1 #define pr_fmt(fmt) "SVM: " fmt
3 #include <linux/kvm_host.h>
7 #include "kvm_cache_regs.h"
12 #include <linux/module.h>
13 #include <linux/mod_devicetable.h>
14 #include <linux/kernel.h>
15 #include <linux/vmalloc.h>
16 #include <linux/highmem.h>
17 #include <linux/amd-iommu.h>
18 #include <linux/sched.h>
19 #include <linux/trace_events.h>
20 #include <linux/slab.h>
21 #include <linux/hashtable.h>
22 #include <linux/objtool.h>
23 #include <linux/psp-sev.h>
24 #include <linux/file.h>
25 #include <linux/pagemap.h>
26 #include <linux/swap.h>
27 #include <linux/rwsem.h>
30 #include <asm/perf_event.h>
31 #include <asm/tlbflush.h>
33 #include <asm/debugreg.h>
34 #include <asm/kvm_para.h>
35 #include <asm/irq_remapping.h>
36 #include <asm/spec-ctrl.h>
37 #include <asm/cpu_device_id.h>
38 #include <asm/traps.h>
40 #include <asm/virtext.h>
46 #include "kvm_onhyperv.h"
47 #include "svm_onhyperv.h"
49 MODULE_AUTHOR("Qumranet");
50 MODULE_LICENSE("GPL");
53 static const struct x86_cpu_id svm_cpu_id[] = {
54 X86_MATCH_FEATURE(X86_FEATURE_SVM, NULL),
57 MODULE_DEVICE_TABLE(x86cpu, svm_cpu_id);
60 #define SEG_TYPE_LDT 2
61 #define SEG_TYPE_BUSY_TSS16 3
63 #define SVM_FEATURE_LBRV (1 << 1)
64 #define SVM_FEATURE_SVML (1 << 2)
65 #define SVM_FEATURE_TSC_RATE (1 << 4)
66 #define SVM_FEATURE_VMCB_CLEAN (1 << 5)
67 #define SVM_FEATURE_FLUSH_ASID (1 << 6)
68 #define SVM_FEATURE_DECODE_ASSIST (1 << 7)
69 #define SVM_FEATURE_PAUSE_FILTER (1 << 10)
71 #define DEBUGCTL_RESERVED_BITS (~(0x3fULL))
73 #define TSC_RATIO_RSVD 0xffffff0000000000ULL
74 #define TSC_RATIO_MIN 0x0000000000000001ULL
75 #define TSC_RATIO_MAX 0x000000ffffffffffULL
77 static bool erratum_383_found __read_mostly;
79 u32 msrpm_offsets[MSRPM_OFFSETS] __read_mostly;
82 * Set osvw_len to higher value when updated Revision Guides
83 * are published and we know what the new status bits are
85 static uint64_t osvw_len = 4, osvw_status;
87 static DEFINE_PER_CPU(u64, current_tsc_ratio);
88 #define TSC_RATIO_DEFAULT 0x0100000000ULL
90 static const struct svm_direct_access_msrs {
91 u32 index; /* Index of the MSR */
92 bool always; /* True if intercept is initially cleared */
93 } direct_access_msrs[MAX_DIRECT_ACCESS_MSRS] = {
94 { .index = MSR_STAR, .always = true },
95 { .index = MSR_IA32_SYSENTER_CS, .always = true },
96 { .index = MSR_IA32_SYSENTER_EIP, .always = false },
97 { .index = MSR_IA32_SYSENTER_ESP, .always = false },
99 { .index = MSR_GS_BASE, .always = true },
100 { .index = MSR_FS_BASE, .always = true },
101 { .index = MSR_KERNEL_GS_BASE, .always = true },
102 { .index = MSR_LSTAR, .always = true },
103 { .index = MSR_CSTAR, .always = true },
104 { .index = MSR_SYSCALL_MASK, .always = true },
106 { .index = MSR_IA32_SPEC_CTRL, .always = false },
107 { .index = MSR_IA32_PRED_CMD, .always = false },
108 { .index = MSR_IA32_LASTBRANCHFROMIP, .always = false },
109 { .index = MSR_IA32_LASTBRANCHTOIP, .always = false },
110 { .index = MSR_IA32_LASTINTFROMIP, .always = false },
111 { .index = MSR_IA32_LASTINTTOIP, .always = false },
112 { .index = MSR_EFER, .always = false },
113 { .index = MSR_IA32_CR_PAT, .always = false },
114 { .index = MSR_AMD64_SEV_ES_GHCB, .always = true },
115 { .index = MSR_INVALID, .always = false },
119 * These 2 parameters are used to config the controls for Pause-Loop Exiting:
120 * pause_filter_count: On processors that support Pause filtering(indicated
121 * by CPUID Fn8000_000A_EDX), the VMCB provides a 16 bit pause filter
122 * count value. On VMRUN this value is loaded into an internal counter.
123 * Each time a pause instruction is executed, this counter is decremented
124 * until it reaches zero at which time a #VMEXIT is generated if pause
125 * intercept is enabled. Refer to AMD APM Vol 2 Section 15.14.4 Pause
126 * Intercept Filtering for more details.
127 * This also indicate if ple logic enabled.
129 * pause_filter_thresh: In addition, some processor families support advanced
130 * pause filtering (indicated by CPUID Fn8000_000A_EDX) upper bound on
131 * the amount of time a guest is allowed to execute in a pause loop.
132 * In this mode, a 16-bit pause filter threshold field is added in the
133 * VMCB. The threshold value is a cycle count that is used to reset the
134 * pause counter. As with simple pause filtering, VMRUN loads the pause
135 * count value from VMCB into an internal counter. Then, on each pause
136 * instruction the hardware checks the elapsed number of cycles since
137 * the most recent pause instruction against the pause filter threshold.
138 * If the elapsed cycle count is greater than the pause filter threshold,
139 * then the internal pause count is reloaded from the VMCB and execution
140 * continues. If the elapsed cycle count is less than the pause filter
141 * threshold, then the internal pause count is decremented. If the count
142 * value is less than zero and PAUSE intercept is enabled, a #VMEXIT is
143 * triggered. If advanced pause filtering is supported and pause filter
144 * threshold field is set to zero, the filter will operate in the simpler,
148 static unsigned short pause_filter_thresh = KVM_DEFAULT_PLE_GAP;
149 module_param(pause_filter_thresh, ushort, 0444);
151 static unsigned short pause_filter_count = KVM_SVM_DEFAULT_PLE_WINDOW;
152 module_param(pause_filter_count, ushort, 0444);
154 /* Default doubles per-vcpu window every exit. */
155 static unsigned short pause_filter_count_grow = KVM_DEFAULT_PLE_WINDOW_GROW;
156 module_param(pause_filter_count_grow, ushort, 0444);
158 /* Default resets per-vcpu window every exit to pause_filter_count. */
159 static unsigned short pause_filter_count_shrink = KVM_DEFAULT_PLE_WINDOW_SHRINK;
160 module_param(pause_filter_count_shrink, ushort, 0444);
162 /* Default is to compute the maximum so we can never overflow. */
163 static unsigned short pause_filter_count_max = KVM_SVM_DEFAULT_PLE_WINDOW_MAX;
164 module_param(pause_filter_count_max, ushort, 0444);
167 * Use nested page tables by default. Note, NPT may get forced off by
168 * svm_hardware_setup() if it's unsupported by hardware or the host kernel.
170 bool npt_enabled = true;
171 module_param_named(npt, npt_enabled, bool, 0444);
173 /* allow nested virtualization in KVM/SVM */
174 static int nested = true;
175 module_param(nested, int, S_IRUGO);
177 /* enable/disable Next RIP Save */
178 static int nrips = true;
179 module_param(nrips, int, 0444);
181 /* enable/disable Virtual VMLOAD VMSAVE */
182 static int vls = true;
183 module_param(vls, int, 0444);
185 /* enable/disable Virtual GIF */
186 static int vgif = true;
187 module_param(vgif, int, 0444);
190 * enable / disable AVIC. Because the defaults differ for APICv
191 * support between VMX and SVM we cannot use module_param_named.
194 module_param(avic, bool, 0444);
196 bool __read_mostly dump_invalid_vmcb;
197 module_param(dump_invalid_vmcb, bool, 0644);
200 bool intercept_smi = true;
201 module_param(intercept_smi, bool, 0444);
204 static bool svm_gp_erratum_intercept = true;
206 static u8 rsm_ins_bytes[] = "\x0f\xaa";
208 static unsigned long iopm_base;
210 struct kvm_ldttss_desc {
213 unsigned base1:8, type:5, dpl:2, p:1;
214 unsigned limit1:4, zero0:3, g:1, base2:8;
217 } __attribute__((packed));
219 DEFINE_PER_CPU(struct svm_cpu_data *, svm_data);
222 * Only MSR_TSC_AUX is switched via the user return hook. EFER is switched via
223 * the VMCB, and the SYSCALL/SYSENTER MSRs are handled by VMLOAD/VMSAVE.
225 * RDTSCP and RDPID are not used in the kernel, specifically to allow KVM to
226 * defer the restoration of TSC_AUX until the CPU returns to userspace.
228 static int tsc_aux_uret_slot __read_mostly = -1;
230 static const u32 msrpm_ranges[] = {0, 0xc0000000, 0xc0010000};
232 #define NUM_MSR_MAPS ARRAY_SIZE(msrpm_ranges)
233 #define MSRS_RANGE_SIZE 2048
234 #define MSRS_IN_RANGE (MSRS_RANGE_SIZE * 8 / 2)
236 u32 svm_msrpm_offset(u32 msr)
241 for (i = 0; i < NUM_MSR_MAPS; i++) {
242 if (msr < msrpm_ranges[i] ||
243 msr >= msrpm_ranges[i] + MSRS_IN_RANGE)
246 offset = (msr - msrpm_ranges[i]) / 4; /* 4 msrs per u8 */
247 offset += (i * MSRS_RANGE_SIZE); /* add range offset */
249 /* Now we have the u8 offset - but need the u32 offset */
253 /* MSR not in any range */
257 #define MAX_INST_SIZE 15
259 static int get_max_npt_level(void)
262 return PT64_ROOT_4LEVEL;
264 return PT32E_ROOT_LEVEL;
268 int svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)
270 struct vcpu_svm *svm = to_svm(vcpu);
271 u64 old_efer = vcpu->arch.efer;
272 vcpu->arch.efer = efer;
275 /* Shadow paging assumes NX to be available. */
278 if (!(efer & EFER_LMA))
282 if ((old_efer & EFER_SVME) != (efer & EFER_SVME)) {
283 if (!(efer & EFER_SVME)) {
284 svm_leave_nested(svm);
285 svm_set_gif(svm, true);
286 /* #GP intercept is still needed for vmware backdoor */
287 if (!enable_vmware_backdoor)
288 clr_exception_intercept(svm, GP_VECTOR);
291 * Free the nested guest state, unless we are in SMM.
292 * In this case we will return to the nested guest
293 * as soon as we leave SMM.
296 svm_free_nested(svm);
299 int ret = svm_allocate_nested(svm);
302 vcpu->arch.efer = old_efer;
306 if (svm_gp_erratum_intercept)
307 set_exception_intercept(svm, GP_VECTOR);
311 svm->vmcb->save.efer = efer | EFER_SVME;
312 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
316 static int is_external_interrupt(u32 info)
318 info &= SVM_EVTINJ_TYPE_MASK | SVM_EVTINJ_VALID;
319 return info == (SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR);
322 static u32 svm_get_interrupt_shadow(struct kvm_vcpu *vcpu)
324 struct vcpu_svm *svm = to_svm(vcpu);
327 if (svm->vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK)
328 ret = KVM_X86_SHADOW_INT_STI | KVM_X86_SHADOW_INT_MOV_SS;
332 static void svm_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
334 struct vcpu_svm *svm = to_svm(vcpu);
337 svm->vmcb->control.int_state &= ~SVM_INTERRUPT_SHADOW_MASK;
339 svm->vmcb->control.int_state |= SVM_INTERRUPT_SHADOW_MASK;
343 static int skip_emulated_instruction(struct kvm_vcpu *vcpu)
345 struct vcpu_svm *svm = to_svm(vcpu);
348 * SEV-ES does not expose the next RIP. The RIP update is controlled by
349 * the type of exit and the #VC handler in the guest.
351 if (sev_es_guest(vcpu->kvm))
354 if (nrips && svm->vmcb->control.next_rip != 0) {
355 WARN_ON_ONCE(!static_cpu_has(X86_FEATURE_NRIPS));
356 svm->next_rip = svm->vmcb->control.next_rip;
359 if (!svm->next_rip) {
360 if (!kvm_emulate_instruction(vcpu, EMULTYPE_SKIP))
363 kvm_rip_write(vcpu, svm->next_rip);
367 svm_set_interrupt_shadow(vcpu, 0);
372 static void svm_queue_exception(struct kvm_vcpu *vcpu)
374 struct vcpu_svm *svm = to_svm(vcpu);
375 unsigned nr = vcpu->arch.exception.nr;
376 bool has_error_code = vcpu->arch.exception.has_error_code;
377 u32 error_code = vcpu->arch.exception.error_code;
379 kvm_deliver_exception_payload(vcpu);
381 if (nr == BP_VECTOR && !nrips) {
382 unsigned long rip, old_rip = kvm_rip_read(vcpu);
385 * For guest debugging where we have to reinject #BP if some
386 * INT3 is guest-owned:
387 * Emulate nRIP by moving RIP forward. Will fail if injection
388 * raises a fault that is not intercepted. Still better than
389 * failing in all cases.
391 (void)skip_emulated_instruction(vcpu);
392 rip = kvm_rip_read(vcpu);
393 svm->int3_rip = rip + svm->vmcb->save.cs.base;
394 svm->int3_injected = rip - old_rip;
397 svm->vmcb->control.event_inj = nr
399 | (has_error_code ? SVM_EVTINJ_VALID_ERR : 0)
400 | SVM_EVTINJ_TYPE_EXEPT;
401 svm->vmcb->control.event_inj_err = error_code;
404 static void svm_init_erratum_383(void)
410 if (!static_cpu_has_bug(X86_BUG_AMD_TLB_MMATCH))
413 /* Use _safe variants to not break nested virtualization */
414 val = native_read_msr_safe(MSR_AMD64_DC_CFG, &err);
420 low = lower_32_bits(val);
421 high = upper_32_bits(val);
423 native_write_msr_safe(MSR_AMD64_DC_CFG, low, high);
425 erratum_383_found = true;
428 static void svm_init_osvw(struct kvm_vcpu *vcpu)
431 * Guests should see errata 400 and 415 as fixed (assuming that
432 * HLT and IO instructions are intercepted).
434 vcpu->arch.osvw.length = (osvw_len >= 3) ? (osvw_len) : 3;
435 vcpu->arch.osvw.status = osvw_status & ~(6ULL);
438 * By increasing VCPU's osvw.length to 3 we are telling the guest that
439 * all osvw.status bits inside that length, including bit 0 (which is
440 * reserved for erratum 298), are valid. However, if host processor's
441 * osvw_len is 0 then osvw_status[0] carries no information. We need to
442 * be conservative here and therefore we tell the guest that erratum 298
443 * is present (because we really don't know).
445 if (osvw_len == 0 && boot_cpu_data.x86 == 0x10)
446 vcpu->arch.osvw.status |= 1;
449 static int has_svm(void)
453 if (!cpu_has_svm(&msg)) {
454 printk(KERN_INFO "has_svm: %s\n", msg);
459 pr_info("KVM is unsupported when running as an SEV guest\n");
463 if (pgtable_l5_enabled()) {
464 pr_info("KVM doesn't yet support 5-level paging on AMD SVM\n");
471 static void svm_hardware_disable(void)
473 /* Make sure we clean up behind us */
474 if (static_cpu_has(X86_FEATURE_TSCRATEMSR))
475 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
479 amd_pmu_disable_virt();
482 static int svm_hardware_enable(void)
485 struct svm_cpu_data *sd;
487 struct desc_struct *gdt;
488 int me = raw_smp_processor_id();
490 rdmsrl(MSR_EFER, efer);
491 if (efer & EFER_SVME)
495 pr_err("%s: err EOPNOTSUPP on %d\n", __func__, me);
498 sd = per_cpu(svm_data, me);
500 pr_err("%s: svm_data is NULL on %d\n", __func__, me);
504 sd->asid_generation = 1;
505 sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
506 sd->next_asid = sd->max_asid + 1;
507 sd->min_asid = max_sev_asid + 1;
509 gdt = get_current_gdt_rw();
510 sd->tss_desc = (struct kvm_ldttss_desc *)(gdt + GDT_ENTRY_TSS);
512 wrmsrl(MSR_EFER, efer | EFER_SVME);
514 wrmsrl(MSR_VM_HSAVE_PA, __sme_page_pa(sd->save_area));
516 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
517 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
518 __this_cpu_write(current_tsc_ratio, TSC_RATIO_DEFAULT);
525 * Note that it is possible to have a system with mixed processor
526 * revisions and therefore different OSVW bits. If bits are not the same
527 * on different processors then choose the worst case (i.e. if erratum
528 * is present on one processor and not on another then assume that the
529 * erratum is present everywhere).
