260 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3262 sd->tss_desc->type = 9; /* available 32/64-bit TSS */
3266 static void pre_svm_run(struct kvm_vcpu *vcpu)
3268 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3269 struct vcpu_svm *svm = to_svm(vcpu);
3272 * If the previous vmrun of the vmcb occurred on a different physical
3273 * cpu, then mark the vmcb dirty and assign a new asid. Hardware's
3274 * vmcb clean bits are per logical CPU, as are KVM's asid assignments.
3276 if (unlikely(svm->current_vmcb->cpu != vcpu->cpu)) {
3277 svm->current_vmcb->asid_generation = 0;
3278 vmcb_mark_all_dirty(svm->vmcb);
3279 svm->current_vmcb->cpu = vcpu->cpu;
3282 if (sev_guest(vcpu->kvm))
3283 return pre_sev_run(svm, vcpu->cpu);
3285 /* FIXME: handle wraparound of asid_generation */
3286 if (svm->current_vmcb->asid_generation != sd->asid_generation)
3290 static void svm_inject_nmi(struct kvm_vcpu *vcpu)
3292 struct vcpu_svm *svm = to_svm(vcpu);
3294 svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
3295 vcpu->arch.hflags |= HF_NMI_MASK;
3296 if (!sev_es_guest(vcpu->kvm))
3297 svm_set_intercept(svm, INTERCEPT_IRET);
3298 ++vcpu->stat.nmi_injections;
3301 static void svm_inject_irq(struct kvm_vcpu *vcpu)
3303 struct vcpu_svm *svm = to_svm(vcpu);
3305 BUG_ON(!(gif_set(svm)));
3307 trace_kvm_inj_virq(vcpu->arch.interrupt.nr);
3308 ++vcpu->stat.irq_injections;
3310 svm->vmcb->control.event_inj = vcpu->arch.interrupt.nr |
3311 SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR;
3314 void svm_complete_interrupt_delivery(struct kvm_vcpu *vcpu, int delivery_mode,
3315 int trig_mode, int vector)
3318 * vcpu->arch.apicv_active must be read after vcpu->mode.
3319 * Pairs with smp_store_release in vcpu_enter_guest.
3321 bool in_guest_mode = (smp_load_acquire(&vcpu->mode) == IN_GUEST_MODE);
3323 if (!READ_ONCE(vcpu->arch.apicv_active)) {
3324 /* Process the interrupt via inject_pending_event */
3325 kvm_make_request(KVM_REQ_EVENT, vcpu);
3326 kvm_vcpu_kick(vcpu);
3330 trace_kvm_apicv_accept_irq(vcpu->vcpu_id, delivery_mode, trig_mode, vector);
3331 if (in_guest_mode) {
3333 * Signal the doorbell to tell hardware to inject the IRQ. If
3334 * the vCPU exits the guest before the doorbell chimes, hardware
3335 * will automatically process AVIC interrupts at the next VMRUN.
3337 avic_ring_doorbell(vcpu);
3340 * Wake the vCPU if it was blocking. KVM will then detect the
3341 * pending IRQ when checking if the vCPU has a wake event.
3343 kvm_vcpu_wake_up(vcpu);
3347 static void svm_deliver_interrupt(struct kvm_lapic *apic, int delivery_mode,
3348 int trig_mode, int vector)
3350 kvm_lapic_set_irr(vector, apic);
3353 * Pairs with the smp_mb_*() after setting vcpu->guest_mode in
3354 * vcpu_enter_guest() to ensure the write to the vIRR is ordered before
3355 * the read of guest_mode. This guarantees that either VMRUN will see
3356 * and process the new vIRR entry, or that svm_complete_interrupt_delivery
3357 * will signal the doorbell if the CPU has already entered the guest.
3359 smp_mb__after_atomic();
3360 svm_complete_interrupt_delivery(apic->vcpu, delivery_mode, trig_mode, vector);
3363 static void svm_update_cr8_intercept(struct kvm_vcpu *vcpu, int tpr, int irr)
3365 struct vcpu_svm *svm = to_svm(vcpu);
3368 * SEV-ES guests must always keep the CR intercepts cleared. CR
3369 * tracking is done using the CR write traps.
3371 if (sev_es_guest(vcpu->kvm))
3374 if (nested_svm_virtualize_tpr(vcpu))
3377 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3383 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
3386 bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
3388 struct vcpu_svm *svm = to_svm(vcpu);
3389 struct vmcb *vmcb = svm->vmcb;
3395 if (is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3398 ret = (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK) ||
3399 (vcpu->arch.hflags & HF_NMI_MASK);
3404 static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3406 struct vcpu_svm *svm = to_svm(vcpu);
3407 if (svm->nested.nested_run_pending)
3410 if (svm_nmi_blocked(vcpu))
3413 /* An NMI must not be injected into L2 if it's supposed to VM-Exit. */
3414 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3419 static bool svm_get_nmi_mask(struct kvm_vcpu *vcpu)
3421 return !!(vcpu->arch.hflags & HF_NMI_MASK);
3424 static void svm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
3426 struct vcpu_svm *svm = to_svm(vcpu);
3429 vcpu->arch.hflags |= HF_NMI_MASK;
3430 if (!sev_es_guest(vcpu->kvm))
3431 svm_set_intercept(svm, INTERCEPT_IRET);
3433 vcpu->arch.hflags &= ~HF_NMI_MASK;
3434 if (!sev_es_guest(vcpu->kvm))
3435 svm_clr_intercept(svm, INTERCEPT_IRET);
3439 bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
3441 struct vcpu_svm *svm = to_svm(vcpu);
3442 struct vmcb *vmcb = svm->vmcb;
3447 if (is_guest_mode(vcpu)) {
3448 /* As long as interrupts are being delivered... */
3449 if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
3450 ? !(svm->vmcb01.ptr->save.rflags & X86_EFLAGS_IF)
3451 : !(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3454 /* ... vmexits aren't blocked by the interrupt shadow */
3455 if (nested_exit_on_intr(svm))
3458 if (!svm_get_if_flag(vcpu))
3462 return (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK);
3465 static int svm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3467 struct vcpu_svm *svm = to_svm(vcpu);
3469 if (svm->nested.nested_run_pending)
3472 if (svm_interrupt_blocked(vcpu))
3476 * An IRQ must not be injected into L2 if it's supposed to VM-Exit,
3477 * e.g. if the IRQ arrived asynchronously after checking nested events.
3479 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_intr(svm))
3485 static void svm_enable_irq_window(struct kvm_vcpu *vcpu)
3487 struct vcpu_svm *svm = to_svm(vcpu);
3490 * In case GIF=0 we can't rely on the CPU to tell us when GIF becomes
3491 * 1, because that's a separate STGI/VMRUN intercept. The next time we
3492 * get that intercept, this function will be called again though and
3493 * we'll get the vintr intercept. However, if the vGIF feature is
3494 * enabled, the STGI interception will not occur. Enable the irq
3495 * window under the assumption that the hardware will set the GIF.
