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);
930 /* Enable INVPCID feature */
931 kvm_cpu_cap_check_and_set(X86_FEATURE_INVPCID);
934 static __init int svm_hardware_setup(void)
937 struct page *iopm_pages;
941 iopm_pages = alloc_pages(GFP_KERNEL, IOPM_ALLOC_ORDER);
946 iopm_va = page_address(iopm_pages);
947 memset(iopm_va, 0xff, PAGE_SIZE * (1 << IOPM_ALLOC_ORDER));
948 iopm_base = page_to_pfn(iopm_pages) << PAGE_SHIFT;
950 init_msrpm_offsets();
952 supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
954 if (boot_cpu_has(X86_FEATURE_NX))
955 kvm_enable_efer_bits(EFER_NX);
957 if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
958 kvm_enable_efer_bits(EFER_FFXSR);
960 if (boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
961 kvm_has_tsc_control = true;
962 kvm_max_tsc_scaling_ratio = TSC_RATIO_MAX;
963 kvm_tsc_scaling_ratio_frac_bits = 32;
966 /* Check for pause filtering support */
967 if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
968 pause_filter_count = 0;
969 pause_filter_thresh = 0;
970 } else if (!boot_cpu_has(X86_FEATURE_PFTHRESHOLD)) {
971 pause_filter_thresh = 0;
975 printk(KERN_INFO "kvm: Nested Virtualization enabled\n");
976 kvm_enable_efer_bits(EFER_SVME | EFER_LMSLE);
979 if (IS_ENABLED(CONFIG_KVM_AMD_SEV) && sev) {
980 sev_hardware_setup();
986 svm_adjust_mmio_mask();
988 for_each_possible_cpu(cpu) {
989 r = svm_cpu_init(cpu);
994 if (!boot_cpu_has(X86_FEATURE_NPT))
997 if (npt_enabled && !npt)
1000 kvm_configure_mmu(npt_enabled, get_max_npt_level(), PG_LEVEL_1G);
1001 pr_info("kvm: Nested Paging %sabled\n", npt_enabled ? "en" : "dis");
1004 if (!boot_cpu_has(X86_FEATURE_NRIPS))
1010 !boot_cpu_has(X86_FEATURE_AVIC) ||
1011 !IS_ENABLED(CONFIG_X86_LOCAL_APIC)) {
1014 pr_info("AVIC enabled\n");
1016 amd_iommu_register_ga_log_notifier(&avic_ga_log_notifier);
1022 !boot_cpu_has(X86_FEATURE_V_VMSAVE_VMLOAD) ||
1023 !IS_ENABLED(CONFIG_X86_64)) {
1026 pr_info("Virtual VMLOAD VMSAVE supported\n");
1030 if (boot_cpu_has(X86_FEATURE_SVME_ADDR_CHK))
1031 svm_gp_erratum_intercept = false;
1034 if (!boot_cpu_has(X86_FEATURE_VGIF))
1037 pr_info("Virtual GIF supported\n");
1043 * It seems that on AMD processors PTE's accessed bit is
1044 * being set by the CPU hardware before the NPF vmexit.
1045 * This is not expected behaviour and our tests fail because
1047 * A workaround here is to disable support for
1048 * GUEST_MAXPHYADDR < HOST_MAXPHYADDR if NPT is enabled.
1049 * In this case userspace can know if there is support using
1050 * KVM_CAP_SMALLER_MAXPHYADDR extension and decide how to handle
1052 * If future AMD CPU models change the behaviour described above,
1053 * this variable can be changed accordingly
1055 allow_smaller_maxphyaddr = !npt_enabled;
1060 svm_hardware_teardown();
1064 static void init_seg(struct vmcb_seg *seg)
1067 seg->attrib = SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK |
1068 SVM_SELECTOR_WRITE_MASK; /* Read/Write Data Segment */
1069 seg->limit = 0xffff;
1073 static void init_sys_seg(struct vmcb_seg *seg, uint32_t type)
1076 seg->attrib = SVM_SELECTOR_P_MASK | type;
1077 seg->limit = 0xffff;
1081 static u64 svm_write_l1_tsc_offset(struct kvm_vcpu *vcpu, u64 offset)
1083 struct vcpu_svm *svm = to_svm(vcpu);
1084 u64 g_tsc_offset = 0;
1086 if (is_guest_mode(vcpu)) {
1087 /* Write L1's TSC offset. */
1088 g_tsc_offset = svm->vmcb->control.tsc_offset -
1089 svm->nested.hsave->control.tsc_offset;
1090 svm->nested.hsave->control.tsc_offset = offset;
1093 trace_kvm_write_tsc_offset(vcpu->vcpu_id,
1094 svm->vmcb->control.tsc_offset - g_tsc_offset,
1097 svm->vmcb->control.tsc_offset = offset + g_tsc_offset;
1099 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1100 return svm->vmcb->control.tsc_offset;
1103 static void svm_check_invpcid(struct vcpu_svm *svm)
1106 * Intercept INVPCID instruction only if shadow page table is
1107 * enabled. Interception is not required with nested page table
1110 if (kvm_cpu_cap_has(X86_FEATURE_INVPCID)) {
1112 svm_set_intercept(svm, INTERCEPT_INVPCID);
1114 svm_clr_intercept(svm, INTERCEPT_INVPCID);
1118 static void init_vmcb(struct vcpu_svm *svm)
1120 struct vmcb_control_area *control = &svm->vmcb->control;
1121 struct vmcb_save_area *save = &svm->vmcb->save;
1123 svm->vcpu.arch.hflags = 0;
1125 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1126 svm_set_intercept(svm, INTERCEPT_CR3_READ);
1127 svm_set_intercept(svm, INTERCEPT_CR4_READ);
1128 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1129 svm_set_intercept(svm, INTERCEPT_CR3_WRITE);
1130 svm_set_intercept(svm, INTERCEPT_CR4_WRITE);
1131 if (!kvm_vcpu_apicv_active(&svm->vcpu))
1132 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
1134 set_dr_intercepts(svm);
1136 set_exception_intercept(svm, PF_VECTOR);
1137 set_exception_intercept(svm, UD_VECTOR);
1138 set_exception_intercept(svm, MC_VECTOR);
1139 set_exception_intercept(svm, AC_VECTOR);
1140 set_exception_intercept(svm, DB_VECTOR);
1142 * Guest access to VMware backdoor ports could legitimately
1143 * trigger #GP because of TSS I/O permission bitmap.