531 if (cpu_has(&boot_cpu_data, X86_FEATURE_OSVW)) {
532 uint64_t len, status = 0;
535 len = native_read_msr_safe(MSR_AMD64_OSVW_ID_LENGTH, &err);
537 status = native_read_msr_safe(MSR_AMD64_OSVW_STATUS,
541 osvw_status = osvw_len = 0;
545 osvw_status |= status;
546 osvw_status &= (1ULL << osvw_len) - 1;
549 osvw_status = osvw_len = 0;
551 svm_init_erratum_383();
553 amd_pmu_enable_virt();
558 static void svm_cpu_uninit(int cpu)
560 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
565 per_cpu(svm_data, cpu) = NULL;
566 kfree(sd->sev_vmcbs);
567 __free_page(sd->save_area);
571 static int svm_cpu_init(int cpu)
573 struct svm_cpu_data *sd;
576 sd = kzalloc(sizeof(struct svm_cpu_data), GFP_KERNEL);
580 sd->save_area = alloc_page(GFP_KERNEL);
584 clear_page(page_address(sd->save_area));
586 ret = sev_cpu_init(sd);
590 per_cpu(svm_data, cpu) = sd;
595 __free_page(sd->save_area);
602 static int direct_access_msr_slot(u32 msr)
606 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++)
607 if (direct_access_msrs[i].index == msr)
613 static void set_shadow_msr_intercept(struct kvm_vcpu *vcpu, u32 msr, int read,
616 struct vcpu_svm *svm = to_svm(vcpu);
617 int slot = direct_access_msr_slot(msr);
622 /* Set the shadow bitmaps to the desired intercept states */
624 set_bit(slot, svm->shadow_msr_intercept.read);
626 clear_bit(slot, svm->shadow_msr_intercept.read);
629 set_bit(slot, svm->shadow_msr_intercept.write);
631 clear_bit(slot, svm->shadow_msr_intercept.write);
634 static bool valid_msr_intercept(u32 index)
636 return direct_access_msr_slot(index) != -ENOENT;
639 static bool msr_write_intercepted(struct kvm_vcpu *vcpu, u32 msr)
646 msrpm = is_guest_mode(vcpu) ? to_svm(vcpu)->nested.msrpm:
649 offset = svm_msrpm_offset(msr);
650 bit_write = 2 * (msr & 0x0f) + 1;
653 BUG_ON(offset == MSR_INVALID);
655 return !!test_bit(bit_write, &tmp);
658 static void set_msr_interception_bitmap(struct kvm_vcpu *vcpu, u32 *msrpm,
659 u32 msr, int read, int write)
661 u8 bit_read, bit_write;
666 * If this warning triggers extend the direct_access_msrs list at the
667 * beginning of the file
669 WARN_ON(!valid_msr_intercept(msr));
671 /* Enforce non allowed MSRs to trap */
672 if (read && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_READ))
675 if (write && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_WRITE))
678 offset = svm_msrpm_offset(msr);
679 bit_read = 2 * (msr & 0x0f);
680 bit_write = 2 * (msr & 0x0f) + 1;
683 BUG_ON(offset == MSR_INVALID);
685 read ? clear_bit(bit_read, &tmp) : set_bit(bit_read, &tmp);
686 write ? clear_bit(bit_write, &tmp) : set_bit(bit_write, &tmp);
690 svm_hv_vmcb_dirty_nested_enlightenments(vcpu);
694 void set_msr_interception(struct kvm_vcpu *vcpu, u32 *msrpm, u32 msr,
697 set_shadow_msr_intercept(vcpu, msr, read, write);
698 set_msr_interception_bitmap(vcpu, msrpm, msr, read, write);
701 u32 *svm_vcpu_alloc_msrpm(void)
703 unsigned int order = get_order(MSRPM_SIZE);
704 struct page *pages = alloc_pages(GFP_KERNEL_ACCOUNT, order);
710 msrpm = page_address(pages);
711 memset(msrpm, 0xff, PAGE_SIZE * (1 << order));
716 void svm_vcpu_init_msrpm(struct kvm_vcpu *vcpu, u32 *msrpm)
720 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
721 if (!direct_access_msrs[i].always)
723 set_msr_interception(vcpu, msrpm, direct_access_msrs[i].index, 1, 1);
728 void svm_vcpu_free_msrpm(u32 *msrpm)
730 __free_pages(virt_to_page(msrpm), get_order(MSRPM_SIZE));
733 static void svm_msr_filter_changed(struct kvm_vcpu *vcpu)
735 struct vcpu_svm *svm = to_svm(vcpu);
739 * Set intercept permissions for all direct access MSRs again. They
740 * will automatically get filtered through the MSR filter, so we are
741 * back in sync after this.
743 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
744 u32 msr = direct_access_msrs[i].index;
745 u32 read = test_bit(i, svm->shadow_msr_intercept.read);
746 u32 write = test_bit(i, svm->shadow_msr_intercept.write);
748 set_msr_interception_bitmap(vcpu, svm->msrpm, msr, read, write);
752 static void add_msr_offset(u32 offset)
756 for (i = 0; i < MSRPM_OFFSETS; ++i) {
758 /* Offset already in list? */
759 if (msrpm_offsets[i] == offset)
762 /* Slot used by another offset? */
763 if (msrpm_offsets[i] != MSR_INVALID)
766 /* Add offset to list */
767 msrpm_offsets[i] = offset;
773 * If this BUG triggers the msrpm_offsets table has an overflow. Just
774 * increase MSRPM_OFFSETS in this case.
779 static void init_msrpm_offsets(void)
783 memset(msrpm_offsets, 0xff, sizeof(msrpm_offsets));
785 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
788 offset = svm_msrpm_offset(direct_access_msrs[i].index);
789 BUG_ON(offset == MSR_INVALID);
791 add_msr_offset(offset);
795 static void svm_enable_lbrv(struct kvm_vcpu *vcpu)
797 struct vcpu_svm *svm = to_svm(vcpu);
799 svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
800 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
801 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
802 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
803 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
806 static void svm_disable_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, 0, 0);
812 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 0, 0);
813 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 0, 0);
814 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 0, 0);
817 void disable_nmi_singlestep(struct vcpu_svm *svm)
819 svm->nmi_singlestep = false;
821 if (!(svm->vcpu.guest_debug & KVM_GUESTDBG_SINGLESTEP)) {
822 /* Clear our flags if they were not set by the guest */
823 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
824 svm->vmcb->save.rflags &= ~X86_EFLAGS_TF;
825 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
826 svm->vmcb->save.rflags &= ~X86_EFLAGS_RF;
830 static void grow_ple_window(struct kvm_vcpu *vcpu)
832 struct vcpu_svm *svm = to_svm(vcpu);
833 struct vmcb_control_area *control = &svm->vmcb->control;
834 int old = control->pause_filter_count;
836 control->pause_filter_count = __grow_ple_window(old,
838 pause_filter_count_grow,
839 pause_filter_count_max);
841 if (control->pause_filter_count != old) {
842 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
843 trace_kvm_ple_window_update(vcpu->vcpu_id,
844 control->pause_filter_count, old);
848 static void shrink_ple_window(struct kvm_vcpu *vcpu)
850 struct vcpu_svm *svm = to_svm(vcpu);
851 struct vmcb_control_area *control = &svm->vmcb->control;
852 int old = control->pause_filter_count;
854 control->pause_filter_count =
855 __shrink_ple_window(old,
857 pause_filter_count_shrink,
859 if (control->pause_filter_count != old) {
860 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
861 trace_kvm_ple_window_update(vcpu->vcpu_id,
862 control->pause_filter_count, old);
867 * The default MMIO mask is a single bit (excluding the present bit),
868 * which could conflict with the memory encryption bit. Check for
869 * memory encryption support and override the default MMIO mask if
870 * memory encryption is enabled.
872 static __init void svm_adjust_mmio_mask(void)
874 unsigned int enc_bit, mask_bit;
877 /* If there is no memory encryption support, use existing mask */
878 if (cpuid_eax(0x80000000) < 0x8000001f)
881 /* If memory encryption is not enabled, use existing mask */
882 rdmsrl(MSR_AMD64_SYSCFG, msr);
883 if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
886 enc_bit = cpuid_ebx(0x8000001f) & 0x3f;
887 mask_bit = boot_cpu_data.x86_phys_bits;
889 /* Increment the mask bit if it is the same as the encryption bit */
890 if (enc_bit == mask_bit)
894 * If the mask bit location is below 52, then some bits above the
895 * physical addressing limit will always be reserved, so use the
896 * rsvd_bits() function to generate the mask. This mask, along with
897 * the present bit, will be used to generate a page fault with
900 * If the mask bit location is 52 (or above), then clear the mask.
902 mask = (mask_bit < 52) ? rsvd_bits(mask_bit, 51) | PT_PRESENT_MASK : 0;
904 kvm_mmu_set_mmio_spte_mask(mask, mask, PT_WRITABLE_MASK | PT_USER_MASK);
907 static void svm_hardware_teardown(void)
911 sev_hardware_teardown();
913 for_each_possible_cpu(cpu)
916 __free_pages(pfn_to_page(iopm_base >> PAGE_SHIFT),
917 get_order(IOPM_SIZE));
921 static __init void svm_set_cpu_caps(void)
927 /* CPUID 0x80000001 and 0x8000000A (SVM features) */
929 kvm_cpu_cap_set(X86_FEATURE_SVM);
932 kvm_cpu_cap_set(X86_FEATURE_NRIPS);
935 kvm_cpu_cap_set(X86_FEATURE_NPT);
937 /* Nested VM can receive #VMEXIT instead of triggering #GP */
938 kvm_cpu_cap_set(X86_FEATURE_SVME_ADDR_CHK);
941 /* CPUID 0x80000008 */
942 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) ||
943 boot_cpu_has(X86_FEATURE_AMD_SSBD))
944 kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
946 /* CPUID 0x8000001F (SME/SEV features) */
950 static __init int svm_hardware_setup(void)
953 struct page *iopm_pages;
956 unsigned int order = get_order(IOPM_SIZE);
959 * NX is required for shadow paging and for NPT if the NX huge pages
960 * mitigation is enabled.
962 if (!boot_cpu_has(X86_FEATURE_NX)) {
963 pr_err_ratelimited("NX (Execute Disable) not supported\n");
966 kvm_enable_efer_bits(EFER_NX);
968 iopm_pages = alloc_pages(GFP_KERNEL, order);
973 iopm_va = page_address(iopm_pages);
974 memset(iopm_va, 0xff, PAGE_SIZE * (1 << order));
975 iopm_base = page_to_pfn(iopm_pages) << PAGE_SHIFT;
977 init_msrpm_offsets();
979 supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
981 if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
982 kvm_enable_efer_bits(EFER_FFXSR);
984 if (boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
985 kvm_has_tsc_control = true;
986 kvm_max_tsc_scaling_ratio = TSC_RATIO_MAX;
987 kvm_tsc_scaling_ratio_frac_bits = 32;
990 tsc_aux_uret_slot = kvm_add_user_return_msr(MSR_TSC_AUX);
992 /* Check for pause filtering support */
993 if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
994 pause_filter_count = 0;
995 pause_filter_thresh = 0;
996 } else if (!boot_cpu_has(X86_FEATURE_PFTHRESHOLD)) {
997 pause_filter_thresh = 0;
1001 printk(KERN_INFO "kvm: Nested Virtualization enabled\n");
1002 kvm_enable_efer_bits(EFER_SVME | EFER_LMSLE);
1006 * KVM's MMU doesn't support using 2-level paging for itself, and thus
1007 * NPT isn't supported if the host is using 2-level paging since host
1008 * CR4 is unchanged on VMRUN.
1010 if (!IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_X86_PAE))
1011 npt_enabled = false;
1013 if (!boot_cpu_has(X86_FEATURE_NPT))
1014 npt_enabled = false;
1016 /* Force VM NPT level equal to the host's max NPT level */
1017 kvm_configure_mmu(npt_enabled, get_max_npt_level(),
1018 get_max_npt_level(), PG_LEVEL_1G);
1019 pr_info("kvm: Nested Paging %sabled\n", npt_enabled ? "en" : "dis");
1021 /* Note, SEV setup consumes npt_enabled. */
1022 sev_hardware_setup();
1024 svm_hv_hardware_setup();
1026 svm_adjust_mmio_mask();
1028 for_each_possible_cpu(cpu) {
1029 r = svm_cpu_init(cpu);
1035 if (!boot_cpu_has(X86_FEATURE_NRIPS))
1039 enable_apicv = avic = avic && npt_enabled && boot_cpu_has(X86_FEATURE_AVIC);
1042 pr_info("AVIC enabled\n");
1044 amd_iommu_register_ga_log_notifier(&avic_ga_log_notifier);
1049 !boot_cpu_has(X86_FEATURE_V_VMSAVE_VMLOAD) ||
1050 !IS_ENABLED(CONFIG_X86_64)) {
1053 pr_info("Virtual VMLOAD VMSAVE supported\n");
1057 if (boot_cpu_has(X86_FEATURE_SVME_ADDR_CHK))
1058 svm_gp_erratum_intercept = false;
1061 if (!boot_cpu_has(X86_FEATURE_VGIF))
1064 pr_info("Virtual GIF supported\n");
1070 * It seems that on AMD processors PTE's accessed bit is
1071 * being set by the CPU hardware before the NPF vmexit.
1072 * This is not expected behaviour and our tests fail because
1074 * A workaround here is to disable support for
1075 * GUEST_MAXPHYADDR < HOST_MAXPHYADDR if NPT is enabled.
1076 * In this case userspace can know if there is support using
1077 * KVM_CAP_SMALLER_MAXPHYADDR extension and decide how to handle
1079 * If future AMD CPU models change the behaviour described above,
1080 * this variable can be changed accordingly
1082 allow_smaller_maxphyaddr = !npt_enabled;
1087 svm_hardware_teardown();
1091 static void init_seg(struct vmcb_seg *seg)
1094 seg->attrib = SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK |
1095 SVM_SELECTOR_WRITE_MASK; /* Read/Write Data Segment */
1096 seg->limit = 0xffff;
1100 static void init_sys_seg(struct vmcb_seg *seg, uint32_t type)
1103 seg->attrib = SVM_SELECTOR_P_MASK | type;
1104 seg->limit = 0xffff;
1108 static u64 svm_get_l2_tsc_offset(struct kvm_vcpu *vcpu)
1110 struct vcpu_svm *svm = to_svm(vcpu);
1112 return svm->nested.ctl.tsc_offset;
1115 static u64 svm_get_l2_tsc_multiplier(struct kvm_vcpu *vcpu)
1117 return kvm_default_tsc_scaling_ratio;
1120 static void svm_write_tsc_offset(struct kvm_vcpu *vcpu, u64 offset)
1122 struct vcpu_svm *svm = to_svm(vcpu);
1124 svm->vmcb01.ptr->control.tsc_offset = vcpu->arch.l1_tsc_offset;
1125 svm->vmcb->control.tsc_offset = offset;
1126 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1129 static void svm_write_tsc_multiplier(struct kvm_vcpu *vcpu, u64 multiplier)
1131 wrmsrl(MSR_AMD64_TSC_RATIO, multiplier);
1134 /* Evaluate instruction intercepts that depend on guest CPUID features. */
1135 static void svm_recalc_instruction_intercepts(struct kvm_vcpu *vcpu,
1136 struct vcpu_svm *svm)
1139 * Intercept INVPCID if shadow paging is enabled to sync/free shadow
1140 * roots, or if INVPCID is disabled in the guest to inject #UD.
1142 if (kvm_cpu_cap_has(X86_FEATURE_INVPCID)) {
1144 !guest_cpuid_has(&svm->vcpu, X86_FEATURE_INVPCID))
1145 svm_set_intercept(svm, INTERCEPT_INVPCID);
1147 svm_clr_intercept(svm, INTERCEPT_INVPCID);
1150 if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP)) {
1151 if (guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP))
1152 svm_clr_intercept(svm, INTERCEPT_RDTSCP);
1154 svm_set_intercept(svm, INTERCEPT_RDTSCP);
1158 static void init_vmcb(struct kvm_vcpu *vcpu)
1160 struct vcpu_svm *svm = to_svm(vcpu);
1161 struct vmcb_control_area *control = &svm->vmcb->control;
1162 struct vmcb_save_area *save = &svm->vmcb->save;
1164 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1165 svm_set_intercept(svm, INTERCEPT_CR3_READ);
1166 svm_set_intercept(svm, INTERCEPT_CR4_READ);
1167 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1168 svm_set_intercept(svm, INTERCEPT_CR3_WRITE);
1169 svm_set_intercept(svm, INTERCEPT_CR4_WRITE);
1170 if (!kvm_vcpu_apicv_active(vcpu))
1171 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
1173 set_dr_intercepts(svm);
1175 set_exception_intercept(svm, PF_VECTOR);
1176 set_exception_intercept(svm, UD_VECTOR);
1177 set_exception_intercept(svm, MC_VECTOR);
1178 set_exception_intercept(svm, AC_VECTOR);
1179 set_exception_intercept(svm, DB_VECTOR);
1181 * Guest access to VMware backdoor ports could legitimately
1182 * trigger #GP because of TSS I/O permission bitmap.
1183 * We intercept those #GP and allow access to them anyway
1186 if (enable_vmware_backdoor)
1187 set_exception_intercept(svm, GP_VECTOR);
1189 svm_set_intercept(svm, INTERCEPT_INTR);
1190 svm_set_intercept(svm, INTERCEPT_NMI);
1193 svm_set_intercept(svm, INTERCEPT_SMI);
1195 svm_set_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1196 svm_set_intercept(svm, INTERCEPT_RDPMC);
1197 svm_set_intercept(svm, INTERCEPT_CPUID);
1198 svm_set_intercept(svm, INTERCEPT_INVD);
1199 svm_set_intercept(svm, INTERCEPT_INVLPG);
1200 svm_set_intercept(svm, INTERCEPT_INVLPGA);
1201 svm_set_intercept(svm, INTERCEPT_IOIO_PROT);
1202 svm_set_intercept(svm, INTERCEPT_MSR_PROT);
1203 svm_set_intercept(svm, INTERCEPT_TASK_SWITCH);
1204 svm_set_intercept(svm, INTERCEPT_SHUTDOWN);
1205 svm_set_intercept(svm, INTERCEPT_VMRUN);
1206 svm_set_intercept(svm, INTERCEPT_VMMCALL);
1207 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1208 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1209 svm_set_intercept(svm, INTERCEPT_STGI);
1210 svm_set_intercept(svm, INTERCEPT_CLGI);
1211 svm_set_intercept(svm, INTERCEPT_SKINIT);
1212 svm_set_intercept(svm, INTERCEPT_WBINVD);
1213 svm_set_intercept(svm, INTERCEPT_XSETBV);
1214 svm_set_intercept(svm, INTERCEPT_RDPRU);
1215 svm_set_intercept(svm, INTERCEPT_RSM);
1217 if (!kvm_mwait_in_guest(vcpu->kvm)) {
1218 svm_set_intercept(svm, INTERCEPT_MONITOR);
1219 svm_set_intercept(svm, INTERCEPT_MWAIT);
1222 if (!kvm_hlt_in_guest(vcpu->kvm))
1223 svm_set_intercept(svm, INTERCEPT_HLT);
1225 control->iopm_base_pa = __sme_set(iopm_base);
1226 control->msrpm_base_pa = __sme_set(__pa(svm->msrpm));
1227 control->int_ctl = V_INTR_MASKING_MASK;
1229 init_seg(&save->es);
1230 init_seg(&save->ss);
1231 init_seg(&save->ds);
1232 init_seg(&save->fs);
1233 init_seg(&save->gs);
1235 save->cs.selector = 0xf000;
1236 save->cs.base = 0xffff0000;
1237 /* Executable/Readable Code Segment */
1238 save->cs.attrib = SVM_SELECTOR_READ_MASK | SVM_SELECTOR_P_MASK |
1239 SVM_SELECTOR_S_MASK | SVM_SELECTOR_CODE_MASK;
1240 save->cs.limit = 0xffff;
1242 save->gdtr.base = 0;
1243 save->gdtr.limit = 0xffff;
1244 save->idtr.base = 0;
1245 save->idtr.limit = 0xffff;
1247 init_sys_seg(&save->ldtr, SEG_TYPE_LDT);
1248 init_sys_seg(&save->tr, SEG_TYPE_BUSY_TSS16);
1251 /* Setup VMCB for Nested Paging */
1252 control->nested_ctl |= SVM_NESTED_CTL_NP_ENABLE;
1253 svm_clr_intercept(svm, INTERCEPT_INVLPG);
1254 clr_exception_intercept(svm, PF_VECTOR);
1255 svm_clr_intercept(svm, INTERCEPT_CR3_READ);
1256 svm_clr_intercept(svm, INTERCEPT_CR3_WRITE);
1257 save->g_pat = vcpu->arch.pat;
1260 svm->current_vmcb->asid_generation = 0;
1263 svm->nested.vmcb12_gpa = INVALID_GPA;
1264 svm->nested.last_vmcb12_gpa = INVALID_GPA;
1266 if (!kvm_pause_in_guest(vcpu->kvm)) {
1267 control->pause_filter_count = pause_filter_count;
1268 if (pause_filter_thresh)
1269 control->pause_filter_thresh = pause_filter_thresh;
1270 svm_set_intercept(svm, INTERCEPT_PAUSE);
1272 svm_clr_intercept(svm, INTERCEPT_PAUSE);
1275 svm_recalc_instruction_intercepts(vcpu, svm);
1278 * If the host supports V_SPEC_CTRL then disable the interception
1279 * of MSR_IA32_SPEC_CTRL.