3497 if (vgif_enabled(svm) || gif_set(svm)) {
3499 * IRQ window is not needed when AVIC is enabled,
3500 * unless we have pending ExtINT since it cannot be injected
3501 * via AVIC. In such case, we need to temporarily disable AVIC,
3502 * and fallback to injecting IRQ via V_IRQ.
3504 kvm_request_apicv_update(vcpu->kvm, false, APICV_INHIBIT_REASON_IRQWIN);
3509 static void svm_enable_nmi_window(struct kvm_vcpu *vcpu)
3511 struct vcpu_svm *svm = to_svm(vcpu);
3513 if ((vcpu->arch.hflags & (HF_NMI_MASK | HF_IRET_MASK)) == HF_NMI_MASK)
3514 return; /* IRET will cause a vm exit */
3516 if (!gif_set(svm)) {
3517 if (vgif_enabled(svm))
3518 svm_set_intercept(svm, INTERCEPT_STGI);
3519 return; /* STGI will cause a vm exit */
3523 * Something prevents NMI from been injected. Single step over possible
3524 * problem (IRET or exception injection or interrupt shadow)
3526 svm->nmi_singlestep_guest_rflags = svm_get_rflags(vcpu);
3527 svm->nmi_singlestep = true;
3528 svm->vmcb->save.rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
3531 static void svm_flush_tlb_current(struct kvm_vcpu *vcpu)
3533 struct vcpu_svm *svm = to_svm(vcpu);
3536 * Flush only the current ASID even if the TLB flush was invoked via
3537 * kvm_flush_remote_tlbs(). Although flushing remote TLBs requires all
3538 * ASIDs to be flushed, KVM uses a single ASID for L1 and L2, and
3539 * unconditionally does a TLB flush on both nested VM-Enter and nested
3540 * VM-Exit (via kvm_mmu_reset_context()).
3542 if (static_cpu_has(X86_FEATURE_FLUSHBYASID))
3543 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3545 svm->current_vmcb->asid_generation--;
3548 static void svm_flush_tlb_gva(struct kvm_vcpu *vcpu, gva_t gva)
3550 struct vcpu_svm *svm = to_svm(vcpu);
3552 invlpga(gva, svm->vmcb->control.asid);
3555 static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)
3557 struct vcpu_svm *svm = to_svm(vcpu);
3559 if (nested_svm_virtualize_tpr(vcpu))
3562 if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
3563 int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
3564 kvm_set_cr8(vcpu, cr8);
3568 static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
3570 struct vcpu_svm *svm = to_svm(vcpu);
3573 if (nested_svm_virtualize_tpr(vcpu) ||
3574 kvm_vcpu_apicv_active(vcpu))
3577 cr8 = kvm_get_cr8(vcpu);
3578 svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
3579 svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
3582 static void svm_complete_interrupts(struct kvm_vcpu *vcpu)
3584 struct vcpu_svm *svm = to_svm(vcpu);
3587 u32 exitintinfo = svm->vmcb->control.exit_int_info;
3588 unsigned int3_injected = svm->int3_injected;
3590 svm->int3_injected = 0;
3593 * If we've made progress since setting HF_IRET_MASK, we've
3594 * executed an IRET and can allow NMI injection.
3596 if ((vcpu->arch.hflags & HF_IRET_MASK) &&
3597 (sev_es_guest(vcpu->kvm) ||
3598 kvm_rip_read(vcpu) != svm->nmi_iret_rip)) {
3599 vcpu->arch.hflags &= ~(HF_NMI_MASK | HF_IRET_MASK);
3600 kvm_make_request(KVM_REQ_EVENT, vcpu);
3603 vcpu->arch.nmi_injected = false;
3604 kvm_clear_exception_queue(vcpu);
3605 kvm_clear_interrupt_queue(vcpu);
3607 if (!(exitintinfo & SVM_EXITINTINFO_VALID))
3610 kvm_make_request(KVM_REQ_EVENT, vcpu);
3612 vector = exitintinfo & SVM_EXITINTINFO_VEC_MASK;
3613 type = exitintinfo & SVM_EXITINTINFO_TYPE_MASK;
3616 case SVM_EXITINTINFO_TYPE_NMI:
3617 vcpu->arch.nmi_injected = true;
3619 case SVM_EXITINTINFO_TYPE_EXEPT:
3621 * Never re-inject a #VC exception.
3623 if (vector == X86_TRAP_VC)
3627 * In case of software exceptions, do not reinject the vector,
3628 * but re-execute the instruction instead. Rewind RIP first
3629 * if we emulated INT3 before.
3631 if (kvm_exception_is_soft(vector)) {
3632 if (vector == BP_VECTOR && int3_injected &&
3633 kvm_is_linear_rip(vcpu, svm->int3_rip))
3635 kvm_rip_read(vcpu) - int3_injected);
3638 if (exitintinfo & SVM_EXITINTINFO_VALID_ERR) {
3639 u32 err = svm->vmcb->control.exit_int_info_err;
3640 kvm_requeue_exception_e(vcpu, vector, err);
3643 kvm_requeue_exception(vcpu, vector);
3645 case SVM_EXITINTINFO_TYPE_INTR:
3646 kvm_queue_interrupt(vcpu, vector, false);
3653 static void svm_cancel_injection(struct kvm_vcpu *vcpu)
3655 struct vcpu_svm *svm = to_svm(vcpu);
3656 struct vmcb_control_area *control = &svm->vmcb->control;
3658 control->exit_int_info = control->event_inj;
3659 control->exit_int_info_err = control->event_inj_err;
3660 control->event_inj = 0;
3661 svm_complete_interrupts(vcpu);
3664 static int svm_vcpu_pre_run(struct kvm_vcpu *vcpu)
3669 static fastpath_t svm_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
3671 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_MSR &&
3672 to_svm(vcpu)->vmcb->control.exit_info_1)
3673 return handle_fastpath_set_msr_irqoff(vcpu);
3675 return EXIT_FASTPATH_NONE;
3678 static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
3680 struct vcpu_svm *svm = to_svm(vcpu);
3681 unsigned long vmcb_pa = svm->current_vmcb->pa;
3683 guest_state_enter_irqoff();
3685 if (sev_es_guest(vcpu->kvm)) {
3686 __svm_sev_es_vcpu_run(vmcb_pa);
3688 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3691 * Use a single vmcb (vmcb01 because it's always valid) for
3692 * context switching guest state via VMLOAD/VMSAVE, that way
3693 * the state doesn't need to be copied between vmcb01 and
3694 * vmcb02 when switching vmcbs for nested virtualization.
3696 vmload(svm->vmcb01.pa);
3697 __svm_vcpu_run(vmcb_pa, (unsigned long *)&vcpu->arch.regs);
3698 vmsave(svm->vmcb01.pa);
3700 vmload(__sme_page_pa(sd->save_area));
3703 guest_state_exit_irqoff();
3706 static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
3708 struct vcpu_svm *svm = to_svm(vcpu);
3710 trace_kvm_entry(vcpu);
3712 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
3713 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
3714 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
3717 * Disable singlestep if we're injecting an interrupt/exception.