1144 * We intercept those #GP and allow access to them anyway
1147 if (enable_vmware_backdoor)
1148 set_exception_intercept(svm, GP_VECTOR);
1150 svm_set_intercept(svm, INTERCEPT_INTR);
1151 svm_set_intercept(svm, INTERCEPT_NMI);
1152 svm_set_intercept(svm, INTERCEPT_SMI);
1153 svm_set_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1154 svm_set_intercept(svm, INTERCEPT_RDPMC);
1155 svm_set_intercept(svm, INTERCEPT_CPUID);
1156 svm_set_intercept(svm, INTERCEPT_INVD);
1157 svm_set_intercept(svm, INTERCEPT_INVLPG);
1158 svm_set_intercept(svm, INTERCEPT_INVLPGA);
1159 svm_set_intercept(svm, INTERCEPT_IOIO_PROT);
1160 svm_set_intercept(svm, INTERCEPT_MSR_PROT);
1161 svm_set_intercept(svm, INTERCEPT_TASK_SWITCH);
1162 svm_set_intercept(svm, INTERCEPT_SHUTDOWN);
1163 svm_set_intercept(svm, INTERCEPT_VMRUN);
1164 svm_set_intercept(svm, INTERCEPT_VMMCALL);
1165 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1166 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1167 svm_set_intercept(svm, INTERCEPT_STGI);
1168 svm_set_intercept(svm, INTERCEPT_CLGI);
1169 svm_set_intercept(svm, INTERCEPT_SKINIT);
1170 svm_set_intercept(svm, INTERCEPT_WBINVD);
1171 svm_set_intercept(svm, INTERCEPT_XSETBV);
1172 svm_set_intercept(svm, INTERCEPT_RDPRU);
1173 svm_set_intercept(svm, INTERCEPT_RSM);
1175 if (!kvm_mwait_in_guest(svm->vcpu.kvm)) {
1176 svm_set_intercept(svm, INTERCEPT_MONITOR);
1177 svm_set_intercept(svm, INTERCEPT_MWAIT);
1180 if (!kvm_hlt_in_guest(svm->vcpu.kvm))
1181 svm_set_intercept(svm, INTERCEPT_HLT);
1183 control->iopm_base_pa = __sme_set(iopm_base);
1184 control->msrpm_base_pa = __sme_set(__pa(svm->msrpm));
1185 control->int_ctl = V_INTR_MASKING_MASK;
1187 init_seg(&save->es);
1188 init_seg(&save->ss);
1189 init_seg(&save->ds);
1190 init_seg(&save->fs);
1191 init_seg(&save->gs);
1193 save->cs.selector = 0xf000;
1194 save->cs.base = 0xffff0000;
1195 /* Executable/Readable Code Segment */
1196 save->cs.attrib = SVM_SELECTOR_READ_MASK | SVM_SELECTOR_P_MASK |
1197 SVM_SELECTOR_S_MASK | SVM_SELECTOR_CODE_MASK;
1198 save->cs.limit = 0xffff;
1200 save->gdtr.limit = 0xffff;
1201 save->idtr.limit = 0xffff;
1203 init_sys_seg(&save->ldtr, SEG_TYPE_LDT);
1204 init_sys_seg(&save->tr, SEG_TYPE_BUSY_TSS16);
1206 svm_set_efer(&svm->vcpu, 0);
1207 save->dr6 = 0xffff0ff0;
1208 kvm_set_rflags(&svm->vcpu, X86_EFLAGS_FIXED);
1209 save->rip = 0x0000fff0;
1210 svm->vcpu.arch.regs[VCPU_REGS_RIP] = save->rip;
1213 * svm_set_cr0() sets PG and WP and clears NW and CD on save->cr0.
1214 * It also updates the guest-visible cr0 value.
1216 svm_set_cr0(&svm->vcpu, X86_CR0_NW | X86_CR0_CD | X86_CR0_ET);
1217 kvm_mmu_reset_context(&svm->vcpu);
1219 save->cr4 = X86_CR4_PAE;
1223 /* Setup VMCB for Nested Paging */
1224 control->nested_ctl |= SVM_NESTED_CTL_NP_ENABLE;
1225 svm_clr_intercept(svm, INTERCEPT_INVLPG);
1226 clr_exception_intercept(svm, PF_VECTOR);
1227 svm_clr_intercept(svm, INTERCEPT_CR3_READ);
1228 svm_clr_intercept(svm, INTERCEPT_CR3_WRITE);
1229 save->g_pat = svm->vcpu.arch.pat;
1233 svm->asid_generation = 0;
1236 svm->nested.vmcb12_gpa = 0;
1237 svm->vcpu.arch.hflags = 0;
1239 if (!kvm_pause_in_guest(svm->vcpu.kvm)) {
1240 control->pause_filter_count = pause_filter_count;
1241 if (pause_filter_thresh)
1242 control->pause_filter_thresh = pause_filter_thresh;
1243 svm_set_intercept(svm, INTERCEPT_PAUSE);
1245 svm_clr_intercept(svm, INTERCEPT_PAUSE);
1248 svm_check_invpcid(svm);
1250 if (kvm_vcpu_apicv_active(&svm->vcpu))
1251 avic_init_vmcb(svm);
1254 * If hardware supports Virtual VMLOAD VMSAVE then enable it
1255 * in VMCB and clear intercepts to avoid #VMEXIT.
1258 svm_clr_intercept(svm, INTERCEPT_VMLOAD);
1259 svm_clr_intercept(svm, INTERCEPT_VMSAVE);
1260 svm->vmcb->control.virt_ext |= VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
1264 svm_clr_intercept(svm, INTERCEPT_STGI);
1265 svm_clr_intercept(svm, INTERCEPT_CLGI);
1266 svm->vmcb->control.int_ctl |= V_GIF_ENABLE_MASK;
1269 if (sev_guest(svm->vcpu.kvm)) {
1270 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
1271 clr_exception_intercept(svm, UD_VECTOR);
1273 if (sev_es_guest(svm->vcpu.kvm)) {
1274 /* Perform SEV-ES specific VMCB updates */
1275 sev_es_init_vmcb(svm);
1279 vmcb_mark_all_dirty(svm->vmcb);
1285 static void svm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
1287 struct vcpu_svm *svm = to_svm(vcpu);
1292 svm->virt_spec_ctrl = 0;
1295 svm->vcpu.arch.apic_base = APIC_DEFAULT_PHYS_BASE |
1296 MSR_IA32_APICBASE_ENABLE;
1297 if (kvm_vcpu_is_reset_bsp(&svm->vcpu))
1298 svm->vcpu.arch.apic_base |= MSR_IA32_APICBASE_BSP;
1302 kvm_cpuid(vcpu, &eax, &dummy, &dummy, &dummy, false);
1303 kvm_rdx_write(vcpu, eax);
1305 if (kvm_vcpu_apicv_active(vcpu) && !init_event)
1306 avic_update_vapic_bar(svm, APIC_DEFAULT_PHYS_BASE);
1309 static int svm_create_vcpu(struct kvm_vcpu *vcpu)
1311 struct vcpu_svm *svm;
1312 struct page *vmcb_page;
1313 struct page *vmsa_page = NULL;
1316 BUILD_BUG_ON(offsetof(struct vcpu_svm, vcpu) != 0);
1320 vmcb_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1324 if (sev_es_guest(svm->vcpu.kvm)) {
1326 * SEV-ES guests require a separate VMSA page used to contain
1327 * the encrypted register state of the guest.
1329 vmsa_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1331 goto error_free_vmcb_page;
1334 * SEV-ES guests maintain an encrypted version of their FPU
1335 * state which is restored and saved on VMRUN and VMEXIT.
1336 * Free the fpu structure to prevent KVM from attempting to
1337 * access the FPU state.
1339 kvm_free_guest_fpu(vcpu);
1342 err = avic_init_vcpu(svm);
1344 goto error_free_vmsa_page;
1346 /* We initialize this flag to true to make sure that the is_running
1347 * bit would be set the first time the vcpu is loaded.
1349 if (irqchip_in_kernel(vcpu->kvm) && kvm_apicv_activated(vcpu->kvm))
1350 svm->avic_is_running = true;
1352 svm->msrpm = svm_vcpu_alloc_msrpm();
1355 goto error_free_vmsa_page;
1358 svm_vcpu_init_msrpm(vcpu, svm->msrpm);
1360 svm->vmcb = page_address(vmcb_page);
1361 svm->vmcb_pa = __sme_set(page_to_pfn(vmcb_page) << PAGE_SHIFT);
1364 svm->vmsa = page_address(vmsa_page);
1366 svm->asid_generation = 0;
1367 svm->guest_state_loaded = false;
1370 svm_init_osvw(vcpu);
1371 vcpu->arch.microcode_version = 0x01000065;
1373 if (sev_es_guest(svm->vcpu.kvm))
1374 /* Perform SEV-ES specific VMCB creation updates */
1375 sev_es_create_vcpu(svm);
1379 error_free_vmsa_page:
1381 __free_page(vmsa_page);
1382 error_free_vmcb_page:
1383 __free_page(vmcb_page);
1388 static void svm_clear_current_vmcb(struct vmcb *vmcb)
1392 for_each_online_cpu(i)
1393 cmpxchg(&per_cpu(svm_data, i)->current_vmcb, vmcb, NULL);
1396 static void svm_free_vcpu(struct kvm_vcpu *vcpu)
1398 struct vcpu_svm *svm = to_svm(vcpu);
1401 * The vmcb page can be recycled, causing a false negative in
1402 * svm_vcpu_load(). So, ensure that no logical CPU has this
1403 * vmcb page recorded as its current vmcb.
1405 svm_clear_current_vmcb(svm->vmcb);
1407 svm_free_nested(svm);
1409 sev_free_vcpu(vcpu);
1411 __free_page(pfn_to_page(__sme_clr(svm->vmcb_pa) >> PAGE_SHIFT));
1412 __free_pages(virt_to_page(svm->msrpm), MSRPM_ALLOC_ORDER);
1415 static void svm_prepare_guest_switch(struct kvm_vcpu *vcpu)
1417 struct vcpu_svm *svm = to_svm(vcpu);
1418 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
1421 if (svm->guest_state_loaded)
1425 * Certain MSRs are restored on VMEXIT (sev-es), or vmload of host save
1426 * area (non-sev-es). Save ones that aren't so we can restore them
1427 * individually later.