1281 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
1282 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
1284 if (kvm_vcpu_apicv_active(vcpu))
1285 avic_init_vmcb(svm);
1288 svm_clr_intercept(svm, INTERCEPT_STGI);
1289 svm_clr_intercept(svm, INTERCEPT_CLGI);
1290 svm->vmcb->control.int_ctl |= V_GIF_ENABLE_MASK;
1293 if (sev_guest(vcpu->kvm)) {
1294 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
1295 clr_exception_intercept(svm, UD_VECTOR);
1297 if (sev_es_guest(vcpu->kvm)) {
1298 /* Perform SEV-ES specific VMCB updates */
1299 sev_es_init_vmcb(svm);
1303 svm_hv_init_vmcb(svm->vmcb);
1305 vmcb_mark_all_dirty(svm->vmcb);
1311 static void svm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
1313 struct vcpu_svm *svm = to_svm(vcpu);
1316 svm->virt_spec_ctrl = 0;
1321 void svm_switch_vmcb(struct vcpu_svm *svm, struct kvm_vmcb_info *target_vmcb)
1323 svm->current_vmcb = target_vmcb;
1324 svm->vmcb = target_vmcb->ptr;
1327 static int svm_create_vcpu(struct kvm_vcpu *vcpu)
1329 struct vcpu_svm *svm;
1330 struct page *vmcb01_page;
1331 struct page *vmsa_page = NULL;
1334 BUILD_BUG_ON(offsetof(struct vcpu_svm, vcpu) != 0);
1338 vmcb01_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1342 if (sev_es_guest(vcpu->kvm)) {
1344 * SEV-ES guests require a separate VMSA page used to contain
1345 * the encrypted register state of the guest.
1347 vmsa_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1349 goto error_free_vmcb_page;
1352 * SEV-ES guests maintain an encrypted version of their FPU
1353 * state which is restored and saved on VMRUN and VMEXIT.
1354 * Free the fpu structure to prevent KVM from attempting to
1355 * access the FPU state.
1357 kvm_free_guest_fpu(vcpu);
1360 err = avic_init_vcpu(svm);
1362 goto error_free_vmsa_page;
1364 /* We initialize this flag to true to make sure that the is_running
1365 * bit would be set the first time the vcpu is loaded.
1367 if (irqchip_in_kernel(vcpu->kvm) && kvm_apicv_activated(vcpu->kvm))
1368 svm->avic_is_running = true;
1370 svm->msrpm = svm_vcpu_alloc_msrpm();
1373 goto error_free_vmsa_page;
1376 svm->vmcb01.ptr = page_address(vmcb01_page);
1377 svm->vmcb01.pa = __sme_set(page_to_pfn(vmcb01_page) << PAGE_SHIFT);
1380 svm->vmsa = page_address(vmsa_page);
1382 svm->guest_state_loaded = false;
1384 svm_switch_vmcb(svm, &svm->vmcb01);
1387 svm_vcpu_init_msrpm(vcpu, svm->msrpm);
1389 svm_init_osvw(vcpu);
1390 vcpu->arch.microcode_version = 0x01000065;
1392 if (sev_es_guest(vcpu->kvm))
1393 /* Perform SEV-ES specific VMCB creation updates */
1394 sev_es_create_vcpu(svm);
1398 error_free_vmsa_page:
1400 __free_page(vmsa_page);
1401 error_free_vmcb_page:
1402 __free_page(vmcb01_page);
1407 static void svm_clear_current_vmcb(struct vmcb *vmcb)
1411 for_each_online_cpu(i)
1412 cmpxchg(&per_cpu(svm_data, i)->current_vmcb, vmcb, NULL);
1415 static void svm_free_vcpu(struct kvm_vcpu *vcpu)
1417 struct vcpu_svm *svm = to_svm(vcpu);
1420 * The vmcb page can be recycled, causing a false negative in
1421 * svm_vcpu_load(). So, ensure that no logical CPU has this
1422 * vmcb page recorded as its current vmcb.
1424 svm_clear_current_vmcb(svm->vmcb);
1426 svm_free_nested(svm);
1428 sev_free_vcpu(vcpu);
1430 __free_page(pfn_to_page(__sme_clr(svm->vmcb01.pa) >> PAGE_SHIFT));
1431 __free_pages(virt_to_page(svm->msrpm), get_order(MSRPM_SIZE));
1434 static void svm_prepare_guest_switch(struct kvm_vcpu *vcpu)
1436 struct vcpu_svm *svm = to_svm(vcpu);
1437 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
1439 if (sev_es_guest(vcpu->kvm))
1440 sev_es_unmap_ghcb(svm);
1442 if (svm->guest_state_loaded)
1446 * Save additional host state that will be restored on VMEXIT (sev-es)
1447 * or subsequent vmload of host save area.
1449 if (sev_es_guest(vcpu->kvm)) {
1450 sev_es_prepare_guest_switch(svm, vcpu->cpu);
1452 vmsave(__sme_page_pa(sd->save_area));
1455 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
1456 u64 tsc_ratio = vcpu->arch.tsc_scaling_ratio;
1457 if (tsc_ratio != __this_cpu_read(current_tsc_ratio)) {
1458 __this_cpu_write(current_tsc_ratio, tsc_ratio);
1459 wrmsrl(MSR_AMD64_TSC_RATIO, tsc_ratio);
1463 if (likely(tsc_aux_uret_slot >= 0))
1464 kvm_set_user_return_msr(tsc_aux_uret_slot, svm->tsc_aux, -1ull);
1466 svm->guest_state_loaded = true;
1469 static void svm_prepare_host_switch(struct kvm_vcpu *vcpu)
1471 to_svm(vcpu)->guest_state_loaded = false;
1474 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1476 struct vcpu_svm *svm = to_svm(vcpu);
1477 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1479 if (sd->current_vmcb != svm->vmcb) {
1480 sd->current_vmcb = svm->vmcb;
1481 indirect_branch_prediction_barrier();
1483 if (kvm_vcpu_apicv_active(vcpu))
1484 avic_vcpu_load(vcpu, cpu);
1487 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
1489 if (kvm_vcpu_apicv_active(vcpu))
1490 avic_vcpu_put(vcpu);
1492 svm_prepare_host_switch(vcpu);
1494 ++vcpu->stat.host_state_reload;
1497 static unsigned long svm_get_rflags(struct kvm_vcpu *vcpu)
1499 struct vcpu_svm *svm = to_svm(vcpu);
1500 unsigned long rflags = svm->vmcb->save.rflags;
1502 if (svm->nmi_singlestep) {
1503 /* Hide our flags if they were not set by the guest */
1504 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
1505 rflags &= ~X86_EFLAGS_TF;
1506 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
1507 rflags &= ~X86_EFLAGS_RF;
1512 static void svm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
1514 if (to_svm(vcpu)->nmi_singlestep)
1515 rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
1518 * Any change of EFLAGS.VM is accompanied by a reload of SS
1519 * (caused by either a task switch or an inter-privilege IRET),
1520 * so we do not need to update the CPL here.
1522 to_svm(vcpu)->vmcb->save.rflags = rflags;
1525 static void svm_cache_reg(struct kvm_vcpu *vcpu, enum kvm_reg reg)
1528 case VCPU_EXREG_PDPTR:
1529 BUG_ON(!npt_enabled);
1530 load_pdptrs(vcpu, vcpu->arch.walk_mmu, kvm_read_cr3(vcpu));
1533 KVM_BUG_ON(1, vcpu->kvm);
1537 static void svm_set_vintr(struct vcpu_svm *svm)
1539 struct vmcb_control_area *control;
1542 * The following fields are ignored when AVIC is enabled
1544 WARN_ON(kvm_apicv_activated(svm->vcpu.kvm));
1546 svm_set_intercept(svm, INTERCEPT_VINTR);
1549 * This is just a dummy VINTR to actually cause a vmexit to happen.
1550 * Actual injection of virtual interrupts happens through EVENTINJ.
1552 control = &svm->vmcb->control;
1553 control->int_vector = 0x0;
1554 control->int_ctl &= ~V_INTR_PRIO_MASK;
1555 control->int_ctl |= V_IRQ_MASK |
1556 ((/*control->int_vector >> 4*/ 0xf) << V_INTR_PRIO_SHIFT);
1557 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1560 static void svm_clear_vintr(struct vcpu_svm *svm)
1562 const u32 mask = V_TPR_MASK | V_GIF_ENABLE_MASK | V_GIF_MASK | V_INTR_MASKING_MASK;
1563 svm_clr_intercept(svm, INTERCEPT_VINTR);
1565 /* Drop int_ctl fields related to VINTR injection. */
1566 svm->vmcb->control.int_ctl &= mask;
1567 if (is_guest_mode(&svm->vcpu)) {
1568 svm->vmcb01.ptr->control.int_ctl &= mask;
1570 WARN_ON((svm->vmcb->control.int_ctl & V_TPR_MASK) !=
1571 (svm->nested.ctl.int_ctl & V_TPR_MASK));
1572 svm->vmcb->control.int_ctl |= svm->nested.ctl.int_ctl & ~mask;
1575 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1578 static struct vmcb_seg *svm_seg(struct kvm_vcpu *vcpu, int seg)
1580 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1581 struct vmcb_save_area *save01 = &to_svm(vcpu)->vmcb01.ptr->save;
1584 case VCPU_SREG_CS: return &save->cs;
1585 case VCPU_SREG_DS: return &save->ds;
1586 case VCPU_SREG_ES: return &save->es;
1587 case VCPU_SREG_FS: return &save01->fs;
1588 case VCPU_SREG_GS: return &save01->gs;
1589 case VCPU_SREG_SS: return &save->ss;
1590 case VCPU_SREG_TR: return &save01->tr;
1591 case VCPU_SREG_LDTR: return &save01->ldtr;
1597 static u64 svm_get_segment_base(struct kvm_vcpu *vcpu, int seg)
1599 struct vmcb_seg *s = svm_seg(vcpu, seg);
1604 static void svm_get_segment(struct kvm_vcpu *vcpu,
1605 struct kvm_segment *var, int seg)
1607 struct vmcb_seg *s = svm_seg(vcpu, seg);
1609 var->base = s->base;
1610 var->limit = s->limit;
1611 var->selector = s->selector;
1612 var->type = s->attrib & SVM_SELECTOR_TYPE_MASK;
1613 var->s = (s->attrib >> SVM_SELECTOR_S_SHIFT) & 1;
1614 var->dpl = (s->attrib >> SVM_SELECTOR_DPL_SHIFT) & 3;
1615 var->present = (s->attrib >> SVM_SELECTOR_P_SHIFT) & 1;
1616 var->avl = (s->attrib >> SVM_SELECTOR_AVL_SHIFT) & 1;
1617 var->l = (s->attrib >> SVM_SELECTOR_L_SHIFT) & 1;
1618 var->db = (s->attrib >> SVM_SELECTOR_DB_SHIFT) & 1;
1621 * AMD CPUs circa 2014 track the G bit for all segments except CS.
1622 * However, the SVM spec states that the G bit is not observed by the
1623 * CPU, and some VMware virtual CPUs drop the G bit for all segments.
1624 * So let's synthesize a legal G bit for all segments, this helps
1625 * running KVM nested. It also helps cross-vendor migration, because
1626 * Intel's vmentry has a check on the 'G' bit.
1628 var->g = s->limit > 0xfffff;
1631 * AMD's VMCB does not have an explicit unusable field, so emulate it
1632 * for cross vendor migration purposes by "not present"
1634 var->unusable = !var->present;
1639 * Work around a bug where the busy flag in the tr selector
1649 * The accessed bit must always be set in the segment
1650 * descriptor cache, although it can be cleared in the
1651 * descriptor, the cached bit always remains at 1. Since
1652 * Intel has a check on this, set it here to support
1653 * cross-vendor migration.
1660 * On AMD CPUs sometimes the DB bit in the segment
1661 * descriptor is left as 1, although the whole segment has
1662 * been made unusable. Clear it here to pass an Intel VMX
1663 * entry check when cross vendor migrating.
1667 /* This is symmetric with svm_set_segment() */
1668 var->dpl = to_svm(vcpu)->vmcb->save.cpl;
1673 static int svm_get_cpl(struct kvm_vcpu *vcpu)
1675 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1680 static void svm_get_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1682 struct vcpu_svm *svm = to_svm(vcpu);
1684 dt->size = svm->vmcb->save.idtr.limit;
1685 dt->address = svm->vmcb->save.idtr.base;
1688 static void svm_set_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1690 struct vcpu_svm *svm = to_svm(vcpu);
1692 svm->vmcb->save.idtr.limit = dt->size;
1693 svm->vmcb->save.idtr.base = dt->address ;
1694 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1697 static void svm_get_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1699 struct vcpu_svm *svm = to_svm(vcpu);
1701 dt->size = svm->vmcb->save.gdtr.limit;
1702 dt->address = svm->vmcb->save.gdtr.base;
1705 static void svm_set_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1707 struct vcpu_svm *svm = to_svm(vcpu);
1709 svm->vmcb->save.gdtr.limit = dt->size;
1710 svm->vmcb->save.gdtr.base = dt->address ;
1711 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1714 void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1716 struct vcpu_svm *svm = to_svm(vcpu);
1719 #ifdef CONFIG_X86_64
1720 if (vcpu->arch.efer & EFER_LME && !vcpu->arch.guest_state_protected) {
1721 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
1722 vcpu->arch.efer |= EFER_LMA;
1723 svm->vmcb->save.efer |= EFER_LMA | EFER_LME;
1726 if (is_paging(vcpu) && !(cr0 & X86_CR0_PG)) {
1727 vcpu->arch.efer &= ~EFER_LMA;
1728 svm->vmcb->save.efer &= ~(EFER_LMA | EFER_LME);
1732 vcpu->arch.cr0 = cr0;
1735 hcr0 |= X86_CR0_PG | X86_CR0_WP;
1738 * re-enable caching here because the QEMU bios
1739 * does not do it - this results in some delay at
1742 if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
1743 hcr0 &= ~(X86_CR0_CD | X86_CR0_NW);
1745 svm->vmcb->save.cr0 = hcr0;
1746 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1749 * SEV-ES guests must always keep the CR intercepts cleared. CR
1750 * tracking is done using the CR write traps.
1752 if (sev_es_guest(vcpu->kvm))
1756 /* Selective CR0 write remains on. */
1757 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1758 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1760 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1761 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1765 static bool svm_is_valid_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1770 void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1772 unsigned long host_cr4_mce = cr4_read_shadow() & X86_CR4_MCE;
1773 unsigned long old_cr4 = vcpu->arch.cr4;
1775 if (npt_enabled && ((old_cr4 ^ cr4) & X86_CR4_PGE))
1776 svm_flush_tlb(vcpu);
1778 vcpu->arch.cr4 = cr4;
1781 cr4 |= host_cr4_mce;
1782 to_svm(vcpu)->vmcb->save.cr4 = cr4;
1783 vmcb_mark_dirty(to_svm(vcpu)->vmcb, VMCB_CR);
1785 if ((cr4 ^ old_cr4) & (X86_CR4_OSXSAVE | X86_CR4_PKE))
1786 kvm_update_cpuid_runtime(vcpu);
1789 static void svm_set_segment(struct kvm_vcpu *vcpu,
1790 struct kvm_segment *var, int seg)
1792 struct vcpu_svm *svm = to_svm(vcpu);
1793 struct vmcb_seg *s = svm_seg(vcpu, seg);
1795 s->base = var->base;
1796 s->limit = var->limit;
1797 s->selector = var->selector;
1798 s->attrib = (var->type & SVM_SELECTOR_TYPE_MASK);
1799 s->attrib |= (var->s & 1) << SVM_SELECTOR_S_SHIFT;
1800 s->attrib |= (var->dpl & 3) << SVM_SELECTOR_DPL_SHIFT;
1801 s->attrib |= ((var->present & 1) && !var->unusable) << SVM_SELECTOR_P_SHIFT;
1802 s->attrib |= (var->avl & 1) << SVM_SELECTOR_AVL_SHIFT;
1803 s->attrib |= (var->l & 1) << SVM_SELECTOR_L_SHIFT;
1804 s->attrib |= (var->db & 1) << SVM_SELECTOR_DB_SHIFT;
1805 s->attrib |= (var->g & 1) << SVM_SELECTOR_G_SHIFT;
1808 * This is always accurate, except if SYSRET returned to a segment
1809 * with SS.DPL != 3. Intel does not have this quirk, and always
1810 * forces SS.DPL to 3 on sysret, so we ignore that case; fixing it
1811 * would entail passing the CPL to userspace and back.
1813 if (seg == VCPU_SREG_SS)
1814 /* This is symmetric with svm_get_segment() */
1815 svm->vmcb->save.cpl = (var->dpl & 3);
1817 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
1820 static void svm_update_exception_bitmap(struct kvm_vcpu *vcpu)
1822 struct vcpu_svm *svm = to_svm(vcpu);
1824 clr_exception_intercept(svm, BP_VECTOR);
1826 if (vcpu->guest_debug & KVM_GUESTDBG_ENABLE) {
1827 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP)
1828 set_exception_intercept(svm, BP_VECTOR);
1832 static void new_asid(struct vcpu_svm *svm, struct svm_cpu_data *sd)
1834 if (sd->next_asid > sd->max_asid) {
1835 ++sd->asid_generation;
1836 sd->next_asid = sd->min_asid;
1837 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
1838 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1841 svm->current_vmcb->asid_generation = sd->asid_generation;
1842 svm->asid = sd->next_asid++;
1845 static void svm_set_dr6(struct vcpu_svm *svm, unsigned long value)
1847 struct vmcb *vmcb = svm->vmcb;
1849 if (svm->vcpu.arch.guest_state_protected)
1852 if (unlikely(value != vmcb->save.dr6)) {
1853 vmcb->save.dr6 = value;
1854 vmcb_mark_dirty(vmcb, VMCB_DR);
1858 static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
1860 struct vcpu_svm *svm = to_svm(vcpu);
1862 if (vcpu->arch.guest_state_protected)
1865 get_debugreg(vcpu->arch.db[0], 0);
1866 get_debugreg(vcpu->arch.db[1], 1);
1867 get_debugreg(vcpu->arch.db[2], 2);
1868 get_debugreg(vcpu->arch.db[3], 3);
1870 * We cannot reset svm->vmcb->save.dr6 to DR6_ACTIVE_LOW here,
1871 * because db_interception might need it. We can do it before vmentry.
