1 #define pr_fmt(fmt) "SVM: " fmt
3 #include <linux/kvm_host.h>
7 #include "kvm_cache_regs.h"
12 #include <linux/module.h>
13 #include <linux/mod_devicetable.h>
14 #include <linux/kernel.h>
15 #include <linux/vmalloc.h>
16 #include <linux/highmem.h>
17 #include <linux/amd-iommu.h>
18 #include <linux/sched.h>
19 #include <linux/trace_events.h>
20 #include <linux/slab.h>
21 #include <linux/hashtable.h>
22 #include <linux/objtool.h>
23 #include <linux/psp-sev.h>
24 #include <linux/file.h>
25 #include <linux/pagemap.h>
26 #include <linux/swap.h>
27 #include <linux/rwsem.h>
30 #include <asm/perf_event.h>
31 #include <asm/tlbflush.h>
33 #include <asm/debugreg.h>
34 #include <asm/kvm_para.h>
35 #include <asm/irq_remapping.h>
36 #include <asm/spec-ctrl.h>
37 #include <asm/cpu_device_id.h>
38 #include <asm/traps.h>
40 #include <asm/virtext.h>
46 #define __ex(x) __kvm_handle_fault_on_reboot(x)
48 MODULE_AUTHOR("Qumranet");
49 MODULE_LICENSE("GPL");
52 static const struct x86_cpu_id svm_cpu_id[] = {
53 X86_MATCH_FEATURE(X86_FEATURE_SVM, NULL),
56 MODULE_DEVICE_TABLE(x86cpu, svm_cpu_id);
59 #define IOPM_ALLOC_ORDER 2
60 #define MSRPM_ALLOC_ORDER 1
62 #define SEG_TYPE_LDT 2
63 #define SEG_TYPE_BUSY_TSS16 3
65 #define SVM_FEATURE_LBRV (1 << 1)
66 #define SVM_FEATURE_SVML (1 << 2)
67 #define SVM_FEATURE_TSC_RATE (1 << 4)
68 #define SVM_FEATURE_VMCB_CLEAN (1 << 5)
69 #define SVM_FEATURE_FLUSH_ASID (1 << 6)
70 #define SVM_FEATURE_DECODE_ASSIST (1 << 7)
71 #define SVM_FEATURE_PAUSE_FILTER (1 << 10)
73 #define DEBUGCTL_RESERVED_BITS (~(0x3fULL))
75 #define TSC_RATIO_RSVD 0xffffff0000000000ULL
76 #define TSC_RATIO_MIN 0x0000000000000001ULL
77 #define TSC_RATIO_MAX 0x000000ffffffffffULL
79 static bool erratum_383_found __read_mostly;
81 u32 msrpm_offsets[MSRPM_OFFSETS] __read_mostly;
84 * Set osvw_len to higher value when updated Revision Guides
85 * are published and we know what the new status bits are
87 static uint64_t osvw_len = 4, osvw_status;
89 static DEFINE_PER_CPU(u64, current_tsc_ratio);
90 #define TSC_RATIO_DEFAULT 0x0100000000ULL
92 static const struct svm_direct_access_msrs {
93 u32 index; /* Index of the MSR */
94 bool always; /* True if intercept is initially cleared */
95 } direct_access_msrs[MAX_DIRECT_ACCESS_MSRS] = {
96 { .index = MSR_STAR, .always = true },
97 { .index = MSR_IA32_SYSENTER_CS, .always = true },
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 },
118 /* enable NPT for AMD64 and X86 with PAE */
119 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
120 bool npt_enabled = true;
126 * These 2 parameters are used to config the controls for Pause-Loop Exiting:
127 * pause_filter_count: On processors that support Pause filtering(indicated
128 * by CPUID Fn8000_000A_EDX), the VMCB provides a 16 bit pause filter
129 * count value. On VMRUN this value is loaded into an internal counter.
130 * Each time a pause instruction is executed, this counter is decremented
131 * until it reaches zero at which time a #VMEXIT is generated if pause
132 * intercept is enabled. Refer to AMD APM Vol 2 Section 15.14.4 Pause
133 * Intercept Filtering for more details.
134 * This also indicate if ple logic enabled.
136 * pause_filter_thresh: In addition, some processor families support advanced
137 * pause filtering (indicated by CPUID Fn8000_000A_EDX) upper bound on
138 * the amount of time a guest is allowed to execute in a pause loop.
139 * In this mode, a 16-bit pause filter threshold field is added in the
140 * VMCB. The threshold value is a cycle count that is used to reset the
141 * pause counter. As with simple pause filtering, VMRUN loads the pause
142 * count value from VMCB into an internal counter. Then, on each pause
143 * instruction the hardware checks the elapsed number of cycles since
144 * the most recent pause instruction against the pause filter threshold.
145 * If the elapsed cycle count is greater than the pause filter threshold,
146 * then the internal pause count is reloaded from the VMCB and execution
147 * continues. If the elapsed cycle count is less than the pause filter
148 * threshold, then the internal pause count is decremented. If the count
149 * value is less than zero and PAUSE intercept is enabled, a #VMEXIT is
150 * triggered. If advanced pause filtering is supported and pause filter
151 * threshold field is set to zero, the filter will operate in the simpler,
155 static unsigned short pause_filter_thresh = KVM_DEFAULT_PLE_GAP;
156 module_param(pause_filter_thresh, ushort, 0444);
158 static unsigned short pause_filter_count = KVM_SVM_DEFAULT_PLE_WINDOW;
159 module_param(pause_filter_count, ushort, 0444);
161 /* Default doubles per-vcpu window every exit. */
162 static unsigned short pause_filter_count_grow = KVM_DEFAULT_PLE_WINDOW_GROW;
163 module_param(pause_filter_count_grow, ushort, 0444);
165 /* Default resets per-vcpu window every exit to pause_filter_count. */
166 static unsigned short pause_filter_count_shrink = KVM_DEFAULT_PLE_WINDOW_SHRINK;
167 module_param(pause_filter_count_shrink, ushort, 0444);
169 /* Default is to compute the maximum so we can never overflow. */
170 static unsigned short pause_filter_count_max = KVM_SVM_DEFAULT_PLE_WINDOW_MAX;
171 module_param(pause_filter_count_max, ushort, 0444);
173 /* allow nested paging (virtualized MMU) for all guests */
174 static int npt = true;
175 module_param(npt, int, S_IRUGO);
177 /* allow nested virtualization in KVM/SVM */
178 static int nested = true;
179 module_param(nested, int, S_IRUGO);
181 /* enable/disable Next RIP Save */
182 static int nrips = true;
183 module_param(nrips, int, 0444);
185 /* enable/disable Virtual VMLOAD VMSAVE */
186 static int vls = true;
187 module_param(vls, int, 0444);
189 /* enable/disable Virtual GIF */
190 static int vgif = true;
191 module_param(vgif, int, 0444);
193 /* enable/disable SEV support */
194 int sev = IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT_ACTIVE_BY_DEFAULT);
195 module_param(sev, int, 0444);
197 /* enable/disable SEV-ES support */
198 int sev_es = IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT_ACTIVE_BY_DEFAULT);
199 module_param(sev_es, int, 0444);
201 bool __read_mostly dump_invalid_vmcb;
202 module_param(dump_invalid_vmcb, bool, 0644);
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);
221 static const u32 msrpm_ranges[] = {0, 0xc0000000, 0xc0010000};
223 #define NUM_MSR_MAPS ARRAY_SIZE(msrpm_ranges)
224 #define MSRS_RANGE_SIZE 2048
225 #define MSRS_IN_RANGE (MSRS_RANGE_SIZE * 8 / 2)
227 u32 svm_msrpm_offset(u32 msr)
232 for (i = 0; i < NUM_MSR_MAPS; i++) {
233 if (msr < msrpm_ranges[i] ||
234 msr >= msrpm_ranges[i] + MSRS_IN_RANGE)
237 offset = (msr - msrpm_ranges[i]) / 4; /* 4 msrs per u8 */
238 offset += (i * MSRS_RANGE_SIZE); /* add range offset */
240 /* Now we have the u8 offset - but need the u32 offset */
244 /* MSR not in any range */
248 #define MAX_INST_SIZE 15
250 static int get_max_npt_level(void)
253 return PT64_ROOT_4LEVEL;
255 return PT32E_ROOT_LEVEL;
259 int svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)
261 struct vcpu_svm *svm = to_svm(vcpu);
262 u64 old_efer = vcpu->arch.efer;
263 vcpu->arch.efer = efer;
266 /* Shadow paging assumes NX to be available. */
269 if (!(efer & EFER_LMA))
273 if ((old_efer & EFER_SVME) != (efer & EFER_SVME)) {
274 if (!(efer & EFER_SVME)) {
275 svm_leave_nested(svm);
276 svm_set_gif(svm, true);
277 /* #GP intercept is still needed for vmware backdoor */
278 if (!enable_vmware_backdoor)
279 clr_exception_intercept(svm, GP_VECTOR);
282 * Free the nested guest state, unless we are in SMM.
283 * In this case we will return to the nested guest
284 * as soon as we leave SMM.
286 if (!is_smm(&svm->vcpu))
287 svm_free_nested(svm);
290 int ret = svm_allocate_nested(svm);
293 vcpu->arch.efer = old_efer;
297 if (svm_gp_erratum_intercept)
298 set_exception_intercept(svm, GP_VECTOR);
302 svm->vmcb->save.efer = efer | EFER_SVME;
303 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
307 static int is_external_interrupt(u32 info)
309 info &= SVM_EVTINJ_TYPE_MASK | SVM_EVTINJ_VALID;
310 return info == (SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR);
313 static u32 svm_get_interrupt_shadow(struct kvm_vcpu *vcpu)
315 struct vcpu_svm *svm = to_svm(vcpu);
318 if (svm->vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK)
319 ret = KVM_X86_SHADOW_INT_STI | KVM_X86_SHADOW_INT_MOV_SS;
323 static void svm_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
325 struct vcpu_svm *svm = to_svm(vcpu);
328 svm->vmcb->control.int_state &= ~SVM_INTERRUPT_SHADOW_MASK;
330 svm->vmcb->control.int_state |= SVM_INTERRUPT_SHADOW_MASK;
334 static int skip_emulated_instruction(struct kvm_vcpu *vcpu)
336 struct vcpu_svm *svm = to_svm(vcpu);
339 * SEV-ES does not expose the next RIP. The RIP update is controlled by
340 * the type of exit and the #VC handler in the guest.
342 if (sev_es_guest(vcpu->kvm))
345 if (nrips && svm->vmcb->control.next_rip != 0) {
346 WARN_ON_ONCE(!static_cpu_has(X86_FEATURE_NRIPS));
347 svm->next_rip = svm->vmcb->control.next_rip;
350 if (!svm->next_rip) {
351 if (!kvm_emulate_instruction(vcpu, EMULTYPE_SKIP))
354 kvm_rip_write(vcpu, svm->next_rip);
358 svm_set_interrupt_shadow(vcpu, 0);
363 static void svm_queue_exception(struct kvm_vcpu *vcpu)
365 struct vcpu_svm *svm = to_svm(vcpu);
366 unsigned nr = vcpu->arch.exception.nr;
367 bool has_error_code = vcpu->arch.exception.has_error_code;
368 u32 error_code = vcpu->arch.exception.error_code;
370 kvm_deliver_exception_payload(&svm->vcpu);
372 if (nr == BP_VECTOR && !nrips) {
373 unsigned long rip, old_rip = kvm_rip_read(&svm->vcpu);
376 * For guest debugging where we have to reinject #BP if some
377 * INT3 is guest-owned:
378 * Emulate nRIP by moving RIP forward. Will fail if injection
379 * raises a fault that is not intercepted. Still better than
380 * failing in all cases.
382 (void)skip_emulated_instruction(&svm->vcpu);
383 rip = kvm_rip_read(&svm->vcpu);
384 svm->int3_rip = rip + svm->vmcb->save.cs.base;
385 svm->int3_injected = rip - old_rip;
388 svm->vmcb->control.event_inj = nr
390 | (has_error_code ? SVM_EVTINJ_VALID_ERR : 0)
391 | SVM_EVTINJ_TYPE_EXEPT;
392 svm->vmcb->control.event_inj_err = error_code;
395 static void svm_init_erratum_383(void)
401 if (!static_cpu_has_bug(X86_BUG_AMD_TLB_MMATCH))
404 /* Use _safe variants to not break nested virtualization */
405 val = native_read_msr_safe(MSR_AMD64_DC_CFG, &err);
411 low = lower_32_bits(val);
412 high = upper_32_bits(val);
414 native_write_msr_safe(MSR_AMD64_DC_CFG, low, high);
416 erratum_383_found = true;
419 static void svm_init_osvw(struct kvm_vcpu *vcpu)
422 * Guests should see errata 400 and 415 as fixed (assuming that
423 * HLT and IO instructions are intercepted).
425 vcpu->arch.osvw.length = (osvw_len >= 3) ? (osvw_len) : 3;
426 vcpu->arch.osvw.status = osvw_status & ~(6ULL);
429 * By increasing VCPU's osvw.length to 3 we are telling the guest that
430 * all osvw.status bits inside that length, including bit 0 (which is
431 * reserved for erratum 298), are valid. However, if host processor's
432 * osvw_len is 0 then osvw_status[0] carries no information. We need to
433 * be conservative here and therefore we tell the guest that erratum 298
434 * is present (because we really don't know).
436 if (osvw_len == 0 && boot_cpu_data.x86 == 0x10)
437 vcpu->arch.osvw.status |= 1;
440 static int has_svm(void)
444 if (!cpu_has_svm(&msg)) {
445 printk(KERN_INFO "has_svm: %s\n", msg);
450 pr_info("KVM is unsupported when running as an SEV guest\n");
457 static void svm_hardware_disable(void)
459 /* Make sure we clean up behind us */
460 if (static_cpu_has(X86_FEATURE_TSCRATEMSR))
461 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
465 amd_pmu_disable_virt();
468 static int svm_hardware_enable(void)
471 struct svm_cpu_data *sd;
473 struct desc_struct *gdt;
474 int me = raw_smp_processor_id();
476 rdmsrl(MSR_EFER, efer);
477 if (efer & EFER_SVME)
481 pr_err("%s: err EOPNOTSUPP on %d\n", __func__, me);
484 sd = per_cpu(svm_data, me);
486 pr_err("%s: svm_data is NULL on %d\n", __func__, me);
490 sd->asid_generation = 1;
491 sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
492 sd->next_asid = sd->max_asid + 1;
493 sd->min_asid = max_sev_asid + 1;
495 gdt = get_current_gdt_rw();
496 sd->tss_desc = (struct kvm_ldttss_desc *)(gdt + GDT_ENTRY_TSS);
498 wrmsrl(MSR_EFER, efer | EFER_SVME);
500 wrmsrl(MSR_VM_HSAVE_PA, __sme_page_pa(sd->save_area));
502 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
503 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
504 __this_cpu_write(current_tsc_ratio, TSC_RATIO_DEFAULT);
511 * Note that it is possible to have a system with mixed processor
512 * revisions and therefore different OSVW bits. If bits are not the same
513 * on different processors then choose the worst case (i.e. if erratum
514 * is present on one processor and not on another then assume that the
515 * erratum is present everywhere).
517 if (cpu_has(&boot_cpu_data, X86_FEATURE_OSVW)) {
518 uint64_t len, status = 0;
521 len = native_read_msr_safe(MSR_AMD64_OSVW_ID_LENGTH, &err);
523 status = native_read_msr_safe(MSR_AMD64_OSVW_STATUS,
527 osvw_status = osvw_len = 0;
531 osvw_status |= status;
532 osvw_status &= (1ULL << osvw_len) - 1;
535 osvw_status = osvw_len = 0;
537 svm_init_erratum_383();
539 amd_pmu_enable_virt();
544 static void svm_cpu_uninit(int cpu)
546 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
551 per_cpu(svm_data, cpu) = NULL;
552 kfree(sd->sev_vmcbs);
553 __free_page(sd->save_area);
557 static int svm_cpu_init(int cpu)
559 struct svm_cpu_data *sd;
561 sd = kzalloc(sizeof(struct svm_cpu_data), GFP_KERNEL);
565 sd->save_area = alloc_page(GFP_KERNEL);
568 clear_page(page_address(sd->save_area));
570 if (svm_sev_enabled()) {
571 sd->sev_vmcbs = kmalloc_array(max_sev_asid + 1,
578 per_cpu(svm_data, cpu) = sd;
583 __free_page(sd->save_area);
590 static int direct_access_msr_slot(u32 msr)
594 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++)
595 if (direct_access_msrs[i].index == msr)
601 static void set_shadow_msr_intercept(struct kvm_vcpu *vcpu, u32 msr, int read,
604 struct vcpu_svm *svm = to_svm(vcpu);
605 int slot = direct_access_msr_slot(msr);
610 /* Set the shadow bitmaps to the desired intercept states */
612 set_bit(slot, svm->shadow_msr_intercept.read);
614 clear_bit(slot, svm->shadow_msr_intercept.read);
617 set_bit(slot, svm->shadow_msr_intercept.write);
619 clear_bit(slot, svm->shadow_msr_intercept.write);
622 static bool valid_msr_intercept(u32 index)
624 return direct_access_msr_slot(index) != -ENOENT;
627 static bool msr_write_intercepted(struct kvm_vcpu *vcpu, u32 msr)
634 msrpm = is_guest_mode(vcpu) ? to_svm(vcpu)->nested.msrpm:
637 offset = svm_msrpm_offset(msr);
638 bit_write = 2 * (msr & 0x0f) + 1;
641 BUG_ON(offset == MSR_INVALID);
643 return !!test_bit(bit_write, &tmp);
646 static void set_msr_interception_bitmap(struct kvm_vcpu *vcpu, u32 *msrpm,
647 u32 msr, int read, int write)
649 u8 bit_read, bit_write;
654 * If this warning triggers extend the direct_access_msrs list at the
655 * beginning of the file
657 WARN_ON(!valid_msr_intercept(msr));
659 /* Enforce non allowed MSRs to trap */
660 if (read && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_READ))
663 if (write && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_WRITE))
666 offset = svm_msrpm_offset(msr);
667 bit_read = 2 * (msr & 0x0f);
668 bit_write = 2 * (msr & 0x0f) + 1;
671 BUG_ON(offset == MSR_INVALID);
673 read ? clear_bit(bit_read, &tmp) : set_bit(bit_read, &tmp);
674 write ? clear_bit(bit_write, &tmp) : set_bit(bit_write, &tmp);
679 void set_msr_interception(struct kvm_vcpu *vcpu, u32 *msrpm, u32 msr,
682 set_shadow_msr_intercept(vcpu, msr, read, write);
683 set_msr_interception_bitmap(vcpu, msrpm, msr, read, write);
686 u32 *svm_vcpu_alloc_msrpm(void)
688 struct page *pages = alloc_pages(GFP_KERNEL_ACCOUNT, MSRPM_ALLOC_ORDER);
694 msrpm = page_address(pages);
695 memset(msrpm, 0xff, PAGE_SIZE * (1 << MSRPM_ALLOC_ORDER));
700 void svm_vcpu_init_msrpm(struct kvm_vcpu *vcpu, u32 *msrpm)
704 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
705 if (!direct_access_msrs[i].always)
707 set_msr_interception(vcpu, msrpm, direct_access_msrs[i].index, 1, 1);
712 void svm_vcpu_free_msrpm(u32 *msrpm)
714 __free_pages(virt_to_page(msrpm), MSRPM_ALLOC_ORDER);
717 static void svm_msr_filter_changed(struct kvm_vcpu *vcpu)
719 struct vcpu_svm *svm = to_svm(vcpu);
723 * Set intercept permissions for all direct access MSRs again. They
724 * will automatically get filtered through the MSR filter, so we are
725 * back in sync after this.