1429 for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
1430 rdmsrl(host_save_user_msrs[i], svm->host_user_msrs[i]);
1433 * Save additional host state that will be restored on VMEXIT (sev-es)
1434 * or subsequent vmload of host save area.
1436 if (sev_es_guest(svm->vcpu.kvm)) {
1437 sev_es_prepare_guest_switch(svm, vcpu->cpu);
1439 vmsave(__sme_page_pa(sd->save_area));
1442 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
1443 u64 tsc_ratio = vcpu->arch.tsc_scaling_ratio;
1444 if (tsc_ratio != __this_cpu_read(current_tsc_ratio)) {
1445 __this_cpu_write(current_tsc_ratio, tsc_ratio);
1446 wrmsrl(MSR_AMD64_TSC_RATIO, tsc_ratio);
1450 /* This assumes that the kernel never uses MSR_TSC_AUX */
1451 if (static_cpu_has(X86_FEATURE_RDTSCP))
1452 wrmsrl(MSR_TSC_AUX, svm->tsc_aux);
1454 svm->guest_state_loaded = true;
1457 static void svm_prepare_host_switch(struct kvm_vcpu *vcpu)
1459 struct vcpu_svm *svm = to_svm(vcpu);
1462 if (!svm->guest_state_loaded)
1466 * Certain MSRs are restored on VMEXIT (sev-es), or vmload of host save
1467 * area (non-sev-es). Restore the ones that weren't.
1469 for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
1470 wrmsrl(host_save_user_msrs[i], svm->host_user_msrs[i]);
1472 svm->guest_state_loaded = false;
1475 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1477 struct vcpu_svm *svm = to_svm(vcpu);
1478 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1480 if (unlikely(cpu != vcpu->cpu)) {
1481 svm->asid_generation = 0;
1482 vmcb_mark_all_dirty(svm->vmcb);
1485 if (sd->current_vmcb != svm->vmcb) {
1486 sd->current_vmcb = svm->vmcb;
1487 indirect_branch_prediction_barrier();
1489 avic_vcpu_load(vcpu, cpu);
1492 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
1494 avic_vcpu_put(vcpu);
1495 svm_prepare_host_switch(vcpu);
1497 ++vcpu->stat.host_state_reload;
1500 static unsigned long svm_get_rflags(struct kvm_vcpu *vcpu)
1502 struct vcpu_svm *svm = to_svm(vcpu);
1503 unsigned long rflags = svm->vmcb->save.rflags;
1505 if (svm->nmi_singlestep) {
1506 /* Hide our flags if they were not set by the guest */
1507 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
1508 rflags &= ~X86_EFLAGS_TF;
1509 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
1510 rflags &= ~X86_EFLAGS_RF;
1515 static void svm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
1517 if (to_svm(vcpu)->nmi_singlestep)
1518 rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
1521 * Any change of EFLAGS.VM is accompanied by a reload of SS
1522 * (caused by either a task switch or an inter-privilege IRET),
1523 * so we do not need to update the CPL here.
1525 to_svm(vcpu)->vmcb->save.rflags = rflags;
1528 static void svm_cache_reg(struct kvm_vcpu *vcpu, enum kvm_reg reg)
1531 case VCPU_EXREG_PDPTR:
1532 BUG_ON(!npt_enabled);
1533 load_pdptrs(vcpu, vcpu->arch.walk_mmu, kvm_read_cr3(vcpu));
1540 static void svm_set_vintr(struct vcpu_svm *svm)
1542 struct vmcb_control_area *control;
1544 /* The following fields are ignored when AVIC is enabled */
1545 WARN_ON(kvm_vcpu_apicv_active(&svm->vcpu));
1546 svm_set_intercept(svm, INTERCEPT_VINTR);
1549 * This is just a dummy VINTR to actually cause a vmexit to happen.
1550 * Actual injection of virtual interrupts happens through EVENTINJ.
1552 control = &svm->vmcb->control;
1553 control->int_vector = 0x0;
1554 control->int_ctl &= ~V_INTR_PRIO_MASK;
1555 control->int_ctl |= V_IRQ_MASK |
1556 ((/*control->int_vector >> 4*/ 0xf) << V_INTR_PRIO_SHIFT);
1557 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1560 static void svm_clear_vintr(struct vcpu_svm *svm)
1562 const u32 mask = V_TPR_MASK | V_GIF_ENABLE_MASK | V_GIF_MASK | V_INTR_MASKING_MASK;
1563 svm_clr_intercept(svm, INTERCEPT_VINTR);
1565 /* Drop int_ctl fields related to VINTR injection. */
1566 svm->vmcb->control.int_ctl &= mask;
1567 if (is_guest_mode(&svm->vcpu)) {
1568 svm->nested.hsave->control.int_ctl &= mask;
1570 WARN_ON((svm->vmcb->control.int_ctl & V_TPR_MASK) !=
1571 (svm->nested.ctl.int_ctl & V_TPR_MASK));
1572 svm->vmcb->control.int_ctl |= svm->nested.ctl.int_ctl & ~mask;
1575 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1578 static struct vmcb_seg *svm_seg(struct kvm_vcpu *vcpu, int seg)
1580 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1583 case VCPU_SREG_CS: return &save->cs;
1584 case VCPU_SREG_DS: return &save->ds;
1585 case VCPU_SREG_ES: return &save->es;
1586 case VCPU_SREG_FS: return &save->fs;
1587 case VCPU_SREG_GS: return &save->gs;
1588 case VCPU_SREG_SS: return &save->ss;
1589 case VCPU_SREG_TR: return &save->tr;
1590 case VCPU_SREG_LDTR: return &save->ldtr;
1596 static u64 svm_get_segment_base(struct kvm_vcpu *vcpu, int seg)
1598 struct vmcb_seg *s = svm_seg(vcpu, seg);
1603 static void svm_get_segment(struct kvm_vcpu *vcpu,
1604 struct kvm_segment *var, int seg)
1606 struct vmcb_seg *s = svm_seg(vcpu, seg);
1608 var->base = s->base;
1609 var->limit = s->limit;
1610 var->selector = s->selector;
1611 var->type = s->attrib & SVM_SELECTOR_TYPE_MASK;
1612 var->s = (s->attrib >> SVM_SELECTOR_S_SHIFT) & 1;
1613 var->dpl = (s->attrib >> SVM_SELECTOR_DPL_SHIFT) & 3;
1614 var->present = (s->attrib >> SVM_SELECTOR_P_SHIFT) & 1;
1615 var->avl = (s->attrib >> SVM_SELECTOR_AVL_SHIFT) & 1;
1616 var->l = (s->attrib >> SVM_SELECTOR_L_SHIFT) & 1;
1617 var->db = (s->attrib >> SVM_SELECTOR_DB_SHIFT) & 1;
1620 * AMD CPUs circa 2014 track the G bit for all segments except CS.
1621 * However, the SVM spec states that the G bit is not observed by the
1622 * CPU, and some VMware virtual CPUs drop the G bit for all segments.
1623 * So let's synthesize a legal G bit for all segments, this helps
1624 * running KVM nested. It also helps cross-vendor migration, because
1625 * Intel's vmentry has a check on the 'G' bit.
1627 var->g = s->limit > 0xfffff;
1630 * AMD's VMCB does not have an explicit unusable field, so emulate it
1631 * for cross vendor migration purposes by "not present"
1633 var->unusable = !var->present;
1638 * Work around a bug where the busy flag in the tr selector
1648 * The accessed bit must always be set in the segment
1649 * descriptor cache, although it can be cleared in the
1650 * descriptor, the cached bit always remains at 1. Since
1651 * Intel has a check on this, set it here to support
1652 * cross-vendor migration.
1659 * On AMD CPUs sometimes the DB bit in the segment
1660 * descriptor is left as 1, although the whole segment has
1661 * been made unusable. Clear it here to pass an Intel VMX
1662 * entry check when cross vendor migrating.