1873 vcpu->arch.dr6 = svm->vmcb->save.dr6;
1874 vcpu->arch.dr7 = svm->vmcb->save.dr7;
1875 vcpu->arch.switch_db_regs &= ~KVM_DEBUGREG_WONT_EXIT;
1876 set_dr_intercepts(svm);
1879 static void svm_set_dr7(struct kvm_vcpu *vcpu, unsigned long value)
1881 struct vcpu_svm *svm = to_svm(vcpu);
1883 if (vcpu->arch.guest_state_protected)
1886 svm->vmcb->save.dr7 = value;
1887 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
1890 static int pf_interception(struct kvm_vcpu *vcpu)
1892 struct vcpu_svm *svm = to_svm(vcpu);
1894 u64 fault_address = svm->vmcb->control.exit_info_2;
1895 u64 error_code = svm->vmcb->control.exit_info_1;
1897 return kvm_handle_page_fault(vcpu, error_code, fault_address,
1898 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1899 svm->vmcb->control.insn_bytes : NULL,
1900 svm->vmcb->control.insn_len);
1903 static int npf_interception(struct kvm_vcpu *vcpu)
1905 struct vcpu_svm *svm = to_svm(vcpu);
1907 u64 fault_address = svm->vmcb->control.exit_info_2;
1908 u64 error_code = svm->vmcb->control.exit_info_1;
1910 trace_kvm_page_fault(fault_address, error_code);
1911 return kvm_mmu_page_fault(vcpu, fault_address, error_code,
1912 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1913 svm->vmcb->control.insn_bytes : NULL,
1914 svm->vmcb->control.insn_len);
1917 static int db_interception(struct kvm_vcpu *vcpu)
1919 struct kvm_run *kvm_run = vcpu->run;
1920 struct vcpu_svm *svm = to_svm(vcpu);
1922 if (!(vcpu->guest_debug &
1923 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) &&
1924 !svm->nmi_singlestep) {
1925 u32 payload = svm->vmcb->save.dr6 ^ DR6_ACTIVE_LOW;
1926 kvm_queue_exception_p(vcpu, DB_VECTOR, payload);
1930 if (svm->nmi_singlestep) {
1931 disable_nmi_singlestep(svm);
1932 /* Make sure we check for pending NMIs upon entry */
1933 kvm_make_request(KVM_REQ_EVENT, vcpu);
1936 if (vcpu->guest_debug &
1937 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) {
1938 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1939 kvm_run->debug.arch.dr6 = svm->vmcb->save.dr6;
1940 kvm_run->debug.arch.dr7 = svm->vmcb->save.dr7;
1941 kvm_run->debug.arch.pc =
1942 svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1943 kvm_run->debug.arch.exception = DB_VECTOR;
1950 static int bp_interception(struct kvm_vcpu *vcpu)
1952 struct vcpu_svm *svm = to_svm(vcpu);
1953 struct kvm_run *kvm_run = vcpu->run;
1955 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1956 kvm_run->debug.arch.pc = svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1957 kvm_run->debug.arch.exception = BP_VECTOR;
1961 static int ud_interception(struct kvm_vcpu *vcpu)
1963 return handle_ud(vcpu);
1966 static int ac_interception(struct kvm_vcpu *vcpu)
1968 kvm_queue_exception_e(vcpu, AC_VECTOR, 0);
1972 static bool is_erratum_383(void)
1977 if (!erratum_383_found)
1980 value = native_read_msr_safe(MSR_IA32_MC0_STATUS, &err);
1984 /* Bit 62 may or may not be set for this mce */
1985 value &= ~(1ULL << 62);
1987 if (value != 0xb600000000010015ULL)
1990 /* Clear MCi_STATUS registers */
1991 for (i = 0; i < 6; ++i)
1992 native_write_msr_safe(MSR_IA32_MCx_STATUS(i), 0, 0);
1994 value = native_read_msr_safe(MSR_IA32_MCG_STATUS, &err);
1998 value &= ~(1ULL << 2);
1999 low = lower_32_bits(value);
2000 high = upper_32_bits(value);
2002 native_write_msr_safe(MSR_IA32_MCG_STATUS, low, high);
2005 /* Flush tlb to evict multi-match entries */
2011 static void svm_handle_mce(struct kvm_vcpu *vcpu)
2013 if (is_erratum_383()) {
2015 * Erratum 383 triggered. Guest state is corrupt so kill the
2018 pr_err("KVM: Guest triggered AMD Erratum 383\n");
2020 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2026 * On an #MC intercept the MCE handler is not called automatically in
2027 * the host. So do it by hand here.
2029 kvm_machine_check();
2032 static int mc_interception(struct kvm_vcpu *vcpu)
2037 static int shutdown_interception(struct kvm_vcpu *vcpu)
2039 struct kvm_run *kvm_run = vcpu->run;
2040 struct vcpu_svm *svm = to_svm(vcpu);
2043 * The VM save area has already been encrypted so it
2044 * cannot be reinitialized - just terminate.
2046 if (sev_es_guest(vcpu->kvm))
2050 * VMCB is undefined after a SHUTDOWN intercept. INIT the vCPU to put
2051 * the VMCB in a known good state. Unfortuately, KVM doesn't have
2052 * KVM_MP_STATE_SHUTDOWN and can't add it without potentially breaking
2053 * userspace. At a platform view, INIT is acceptable behavior as
2054 * there exist bare metal platforms that automatically INIT the CPU
2055 * in response to shutdown.
2057 clear_page(svm->vmcb);
2058 kvm_vcpu_reset(vcpu, true);
2060 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
2064 static int io_interception(struct kvm_vcpu *vcpu)
2066 struct vcpu_svm *svm = to_svm(vcpu);
2067 u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
2068 int size, in, string;
2071 ++vcpu->stat.io_exits;
2072 string = (io_info & SVM_IOIO_STR_MASK) != 0;
2073 in = (io_info & SVM_IOIO_TYPE_MASK) != 0;
2074 port = io_info >> 16;
2075 size = (io_info & SVM_IOIO_SIZE_MASK) >> SVM_IOIO_SIZE_SHIFT;
2078 if (sev_es_guest(vcpu->kvm))
2079 return sev_es_string_io(svm, size, port, in);
2081 return kvm_emulate_instruction(vcpu, 0);
2084 svm->next_rip = svm->vmcb->control.exit_info_2;
2086 return kvm_fast_pio(vcpu, size, port, in);
2089 static int nmi_interception(struct kvm_vcpu *vcpu)
2094 static int smi_interception(struct kvm_vcpu *vcpu)
2099 static int intr_interception(struct kvm_vcpu *vcpu)
2101 ++vcpu->stat.irq_exits;
2105 static int vmload_vmsave_interception(struct kvm_vcpu *vcpu, bool vmload)
2107 struct vcpu_svm *svm = to_svm(vcpu);
2108 struct vmcb *vmcb12;
2109 struct kvm_host_map map;
2112 if (nested_svm_check_permissions(vcpu))
2115 ret = kvm_vcpu_map(vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2118 kvm_inject_gp(vcpu, 0);
2124 ret = kvm_skip_emulated_instruction(vcpu);
2127 svm_copy_vmloadsave_state(svm->vmcb, vmcb12);
2128 svm->sysenter_eip_hi = 0;
2129 svm->sysenter_esp_hi = 0;
2131 svm_copy_vmloadsave_state(vmcb12, svm->vmcb);
2134 kvm_vcpu_unmap(vcpu, &map, true);
2139 static int vmload_interception(struct kvm_vcpu *vcpu)
2141 return vmload_vmsave_interception(vcpu, true);
2144 static int vmsave_interception(struct kvm_vcpu *vcpu)
2146 return vmload_vmsave_interception(vcpu, false);
2149 static int vmrun_interception(struct kvm_vcpu *vcpu)
2151 if (nested_svm_check_permissions(vcpu))
2154 return nested_svm_vmrun(vcpu);
2164 /* Return NONE_SVM_INSTR if not SVM instrs, otherwise return decode result */
2165 static int svm_instr_opcode(struct kvm_vcpu *vcpu)
2167 struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt;
2169 if (ctxt->b != 0x1 || ctxt->opcode_len != 2)
2170 return NONE_SVM_INSTR;
2172 switch (ctxt->modrm) {
2173 case 0xd8: /* VMRUN */
2174 return SVM_INSTR_VMRUN;
2175 case 0xda: /* VMLOAD */
2176 return SVM_INSTR_VMLOAD;
2177 case 0xdb: /* VMSAVE */
2178 return SVM_INSTR_VMSAVE;
2183 return NONE_SVM_INSTR;
2186 static int emulate_svm_instr(struct kvm_vcpu *vcpu, int opcode)
2188 const int guest_mode_exit_codes[] = {
2189 [SVM_INSTR_VMRUN] = SVM_EXIT_VMRUN,
2190 [SVM_INSTR_VMLOAD] = SVM_EXIT_VMLOAD,
2191 [SVM_INSTR_VMSAVE] = SVM_EXIT_VMSAVE,
2193 int (*const svm_instr_handlers[])(struct kvm_vcpu *vcpu) = {
2194 [SVM_INSTR_VMRUN] = vmrun_interception,
2195 [SVM_INSTR_VMLOAD] = vmload_interception,
2196 [SVM_INSTR_VMSAVE] = vmsave_interception,
2198 struct vcpu_svm *svm = to_svm(vcpu);
2201 if (is_guest_mode(vcpu)) {
2202 /* Returns '1' or -errno on failure, '0' on success. */
2203 ret = nested_svm_simple_vmexit(svm, guest_mode_exit_codes[opcode]);
2208 return svm_instr_handlers[opcode](vcpu);
2212 * #GP handling code. Note that #GP can be triggered under the following two
2214 * 1) SVM VM-related instructions (VMRUN/VMSAVE/VMLOAD) that trigger #GP on
2215 * some AMD CPUs when EAX of these instructions are in the reserved memory
2216 * regions (e.g. SMM memory on host).
2217 * 2) VMware backdoor
2219 static int gp_interception(struct kvm_vcpu *vcpu)
2221 struct vcpu_svm *svm = to_svm(vcpu);
2222 u32 error_code = svm->vmcb->control.exit_info_1;
2225 /* Both #GP cases have zero error_code */
2229 /* Decode the instruction for usage later */
2230 if (x86_decode_emulated_instruction(vcpu, 0, NULL, 0) != EMULATION_OK)
2233 opcode = svm_instr_opcode(vcpu);
2235 if (opcode == NONE_SVM_INSTR) {
2236 if (!enable_vmware_backdoor)
2240 * VMware backdoor emulation on #GP interception only handles
2241 * IN{S}, OUT{S}, and RDPMC.
2243 if (!is_guest_mode(vcpu))
2244 return kvm_emulate_instruction(vcpu,
2245 EMULTYPE_VMWARE_GP | EMULTYPE_NO_DECODE);
2247 return emulate_svm_instr(vcpu, opcode);
2250 kvm_queue_exception_e(vcpu, GP_VECTOR, error_code);
2254 void svm_set_gif(struct vcpu_svm *svm, bool value)
2258 * If VGIF is enabled, the STGI intercept is only added to
2259 * detect the opening of the SMI/NMI window; remove it now.
2260 * Likewise, clear the VINTR intercept, we will set it
2261 * again while processing KVM_REQ_EVENT if needed.
2263 if (vgif_enabled(svm))
2264 svm_clr_intercept(svm, INTERCEPT_STGI);
2265 if (svm_is_intercept(svm, INTERCEPT_VINTR))
2266 svm_clear_vintr(svm);
2269 if (svm->vcpu.arch.smi_pending ||
2270 svm->vcpu.arch.nmi_pending ||
2271 kvm_cpu_has_injectable_intr(&svm->vcpu))
2272 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2277 * After a CLGI no interrupts should come. But if vGIF is
2278 * in use, we still rely on the VINTR intercept (rather than
2279 * STGI) to detect an open interrupt window.
2281 if (!vgif_enabled(svm))
2282 svm_clear_vintr(svm);
2286 static int stgi_interception(struct kvm_vcpu *vcpu)
2290 if (nested_svm_check_permissions(vcpu))
2293 ret = kvm_skip_emulated_instruction(vcpu);
2294 svm_set_gif(to_svm(vcpu), true);
2298 static int clgi_interception(struct kvm_vcpu *vcpu)
2302 if (nested_svm_check_permissions(vcpu))
2305 ret = kvm_skip_emulated_instruction(vcpu);
2306 svm_set_gif(to_svm(vcpu), false);
2310 static int invlpga_interception(struct kvm_vcpu *vcpu)
2312 gva_t gva = kvm_rax_read(vcpu);
2313 u32 asid = kvm_rcx_read(vcpu);
2315 /* FIXME: Handle an address size prefix. */
2316 if (!is_long_mode(vcpu))
2319 trace_kvm_invlpga(to_svm(vcpu)->vmcb->save.rip, asid, gva);
2321 /* Let's treat INVLPGA the same as INVLPG (can be optimized!) */
2322 kvm_mmu_invlpg(vcpu, gva);
2324 return kvm_skip_emulated_instruction(vcpu);
2327 static int skinit_interception(struct kvm_vcpu *vcpu)
2329 trace_kvm_skinit(to_svm(vcpu)->vmcb->save.rip, kvm_rax_read(vcpu));
2331 kvm_queue_exception(vcpu, UD_VECTOR);
2335 static int task_switch_interception(struct kvm_vcpu *vcpu)
2337 struct vcpu_svm *svm = to_svm(vcpu);
2340 int int_type = svm->vmcb->control.exit_int_info &
2341 SVM_EXITINTINFO_TYPE_MASK;
2342 int int_vec = svm->vmcb->control.exit_int_info & SVM_EVTINJ_VEC_MASK;
2344 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_TYPE_MASK;
2346 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID;
2347 bool has_error_code = false;
2350 tss_selector = (u16)svm->vmcb->control.exit_info_1;
2352 if (svm->vmcb->control.exit_info_2 &
2353 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
2354 reason = TASK_SWITCH_IRET;
2355 else if (svm->vmcb->control.exit_info_2 &
2356 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
2357 reason = TASK_SWITCH_JMP;
2359 reason = TASK_SWITCH_GATE;
2361 reason = TASK_SWITCH_CALL;
2363 if (reason == TASK_SWITCH_GATE) {
2365 case SVM_EXITINTINFO_TYPE_NMI:
2366 vcpu->arch.nmi_injected = false;
2368 case SVM_EXITINTINFO_TYPE_EXEPT:
2369 if (svm->vmcb->control.exit_info_2 &
2370 (1ULL << SVM_EXITINFOSHIFT_TS_HAS_ERROR_CODE)) {
2371 has_error_code = true;
2373 (u32)svm->vmcb->control.exit_info_2;
2375 kvm_clear_exception_queue(vcpu);
2377 case SVM_EXITINTINFO_TYPE_INTR:
2378 kvm_clear_interrupt_queue(vcpu);
2385 if (reason != TASK_SWITCH_GATE ||
2386 int_type == SVM_EXITINTINFO_TYPE_SOFT ||
2387 (int_type == SVM_EXITINTINFO_TYPE_EXEPT &&
2388 (int_vec == OF_VECTOR || int_vec == BP_VECTOR))) {
2389 if (!skip_emulated_instruction(vcpu))
2393 if (int_type != SVM_EXITINTINFO_TYPE_SOFT)
2396 return kvm_task_switch(vcpu, tss_selector, int_vec, reason,
2397 has_error_code, error_code);
2400 static int iret_interception(struct kvm_vcpu *vcpu)
2402 struct vcpu_svm *svm = to_svm(vcpu);
2404 ++vcpu->stat.nmi_window_exits;
2405 vcpu->arch.hflags |= HF_IRET_MASK;
2406 if (!sev_es_guest(vcpu->kvm)) {
2407 svm_clr_intercept(svm, INTERCEPT_IRET);
2408 svm->nmi_iret_rip = kvm_rip_read(vcpu);
2410 kvm_make_request(KVM_REQ_EVENT, vcpu);
2414 static int invlpg_interception(struct kvm_vcpu *vcpu)
2416 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2417 return kvm_emulate_instruction(vcpu, 0);
2419 kvm_mmu_invlpg(vcpu, to_svm(vcpu)->vmcb->control.exit_info_1);
2420 return kvm_skip_emulated_instruction(vcpu);
2423 static int emulate_on_interception(struct kvm_vcpu *vcpu)
2425 return kvm_emulate_instruction(vcpu, 0);
2428 static int rsm_interception(struct kvm_vcpu *vcpu)
2430 return kvm_emulate_instruction_from_buffer(vcpu, rsm_ins_bytes, 2);
2433 static bool check_selective_cr0_intercepted(struct kvm_vcpu *vcpu,
2436 struct vcpu_svm *svm = to_svm(vcpu);
2437 unsigned long cr0 = vcpu->arch.cr0;
2440 if (!is_guest_mode(vcpu) ||
2441 (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_SELECTIVE_CR0))))
2444 cr0 &= ~SVM_CR0_SELECTIVE_MASK;
2445 val &= ~SVM_CR0_SELECTIVE_MASK;
2448 svm->vmcb->control.exit_code = SVM_EXIT_CR0_SEL_WRITE;
2449 ret = (nested_svm_exit_handled(svm) == NESTED_EXIT_DONE);
2455 #define CR_VALID (1ULL << 63)
2457 static int cr_interception(struct kvm_vcpu *vcpu)
2459 struct vcpu_svm *svm = to_svm(vcpu);
2464 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2465 return emulate_on_interception(vcpu);
2467 if (unlikely((svm->vmcb->control.exit_info_1 & CR_VALID) == 0))
2468 return emulate_on_interception(vcpu);
2470 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2471 if (svm->vmcb->control.exit_code == SVM_EXIT_CR0_SEL_WRITE)
2472 cr = SVM_EXIT_WRITE_CR0 - SVM_EXIT_READ_CR0;
2474 cr = svm->vmcb->control.exit_code - SVM_EXIT_READ_CR0;
2477 if (cr >= 16) { /* mov to cr */
2479 val = kvm_register_read(vcpu, reg);
2480 trace_kvm_cr_write(cr, val);
2483 if (!check_selective_cr0_intercepted(vcpu, val))
2484 err = kvm_set_cr0(vcpu, val);
2490 err = kvm_set_cr3(vcpu, val);
2493 err = kvm_set_cr4(vcpu, val);
2496 err = kvm_set_cr8(vcpu, val);
2499 WARN(1, "unhandled write to CR%d", cr);
2500 kvm_queue_exception(vcpu, UD_VECTOR);
2503 } else { /* mov from cr */
2506 val = kvm_read_cr0(vcpu);
2509 val = vcpu->arch.cr2;
2512 val = kvm_read_cr3(vcpu);
2515 val = kvm_read_cr4(vcpu);
2518 val = kvm_get_cr8(vcpu);
2521 WARN(1, "unhandled read from CR%d", cr);
2522 kvm_queue_exception(vcpu, UD_VECTOR);
2525 kvm_register_write(vcpu, reg, val);
2526 trace_kvm_cr_read(cr, val);
2528 return kvm_complete_insn_gp(vcpu, err);
2531 static int cr_trap(struct kvm_vcpu *vcpu)
2533 struct vcpu_svm *svm = to_svm(vcpu);
2534 unsigned long old_value, new_value;
2538 new_value = (unsigned long)svm->vmcb->control.exit_info_1;
2540 cr = svm->vmcb->control.exit_code - SVM_EXIT_CR0_WRITE_TRAP;
2543 old_value = kvm_read_cr0(vcpu);
2544 svm_set_cr0(vcpu, new_value);
2546 kvm_post_set_cr0(vcpu, old_value, new_value);
2549 old_value = kvm_read_cr4(vcpu);
2550 svm_set_cr4(vcpu, new_value);
2552 kvm_post_set_cr4(vcpu, old_value, new_value);
2555 ret = kvm_set_cr8(vcpu, new_value);
2558 WARN(1, "unhandled CR%d write trap", cr);
2559 kvm_queue_exception(vcpu, UD_VECTOR);
2563 return kvm_complete_insn_gp(vcpu, ret);
2566 static int dr_interception(struct kvm_vcpu *vcpu)
2568 struct vcpu_svm *svm = to_svm(vcpu);
2573 if (vcpu->guest_debug == 0) {
2575 * No more DR vmexits; force a reload of the debug registers
2576 * and reenter on this instruction. The next vmexit will
2577 * retrieve the full state of the debug registers.