3718 * We don't want our modified rflags to be pushed on the stack where
3719 * we might not be able to easily reset them if we disabled NMI
3722 if (svm->nmi_singlestep && svm->vmcb->control.event_inj) {
3724 * Event injection happens before external interrupts cause a
3725 * vmexit and interrupts are disabled here, so smp_send_reschedule
3726 * is enough to force an immediate vmexit.
3728 disable_nmi_singlestep(svm);
3729 smp_send_reschedule(vcpu->cpu);
3734 sync_lapic_to_cr8(vcpu);
3736 if (unlikely(svm->asid != svm->vmcb->control.asid)) {
3737 svm->vmcb->control.asid = svm->asid;
3738 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
3740 svm->vmcb->save.cr2 = vcpu->arch.cr2;
3742 svm_hv_update_vp_id(svm->vmcb, vcpu);
3745 * Run with all-zero DR6 unless needed, so that we can get the exact cause
3748 if (unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT))
3749 svm_set_dr6(svm, vcpu->arch.dr6);
3751 svm_set_dr6(svm, DR6_ACTIVE_LOW);
3754 kvm_load_guest_xsave_state(vcpu);
3756 kvm_wait_lapic_expire(vcpu);
3759 * If this vCPU has touched SPEC_CTRL, restore the guest's value if
3760 * it's non-zero. Since vmentry is serialising on affected CPUs, there
3761 * is no need to worry about the conditional branch over the wrmsr
3762 * being speculatively taken.
3764 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3765 x86_spec_ctrl_set_guest(svm->spec_ctrl, svm->virt_spec_ctrl);
3767 svm_vcpu_enter_exit(vcpu);
3770 * We do not use IBRS in the kernel. If this vCPU has used the
3771 * SPEC_CTRL MSR it may have left it on; save the value and
3772 * turn it off. This is much more efficient than blindly adding
3773 * it to the atomic save/restore list. Especially as the former
3774 * (Saving guest MSRs on vmexit) doesn't even exist in KVM.
3776 * For non-nested case:
3777 * If the L01 MSR bitmap does not intercept the MSR, then we need to
3781 * If the L02 MSR bitmap does not intercept the MSR, then we need to
3784 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL) &&
3785 unlikely(!msr_write_intercepted(vcpu, MSR_IA32_SPEC_CTRL)))
3786 svm->spec_ctrl = native_read_msr(MSR_IA32_SPEC_CTRL);
3788 if (!sev_es_guest(vcpu->kvm))
3791 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3792 x86_spec_ctrl_restore_host(svm->spec_ctrl, svm->virt_spec_ctrl);
3794 if (!sev_es_guest(vcpu->kvm)) {
3795 vcpu->arch.cr2 = svm->vmcb->save.cr2;
3796 vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
3797 vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
3798 vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
3800 vcpu->arch.regs_dirty = 0;
3802 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3803 kvm_before_interrupt(vcpu, KVM_HANDLING_NMI);
3805 kvm_load_host_xsave_state(vcpu);
3808 /* Any pending NMI will happen here */
3810 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3811 kvm_after_interrupt(vcpu);
3813 sync_cr8_to_lapic(vcpu);
3816 if (is_guest_mode(vcpu)) {
3817 nested_sync_control_from_vmcb02(svm);
3819 /* Track VMRUNs that have made past consistency checking */
3820 if (svm->nested.nested_run_pending &&
3821 svm->vmcb->control.exit_code != SVM_EXIT_ERR)
3822 ++vcpu->stat.nested_run;
3824 svm->nested.nested_run_pending = 0;
3827 svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
3828 vmcb_mark_all_clean(svm->vmcb);
3830 /* if exit due to PF check for async PF */
3831 if (svm->vmcb->control.exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR)
3832 vcpu->arch.apf.host_apf_flags =
3833 kvm_read_and_reset_apf_flags();
3835 vcpu->arch.regs_avail &= ~SVM_REGS_LAZY_LOAD_SET;
3838 * We need to handle MC intercepts here before the vcpu has a chance to
3839 * change the physical cpu
3841 if (unlikely(svm->vmcb->control.exit_code ==
3842 SVM_EXIT_EXCP_BASE + MC_VECTOR))
3843 svm_handle_mce(vcpu);
3845 svm_complete_interrupts(vcpu);
3847 if (is_guest_mode(vcpu))
3848 return EXIT_FASTPATH_NONE;
3850 return svm_exit_handlers_fastpath(vcpu);
3853 static void svm_load_mmu_pgd(struct kvm_vcpu *vcpu, hpa_t root_hpa,
3856 struct vcpu_svm *svm = to_svm(vcpu);
3860 svm->vmcb->control.nested_cr3 = __sme_set(root_hpa);
3861 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
3863 hv_track_root_tdp(vcpu, root_hpa);
3865 cr3 = vcpu->arch.cr3;
3866 } else if (vcpu->arch.mmu->shadow_root_level >= PT64_ROOT_4LEVEL) {
3867 cr3 = __sme_set(root_hpa) | kvm_get_active_pcid(vcpu);
3869 /* PCID in the guest should be impossible with a 32-bit MMU. */
3870 WARN_ON_ONCE(kvm_get_active_pcid(vcpu));
3874 svm->vmcb->save.cr3 = cr3;
3875 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
3878 static int is_disabled(void)
3882 rdmsrl(MSR_VM_CR, vm_cr);
3883 if (vm_cr & (1 << SVM_VM_CR_SVM_DISABLE))
3890 svm_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
3893 * Patch in the VMMCALL instruction:
3895 hypercall[0] = 0x0f;
3896 hypercall[1] = 0x01;
3897 hypercall[2] = 0xd9;
3900 static int __init svm_check_processor_compat(void)
3906 * The kvm parameter can be NULL (module initialization, or invocation before
3907 * VM creation). Be sure to check the kvm parameter before using it.
3909 static bool svm_has_emulated_msr(struct kvm *kvm, u32 index)
3912 case MSR_IA32_MCG_EXT_CTL:
3913 case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:
3915 case MSR_IA32_SMBASE:
3916 /* SEV-ES guests do not support SMM, so report false */
3917 if (kvm && sev_es_guest(kvm))
3927 static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
3929 struct vcpu_svm *svm = to_svm(vcpu);
3930 struct kvm_cpuid_entry2 *best;
3932 vcpu->arch.xsaves_enabled = guest_cpuid_has(vcpu, X86_FEATURE_XSAVE) &&
3933 boot_cpu_has(X86_FEATURE_XSAVE) &&
3934 boot_cpu_has(X86_FEATURE_XSAVES);
3936 /* Update nrips enabled cache */
3937 svm->nrips_enabled = kvm_cpu_cap_has(X86_FEATURE_NRIPS) &&
3938 guest_cpuid_has(vcpu, X86_FEATURE_NRIPS);
3940 svm->tsc_scaling_enabled = tsc_scaling && guest_cpuid_has(vcpu, X86_FEATURE_TSCRATEMSR);
3942 svm_recalc_instruction_intercepts(vcpu, svm);
3944 /* For sev guests, the memory encryption bit is not reserved in CR3. */
3945 if (sev_guest(vcpu->kvm)) {
3946 best = kvm_find_cpuid_entry(vcpu, 0x8000001F, 0);
3948 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
3951 if (kvm_vcpu_apicv_active(vcpu)) {
3953 * AVIC does not work with an x2APIC mode guest. If the X2APIC feature
3954 * is exposed to the guest, disable AVIC.