1666 /* This is symmetric with svm_set_segment() */
1667 var->dpl = to_svm(vcpu)->vmcb->save.cpl;
1672 static int svm_get_cpl(struct kvm_vcpu *vcpu)
1674 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1679 static void svm_get_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1681 struct vcpu_svm *svm = to_svm(vcpu);
1683 dt->size = svm->vmcb->save.idtr.limit;
1684 dt->address = svm->vmcb->save.idtr.base;
1687 static void svm_set_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1689 struct vcpu_svm *svm = to_svm(vcpu);
1691 svm->vmcb->save.idtr.limit = dt->size;
1692 svm->vmcb->save.idtr.base = dt->address ;
1693 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1696 static void svm_get_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1698 struct vcpu_svm *svm = to_svm(vcpu);
1700 dt->size = svm->vmcb->save.gdtr.limit;
1701 dt->address = svm->vmcb->save.gdtr.base;
1704 static void svm_set_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1706 struct vcpu_svm *svm = to_svm(vcpu);
1708 svm->vmcb->save.gdtr.limit = dt->size;
1709 svm->vmcb->save.gdtr.base = dt->address ;
1710 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1713 static void update_cr0_intercept(struct vcpu_svm *svm)
1719 * SEV-ES guests must always keep the CR intercepts cleared. CR
1720 * tracking is done using the CR write traps.
1722 if (sev_es_guest(svm->vcpu.kvm))
1725 gcr0 = svm->vcpu.arch.cr0;
1726 hcr0 = &svm->vmcb->save.cr0;
1727 *hcr0 = (*hcr0 & ~SVM_CR0_SELECTIVE_MASK)
1728 | (gcr0 & SVM_CR0_SELECTIVE_MASK);
1730 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1732 if (gcr0 == *hcr0) {
1733 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1734 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1736 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1737 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1741 void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1743 struct vcpu_svm *svm = to_svm(vcpu);
1745 #ifdef CONFIG_X86_64
1746 if (vcpu->arch.efer & EFER_LME && !vcpu->arch.guest_state_protected) {
1747 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
1748 vcpu->arch.efer |= EFER_LMA;
1749 svm->vmcb->save.efer |= EFER_LMA | EFER_LME;
1752 if (is_paging(vcpu) && !(cr0 & X86_CR0_PG)) {
1753 vcpu->arch.efer &= ~EFER_LMA;
1754 svm->vmcb->save.efer &= ~(EFER_LMA | EFER_LME);
1758 vcpu->arch.cr0 = cr0;
1761 cr0 |= X86_CR0_PG | X86_CR0_WP;
1764 * re-enable caching here because the QEMU bios
1765 * does not do it - this results in some delay at
1768 if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
1769 cr0 &= ~(X86_CR0_CD | X86_CR0_NW);
1770 svm->vmcb->save.cr0 = cr0;
1771 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1772 update_cr0_intercept(svm);
1775 static bool svm_is_valid_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1780 void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1782 unsigned long host_cr4_mce = cr4_read_shadow() & X86_CR4_MCE;
1783 unsigned long old_cr4 = vcpu->arch.cr4;
1785 if (npt_enabled && ((old_cr4 ^ cr4) & X86_CR4_PGE))
1786 svm_flush_tlb(vcpu);
1788 vcpu->arch.cr4 = cr4;
1791 cr4 |= host_cr4_mce;
1792 to_svm(vcpu)->vmcb->save.cr4 = cr4;
1793 vmcb_mark_dirty(to_svm(vcpu)->vmcb, VMCB_CR);
1795 if ((cr4 ^ old_cr4) & (X86_CR4_OSXSAVE | X86_CR4_PKE))
1796 kvm_update_cpuid_runtime(vcpu);
1799 static void svm_set_segment(struct kvm_vcpu *vcpu,
1800 struct kvm_segment *var, int seg)
1802 struct vcpu_svm *svm = to_svm(vcpu);
1803 struct vmcb_seg *s = svm_seg(vcpu, seg);
1805 s->base = var->base;
1806 s->limit = var->limit;
1807 s->selector = var->selector;
1808 s->attrib = (var->type & SVM_SELECTOR_TYPE_MASK);
1809 s->attrib |= (var->s & 1) << SVM_SELECTOR_S_SHIFT;
1810 s->attrib |= (var->dpl & 3) << SVM_SELECTOR_DPL_SHIFT;
1811 s->attrib |= ((var->present & 1) && !var->unusable) << SVM_SELECTOR_P_SHIFT;
1812 s->attrib |= (var->avl & 1) << SVM_SELECTOR_AVL_SHIFT;
1813 s->attrib |= (var->l & 1) << SVM_SELECTOR_L_SHIFT;
1814 s->attrib |= (var->db & 1) << SVM_SELECTOR_DB_SHIFT;
1815 s->attrib |= (var->g & 1) << SVM_SELECTOR_G_SHIFT;
1818 * This is always accurate, except if SYSRET returned to a segment
1819 * with SS.DPL != 3. Intel does not have this quirk, and always
1820 * forces SS.DPL to 3 on sysret, so we ignore that case; fixing it
1821 * would entail passing the CPL to userspace and back.
1823 if (seg == VCPU_SREG_SS)
1824 /* This is symmetric with svm_get_segment() */
1825 svm->vmcb->save.cpl = (var->dpl & 3);
1827 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
1830 static void svm_update_exception_bitmap(struct kvm_vcpu *vcpu)
1832 struct vcpu_svm *svm = to_svm(vcpu);
1834 clr_exception_intercept(svm, BP_VECTOR);
1836 if (vcpu->guest_debug & KVM_GUESTDBG_ENABLE) {
1837 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP)
1838 set_exception_intercept(svm, BP_VECTOR);
1842 static void new_asid(struct vcpu_svm *svm, struct svm_cpu_data *sd)
1844 if (sd->next_asid > sd->max_asid) {
1845 ++sd->asid_generation;
1846 sd->next_asid = sd->min_asid;
1847 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
1848 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1851 svm->asid_generation = sd->asid_generation;
1852 svm->asid = sd->next_asid++;
1855 static void svm_set_dr6(struct vcpu_svm *svm, unsigned long value)
1857 struct vmcb *vmcb = svm->vmcb;
1859 if (svm->vcpu.arch.guest_state_protected)
1862 if (unlikely(value != vmcb->save.dr6)) {
1863 vmcb->save.dr6 = value;
1864 vmcb_mark_dirty(vmcb, VMCB_DR);
1868 static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
1870 struct vcpu_svm *svm = to_svm(vcpu);
1872 if (vcpu->arch.guest_state_protected)
1875 get_debugreg(vcpu->arch.db[0], 0);
1876 get_debugreg(vcpu->arch.db[1], 1);
1877 get_debugreg(vcpu->arch.db[2], 2);
1878 get_debugreg(vcpu->arch.db[3], 3);
1880 * We cannot reset svm->vmcb->save.dr6 to DR6_ACTIVE_LOW here,
1881 * because db_interception might need it. We can do it before vmentry.