2579 clr_dr_intercepts(svm);
2580 vcpu->arch.switch_db_regs |= KVM_DEBUGREG_WONT_EXIT;
2584 if (!boot_cpu_has(X86_FEATURE_DECODEASSISTS))
2585 return emulate_on_interception(vcpu);
2587 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2588 dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
2589 if (dr >= 16) { /* mov to DRn */
2591 val = kvm_register_read(vcpu, reg);
2592 err = kvm_set_dr(vcpu, dr, val);
2594 kvm_get_dr(vcpu, dr, &val);
2595 kvm_register_write(vcpu, reg, val);
2598 return kvm_complete_insn_gp(vcpu, err);
2601 static int cr8_write_interception(struct kvm_vcpu *vcpu)
2605 u8 cr8_prev = kvm_get_cr8(vcpu);
2606 /* instruction emulation calls kvm_set_cr8() */
2607 r = cr_interception(vcpu);
2608 if (lapic_in_kernel(vcpu))
2610 if (cr8_prev <= kvm_get_cr8(vcpu))
2612 vcpu->run->exit_reason = KVM_EXIT_SET_TPR;
2616 static int efer_trap(struct kvm_vcpu *vcpu)
2618 struct msr_data msr_info;
2622 * Clear the EFER_SVME bit from EFER. The SVM code always sets this
2623 * bit in svm_set_efer(), but __kvm_valid_efer() checks it against
2624 * whether the guest has X86_FEATURE_SVM - this avoids a failure if
2625 * the guest doesn't have X86_FEATURE_SVM.
2627 msr_info.host_initiated = false;
2628 msr_info.index = MSR_EFER;
2629 msr_info.data = to_svm(vcpu)->vmcb->control.exit_info_1 & ~EFER_SVME;
2630 ret = kvm_set_msr_common(vcpu, &msr_info);
2632 return kvm_complete_insn_gp(vcpu, ret);
2635 static int svm_get_msr_feature(struct kvm_msr_entry *msr)
2639 switch (msr->index) {
2640 case MSR_F10H_DECFG:
2641 if (boot_cpu_has(X86_FEATURE_LFENCE_RDTSC))
2642 msr->data |= MSR_F10H_DECFG_LFENCE_SERIALIZE;
2644 case MSR_IA32_PERF_CAPABILITIES:
2647 return KVM_MSR_RET_INVALID;
2653 static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
2655 struct vcpu_svm *svm = to_svm(vcpu);
2657 switch (msr_info->index) {
2659 msr_info->data = svm->vmcb01.ptr->save.star;
2661 #ifdef CONFIG_X86_64
2663 msr_info->data = svm->vmcb01.ptr->save.lstar;
2666 msr_info->data = svm->vmcb01.ptr->save.cstar;
2668 case MSR_KERNEL_GS_BASE:
2669 msr_info->data = svm->vmcb01.ptr->save.kernel_gs_base;
2671 case MSR_SYSCALL_MASK:
2672 msr_info->data = svm->vmcb01.ptr->save.sfmask;
2675 case MSR_IA32_SYSENTER_CS:
2676 msr_info->data = svm->vmcb01.ptr->save.sysenter_cs;
2678 case MSR_IA32_SYSENTER_EIP:
2679 msr_info->data = (u32)svm->vmcb01.ptr->save.sysenter_eip;
2680 if (guest_cpuid_is_intel(vcpu))
2681 msr_info->data |= (u64)svm->sysenter_eip_hi << 32;
2683 case MSR_IA32_SYSENTER_ESP:
2684 msr_info->data = svm->vmcb01.ptr->save.sysenter_esp;
2685 if (guest_cpuid_is_intel(vcpu))
2686 msr_info->data |= (u64)svm->sysenter_esp_hi << 32;
2689 msr_info->data = svm->tsc_aux;
2692 * Nobody will change the following 5 values in the VMCB so we can
2693 * safely return them on rdmsr. They will always be 0 until LBRV is
2696 case MSR_IA32_DEBUGCTLMSR:
2697 msr_info->data = svm->vmcb->save.dbgctl;
2699 case MSR_IA32_LASTBRANCHFROMIP:
2700 msr_info->data = svm->vmcb->save.br_from;
2702 case MSR_IA32_LASTBRANCHTOIP:
2703 msr_info->data = svm->vmcb->save.br_to;
2705 case MSR_IA32_LASTINTFROMIP:
2706 msr_info->data = svm->vmcb->save.last_excp_from;
2708 case MSR_IA32_LASTINTTOIP:
2709 msr_info->data = svm->vmcb->save.last_excp_to;
2711 case MSR_VM_HSAVE_PA:
2712 msr_info->data = svm->nested.hsave_msr;
2715 msr_info->data = svm->nested.vm_cr_msr;
2717 case MSR_IA32_SPEC_CTRL:
2718 if (!msr_info->host_initiated &&
2719 !guest_has_spec_ctrl_msr(vcpu))
2722 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2723 msr_info->data = svm->vmcb->save.spec_ctrl;
2725 msr_info->data = svm->spec_ctrl;
2727 case MSR_AMD64_VIRT_SPEC_CTRL:
2728 if (!msr_info->host_initiated &&
2729 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2732 msr_info->data = svm->virt_spec_ctrl;
2734 case MSR_F15H_IC_CFG: {
2738 family = guest_cpuid_family(vcpu);
2739 model = guest_cpuid_model(vcpu);
2741 if (family < 0 || model < 0)
2742 return kvm_get_msr_common(vcpu, msr_info);
2746 if (family == 0x15 &&
2747 (model >= 0x2 && model < 0x20))
2748 msr_info->data = 0x1E;
2751 case MSR_F10H_DECFG:
2752 msr_info->data = svm->msr_decfg;
2755 return kvm_get_msr_common(vcpu, msr_info);
2760 static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
2762 struct vcpu_svm *svm = to_svm(vcpu);
2763 if (!err || !sev_es_guest(vcpu->kvm) || WARN_ON_ONCE(!svm->ghcb))
2764 return kvm_complete_insn_gp(vcpu, err);
2766 ghcb_set_sw_exit_info_1(svm->ghcb, 1);
2767 ghcb_set_sw_exit_info_2(svm->ghcb,
2769 SVM_EVTINJ_TYPE_EXEPT |
2774 static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
2776 struct vcpu_svm *svm = to_svm(vcpu);
2777 int svm_dis, chg_mask;
2779 if (data & ~SVM_VM_CR_VALID_MASK)
2782 chg_mask = SVM_VM_CR_VALID_MASK;
2784 if (svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK)
2785 chg_mask &= ~(SVM_VM_CR_SVM_LOCK_MASK | SVM_VM_CR_SVM_DIS_MASK);
2787 svm->nested.vm_cr_msr &= ~chg_mask;
2788 svm->nested.vm_cr_msr |= (data & chg_mask);
2790 svm_dis = svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK;
2792 /* check for svm_disable while efer.svme is set */
2793 if (svm_dis && (vcpu->arch.efer & EFER_SVME))
2799 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
2801 struct vcpu_svm *svm = to_svm(vcpu);
2804 u32 ecx = msr->index;
2805 u64 data = msr->data;
2807 case MSR_IA32_CR_PAT:
2808 if (!kvm_mtrr_valid(vcpu, MSR_IA32_CR_PAT, data))
2810 vcpu->arch.pat = data;
2811 svm->vmcb01.ptr->save.g_pat = data;
2812 if (is_guest_mode(vcpu))
2813 nested_vmcb02_compute_g_pat(svm);
2814 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
2816 case MSR_IA32_SPEC_CTRL:
2817 if (!msr->host_initiated &&
2818 !guest_has_spec_ctrl_msr(vcpu))
2821 if (kvm_spec_ctrl_test_value(data))
2824 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2825 svm->vmcb->save.spec_ctrl = data;
2827 svm->spec_ctrl = data;
2833 * When it's written (to non-zero) for the first time, pass
2837 * The handling of the MSR bitmap for L2 guests is done in
2838 * nested_svm_vmrun_msrpm.
2839 * We update the L1 MSR bit as well since it will end up
2840 * touching the MSR anyway now.
2842 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
2844 case MSR_IA32_PRED_CMD:
2845 if (!msr->host_initiated &&
2846 !guest_has_pred_cmd_msr(vcpu))
2849 if (data & ~PRED_CMD_IBPB)
2851 if (!boot_cpu_has(X86_FEATURE_IBPB))
2856 wrmsrl(MSR_IA32_PRED_CMD, PRED_CMD_IBPB);
2857 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_PRED_CMD, 0, 1);
2859 case MSR_AMD64_VIRT_SPEC_CTRL:
2860 if (!msr->host_initiated &&
2861 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2864 if (data & ~SPEC_CTRL_SSBD)
2867 svm->virt_spec_ctrl = data;
2870 svm->vmcb01.ptr->save.star = data;
2872 #ifdef CONFIG_X86_64
2874 svm->vmcb01.ptr->save.lstar = data;
2877 svm->vmcb01.ptr->save.cstar = data;
2879 case MSR_KERNEL_GS_BASE:
2880 svm->vmcb01.ptr->save.kernel_gs_base = data;
2882 case MSR_SYSCALL_MASK:
2883 svm->vmcb01.ptr->save.sfmask = data;
2886 case MSR_IA32_SYSENTER_CS:
2887 svm->vmcb01.ptr->save.sysenter_cs = data;
2889 case MSR_IA32_SYSENTER_EIP:
2890 svm->vmcb01.ptr->save.sysenter_eip = (u32)data;
2892 * We only intercept the MSR_IA32_SYSENTER_{EIP|ESP} msrs
2893 * when we spoof an Intel vendor ID (for cross vendor migration).
2894 * In this case we use this intercept to track the high
2895 * 32 bit part of these msrs to support Intel's
2896 * implementation of SYSENTER/SYSEXIT.
2898 svm->sysenter_eip_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2900 case MSR_IA32_SYSENTER_ESP:
2901 svm->vmcb01.ptr->save.sysenter_esp = (u32)data;
2902 svm->sysenter_esp_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2906 * TSC_AUX is usually changed only during boot and never read
2907 * directly. Intercept TSC_AUX instead of exposing it to the
2908 * guest via direct_access_msrs, and switch it via user return.
2911 r = kvm_set_user_return_msr(tsc_aux_uret_slot, data, -1ull);
2916 svm->tsc_aux = data;
2918 case MSR_IA32_DEBUGCTLMSR:
2919 if (!boot_cpu_has(X86_FEATURE_LBRV)) {
2920 vcpu_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTL 0x%llx, nop\n",
2924 if (data & DEBUGCTL_RESERVED_BITS)
2927 svm->vmcb->save.dbgctl = data;
2928 vmcb_mark_dirty(svm->vmcb, VMCB_LBR);
2929 if (data & (1ULL<<0))
2930 svm_enable_lbrv(vcpu);
2932 svm_disable_lbrv(vcpu);
2934 case MSR_VM_HSAVE_PA:
2936 * Old kernels did not validate the value written to
2937 * MSR_VM_HSAVE_PA. Allow KVM_SET_MSR to set an invalid
2938 * value to allow live migrating buggy or malicious guests
2939 * originating from those kernels.
2941 if (!msr->host_initiated && !page_address_valid(vcpu, data))
2944 svm->nested.hsave_msr = data & PAGE_MASK;
2947 return svm_set_vm_cr(vcpu, data);
2949 vcpu_unimpl(vcpu, "unimplemented wrmsr: 0x%x data 0x%llx\n", ecx, data);
2951 case MSR_F10H_DECFG: {
2952 struct kvm_msr_entry msr_entry;
2954 msr_entry.index = msr->index;
2955 if (svm_get_msr_feature(&msr_entry))
2958 /* Check the supported bits */
2959 if (data & ~msr_entry.data)
2962 /* Don't allow the guest to change a bit, #GP */
2963 if (!msr->host_initiated && (data ^ msr_entry.data))
2966 svm->msr_decfg = data;
2970 return kvm_set_msr_common(vcpu, msr);
2975 static int msr_interception(struct kvm_vcpu *vcpu)
2977 if (to_svm(vcpu)->vmcb->control.exit_info_1)
2978 return kvm_emulate_wrmsr(vcpu);
2980 return kvm_emulate_rdmsr(vcpu);
2983 static int interrupt_window_interception(struct kvm_vcpu *vcpu)
2985 kvm_make_request(KVM_REQ_EVENT, vcpu);
2986 svm_clear_vintr(to_svm(vcpu));
2989 * For AVIC, the only reason to end up here is ExtINTs.
2990 * In this case AVIC was temporarily disabled for
2991 * requesting the IRQ window and we have to re-enable it.
2993 kvm_request_apicv_update(vcpu->kvm, true, APICV_INHIBIT_REASON_IRQWIN);
2995 ++vcpu->stat.irq_window_exits;
2999 static int pause_interception(struct kvm_vcpu *vcpu)
3004 * CPL is not made available for an SEV-ES guest, therefore
3005 * vcpu->arch.preempted_in_kernel can never be true. Just
3006 * set in_kernel to false as well.
3008 in_kernel = !sev_es_guest(vcpu->kvm) && svm_get_cpl(vcpu) == 0;
3010 if (!kvm_pause_in_guest(vcpu->kvm))
3011 grow_ple_window(vcpu);
3013 kvm_vcpu_on_spin(vcpu, in_kernel);
3014 return kvm_skip_emulated_instruction(vcpu);
3017 static int invpcid_interception(struct kvm_vcpu *vcpu)
3019 struct vcpu_svm *svm = to_svm(vcpu);
3023 if (!guest_cpuid_has(vcpu, X86_FEATURE_INVPCID)) {
3024 kvm_queue_exception(vcpu, UD_VECTOR);
3029 * For an INVPCID intercept:
3030 * EXITINFO1 provides the linear address of the memory operand.
3031 * EXITINFO2 provides the contents of the register operand.