3956 if (guest_cpuid_has(vcpu, X86_FEATURE_X2APIC))
3957 kvm_request_apicv_update(vcpu->kvm, false,
3958 APICV_INHIBIT_REASON_X2APIC);
3961 * Currently, AVIC does not work with nested virtualization.
3962 * So, we disable AVIC when cpuid for SVM is set in the L1 guest.
3964 if (nested && guest_cpuid_has(vcpu, X86_FEATURE_SVM))
3965 kvm_request_apicv_update(vcpu->kvm, false,
3966 APICV_INHIBIT_REASON_NESTED);
3968 init_vmcb_after_set_cpuid(vcpu);
3971 static bool svm_has_wbinvd_exit(void)
3976 #define PRE_EX(exit) { .exit_code = (exit), \
3977 .stage = X86_ICPT_PRE_EXCEPT, }
3978 #define POST_EX(exit) { .exit_code = (exit), \
3979 .stage = X86_ICPT_POST_EXCEPT, }
3980 #define POST_MEM(exit) { .exit_code = (exit), \
3981 .stage = X86_ICPT_POST_MEMACCESS, }
3983 static const struct __x86_intercept {
3985 enum x86_intercept_stage stage;
3986 } x86_intercept_map[] = {
3987 [x86_intercept_cr_read] = POST_EX(SVM_EXIT_READ_CR0),
3988 [x86_intercept_cr_write] = POST_EX(SVM_EXIT_WRITE_CR0),
3989 [x86_intercept_clts] = POST_EX(SVM_EXIT_WRITE_CR0),
3990 [x86_intercept_lmsw] = POST_EX(SVM_EXIT_WRITE_CR0),
3991 [x86_intercept_smsw] = POST_EX(SVM_EXIT_READ_CR0),
3992 [x86_intercept_dr_read] = POST_EX(SVM_EXIT_READ_DR0),
3993 [x86_intercept_dr_write] = POST_EX(SVM_EXIT_WRITE_DR0),
3994 [x86_intercept_sldt] = POST_EX(SVM_EXIT_LDTR_READ),
3995 [x86_intercept_str] = POST_EX(SVM_EXIT_TR_READ),
3996 [x86_intercept_lldt] = POST_EX(SVM_EXIT_LDTR_WRITE),
3997 [x86_intercept_ltr] = POST_EX(SVM_EXIT_TR_WRITE),
3998 [x86_intercept_sgdt] = POST_EX(SVM_EXIT_GDTR_READ),
3999 [x86_intercept_sidt] = POST_EX(SVM_EXIT_IDTR_READ),
4000 [x86_intercept_lgdt] = POST_EX(SVM_EXIT_GDTR_WRITE),
4001 [x86_intercept_lidt] = POST_EX(SVM_EXIT_IDTR_WRITE),
4002 [x86_intercept_vmrun] = POST_EX(SVM_EXIT_VMRUN),
4003 [x86_intercept_vmmcall] = POST_EX(SVM_EXIT_VMMCALL),
4004 [x86_intercept_vmload] = POST_EX(SVM_EXIT_VMLOAD),
4005 [x86_intercept_vmsave] = POST_EX(SVM_EXIT_VMSAVE),
4006 [x86_intercept_stgi] = POST_EX(SVM_EXIT_STGI),
4007 [x86_intercept_clgi] = POST_EX(SVM_EXIT_CLGI),
4008 [x86_intercept_skinit] = POST_EX(SVM_EXIT_SKINIT),
4009 [x86_intercept_invlpga] = POST_EX(SVM_EXIT_INVLPGA),
4010 [x86_intercept_rdtscp] = POST_EX(SVM_EXIT_RDTSCP),
4011 [x86_intercept_monitor] = POST_MEM(SVM_EXIT_MONITOR),
4012 [x86_intercept_mwait] = POST_EX(SVM_EXIT_MWAIT),
4013 [x86_intercept_invlpg] = POST_EX(SVM_EXIT_INVLPG),
4014 [x86_intercept_invd] = POST_EX(SVM_EXIT_INVD),
4015 [x86_intercept_wbinvd] = POST_EX(SVM_EXIT_WBINVD),
4016 [x86_intercept_wrmsr] = POST_EX(SVM_EXIT_MSR),
4017 [x86_intercept_rdtsc] = POST_EX(SVM_EXIT_RDTSC),
4018 [x86_intercept_rdmsr] = POST_EX(SVM_EXIT_MSR),
4019 [x86_intercept_rdpmc] = POST_EX(SVM_EXIT_RDPMC),
4020 [x86_intercept_cpuid] = PRE_EX(SVM_EXIT_CPUID),
4021 [x86_intercept_rsm] = PRE_EX(SVM_EXIT_RSM),
4022 [x86_intercept_pause] = PRE_EX(SVM_EXIT_PAUSE),
4023 [x86_intercept_pushf] = PRE_EX(SVM_EXIT_PUSHF),
4024 [x86_intercept_popf] = PRE_EX(SVM_EXIT_POPF),
4025 [x86_intercept_intn] = PRE_EX(SVM_EXIT_SWINT),
4026 [x86_intercept_iret] = PRE_EX(SVM_EXIT_IRET),
4027 [x86_intercept_icebp] = PRE_EX(SVM_EXIT_ICEBP),
4028 [x86_intercept_hlt] = POST_EX(SVM_EXIT_HLT),
4029 [x86_intercept_in] = POST_EX(SVM_EXIT_IOIO),
4030 [x86_intercept_ins] = POST_EX(SVM_EXIT_IOIO),
4031 [x86_intercept_out] = POST_EX(SVM_EXIT_IOIO),
4032 [x86_intercept_outs] = POST_EX(SVM_EXIT_IOIO),
4033 [x86_intercept_xsetbv] = PRE_EX(SVM_EXIT_XSETBV),
4040 static int svm_check_intercept(struct kvm_vcpu *vcpu,
4041 struct x86_instruction_info *info,
4042 enum x86_intercept_stage stage,
4043 struct x86_exception *exception)
4045 struct vcpu_svm *svm = to_svm(vcpu);
4046 int vmexit, ret = X86EMUL_CONTINUE;
4047 struct __x86_intercept icpt_info;
4048 struct vmcb *vmcb = svm->vmcb;
4050 if (info->intercept >= ARRAY_SIZE(x86_intercept_map))
4053 icpt_info = x86_intercept_map[info->intercept];
4055 if (stage != icpt_info.stage)
4058 switch (icpt_info.exit_code) {
4059 case SVM_EXIT_READ_CR0:
4060 if (info->intercept == x86_intercept_cr_read)
4061 icpt_info.exit_code += info->modrm_reg;
4063 case SVM_EXIT_WRITE_CR0: {
4064 unsigned long cr0, val;
4066 if (info->intercept == x86_intercept_cr_write)
4067 icpt_info.exit_code += info->modrm_reg;
4069 if (icpt_info.exit_code != SVM_EXIT_WRITE_CR0 ||
4070 info->intercept == x86_intercept_clts)
4073 if (!(vmcb12_is_intercept(&svm->nested.ctl,
4074 INTERCEPT_SELECTIVE_CR0)))
4077 cr0 = vcpu->arch.cr0 & ~SVM_CR0_SELECTIVE_MASK;
4078 val = info->src_val & ~SVM_CR0_SELECTIVE_MASK;
4080 if (info->intercept == x86_intercept_lmsw) {
4083 /* lmsw can't clear PE - catch this here */
4084 if (cr0 & X86_CR0_PE)
4089 icpt_info.exit_code = SVM_EXIT_CR0_SEL_WRITE;
4093 case SVM_EXIT_READ_DR0:
4094 case SVM_EXIT_WRITE_DR0:
4095 icpt_info.exit_code += info->modrm_reg;
4098 if (info->intercept == x86_intercept_wrmsr)
4099 vmcb->control.exit_info_1 = 1;
4101 vmcb->control.exit_info_1 = 0;
4103 case SVM_EXIT_PAUSE:
4105 * We get this for NOP only, but pause
4106 * is rep not, check this here
4108 if (info->rep_prefix != REPE_PREFIX)
4111 case SVM_EXIT_IOIO: {
4115 if (info->intercept == x86_intercept_in ||
4116 info->intercept == x86_intercept_ins) {
4117 exit_info = ((info->src_val & 0xffff) << 16) |
4119 bytes = info->dst_bytes;