727 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
728 u32 msr = direct_access_msrs[i].index;
729 u32 read = test_bit(i, svm->shadow_msr_intercept.read);
730 u32 write = test_bit(i, svm->shadow_msr_intercept.write);
732 set_msr_interception_bitmap(vcpu, svm->msrpm, msr, read, write);
736 static void add_msr_offset(u32 offset)
740 for (i = 0; i < MSRPM_OFFSETS; ++i) {
742 /* Offset already in list? */
743 if (msrpm_offsets[i] == offset)
746 /* Slot used by another offset? */
747 if (msrpm_offsets[i] != MSR_INVALID)
750 /* Add offset to list */
751 msrpm_offsets[i] = offset;
757 * If this BUG triggers the msrpm_offsets table has an overflow. Just
758 * increase MSRPM_OFFSETS in this case.
763 static void init_msrpm_offsets(void)
767 memset(msrpm_offsets, 0xff, sizeof(msrpm_offsets));
769 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
772 offset = svm_msrpm_offset(direct_access_msrs[i].index);
773 BUG_ON(offset == MSR_INVALID);
775 add_msr_offset(offset);
779 static void svm_enable_lbrv(struct kvm_vcpu *vcpu)
781 struct vcpu_svm *svm = to_svm(vcpu);
783 svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
784 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
785 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
786 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
787 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
790 static void svm_disable_lbrv(struct kvm_vcpu *vcpu)
792 struct vcpu_svm *svm = to_svm(vcpu);
794 svm->vmcb->control.virt_ext &= ~LBR_CTL_ENABLE_MASK;
795 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 0, 0);
796 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 0, 0);
797 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 0, 0);
798 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 0, 0);
801 void disable_nmi_singlestep(struct vcpu_svm *svm)
803 svm->nmi_singlestep = false;
805 if (!(svm->vcpu.guest_debug & KVM_GUESTDBG_SINGLESTEP)) {
806 /* Clear our flags if they were not set by the guest */
807 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
808 svm->vmcb->save.rflags &= ~X86_EFLAGS_TF;
809 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
810 svm->vmcb->save.rflags &= ~X86_EFLAGS_RF;
814 static void grow_ple_window(struct kvm_vcpu *vcpu)
816 struct vcpu_svm *svm = to_svm(vcpu);
817 struct vmcb_control_area *control = &svm->vmcb->control;
818 int old = control->pause_filter_count;
820 control->pause_filter_count = __grow_ple_window(old,
822 pause_filter_count_grow,
823 pause_filter_count_max);
825 if (control->pause_filter_count != old) {
826 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
827 trace_kvm_ple_window_update(vcpu->vcpu_id,
828 control->pause_filter_count, old);
832 static void shrink_ple_window(struct kvm_vcpu *vcpu)
834 struct vcpu_svm *svm = to_svm(vcpu);
835 struct vmcb_control_area *control = &svm->vmcb->control;
836 int old = control->pause_filter_count;
838 control->pause_filter_count =
839 __shrink_ple_window(old,
841 pause_filter_count_shrink,
843 if (control->pause_filter_count != old) {
844 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
845 trace_kvm_ple_window_update(vcpu->vcpu_id,
846 control->pause_filter_count, old);
851 * The default MMIO mask is a single bit (excluding the present bit),
852 * which could conflict with the memory encryption bit. Check for
853 * memory encryption support and override the default MMIO mask if
854 * memory encryption is enabled.
856 static __init void svm_adjust_mmio_mask(void)
858 unsigned int enc_bit, mask_bit;
861 /* If there is no memory encryption support, use existing mask */
862 if (cpuid_eax(0x80000000) < 0x8000001f)
865 /* If memory encryption is not enabled, use existing mask */
866 rdmsrl(MSR_K8_SYSCFG, msr);
867 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
870 enc_bit = cpuid_ebx(0x8000001f) & 0x3f;
871 mask_bit = boot_cpu_data.x86_phys_bits;
873 /* Increment the mask bit if it is the same as the encryption bit */
874 if (enc_bit == mask_bit)
878 * If the mask bit location is below 52, then some bits above the
879 * physical addressing limit will always be reserved, so use the
880 * rsvd_bits() function to generate the mask. This mask, along with
881 * the present bit, will be used to generate a page fault with
884 * If the mask bit location is 52 (or above), then clear the mask.
886 mask = (mask_bit < 52) ? rsvd_bits(mask_bit, 51) | PT_PRESENT_MASK : 0;
888 kvm_mmu_set_mmio_spte_mask(mask, PT_WRITABLE_MASK | PT_USER_MASK);
891 static void svm_hardware_teardown(void)
895 if (svm_sev_enabled())
896 sev_hardware_teardown();
898 for_each_possible_cpu(cpu)
901 __free_pages(pfn_to_page(iopm_base >> PAGE_SHIFT), IOPM_ALLOC_ORDER);
905 static __init void svm_set_cpu_caps(void)
911 /* CPUID 0x80000001 and 0x8000000A (SVM features) */
913 kvm_cpu_cap_set(X86_FEATURE_SVM);
916 kvm_cpu_cap_set(X86_FEATURE_NRIPS);
919 kvm_cpu_cap_set(X86_FEATURE_NPT);
921 /* Nested VM can receive #VMEXIT instead of triggering #GP */
922 kvm_cpu_cap_set(X86_FEATURE_SVME_ADDR_CHK);
925 /* CPUID 0x80000008 */
926 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) ||
927 boot_cpu_has(X86_FEATURE_AMD_SSBD))
928 kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
931 static __init int svm_hardware_setup(void)
934 struct page *iopm_pages;
938 iopm_pages = alloc_pages(GFP_KERNEL, IOPM_ALLOC_ORDER);
943 iopm_va = page_address(iopm_pages);
944 memset(iopm_va, 0xff, PAGE_SIZE * (1 << IOPM_ALLOC_ORDER));
945 iopm_base = page_to_pfn(iopm_pages) << PAGE_SHIFT;
947 init_msrpm_offsets();
949 supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
951 if (boot_cpu_has(X86_FEATURE_NX))
952 kvm_enable_efer_bits(EFER_NX);
954 if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
955 kvm_enable_efer_bits(EFER_FFXSR);
957 if (boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
958 kvm_has_tsc_control = true;
959 kvm_max_tsc_scaling_ratio = TSC_RATIO_MAX;
960 kvm_tsc_scaling_ratio_frac_bits = 32;
963 /* Check for pause filtering support */
964 if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
965 pause_filter_count = 0;
966 pause_filter_thresh = 0;
967 } else if (!boot_cpu_has(X86_FEATURE_PFTHRESHOLD)) {
968 pause_filter_thresh = 0;
972 printk(KERN_INFO "kvm: Nested Virtualization enabled\n");
973 kvm_enable_efer_bits(EFER_SVME | EFER_LMSLE);
976 if (IS_ENABLED(CONFIG_KVM_AMD_SEV) && sev) {
977 sev_hardware_setup();
983 svm_adjust_mmio_mask();
985 for_each_possible_cpu(cpu) {
986 r = svm_cpu_init(cpu);
991 if (!boot_cpu_has(X86_FEATURE_NPT))
994 if (npt_enabled && !npt)
997 kvm_configure_mmu(npt_enabled, get_max_npt_level(), PG_LEVEL_1G);
998 pr_info("kvm: Nested Paging %sabled\n", npt_enabled ? "en" : "dis");
1001 if (!boot_cpu_has(X86_FEATURE_NRIPS))
1007 !boot_cpu_has(X86_FEATURE_AVIC) ||
1008 !IS_ENABLED(CONFIG_X86_LOCAL_APIC)) {
1011 pr_info("AVIC enabled\n");
1013 amd_iommu_register_ga_log_notifier(&avic_ga_log_notifier);
1019 !boot_cpu_has(X86_FEATURE_V_VMSAVE_VMLOAD) ||
1020 !IS_ENABLED(CONFIG_X86_64)) {
1023 pr_info("Virtual VMLOAD VMSAVE supported\n");
1027 if (boot_cpu_has(X86_FEATURE_SVME_ADDR_CHK))
1028 svm_gp_erratum_intercept = false;
1031 if (!boot_cpu_has(X86_FEATURE_VGIF))
1034 pr_info("Virtual GIF supported\n");
1040 * It seems that on AMD processors PTE's accessed bit is
1041 * being set by the CPU hardware before the NPF vmexit.
1042 * This is not expected behaviour and our tests fail because
1044 * A workaround here is to disable support for
1045 * GUEST_MAXPHYADDR < HOST_MAXPHYADDR if NPT is enabled.
1046 * In this case userspace can know if there is support using
1047 * KVM_CAP_SMALLER_MAXPHYADDR extension and decide how to handle
1049 * If future AMD CPU models change the behaviour described above,
1050 * this variable can be changed accordingly
1052 allow_smaller_maxphyaddr = !npt_enabled;
1057 svm_hardware_teardown();
1061 static void init_seg(struct vmcb_seg *seg)
1064 seg->attrib = SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK |
1065 SVM_SELECTOR_WRITE_MASK; /* Read/Write Data Segment */
1066 seg->limit = 0xffff;
1070 static void init_sys_seg(struct vmcb_seg *seg, uint32_t type)
1073 seg->attrib = SVM_SELECTOR_P_MASK | type;
1074 seg->limit = 0xffff;
1078 static u64 svm_write_l1_tsc_offset(struct kvm_vcpu *vcpu, u64 offset)
1080 struct vcpu_svm *svm = to_svm(vcpu);
1081 u64 g_tsc_offset = 0;
1083 if (is_guest_mode(vcpu)) {
1084 /* Write L1's TSC offset. */
1085 g_tsc_offset = svm->vmcb->control.tsc_offset -
1086 svm->nested.hsave->control.tsc_offset;
1087 svm->nested.hsave->control.tsc_offset = offset;
1090 trace_kvm_write_tsc_offset(vcpu->vcpu_id,
1091 svm->vmcb->control.tsc_offset - g_tsc_offset,
1094 svm->vmcb->control.tsc_offset = offset + g_tsc_offset;
1096 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1097 return svm->vmcb->control.tsc_offset;
1100 static void svm_check_invpcid(struct vcpu_svm *svm)
1103 * Intercept INVPCID if shadow paging is enabled to sync/free shadow
1104 * roots, or if INVPCID is disabled in the guest to inject #UD.
1106 if (kvm_cpu_cap_has(X86_FEATURE_INVPCID)) {
1108 !guest_cpuid_has(&svm->vcpu, X86_FEATURE_INVPCID))
1109 svm_set_intercept(svm, INTERCEPT_INVPCID);
1111 svm_clr_intercept(svm, INTERCEPT_INVPCID);
1115 static void init_vmcb(struct vcpu_svm *svm)
1117 struct vmcb_control_area *control = &svm->vmcb->control;
1118 struct vmcb_save_area *save = &svm->vmcb->save;
1120 svm->vcpu.arch.hflags = 0;
1122 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1123 svm_set_intercept(svm, INTERCEPT_CR3_READ);
1124 svm_set_intercept(svm, INTERCEPT_CR4_READ);
1125 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1126 svm_set_intercept(svm, INTERCEPT_CR3_WRITE);
1127 svm_set_intercept(svm, INTERCEPT_CR4_WRITE);
1128 if (!kvm_vcpu_apicv_active(&svm->vcpu))
1129 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
1131 set_dr_intercepts(svm);
1133 set_exception_intercept(svm, PF_VECTOR);
1134 set_exception_intercept(svm, UD_VECTOR);
1135 set_exception_intercept(svm, MC_VECTOR);
1136 set_exception_intercept(svm, AC_VECTOR);
1137 set_exception_intercept(svm, DB_VECTOR);
1139 * Guest access to VMware backdoor ports could legitimately
1140 * trigger #GP because of TSS I/O permission bitmap.
1141 * We intercept those #GP and allow access to them anyway
1144 if (enable_vmware_backdoor)
1145 set_exception_intercept(svm, GP_VECTOR);
1147 svm_set_intercept(svm, INTERCEPT_INTR);
1148 svm_set_intercept(svm, INTERCEPT_NMI);
1149 svm_set_intercept(svm, INTERCEPT_SMI);
1150 svm_set_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1151 svm_set_intercept(svm, INTERCEPT_RDPMC);
1152 svm_set_intercept(svm, INTERCEPT_CPUID);
1153 svm_set_intercept(svm, INTERCEPT_INVD);
1154 svm_set_intercept(svm, INTERCEPT_INVLPG);
1155 svm_set_intercept(svm, INTERCEPT_INVLPGA);
1156 svm_set_intercept(svm, INTERCEPT_IOIO_PROT);
1157 svm_set_intercept(svm, INTERCEPT_MSR_PROT);
1158 svm_set_intercept(svm, INTERCEPT_TASK_SWITCH);
1159 svm_set_intercept(svm, INTERCEPT_SHUTDOWN);
1160 svm_set_intercept(svm, INTERCEPT_VMRUN);
1161 svm_set_intercept(svm, INTERCEPT_VMMCALL);
1162 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1163 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1164 svm_set_intercept(svm, INTERCEPT_STGI);
1165 svm_set_intercept(svm, INTERCEPT_CLGI);
1166 svm_set_intercept(svm, INTERCEPT_SKINIT);
1167 svm_set_intercept(svm, INTERCEPT_WBINVD);
1168 svm_set_intercept(svm, INTERCEPT_XSETBV);
1169 svm_set_intercept(svm, INTERCEPT_RDPRU);
1170 svm_set_intercept(svm, INTERCEPT_RSM);
1172 if (!kvm_mwait_in_guest(svm->vcpu.kvm)) {
1173 svm_set_intercept(svm, INTERCEPT_MONITOR);
1174 svm_set_intercept(svm, INTERCEPT_MWAIT);
1177 if (!kvm_hlt_in_guest(svm->vcpu.kvm))
1178 svm_set_intercept(svm, INTERCEPT_HLT);
1180 control->iopm_base_pa = __sme_set(iopm_base);
1181 control->msrpm_base_pa = __sme_set(__pa(svm->msrpm));
1182 control->int_ctl = V_INTR_MASKING_MASK;
1184 init_seg(&save->es);
1185 init_seg(&save->ss);
1186 init_seg(&save->ds);
1187 init_seg(&save->fs);
1188 init_seg(&save->gs);
1190 save->cs.selector = 0xf000;
1191 save->cs.base = 0xffff0000;
1192 /* Executable/Readable Code Segment */
1193 save->cs.attrib = SVM_SELECTOR_READ_MASK | SVM_SELECTOR_P_MASK |
1194 SVM_SELECTOR_S_MASK | SVM_SELECTOR_CODE_MASK;
1195 save->cs.limit = 0xffff;
1197 save->gdtr.limit = 0xffff;
1198 save->idtr.limit = 0xffff;
1200 init_sys_seg(&save->ldtr, SEG_TYPE_LDT);
1201 init_sys_seg(&save->tr, SEG_TYPE_BUSY_TSS16);
1203 svm_set_efer(&svm->vcpu, 0);
1204 save->dr6 = 0xffff0ff0;
1205 kvm_set_rflags(&svm->vcpu, X86_EFLAGS_FIXED);
1206 save->rip = 0x0000fff0;
1207 svm->vcpu.arch.regs[VCPU_REGS_RIP] = save->rip;
1210 * svm_set_cr0() sets PG and WP and clears NW and CD on save->cr0.
1211 * It also updates the guest-visible cr0 value.
1213 svm_set_cr0(&svm->vcpu, X86_CR0_NW | X86_CR0_CD | X86_CR0_ET);
1214 kvm_mmu_reset_context(&svm->vcpu);
1216 save->cr4 = X86_CR4_PAE;
1220 /* Setup VMCB for Nested Paging */
1221 control->nested_ctl |= SVM_NESTED_CTL_NP_ENABLE;
1222 svm_clr_intercept(svm, INTERCEPT_INVLPG);
1223 clr_exception_intercept(svm, PF_VECTOR);
1224 svm_clr_intercept(svm, INTERCEPT_CR3_READ);
1225 svm_clr_intercept(svm, INTERCEPT_CR3_WRITE);
1226 save->g_pat = svm->vcpu.arch.pat;
1230 svm->asid_generation = 0;
1233 svm->nested.vmcb12_gpa = 0;
1234 svm->vcpu.arch.hflags = 0;
1236 if (!kvm_pause_in_guest(svm->vcpu.kvm)) {
1237 control->pause_filter_count = pause_filter_count;
1238 if (pause_filter_thresh)
1239 control->pause_filter_thresh = pause_filter_thresh;
1240 svm_set_intercept(svm, INTERCEPT_PAUSE);
1242 svm_clr_intercept(svm, INTERCEPT_PAUSE);
1245 svm_check_invpcid(svm);
1247 if (kvm_vcpu_apicv_active(&svm->vcpu))
1248 avic_init_vmcb(svm);
1251 * If hardware supports Virtual VMLOAD VMSAVE then enable it
1252 * in VMCB and clear intercepts to avoid #VMEXIT.