3033 type = svm->vmcb->control.exit_info_2;
3034 gva = svm->vmcb->control.exit_info_1;
3037 kvm_inject_gp(vcpu, 0);
3041 return kvm_handle_invpcid(vcpu, type, gva);
3044 static int (*const svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {
3045 [SVM_EXIT_READ_CR0] = cr_interception,
3046 [SVM_EXIT_READ_CR3] = cr_interception,
3047 [SVM_EXIT_READ_CR4] = cr_interception,
3048 [SVM_EXIT_READ_CR8] = cr_interception,
3049 [SVM_EXIT_CR0_SEL_WRITE] = cr_interception,
3050 [SVM_EXIT_WRITE_CR0] = cr_interception,
3051 [SVM_EXIT_WRITE_CR3] = cr_interception,
3052 [SVM_EXIT_WRITE_CR4] = cr_interception,
3053 [SVM_EXIT_WRITE_CR8] = cr8_write_interception,
3054 [SVM_EXIT_READ_DR0] = dr_interception,
3055 [SVM_EXIT_READ_DR1] = dr_interception,
3056 [SVM_EXIT_READ_DR2] = dr_interception,
3057 [SVM_EXIT_READ_DR3] = dr_interception,
3058 [SVM_EXIT_READ_DR4] = dr_interception,
3059 [SVM_EXIT_READ_DR5] = dr_interception,
3060 [SVM_EXIT_READ_DR6] = dr_interception,
3061 [SVM_EXIT_READ_DR7] = dr_interception,
3062 [SVM_EXIT_WRITE_DR0] = dr_interception,
3063 [SVM_EXIT_WRITE_DR1] = dr_interception,
3064 [SVM_EXIT_WRITE_DR2] = dr_interception,
3065 [SVM_EXIT_WRITE_DR3] = dr_interception,
3066 [SVM_EXIT_WRITE_DR4] = dr_interception,
3067 [SVM_EXIT_WRITE_DR5] = dr_interception,
3068 [SVM_EXIT_WRITE_DR6] = dr_interception,
3069 [SVM_EXIT_WRITE_DR7] = dr_interception,
3070 [SVM_EXIT_EXCP_BASE + DB_VECTOR] = db_interception,
3071 [SVM_EXIT_EXCP_BASE + BP_VECTOR] = bp_interception,
3072 [SVM_EXIT_EXCP_BASE + UD_VECTOR] = ud_interception,
3073 [SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
3074 [SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
3075 [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
3076 [SVM_EXIT_EXCP_BASE + GP_VECTOR] = gp_interception,
3077 [SVM_EXIT_INTR] = intr_interception,
3078 [SVM_EXIT_NMI] = nmi_interception,
3079 [SVM_EXIT_SMI] = smi_interception,
3080 [SVM_EXIT_VINTR] = interrupt_window_interception,
3081 [SVM_EXIT_RDPMC] = kvm_emulate_rdpmc,
3082 [SVM_EXIT_CPUID] = kvm_emulate_cpuid,
3083 [SVM_EXIT_IRET] = iret_interception,
3084 [SVM_EXIT_INVD] = kvm_emulate_invd,
3085 [SVM_EXIT_PAUSE] = pause_interception,
3086 [SVM_EXIT_HLT] = kvm_emulate_halt,
3087 [SVM_EXIT_INVLPG] = invlpg_interception,
3088 [SVM_EXIT_INVLPGA] = invlpga_interception,
3089 [SVM_EXIT_IOIO] = io_interception,
3090 [SVM_EXIT_MSR] = msr_interception,
3091 [SVM_EXIT_TASK_SWITCH] = task_switch_interception,
3092 [SVM_EXIT_SHUTDOWN] = shutdown_interception,
3093 [SVM_EXIT_VMRUN] = vmrun_interception,
3094 [SVM_EXIT_VMMCALL] = kvm_emulate_hypercall,
3095 [SVM_EXIT_VMLOAD] = vmload_interception,
3096 [SVM_EXIT_VMSAVE] = vmsave_interception,
3097 [SVM_EXIT_STGI] = stgi_interception,
3098 [SVM_EXIT_CLGI] = clgi_interception,
3099 [SVM_EXIT_SKINIT] = skinit_interception,
3100 [SVM_EXIT_RDTSCP] = kvm_handle_invalid_op,
3101 [SVM_EXIT_WBINVD] = kvm_emulate_wbinvd,
3102 [SVM_EXIT_MONITOR] = kvm_emulate_monitor,
3103 [SVM_EXIT_MWAIT] = kvm_emulate_mwait,
3104 [SVM_EXIT_XSETBV] = kvm_emulate_xsetbv,
3105 [SVM_EXIT_RDPRU] = kvm_handle_invalid_op,
3106 [SVM_EXIT_EFER_WRITE_TRAP] = efer_trap,
3107 [SVM_EXIT_CR0_WRITE_TRAP] = cr_trap,
3108 [SVM_EXIT_CR4_WRITE_TRAP] = cr_trap,
3109 [SVM_EXIT_CR8_WRITE_TRAP] = cr_trap,
3110 [SVM_EXIT_INVPCID] = invpcid_interception,
3111 [SVM_EXIT_NPF] = npf_interception,
3112 [SVM_EXIT_RSM] = rsm_interception,
3113 [SVM_EXIT_AVIC_INCOMPLETE_IPI] = avic_incomplete_ipi_interception,
3114 [SVM_EXIT_AVIC_UNACCELERATED_ACCESS] = avic_unaccelerated_access_interception,
3115 [SVM_EXIT_VMGEXIT] = sev_handle_vmgexit,
3118 static void dump_vmcb(struct kvm_vcpu *vcpu)
3120 struct vcpu_svm *svm = to_svm(vcpu);
3121 struct vmcb_control_area *control = &svm->vmcb->control;
3122 struct vmcb_save_area *save = &svm->vmcb->save;
3123 struct vmcb_save_area *save01 = &svm->vmcb01.ptr->save;
3125 if (!dump_invalid_vmcb) {
3126 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3130 pr_err("VMCB %p, last attempted VMRUN on CPU %d\n",
3131 svm->current_vmcb->ptr, vcpu->arch.last_vmentry_cpu);
3132 pr_err("VMCB Control Area:\n");
3133 pr_err("%-20s%04x\n", "cr_read:", control->intercepts[INTERCEPT_CR] & 0xffff);
3134 pr_err("%-20s%04x\n", "cr_write:", control->intercepts[INTERCEPT_CR] >> 16);
3135 pr_err("%-20s%04x\n", "dr_read:", control->intercepts[INTERCEPT_DR] & 0xffff);
3136 pr_err("%-20s%04x\n", "dr_write:", control->intercepts[INTERCEPT_DR] >> 16);
3137 pr_err("%-20s%08x\n", "exceptions:", control->intercepts[INTERCEPT_EXCEPTION]);
3138 pr_err("%-20s%08x %08x\n", "intercepts:",
3139 control->intercepts[INTERCEPT_WORD3],
3140 control->intercepts[INTERCEPT_WORD4]);
3141 pr_err("%-20s%d\n", "pause filter count:", control->pause_filter_count);
3142 pr_err("%-20s%d\n", "pause filter threshold:",
3143 control->pause_filter_thresh);
3144 pr_err("%-20s%016llx\n", "iopm_base_pa:", control->iopm_base_pa);
3145 pr_err("%-20s%016llx\n", "msrpm_base_pa:", control->msrpm_base_pa);
3146 pr_err("%-20s%016llx\n", "tsc_offset:", control->tsc_offset);
3147 pr_err("%-20s%d\n", "asid:", control->asid);
3148 pr_err("%-20s%d\n", "tlb_ctl:", control->tlb_ctl);
3149 pr_err("%-20s%08x\n", "int_ctl:", control->int_ctl);
3150 pr_err("%-20s%08x\n", "int_vector:", control->int_vector);
3151 pr_err("%-20s%08x\n", "int_state:", control->int_state);
3152 pr_err("%-20s%08x\n", "exit_code:", control->exit_code);
3153 pr_err("%-20s%016llx\n", "exit_info1:", control->exit_info_1);
3154 pr_err("%-20s%016llx\n", "exit_info2:", control->exit_info_2);
3155 pr_err("%-20s%08x\n", "exit_int_info:", control->exit_int_info);
3156 pr_err("%-20s%08x\n", "exit_int_info_err:", control->exit_int_info_err);
3157 pr_err("%-20s%lld\n", "nested_ctl:", control->nested_ctl);
3158 pr_err("%-20s%016llx\n", "nested_cr3:", control->nested_cr3);
3159 pr_err("%-20s%016llx\n", "avic_vapic_bar:", control->avic_vapic_bar);
3160 pr_err("%-20s%016llx\n", "ghcb:", control->ghcb_gpa);
3161 pr_err("%-20s%08x\n", "event_inj:", control->event_inj);
3162 pr_err("%-20s%08x\n", "event_inj_err:", control->event_inj_err);
3163 pr_err("%-20s%lld\n", "virt_ext:", control->virt_ext);
3164 pr_err("%-20s%016llx\n", "next_rip:", control->next_rip);
3165 pr_err("%-20s%016llx\n", "avic_backing_page:", control->avic_backing_page);
3166 pr_err("%-20s%016llx\n", "avic_logical_id:", control->avic_logical_id);
3167 pr_err("%-20s%016llx\n", "avic_physical_id:", control->avic_physical_id);
3168 pr_err("%-20s%016llx\n", "vmsa_pa:", control->vmsa_pa);
3169 pr_err("VMCB State Save Area:\n");
3170 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3172 save->es.selector, save->es.attrib,
3173 save->es.limit, save->es.base);
3174 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3176 save->cs.selector, save->cs.attrib,
3177 save->cs.limit, save->cs.base);
3178 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3180 save->ss.selector, save->ss.attrib,
3181 save->ss.limit, save->ss.base);
3182 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3184 save->ds.selector, save->ds.attrib,
3185 save->ds.limit, save->ds.base);
3186 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3188 save01->fs.selector, save01->fs.attrib,
3189 save01->fs.limit, save01->fs.base);
3190 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3192 save01->gs.selector, save01->gs.attrib,
3193 save01->gs.limit, save01->gs.base);
3194 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3196 save->gdtr.selector, save->gdtr.attrib,
3197 save->gdtr.limit, save->gdtr.base);
3198 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3200 save01->ldtr.selector, save01->ldtr.attrib,
3201 save01->ldtr.limit, save01->ldtr.base);
3202 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3204 save->idtr.selector, save->idtr.attrib,
3205 save->idtr.limit, save->idtr.base);
3206 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3208 save01->tr.selector, save01->tr.attrib,
3209 save01->tr.limit, save01->tr.base);
3210 pr_err("cpl: %d efer: %016llx\n",
3211 save->cpl, save->efer);
3212 pr_err("%-15s %016llx %-13s %016llx\n",
3213 "cr0:", save->cr0, "cr2:", save->cr2);
3214 pr_err("%-15s %016llx %-13s %016llx\n",
3215 "cr3:", save->cr3, "cr4:", save->cr4);
3216 pr_err("%-15s %016llx %-13s %016llx\n",
3217 "dr6:", save->dr6, "dr7:", save->dr7);
3218 pr_err("%-15s %016llx %-13s %016llx\n",
3219 "rip:", save->rip, "rflags:", save->rflags);
3220 pr_err("%-15s %016llx %-13s %016llx\n",
3221 "rsp:", save->rsp, "rax:", save->rax);
3222 pr_err("%-15s %016llx %-13s %016llx\n",
3223 "star:", save01->star, "lstar:", save01->lstar);
3224 pr_err("%-15s %016llx %-13s %016llx\n",
3225 "cstar:", save01->cstar, "sfmask:", save01->sfmask);
3226 pr_err("%-15s %016llx %-13s %016llx\n",
3227 "kernel_gs_base:", save01->kernel_gs_base,
3228 "sysenter_cs:", save01->sysenter_cs);
3229 pr_err("%-15s %016llx %-13s %016llx\n",
3230 "sysenter_esp:", save01->sysenter_esp,
3231 "sysenter_eip:", save01->sysenter_eip);
3232 pr_err("%-15s %016llx %-13s %016llx\n",
3233 "gpat:", save->g_pat, "dbgctl:", save->dbgctl);
3234 pr_err("%-15s %016llx %-13s %016llx\n",
3235 "br_from:", save->br_from, "br_to:", save->br_to);
3236 pr_err("%-15s %016llx %-13s %016llx\n",
3237 "excp_from:", save->last_excp_from,
3238 "excp_to:", save->last_excp_to);
3241 static bool svm_check_exit_valid(struct kvm_vcpu *vcpu, u64 exit_code)
3243 return (exit_code < ARRAY_SIZE(svm_exit_handlers) &&
3244 svm_exit_handlers[exit_code]);
3247 static int svm_handle_invalid_exit(struct kvm_vcpu *vcpu, u64 exit_code)
3249 vcpu_unimpl(vcpu, "svm: unexpected exit reason 0x%llx\n", exit_code);
3251 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3252 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
3253 vcpu->run->internal.ndata = 2;
3254 vcpu->run->internal.data[0] = exit_code;
3255 vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
3259 int svm_invoke_exit_handler(struct kvm_vcpu *vcpu, u64 exit_code)
3261 if (!svm_check_exit_valid(vcpu, exit_code))
3262 return svm_handle_invalid_exit(vcpu, exit_code);
3264 #ifdef CONFIG_RETPOLINE
3265 if (exit_code == SVM_EXIT_MSR)
3266 return msr_interception(vcpu);
3267 else if (exit_code == SVM_EXIT_VINTR)
3268 return interrupt_window_interception(vcpu);
3269 else if (exit_code == SVM_EXIT_INTR)
3270 return intr_interception(vcpu);
3271 else if (exit_code == SVM_EXIT_HLT)
3272 return kvm_emulate_halt(vcpu);
3273 else if (exit_code == SVM_EXIT_NPF)
3274 return npf_interception(vcpu);
3276 return svm_exit_handlers[exit_code](vcpu);
3279 static void svm_get_exit_info(struct kvm_vcpu *vcpu, u64 *info1, u64 *info2,
3280 u32 *intr_info, u32 *error_code)
3282 struct vmcb_control_area *control = &to_svm(vcpu)->vmcb->control;
3284 *info1 = control->exit_info_1;
3285 *info2 = control->exit_info_2;
3286 *intr_info = control->exit_int_info;
3287 if ((*intr_info & SVM_EXITINTINFO_VALID) &&
3288 (*intr_info & SVM_EXITINTINFO_VALID_ERR))
3289 *error_code = control->exit_int_info_err;
3294 static int handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
3296 struct vcpu_svm *svm = to_svm(vcpu);
3297 struct kvm_run *kvm_run = vcpu->run;
3298 u32 exit_code = svm->vmcb->control.exit_code;
3300 trace_kvm_exit(exit_code, vcpu, KVM_ISA_SVM);
3302 /* SEV-ES guests must use the CR write traps to track CR registers. */
3303 if (!sev_es_guest(vcpu->kvm)) {
3304 if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
3305 vcpu->arch.cr0 = svm->vmcb->save.cr0;
3307 vcpu->arch.cr3 = svm->vmcb->save.cr3;
3310 if (is_guest_mode(vcpu)) {
3313 trace_kvm_nested_vmexit(exit_code, vcpu, KVM_ISA_SVM);
3315 vmexit = nested_svm_exit_special(svm);
3317 if (vmexit == NESTED_EXIT_CONTINUE)
3318 vmexit = nested_svm_exit_handled(svm);
3320 if (vmexit == NESTED_EXIT_DONE)
3324 if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
3325 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3326 kvm_run->fail_entry.hardware_entry_failure_reason
3327 = svm->vmcb->control.exit_code;
3328 kvm_run->fail_entry.cpu = vcpu->arch.last_vmentry_cpu;
3333 if (is_external_interrupt(svm->vmcb->control.exit_int_info) &&
3334 exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR &&
3335 exit_code != SVM_EXIT_NPF && exit_code != SVM_EXIT_TASK_SWITCH &&
3336 exit_code != SVM_EXIT_INTR && exit_code != SVM_EXIT_NMI)
3337 printk(KERN_ERR "%s: unexpected exit_int_info 0x%x "
3339 __func__, svm->vmcb->control.exit_int_info,
3342 if (exit_fastpath != EXIT_FASTPATH_NONE)
3345 return svm_invoke_exit_handler(vcpu, exit_code);
3348 static void reload_tss(struct kvm_vcpu *vcpu)
3350 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3352 sd->tss_desc->type = 9; /* available 32/64-bit TSS */
3356 static void pre_svm_run(struct kvm_vcpu *vcpu)
3358 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3359 struct vcpu_svm *svm = to_svm(vcpu);
3362 * If the previous vmrun of the vmcb occurred on a different physical
3363 * cpu, then mark the vmcb dirty and assign a new asid. Hardware's
3364 * vmcb clean bits are per logical CPU, as are KVM's asid assignments.
3366 if (unlikely(svm->current_vmcb->cpu != vcpu->cpu)) {
3367 svm->current_vmcb->asid_generation = 0;
3368 vmcb_mark_all_dirty(svm->vmcb);
3369 svm->current_vmcb->cpu = vcpu->cpu;
3372 if (sev_guest(vcpu->kvm))
3373 return pre_sev_run(svm, vcpu->cpu);
3375 /* FIXME: handle wraparound of asid_generation */
3376 if (svm->current_vmcb->asid_generation != sd->asid_generation)
3380 static void svm_inject_nmi(struct kvm_vcpu *vcpu)
3382 struct vcpu_svm *svm = to_svm(vcpu);
3384 svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
3385 vcpu->arch.hflags |= HF_NMI_MASK;
3386 if (!sev_es_guest(vcpu->kvm))
3387 svm_set_intercept(svm, INTERCEPT_IRET);
3388 ++vcpu->stat.nmi_injections;
3391 static void svm_set_irq(struct kvm_vcpu *vcpu)
3393 struct vcpu_svm *svm = to_svm(vcpu);
3395 BUG_ON(!(gif_set(svm)));
3397 trace_kvm_inj_virq(vcpu->arch.interrupt.nr);
3398 ++vcpu->stat.irq_injections;
3400 svm->vmcb->control.event_inj = vcpu->arch.interrupt.nr |
3401 SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR;
3404 static void svm_update_cr8_intercept(struct kvm_vcpu *vcpu, int tpr, int irr)
3406 struct vcpu_svm *svm = to_svm(vcpu);
3409 * SEV-ES guests must always keep the CR intercepts cleared. CR
3410 * tracking is done using the CR write traps.
3412 if (sev_es_guest(vcpu->kvm))
3415 if (nested_svm_virtualize_tpr(vcpu))
3418 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3424 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
3427 bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
3429 struct vcpu_svm *svm = to_svm(vcpu);
3430 struct vmcb *vmcb = svm->vmcb;
3436 if (is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3439 ret = (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK) ||
3440 (vcpu->arch.hflags & HF_NMI_MASK);
3445 static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3447 struct vcpu_svm *svm = to_svm(vcpu);
3448 if (svm->nested.nested_run_pending)
3451 /* An NMI must not be injected into L2 if it's supposed to VM-Exit. */
3452 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3455 return !svm_nmi_blocked(vcpu);
3458 static bool svm_get_nmi_mask(struct kvm_vcpu *vcpu)
3460 return !!(vcpu->arch.hflags & HF_NMI_MASK);
3463 static void svm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
3465 struct vcpu_svm *svm = to_svm(vcpu);
3468 vcpu->arch.hflags |= HF_NMI_MASK;
3469 if (!sev_es_guest(vcpu->kvm))
3470 svm_set_intercept(svm, INTERCEPT_IRET);
3472 vcpu->arch.hflags &= ~HF_NMI_MASK;
3473 if (!sev_es_guest(vcpu->kvm))
3474 svm_clr_intercept(svm, INTERCEPT_IRET);
3478 bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
3480 struct vcpu_svm *svm = to_svm(vcpu);
3481 struct vmcb *vmcb = svm->vmcb;
3486 if (sev_es_guest(vcpu->kvm)) {
3488 * SEV-ES guests to not expose RFLAGS. Use the VMCB interrupt mask
3489 * bit to determine the state of the IF flag.
3491 if (!(vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK))
3493 } else if (is_guest_mode(vcpu)) {
3494 /* As long as interrupts are being delivered... */
3495 if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
3496 ? !(svm->vmcb01.ptr->save.rflags & X86_EFLAGS_IF)
3497 : !(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3500 /* ... vmexits aren't blocked by the interrupt shadow */
3501 if (nested_exit_on_intr(svm))
3504 if (!(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3508 return (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK);
3511 static int svm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3513 struct vcpu_svm *svm = to_svm(vcpu);
3514 if (svm->nested.nested_run_pending)
3518 * An IRQ must not be injected into L2 if it's supposed to VM-Exit,
3519 * e.g. if the IRQ arrived asynchronously after checking nested events.
3521 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_intr(svm))
3524 return !svm_interrupt_blocked(vcpu);
3527 static void svm_enable_irq_window(struct kvm_vcpu *vcpu)
3529 struct vcpu_svm *svm = to_svm(vcpu);
3532 * In case GIF=0 we can't rely on the CPU to tell us when GIF becomes
3533 * 1, because that's a separate STGI/VMRUN intercept. The next time we
3534 * get that intercept, this function will be called again though and
3535 * we'll get the vintr intercept. However, if the vGIF feature is
3536 * enabled, the STGI interception will not occur. Enable the irq
3537 * window under the assumption that the hardware will set the GIF.
3539 if (vgif_enabled(svm) || gif_set(svm)) {
3541 * IRQ window is not needed when AVIC is enabled,
3542 * unless we have pending ExtINT since it cannot be injected
3543 * via AVIC. In such case, we need to temporarily disable AVIC,
3544 * and fallback to injecting IRQ via V_IRQ.
3546 kvm_request_apicv_update(vcpu->kvm, false, APICV_INHIBIT_REASON_IRQWIN);
3551 static void svm_enable_nmi_window(struct kvm_vcpu *vcpu)
3553 struct vcpu_svm *svm = to_svm(vcpu);
3555 if ((vcpu->arch.hflags & (HF_NMI_MASK | HF_IRET_MASK)) == HF_NMI_MASK)
3556 return; /* IRET will cause a vm exit */
3558 if (!gif_set(svm)) {
3559 if (vgif_enabled(svm))
3560 svm_set_intercept(svm, INTERCEPT_STGI);
3561 return; /* STGI will cause a vm exit */
3565 * Something prevents NMI from been injected. Single step over possible
3566 * problem (IRET or exception injection or interrupt shadow)
3568 svm->nmi_singlestep_guest_rflags = svm_get_rflags(vcpu);
3569 svm->nmi_singlestep = true;
3570 svm->vmcb->save.rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
3573 static int svm_set_tss_addr(struct kvm *kvm, unsigned int addr)
3578 static int svm_set_identity_map_addr(struct kvm *kvm, u64 ident_addr)
3583 void svm_flush_tlb(struct kvm_vcpu *vcpu)
3585 struct vcpu_svm *svm = to_svm(vcpu);
3588 * Flush only the current ASID even if the TLB flush was invoked via
3589 * kvm_flush_remote_tlbs(). Although flushing remote TLBs requires all
3590 * ASIDs to be flushed, KVM uses a single ASID for L1 and L2, and
3591 * unconditionally does a TLB flush on both nested VM-Enter and nested
3592 * VM-Exit (via kvm_mmu_reset_context()).