4121 exit_info = (info->dst_val & 0xffff) << 16;
4122 bytes = info->src_bytes;
4125 if (info->intercept == x86_intercept_outs ||
4126 info->intercept == x86_intercept_ins)
4127 exit_info |= SVM_IOIO_STR_MASK;
4129 if (info->rep_prefix)
4130 exit_info |= SVM_IOIO_REP_MASK;
4132 bytes = min(bytes, 4u);
4134 exit_info |= bytes << SVM_IOIO_SIZE_SHIFT;
4136 exit_info |= (u32)info->ad_bytes << (SVM_IOIO_ASIZE_SHIFT - 1);
4138 vmcb->control.exit_info_1 = exit_info;
4139 vmcb->control.exit_info_2 = info->next_rip;
4147 /* TODO: Advertise NRIPS to guest hypervisor unconditionally */
4148 if (static_cpu_has(X86_FEATURE_NRIPS))
4149 vmcb->control.next_rip = info->next_rip;
4150 vmcb->control.exit_code = icpt_info.exit_code;
4151 vmexit = nested_svm_exit_handled(svm);
4153 ret = (vmexit == NESTED_EXIT_DONE) ? X86EMUL_INTERCEPTED
4160 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
4164 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
4166 if (!kvm_pause_in_guest(vcpu->kvm))
4167 shrink_ple_window(vcpu);
4170 static void svm_setup_mce(struct kvm_vcpu *vcpu)
4172 /* [63:9] are reserved. */
4173 vcpu->arch.mcg_cap &= 0x1ff;
4176 bool svm_smi_blocked(struct kvm_vcpu *vcpu)
4178 struct vcpu_svm *svm = to_svm(vcpu);
4180 /* Per APM Vol.2 15.22.2 "Response to SMI" */
4184 return is_smm(vcpu);
4187 static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
4189 struct vcpu_svm *svm = to_svm(vcpu);
4190 if (svm->nested.nested_run_pending)
4193 if (svm_smi_blocked(vcpu))
4196 /* An SMI must not be injected into L2 if it's supposed to VM-Exit. */
4197 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_smi(svm))
4203 static int svm_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
4205 struct vcpu_svm *svm = to_svm(vcpu);
4206 struct kvm_host_map map_save;
4209 if (!is_guest_mode(vcpu))
4212 /* FED8h - SVM Guest */
4213 put_smstate(u64, smstate, 0x7ed8, 1);
4214 /* FEE0h - SVM Guest VMCB Physical Address */
4215 put_smstate(u64, smstate, 0x7ee0, svm->nested.vmcb12_gpa);
4217 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4218 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4219 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4221 ret = nested_svm_vmexit(svm);
4226 * KVM uses VMCB01 to store L1 host state while L2 runs but
4227 * VMCB01 is going to be used during SMM and thus the state will
4228 * be lost. Temporary save non-VMLOAD/VMSAVE state to the host save
4229 * area pointed to by MSR_VM_HSAVE_PA. APM guarantees that the
4230 * format of the area is identical to guest save area offsetted
4231 * by 0x400 (matches the offset of 'struct vmcb_save_area'
4232 * within 'struct vmcb'). Note: HSAVE area may also be used by
4233 * L1 hypervisor to save additional host context (e.g. KVM does
4234 * that, see svm_prepare_switch_to_guest()) which must be
4237 if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr),
4238 &map_save) == -EINVAL)
4241 BUILD_BUG_ON(offsetof(struct vmcb, save) != 0x400);
4243 svm_copy_vmrun_state(map_save.hva + 0x400,
4244 &svm->vmcb01.ptr->save);
4246 kvm_vcpu_unmap(vcpu, &map_save, true);
4250 static int svm_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
4252 struct vcpu_svm *svm = to_svm(vcpu);
4253 struct kvm_host_map map, map_save;
4254 u64 saved_efer, vmcb12_gpa;
4255 struct vmcb *vmcb12;
4258 if (!guest_cpuid_has(vcpu, X86_FEATURE_LM))
4261 /* Non-zero if SMI arrived while vCPU was in guest mode. */
4262 if (!GET_SMSTATE(u64, smstate, 0x7ed8))
4265 if (!guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4268 saved_efer = GET_SMSTATE(u64, smstate, 0x7ed0);
4269 if (!(saved_efer & EFER_SVME))
4272 vmcb12_gpa = GET_SMSTATE(u64, smstate, 0x7ee0);
4273 if (kvm_vcpu_map(vcpu, gpa_to_gfn(vmcb12_gpa), &map) == -EINVAL)
4277 if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr), &map_save) == -EINVAL)
4280 if (svm_allocate_nested(svm))
4284 * Restore L1 host state from L1 HSAVE area as VMCB01 was
4285 * used during SMM (see svm_enter_smm())
4288 svm_copy_vmrun_state(&svm->vmcb01.ptr->save, map_save.hva + 0x400);
4291 * Enter the nested guest now
4294 vmcb_mark_all_dirty(svm->vmcb01.ptr);
4297 nested_copy_vmcb_control_to_cache(svm, &vmcb12->control);
4298 nested_copy_vmcb_save_to_cache(svm, &vmcb12->save);
4299 ret = enter_svm_guest_mode(vcpu, vmcb12_gpa, vmcb12, false);
4304 svm->nested.nested_run_pending = 1;
4307 kvm_vcpu_unmap(vcpu, &map_save, true);
4309 kvm_vcpu_unmap(vcpu, &map, true);
4313 static void svm_enable_smi_window(struct kvm_vcpu *vcpu)
4315 struct vcpu_svm *svm = to_svm(vcpu);
4317 if (!gif_set(svm)) {
4318 if (vgif_enabled(svm))
4319 svm_set_intercept(svm, INTERCEPT_STGI);
4320 /* STGI will cause a vm exit */
4322 /* We must be in SMM; RSM will cause a vmexit anyway. */
4326 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, int emul_type,
4327 void *insn, int insn_len)
4329 bool smep, smap, is_user;
4333 /* Emulation is always possible when KVM has access to all guest state. */
4334 if (!sev_guest(vcpu->kvm))
4337 /* #UD and #GP should never be intercepted for SEV guests. */
4338 WARN_ON_ONCE(emul_type & (EMULTYPE_TRAP_UD |
4339 EMULTYPE_TRAP_UD_FORCED |
4340 EMULTYPE_VMWARE_GP));
4343 * Emulation is impossible for SEV-ES guests as KVM doesn't have access
4344 * to guest register state.