1255 svm_clr_intercept(svm, INTERCEPT_VMLOAD);
1256 svm_clr_intercept(svm, INTERCEPT_VMSAVE);
1257 svm->vmcb->control.virt_ext |= VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
1261 svm_clr_intercept(svm, INTERCEPT_STGI);
1262 svm_clr_intercept(svm, INTERCEPT_CLGI);
1263 svm->vmcb->control.int_ctl |= V_GIF_ENABLE_MASK;
1266 if (sev_guest(svm->vcpu.kvm)) {
1267 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
1268 clr_exception_intercept(svm, UD_VECTOR);
1270 if (sev_es_guest(svm->vcpu.kvm)) {
1271 /* Perform SEV-ES specific VMCB updates */
1272 sev_es_init_vmcb(svm);
1276 vmcb_mark_all_dirty(svm->vmcb);
1282 static void svm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
1284 struct vcpu_svm *svm = to_svm(vcpu);
1289 svm->virt_spec_ctrl = 0;
1292 svm->vcpu.arch.apic_base = APIC_DEFAULT_PHYS_BASE |
1293 MSR_IA32_APICBASE_ENABLE;
1294 if (kvm_vcpu_is_reset_bsp(&svm->vcpu))
1295 svm->vcpu.arch.apic_base |= MSR_IA32_APICBASE_BSP;
1299 kvm_cpuid(vcpu, &eax, &dummy, &dummy, &dummy, false);
1300 kvm_rdx_write(vcpu, eax);
1302 if (kvm_vcpu_apicv_active(vcpu) && !init_event)
1303 avic_update_vapic_bar(svm, APIC_DEFAULT_PHYS_BASE);
1306 static int svm_create_vcpu(struct kvm_vcpu *vcpu)
1308 struct vcpu_svm *svm;
1309 struct page *vmcb_page;
1310 struct page *vmsa_page = NULL;
1313 BUILD_BUG_ON(offsetof(struct vcpu_svm, vcpu) != 0);
1317 vmcb_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1321 if (sev_es_guest(svm->vcpu.kvm)) {
1323 * SEV-ES guests require a separate VMSA page used to contain
1324 * the encrypted register state of the guest.
1326 vmsa_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1328 goto error_free_vmcb_page;
1331 * SEV-ES guests maintain an encrypted version of their FPU
1332 * state which is restored and saved on VMRUN and VMEXIT.
1333 * Free the fpu structure to prevent KVM from attempting to
1334 * access the FPU state.
1336 kvm_free_guest_fpu(vcpu);
1339 err = avic_init_vcpu(svm);
1341 goto error_free_vmsa_page;
1343 /* We initialize this flag to true to make sure that the is_running
1344 * bit would be set the first time the vcpu is loaded.
1346 if (irqchip_in_kernel(vcpu->kvm) && kvm_apicv_activated(vcpu->kvm))
1347 svm->avic_is_running = true;
1349 svm->msrpm = svm_vcpu_alloc_msrpm();
1352 goto error_free_vmsa_page;
1355 svm_vcpu_init_msrpm(vcpu, svm->msrpm);
1357 svm->vmcb = page_address(vmcb_page);
1358 svm->vmcb_pa = __sme_set(page_to_pfn(vmcb_page) << PAGE_SHIFT);
1361 svm->vmsa = page_address(vmsa_page);
1363 svm->asid_generation = 0;
1364 svm->guest_state_loaded = false;
1367 svm_init_osvw(vcpu);
1368 vcpu->arch.microcode_version = 0x01000065;
1370 if (sev_es_guest(svm->vcpu.kvm))
1371 /* Perform SEV-ES specific VMCB creation updates */
1372 sev_es_create_vcpu(svm);
1376 error_free_vmsa_page:
1378 __free_page(vmsa_page);
1379 error_free_vmcb_page:
1380 __free_page(vmcb_page);
1385 static void svm_clear_current_vmcb(struct vmcb *vmcb)
1389 for_each_online_cpu(i)
1390 cmpxchg(&per_cpu(svm_data, i)->current_vmcb, vmcb, NULL);
1393 static void svm_free_vcpu(struct kvm_vcpu *vcpu)
1395 struct vcpu_svm *svm = to_svm(vcpu);
1398 * The vmcb page can be recycled, causing a false negative in
1399 * svm_vcpu_load(). So, ensure that no logical CPU has this
1400 * vmcb page recorded as its current vmcb.
1402 svm_clear_current_vmcb(svm->vmcb);
1404 svm_free_nested(svm);
1406 sev_free_vcpu(vcpu);
1408 __free_page(pfn_to_page(__sme_clr(svm->vmcb_pa) >> PAGE_SHIFT));
1409 __free_pages(virt_to_page(svm->msrpm), MSRPM_ALLOC_ORDER);
1412 static void svm_prepare_guest_switch(struct kvm_vcpu *vcpu)
1414 struct vcpu_svm *svm = to_svm(vcpu);
1415 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
1418 if (svm->guest_state_loaded)
1422 * Certain MSRs are restored on VMEXIT (sev-es), or vmload of host save
1423 * area (non-sev-es). Save ones that aren't so we can restore them
1424 * individually later.
1426 for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
1427 rdmsrl(host_save_user_msrs[i], svm->host_user_msrs[i]);
1430 * Save additional host state that will be restored on VMEXIT (sev-es)
1431 * or subsequent vmload of host save area.
1433 if (sev_es_guest(svm->vcpu.kvm)) {
1434 sev_es_prepare_guest_switch(svm, vcpu->cpu);
1436 vmsave(__sme_page_pa(sd->save_area));
1439 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
1440 u64 tsc_ratio = vcpu->arch.tsc_scaling_ratio;
1441 if (tsc_ratio != __this_cpu_read(current_tsc_ratio)) {
1442 __this_cpu_write(current_tsc_ratio, tsc_ratio);
1443 wrmsrl(MSR_AMD64_TSC_RATIO, tsc_ratio);
1447 /* This assumes that the kernel never uses MSR_TSC_AUX */
1448 if (static_cpu_has(X86_FEATURE_RDTSCP))
1449 wrmsrl(MSR_TSC_AUX, svm->tsc_aux);
1451 svm->guest_state_loaded = true;
1454 static void svm_prepare_host_switch(struct kvm_vcpu *vcpu)
1456 struct vcpu_svm *svm = to_svm(vcpu);
1459 if (!svm->guest_state_loaded)
1463 * Certain MSRs are restored on VMEXIT (sev-es), or vmload of host save
1464 * area (non-sev-es). Restore the ones that weren't.
1466 for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
1467 wrmsrl(host_save_user_msrs[i], svm->host_user_msrs[i]);
1469 svm->guest_state_loaded = false;
1472 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1474 struct vcpu_svm *svm = to_svm(vcpu);
1475 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1477 if (unlikely(cpu != vcpu->cpu)) {
1478 svm->asid_generation = 0;
1479 vmcb_mark_all_dirty(svm->vmcb);
1482 if (sd->current_vmcb != svm->vmcb) {
1483 sd->current_vmcb = svm->vmcb;
1484 indirect_branch_prediction_barrier();
1486 avic_vcpu_load(vcpu, cpu);
1489 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
1491 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));
1537 static void svm_set_vintr(struct vcpu_svm *svm)
1539 struct vmcb_control_area *control;
1541 /* The following fields are ignored when AVIC is enabled */
1542 WARN_ON(kvm_vcpu_apicv_active(&svm->vcpu));
1543 svm_set_intercept(svm, INTERCEPT_VINTR);
1546 * This is just a dummy VINTR to actually cause a vmexit to happen.
1547 * Actual injection of virtual interrupts happens through EVENTINJ.
1549 control = &svm->vmcb->control;
1550 control->int_vector = 0x0;
1551 control->int_ctl &= ~V_INTR_PRIO_MASK;
1552 control->int_ctl |= V_IRQ_MASK |
1553 ((/*control->int_vector >> 4*/ 0xf) << V_INTR_PRIO_SHIFT);
1554 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1557 static void svm_clear_vintr(struct vcpu_svm *svm)
1559 const u32 mask = V_TPR_MASK | V_GIF_ENABLE_MASK | V_GIF_MASK | V_INTR_MASKING_MASK;
1560 svm_clr_intercept(svm, INTERCEPT_VINTR);
1562 /* Drop int_ctl fields related to VINTR injection. */
1563 svm->vmcb->control.int_ctl &= mask;
1564 if (is_guest_mode(&svm->vcpu)) {
1565 svm->nested.hsave->control.int_ctl &= mask;
1567 WARN_ON((svm->vmcb->control.int_ctl & V_TPR_MASK) !=
1568 (svm->nested.ctl.int_ctl & V_TPR_MASK));
1569 svm->vmcb->control.int_ctl |= svm->nested.ctl.int_ctl & ~mask;
1572 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1575 static struct vmcb_seg *svm_seg(struct kvm_vcpu *vcpu, int seg)
1577 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1580 case VCPU_SREG_CS: return &save->cs;
1581 case VCPU_SREG_DS: return &save->ds;
1582 case VCPU_SREG_ES: return &save->es;
1583 case VCPU_SREG_FS: return &save->fs;
1584 case VCPU_SREG_GS: return &save->gs;
1585 case VCPU_SREG_SS: return &save->ss;
1586 case VCPU_SREG_TR: return &save->tr;
1587 case VCPU_SREG_LDTR: return &save->ldtr;
1593 static u64 svm_get_segment_base(struct kvm_vcpu *vcpu, int seg)
1595 struct vmcb_seg *s = svm_seg(vcpu, seg);
1600 static void svm_get_segment(struct kvm_vcpu *vcpu,
1601 struct kvm_segment *var, int seg)
1603 struct vmcb_seg *s = svm_seg(vcpu, seg);
1605 var->base = s->base;
1606 var->limit = s->limit;
1607 var->selector = s->selector;
1608 var->type = s->attrib & SVM_SELECTOR_TYPE_MASK;
1609 var->s = (s->attrib >> SVM_SELECTOR_S_SHIFT) & 1;
1610 var->dpl = (s->attrib >> SVM_SELECTOR_DPL_SHIFT) & 3;
1611 var->present = (s->attrib >> SVM_SELECTOR_P_SHIFT) & 1;
1612 var->avl = (s->attrib >> SVM_SELECTOR_AVL_SHIFT) & 1;
1613 var->l = (s->attrib >> SVM_SELECTOR_L_SHIFT) & 1;
1614 var->db = (s->attrib >> SVM_SELECTOR_DB_SHIFT) & 1;
1617 * AMD CPUs circa 2014 track the G bit for all segments except CS.
1618 * However, the SVM spec states that the G bit is not observed by the
1619 * CPU, and some VMware virtual CPUs drop the G bit for all segments.
1620 * So let's synthesize a legal G bit for all segments, this helps
1621 * running KVM nested. It also helps cross-vendor migration, because
1622 * Intel's vmentry has a check on the 'G' bit.
1624 var->g = s->limit > 0xfffff;
1627 * AMD's VMCB does not have an explicit unusable field, so emulate it
1628 * for cross vendor migration purposes by "not present"
1630 var->unusable = !var->present;
1635 * Work around a bug where the busy flag in the tr selector
1645 * The accessed bit must always be set in the segment
1646 * descriptor cache, although it can be cleared in the
1647 * descriptor, the cached bit always remains at 1. Since
1648 * Intel has a check on this, set it here to support
1649 * cross-vendor migration.
1656 * On AMD CPUs sometimes the DB bit in the segment
1657 * descriptor is left as 1, although the whole segment has
1658 * been made unusable. Clear it here to pass an Intel VMX
1659 * entry check when cross vendor migrating.
1663 /* This is symmetric with svm_set_segment() */
1664 var->dpl = to_svm(vcpu)->vmcb->save.cpl;
1669 static int svm_get_cpl(struct kvm_vcpu *vcpu)
1671 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1676 static void svm_get_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1678 struct vcpu_svm *svm = to_svm(vcpu);
1680 dt->size = svm->vmcb->save.idtr.limit;
1681 dt->address = svm->vmcb->save.idtr.base;
1684 static void svm_set_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1686 struct vcpu_svm *svm = to_svm(vcpu);
1688 svm->vmcb->save.idtr.limit = dt->size;
1689 svm->vmcb->save.idtr.base = dt->address ;
1690 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1693 static void svm_get_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1695 struct vcpu_svm *svm = to_svm(vcpu);
1697 dt->size = svm->vmcb->save.gdtr.limit;
1698 dt->address = svm->vmcb->save.gdtr.base;
1701 static void svm_set_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1703 struct vcpu_svm *svm = to_svm(vcpu);
1705 svm->vmcb->save.gdtr.limit = dt->size;
1706 svm->vmcb->save.gdtr.base = dt->address ;
1707 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1710 static void update_cr0_intercept(struct vcpu_svm *svm)
1716 * SEV-ES guests must always keep the CR intercepts cleared. CR
1717 * tracking is done using the CR write traps.
1719 if (sev_es_guest(svm->vcpu.kvm))
1722 gcr0 = svm->vcpu.arch.cr0;
1723 hcr0 = &svm->vmcb->save.cr0;
1724 *hcr0 = (*hcr0 & ~SVM_CR0_SELECTIVE_MASK)
1725 | (gcr0 & SVM_CR0_SELECTIVE_MASK);
1727 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1729 if (gcr0 == *hcr0) {
1730 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1731 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1733 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1734 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1738 void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1740 struct vcpu_svm *svm = to_svm(vcpu);
1742 #ifdef CONFIG_X86_64
1743 if (vcpu->arch.efer & EFER_LME && !vcpu->arch.guest_state_protected) {
1744 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
1745 vcpu->arch.efer |= EFER_LMA;
1746 svm->vmcb->save.efer |= EFER_LMA | EFER_LME;
1749 if (is_paging(vcpu) && !(cr0 & X86_CR0_PG)) {
1750 vcpu->arch.efer &= ~EFER_LMA;
1751 svm->vmcb->save.efer &= ~(EFER_LMA | EFER_LME);
1755 vcpu->arch.cr0 = cr0;
1758 cr0 |= X86_CR0_PG | X86_CR0_WP;
1761 * re-enable caching here because the QEMU bios
1762 * does not do it - this results in some delay at
1765 if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
1766 cr0 &= ~(X86_CR0_CD | X86_CR0_NW);
1767 svm->vmcb->save.cr0 = cr0;
1768 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1769 update_cr0_intercept(svm);
1772 static bool svm_is_valid_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1777 void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1779 unsigned long host_cr4_mce = cr4_read_shadow() & X86_CR4_MCE;
1780 unsigned long old_cr4 = vcpu->arch.cr4;
1782 if (npt_enabled && ((old_cr4 ^ cr4) & X86_CR4_PGE))
1783 svm_flush_tlb(vcpu);
1785 vcpu->arch.cr4 = cr4;
1788 cr4 |= host_cr4_mce;
1789 to_svm(vcpu)->vmcb->save.cr4 = cr4;
1790 vmcb_mark_dirty(to_svm(vcpu)->vmcb, VMCB_CR);
1792 if ((cr4 ^ old_cr4) & (X86_CR4_OSXSAVE | X86_CR4_PKE))
1793 kvm_update_cpuid_runtime(vcpu);
1796 static void svm_set_segment(struct kvm_vcpu *vcpu,
1797 struct kvm_segment *var, int seg)
1799 struct vcpu_svm *svm = to_svm(vcpu);
1800 struct vmcb_seg *s = svm_seg(vcpu, seg);
1802 s->base = var->base;
1803 s->limit = var->limit;
1804 s->selector = var->selector;
1805 s->attrib = (var->type & SVM_SELECTOR_TYPE_MASK);
1806 s->attrib |= (var->s & 1) << SVM_SELECTOR_S_SHIFT;
1807 s->attrib |= (var->dpl & 3) << SVM_SELECTOR_DPL_SHIFT;
1808 s->attrib |= ((var->present & 1) && !var->unusable) << SVM_SELECTOR_P_SHIFT;
1809 s->attrib |= (var->avl & 1) << SVM_SELECTOR_AVL_SHIFT;
1810 s->attrib |= (var->l & 1) << SVM_SELECTOR_L_SHIFT;
1811 s->attrib |= (var->db & 1) << SVM_SELECTOR_DB_SHIFT;
1812 s->attrib |= (var->g & 1) << SVM_SELECTOR_G_SHIFT;
1815 * This is always accurate, except if SYSRET returned to a segment
1816 * with SS.DPL != 3. Intel does not have this quirk, and always
1817 * forces SS.DPL to 3 on sysret, so we ignore that case; fixing it
1818 * would entail passing the CPL to userspace and back.
1820 if (seg == VCPU_SREG_SS)
1821 /* This is symmetric with svm_get_segment() */
1822 svm->vmcb->save.cpl = (var->dpl & 3);
1824 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
1827 static void svm_update_exception_bitmap(struct kvm_vcpu *vcpu)
1829 struct vcpu_svm *svm = to_svm(vcpu);
1831 clr_exception_intercept(svm, BP_VECTOR);
1833 if (vcpu->guest_debug & KVM_GUESTDBG_ENABLE) {
1834 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP)
1835 set_exception_intercept(svm, BP_VECTOR);
1839 static void new_asid(struct vcpu_svm *svm, struct svm_cpu_data *sd)
1841 if (sd->next_asid > sd->max_asid) {
1842 ++sd->asid_generation;
1843 sd->next_asid = sd->min_asid;
1844 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
1845 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1848 svm->asid_generation = sd->asid_generation;
1849 svm->asid = sd->next_asid++;
1852 static void svm_set_dr6(struct vcpu_svm *svm, unsigned long value)
1854 struct vmcb *vmcb = svm->vmcb;
1856 if (svm->vcpu.arch.guest_state_protected)
1859 if (unlikely(value != vmcb->save.dr6)) {
1860 vmcb->save.dr6 = value;
1861 vmcb_mark_dirty(vmcb, VMCB_DR);
1865 static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
1867 struct vcpu_svm *svm = to_svm(vcpu);
1869 if (vcpu->arch.guest_state_protected)
1872 get_debugreg(vcpu->arch.db[0], 0);
1873 get_debugreg(vcpu->arch.db[1], 1);
1874 get_debugreg(vcpu->arch.db[2], 2);
1875 get_debugreg(vcpu->arch.db[3], 3);
1877 * We cannot reset svm->vmcb->save.dr6 to DR6_ACTIVE_LOW here,
1878 * because db_interception might need it. We can do it before vmentry.