3594 if (static_cpu_has(X86_FEATURE_FLUSHBYASID))
3595 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3597 svm->current_vmcb->asid_generation--;
3600 static void svm_flush_tlb_gva(struct kvm_vcpu *vcpu, gva_t gva)
3602 struct vcpu_svm *svm = to_svm(vcpu);
3604 invlpga(gva, svm->vmcb->control.asid);
3607 static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)
3609 struct vcpu_svm *svm = to_svm(vcpu);
3611 if (nested_svm_virtualize_tpr(vcpu))
3614 if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
3615 int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
3616 kvm_set_cr8(vcpu, cr8);
3620 static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
3622 struct vcpu_svm *svm = to_svm(vcpu);
3625 if (nested_svm_virtualize_tpr(vcpu) ||
3626 kvm_vcpu_apicv_active(vcpu))
3629 cr8 = kvm_get_cr8(vcpu);
3630 svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
3631 svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
3634 static void svm_complete_interrupts(struct kvm_vcpu *vcpu)
3636 struct vcpu_svm *svm = to_svm(vcpu);
3639 u32 exitintinfo = svm->vmcb->control.exit_int_info;
3640 unsigned int3_injected = svm->int3_injected;
3642 svm->int3_injected = 0;
3645 * If we've made progress since setting HF_IRET_MASK, we've
3646 * executed an IRET and can allow NMI injection.
3648 if ((vcpu->arch.hflags & HF_IRET_MASK) &&
3649 (sev_es_guest(vcpu->kvm) ||
3650 kvm_rip_read(vcpu) != svm->nmi_iret_rip)) {
3651 vcpu->arch.hflags &= ~(HF_NMI_MASK | HF_IRET_MASK);
3652 kvm_make_request(KVM_REQ_EVENT, vcpu);
3655 vcpu->arch.nmi_injected = false;
3656 kvm_clear_exception_queue(vcpu);
3657 kvm_clear_interrupt_queue(vcpu);
3659 if (!(exitintinfo & SVM_EXITINTINFO_VALID))
3662 kvm_make_request(KVM_REQ_EVENT, vcpu);
3664 vector = exitintinfo & SVM_EXITINTINFO_VEC_MASK;
3665 type = exitintinfo & SVM_EXITINTINFO_TYPE_MASK;
3668 case SVM_EXITINTINFO_TYPE_NMI:
3669 vcpu->arch.nmi_injected = true;
3671 case SVM_EXITINTINFO_TYPE_EXEPT:
3673 * Never re-inject a #VC exception.
3675 if (vector == X86_TRAP_VC)
3679 * In case of software exceptions, do not reinject the vector,
3680 * but re-execute the instruction instead. Rewind RIP first
3681 * if we emulated INT3 before.
3683 if (kvm_exception_is_soft(vector)) {
3684 if (vector == BP_VECTOR && int3_injected &&
3685 kvm_is_linear_rip(vcpu, svm->int3_rip))
3687 kvm_rip_read(vcpu) - int3_injected);
3690 if (exitintinfo & SVM_EXITINTINFO_VALID_ERR) {
3691 u32 err = svm->vmcb->control.exit_int_info_err;
3692 kvm_requeue_exception_e(vcpu, vector, err);
3695 kvm_requeue_exception(vcpu, vector);
3697 case SVM_EXITINTINFO_TYPE_INTR:
3698 kvm_queue_interrupt(vcpu, vector, false);
3705 static void svm_cancel_injection(struct kvm_vcpu *vcpu)
3707 struct vcpu_svm *svm = to_svm(vcpu);
3708 struct vmcb_control_area *control = &svm->vmcb->control;
3710 control->exit_int_info = control->event_inj;
3711 control->exit_int_info_err = control->event_inj_err;
3712 control->event_inj = 0;
3713 svm_complete_interrupts(vcpu);
3716 static fastpath_t svm_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
3718 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_MSR &&
3719 to_svm(vcpu)->vmcb->control.exit_info_1)
3720 return handle_fastpath_set_msr_irqoff(vcpu);
3722 return EXIT_FASTPATH_NONE;
3725 static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
3727 struct vcpu_svm *svm = to_svm(vcpu);
3728 unsigned long vmcb_pa = svm->current_vmcb->pa;
3730 kvm_guest_enter_irqoff();
3732 if (sev_es_guest(vcpu->kvm)) {
3733 __svm_sev_es_vcpu_run(vmcb_pa);
3735 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3738 * Use a single vmcb (vmcb01 because it's always valid) for
3739 * context switching guest state via VMLOAD/VMSAVE, that way
3740 * the state doesn't need to be copied between vmcb01 and
3741 * vmcb02 when switching vmcbs for nested virtualization.
3743 vmload(svm->vmcb01.pa);
3744 __svm_vcpu_run(vmcb_pa, (unsigned long *)&vcpu->arch.regs);
3745 vmsave(svm->vmcb01.pa);
3747 vmload(__sme_page_pa(sd->save_area));
3750 kvm_guest_exit_irqoff();
3753 static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
3755 struct vcpu_svm *svm = to_svm(vcpu);
3757 trace_kvm_entry(vcpu);
3759 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
3760 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
3761 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
3764 * Disable singlestep if we're injecting an interrupt/exception.
3765 * We don't want our modified rflags to be pushed on the stack where
3766 * we might not be able to easily reset them if we disabled NMI
3769 if (svm->nmi_singlestep && svm->vmcb->control.event_inj) {
3771 * Event injection happens before external interrupts cause a
3772 * vmexit and interrupts are disabled here, so smp_send_reschedule
3773 * is enough to force an immediate vmexit.
3775 disable_nmi_singlestep(svm);
3776 smp_send_reschedule(vcpu->cpu);
3781 WARN_ON_ONCE(kvm_apicv_activated(vcpu->kvm) != kvm_vcpu_apicv_active(vcpu));
3783 sync_lapic_to_cr8(vcpu);
3785 if (unlikely(svm->asid != svm->vmcb->control.asid)) {
3786 svm->vmcb->control.asid = svm->asid;
3787 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
3789 svm->vmcb->save.cr2 = vcpu->arch.cr2;
3791 svm_hv_update_vp_id(svm->vmcb, vcpu);
3794 * Run with all-zero DR6 unless needed, so that we can get the exact cause
3797 if (unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT))
3798 svm_set_dr6(svm, vcpu->arch.dr6);
3800 svm_set_dr6(svm, DR6_ACTIVE_LOW);
3803 kvm_load_guest_xsave_state(vcpu);
3805 kvm_wait_lapic_expire(vcpu);
3808 * If this vCPU has touched SPEC_CTRL, restore the guest's value if
3809 * it's non-zero. Since vmentry is serialising on affected CPUs, there
3810 * is no need to worry about the conditional branch over the wrmsr
3811 * being speculatively taken.
3813 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3814 x86_spec_ctrl_set_guest(svm->spec_ctrl, svm->virt_spec_ctrl);
3816 svm_vcpu_enter_exit(vcpu);
3819 * We do not use IBRS in the kernel. If this vCPU has used the
3820 * SPEC_CTRL MSR it may have left it on; save the value and
3821 * turn it off. This is much more efficient than blindly adding
3822 * it to the atomic save/restore list. Especially as the former
3823 * (Saving guest MSRs on vmexit) doesn't even exist in KVM.
3825 * For non-nested case:
3826 * If the L01 MSR bitmap does not intercept the MSR, then we need to
3830 * If the L02 MSR bitmap does not intercept the MSR, then we need to
3833 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL) &&
3834 unlikely(!msr_write_intercepted(vcpu, MSR_IA32_SPEC_CTRL)))
3835 svm->spec_ctrl = native_read_msr(MSR_IA32_SPEC_CTRL);
3837 if (!sev_es_guest(vcpu->kvm))
3840 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3841 x86_spec_ctrl_restore_host(svm->spec_ctrl, svm->virt_spec_ctrl);
3843 if (!sev_es_guest(vcpu->kvm)) {
3844 vcpu->arch.cr2 = svm->vmcb->save.cr2;
3845 vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
3846 vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
3847 vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
3850 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3851 kvm_before_interrupt(vcpu);
3853 kvm_load_host_xsave_state(vcpu);
3856 /* Any pending NMI will happen here */
3858 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3859 kvm_after_interrupt(vcpu);
3861 sync_cr8_to_lapic(vcpu);
3864 if (is_guest_mode(vcpu)) {
3865 nested_sync_control_from_vmcb02(svm);
3867 /* Track VMRUNs that have made past consistency checking */
3868 if (svm->nested.nested_run_pending &&
3869 svm->vmcb->control.exit_code != SVM_EXIT_ERR)
3870 ++vcpu->stat.nested_run;
3872 svm->nested.nested_run_pending = 0;
3875 svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
3876 vmcb_mark_all_clean(svm->vmcb);
3878 /* if exit due to PF check for async PF */
3879 if (svm->vmcb->control.exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR)
3880 vcpu->arch.apf.host_apf_flags =
3881 kvm_read_and_reset_apf_flags();
3884 kvm_register_clear_available(vcpu, VCPU_EXREG_PDPTR);
3887 * We need to handle MC intercepts here before the vcpu has a chance to
3888 * change the physical cpu
3890 if (unlikely(svm->vmcb->control.exit_code ==
3891 SVM_EXIT_EXCP_BASE + MC_VECTOR))
3892 svm_handle_mce(vcpu);
3894 svm_complete_interrupts(vcpu);
3896 if (is_guest_mode(vcpu))
3897 return EXIT_FASTPATH_NONE;
3899 return svm_exit_handlers_fastpath(vcpu);
3902 static void svm_load_mmu_pgd(struct kvm_vcpu *vcpu, hpa_t root_hpa,
3905 struct vcpu_svm *svm = to_svm(vcpu);
3909 svm->vmcb->control.nested_cr3 = __sme_set(root_hpa);
3910 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
3912 hv_track_root_tdp(vcpu, root_hpa);
3914 /* Loading L2's CR3 is handled by enter_svm_guest_mode. */
3915 if (!test_bit(VCPU_EXREG_CR3, (ulong *)&vcpu->arch.regs_avail))
3917 cr3 = vcpu->arch.cr3;
3918 } else if (vcpu->arch.mmu->shadow_root_level >= PT64_ROOT_4LEVEL) {
3919 cr3 = __sme_set(root_hpa) | kvm_get_active_pcid(vcpu);
3921 /* PCID in the guest should be impossible with a 32-bit MMU. */
3922 WARN_ON_ONCE(kvm_get_active_pcid(vcpu));
3926 svm->vmcb->save.cr3 = cr3;
3927 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
3930 static int is_disabled(void)
3934 rdmsrl(MSR_VM_CR, vm_cr);
3935 if (vm_cr & (1 << SVM_VM_CR_SVM_DISABLE))
3942 svm_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
3945 * Patch in the VMMCALL instruction:
3947 hypercall[0] = 0x0f;
3948 hypercall[1] = 0x01;
3949 hypercall[2] = 0xd9;
3952 static int __init svm_check_processor_compat(void)
3957 static bool svm_cpu_has_accelerated_tpr(void)
3963 * The kvm parameter can be NULL (module initialization, or invocation before
3964 * VM creation). Be sure to check the kvm parameter before using it.
3966 static bool svm_has_emulated_msr(struct kvm *kvm, u32 index)
3969 case MSR_IA32_MCG_EXT_CTL:
3970 case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:
3972 case MSR_IA32_SMBASE:
3973 /* SEV-ES guests do not support SMM, so report false */
3974 if (kvm && sev_es_guest(kvm))
3984 static u64 svm_get_mt_mask(struct kvm_vcpu *vcpu, gfn_t gfn, bool is_mmio)
3989 static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
3991 struct vcpu_svm *svm = to_svm(vcpu);
3992 struct kvm_cpuid_entry2 *best;
3994 vcpu->arch.xsaves_enabled = guest_cpuid_has(vcpu, X86_FEATURE_XSAVE) &&
3995 boot_cpu_has(X86_FEATURE_XSAVE) &&
3996 boot_cpu_has(X86_FEATURE_XSAVES);
3998 /* Update nrips enabled cache */
3999 svm->nrips_enabled = kvm_cpu_cap_has(X86_FEATURE_NRIPS) &&
4000 guest_cpuid_has(vcpu, X86_FEATURE_NRIPS);
4002 svm_recalc_instruction_intercepts(vcpu, svm);
4004 /* For sev guests, the memory encryption bit is not reserved in CR3. */
4005 if (sev_guest(vcpu->kvm)) {
4006 best = kvm_find_cpuid_entry(vcpu, 0x8000001F, 0);
4008 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
4011 if (kvm_vcpu_apicv_active(vcpu)) {
4013 * AVIC does not work with an x2APIC mode guest. If the X2APIC feature
4014 * is exposed to the guest, disable AVIC.
4016 if (guest_cpuid_has(vcpu, X86_FEATURE_X2APIC))
4017 kvm_request_apicv_update(vcpu->kvm, false,
4018 APICV_INHIBIT_REASON_X2APIC);
4021 * Currently, AVIC does not work with nested virtualization.
4022 * So, we disable AVIC when cpuid for SVM is set in the L1 guest.
4024 if (nested && guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4025 kvm_request_apicv_update(vcpu->kvm, false,
4026 APICV_INHIBIT_REASON_NESTED);
4029 if (guest_cpuid_is_intel(vcpu)) {
4031 * We must intercept SYSENTER_EIP and SYSENTER_ESP
4032 * accesses because the processor only stores 32 bits.
4033 * For the same reason we cannot use virtual VMLOAD/VMSAVE.
4035 svm_set_intercept(svm, INTERCEPT_VMLOAD);
4036 svm_set_intercept(svm, INTERCEPT_VMSAVE);
4037 svm->vmcb->control.virt_ext &= ~VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
4039 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 0, 0);
4040 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 0, 0);
4043 * If hardware supports Virtual VMLOAD VMSAVE then enable it
4044 * in VMCB and clear intercepts to avoid #VMEXIT.