4346 if (sev_es_guest(vcpu->kvm))
4350 * Emulation is possible if the instruction is already decoded, e.g.
4351 * when completing I/O after returning from userspace.
4353 if (emul_type & EMULTYPE_NO_DECODE)
4357 * Emulation is possible for SEV guests if and only if a prefilled
4358 * buffer containing the bytes of the intercepted instruction is
4359 * available. SEV guest memory is encrypted with a guest specific key
4360 * and cannot be decrypted by KVM, i.e. KVM would read cyphertext and
4363 * Inject #UD if KVM reached this point without an instruction buffer.
4364 * In practice, this path should never be hit by a well-behaved guest,
4365 * e.g. KVM doesn't intercept #UD or #GP for SEV guests, but this path
4366 * is still theoretically reachable, e.g. via unaccelerated fault-like
4367 * AVIC access, and needs to be handled by KVM to avoid putting the
4368 * guest into an infinite loop. Injecting #UD is somewhat arbitrary,
4369 * but its the least awful option given lack of insight into the guest.
4371 if (unlikely(!insn)) {
4372 kvm_queue_exception(vcpu, UD_VECTOR);
4377 * Emulate for SEV guests if the insn buffer is not empty. The buffer
4378 * will be empty if the DecodeAssist microcode cannot fetch bytes for
4379 * the faulting instruction because the code fetch itself faulted, e.g.
4380 * the guest attempted to fetch from emulated MMIO or a guest page
4381 * table used to translate CS:RIP resides in emulated MMIO.
4383 if (likely(insn_len))
4387 * Detect and workaround Errata 1096 Fam_17h_00_0Fh.
4390 * When CPU raises #NPF on guest data access and vCPU CR4.SMAP=1, it is
4391 * possible that CPU microcode implementing DecodeAssist will fail to
4392 * read guest memory at CS:RIP and vmcb.GuestIntrBytes will incorrectly
4393 * be '0'. This happens because microcode reads CS:RIP using a _data_
4394 * loap uop with CPL=0 privileges. If the load hits a SMAP #PF, ucode
4395 * gives up and does not fill the instruction bytes buffer.
4397 * As above, KVM reaches this point iff the VM is an SEV guest, the CPU
4398 * supports DecodeAssist, a #NPF was raised, KVM's page fault handler
4399 * triggered emulation (e.g. for MMIO), and the CPU returned 0 in the
4400 * GuestIntrBytes field of the VMCB.
4402 * This does _not_ mean that the erratum has been encountered, as the
4403 * DecodeAssist will also fail if the load for CS:RIP hits a legitimate
4404 * #PF, e.g. if the guest attempt to execute from emulated MMIO and
4405 * encountered a reserved/not-present #PF.
4407 * To hit the erratum, the following conditions must be true:
4408 * 1. CR4.SMAP=1 (obviously).
4409 * 2. CR4.SMEP=0 || CPL=3. If SMEP=1 and CPL<3, the erratum cannot
4410 * have been hit as the guest would have encountered a SMEP
4411 * violation #PF, not a #NPF.
4412 * 3. The #NPF is not due to a code fetch, in which case failure to
4413 * retrieve the instruction bytes is legitimate (see abvoe).
4415 * In addition, don't apply the erratum workaround if the #NPF occurred
4416 * while translating guest page tables (see below).
4418 error_code = to_svm(vcpu)->vmcb->control.exit_info_1;
4419 if (error_code & (PFERR_GUEST_PAGE_MASK | PFERR_FETCH_MASK))
4422 cr4 = kvm_read_cr4(vcpu);
4423 smep = cr4 & X86_CR4_SMEP;
4424 smap = cr4 & X86_CR4_SMAP;
4425 is_user = svm_get_cpl(vcpu) == 3;
4426 if (smap && (!smep || is_user)) {
4427 pr_err_ratelimited("KVM: SEV Guest triggered AMD Erratum 1096\n");
4430 * If the fault occurred in userspace, arbitrarily inject #GP
4431 * to avoid killing the guest and to hopefully avoid confusing
4432 * the guest kernel too much, e.g. injecting #PF would not be
4433 * coherent with respect to the guest's page tables. Request
4434 * triple fault if the fault occurred in the kernel as there's
4435 * no fault that KVM can inject without confusing the guest.
4436 * In practice, the triple fault is moot as no sane SEV kernel
4437 * will execute from user memory while also running with SMAP=1.
4440 kvm_inject_gp(vcpu, 0);
4442 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
4447 * If the erratum was not hit, simply resume the guest and let it fault
4448 * again. While awful, e.g. the vCPU may get stuck in an infinite loop
4449 * if the fault is at CPL=0, it's the lesser of all evils. Exiting to
4450 * userspace will kill the guest, and letting the emulator read garbage
4451 * will yield random behavior and potentially corrupt the guest.
4453 * Simply resuming the guest is technically not a violation of the SEV
4454 * architecture. AMD's APM states that all code fetches and page table
4455 * accesses for SEV guest are encrypted, regardless of the C-Bit. The
4456 * APM also states that encrypted accesses to MMIO are "ignored", but
4457 * doesn't explicitly define "ignored", i.e. doing nothing and letting
4458 * the guest spin is technically "ignoring" the access.
4463 static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
4465 struct vcpu_svm *svm = to_svm(vcpu);
4468 * TODO: Last condition latch INIT signals on vCPU when
4469 * vCPU is in guest-mode and vmcb12 defines intercept on INIT.
4470 * To properly emulate the INIT intercept,
4471 * svm_check_nested_events() should call nested_svm_vmexit()
4472 * if an INIT signal is pending.