1880 vcpu->arch.dr6 = svm->vmcb->save.dr6;
1881 vcpu->arch.dr7 = svm->vmcb->save.dr7;
1882 vcpu->arch.switch_db_regs &= ~KVM_DEBUGREG_WONT_EXIT;
1883 set_dr_intercepts(svm);
1886 static void svm_set_dr7(struct kvm_vcpu *vcpu, unsigned long value)
1888 struct vcpu_svm *svm = to_svm(vcpu);
1890 if (vcpu->arch.guest_state_protected)
1893 svm->vmcb->save.dr7 = value;
1894 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
1897 static int pf_interception(struct vcpu_svm *svm)
1899 u64 fault_address = __sme_clr(svm->vmcb->control.exit_info_2);
1900 u64 error_code = svm->vmcb->control.exit_info_1;
1902 return kvm_handle_page_fault(&svm->vcpu, error_code, fault_address,
1903 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1904 svm->vmcb->control.insn_bytes : NULL,
1905 svm->vmcb->control.insn_len);
1908 static int npf_interception(struct vcpu_svm *svm)
1910 u64 fault_address = __sme_clr(svm->vmcb->control.exit_info_2);
1911 u64 error_code = svm->vmcb->control.exit_info_1;
1913 trace_kvm_page_fault(fault_address, error_code);
1914 return kvm_mmu_page_fault(&svm->vcpu, fault_address, error_code,
1915 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1916 svm->vmcb->control.insn_bytes : NULL,
1917 svm->vmcb->control.insn_len);
1920 static int db_interception(struct vcpu_svm *svm)
1922 struct kvm_run *kvm_run = svm->vcpu.run;
1923 struct kvm_vcpu *vcpu = &svm->vcpu;
1925 if (!(svm->vcpu.guest_debug &
1926 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) &&
1927 !svm->nmi_singlestep) {
1928 u32 payload = svm->vmcb->save.dr6 ^ DR6_ACTIVE_LOW;
1929 kvm_queue_exception_p(&svm->vcpu, DB_VECTOR, payload);
1933 if (svm->nmi_singlestep) {
1934 disable_nmi_singlestep(svm);
1935 /* Make sure we check for pending NMIs upon entry */
1936 kvm_make_request(KVM_REQ_EVENT, vcpu);
1939 if (svm->vcpu.guest_debug &
1940 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) {
1941 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1942 kvm_run->debug.arch.dr6 = svm->vmcb->save.dr6;
1943 kvm_run->debug.arch.dr7 = svm->vmcb->save.dr7;
1944 kvm_run->debug.arch.pc =
1945 svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1946 kvm_run->debug.arch.exception = DB_VECTOR;
1953 static int bp_interception(struct vcpu_svm *svm)
1955 struct kvm_run *kvm_run = svm->vcpu.run;
1957 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1958 kvm_run->debug.arch.pc = svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1959 kvm_run->debug.arch.exception = BP_VECTOR;
1963 static int ud_interception(struct vcpu_svm *svm)
1965 return handle_ud(&svm->vcpu);
1968 static int ac_interception(struct vcpu_svm *svm)
1970 kvm_queue_exception_e(&svm->vcpu, AC_VECTOR, 0);
1974 static bool is_erratum_383(void)
1979 if (!erratum_383_found)
1982 value = native_read_msr_safe(MSR_IA32_MC0_STATUS, &err);
1986 /* Bit 62 may or may not be set for this mce */
1987 value &= ~(1ULL << 62);
1989 if (value != 0xb600000000010015ULL)
1992 /* Clear MCi_STATUS registers */
1993 for (i = 0; i < 6; ++i)
1994 native_write_msr_safe(MSR_IA32_MCx_STATUS(i), 0, 0);
1996 value = native_read_msr_safe(MSR_IA32_MCG_STATUS, &err);
2000 value &= ~(1ULL << 2);
2001 low = lower_32_bits(value);
2002 high = upper_32_bits(value);
2004 native_write_msr_safe(MSR_IA32_MCG_STATUS, low, high);
2007 /* Flush tlb to evict multi-match entries */
2013 static void svm_handle_mce(struct vcpu_svm *svm)
2015 if (is_erratum_383()) {
2017 * Erratum 383 triggered. Guest state is corrupt so kill the
2020 pr_err("KVM: Guest triggered AMD Erratum 383\n");
2022 kvm_make_request(KVM_REQ_TRIPLE_FAULT, &svm->vcpu);
2028 * On an #MC intercept the MCE handler is not called automatically in
2029 * the host. So do it by hand here.
2031 kvm_machine_check();
2034 static int mc_interception(struct vcpu_svm *svm)
2039 static int shutdown_interception(struct vcpu_svm *svm)
2041 struct kvm_run *kvm_run = svm->vcpu.run;
2044 * The VM save area has already been encrypted so it
2045 * cannot be reinitialized - just terminate.
2047 if (sev_es_guest(svm->vcpu.kvm))
2051 * VMCB is undefined after a SHUTDOWN intercept
2052 * so reinitialize it.
2054 clear_page(svm->vmcb);
2057 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
2061 static int io_interception(struct vcpu_svm *svm)
2063 struct kvm_vcpu *vcpu = &svm->vcpu;
2064 u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
2065 int size, in, string;
2068 ++svm->vcpu.stat.io_exits;
2069 string = (io_info & SVM_IOIO_STR_MASK) != 0;
2070 in = (io_info & SVM_IOIO_TYPE_MASK) != 0;
2071 port = io_info >> 16;
2072 size = (io_info & SVM_IOIO_SIZE_MASK) >> SVM_IOIO_SIZE_SHIFT;
2075 if (sev_es_guest(vcpu->kvm))
2076 return sev_es_string_io(svm, size, port, in);
2078 return kvm_emulate_instruction(vcpu, 0);
2081 svm->next_rip = svm->vmcb->control.exit_info_2;
2083 return kvm_fast_pio(&svm->vcpu, size, port, in);
2086 static int nmi_interception(struct vcpu_svm *svm)
2091 static int intr_interception(struct vcpu_svm *svm)
2093 ++svm->vcpu.stat.irq_exits;
2097 static int nop_on_interception(struct vcpu_svm *svm)
2102 static int halt_interception(struct vcpu_svm *svm)
2104 return kvm_emulate_halt(&svm->vcpu);
2107 static int vmmcall_interception(struct vcpu_svm *svm)
2109 return kvm_emulate_hypercall(&svm->vcpu);
2112 static int vmload_interception(struct vcpu_svm *svm)
2114 struct vmcb *nested_vmcb;
2115 struct kvm_host_map map;
2118 if (nested_svm_check_permissions(svm))
2121 ret = kvm_vcpu_map(&svm->vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2124 kvm_inject_gp(&svm->vcpu, 0);
2128 nested_vmcb = map.hva;
2130 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2132 nested_svm_vmloadsave(nested_vmcb, svm->vmcb);
2133 kvm_vcpu_unmap(&svm->vcpu, &map, true);
2138 static int vmsave_interception(struct vcpu_svm *svm)
2140 struct vmcb *nested_vmcb;
2141 struct kvm_host_map map;
2144 if (nested_svm_check_permissions(svm))
2147 ret = kvm_vcpu_map(&svm->vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2150 kvm_inject_gp(&svm->vcpu, 0);
2154 nested_vmcb = map.hva;
2156 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2158 nested_svm_vmloadsave(svm->vmcb, nested_vmcb);
2159 kvm_vcpu_unmap(&svm->vcpu, &map, true);
2164 static int vmrun_interception(struct vcpu_svm *svm)
2166 if (nested_svm_check_permissions(svm))
2169 return nested_svm_vmrun(svm);
2179 /* Return NONE_SVM_INSTR if not SVM instrs, otherwise return decode result */
2180 static int svm_instr_opcode(struct kvm_vcpu *vcpu)
2182 struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt;
2184 if (ctxt->b != 0x1 || ctxt->opcode_len != 2)
2185 return NONE_SVM_INSTR;
2187 switch (ctxt->modrm) {
2188 case 0xd8: /* VMRUN */
2189 return SVM_INSTR_VMRUN;
2190 case 0xda: /* VMLOAD */
2191 return SVM_INSTR_VMLOAD;
2192 case 0xdb: /* VMSAVE */
2193 return SVM_INSTR_VMSAVE;
2198 return NONE_SVM_INSTR;
2201 static int emulate_svm_instr(struct kvm_vcpu *vcpu, int opcode)
2203 const int guest_mode_exit_codes[] = {
2204 [SVM_INSTR_VMRUN] = SVM_EXIT_VMRUN,
2205 [SVM_INSTR_VMLOAD] = SVM_EXIT_VMLOAD,
2206 [SVM_INSTR_VMSAVE] = SVM_EXIT_VMSAVE,
2208 int (*const svm_instr_handlers[])(struct vcpu_svm *svm) = {
2209 [SVM_INSTR_VMRUN] = vmrun_interception,
2210 [SVM_INSTR_VMLOAD] = vmload_interception,
2211 [SVM_INSTR_VMSAVE] = vmsave_interception,
2213 struct vcpu_svm *svm = to_svm(vcpu);
2216 if (is_guest_mode(vcpu)) {
2217 svm->vmcb->control.exit_code = guest_mode_exit_codes[opcode];
2218 svm->vmcb->control.exit_info_1 = 0;
2219 svm->vmcb->control.exit_info_2 = 0;
2221 /* Returns '1' or -errno on failure, '0' on success. */
2222 ret = nested_svm_vmexit(svm);
2227 return svm_instr_handlers[opcode](svm);
2231 * #GP handling code. Note that #GP can be triggered under the following two
2233 * 1) SVM VM-related instructions (VMRUN/VMSAVE/VMLOAD) that trigger #GP on
2234 * some AMD CPUs when EAX of these instructions are in the reserved memory
2235 * regions (e.g. SMM memory on host).
2236 * 2) VMware backdoor
2238 static int gp_interception(struct vcpu_svm *svm)
2240 struct kvm_vcpu *vcpu = &svm->vcpu;
2241 u32 error_code = svm->vmcb->control.exit_info_1;
2244 /* Both #GP cases have zero error_code */
2248 /* Decode the instruction for usage later */
2249 if (x86_decode_emulated_instruction(vcpu, 0, NULL, 0) != EMULATION_OK)
2252 opcode = svm_instr_opcode(vcpu);
2254 if (opcode == NONE_SVM_INSTR) {
2255 if (!enable_vmware_backdoor)
2259 * VMware backdoor emulation on #GP interception only handles
2260 * IN{S}, OUT{S}, and RDPMC.
2262 if (!is_guest_mode(vcpu))
2263 return kvm_emulate_instruction(vcpu,
2264 EMULTYPE_VMWARE_GP | EMULTYPE_NO_DECODE);
2266 return emulate_svm_instr(vcpu, opcode);
2269 kvm_queue_exception_e(vcpu, GP_VECTOR, error_code);
2273 void svm_set_gif(struct vcpu_svm *svm, bool value)
2277 * If VGIF is enabled, the STGI intercept is only added to
2278 * detect the opening of the SMI/NMI window; remove it now.
2279 * Likewise, clear the VINTR intercept, we will set it
2280 * again while processing KVM_REQ_EVENT if needed.
2282 if (vgif_enabled(svm))
2283 svm_clr_intercept(svm, INTERCEPT_STGI);
2284 if (svm_is_intercept(svm, INTERCEPT_VINTR))
2285 svm_clear_vintr(svm);
2288 if (svm->vcpu.arch.smi_pending ||
2289 svm->vcpu.arch.nmi_pending ||
2290 kvm_cpu_has_injectable_intr(&svm->vcpu))
2291 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2296 * After a CLGI no interrupts should come. But if vGIF is
2297 * in use, we still rely on the VINTR intercept (rather than
2298 * STGI) to detect an open interrupt window.
2300 if (!vgif_enabled(svm))
2301 svm_clear_vintr(svm);
2305 static int stgi_interception(struct vcpu_svm *svm)
2309 if (nested_svm_check_permissions(svm))
2312 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2313 svm_set_gif(svm, true);
2317 static int clgi_interception(struct vcpu_svm *svm)
2321 if (nested_svm_check_permissions(svm))
2324 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2325 svm_set_gif(svm, false);
2329 static int invlpga_interception(struct vcpu_svm *svm)
2331 struct kvm_vcpu *vcpu = &svm->vcpu;
2333 trace_kvm_invlpga(svm->vmcb->save.rip, kvm_rcx_read(&svm->vcpu),
2334 kvm_rax_read(&svm->vcpu));
2336 /* Let's treat INVLPGA the same as INVLPG (can be optimized!) */
2337 kvm_mmu_invlpg(vcpu, kvm_rax_read(&svm->vcpu));
2339 return kvm_skip_emulated_instruction(&svm->vcpu);
2342 static int skinit_interception(struct vcpu_svm *svm)
2344 trace_kvm_skinit(svm->vmcb->save.rip, kvm_rax_read(&svm->vcpu));
2346 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2350 static int wbinvd_interception(struct vcpu_svm *svm)
2352 return kvm_emulate_wbinvd(&svm->vcpu);
2355 static int xsetbv_interception(struct vcpu_svm *svm)
2357 u64 new_bv = kvm_read_edx_eax(&svm->vcpu);
2358 u32 index = kvm_rcx_read(&svm->vcpu);
2360 int err = kvm_set_xcr(&svm->vcpu, index, new_bv);
2361 return kvm_complete_insn_gp(&svm->vcpu, err);
2364 static int rdpru_interception(struct vcpu_svm *svm)
2366 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2370 static int task_switch_interception(struct vcpu_svm *svm)
2374 int int_type = svm->vmcb->control.exit_int_info &
2375 SVM_EXITINTINFO_TYPE_MASK;
2376 int int_vec = svm->vmcb->control.exit_int_info & SVM_EVTINJ_VEC_MASK;
2378 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_TYPE_MASK;
2380 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID;
2381 bool has_error_code = false;
2384 tss_selector = (u16)svm->vmcb->control.exit_info_1;
2386 if (svm->vmcb->control.exit_info_2 &
2387 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
2388 reason = TASK_SWITCH_IRET;
2389 else if (svm->vmcb->control.exit_info_2 &
2390 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
2391 reason = TASK_SWITCH_JMP;
2393 reason = TASK_SWITCH_GATE;
2395 reason = TASK_SWITCH_CALL;
2397 if (reason == TASK_SWITCH_GATE) {
2399 case SVM_EXITINTINFO_TYPE_NMI:
2400 svm->vcpu.arch.nmi_injected = false;
2402 case SVM_EXITINTINFO_TYPE_EXEPT:
2403 if (svm->vmcb->control.exit_info_2 &
2404 (1ULL << SVM_EXITINFOSHIFT_TS_HAS_ERROR_CODE)) {
2405 has_error_code = true;
2407 (u32)svm->vmcb->control.exit_info_2;
2409 kvm_clear_exception_queue(&svm->vcpu);
2411 case SVM_EXITINTINFO_TYPE_INTR:
2412 kvm_clear_interrupt_queue(&svm->vcpu);
2419 if (reason != TASK_SWITCH_GATE ||
2420 int_type == SVM_EXITINTINFO_TYPE_SOFT ||
2421 (int_type == SVM_EXITINTINFO_TYPE_EXEPT &&
2422 (int_vec == OF_VECTOR || int_vec == BP_VECTOR))) {
2423 if (!skip_emulated_instruction(&svm->vcpu))
2427 if (int_type != SVM_EXITINTINFO_TYPE_SOFT)
2430 return kvm_task_switch(&svm->vcpu, tss_selector, int_vec, reason,
2431 has_error_code, error_code);
2434 static int cpuid_interception(struct vcpu_svm *svm)
2436 return kvm_emulate_cpuid(&svm->vcpu);
2439 static int iret_interception(struct vcpu_svm *svm)
2441 ++svm->vcpu.stat.nmi_window_exits;
2442 svm->vcpu.arch.hflags |= HF_IRET_MASK;
2443 if (!sev_es_guest(svm->vcpu.kvm)) {
2444 svm_clr_intercept(svm, INTERCEPT_IRET);
2445 svm->nmi_iret_rip = kvm_rip_read(&svm->vcpu);
2447 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2451 static int invd_interception(struct vcpu_svm *svm)
2453 /* Treat an INVD instruction as a NOP and just skip it. */
2454 return kvm_skip_emulated_instruction(&svm->vcpu);
2457 static int invlpg_interception(struct vcpu_svm *svm)
2459 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2460 return kvm_emulate_instruction(&svm->vcpu, 0);
2462 kvm_mmu_invlpg(&svm->vcpu, svm->vmcb->control.exit_info_1);
2463 return kvm_skip_emulated_instruction(&svm->vcpu);
2466 static int emulate_on_interception(struct vcpu_svm *svm)
2468 return kvm_emulate_instruction(&svm->vcpu, 0);
2471 static int rsm_interception(struct vcpu_svm *svm)
2473 return kvm_emulate_instruction_from_buffer(&svm->vcpu, rsm_ins_bytes, 2);
2476 static int rdpmc_interception(struct vcpu_svm *svm)
2481 return emulate_on_interception(svm);
2483 err = kvm_rdpmc(&svm->vcpu);
2484 return kvm_complete_insn_gp(&svm->vcpu, err);
2487 static bool check_selective_cr0_intercepted(struct vcpu_svm *svm,
2490 unsigned long cr0 = svm->vcpu.arch.cr0;
2493 if (!is_guest_mode(&svm->vcpu) ||
2494 (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_SELECTIVE_CR0))))
2497 cr0 &= ~SVM_CR0_SELECTIVE_MASK;
2498 val &= ~SVM_CR0_SELECTIVE_MASK;
2501 svm->vmcb->control.exit_code = SVM_EXIT_CR0_SEL_WRITE;
2502 ret = (nested_svm_exit_handled(svm) == NESTED_EXIT_DONE);
2508 #define CR_VALID (1ULL << 63)
2510 static int cr_interception(struct vcpu_svm *svm)
2516 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2517 return emulate_on_interception(svm);
2519 if (unlikely((svm->vmcb->control.exit_info_1 & CR_VALID) == 0))
2520 return emulate_on_interception(svm);
2522 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2523 if (svm->vmcb->control.exit_code == SVM_EXIT_CR0_SEL_WRITE)
2524 cr = SVM_EXIT_WRITE_CR0 - SVM_EXIT_READ_CR0;
2526 cr = svm->vmcb->control.exit_code - SVM_EXIT_READ_CR0;
2529 if (cr >= 16) { /* mov to cr */
2531 val = kvm_register_read(&svm->vcpu, reg);
2532 trace_kvm_cr_write(cr, val);
2535 if (!check_selective_cr0_intercepted(svm, val))
2536 err = kvm_set_cr0(&svm->vcpu, val);
2542 err = kvm_set_cr3(&svm->vcpu, val);
2545 err = kvm_set_cr4(&svm->vcpu, val);
2548 err = kvm_set_cr8(&svm->vcpu, val);
2551 WARN(1, "unhandled write to CR%d", cr);
2552 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2555 } else { /* mov from cr */
2558 val = kvm_read_cr0(&svm->vcpu);
2561 val = svm->vcpu.arch.cr2;
2564 val = kvm_read_cr3(&svm->vcpu);
2567 val = kvm_read_cr4(&svm->vcpu);
2570 val = kvm_get_cr8(&svm->vcpu);
2573 WARN(1, "unhandled read from CR%d", cr);
2574 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2577 kvm_register_write(&svm->vcpu, reg, val);
2578 trace_kvm_cr_read(cr, val);
2580 return kvm_complete_insn_gp(&svm->vcpu, err);
2583 static int cr_trap(struct vcpu_svm *svm)
2585 struct kvm_vcpu *vcpu = &svm->vcpu;
2586 unsigned long old_value, new_value;
2590 new_value = (unsigned long)svm->vmcb->control.exit_info_1;
2592 cr = svm->vmcb->control.exit_code - SVM_EXIT_CR0_WRITE_TRAP;
2595 old_value = kvm_read_cr0(vcpu);
2596 svm_set_cr0(vcpu, new_value);
2598 kvm_post_set_cr0(vcpu, old_value, new_value);
2601 old_value = kvm_read_cr4(vcpu);
2602 svm_set_cr4(vcpu, new_value);
2604 kvm_post_set_cr4(vcpu, old_value, new_value);
2607 ret = kvm_set_cr8(&svm->vcpu, new_value);
2610 WARN(1, "unhandled CR%d write trap", cr);
2611 kvm_queue_exception(vcpu, UD_VECTOR);
2615 return kvm_complete_insn_gp(vcpu, ret);
2618 static int dr_interception(struct vcpu_svm *svm)
2624 if (svm->vcpu.guest_debug == 0) {
2626 * No more DR vmexits; force a reload of the debug registers
2627 * and reenter on this instruction. The next vmexit will
2628 * retrieve the full state of the debug registers.