1883 vcpu->arch.dr6 = svm->vmcb->save.dr6;
1884 vcpu->arch.dr7 = svm->vmcb->save.dr7;
1885 vcpu->arch.switch_db_regs &= ~KVM_DEBUGREG_WONT_EXIT;
1886 set_dr_intercepts(svm);
1889 static void svm_set_dr7(struct kvm_vcpu *vcpu, unsigned long value)
1891 struct vcpu_svm *svm = to_svm(vcpu);
1893 if (vcpu->arch.guest_state_protected)
1896 svm->vmcb->save.dr7 = value;
1897 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
1900 static int pf_interception(struct vcpu_svm *svm)
1902 u64 fault_address = __sme_clr(svm->vmcb->control.exit_info_2);
1903 u64 error_code = svm->vmcb->control.exit_info_1;
1905 return kvm_handle_page_fault(&svm->vcpu, error_code, fault_address,
1906 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1907 svm->vmcb->control.insn_bytes : NULL,
1908 svm->vmcb->control.insn_len);
1911 static int npf_interception(struct vcpu_svm *svm)
1913 u64 fault_address = __sme_clr(svm->vmcb->control.exit_info_2);
1914 u64 error_code = svm->vmcb->control.exit_info_1;
1916 trace_kvm_page_fault(fault_address, error_code);
1917 return kvm_mmu_page_fault(&svm->vcpu, fault_address, error_code,
1918 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1919 svm->vmcb->control.insn_bytes : NULL,
1920 svm->vmcb->control.insn_len);
1923 static int db_interception(struct vcpu_svm *svm)
1925 struct kvm_run *kvm_run = svm->vcpu.run;
1926 struct kvm_vcpu *vcpu = &svm->vcpu;
1928 if (!(svm->vcpu.guest_debug &
1929 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) &&
1930 !svm->nmi_singlestep) {
1931 u32 payload = svm->vmcb->save.dr6 ^ DR6_ACTIVE_LOW;
1932 kvm_queue_exception_p(&svm->vcpu, DB_VECTOR, payload);
1936 if (svm->nmi_singlestep) {
1937 disable_nmi_singlestep(svm);
1938 /* Make sure we check for pending NMIs upon entry */
1939 kvm_make_request(KVM_REQ_EVENT, vcpu);
1942 if (svm->vcpu.guest_debug &
1943 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) {
1944 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1945 kvm_run->debug.arch.dr6 = svm->vmcb->save.dr6;
1946 kvm_run->debug.arch.dr7 = svm->vmcb->save.dr7;
1947 kvm_run->debug.arch.pc =
1948 svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1949 kvm_run->debug.arch.exception = DB_VECTOR;
1956 static int bp_interception(struct vcpu_svm *svm)
1958 struct kvm_run *kvm_run = svm->vcpu.run;
1960 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1961 kvm_run->debug.arch.pc = svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1962 kvm_run->debug.arch.exception = BP_VECTOR;
1966 static int ud_interception(struct vcpu_svm *svm)
1968 return handle_ud(&svm->vcpu);
1971 static int ac_interception(struct vcpu_svm *svm)
1973 kvm_queue_exception_e(&svm->vcpu, AC_VECTOR, 0);
1977 static bool is_erratum_383(void)
1982 if (!erratum_383_found)
1985 value = native_read_msr_safe(MSR_IA32_MC0_STATUS, &err);
1989 /* Bit 62 may or may not be set for this mce */
1990 value &= ~(1ULL << 62);
1992 if (value != 0xb600000000010015ULL)
1995 /* Clear MCi_STATUS registers */
1996 for (i = 0; i < 6; ++i)
1997 native_write_msr_safe(MSR_IA32_MCx_STATUS(i), 0, 0);
1999 value = native_read_msr_safe(MSR_IA32_MCG_STATUS, &err);
2003 value &= ~(1ULL << 2);
2004 low = lower_32_bits(value);
2005 high = upper_32_bits(value);
2007 native_write_msr_safe(MSR_IA32_MCG_STATUS, low, high);
2010 /* Flush tlb to evict multi-match entries */
2016 static void svm_handle_mce(struct vcpu_svm *svm)
2018 if (is_erratum_383()) {
2020 * Erratum 383 triggered. Guest state is corrupt so kill the
2023 pr_err("KVM: Guest triggered AMD Erratum 383\n");
2025 kvm_make_request(KVM_REQ_TRIPLE_FAULT, &svm->vcpu);
2031 * On an #MC intercept the MCE handler is not called automatically in
2032 * the host. So do it by hand here.
2034 kvm_machine_check();
2037 static int mc_interception(struct vcpu_svm *svm)
2042 static int shutdown_interception(struct vcpu_svm *svm)
2044 struct kvm_run *kvm_run = svm->vcpu.run;
2047 * The VM save area has already been encrypted so it
2048 * cannot be reinitialized - just terminate.
2050 if (sev_es_guest(svm->vcpu.kvm))
2054 * VMCB is undefined after a SHUTDOWN intercept
2055 * so reinitialize it.
2057 clear_page(svm->vmcb);
2060 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
2064 static int io_interception(struct vcpu_svm *svm)
2066 struct kvm_vcpu *vcpu = &svm->vcpu;
2067 u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
2068 int size, in, string;
2071 ++svm->vcpu.stat.io_exits;
2072 string = (io_info & SVM_IOIO_STR_MASK) != 0;
2073 in = (io_info & SVM_IOIO_TYPE_MASK) != 0;
2074 port = io_info >> 16;
2075 size = (io_info & SVM_IOIO_SIZE_MASK) >> SVM_IOIO_SIZE_SHIFT;
2078 if (sev_es_guest(vcpu->kvm))
2079 return sev_es_string_io(svm, size, port, in);
2081 return kvm_emulate_instruction(vcpu, 0);
2084 svm->next_rip = svm->vmcb->control.exit_info_2;
2086 return kvm_fast_pio(&svm->vcpu, size, port, in);
2089 static int nmi_interception(struct vcpu_svm *svm)
2094 static int intr_interception(struct vcpu_svm *svm)
2096 ++svm->vcpu.stat.irq_exits;
2100 static int nop_on_interception(struct vcpu_svm *svm)
2105 static int halt_interception(struct vcpu_svm *svm)
2107 return kvm_emulate_halt(&svm->vcpu);
2110 static int vmmcall_interception(struct vcpu_svm *svm)
2112 return kvm_emulate_hypercall(&svm->vcpu);
2115 static int vmload_interception(struct vcpu_svm *svm)
2117 struct vmcb *nested_vmcb;
2118 struct kvm_host_map map;
2121 if (nested_svm_check_permissions(svm))
2124 ret = kvm_vcpu_map(&svm->vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2127 kvm_inject_gp(&svm->vcpu, 0);
2131 nested_vmcb = map.hva;
2133 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2135 nested_svm_vmloadsave(nested_vmcb, svm->vmcb);
2136 kvm_vcpu_unmap(&svm->vcpu, &map, true);
2141 static int vmsave_interception(struct vcpu_svm *svm)
2143 struct vmcb *nested_vmcb;
2144 struct kvm_host_map map;
2147 if (nested_svm_check_permissions(svm))
2150 ret = kvm_vcpu_map(&svm->vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2153 kvm_inject_gp(&svm->vcpu, 0);
2157 nested_vmcb = map.hva;
2159 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2161 nested_svm_vmloadsave(svm->vmcb, nested_vmcb);
2162 kvm_vcpu_unmap(&svm->vcpu, &map, true);
2167 static int vmrun_interception(struct vcpu_svm *svm)
2169 if (nested_svm_check_permissions(svm))
2172 return nested_svm_vmrun(svm);
2182 /* Return NONE_SVM_INSTR if not SVM instrs, otherwise return decode result */
2183 static int svm_instr_opcode(struct kvm_vcpu *vcpu)
2185 struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt;
2187 if (ctxt->b != 0x1 || ctxt->opcode_len != 2)
2188 return NONE_SVM_INSTR;
2190 switch (ctxt->modrm) {
2191 case 0xd8: /* VMRUN */
2192 return SVM_INSTR_VMRUN;
2193 case 0xda: /* VMLOAD */
2194 return SVM_INSTR_VMLOAD;
2195 case 0xdb: /* VMSAVE */
2196 return SVM_INSTR_VMSAVE;
2201 return NONE_SVM_INSTR;
2204 static int emulate_svm_instr(struct kvm_vcpu *vcpu, int opcode)
2206 const int guest_mode_exit_codes[] = {
2207 [SVM_INSTR_VMRUN] = SVM_EXIT_VMRUN,
2208 [SVM_INSTR_VMLOAD] = SVM_EXIT_VMLOAD,
2209 [SVM_INSTR_VMSAVE] = SVM_EXIT_VMSAVE,
2211 int (*const svm_instr_handlers[])(struct vcpu_svm *svm) = {
2212 [SVM_INSTR_VMRUN] = vmrun_interception,
2213 [SVM_INSTR_VMLOAD] = vmload_interception,
2214 [SVM_INSTR_VMSAVE] = vmsave_interception,
2216 struct vcpu_svm *svm = to_svm(vcpu);
2218 if (is_guest_mode(vcpu)) {
2219 svm->vmcb->control.exit_code = guest_mode_exit_codes[opcode];
2220 svm->vmcb->control.exit_info_1 = 0;
2221 svm->vmcb->control.exit_info_2 = 0;
2223 return nested_svm_vmexit(svm);
2225 return svm_instr_handlers[opcode](svm);
2229 * #GP handling code. Note that #GP can be triggered under the following two
2231 * 1) SVM VM-related instructions (VMRUN/VMSAVE/VMLOAD) that trigger #GP on
2232 * some AMD CPUs when EAX of these instructions are in the reserved memory
2233 * regions (e.g. SMM memory on host).