4474 return !gif_set(svm) ||
4475 (vmcb_is_intercept(&svm->vmcb->control, INTERCEPT_INIT));
4478 static void svm_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
4480 if (!sev_es_guest(vcpu->kvm))
4481 return kvm_vcpu_deliver_sipi_vector(vcpu, vector);
4483 sev_vcpu_deliver_sipi_vector(vcpu, vector);
4486 static void svm_vm_destroy(struct kvm *kvm)
4488 avic_vm_destroy(kvm);
4489 sev_vm_destroy(kvm);
4492 static int svm_vm_init(struct kvm *kvm)
4494 if (!pause_filter_count || !pause_filter_thresh)
4495 kvm->arch.pause_in_guest = true;
4498 int ret = avic_vm_init(kvm);
4506 static struct kvm_x86_ops svm_x86_ops __initdata = {
4509 .hardware_unsetup = svm_hardware_unsetup,
4510 .hardware_enable = svm_hardware_enable,
4511 .hardware_disable = svm_hardware_disable,
4512 .has_emulated_msr = svm_has_emulated_msr,
4514 .vcpu_create = svm_vcpu_create,
4515 .vcpu_free = svm_vcpu_free,
4516 .vcpu_reset = svm_vcpu_reset,
4518 .vm_size = sizeof(struct kvm_svm),
4519 .vm_init = svm_vm_init,
4520 .vm_destroy = svm_vm_destroy,
4522 .prepare_switch_to_guest = svm_prepare_switch_to_guest,
4523 .vcpu_load = svm_vcpu_load,
4524 .vcpu_put = svm_vcpu_put,
4525 .vcpu_blocking = avic_vcpu_blocking,
4526 .vcpu_unblocking = avic_vcpu_unblocking,
4528 .update_exception_bitmap = svm_update_exception_bitmap,
4529 .get_msr_feature = svm_get_msr_feature,
4530 .get_msr = svm_get_msr,
4531 .set_msr = svm_set_msr,
4532 .get_segment_base = svm_get_segment_base,
4533 .get_segment = svm_get_segment,
4534 .set_segment = svm_set_segment,
4535 .get_cpl = svm_get_cpl,
4536 .get_cs_db_l_bits = svm_get_cs_db_l_bits,
4537 .set_cr0 = svm_set_cr0,
4538 .post_set_cr3 = sev_post_set_cr3,
4539 .is_valid_cr4 = svm_is_valid_cr4,
4540 .set_cr4 = svm_set_cr4,
4541 .set_efer = svm_set_efer,
4542 .get_idt = svm_get_idt,
4543 .set_idt = svm_set_idt,
4544 .get_gdt = svm_get_gdt,
4545 .set_gdt = svm_set_gdt,
4546 .set_dr7 = svm_set_dr7,
4547 .sync_dirty_debug_regs = svm_sync_dirty_debug_regs,
4548 .cache_reg = svm_cache_reg,
4549 .get_rflags = svm_get_rflags,
4550 .set_rflags = svm_set_rflags,
4551 .get_if_flag = svm_get_if_flag,
4553 .flush_tlb_all = svm_flush_tlb_current,
4554 .flush_tlb_current = svm_flush_tlb_current,
4555 .flush_tlb_gva = svm_flush_tlb_gva,
4556 .flush_tlb_guest = svm_flush_tlb_current,
4558 .vcpu_pre_run = svm_vcpu_pre_run,
4559 .vcpu_run = svm_vcpu_run,
4560 .handle_exit = svm_handle_exit,
4561 .skip_emulated_instruction = svm_skip_emulated_instruction,
4562 .update_emulated_instruction = NULL,
4563 .set_interrupt_shadow = svm_set_interrupt_shadow,
4564 .get_interrupt_shadow = svm_get_interrupt_shadow,
4565 .patch_hypercall = svm_patch_hypercall,
4566 .inject_irq = svm_inject_irq,
4567 .inject_nmi = svm_inject_nmi,
4568 .queue_exception = svm_queue_exception,
4569 .cancel_injection = svm_cancel_injection,
4570 .interrupt_allowed = svm_interrupt_allowed,
4571 .nmi_allowed = svm_nmi_allowed,
4572 .get_nmi_mask = svm_get_nmi_mask,
4573 .set_nmi_mask = svm_set_nmi_mask,
4574 .enable_nmi_window = svm_enable_nmi_window,
4575 .enable_irq_window = svm_enable_irq_window,
4576 .update_cr8_intercept = svm_update_cr8_intercept,
4577 .refresh_apicv_exec_ctrl = avic_refresh_apicv_exec_ctrl,
4578 .check_apicv_inhibit_reasons = avic_check_apicv_inhibit_reasons,
4579 .apicv_post_state_restore = avic_apicv_post_state_restore,
4581 .get_exit_info = svm_get_exit_info,
4583 .vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
4585 .has_wbinvd_exit = svm_has_wbinvd_exit,
4587 .get_l2_tsc_offset = svm_get_l2_tsc_offset,
4588 .get_l2_tsc_multiplier = svm_get_l2_tsc_multiplier,
4589 .write_tsc_offset = svm_write_tsc_offset,
4590 .write_tsc_multiplier = svm_write_tsc_multiplier,
4592 .load_mmu_pgd = svm_load_mmu_pgd,
4594 .check_intercept = svm_check_intercept,
4595 .handle_exit_irqoff = svm_handle_exit_irqoff,
4597 .request_immediate_exit = __kvm_request_immediate_exit,
4599 .sched_in = svm_sched_in,
4601 .pmu_ops = &amd_pmu_ops,
4602 .nested_ops = &svm_nested_ops,
4604 .deliver_interrupt = svm_deliver_interrupt,
4605 .pi_update_irte = avic_pi_update_irte,
4606 .setup_mce = svm_setup_mce,
4608 .smi_allowed = svm_smi_allowed,
4609 .enter_smm = svm_enter_smm,
4610 .leave_smm = svm_leave_smm,
4611 .enable_smi_window = svm_enable_smi_window,
4613 .mem_enc_ioctl = sev_mem_enc_ioctl,
4614 .mem_enc_register_region = sev_mem_enc_register_region,
4615 .mem_enc_unregister_region = sev_mem_enc_unregister_region,
4617 .vm_copy_enc_context_from = sev_vm_copy_enc_context_from,
4618 .vm_move_enc_context_from = sev_vm_move_enc_context_from,
4620 .can_emulate_instruction = svm_can_emulate_instruction,
4622 .apic_init_signal_blocked = svm_apic_init_signal_blocked,
4624 .msr_filter_changed = svm_msr_filter_changed,
4625 .complete_emulated_msr = svm_complete_emulated_msr,
4627 .vcpu_deliver_sipi_vector = svm_vcpu_deliver_sipi_vector,
4631 * The default MMIO mask is a single bit (excluding the present bit),
4632 * which could conflict with the memory encryption bit. Check for
4633 * memory encryption support and override the default MMIO mask if
4634 * memory encryption is enabled.
4636 static __init void svm_adjust_mmio_mask(void)
4638 unsigned int enc_bit, mask_bit;
4641 /* If there is no memory encryption support, use existing mask */
4642 if (cpuid_eax(0x80000000) < 0x8000001f)
4645 /* If memory encryption is not enabled, use existing mask */
4646 rdmsrl(MSR_AMD64_SYSCFG, msr);
4647 if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
4650 enc_bit = cpuid_ebx(0x8000001f) & 0x3f;
4651 mask_bit = boot_cpu_data.x86_phys_bits;
4653 /* Increment the mask bit if it is the same as the encryption bit */
4654 if (enc_bit == mask_bit)
4658 * If the mask bit location is below 52, then some bits above the
4659 * physical addressing limit will always be reserved, so use the
4660 * rsvd_bits() function to generate the mask. This mask, along with
4661 * the present bit, will be used to generate a page fault with
4664 * If the mask bit location is 52 (or above), then clear the mask.