2630 clr_dr_intercepts(svm);
2631 svm->vcpu.arch.switch_db_regs |= KVM_DEBUGREG_WONT_EXIT;
2635 if (!boot_cpu_has(X86_FEATURE_DECODEASSISTS))
2636 return emulate_on_interception(svm);
2638 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2639 dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
2640 if (dr >= 16) { /* mov to DRn */
2642 val = kvm_register_read(&svm->vcpu, reg);
2643 err = kvm_set_dr(&svm->vcpu, dr, val);
2645 kvm_get_dr(&svm->vcpu, dr, &val);
2646 kvm_register_write(&svm->vcpu, reg, val);
2649 return kvm_complete_insn_gp(&svm->vcpu, err);
2652 static int cr8_write_interception(struct vcpu_svm *svm)
2654 struct kvm_run *kvm_run = svm->vcpu.run;
2657 u8 cr8_prev = kvm_get_cr8(&svm->vcpu);
2658 /* instruction emulation calls kvm_set_cr8() */
2659 r = cr_interception(svm);
2660 if (lapic_in_kernel(&svm->vcpu))
2662 if (cr8_prev <= kvm_get_cr8(&svm->vcpu))
2664 kvm_run->exit_reason = KVM_EXIT_SET_TPR;
2668 static int efer_trap(struct vcpu_svm *svm)
2670 struct msr_data msr_info;
2674 * Clear the EFER_SVME bit from EFER. The SVM code always sets this
2675 * bit in svm_set_efer(), but __kvm_valid_efer() checks it against
2676 * whether the guest has X86_FEATURE_SVM - this avoids a failure if
2677 * the guest doesn't have X86_FEATURE_SVM.
2679 msr_info.host_initiated = false;
2680 msr_info.index = MSR_EFER;
2681 msr_info.data = svm->vmcb->control.exit_info_1 & ~EFER_SVME;
2682 ret = kvm_set_msr_common(&svm->vcpu, &msr_info);
2684 return kvm_complete_insn_gp(&svm->vcpu, ret);
2687 static int svm_get_msr_feature(struct kvm_msr_entry *msr)
2691 switch (msr->index) {
2692 case MSR_F10H_DECFG:
2693 if (boot_cpu_has(X86_FEATURE_LFENCE_RDTSC))
2694 msr->data |= MSR_F10H_DECFG_LFENCE_SERIALIZE;
2696 case MSR_IA32_PERF_CAPABILITIES:
2699 return KVM_MSR_RET_INVALID;
2705 static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
2707 struct vcpu_svm *svm = to_svm(vcpu);
2709 switch (msr_info->index) {
2711 msr_info->data = svm->vmcb->save.star;
2713 #ifdef CONFIG_X86_64
2715 msr_info->data = svm->vmcb->save.lstar;
2718 msr_info->data = svm->vmcb->save.cstar;
2720 case MSR_KERNEL_GS_BASE:
2721 msr_info->data = svm->vmcb->save.kernel_gs_base;
2723 case MSR_SYSCALL_MASK:
2724 msr_info->data = svm->vmcb->save.sfmask;
2727 case MSR_IA32_SYSENTER_CS:
2728 msr_info->data = svm->vmcb->save.sysenter_cs;
2730 case MSR_IA32_SYSENTER_EIP:
2731 msr_info->data = svm->sysenter_eip;
2733 case MSR_IA32_SYSENTER_ESP:
2734 msr_info->data = svm->sysenter_esp;
2737 if (!boot_cpu_has(X86_FEATURE_RDTSCP))
2739 msr_info->data = svm->tsc_aux;
2742 * Nobody will change the following 5 values in the VMCB so we can
2743 * safely return them on rdmsr. They will always be 0 until LBRV is
2746 case MSR_IA32_DEBUGCTLMSR:
2747 msr_info->data = svm->vmcb->save.dbgctl;
2749 case MSR_IA32_LASTBRANCHFROMIP:
2750 msr_info->data = svm->vmcb->save.br_from;
2752 case MSR_IA32_LASTBRANCHTOIP:
2753 msr_info->data = svm->vmcb->save.br_to;
2755 case MSR_IA32_LASTINTFROMIP:
2756 msr_info->data = svm->vmcb->save.last_excp_from;
2758 case MSR_IA32_LASTINTTOIP:
2759 msr_info->data = svm->vmcb->save.last_excp_to;
2761 case MSR_VM_HSAVE_PA:
2762 msr_info->data = svm->nested.hsave_msr;
2765 msr_info->data = svm->nested.vm_cr_msr;
2767 case MSR_IA32_SPEC_CTRL:
2768 if (!msr_info->host_initiated &&
2769 !guest_has_spec_ctrl_msr(vcpu))
2772 msr_info->data = svm->spec_ctrl;
2774 case MSR_AMD64_VIRT_SPEC_CTRL:
2775 if (!msr_info->host_initiated &&
2776 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2779 msr_info->data = svm->virt_spec_ctrl;
2781 case MSR_F15H_IC_CFG: {
2785 family = guest_cpuid_family(vcpu);
2786 model = guest_cpuid_model(vcpu);
2788 if (family < 0 || model < 0)
2789 return kvm_get_msr_common(vcpu, msr_info);
2793 if (family == 0x15 &&
2794 (model >= 0x2 && model < 0x20))
2795 msr_info->data = 0x1E;
2798 case MSR_F10H_DECFG:
2799 msr_info->data = svm->msr_decfg;
2802 return kvm_get_msr_common(vcpu, msr_info);
2807 static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
2809 struct vcpu_svm *svm = to_svm(vcpu);
2810 if (!sev_es_guest(svm->vcpu.kvm) || !err)
2811 return kvm_complete_insn_gp(&svm->vcpu, err);
2813 ghcb_set_sw_exit_info_1(svm->ghcb, 1);
2814 ghcb_set_sw_exit_info_2(svm->ghcb,
2816 SVM_EVTINJ_TYPE_EXEPT |
2821 static int rdmsr_interception(struct vcpu_svm *svm)
2823 return kvm_emulate_rdmsr(&svm->vcpu);
2826 static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
2828 struct vcpu_svm *svm = to_svm(vcpu);
2829 int svm_dis, chg_mask;
2831 if (data & ~SVM_VM_CR_VALID_MASK)
2834 chg_mask = SVM_VM_CR_VALID_MASK;
2836 if (svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK)
2837 chg_mask &= ~(SVM_VM_CR_SVM_LOCK_MASK | SVM_VM_CR_SVM_DIS_MASK);
2839 svm->nested.vm_cr_msr &= ~chg_mask;
2840 svm->nested.vm_cr_msr |= (data & chg_mask);
2842 svm_dis = svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK;
2844 /* check for svm_disable while efer.svme is set */
2845 if (svm_dis && (vcpu->arch.efer & EFER_SVME))
2851 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
2853 struct vcpu_svm *svm = to_svm(vcpu);
2855 u32 ecx = msr->index;
2856 u64 data = msr->data;
2858 case MSR_IA32_CR_PAT:
2859 if (!kvm_mtrr_valid(vcpu, MSR_IA32_CR_PAT, data))
2861 vcpu->arch.pat = data;
2862 svm->vmcb->save.g_pat = data;
2863 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
2865 case MSR_IA32_SPEC_CTRL:
2866 if (!msr->host_initiated &&
2867 !guest_has_spec_ctrl_msr(vcpu))
2870 if (kvm_spec_ctrl_test_value(data))
2873 svm->spec_ctrl = data;
2879 * When it's written (to non-zero) for the first time, pass
2883 * The handling of the MSR bitmap for L2 guests is done in
2884 * nested_svm_vmrun_msrpm.
2885 * We update the L1 MSR bit as well since it will end up
2886 * touching the MSR anyway now.
2888 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
2890 case MSR_IA32_PRED_CMD:
2891 if (!msr->host_initiated &&
2892 !guest_has_pred_cmd_msr(vcpu))
2895 if (data & ~PRED_CMD_IBPB)
2897 if (!boot_cpu_has(X86_FEATURE_IBPB))
2902 wrmsrl(MSR_IA32_PRED_CMD, PRED_CMD_IBPB);
2903 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_PRED_CMD, 0, 1);
2905 case MSR_AMD64_VIRT_SPEC_CTRL:
2906 if (!msr->host_initiated &&
2907 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2910 if (data & ~SPEC_CTRL_SSBD)
2913 svm->virt_spec_ctrl = data;
2916 svm->vmcb->save.star = data;
2918 #ifdef CONFIG_X86_64
2920 svm->vmcb->save.lstar = data;
2923 svm->vmcb->save.cstar = data;
2925 case MSR_KERNEL_GS_BASE:
2926 svm->vmcb->save.kernel_gs_base = data;
2928 case MSR_SYSCALL_MASK:
2929 svm->vmcb->save.sfmask = data;
2932 case MSR_IA32_SYSENTER_CS:
2933 svm->vmcb->save.sysenter_cs = data;
2935 case MSR_IA32_SYSENTER_EIP:
2936 svm->sysenter_eip = data;
2937 svm->vmcb->save.sysenter_eip = data;
2939 case MSR_IA32_SYSENTER_ESP:
2940 svm->sysenter_esp = data;
2941 svm->vmcb->save.sysenter_esp = data;
2944 if (!boot_cpu_has(X86_FEATURE_RDTSCP))
2948 * This is rare, so we update the MSR here instead of using
2949 * direct_access_msrs. Doing that would require a rdmsr in
2952 svm->tsc_aux = data;
2953 wrmsrl(MSR_TSC_AUX, svm->tsc_aux);
2955 case MSR_IA32_DEBUGCTLMSR:
2956 if (!boot_cpu_has(X86_FEATURE_LBRV)) {
2957 vcpu_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTL 0x%llx, nop\n",
2961 if (data & DEBUGCTL_RESERVED_BITS)
2964 svm->vmcb->save.dbgctl = data;
2965 vmcb_mark_dirty(svm->vmcb, VMCB_LBR);
2966 if (data & (1ULL<<0))
2967 svm_enable_lbrv(vcpu);
2969 svm_disable_lbrv(vcpu);
2971 case MSR_VM_HSAVE_PA:
2972 svm->nested.hsave_msr = data;
2975 return svm_set_vm_cr(vcpu, data);
2977 vcpu_unimpl(vcpu, "unimplemented wrmsr: 0x%x data 0x%llx\n", ecx, data);
2979 case MSR_F10H_DECFG: {
2980 struct kvm_msr_entry msr_entry;
2982 msr_entry.index = msr->index;
2983 if (svm_get_msr_feature(&msr_entry))
2986 /* Check the supported bits */
2987 if (data & ~msr_entry.data)
2990 /* Don't allow the guest to change a bit, #GP */
2991 if (!msr->host_initiated && (data ^ msr_entry.data))
2994 svm->msr_decfg = data;
2997 case MSR_IA32_APICBASE:
2998 if (kvm_vcpu_apicv_active(vcpu))
2999 avic_update_vapic_bar(to_svm(vcpu), data);
3002 return kvm_set_msr_common(vcpu, msr);
3007 static int wrmsr_interception(struct vcpu_svm *svm)
3009 return kvm_emulate_wrmsr(&svm->vcpu);
3012 static int msr_interception(struct vcpu_svm *svm)
3014 if (svm->vmcb->control.exit_info_1)
3015 return wrmsr_interception(svm);
3017 return rdmsr_interception(svm);
3020 static int interrupt_window_interception(struct vcpu_svm *svm)
3022 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
3023 svm_clear_vintr(svm);
3026 * For AVIC, the only reason to end up here is ExtINTs.
3027 * In this case AVIC was temporarily disabled for
3028 * requesting the IRQ window and we have to re-enable it.
3030 svm_toggle_avic_for_irq_window(&svm->vcpu, true);
3032 ++svm->vcpu.stat.irq_window_exits;
3036 static int pause_interception(struct vcpu_svm *svm)
3038 struct kvm_vcpu *vcpu = &svm->vcpu;
3042 * CPL is not made available for an SEV-ES guest, therefore
3043 * vcpu->arch.preempted_in_kernel can never be true. Just
3044 * set in_kernel to false as well.
3046 in_kernel = !sev_es_guest(svm->vcpu.kvm) && svm_get_cpl(vcpu) == 0;
3048 if (!kvm_pause_in_guest(vcpu->kvm))
3049 grow_ple_window(vcpu);
3051 kvm_vcpu_on_spin(vcpu, in_kernel);
3055 static int nop_interception(struct vcpu_svm *svm)
3057 return kvm_skip_emulated_instruction(&(svm->vcpu));
3060 static int monitor_interception(struct vcpu_svm *svm)
3062 printk_once(KERN_WARNING "kvm: MONITOR instruction emulated as NOP!\n");
3063 return nop_interception(svm);
3066 static int mwait_interception(struct vcpu_svm *svm)
3068 printk_once(KERN_WARNING "kvm: MWAIT instruction emulated as NOP!\n");
3069 return nop_interception(svm);
3072 static int invpcid_interception(struct vcpu_svm *svm)
3074 struct kvm_vcpu *vcpu = &svm->vcpu;
3078 if (!guest_cpuid_has(vcpu, X86_FEATURE_INVPCID)) {
3079 kvm_queue_exception(vcpu, UD_VECTOR);
3084 * For an INVPCID intercept:
3085 * EXITINFO1 provides the linear address of the memory operand.
3086 * EXITINFO2 provides the contents of the register operand.