2234 * 2) VMware backdoor
2236 static int gp_interception(struct vcpu_svm *svm)
2238 struct kvm_vcpu *vcpu = &svm->vcpu;
2239 u32 error_code = svm->vmcb->control.exit_info_1;
2242 /* Both #GP cases have zero error_code */
2246 /* Decode the instruction for usage later */
2247 if (x86_decode_emulated_instruction(vcpu, 0, NULL, 0) != EMULATION_OK)
2250 opcode = svm_instr_opcode(vcpu);
2252 if (opcode == NONE_SVM_INSTR) {
2253 if (!enable_vmware_backdoor)
2257 * VMware backdoor emulation on #GP interception only handles
2258 * IN{S}, OUT{S}, and RDPMC.
2260 if (!is_guest_mode(vcpu))
2261 return kvm_emulate_instruction(vcpu,
2262 EMULTYPE_VMWARE_GP | EMULTYPE_NO_DECODE);
2264 return emulate_svm_instr(vcpu, opcode);
2267 kvm_queue_exception_e(vcpu, GP_VECTOR, error_code);
2271 void svm_set_gif(struct vcpu_svm *svm, bool value)
2275 * If VGIF is enabled, the STGI intercept is only added to
2276 * detect the opening of the SMI/NMI window; remove it now.
2277 * Likewise, clear the VINTR intercept, we will set it
2278 * again while processing KVM_REQ_EVENT if needed.
2280 if (vgif_enabled(svm))
2281 svm_clr_intercept(svm, INTERCEPT_STGI);
2282 if (svm_is_intercept(svm, INTERCEPT_VINTR))
2283 svm_clear_vintr(svm);
2286 if (svm->vcpu.arch.smi_pending ||
2287 svm->vcpu.arch.nmi_pending ||
2288 kvm_cpu_has_injectable_intr(&svm->vcpu))
2289 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2294 * After a CLGI no interrupts should come. But if vGIF is
2295 * in use, we still rely on the VINTR intercept (rather than
2296 * STGI) to detect an open interrupt window.
2298 if (!vgif_enabled(svm))
2299 svm_clear_vintr(svm);
2303 static int stgi_interception(struct vcpu_svm *svm)
2307 if (nested_svm_check_permissions(svm))
2310 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2311 svm_set_gif(svm, true);
2315 static int clgi_interception(struct vcpu_svm *svm)
2319 if (nested_svm_check_permissions(svm))
2322 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2323 svm_set_gif(svm, false);
2327 static int invlpga_interception(struct vcpu_svm *svm)
2329 struct kvm_vcpu *vcpu = &svm->vcpu;
2331 trace_kvm_invlpga(svm->vmcb->save.rip, kvm_rcx_read(&svm->vcpu),
2332 kvm_rax_read(&svm->vcpu));
2334 /* Let's treat INVLPGA the same as INVLPG (can be optimized!) */
2335 kvm_mmu_invlpg(vcpu, kvm_rax_read(&svm->vcpu));
2337 return kvm_skip_emulated_instruction(&svm->vcpu);
2340 static int skinit_interception(struct vcpu_svm *svm)
2342 trace_kvm_skinit(svm->vmcb->save.rip, kvm_rax_read(&svm->vcpu));
2344 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2348 static int wbinvd_interception(struct vcpu_svm *svm)
2350 return kvm_emulate_wbinvd(&svm->vcpu);
2353 static int xsetbv_interception(struct vcpu_svm *svm)
2355 u64 new_bv = kvm_read_edx_eax(&svm->vcpu);
2356 u32 index = kvm_rcx_read(&svm->vcpu);
2358 int err = kvm_set_xcr(&svm->vcpu, index, new_bv);
2359 return kvm_complete_insn_gp(&svm->vcpu, err);
2362 static int rdpru_interception(struct vcpu_svm *svm)
2364 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2368 static int task_switch_interception(struct vcpu_svm *svm)
2372 int int_type = svm->vmcb->control.exit_int_info &
2373 SVM_EXITINTINFO_TYPE_MASK;
2374 int int_vec = svm->vmcb->control.exit_int_info & SVM_EVTINJ_VEC_MASK;
2376 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_TYPE_MASK;
2378 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID;
2379 bool has_error_code = false;
2382 tss_selector = (u16)svm->vmcb->control.exit_info_1;
2384 if (svm->vmcb->control.exit_info_2 &
2385 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
2386 reason = TASK_SWITCH_IRET;
2387 else if (svm->vmcb->control.exit_info_2 &
2388 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
2389 reason = TASK_SWITCH_JMP;
2391 reason = TASK_SWITCH_GATE;
2393 reason = TASK_SWITCH_CALL;
2395 if (reason == TASK_SWITCH_GATE) {
2397 case SVM_EXITINTINFO_TYPE_NMI:
2398 svm->vcpu.arch.nmi_injected = false;
2400 case SVM_EXITINTINFO_TYPE_EXEPT:
2401 if (svm->vmcb->control.exit_info_2 &
2402 (1ULL << SVM_EXITINFOSHIFT_TS_HAS_ERROR_CODE)) {
2403 has_error_code = true;
2405 (u32)svm->vmcb->control.exit_info_2;
2407 kvm_clear_exception_queue(&svm->vcpu);
2409 case SVM_EXITINTINFO_TYPE_INTR:
2410 kvm_clear_interrupt_queue(&svm->vcpu);
2417 if (reason != TASK_SWITCH_GATE ||
2418 int_type == SVM_EXITINTINFO_TYPE_SOFT ||
2419 (int_type == SVM_EXITINTINFO_TYPE_EXEPT &&
2420 (int_vec == OF_VECTOR || int_vec == BP_VECTOR))) {
2421 if (!skip_emulated_instruction(&svm->vcpu))
2425 if (int_type != SVM_EXITINTINFO_TYPE_SOFT)
2428 return kvm_task_switch(&svm->vcpu, tss_selector, int_vec, reason,
2429 has_error_code, error_code);
2432 static int cpuid_interception(struct vcpu_svm *svm)
2434 return kvm_emulate_cpuid(&svm->vcpu);
2437 static int iret_interception(struct vcpu_svm *svm)
2439 ++svm->vcpu.stat.nmi_window_exits;
2440 svm->vcpu.arch.hflags |= HF_IRET_MASK;
2441 if (!sev_es_guest(svm->vcpu.kvm)) {
2442 svm_clr_intercept(svm, INTERCEPT_IRET);
2443 svm->nmi_iret_rip = kvm_rip_read(&svm->vcpu);
2445 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2449 static int invd_interception(struct vcpu_svm *svm)
2451 /* Treat an INVD instruction as a NOP and just skip it. */
2452 return kvm_skip_emulated_instruction(&svm->vcpu);
2455 static int invlpg_interception(struct vcpu_svm *svm)
2457 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2458 return kvm_emulate_instruction(&svm->vcpu, 0);
2460 kvm_mmu_invlpg(&svm->vcpu, svm->vmcb->control.exit_info_1);
2461 return kvm_skip_emulated_instruction(&svm->vcpu);
2464 static int emulate_on_interception(struct vcpu_svm *svm)
2466 return kvm_emulate_instruction(&svm->vcpu, 0);
2469 static int rsm_interception(struct vcpu_svm *svm)
2471 return kvm_emulate_instruction_from_buffer(&svm->vcpu, rsm_ins_bytes, 2);
2474 static int rdpmc_interception(struct vcpu_svm *svm)
2479 return emulate_on_interception(svm);
2481 err = kvm_rdpmc(&svm->vcpu);
2482 return kvm_complete_insn_gp(&svm->vcpu, err);
2485 static bool check_selective_cr0_intercepted(struct vcpu_svm *svm,
2488 unsigned long cr0 = svm->vcpu.arch.cr0;
2491 if (!is_guest_mode(&svm->vcpu) ||
2492 (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_SELECTIVE_CR0))))