4047 svm_clr_intercept(svm, INTERCEPT_VMLOAD);
4048 svm_clr_intercept(svm, INTERCEPT_VMSAVE);
4049 svm->vmcb->control.virt_ext |= VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
4051 /* No need to intercept these MSRs */
4052 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 1, 1);
4053 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 1, 1);
4057 static bool svm_has_wbinvd_exit(void)
4062 #define PRE_EX(exit) { .exit_code = (exit), \
4063 .stage = X86_ICPT_PRE_EXCEPT, }
4064 #define POST_EX(exit) { .exit_code = (exit), \
4065 .stage = X86_ICPT_POST_EXCEPT, }
4066 #define POST_MEM(exit) { .exit_code = (exit), \
4067 .stage = X86_ICPT_POST_MEMACCESS, }
4069 static const struct __x86_intercept {
4071 enum x86_intercept_stage stage;
4072 } x86_intercept_map[] = {
4073 [x86_intercept_cr_read] = POST_EX(SVM_EXIT_READ_CR0),
4074 [x86_intercept_cr_write] = POST_EX(SVM_EXIT_WRITE_CR0),
4075 [x86_intercept_clts] = POST_EX(SVM_EXIT_WRITE_CR0),
4076 [x86_intercept_lmsw] = POST_EX(SVM_EXIT_WRITE_CR0),
4077 [x86_intercept_smsw] = POST_EX(SVM_EXIT_READ_CR0),
4078 [x86_intercept_dr_read] = POST_EX(SVM_EXIT_READ_DR0),
4079 [x86_intercept_dr_write] = POST_EX(SVM_EXIT_WRITE_DR0),
4080 [x86_intercept_sldt] = POST_EX(SVM_EXIT_LDTR_READ),
4081 [x86_intercept_str] = POST_EX(SVM_EXIT_TR_READ),
4082 [x86_intercept_lldt] = POST_EX(SVM_EXIT_LDTR_WRITE),
4083 [x86_intercept_ltr] = POST_EX(SVM_EXIT_TR_WRITE),
4084 [x86_intercept_sgdt] = POST_EX(SVM_EXIT_GDTR_READ),
4085 [x86_intercept_sidt] = POST_EX(SVM_EXIT_IDTR_READ),
4086 [x86_intercept_lgdt] = POST_EX(SVM_EXIT_GDTR_WRITE),
4087 [x86_intercept_lidt] = POST_EX(SVM_EXIT_IDTR_WRITE),
4088 [x86_intercept_vmrun] = POST_EX(SVM_EXIT_VMRUN),
4089 [x86_intercept_vmmcall] = POST_EX(SVM_EXIT_VMMCALL),
4090 [x86_intercept_vmload] = POST_EX(SVM_EXIT_VMLOAD),
4091 [x86_intercept_vmsave] = POST_EX(SVM_EXIT_VMSAVE),
4092 [x86_intercept_stgi] = POST_EX(SVM_EXIT_STGI),
4093 [x86_intercept_clgi] = POST_EX(SVM_EXIT_CLGI),
4094 [x86_intercept_skinit] = POST_EX(SVM_EXIT_SKINIT),
4095 [x86_intercept_invlpga] = POST_EX(SVM_EXIT_INVLPGA),
4096 [x86_intercept_rdtscp] = POST_EX(SVM_EXIT_RDTSCP),
4097 [x86_intercept_monitor] = POST_MEM(SVM_EXIT_MONITOR),
4098 [x86_intercept_mwait] = POST_EX(SVM_EXIT_MWAIT),
4099 [x86_intercept_invlpg] = POST_EX(SVM_EXIT_INVLPG),
4100 [x86_intercept_invd] = POST_EX(SVM_EXIT_INVD),
4101 [x86_intercept_wbinvd] = POST_EX(SVM_EXIT_WBINVD),
4102 [x86_intercept_wrmsr] = POST_EX(SVM_EXIT_MSR),
4103 [x86_intercept_rdtsc] = POST_EX(SVM_EXIT_RDTSC),
4104 [x86_intercept_rdmsr] = POST_EX(SVM_EXIT_MSR),
4105 [x86_intercept_rdpmc] = POST_EX(SVM_EXIT_RDPMC),
4106 [x86_intercept_cpuid] = PRE_EX(SVM_EXIT_CPUID),
4107 [x86_intercept_rsm] = PRE_EX(SVM_EXIT_RSM),
4108 [x86_intercept_pause] = PRE_EX(SVM_EXIT_PAUSE),
4109 [x86_intercept_pushf] = PRE_EX(SVM_EXIT_PUSHF),
4110 [x86_intercept_popf] = PRE_EX(SVM_EXIT_POPF),
4111 [x86_intercept_intn] = PRE_EX(SVM_EXIT_SWINT),
4112 [x86_intercept_iret] = PRE_EX(SVM_EXIT_IRET),
4113 [x86_intercept_icebp] = PRE_EX(SVM_EXIT_ICEBP),
4114 [x86_intercept_hlt] = POST_EX(SVM_EXIT_HLT),
4115 [x86_intercept_in] = POST_EX(SVM_EXIT_IOIO),
4116 [x86_intercept_ins] = POST_EX(SVM_EXIT_IOIO),
4117 [x86_intercept_out] = POST_EX(SVM_EXIT_IOIO),
4118 [x86_intercept_outs] = POST_EX(SVM_EXIT_IOIO),
4119 [x86_intercept_xsetbv] = PRE_EX(SVM_EXIT_XSETBV),
4126 static int svm_check_intercept(struct kvm_vcpu *vcpu,
4127 struct x86_instruction_info *info,
4128 enum x86_intercept_stage stage,
4129 struct x86_exception *exception)
4131 struct vcpu_svm *svm = to_svm(vcpu);
4132 int vmexit, ret = X86EMUL_CONTINUE;
4133 struct __x86_intercept icpt_info;
4134 struct vmcb *vmcb = svm->vmcb;
4136 if (info->intercept >= ARRAY_SIZE(x86_intercept_map))
4139 icpt_info = x86_intercept_map[info->intercept];
4141 if (stage != icpt_info.stage)
4144 switch (icpt_info.exit_code) {
4145 case SVM_EXIT_READ_CR0:
4146 if (info->intercept == x86_intercept_cr_read)
4147 icpt_info.exit_code += info->modrm_reg;
4149 case SVM_EXIT_WRITE_CR0: {
4150 unsigned long cr0, val;
4152 if (info->intercept == x86_intercept_cr_write)
4153 icpt_info.exit_code += info->modrm_reg;
4155 if (icpt_info.exit_code != SVM_EXIT_WRITE_CR0 ||
4156 info->intercept == x86_intercept_clts)
4159 if (!(vmcb_is_intercept(&svm->nested.ctl,
4160 INTERCEPT_SELECTIVE_CR0)))
4163 cr0 = vcpu->arch.cr0 & ~SVM_CR0_SELECTIVE_MASK;
4164 val = info->src_val & ~SVM_CR0_SELECTIVE_MASK;
4166 if (info->intercept == x86_intercept_lmsw) {
4169 /* lmsw can't clear PE - catch this here */
4170 if (cr0 & X86_CR0_PE)
4175 icpt_info.exit_code = SVM_EXIT_CR0_SEL_WRITE;
4179 case SVM_EXIT_READ_DR0:
4180 case SVM_EXIT_WRITE_DR0:
4181 icpt_info.exit_code += info->modrm_reg;
4184 if (info->intercept == x86_intercept_wrmsr)
4185 vmcb->control.exit_info_1 = 1;
4187 vmcb->control.exit_info_1 = 0;
4189 case SVM_EXIT_PAUSE:
4191 * We get this for NOP only, but pause
4192 * is rep not, check this here
4194 if (info->rep_prefix != REPE_PREFIX)
4197 case SVM_EXIT_IOIO: {
4201 if (info->intercept == x86_intercept_in ||
4202 info->intercept == x86_intercept_ins) {
4203 exit_info = ((info->src_val & 0xffff) << 16) |
4205 bytes = info->dst_bytes;
4207 exit_info = (info->dst_val & 0xffff) << 16;
4208 bytes = info->src_bytes;
4211 if (info->intercept == x86_intercept_outs ||
4212 info->intercept == x86_intercept_ins)
4213 exit_info |= SVM_IOIO_STR_MASK;
4215 if (info->rep_prefix)
4216 exit_info |= SVM_IOIO_REP_MASK;
4218 bytes = min(bytes, 4u);
4220 exit_info |= bytes << SVM_IOIO_SIZE_SHIFT;
4222 exit_info |= (u32)info->ad_bytes << (SVM_IOIO_ASIZE_SHIFT - 1);
4224 vmcb->control.exit_info_1 = exit_info;
4225 vmcb->control.exit_info_2 = info->next_rip;
4233 /* TODO: Advertise NRIPS to guest hypervisor unconditionally */
4234 if (static_cpu_has(X86_FEATURE_NRIPS))
4235 vmcb->control.next_rip = info->next_rip;
4236 vmcb->control.exit_code = icpt_info.exit_code;
4237 vmexit = nested_svm_exit_handled(svm);
4239 ret = (vmexit == NESTED_EXIT_DONE) ? X86EMUL_INTERCEPTED
4246 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
4250 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
4252 if (!kvm_pause_in_guest(vcpu->kvm))
4253 shrink_ple_window(vcpu);
4256 static void svm_setup_mce(struct kvm_vcpu *vcpu)
4258 /* [63:9] are reserved. */
4259 vcpu->arch.mcg_cap &= 0x1ff;
4262 bool svm_smi_blocked(struct kvm_vcpu *vcpu)
4264 struct vcpu_svm *svm = to_svm(vcpu);
4266 /* Per APM Vol.2 15.22.2 "Response to SMI" */
4270 return is_smm(vcpu);
4273 static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
4275 struct vcpu_svm *svm = to_svm(vcpu);
4276 if (svm->nested.nested_run_pending)
4279 /* An SMI must not be injected into L2 if it's supposed to VM-Exit. */
4280 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_smi(svm))
4283 return !svm_smi_blocked(vcpu);
4286 static int svm_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
4288 struct vcpu_svm *svm = to_svm(vcpu);
4289 struct kvm_host_map map_save;
4292 if (is_guest_mode(vcpu)) {
4293 /* FED8h - SVM Guest */
4294 put_smstate(u64, smstate, 0x7ed8, 1);
4295 /* FEE0h - SVM Guest VMCB Physical Address */
4296 put_smstate(u64, smstate, 0x7ee0, svm->nested.vmcb12_gpa);
4298 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4299 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4300 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4302 ret = nested_svm_vmexit(svm);
4307 * KVM uses VMCB01 to store L1 host state while L2 runs but
4308 * VMCB01 is going to be used during SMM and thus the state will
4309 * be lost. Temporary save non-VMLOAD/VMSAVE state to the host save
4310 * area pointed to by MSR_VM_HSAVE_PA. APM guarantees that the
4311 * format of the area is identical to guest save area offsetted
4312 * by 0x400 (matches the offset of 'struct vmcb_save_area'
4313 * within 'struct vmcb'). Note: HSAVE area may also be used by
4314 * L1 hypervisor to save additional host context (e.g. KVM does
4315 * that, see svm_prepare_guest_switch()) which must be
4318 if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr),
4319 &map_save) == -EINVAL)
4322 BUILD_BUG_ON(offsetof(struct vmcb, save) != 0x400);
4324 svm_copy_vmrun_state(map_save.hva + 0x400,
4325 &svm->vmcb01.ptr->save);
4327 kvm_vcpu_unmap(vcpu, &map_save, true);
4332 static int svm_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
4334 struct vcpu_svm *svm = to_svm(vcpu);
4335 struct kvm_host_map map, map_save;
4338 if (guest_cpuid_has(vcpu, X86_FEATURE_LM)) {
4339 u64 saved_efer = GET_SMSTATE(u64, smstate, 0x7ed0);
4340 u64 guest = GET_SMSTATE(u64, smstate, 0x7ed8);
4341 u64 vmcb12_gpa = GET_SMSTATE(u64, smstate, 0x7ee0);
4342 struct vmcb *vmcb12;
4345 if (!guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4348 if (!(saved_efer & EFER_SVME))
4351 if (kvm_vcpu_map(vcpu,
4352 gpa_to_gfn(vmcb12_gpa), &map) == -EINVAL)
4355 if (svm_allocate_nested(svm))
4360 nested_load_control_from_vmcb12(svm, &vmcb12->control);
4362 ret = enter_svm_guest_mode(vcpu, vmcb12_gpa, vmcb12);
4363 kvm_vcpu_unmap(vcpu, &map, true);
4366 * Restore L1 host state from L1 HSAVE area as VMCB01 was
4367 * used during SMM (see svm_enter_smm())
4369 if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr),
4370 &map_save) == -EINVAL)
4373 svm_copy_vmrun_state(&svm->vmcb01.ptr->save,
4374 map_save.hva + 0x400);
4376 kvm_vcpu_unmap(vcpu, &map_save, true);
4383 static void svm_enable_smi_window(struct kvm_vcpu *vcpu)
4385 struct vcpu_svm *svm = to_svm(vcpu);
4387 if (!gif_set(svm)) {
4388 if (vgif_enabled(svm))
4389 svm_set_intercept(svm, INTERCEPT_STGI);
4390 /* STGI will cause a vm exit */
4392 /* We must be in SMM; RSM will cause a vmexit anyway. */
4396 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, void *insn, int insn_len)
4398 bool smep, smap, is_user;
4402 * When the guest is an SEV-ES guest, emulation is not possible.
4404 if (sev_es_guest(vcpu->kvm))
4408 * Detect and workaround Errata 1096 Fam_17h_00_0Fh.
4411 * When CPU raise #NPF on guest data access and vCPU CR4.SMAP=1, it is
4412 * possible that CPU microcode implementing DecodeAssist will fail
4413 * to read bytes of instruction which caused #NPF. In this case,
4414 * GuestIntrBytes field of the VMCB on a VMEXIT will incorrectly
4415 * return 0 instead of the correct guest instruction bytes.
4417 * This happens because CPU microcode reading instruction bytes
4418 * uses a special opcode which attempts to read data using CPL=0
4419 * privileges. The microcode reads CS:RIP and if it hits a SMAP
4420 * fault, it gives up and returns no instruction bytes.
4423 * We reach here in case CPU supports DecodeAssist, raised #NPF and
4424 * returned 0 in GuestIntrBytes field of the VMCB.
4425 * First, errata can only be triggered in case vCPU CR4.SMAP=1.
4426 * Second, if vCPU CR4.SMEP=1, errata could only be triggered
4427 * in case vCPU CPL==3 (Because otherwise guest would have triggered
4428 * a SMEP fault instead of #NPF).
4429 * Otherwise, vCPU CR4.SMEP=0, errata could be triggered by any vCPU CPL.
4430 * As most guests enable SMAP if they have also enabled SMEP, use above
4431 * logic in order to attempt minimize false-positive of detecting errata
4432 * while still preserving all cases semantic correctness.
4435 * To determine what instruction the guest was executing, the hypervisor
4436 * will have to decode the instruction at the instruction pointer.
4438 * In non SEV guest, hypervisor will be able to read the guest
4439 * memory to decode the instruction pointer when insn_len is zero
4440 * so we return true to indicate that decoding is possible.
4442 * But in the SEV guest, the guest memory is encrypted with the
4443 * guest specific key and hypervisor will not be able to decode the
4444 * instruction pointer so we will not able to workaround it. Lets
4445 * print the error and request to kill the guest.
4447 if (likely(!insn || insn_len))
4451 * If RIP is invalid, go ahead with emulation which will cause an
4452 * internal error exit.
4454 if (!kvm_vcpu_gfn_to_memslot(vcpu, kvm_rip_read(vcpu) >> PAGE_SHIFT))
4457 cr4 = kvm_read_cr4(vcpu);
4458 smep = cr4 & X86_CR4_SMEP;
4459 smap = cr4 & X86_CR4_SMAP;
4460 is_user = svm_get_cpl(vcpu) == 3;
4461 if (smap && (!smep || is_user)) {
4462 if (!sev_guest(vcpu->kvm))
4465 pr_err_ratelimited("KVM: SEV Guest triggered AMD Erratum 1096\n");
4466 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
4472 static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
4474 struct vcpu_svm *svm = to_svm(vcpu);
4477 * TODO: Last condition latch INIT signals on vCPU when
4478 * vCPU is in guest-mode and vmcb12 defines intercept on INIT.
4479 * To properly emulate the INIT intercept,
4480 * svm_check_nested_events() should call nested_svm_vmexit()
4481 * if an INIT signal is pending.
4483 return !gif_set(svm) ||
4484 (vmcb_is_intercept(&svm->vmcb->control, INTERCEPT_INIT));
4487 static void svm_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
4489 if (!sev_es_guest(vcpu->kvm))
4490 return kvm_vcpu_deliver_sipi_vector(vcpu, vector);
4492 sev_vcpu_deliver_sipi_vector(vcpu, vector);
4495 static void svm_vm_destroy(struct kvm *kvm)
4497 avic_vm_destroy(kvm);
4498 sev_vm_destroy(kvm);
4501 static int svm_vm_init(struct kvm *kvm)
4503 if (!pause_filter_count || !pause_filter_thresh)
4504 kvm->arch.pause_in_guest = true;
4507 int ret = avic_vm_init(kvm);
4515 static struct kvm_x86_ops svm_x86_ops __initdata = {
4516 .hardware_unsetup = svm_hardware_teardown,
4517 .hardware_enable = svm_hardware_enable,
4518 .hardware_disable = svm_hardware_disable,
4519 .cpu_has_accelerated_tpr = svm_cpu_has_accelerated_tpr,
4520 .has_emulated_msr = svm_has_emulated_msr,
4522 .vcpu_create = svm_create_vcpu,
4523 .vcpu_free = svm_free_vcpu,
4524 .vcpu_reset = svm_vcpu_reset,
4526 .vm_size = sizeof(struct kvm_svm),
4527 .vm_init = svm_vm_init,
4528 .vm_destroy = svm_vm_destroy,
4530 .prepare_guest_switch = svm_prepare_guest_switch,
4531 .vcpu_load = svm_vcpu_load,
4532 .vcpu_put = svm_vcpu_put,
4533 .vcpu_blocking = svm_vcpu_blocking,
4534 .vcpu_unblocking = svm_vcpu_unblocking,
4536 .update_exception_bitmap = svm_update_exception_bitmap,
4537 .get_msr_feature = svm_get_msr_feature,
4538 .get_msr = svm_get_msr,
4539 .set_msr = svm_set_msr,
4540 .get_segment_base = svm_get_segment_base,
4541 .get_segment = svm_get_segment,
4542 .set_segment = svm_set_segment,
4543 .get_cpl = svm_get_cpl,
4544 .get_cs_db_l_bits = kvm_get_cs_db_l_bits,
4545 .set_cr0 = svm_set_cr0,
4546 .is_valid_cr4 = svm_is_valid_cr4,
4547 .set_cr4 = svm_set_cr4,
4548 .set_efer = svm_set_efer,
4549 .get_idt = svm_get_idt,
4550 .set_idt = svm_set_idt,
4551 .get_gdt = svm_get_gdt,
4552 .set_gdt = svm_set_gdt,
4553 .set_dr7 = svm_set_dr7,
4554 .sync_dirty_debug_regs = svm_sync_dirty_debug_regs,
4555 .cache_reg = svm_cache_reg,
4556 .get_rflags = svm_get_rflags,
4557 .set_rflags = svm_set_rflags,
4559 .tlb_flush_all = svm_flush_tlb,
4560 .tlb_flush_current = svm_flush_tlb,
4561 .tlb_flush_gva = svm_flush_tlb_gva,
4562 .tlb_flush_guest = svm_flush_tlb,
4564 .run = svm_vcpu_run,
4565 .handle_exit = handle_exit,
4566 .skip_emulated_instruction = skip_emulated_instruction,
4567 .update_emulated_instruction = NULL,
4568 .set_interrupt_shadow = svm_set_interrupt_shadow,
4569 .get_interrupt_shadow = svm_get_interrupt_shadow,
4570 .patch_hypercall = svm_patch_hypercall,
4571 .set_irq = svm_set_irq,
4572 .set_nmi = svm_inject_nmi,
4573 .queue_exception = svm_queue_exception,
4574 .cancel_injection = svm_cancel_injection,
4575 .interrupt_allowed = svm_interrupt_allowed,
4576 .nmi_allowed = svm_nmi_allowed,
4577 .get_nmi_mask = svm_get_nmi_mask,
4578 .set_nmi_mask = svm_set_nmi_mask,
4579 .enable_nmi_window = svm_enable_nmi_window,
4580 .enable_irq_window = svm_enable_irq_window,
4581 .update_cr8_intercept = svm_update_cr8_intercept,
4582 .set_virtual_apic_mode = svm_set_virtual_apic_mode,
4583 .refresh_apicv_exec_ctrl = svm_refresh_apicv_exec_ctrl,
4584 .check_apicv_inhibit_reasons = svm_check_apicv_inhibit_reasons,
4585 .load_eoi_exitmap = svm_load_eoi_exitmap,
4586 .hwapic_irr_update = svm_hwapic_irr_update,
4587 .hwapic_isr_update = svm_hwapic_isr_update,
4588 .sync_pir_to_irr = kvm_lapic_find_highest_irr,
4589 .apicv_post_state_restore = avic_post_state_restore,
4591 .set_tss_addr = svm_set_tss_addr,
4592 .set_identity_map_addr = svm_set_identity_map_addr,
4593 .get_mt_mask = svm_get_mt_mask,
4595 .get_exit_info = svm_get_exit_info,
4597 .vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
4599 .has_wbinvd_exit = svm_has_wbinvd_exit,
4601 .get_l2_tsc_offset = svm_get_l2_tsc_offset,
4602 .get_l2_tsc_multiplier = svm_get_l2_tsc_multiplier,
4603 .write_tsc_offset = svm_write_tsc_offset,
4604 .write_tsc_multiplier = svm_write_tsc_multiplier,
4606 .load_mmu_pgd = svm_load_mmu_pgd,
4608 .check_intercept = svm_check_intercept,
4609 .handle_exit_irqoff = svm_handle_exit_irqoff,
4611 .request_immediate_exit = __kvm_request_immediate_exit,
4613 .sched_in = svm_sched_in,
4615 .pmu_ops = &amd_pmu_ops,
4616 .nested_ops = &svm_nested_ops,
4618 .deliver_posted_interrupt = svm_deliver_avic_intr,
4619 .dy_apicv_has_pending_interrupt = svm_dy_apicv_has_pending_interrupt,
4620 .update_pi_irte = svm_update_pi_irte,
4621 .setup_mce = svm_setup_mce,
4623 .smi_allowed = svm_smi_allowed,
4624 .enter_smm = svm_enter_smm,
4625 .leave_smm = svm_leave_smm,
4626 .enable_smi_window = svm_enable_smi_window,
4628 .mem_enc_op = svm_mem_enc_op,
4629 .mem_enc_reg_region = svm_register_enc_region,
4630 .mem_enc_unreg_region = svm_unregister_enc_region,
4632 .vm_copy_enc_context_from = svm_vm_copy_asid_from,
4634 .can_emulate_instruction = svm_can_emulate_instruction,
4636 .apic_init_signal_blocked = svm_apic_init_signal_blocked,
4638 .msr_filter_changed = svm_msr_filter_changed,
4639 .complete_emulated_msr = svm_complete_emulated_msr,
4641 .vcpu_deliver_sipi_vector = svm_vcpu_deliver_sipi_vector,
4644 static struct kvm_x86_init_ops svm_init_ops __initdata = {
4645 .cpu_has_kvm_support = has_svm,
4646 .disabled_by_bios = is_disabled,
4647 .hardware_setup = svm_hardware_setup,
4648 .check_processor_compatibility = svm_check_processor_compat,
4650 .runtime_ops = &svm_x86_ops,
4653 static int __init svm_init(void)
4655 __unused_size_checks();
4657 return kvm_init(&svm_init_ops, sizeof(struct vcpu_svm),
4658 __alignof__(struct vcpu_svm), THIS_MODULE);
4661 static void __exit svm_exit(void)
4666 module_init(svm_init)
4667 module_exit(svm_exit)