3088 type = svm->vmcb->control.exit_info_2;
3089 gva = svm->vmcb->control.exit_info_1;
3092 kvm_inject_gp(vcpu, 0);
3096 return kvm_handle_invpcid(vcpu, type, gva);
3099 static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = {
3100 [SVM_EXIT_READ_CR0] = cr_interception,
3101 [SVM_EXIT_READ_CR3] = cr_interception,
3102 [SVM_EXIT_READ_CR4] = cr_interception,
3103 [SVM_EXIT_READ_CR8] = cr_interception,
3104 [SVM_EXIT_CR0_SEL_WRITE] = cr_interception,
3105 [SVM_EXIT_WRITE_CR0] = cr_interception,
3106 [SVM_EXIT_WRITE_CR3] = cr_interception,
3107 [SVM_EXIT_WRITE_CR4] = cr_interception,
3108 [SVM_EXIT_WRITE_CR8] = cr8_write_interception,
3109 [SVM_EXIT_READ_DR0] = dr_interception,
3110 [SVM_EXIT_READ_DR1] = dr_interception,
3111 [SVM_EXIT_READ_DR2] = dr_interception,
3112 [SVM_EXIT_READ_DR3] = dr_interception,
3113 [SVM_EXIT_READ_DR4] = dr_interception,
3114 [SVM_EXIT_READ_DR5] = dr_interception,
3115 [SVM_EXIT_READ_DR6] = dr_interception,
3116 [SVM_EXIT_READ_DR7] = dr_interception,
3117 [SVM_EXIT_WRITE_DR0] = dr_interception,
3118 [SVM_EXIT_WRITE_DR1] = dr_interception,
3119 [SVM_EXIT_WRITE_DR2] = dr_interception,
3120 [SVM_EXIT_WRITE_DR3] = dr_interception,
3121 [SVM_EXIT_WRITE_DR4] = dr_interception,
3122 [SVM_EXIT_WRITE_DR5] = dr_interception,
3123 [SVM_EXIT_WRITE_DR6] = dr_interception,
3124 [SVM_EXIT_WRITE_DR7] = dr_interception,
3125 [SVM_EXIT_EXCP_BASE + DB_VECTOR] = db_interception,
3126 [SVM_EXIT_EXCP_BASE + BP_VECTOR] = bp_interception,
3127 [SVM_EXIT_EXCP_BASE + UD_VECTOR] = ud_interception,
3128 [SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
3129 [SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
3130 [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
3131 [SVM_EXIT_EXCP_BASE + GP_VECTOR] = gp_interception,
3132 [SVM_EXIT_INTR] = intr_interception,
3133 [SVM_EXIT_NMI] = nmi_interception,
3134 [SVM_EXIT_SMI] = nop_on_interception,
3135 [SVM_EXIT_INIT] = nop_on_interception,
3136 [SVM_EXIT_VINTR] = interrupt_window_interception,
3137 [SVM_EXIT_RDPMC] = rdpmc_interception,
3138 [SVM_EXIT_CPUID] = cpuid_interception,
3139 [SVM_EXIT_IRET] = iret_interception,
3140 [SVM_EXIT_INVD] = invd_interception,
3141 [SVM_EXIT_PAUSE] = pause_interception,
3142 [SVM_EXIT_HLT] = halt_interception,
3143 [SVM_EXIT_INVLPG] = invlpg_interception,
3144 [SVM_EXIT_INVLPGA] = invlpga_interception,
3145 [SVM_EXIT_IOIO] = io_interception,
3146 [SVM_EXIT_MSR] = msr_interception,
3147 [SVM_EXIT_TASK_SWITCH] = task_switch_interception,
3148 [SVM_EXIT_SHUTDOWN] = shutdown_interception,
3149 [SVM_EXIT_VMRUN] = vmrun_interception,
3150 [SVM_EXIT_VMMCALL] = vmmcall_interception,
3151 [SVM_EXIT_VMLOAD] = vmload_interception,
3152 [SVM_EXIT_VMSAVE] = vmsave_interception,
3153 [SVM_EXIT_STGI] = stgi_interception,
3154 [SVM_EXIT_CLGI] = clgi_interception,
3155 [SVM_EXIT_SKINIT] = skinit_interception,
3156 [SVM_EXIT_WBINVD] = wbinvd_interception,
3157 [SVM_EXIT_MONITOR] = monitor_interception,
3158 [SVM_EXIT_MWAIT] = mwait_interception,
3159 [SVM_EXIT_XSETBV] = xsetbv_interception,
3160 [SVM_EXIT_RDPRU] = rdpru_interception,
3161 [SVM_EXIT_EFER_WRITE_TRAP] = efer_trap,
3162 [SVM_EXIT_CR0_WRITE_TRAP] = cr_trap,
3163 [SVM_EXIT_CR4_WRITE_TRAP] = cr_trap,
3164 [SVM_EXIT_CR8_WRITE_TRAP] = cr_trap,
3165 [SVM_EXIT_INVPCID] = invpcid_interception,
3166 [SVM_EXIT_NPF] = npf_interception,
3167 [SVM_EXIT_RSM] = rsm_interception,
3168 [SVM_EXIT_AVIC_INCOMPLETE_IPI] = avic_incomplete_ipi_interception,
3169 [SVM_EXIT_AVIC_UNACCELERATED_ACCESS] = avic_unaccelerated_access_interception,
3170 [SVM_EXIT_VMGEXIT] = sev_handle_vmgexit,
3173 static void dump_vmcb(struct kvm_vcpu *vcpu)
3175 struct vcpu_svm *svm = to_svm(vcpu);
3176 struct vmcb_control_area *control = &svm->vmcb->control;
3177 struct vmcb_save_area *save = &svm->vmcb->save;
3179 if (!dump_invalid_vmcb) {
3180 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3184 pr_err("VMCB Control Area:\n");
3185 pr_err("%-20s%04x\n", "cr_read:", control->intercepts[INTERCEPT_CR] & 0xffff);
3186 pr_err("%-20s%04x\n", "cr_write:", control->intercepts[INTERCEPT_CR] >> 16);
3187 pr_err("%-20s%04x\n", "dr_read:", control->intercepts[INTERCEPT_DR] & 0xffff);
3188 pr_err("%-20s%04x\n", "dr_write:", control->intercepts[INTERCEPT_DR] >> 16);
3189 pr_err("%-20s%08x\n", "exceptions:", control->intercepts[INTERCEPT_EXCEPTION]);
3190 pr_err("%-20s%08x %08x\n", "intercepts:",
3191 control->intercepts[INTERCEPT_WORD3],
3192 control->intercepts[INTERCEPT_WORD4]);
3193 pr_err("%-20s%d\n", "pause filter count:", control->pause_filter_count);
3194 pr_err("%-20s%d\n", "pause filter threshold:",
3195 control->pause_filter_thresh);
3196 pr_err("%-20s%016llx\n", "iopm_base_pa:", control->iopm_base_pa);
3197 pr_err("%-20s%016llx\n", "msrpm_base_pa:", control->msrpm_base_pa);
3198 pr_err("%-20s%016llx\n", "tsc_offset:", control->tsc_offset);
3199 pr_err("%-20s%d\n", "asid:", control->asid);
3200 pr_err("%-20s%d\n", "tlb_ctl:", control->tlb_ctl);
3201 pr_err("%-20s%08x\n", "int_ctl:", control->int_ctl);
3202 pr_err("%-20s%08x\n", "int_vector:", control->int_vector);
3203 pr_err("%-20s%08x\n", "int_state:", control->int_state);
3204 pr_err("%-20s%08x\n", "exit_code:", control->exit_code);
3205 pr_err("%-20s%016llx\n", "exit_info1:", control->exit_info_1);
3206 pr_err("%-20s%016llx\n", "exit_info2:", control->exit_info_2);
3207 pr_err("%-20s%08x\n", "exit_int_info:", control->exit_int_info);
3208 pr_err("%-20s%08x\n", "exit_int_info_err:", control->exit_int_info_err);
3209 pr_err("%-20s%lld\n", "nested_ctl:", control->nested_ctl);
3210 pr_err("%-20s%016llx\n", "nested_cr3:", control->nested_cr3);
3211 pr_err("%-20s%016llx\n", "avic_vapic_bar:", control->avic_vapic_bar);
3212 pr_err("%-20s%016llx\n", "ghcb:", control->ghcb_gpa);
3213 pr_err("%-20s%08x\n", "event_inj:", control->event_inj);
3214 pr_err("%-20s%08x\n", "event_inj_err:", control->event_inj_err);
3215 pr_err("%-20s%lld\n", "virt_ext:", control->virt_ext);
3216 pr_err("%-20s%016llx\n", "next_rip:", control->next_rip);
3217 pr_err("%-20s%016llx\n", "avic_backing_page:", control->avic_backing_page);
3218 pr_err("%-20s%016llx\n", "avic_logical_id:", control->avic_logical_id);
3219 pr_err("%-20s%016llx\n", "avic_physical_id:", control->avic_physical_id);
3220 pr_err("%-20s%016llx\n", "vmsa_pa:", control->vmsa_pa);
3221 pr_err("VMCB State Save Area:\n");
3222 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3224 save->es.selector, save->es.attrib,
3225 save->es.limit, save->es.base);
3226 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3228 save->cs.selector, save->cs.attrib,
3229 save->cs.limit, save->cs.base);
3230 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3232 save->ss.selector, save->ss.attrib,
3233 save->ss.limit, save->ss.base);
3234 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3236 save->ds.selector, save->ds.attrib,
3237 save->ds.limit, save->ds.base);
3238 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3240 save->fs.selector, save->fs.attrib,
3241 save->fs.limit, save->fs.base);
3242 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3244 save->gs.selector, save->gs.attrib,
3245 save->gs.limit, save->gs.base);
3246 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3248 save->gdtr.selector, save->gdtr.attrib,
3249 save->gdtr.limit, save->gdtr.base);
3250 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3252 save->ldtr.selector, save->ldtr.attrib,
3253 save->ldtr.limit, save->ldtr.base);
3254 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3256 save->idtr.selector, save->idtr.attrib,
3257 save->idtr.limit, save->idtr.base);
3258 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3260 save->tr.selector, save->tr.attrib,
3261 save->tr.limit, save->tr.base);
3262 pr_err("cpl: %d efer: %016llx\n",
3263 save->cpl, save->efer);
3264 pr_err("%-15s %016llx %-13s %016llx\n",
3265 "cr0:", save->cr0, "cr2:", save->cr2);
3266 pr_err("%-15s %016llx %-13s %016llx\n",
3267 "cr3:", save->cr3, "cr4:", save->cr4);
3268 pr_err("%-15s %016llx %-13s %016llx\n",
3269 "dr6:", save->dr6, "dr7:", save->dr7);
3270 pr_err("%-15s %016llx %-13s %016llx\n",
3271 "rip:", save->rip, "rflags:", save->rflags);
3272 pr_err("%-15s %016llx %-13s %016llx\n",
3273 "rsp:", save->rsp, "rax:", save->rax);
3274 pr_err("%-15s %016llx %-13s %016llx\n",
3275 "star:", save->star, "lstar:", save->lstar);
3276 pr_err("%-15s %016llx %-13s %016llx\n",
3277 "cstar:", save->cstar, "sfmask:", save->sfmask);
3278 pr_err("%-15s %016llx %-13s %016llx\n",
3279 "kernel_gs_base:", save->kernel_gs_base,
3280 "sysenter_cs:", save->sysenter_cs);
3281 pr_err("%-15s %016llx %-13s %016llx\n",
3282 "sysenter_esp:", save->sysenter_esp,
3283 "sysenter_eip:", save->sysenter_eip);
3284 pr_err("%-15s %016llx %-13s %016llx\n",
3285 "gpat:", save->g_pat, "dbgctl:", save->dbgctl);
3286 pr_err("%-15s %016llx %-13s %016llx\n",
3287 "br_from:", save->br_from, "br_to:", save->br_to);
3288 pr_err("%-15s %016llx %-13s %016llx\n",
3289 "excp_from:", save->last_excp_from,
3290 "excp_to:", save->last_excp_to);
3293 static int svm_handle_invalid_exit(struct kvm_vcpu *vcpu, u64 exit_code)
3295 if (exit_code < ARRAY_SIZE(svm_exit_handlers) &&
3296 svm_exit_handlers[exit_code])
3299 vcpu_unimpl(vcpu, "svm: unexpected exit reason 0x%llx\n", exit_code);
3301 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3302 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
3303 vcpu->run->internal.ndata = 2;
3304 vcpu->run->internal.data[0] = exit_code;
3305 vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
3310 int svm_invoke_exit_handler(struct vcpu_svm *svm, u64 exit_code)
3312 if (svm_handle_invalid_exit(&svm->vcpu, exit_code))
3315 #ifdef CONFIG_RETPOLINE
3316 if (exit_code == SVM_EXIT_MSR)
3317 return msr_interception(svm);
3318 else if (exit_code == SVM_EXIT_VINTR)
3319 return interrupt_window_interception(svm);
3320 else if (exit_code == SVM_EXIT_INTR)
3321 return intr_interception(svm);
3322 else if (exit_code == SVM_EXIT_HLT)
3323 return halt_interception(svm);
3324 else if (exit_code == SVM_EXIT_NPF)
3325 return npf_interception(svm);
3327 return svm_exit_handlers[exit_code](svm);
3330 static void svm_get_exit_info(struct kvm_vcpu *vcpu, u64 *info1, u64 *info2,
3331 u32 *intr_info, u32 *error_code)
3333 struct vmcb_control_area *control = &to_svm(vcpu)->vmcb->control;
3335 *info1 = control->exit_info_1;
3336 *info2 = control->exit_info_2;
3337 *intr_info = control->exit_int_info;
3338 if ((*intr_info & SVM_EXITINTINFO_VALID) &&
3339 (*intr_info & SVM_EXITINTINFO_VALID_ERR))
3340 *error_code = control->exit_int_info_err;
3345 static int handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
3347 struct vcpu_svm *svm = to_svm(vcpu);
3348 struct kvm_run *kvm_run = vcpu->run;
3349 u32 exit_code = svm->vmcb->control.exit_code;
3351 trace_kvm_exit(exit_code, vcpu, KVM_ISA_SVM);
3353 /* SEV-ES guests must use the CR write traps to track CR registers. */
3354 if (!sev_es_guest(vcpu->kvm)) {
3355 if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
3356 vcpu->arch.cr0 = svm->vmcb->save.cr0;
3358 vcpu->arch.cr3 = svm->vmcb->save.cr3;
3361 if (is_guest_mode(vcpu)) {
3364 trace_kvm_nested_vmexit(exit_code, vcpu, KVM_ISA_SVM);
3366 vmexit = nested_svm_exit_special(svm);
3368 if (vmexit == NESTED_EXIT_CONTINUE)
3369 vmexit = nested_svm_exit_handled(svm);
3371 if (vmexit == NESTED_EXIT_DONE)
3375 if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
3376 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3377 kvm_run->fail_entry.hardware_entry_failure_reason
3378 = svm->vmcb->control.exit_code;
3379 kvm_run->fail_entry.cpu = vcpu->arch.last_vmentry_cpu;
3384 if (is_external_interrupt(svm->vmcb->control.exit_int_info) &&
3385 exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR &&
3386 exit_code != SVM_EXIT_NPF && exit_code != SVM_EXIT_TASK_SWITCH &&
3387 exit_code != SVM_EXIT_INTR && exit_code != SVM_EXIT_NMI)
3388 printk(KERN_ERR "%s: unexpected exit_int_info 0x%x "
3390 __func__, svm->vmcb->control.exit_int_info,
3393 if (exit_fastpath != EXIT_FASTPATH_NONE)
3396 return svm_invoke_exit_handler(svm, exit_code);
3399 static void reload_tss(struct kvm_vcpu *vcpu)
3401 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3403 sd->tss_desc->type = 9; /* available 32/64-bit TSS */
3407 static void pre_svm_run(struct vcpu_svm *svm)
3409 struct svm_cpu_data *sd = per_cpu(svm_data, svm->vcpu.cpu);
3411 if (sev_guest(svm->vcpu.kvm))
3412 return pre_sev_run(svm, svm->vcpu.cpu);
3414 /* FIXME: handle wraparound of asid_generation */
3415 if (svm->asid_generation != sd->asid_generation)
3419 static void svm_inject_nmi(struct kvm_vcpu *vcpu)
3421 struct vcpu_svm *svm = to_svm(vcpu);
3423 svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
3424 vcpu->arch.hflags |= HF_NMI_MASK;
3425 if (!sev_es_guest(svm->vcpu.kvm))
3426 svm_set_intercept(svm, INTERCEPT_IRET);
3427 ++vcpu->stat.nmi_injections;
3430 static void svm_set_irq(struct kvm_vcpu *vcpu)
3432 struct vcpu_svm *svm = to_svm(vcpu);
3434 BUG_ON(!(gif_set(svm)));
3436 trace_kvm_inj_virq(vcpu->arch.interrupt.nr);
3437 ++vcpu->stat.irq_injections;
3439 svm->vmcb->control.event_inj = vcpu->arch.interrupt.nr |
3440 SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR;
3443 static void svm_update_cr8_intercept(struct kvm_vcpu *vcpu, int tpr, int irr)
3445 struct vcpu_svm *svm = to_svm(vcpu);
3448 * SEV-ES guests must always keep the CR intercepts cleared. CR
3449 * tracking is done using the CR write traps.
3451 if (sev_es_guest(vcpu->kvm))
3454 if (nested_svm_virtualize_tpr(vcpu))
3457 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3463 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
3466 bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
3468 struct vcpu_svm *svm = to_svm(vcpu);
3469 struct vmcb *vmcb = svm->vmcb;
3475 if (is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3478 ret = (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK) ||
3479 (svm->vcpu.arch.hflags & HF_NMI_MASK);
3484 static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3486 struct vcpu_svm *svm = to_svm(vcpu);
3487 if (svm->nested.nested_run_pending)
3490 /* An NMI must not be injected into L2 if it's supposed to VM-Exit. */
3491 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3494 return !svm_nmi_blocked(vcpu);
3497 static bool svm_get_nmi_mask(struct kvm_vcpu *vcpu)
3499 struct vcpu_svm *svm = to_svm(vcpu);
3501 return !!(svm->vcpu.arch.hflags & HF_NMI_MASK);
3504 static void svm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
3506 struct vcpu_svm *svm = to_svm(vcpu);
3509 svm->vcpu.arch.hflags |= HF_NMI_MASK;
3510 if (!sev_es_guest(svm->vcpu.kvm))
3511 svm_set_intercept(svm, INTERCEPT_IRET);
3513 svm->vcpu.arch.hflags &= ~HF_NMI_MASK;
3514 if (!sev_es_guest(svm->vcpu.kvm))
3515 svm_clr_intercept(svm, INTERCEPT_IRET);
3519 bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
3521 struct vcpu_svm *svm = to_svm(vcpu);
3522 struct vmcb *vmcb = svm->vmcb;
3527 if (sev_es_guest(svm->vcpu.kvm)) {
3529 * SEV-ES guests to not expose RFLAGS. Use the VMCB interrupt mask
3530 * bit to determine the state of the IF flag.
3532 if (!(vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK))
3534 } else if (is_guest_mode(vcpu)) {
3535 /* As long as interrupts are being delivered... */
3536 if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
3537 ? !(svm->nested.hsave->save.rflags & X86_EFLAGS_IF)
3538 : !(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3541 /* ... vmexits aren't blocked by the interrupt shadow */
3542 if (nested_exit_on_intr(svm))
3545 if (!(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3549 return (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK);
3552 static int svm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3554 struct vcpu_svm *svm = to_svm(vcpu);
3555 if (svm->nested.nested_run_pending)
3559 * An IRQ must not be injected into L2 if it's supposed to VM-Exit,
3560 * e.g. if the IRQ arrived asynchronously after checking nested events.
3562 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_intr(svm))
3565 return !svm_interrupt_blocked(vcpu);
3568 static void svm_enable_irq_window(struct kvm_vcpu *vcpu)
3570 struct vcpu_svm *svm = to_svm(vcpu);
3573 * In case GIF=0 we can't rely on the CPU to tell us when GIF becomes
3574 * 1, because that's a separate STGI/VMRUN intercept. The next time we
3575 * get that intercept, this function will be called again though and
3576 * we'll get the vintr intercept. However, if the vGIF feature is
3577 * enabled, the STGI interception will not occur. Enable the irq
3578 * window under the assumption that the hardware will set the GIF.
3580 if (vgif_enabled(svm) || gif_set(svm)) {
3582 * IRQ window is not needed when AVIC is enabled,
3583 * unless we have pending ExtINT since it cannot be injected
3584 * via AVIC. In such case, we need to temporarily disable AVIC,
3585 * and fallback to injecting IRQ via V_IRQ.