4666 mask = (mask_bit < 52) ? rsvd_bits(mask_bit, 51) | PT_PRESENT_MASK : 0;
4668 kvm_mmu_set_mmio_spte_mask(mask, mask, PT_WRITABLE_MASK | PT_USER_MASK);
4671 static __init void svm_set_cpu_caps(void)
4677 /* CPUID 0x80000001 and 0x8000000A (SVM features) */
4679 kvm_cpu_cap_set(X86_FEATURE_SVM);
4680 kvm_cpu_cap_set(X86_FEATURE_VMCBCLEAN);
4683 kvm_cpu_cap_set(X86_FEATURE_NRIPS);
4686 kvm_cpu_cap_set(X86_FEATURE_NPT);
4689 kvm_cpu_cap_set(X86_FEATURE_TSCRATEMSR);
4691 /* Nested VM can receive #VMEXIT instead of triggering #GP */
4692 kvm_cpu_cap_set(X86_FEATURE_SVME_ADDR_CHK);
4695 /* CPUID 0x80000008 */
4696 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) ||
4697 boot_cpu_has(X86_FEATURE_AMD_SSBD))
4698 kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
4700 /* AMD PMU PERFCTR_CORE CPUID */
4701 if (enable_pmu && boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
4702 kvm_cpu_cap_set(X86_FEATURE_PERFCTR_CORE);
4704 /* CPUID 0x8000001F (SME/SEV features) */
4708 static __init int svm_hardware_setup(void)
4711 struct page *iopm_pages;
4714 unsigned int order = get_order(IOPM_SIZE);
4717 * NX is required for shadow paging and for NPT if the NX huge pages
4718 * mitigation is enabled.
4720 if (!boot_cpu_has(X86_FEATURE_NX)) {
4721 pr_err_ratelimited("NX (Execute Disable) not supported\n");
4724 kvm_enable_efer_bits(EFER_NX);
4726 iopm_pages = alloc_pages(GFP_KERNEL, order);
4731 iopm_va = page_address(iopm_pages);
4732 memset(iopm_va, 0xff, PAGE_SIZE * (1 << order));
4733 iopm_base = page_to_pfn(iopm_pages) << PAGE_SHIFT;
4735 init_msrpm_offsets();
4737 supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
4739 if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
4740 kvm_enable_efer_bits(EFER_FFXSR);
4743 if (!boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
4744 tsc_scaling = false;
4746 pr_info("TSC scaling supported\n");
4747 kvm_has_tsc_control = true;
4748 kvm_max_tsc_scaling_ratio = TSC_RATIO_MAX;
4749 kvm_tsc_scaling_ratio_frac_bits = 32;
4753 tsc_aux_uret_slot = kvm_add_user_return_msr(MSR_TSC_AUX);
4755 /* Check for pause filtering support */
4756 if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
4757 pause_filter_count = 0;
4758 pause_filter_thresh = 0;
4759 } else if (!boot_cpu_has(X86_FEATURE_PFTHRESHOLD)) {
4760 pause_filter_thresh = 0;
4764 printk(KERN_INFO "kvm: Nested Virtualization enabled\n");
4765 kvm_enable_efer_bits(EFER_SVME | EFER_LMSLE);
4769 * KVM's MMU doesn't support using 2-level paging for itself, and thus
4770 * NPT isn't supported if the host is using 2-level paging since host
4771 * CR4 is unchanged on VMRUN.
4773 if (!IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_X86_PAE))
4774 npt_enabled = false;
4776 if (!boot_cpu_has(X86_FEATURE_NPT))
4777 npt_enabled = false;
4779 /* Force VM NPT level equal to the host's paging level */
4780 kvm_configure_mmu(npt_enabled, get_npt_level(),
4781 get_npt_level(), PG_LEVEL_1G);
4782 pr_info("kvm: Nested Paging %sabled\n", npt_enabled ? "en" : "dis");
4784 /* Note, SEV setup consumes npt_enabled. */
4785 sev_hardware_setup();
4787 svm_hv_hardware_setup();
4789 svm_adjust_mmio_mask();
4791 for_each_possible_cpu(cpu) {
4792 r = svm_cpu_init(cpu);
4798 if (!boot_cpu_has(X86_FEATURE_NRIPS))
4802 enable_apicv = avic = avic && npt_enabled && boot_cpu_has(X86_FEATURE_AVIC);
4805 pr_info("AVIC enabled\n");
4807 amd_iommu_register_ga_log_notifier(&avic_ga_log_notifier);
4809 svm_x86_ops.vcpu_blocking = NULL;
4810 svm_x86_ops.vcpu_unblocking = NULL;
4815 !boot_cpu_has(X86_FEATURE_V_VMSAVE_VMLOAD) ||
4816 !IS_ENABLED(CONFIG_X86_64)) {
4819 pr_info("Virtual VMLOAD VMSAVE supported\n");
4823 if (boot_cpu_has(X86_FEATURE_SVME_ADDR_CHK))
4824 svm_gp_erratum_intercept = false;
4827 if (!boot_cpu_has(X86_FEATURE_VGIF))
4830 pr_info("Virtual GIF supported\n");
4834 if (!boot_cpu_has(X86_FEATURE_LBRV))
4837 pr_info("LBR virtualization supported\n");
4841 pr_info("PMU virtualization is disabled\n");
4846 * It seems that on AMD processors PTE's accessed bit is
4847 * being set by the CPU hardware before the NPF vmexit.
4848 * This is not expected behaviour and our tests fail because
4850 * A workaround here is to disable support for
4851 * GUEST_MAXPHYADDR < HOST_MAXPHYADDR if NPT is enabled.
4852 * In this case userspace can know if there is support using
4853 * KVM_CAP_SMALLER_MAXPHYADDR extension and decide how to handle
4855 * If future AMD CPU models change the behaviour described above,
4856 * this variable can be changed accordingly
4858 allow_smaller_maxphyaddr = !npt_enabled;
4863 svm_hardware_unsetup();
4868 static struct kvm_x86_init_ops svm_init_ops __initdata = {
4869 .cpu_has_kvm_support = has_svm,
4870 .disabled_by_bios = is_disabled,
4871 .hardware_setup = svm_hardware_setup,
4872 .check_processor_compatibility = svm_check_processor_compat,
4874 .runtime_ops = &svm_x86_ops,
4877 static int __init svm_init(void)
4879 __unused_size_checks();
4881 return kvm_init(&svm_init_ops, sizeof(struct vcpu_svm),
4882 __alignof__(struct vcpu_svm), THIS_MODULE);
4885 static void __exit svm_exit(void)
4890 module_init(svm_init)
4891 module_exit(svm_exit)