2495 cr0 &= ~SVM_CR0_SELECTIVE_MASK;
2496 val &= ~SVM_CR0_SELECTIVE_MASK;
2499 svm->vmcb->control.exit_code = SVM_EXIT_CR0_SEL_WRITE;
2500 ret = (nested_svm_exit_handled(svm) == NESTED_EXIT_DONE);
2506 #define CR_VALID (1ULL << 63)
2508 static int cr_interception(struct vcpu_svm *svm)
2514 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2515 return emulate_on_interception(svm);
2517 if (unlikely((svm->vmcb->control.exit_info_1 & CR_VALID) == 0))
2518 return emulate_on_interception(svm);
2520 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2521 if (svm->vmcb->control.exit_code == SVM_EXIT_CR0_SEL_WRITE)
2522 cr = SVM_EXIT_WRITE_CR0 - SVM_EXIT_READ_CR0;
2524 cr = svm->vmcb->control.exit_code - SVM_EXIT_READ_CR0;
2527 if (cr >= 16) { /* mov to cr */
2529 val = kvm_register_read(&svm->vcpu, reg);
2530 trace_kvm_cr_write(cr, val);
2533 if (!check_selective_cr0_intercepted(svm, val))
2534 err = kvm_set_cr0(&svm->vcpu, val);
2540 err = kvm_set_cr3(&svm->vcpu, val);
2543 err = kvm_set_cr4(&svm->vcpu, val);
2546 err = kvm_set_cr8(&svm->vcpu, val);
2549 WARN(1, "unhandled write to CR%d", cr);
2550 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2553 } else { /* mov from cr */
2556 val = kvm_read_cr0(&svm->vcpu);
2559 val = svm->vcpu.arch.cr2;
2562 val = kvm_read_cr3(&svm->vcpu);
2565 val = kvm_read_cr4(&svm->vcpu);
2568 val = kvm_get_cr8(&svm->vcpu);
2571 WARN(1, "unhandled read from CR%d", cr);
2572 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2575 kvm_register_write(&svm->vcpu, reg, val);
2576 trace_kvm_cr_read(cr, val);
2578 return kvm_complete_insn_gp(&svm->vcpu, err);
2581 static int cr_trap(struct vcpu_svm *svm)
2583 struct kvm_vcpu *vcpu = &svm->vcpu;
2584 unsigned long old_value, new_value;
2588 new_value = (unsigned long)svm->vmcb->control.exit_info_1;
2590 cr = svm->vmcb->control.exit_code - SVM_EXIT_CR0_WRITE_TRAP;
2593 old_value = kvm_read_cr0(vcpu);
2594 svm_set_cr0(vcpu, new_value);
2596 kvm_post_set_cr0(vcpu, old_value, new_value);
2599 old_value = kvm_read_cr4(vcpu);
2600 svm_set_cr4(vcpu, new_value);
2602 kvm_post_set_cr4(vcpu, old_value, new_value);
2605 ret = kvm_set_cr8(&svm->vcpu, new_value);
2608 WARN(1, "unhandled CR%d write trap", cr);
2609 kvm_queue_exception(vcpu, UD_VECTOR);
2613 return kvm_complete_insn_gp(vcpu, ret);
2616 static int dr_interception(struct vcpu_svm *svm)
2622 if (svm->vcpu.guest_debug == 0) {
2624 * No more DR vmexits; force a reload of the debug registers
2625 * and reenter on this instruction. The next vmexit will
2626 * retrieve the full state of the debug registers.
2628 clr_dr_intercepts(svm);
2629 svm->vcpu.arch.switch_db_regs |= KVM_DEBUGREG_WONT_EXIT;
2633 if (!boot_cpu_has(X86_FEATURE_DECODEASSISTS))
2634 return emulate_on_interception(svm);
2636 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2637 dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
2638 if (dr >= 16) { /* mov to DRn */
2640 val = kvm_register_read(&svm->vcpu, reg);
2641 err = kvm_set_dr(&svm->vcpu, dr, val);
2643 kvm_get_dr(&svm->vcpu, dr, &val);
2644 kvm_register_write(&svm->vcpu, reg, val);
2647 return kvm_complete_insn_gp(&svm->vcpu, err);
2650 static int cr8_write_interception(struct vcpu_svm *svm)
2652 struct kvm_run *kvm_run = svm->vcpu.run;
2655 u8 cr8_prev = kvm_get_cr8(&svm->vcpu);
2656 /* instruction emulation calls kvm_set_cr8() */
2657 r = cr_interception(svm);
2658 if (lapic_in_kernel(&svm->vcpu))
2660 if (cr8_prev <= kvm_get_cr8(&svm->vcpu))
2662 kvm_run->exit_reason = KVM_EXIT_SET_TPR;
2666 static int efer_trap(struct vcpu_svm *svm)
2668 struct msr_data msr_info;
2672 * Clear the EFER_SVME bit from EFER. The SVM code always sets this
2673 * bit in svm_set_efer(), but __kvm_valid_efer() checks it against
2674 * whether the guest has X86_FEATURE_SVM - this avoids a failure if
2675 * the guest doesn't have X86_FEATURE_SVM.
2677 msr_info.host_initiated = false;
2678 msr_info.index = MSR_EFER;
2679 msr_info.data = svm->vmcb->control.exit_info_1 & ~EFER_SVME;
2680 ret = kvm_set_msr_common(&svm->vcpu, &msr_info);
2682 return kvm_complete_insn_gp(&svm->vcpu, ret);
2685 static int svm_get_msr_feature(struct kvm_msr_entry *msr)
2689 switch (msr->index) {
2690 case MSR_F10H_DECFG:
2691 if (boot_cpu_has(X86_FEATURE_LFENCE_RDTSC))
2692 msr->data |= MSR_F10H_DECFG_LFENCE_SERIALIZE;
2694 case MSR_IA32_PERF_CAPABILITIES:
2697 return KVM_MSR_RET_INVALID;
2703 static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
2705 struct vcpu_svm *svm = to_svm(vcpu);
2707 switch (msr_info->index) {
2709 msr_info->data = svm->vmcb->save.star;
2711 #ifdef CONFIG_X86_64
2713 msr_info->data = svm->vmcb->save.lstar;
2716 msr_info->data = svm->vmcb->save.cstar;
2718 case MSR_KERNEL_GS_BASE:
2719 msr_info->data = svm->vmcb->save.kernel_gs_base;
2721 case MSR_SYSCALL_MASK:
2722 msr_info->data = svm->vmcb->save.sfmask;
2725 case MSR_IA32_SYSENTER_CS:
2726 msr_info->data = svm->vmcb->save.sysenter_cs;
2728 case MSR_IA32_SYSENTER_EIP:
2729 msr_info->data = svm->sysenter_eip;
2731 case MSR_IA32_SYSENTER_ESP:
2732 msr_info->data = svm->sysenter_esp;
2735 if (!boot_cpu_has(X86_FEATURE_RDTSCP))
2737 msr_info->data = svm->tsc_aux;
2740 * Nobody will change the following 5 values in the VMCB so we can
2741 * safely return them on rdmsr. They will always be 0 until LBRV is
2744 case MSR_IA32_DEBUGCTLMSR:
2745 msr_info->data = svm->vmcb->save.dbgctl;
2747 case MSR_IA32_LASTBRANCHFROMIP:
2748 msr_info->data = svm->vmcb->save.br_from;
2750 case MSR_IA32_LASTBRANCHTOIP:
2751 msr_info->data = svm->vmcb->save.br_to;
2753 case MSR_IA32_LASTINTFROMIP:
2754 msr_info->data = svm->vmcb->save.last_excp_from;
2756 case MSR_IA32_LASTINTTOIP:
2757 msr_info->data = svm->vmcb->save.last_excp_to;
2759 case MSR_VM_HSAVE_PA:
2760 msr_info->data = svm->nested.hsave_msr;
2763 msr_info->data = svm->nested.vm_cr_msr;
2765 case MSR_IA32_SPEC_CTRL:
2766 if (!msr_info->host_initiated &&
2767 !guest_has_spec_ctrl_msr(vcpu))
2770 msr_info->data = svm->spec_ctrl;
2772 case MSR_AMD64_VIRT_SPEC_CTRL:
2773 if (!msr_info->host_initiated &&
2774 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2777 msr_info->data = svm->virt_spec_ctrl;
2779 case MSR_F15H_IC_CFG: {