3587 svm_toggle_avic_for_irq_window(vcpu, false);
3592 static void svm_enable_nmi_window(struct kvm_vcpu *vcpu)
3594 struct vcpu_svm *svm = to_svm(vcpu);
3596 if ((svm->vcpu.arch.hflags & (HF_NMI_MASK | HF_IRET_MASK))
3598 return; /* IRET will cause a vm exit */
3600 if (!gif_set(svm)) {
3601 if (vgif_enabled(svm))
3602 svm_set_intercept(svm, INTERCEPT_STGI);
3603 return; /* STGI will cause a vm exit */
3607 * Something prevents NMI from been injected. Single step over possible
3608 * problem (IRET or exception injection or interrupt shadow)
3610 svm->nmi_singlestep_guest_rflags = svm_get_rflags(vcpu);
3611 svm->nmi_singlestep = true;
3612 svm->vmcb->save.rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
3615 static int svm_set_tss_addr(struct kvm *kvm, unsigned int addr)
3620 static int svm_set_identity_map_addr(struct kvm *kvm, u64 ident_addr)
3625 void svm_flush_tlb(struct kvm_vcpu *vcpu)
3627 struct vcpu_svm *svm = to_svm(vcpu);
3630 * Flush only the current ASID even if the TLB flush was invoked via
3631 * kvm_flush_remote_tlbs(). Although flushing remote TLBs requires all
3632 * ASIDs to be flushed, KVM uses a single ASID for L1 and L2, and
3633 * unconditionally does a TLB flush on both nested VM-Enter and nested
3634 * VM-Exit (via kvm_mmu_reset_context()).
3636 if (static_cpu_has(X86_FEATURE_FLUSHBYASID))
3637 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3639 svm->asid_generation--;
3642 static void svm_flush_tlb_gva(struct kvm_vcpu *vcpu, gva_t gva)
3644 struct vcpu_svm *svm = to_svm(vcpu);
3646 invlpga(gva, svm->vmcb->control.asid);
3649 static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)
3651 struct vcpu_svm *svm = to_svm(vcpu);
3653 if (nested_svm_virtualize_tpr(vcpu))
3656 if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
3657 int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
3658 kvm_set_cr8(vcpu, cr8);
3662 static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
3664 struct vcpu_svm *svm = to_svm(vcpu);
3667 if (nested_svm_virtualize_tpr(vcpu) ||
3668 kvm_vcpu_apicv_active(vcpu))
3671 cr8 = kvm_get_cr8(vcpu);
3672 svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
3673 svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
3676 static void svm_complete_interrupts(struct vcpu_svm *svm)
3680 u32 exitintinfo = svm->vmcb->control.exit_int_info;
3681 unsigned int3_injected = svm->int3_injected;
3683 svm->int3_injected = 0;
3686 * If we've made progress since setting HF_IRET_MASK, we've
3687 * executed an IRET and can allow NMI injection.
3689 if ((svm->vcpu.arch.hflags & HF_IRET_MASK) &&
3690 (sev_es_guest(svm->vcpu.kvm) ||
3691 kvm_rip_read(&svm->vcpu) != svm->nmi_iret_rip)) {
3692 svm->vcpu.arch.hflags &= ~(HF_NMI_MASK | HF_IRET_MASK);
3693 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
3696 svm->vcpu.arch.nmi_injected = false;
3697 kvm_clear_exception_queue(&svm->vcpu);
3698 kvm_clear_interrupt_queue(&svm->vcpu);
3700 if (!(exitintinfo & SVM_EXITINTINFO_VALID))
3703 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
3705 vector = exitintinfo & SVM_EXITINTINFO_VEC_MASK;
3706 type = exitintinfo & SVM_EXITINTINFO_TYPE_MASK;
3709 case SVM_EXITINTINFO_TYPE_NMI:
3710 svm->vcpu.arch.nmi_injected = true;
3712 case SVM_EXITINTINFO_TYPE_EXEPT:
3714 * Never re-inject a #VC exception.
3716 if (vector == X86_TRAP_VC)
3720 * In case of software exceptions, do not reinject the vector,
3721 * but re-execute the instruction instead. Rewind RIP first
3722 * if we emulated INT3 before.
3724 if (kvm_exception_is_soft(vector)) {
3725 if (vector == BP_VECTOR && int3_injected &&
3726 kvm_is_linear_rip(&svm->vcpu, svm->int3_rip))
3727 kvm_rip_write(&svm->vcpu,
3728 kvm_rip_read(&svm->vcpu) -
3732 if (exitintinfo & SVM_EXITINTINFO_VALID_ERR) {
3733 u32 err = svm->vmcb->control.exit_int_info_err;
3734 kvm_requeue_exception_e(&svm->vcpu, vector, err);
3737 kvm_requeue_exception(&svm->vcpu, vector);
3739 case SVM_EXITINTINFO_TYPE_INTR:
3740 kvm_queue_interrupt(&svm->vcpu, vector, false);
3747 static void svm_cancel_injection(struct kvm_vcpu *vcpu)
3749 struct vcpu_svm *svm = to_svm(vcpu);
3750 struct vmcb_control_area *control = &svm->vmcb->control;
3752 control->exit_int_info = control->event_inj;
3753 control->exit_int_info_err = control->event_inj_err;
3754 control->event_inj = 0;
3755 svm_complete_interrupts(svm);
3758 static fastpath_t svm_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
3760 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_MSR &&
3761 to_svm(vcpu)->vmcb->control.exit_info_1)
3762 return handle_fastpath_set_msr_irqoff(vcpu);
3764 return EXIT_FASTPATH_NONE;
3767 static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu,
3768 struct vcpu_svm *svm)
3771 * VMENTER enables interrupts (host state), but the kernel state is
3772 * interrupts disabled when this is invoked. Also tell RCU about
3773 * it. This is the same logic as for exit_to_user_mode().
3775 * This ensures that e.g. latency analysis on the host observes
3776 * guest mode as interrupt enabled.
3778 * guest_enter_irqoff() informs context tracking about the
3779 * transition to guest mode and if enabled adjusts RCU state
3782 instrumentation_begin();
3783 trace_hardirqs_on_prepare();
3784 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
3785 instrumentation_end();
3787 guest_enter_irqoff();
3788 lockdep_hardirqs_on(CALLER_ADDR0);
3790 if (sev_es_guest(svm->vcpu.kvm)) {
3791 __svm_sev_es_vcpu_run(svm->vmcb_pa);
3793 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3795 __svm_vcpu_run(svm->vmcb_pa, (unsigned long *)&svm->vcpu.arch.regs);
3797 vmload(__sme_page_pa(sd->save_area));
3801 * VMEXIT disables interrupts (host state), but tracing and lockdep
3802 * have them in state 'on' as recorded before entering guest mode.
3803 * Same as enter_from_user_mode().
3805 * guest_exit_irqoff() restores host context and reinstates RCU if
3806 * enabled and required.
3808 * This needs to be done before the below as native_read_msr()
3809 * contains a tracepoint and x86_spec_ctrl_restore_host() calls
3810 * into world and some more.
3812 lockdep_hardirqs_off(CALLER_ADDR0);
3813 guest_exit_irqoff();
3815 instrumentation_begin();
3816 trace_hardirqs_off_finish();
3817 instrumentation_end();
3820 static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
3822 struct vcpu_svm *svm = to_svm(vcpu);
3824 trace_kvm_entry(vcpu);
3826 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
3827 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
3828 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
3831 * Disable singlestep if we're injecting an interrupt/exception.
3832 * We don't want our modified rflags to be pushed on the stack where
3833 * we might not be able to easily reset them if we disabled NMI
3836 if (svm->nmi_singlestep && svm->vmcb->control.event_inj) {
3838 * Event injection happens before external interrupts cause a
3839 * vmexit and interrupts are disabled here, so smp_send_reschedule
3840 * is enough to force an immediate vmexit.
3842 disable_nmi_singlestep(svm);
3843 smp_send_reschedule(vcpu->cpu);
3848 sync_lapic_to_cr8(vcpu);
3850 if (unlikely(svm->asid != svm->vmcb->control.asid)) {
3851 svm->vmcb->control.asid = svm->asid;
3852 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
3854 svm->vmcb->save.cr2 = vcpu->arch.cr2;
3857 * Run with all-zero DR6 unless needed, so that we can get the exact cause
3860 if (unlikely(svm->vcpu.arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT))
3861 svm_set_dr6(svm, vcpu->arch.dr6);
3863 svm_set_dr6(svm, DR6_ACTIVE_LOW);
3866 kvm_load_guest_xsave_state(vcpu);
3868 kvm_wait_lapic_expire(vcpu);
3871 * If this vCPU has touched SPEC_CTRL, restore the guest's value if
3872 * it's non-zero. Since vmentry is serialising on affected CPUs, there
3873 * is no need to worry about the conditional branch over the wrmsr
3874 * being speculatively taken.
3876 x86_spec_ctrl_set_guest(svm->spec_ctrl, svm->virt_spec_ctrl);
3878 svm_vcpu_enter_exit(vcpu, svm);
3881 * We do not use IBRS in the kernel. If this vCPU has used the
3882 * SPEC_CTRL MSR it may have left it on; save the value and
3883 * turn it off. This is much more efficient than blindly adding
3884 * it to the atomic save/restore list. Especially as the former
3885 * (Saving guest MSRs on vmexit) doesn't even exist in KVM.
3887 * For non-nested case:
3888 * If the L01 MSR bitmap does not intercept the MSR, then we need to
3892 * If the L02 MSR bitmap does not intercept the MSR, then we need to
3895 if (unlikely(!msr_write_intercepted(vcpu, MSR_IA32_SPEC_CTRL)))
3896 svm->spec_ctrl = native_read_msr(MSR_IA32_SPEC_CTRL);
3898 if (!sev_es_guest(svm->vcpu.kvm))
3901 x86_spec_ctrl_restore_host(svm->spec_ctrl, svm->virt_spec_ctrl);
3903 if (!sev_es_guest(svm->vcpu.kvm)) {
3904 vcpu->arch.cr2 = svm->vmcb->save.cr2;
3905 vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
3906 vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
3907 vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
3910 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3911 kvm_before_interrupt(&svm->vcpu);
3913 kvm_load_host_xsave_state(vcpu);
3916 /* Any pending NMI will happen here */
3918 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3919 kvm_after_interrupt(&svm->vcpu);
3921 sync_cr8_to_lapic(vcpu);
3924 if (is_guest_mode(&svm->vcpu)) {
3925 sync_nested_vmcb_control(svm);
3926 svm->nested.nested_run_pending = 0;
3929 svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
3930 vmcb_mark_all_clean(svm->vmcb);
3932 /* if exit due to PF check for async PF */
3933 if (svm->vmcb->control.exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR)
3934 svm->vcpu.arch.apf.host_apf_flags =
3935 kvm_read_and_reset_apf_flags();
3938 vcpu->arch.regs_avail &= ~(1 << VCPU_EXREG_PDPTR);
3939 vcpu->arch.regs_dirty &= ~(1 << VCPU_EXREG_PDPTR);
3943 * We need to handle MC intercepts here before the vcpu has a chance to
3944 * change the physical cpu
3946 if (unlikely(svm->vmcb->control.exit_code ==
3947 SVM_EXIT_EXCP_BASE + MC_VECTOR))
3948 svm_handle_mce(svm);
3950 svm_complete_interrupts(svm);
3952 if (is_guest_mode(vcpu))
3953 return EXIT_FASTPATH_NONE;
3955 return svm_exit_handlers_fastpath(vcpu);
3958 static void svm_load_mmu_pgd(struct kvm_vcpu *vcpu, unsigned long root,
3961 struct vcpu_svm *svm = to_svm(vcpu);
3964 cr3 = __sme_set(root);
3966 svm->vmcb->control.nested_cr3 = cr3;
3967 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
3969 /* Loading L2's CR3 is handled by enter_svm_guest_mode. */
3970 if (!test_bit(VCPU_EXREG_CR3, (ulong *)&vcpu->arch.regs_avail))
3972 cr3 = vcpu->arch.cr3;
3975 svm->vmcb->save.cr3 = cr3;
3976 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
3979 static int is_disabled(void)
3983 rdmsrl(MSR_VM_CR, vm_cr);
3984 if (vm_cr & (1 << SVM_VM_CR_SVM_DISABLE))
3991 svm_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
3994 * Patch in the VMMCALL instruction:
3996 hypercall[0] = 0x0f;
3997 hypercall[1] = 0x01;
3998 hypercall[2] = 0xd9;
4001 static int __init svm_check_processor_compat(void)
4006 static bool svm_cpu_has_accelerated_tpr(void)
4012 * The kvm parameter can be NULL (module initialization, or invocation before
4013 * VM creation). Be sure to check the kvm parameter before using it.
4015 static bool svm_has_emulated_msr(struct kvm *kvm, u32 index)
4018 case MSR_IA32_MCG_EXT_CTL:
4019 case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:
4021 case MSR_IA32_SMBASE:
4022 /* SEV-ES guests do not support SMM, so report false */
4023 if (kvm && sev_es_guest(kvm))
4033 static u64 svm_get_mt_mask(struct kvm_vcpu *vcpu, gfn_t gfn, bool is_mmio)
4038 static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
4040 struct vcpu_svm *svm = to_svm(vcpu);
4041 struct kvm_cpuid_entry2 *best;
4043 vcpu->arch.xsaves_enabled = guest_cpuid_has(vcpu, X86_FEATURE_XSAVE) &&
4044 boot_cpu_has(X86_FEATURE_XSAVE) &&
4045 boot_cpu_has(X86_FEATURE_XSAVES);
4047 /* Update nrips enabled cache */
4048 svm->nrips_enabled = kvm_cpu_cap_has(X86_FEATURE_NRIPS) &&
4049 guest_cpuid_has(&svm->vcpu, X86_FEATURE_NRIPS);
4051 /* Check again if INVPCID interception if required */
4052 svm_check_invpcid(svm);
4054 /* For sev guests, the memory encryption bit is not reserved in CR3. */
4055 if (sev_guest(vcpu->kvm)) {
4056 best = kvm_find_cpuid_entry(vcpu, 0x8000001F, 0);
4058 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
4061 if (!kvm_vcpu_apicv_active(vcpu))
4065 * AVIC does not work with an x2APIC mode guest. If the X2APIC feature
4066 * is exposed to the guest, disable AVIC.
4068 if (guest_cpuid_has(vcpu, X86_FEATURE_X2APIC))
4069 kvm_request_apicv_update(vcpu->kvm, false,
4070 APICV_INHIBIT_REASON_X2APIC);
4073 * Currently, AVIC does not work with nested virtualization.
4074 * So, we disable AVIC when cpuid for SVM is set in the L1 guest.
4076 if (nested && guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4077 kvm_request_apicv_update(vcpu->kvm, false,
4078 APICV_INHIBIT_REASON_NESTED);
4081 static bool svm_has_wbinvd_exit(void)
4086 #define PRE_EX(exit) { .exit_code = (exit), \
4087 .stage = X86_ICPT_PRE_EXCEPT, }
4088 #define POST_EX(exit) { .exit_code = (exit), \
4089 .stage = X86_ICPT_POST_EXCEPT, }
4090 #define POST_MEM(exit) { .exit_code = (exit), \
4091 .stage = X86_ICPT_POST_MEMACCESS, }
4093 static const struct __x86_intercept {
4095 enum x86_intercept_stage stage;
4096 } x86_intercept_map[] = {
4097 [x86_intercept_cr_read] = POST_EX(SVM_EXIT_READ_CR0),
4098 [x86_intercept_cr_write] = POST_EX(SVM_EXIT_WRITE_CR0),
4099 [x86_intercept_clts] = POST_EX(SVM_EXIT_WRITE_CR0),
4100 [x86_intercept_lmsw] = POST_EX(SVM_EXIT_WRITE_CR0),
4101 [x86_intercept_smsw] = POST_EX(SVM_EXIT_READ_CR0),
4102 [x86_intercept_dr_read] = POST_EX(SVM_EXIT_READ_DR0),
4103 [x86_intercept_dr_write] = POST_EX(SVM_EXIT_WRITE_DR0),
4104 [x86_intercept_sldt] = POST_EX(SVM_EXIT_LDTR_READ),
4105 [x86_intercept_str] = POST_EX(SVM_EXIT_TR_READ),
4106 [x86_intercept_lldt] = POST_EX(SVM_EXIT_LDTR_WRITE),
4107 [x86_intercept_ltr] = POST_EX(SVM_EXIT_TR_WRITE),
4108 [x86_intercept_sgdt] = POST_EX(SVM_EXIT_GDTR_READ),
4109 [x86_intercept_sidt] = POST_EX(SVM_EXIT_IDTR_READ),
4110 [x86_intercept_lgdt] = POST_EX(SVM_EXIT_GDTR_WRITE),
4111 [x86_intercept_lidt] = POST_EX(SVM_EXIT_IDTR_WRITE),
4112 [x86_intercept_vmrun] = POST_EX(SVM_EXIT_VMRUN),
4113 [x86_intercept_vmmcall] = POST_EX(SVM_EXIT_VMMCALL),
4114 [x86_intercept_vmload] = POST_EX(SVM_EXIT_VMLOAD),
4115 [x86_intercept_vmsave] = POST_EX(SVM_EXIT_VMSAVE),
4116 [x86_intercept_stgi] = POST_EX(SVM_EXIT_STGI),
4117 [x86_intercept_clgi] = POST_EX(SVM_EXIT_CLGI),
4118 [x86_intercept_skinit] = POST_EX(SVM_EXIT_SKINIT),
4119 [x86_intercept_invlpga] = POST_EX(SVM_EXIT_INVLPGA),
4120 [x86_intercept_rdtscp] = POST_EX(SVM_EXIT_RDTSCP),
4121 [x86_intercept_monitor] = POST_MEM(SVM_EXIT_MONITOR),
4122 [x86_intercept_mwait] = POST_EX(SVM_EXIT_MWAIT),
4123 [x86_intercept_invlpg] = POST_EX(SVM_EXIT_INVLPG),
4124 [x86_intercept_invd] = POST_EX(SVM_EXIT_INVD),
4125 [x86_intercept_wbinvd] = POST_EX(SVM_EXIT_WBINVD),
4126 [x86_intercept_wrmsr] = POST_EX(SVM_EXIT_MSR),
4127 [x86_intercept_rdtsc] = POST_EX(SVM_EXIT_RDTSC),
4128 [x86_intercept_rdmsr] = POST_EX(SVM_EXIT_MSR),
4129 [x86_intercept_rdpmc] = POST_EX(SVM_EXIT_RDPMC),
4130 [x86_intercept_cpuid] = PRE_EX(SVM_EXIT_CPUID),
4131 [x86_intercept_rsm] = PRE_EX(SVM_EXIT_RSM),
4132 [x86_intercept_pause] = PRE_EX(SVM_EXIT_PAUSE),
4133 [x86_intercept_pushf] = PRE_EX(SVM_EXIT_PUSHF),
4134 [x86_intercept_popf] = PRE_EX(SVM_EXIT_POPF),
4135 [x86_intercept_intn] = PRE_EX(SVM_EXIT_SWINT),
4136 [x86_intercept_iret] = PRE_EX(SVM_EXIT_IRET),
4137 [x86_intercept_icebp] = PRE_EX(SVM_EXIT_ICEBP),
4138 [x86_intercept_hlt] = POST_EX(SVM_EXIT_HLT),
4139 [x86_intercept_in] = POST_EX(SVM_EXIT_IOIO),
4140 [x86_intercept_ins] = POST_EX(SVM_EXIT_IOIO),
4141 [x86_intercept_out] = POST_EX(SVM_EXIT_IOIO),
4142 [x86_intercept_outs] = POST_EX(SVM_EXIT_IOIO),
4143 [x86_intercept_xsetbv] = PRE_EX(SVM_EXIT_XSETBV),
4150 static int svm_check_intercept(struct kvm_vcpu *vcpu,
4151 struct x86_instruction_info *info,
4152 enum x86_intercept_stage stage,
4153 struct x86_exception *exception)
4155 struct vcpu_svm *svm = to_svm(vcpu);
4156 int vmexit, ret = X86EMUL_CONTINUE;
4157 struct __x86_intercept icpt_info;
4158 struct vmcb *vmcb = svm->vmcb;
4160 if (info->intercept >= ARRAY_SIZE(x86_intercept_map))
4163 icpt_info = x86_intercept_map[info->intercept];
4165 if (stage != icpt_info.stage)
4168 switch (icpt_info.exit_code) {
4169 case SVM_EXIT_READ_CR0:
4170 if (info->intercept == x86_intercept_cr_read)
4171 icpt_info.exit_code += info->modrm_reg;
4173 case SVM_EXIT_WRITE_CR0: {
4174 unsigned long cr0, val;
4176 if (info->intercept == x86_intercept_cr_write)
4177 icpt_info.exit_code += info->modrm_reg;
4179 if (icpt_info.exit_code != SVM_EXIT_WRITE_CR0 ||
4180 info->intercept == x86_intercept_clts)
4183 if (!(vmcb_is_intercept(&svm->nested.ctl,
4184 INTERCEPT_SELECTIVE_CR0)))
4187 cr0 = vcpu->arch.cr0 & ~SVM_CR0_SELECTIVE_MASK;
4188 val = info->src_val & ~SVM_CR0_SELECTIVE_MASK;
4190 if (info->intercept == x86_intercept_lmsw) {
4193 /* lmsw can't clear PE - catch this here */
4194 if (cr0 & X86_CR0_PE)
4199 icpt_info.exit_code = SVM_EXIT_CR0_SEL_WRITE;
4203 case SVM_EXIT_READ_DR0:
4204 case SVM_EXIT_WRITE_DR0:
4205 icpt_info.exit_code += info->modrm_reg;
4208 if (info->intercept == x86_intercept_wrmsr)
4209 vmcb->control.exit_info_1 = 1;
4211 vmcb->control.exit_info_1 = 0;
4213 case SVM_EXIT_PAUSE:
4215 * We get this for NOP only, but pause
4216 * is rep not, check this here
4218 if (info->rep_prefix != REPE_PREFIX)
4221 case SVM_EXIT_IOIO: {
4225 if (info->intercept == x86_intercept_in ||
4226 info->intercept == x86_intercept_ins) {
4227 exit_info = ((info->src_val & 0xffff) << 16) |
4229 bytes = info->dst_bytes;
4231 exit_info = (info->dst_val & 0xffff) << 16;
4232 bytes = info->src_bytes;
4235 if (info->intercept == x86_intercept_outs ||
4236 info->intercept == x86_intercept_ins)
4237 exit_info |= SVM_IOIO_STR_MASK;
4239 if (info->rep_prefix)
4240 exit_info |= SVM_IOIO_REP_MASK;
4242 bytes = min(bytes, 4u);
4244 exit_info |= bytes << SVM_IOIO_SIZE_SHIFT;
4246 exit_info |= (u32)info->ad_bytes << (SVM_IOIO_ASIZE_SHIFT - 1);
4248 vmcb->control.exit_info_1 = exit_info;
4249 vmcb->control.exit_info_2 = info->next_rip;
4257 /* TODO: Advertise NRIPS to guest hypervisor unconditionally */
4258 if (static_cpu_has(X86_FEATURE_NRIPS))
4259 vmcb->control.next_rip = info->next_rip;
4260 vmcb->control.exit_code = icpt_info.exit_code;
4261 vmexit = nested_svm_exit_handled(svm);
4263 ret = (vmexit == NESTED_EXIT_DONE) ? X86EMUL_INTERCEPTED
4270 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
4274 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
4276 if (!kvm_pause_in_guest(vcpu->kvm))
4277 shrink_ple_window(vcpu);
4280 static void svm_setup_mce(struct kvm_vcpu *vcpu)
4282 /* [63:9] are reserved. */
4283 vcpu->arch.mcg_cap &= 0x1ff;
4286 bool svm_smi_blocked(struct kvm_vcpu *vcpu)
4288 struct vcpu_svm *svm = to_svm(vcpu);
4290 /* Per APM Vol.2 15.22.2 "Response to SMI" */
4294 return is_smm(vcpu);
4297 static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
4299 struct vcpu_svm *svm = to_svm(vcpu);
4300 if (svm->nested.nested_run_pending)
4303 /* An SMI must not be injected into L2 if it's supposed to VM-Exit. */
4304 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_smi(svm))
4307 return !svm_smi_blocked(vcpu);
4310 static int svm_pre_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
4312 struct vcpu_svm *svm = to_svm(vcpu);
4315 if (is_guest_mode(vcpu)) {
4316 /* FED8h - SVM Guest */
4317 put_smstate(u64, smstate, 0x7ed8, 1);
4318 /* FEE0h - SVM Guest VMCB Physical Address */
4319 put_smstate(u64, smstate, 0x7ee0, svm->nested.vmcb12_gpa);
4321 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4322 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4323 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4325 ret = nested_svm_vmexit(svm);
4332 static int svm_pre_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
4334 struct vcpu_svm *svm = to_svm(vcpu);
4335 struct kvm_host_map map;
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);
4344 if (!guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4347 if (!(saved_efer & EFER_SVME))
4350 if (kvm_vcpu_map(&svm->vcpu,
4351 gpa_to_gfn(vmcb12_gpa), &map) == -EINVAL)
4354 if (svm_allocate_nested(svm))
4357 ret = enter_svm_guest_mode(svm, vmcb12_gpa, map.hva);
4358 kvm_vcpu_unmap(&svm->vcpu, &map, true);
4365 static void svm_enable_smi_window(struct kvm_vcpu *vcpu)
4367 struct vcpu_svm *svm = to_svm(vcpu);
4369 if (!gif_set(svm)) {
4370 if (vgif_enabled(svm))
4371 svm_set_intercept(svm, INTERCEPT_STGI);
4372 /* STGI will cause a vm exit */
4374 /* We must be in SMM; RSM will cause a vmexit anyway. */
4378 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, void *insn, int insn_len)
4380 bool smep, smap, is_user;
4384 * When the guest is an SEV-ES guest, emulation is not possible.
4386 if (sev_es_guest(vcpu->kvm))
4390 * Detect and workaround Errata 1096 Fam_17h_00_0Fh.
4393 * When CPU raise #NPF on guest data access and vCPU CR4.SMAP=1, it is
4394 * possible that CPU microcode implementing DecodeAssist will fail
4395 * to read bytes of instruction which caused #NPF. In this case,
4396 * GuestIntrBytes field of the VMCB on a VMEXIT will incorrectly
4397 * return 0 instead of the correct guest instruction bytes.
4399 * This happens because CPU microcode reading instruction bytes
4400 * uses a special opcode which attempts to read data using CPL=0
4401 * priviledges. The microcode reads CS:RIP and if it hits a SMAP
4402 * fault, it gives up and returns no instruction bytes.
4405 * We reach here in case CPU supports DecodeAssist, raised #NPF and
4406 * returned 0 in GuestIntrBytes field of the VMCB.
4407 * First, errata can only be triggered in case vCPU CR4.SMAP=1.
4408 * Second, if vCPU CR4.SMEP=1, errata could only be triggered
4409 * in case vCPU CPL==3 (Because otherwise guest would have triggered
4410 * a SMEP fault instead of #NPF).
4411 * Otherwise, vCPU CR4.SMEP=0, errata could be triggered by any vCPU CPL.
4412 * As most guests enable SMAP if they have also enabled SMEP, use above
4413 * logic in order to attempt minimize false-positive of detecting errata
4414 * while still preserving all cases semantic correctness.
4417 * To determine what instruction the guest was executing, the hypervisor
4418 * will have to decode the instruction at the instruction pointer.
4420 * In non SEV guest, hypervisor will be able to read the guest
4421 * memory to decode the instruction pointer when insn_len is zero
4422 * so we return true to indicate that decoding is possible.
4424 * But in the SEV guest, the guest memory is encrypted with the
4425 * guest specific key and hypervisor will not be able to decode the
4426 * instruction pointer so we will not able to workaround it. Lets
4427 * print the error and request to kill the guest.
4429 if (likely(!insn || insn_len))
4433 * If RIP is invalid, go ahead with emulation which will cause an
4434 * internal error exit.
4436 if (!kvm_vcpu_gfn_to_memslot(vcpu, kvm_rip_read(vcpu) >> PAGE_SHIFT))
4439 cr4 = kvm_read_cr4(vcpu);
4440 smep = cr4 & X86_CR4_SMEP;
4441 smap = cr4 & X86_CR4_SMAP;
4442 is_user = svm_get_cpl(vcpu) == 3;
4443 if (smap && (!smep || is_user)) {
4444 if (!sev_guest(vcpu->kvm))
4447 pr_err_ratelimited("KVM: SEV Guest triggered AMD Erratum 1096\n");
4448 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
4454 static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
4456 struct vcpu_svm *svm = to_svm(vcpu);
4459 * TODO: Last condition latch INIT signals on vCPU when
4460 * vCPU is in guest-mode and vmcb12 defines intercept on INIT.
4461 * To properly emulate the INIT intercept,
4462 * svm_check_nested_events() should call nested_svm_vmexit()
4463 * if an INIT signal is pending.
4465 return !gif_set(svm) ||
4466 (vmcb_is_intercept(&svm->vmcb->control, INTERCEPT_INIT));
4469 static void svm_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
4471 if (!sev_es_guest(vcpu->kvm))
4472 return kvm_vcpu_deliver_sipi_vector(vcpu, vector);
4474 sev_vcpu_deliver_sipi_vector(vcpu, vector);
4477 static void svm_vm_destroy(struct kvm *kvm)
4479 avic_vm_destroy(kvm);
4480 sev_vm_destroy(kvm);
4483 static int svm_vm_init(struct kvm *kvm)
4485 if (!pause_filter_count || !pause_filter_thresh)
4486 kvm->arch.pause_in_guest = true;
4489 int ret = avic_vm_init(kvm);
4494 kvm_apicv_init(kvm, avic);
4498 static struct kvm_x86_ops svm_x86_ops __initdata = {
4499 .hardware_unsetup = svm_hardware_teardown,
4500 .hardware_enable = svm_hardware_enable,
4501 .hardware_disable = svm_hardware_disable,
4502 .cpu_has_accelerated_tpr = svm_cpu_has_accelerated_tpr,
4503 .has_emulated_msr = svm_has_emulated_msr,
4505 .vcpu_create = svm_create_vcpu,
4506 .vcpu_free = svm_free_vcpu,
4507 .vcpu_reset = svm_vcpu_reset,
4509 .vm_size = sizeof(struct kvm_svm),
4510 .vm_init = svm_vm_init,
4511 .vm_destroy = svm_vm_destroy,
4513 .prepare_guest_switch = svm_prepare_guest_switch,
4514 .vcpu_load = svm_vcpu_load,
4515 .vcpu_put = svm_vcpu_put,
4516 .vcpu_blocking = svm_vcpu_blocking,
4517 .vcpu_unblocking = svm_vcpu_unblocking,
4519 .update_exception_bitmap = svm_update_exception_bitmap,
4520 .get_msr_feature = svm_get_msr_feature,
4521 .get_msr = svm_get_msr,
4522 .set_msr = svm_set_msr,
4523 .get_segment_base = svm_get_segment_base,
4524 .get_segment = svm_get_segment,
4525 .set_segment = svm_set_segment,
4526 .get_cpl = svm_get_cpl,
4527 .get_cs_db_l_bits = kvm_get_cs_db_l_bits,
4528 .set_cr0 = svm_set_cr0,
4529 .is_valid_cr4 = svm_is_valid_cr4,
4530 .set_cr4 = svm_set_cr4,
4531 .set_efer = svm_set_efer,
4532 .get_idt = svm_get_idt,
4533 .set_idt = svm_set_idt,
4534 .get_gdt = svm_get_gdt,
4535 .set_gdt = svm_set_gdt,
4536 .set_dr7 = svm_set_dr7,
4537 .sync_dirty_debug_regs = svm_sync_dirty_debug_regs,
4538 .cache_reg = svm_cache_reg,
4539 .get_rflags = svm_get_rflags,
4540 .set_rflags = svm_set_rflags,
4542 .tlb_flush_all = svm_flush_tlb,
4543 .tlb_flush_current = svm_flush_tlb,
4544 .tlb_flush_gva = svm_flush_tlb_gva,
4545 .tlb_flush_guest = svm_flush_tlb,
4547 .run = svm_vcpu_run,
4548 .handle_exit = handle_exit,
4549 .skip_emulated_instruction = skip_emulated_instruction,
4550 .update_emulated_instruction = NULL,
4551 .set_interrupt_shadow = svm_set_interrupt_shadow,
4552 .get_interrupt_shadow = svm_get_interrupt_shadow,
4553 .patch_hypercall = svm_patch_hypercall,
4554 .set_irq = svm_set_irq,
4555 .set_nmi = svm_inject_nmi,
4556 .queue_exception = svm_queue_exception,
4557 .cancel_injection = svm_cancel_injection,
4558 .interrupt_allowed = svm_interrupt_allowed,
4559 .nmi_allowed = svm_nmi_allowed,
4560 .get_nmi_mask = svm_get_nmi_mask,
4561 .set_nmi_mask = svm_set_nmi_mask,
4562 .enable_nmi_window = svm_enable_nmi_window,
4563 .enable_irq_window = svm_enable_irq_window,
4564 .update_cr8_intercept = svm_update_cr8_intercept,
4565 .set_virtual_apic_mode = svm_set_virtual_apic_mode,
4566 .refresh_apicv_exec_ctrl = svm_refresh_apicv_exec_ctrl,
4567 .check_apicv_inhibit_reasons = svm_check_apicv_inhibit_reasons,
4568 .pre_update_apicv_exec_ctrl = svm_pre_update_apicv_exec_ctrl,
4569 .load_eoi_exitmap = svm_load_eoi_exitmap,
4570 .hwapic_irr_update = svm_hwapic_irr_update,
4571 .hwapic_isr_update = svm_hwapic_isr_update,
4572 .sync_pir_to_irr = kvm_lapic_find_highest_irr,
4573 .apicv_post_state_restore = avic_post_state_restore,
4575 .set_tss_addr = svm_set_tss_addr,
4576 .set_identity_map_addr = svm_set_identity_map_addr,
4577 .get_mt_mask = svm_get_mt_mask,
4579 .get_exit_info = svm_get_exit_info,
4581 .vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
4583 .has_wbinvd_exit = svm_has_wbinvd_exit,
4585 .write_l1_tsc_offset = svm_write_l1_tsc_offset,
4587 .load_mmu_pgd = svm_load_mmu_pgd,
4589 .check_intercept = svm_check_intercept,
4590 .handle_exit_irqoff = svm_handle_exit_irqoff,
4592 .request_immediate_exit = __kvm_request_immediate_exit,
4594 .sched_in = svm_sched_in,
4596 .pmu_ops = &amd_pmu_ops,
4597 .nested_ops = &svm_nested_ops,
4599 .deliver_posted_interrupt = svm_deliver_avic_intr,
4600 .dy_apicv_has_pending_interrupt = svm_dy_apicv_has_pending_interrupt,
4601 .update_pi_irte = svm_update_pi_irte,
4602 .setup_mce = svm_setup_mce,
4604 .smi_allowed = svm_smi_allowed,
4605 .pre_enter_smm = svm_pre_enter_smm,
4606 .pre_leave_smm = svm_pre_leave_smm,
4607 .enable_smi_window = svm_enable_smi_window,
4609 .mem_enc_op = svm_mem_enc_op,
4610 .mem_enc_reg_region = svm_register_enc_region,
4611 .mem_enc_unreg_region = svm_unregister_enc_region,
4613 .can_emulate_instruction = svm_can_emulate_instruction,
4615 .apic_init_signal_blocked = svm_apic_init_signal_blocked,
4617 .msr_filter_changed = svm_msr_filter_changed,
4618 .complete_emulated_msr = svm_complete_emulated_msr,
4620 .vcpu_deliver_sipi_vector = svm_vcpu_deliver_sipi_vector,
4623 static struct kvm_x86_init_ops svm_init_ops __initdata = {
4624 .cpu_has_kvm_support = has_svm,
4625 .disabled_by_bios = is_disabled,
4626 .hardware_setup = svm_hardware_setup,
4627 .check_processor_compatibility = svm_check_processor_compat,
4629 .runtime_ops = &svm_x86_ops,
4632 static int __init svm_init(void)
4634 __unused_size_checks();
4636 return kvm_init(&svm_init_ops, sizeof(struct vcpu_svm),
4637 __alignof__(struct vcpu_svm), THIS_MODULE);
4640 static void __exit svm_exit(void)
4645 module_init(svm_init)
4646 module_exit(svm_exit)