2783 family = guest_cpuid_family(vcpu);
2784 model = guest_cpuid_model(vcpu);
2786 if (family < 0 || model < 0)
2787 return kvm_get_msr_common(vcpu, msr_info);
2791 if (family == 0x15 &&
2792 (model >= 0x2 && model < 0x20))
2793 msr_info->data = 0x1E;
2796 case MSR_F10H_DECFG:
2797 msr_info->data = svm->msr_decfg;
2800 return kvm_get_msr_common(vcpu, msr_info);
2805 static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
2807 struct vcpu_svm *svm = to_svm(vcpu);
2808 if (!sev_es_guest(svm->vcpu.kvm) || !err)
2809 return kvm_complete_insn_gp(&svm->vcpu, err);
2811 ghcb_set_sw_exit_info_1(svm->ghcb, 1);
2812 ghcb_set_sw_exit_info_2(svm->ghcb,
2814 SVM_EVTINJ_TYPE_EXEPT |
2819 static int rdmsr_interception(struct vcpu_svm *svm)
2821 return kvm_emulate_rdmsr(&svm->vcpu);
2824 static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
2826 struct vcpu_svm *svm = to_svm(vcpu);
2827 int svm_dis, chg_mask;
2829 if (data & ~SVM_VM_CR_VALID_MASK)
2832 chg_mask = SVM_VM_CR_VALID_MASK;
2834 if (svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK)
2835 chg_mask &= ~(SVM_VM_CR_SVM_LOCK_MASK | SVM_VM_CR_SVM_DIS_MASK);
2837 svm->nested.vm_cr_msr &= ~chg_mask;
2838 svm->nested.vm_cr_msr |= (data & chg_mask);
2840 svm_dis = svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK;
2842 /* check for svm_disable while efer.svme is set */
2843 if (svm_dis && (vcpu->arch.efer & EFER_SVME))
2849 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
2851 struct vcpu_svm *svm = to_svm(vcpu);
2853 u32 ecx = msr->index;
2854 u64 data = msr->data;
2856 case MSR_IA32_CR_PAT:
2857 if (!kvm_mtrr_valid(vcpu, MSR_IA32_CR_PAT, data))
2859 vcpu->arch.pat = data;
2860 svm->vmcb->save.g_pat = data;
2861 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
2863 case MSR_IA32_SPEC_CTRL:
2864 if (!msr->host_initiated &&
2865 !guest_has_spec_ctrl_msr(vcpu))
2868 if (kvm_spec_ctrl_test_value(data))
2871 svm->spec_ctrl = data;
2877 * When it's written (to non-zero) for the first time, pass
2881 * The handling of the MSR bitmap for L2 guests is done in
2882 * nested_svm_vmrun_msrpm.
2883 * We update the L1 MSR bit as well since it will end up
2884 * touching the MSR anyway now.
2886 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
2888 case MSR_IA32_PRED_CMD:
2889 if (!msr->host_initiated &&
2890 !guest_has_pred_cmd_msr(vcpu))
2893 if (data & ~PRED_CMD_IBPB)
2895 if (!boot_cpu_has(X86_FEATURE_IBPB))
2900 wrmsrl(MSR_IA32_PRED_CMD, PRED_CMD_IBPB);
2901 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_PRED_CMD, 0, 1);
2903 case MSR_AMD64_VIRT_SPEC_CTRL:
2904 if (!msr->host_initiated &&
2905 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2908 if (data & ~SPEC_CTRL_SSBD)
2911 svm->virt_spec_ctrl = data;
2914 svm->vmcb->save.star = data;
2916 #ifdef CONFIG_X86_64
2918 svm->vmcb->save.lstar = data;
2921 svm->vmcb->save.cstar = data;
2923 case MSR_KERNEL_GS_BASE:
2924 svm->vmcb->save.kernel_gs_base = data;
2926 case MSR_SYSCALL_MASK:
2927 svm->vmcb->save.sfmask = data;
2930 case MSR_IA32_SYSENTER_CS:
2931 svm->vmcb->save.sysenter_cs = data;
2933 case MSR_IA32_SYSENTER_EIP:
2934 svm->sysenter_eip = data;
2935 svm->vmcb->save.sysenter_eip = data;
2937 case MSR_IA32_SYSENTER_ESP:
2938 svm->sysenter_esp = data;
2939 svm->vmcb->save.sysenter_esp = data;
2942 if (!boot_cpu_has(X86_FEATURE_RDTSCP))
2946 * This is rare, so we update the MSR here instead of using
2947 * direct_access_msrs. Doing that would require a rdmsr in
2950 svm->tsc_aux = data;
2951 wrmsrl(MSR_TSC_AUX, svm->tsc_aux);
2953 case MSR_IA32_DEBUGCTLMSR:
2954 if (!boot_cpu_has(X86_FEATURE_LBRV)) {
2955 vcpu_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTL 0x%llx, nop\n",
2959 if (data & DEBUGCTL_RESERVED_BITS)
2962 svm->vmcb->save.dbgctl = data;
2963 vmcb_mark_dirty(svm->vmcb, VMCB_LBR);
2964 if (data & (1ULL<<0))
2965 svm_enable_lbrv(vcpu);
2967 svm_disable_lbrv(vcpu);
2969 case MSR_VM_HSAVE_PA:
2970 svm->nested.hsave_msr = data;
2973 return svm_set_vm_cr(vcpu, data);
2975 vcpu_unimpl(vcpu, "unimplemented wrmsr: 0x%x data 0x%llx\n", ecx, data);
2977 case MSR_F10H_DECFG: {
2978 struct kvm_msr_entry msr_entry;
2980 msr_entry.index = msr->index;
2981 if (svm_get_msr_feature(&msr_entry))
2984 /* Check the supported bits */
2985 if (data & ~msr_entry.data)
2988 /* Don't allow the guest to change a bit, #GP */
2989 if (!msr->host_initiated && (data ^ msr_entry.data))
2992 svm->msr_decfg = data;
2995 case MSR_IA32_APICBASE:
2996 if (kvm_vcpu_apicv_active(vcpu))
2997 avic_update_vapic_bar(to_svm(vcpu), data);
3000 return kvm_set_msr_common(vcpu, msr);
3005 static int wrmsr_interception(struct vcpu_svm *svm)
3007 return kvm_emulate_wrmsr(&svm->vcpu);
3010 static int msr_interception(struct vcpu_svm *svm)
3012 if (svm->vmcb->control.exit_info_1)
3013 return wrmsr_interception(svm);
3015 return rdmsr_interception(svm);
3018 static int interrupt_window_interception(struct vcpu_svm *svm)
3020 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
3021 svm_clear_vintr(svm);
3024 * For AVIC, the only reason to end up here is ExtINTs.
3025 * In this case AVIC was temporarily disabled for
3026 * requesting the IRQ window and we have to re-enable it.
3028 svm_toggle_avic_for_irq_window(&svm->vcpu, true);
3030 ++svm->vcpu.stat.irq_window_exits;
3034 static int pause_interception(struct vcpu_svm *svm)
3036 struct kvm_vcpu *vcpu = &svm->vcpu;
3040 * CPL is not made available for an SEV-ES guest, therefore
3041 * vcpu->arch.preempted_in_kernel can never be true. Just
3042 * set in_kernel to false as well.
3044 in_kernel = !sev_es_guest(svm->vcpu.kvm) && svm_get_cpl(vcpu) == 0;
3046 if (!kvm_pause_in_guest(vcpu->kvm))
3047 grow_ple_window(vcpu);
3049 kvm_vcpu_on_spin(vcpu, in_kernel);
3053 static int nop_interception(struct vcpu_svm *svm)
3055 return kvm_skip_emulated_instruction(&(svm->vcpu));
3058 static int monitor_interception(struct vcpu_svm *svm)
3060 printk_once(KERN_WARNING "kvm: MONITOR instruction emulated as NOP!\n");
3061 return nop_interception(svm);
3064 static int mwait_interception(struct vcpu_svm *svm)
3066 printk_once(KERN_WARNING "kvm: MWAIT instruction emulated as NOP!\n");
3067 return nop_interception(svm);
3070 static int invpcid_interception(struct vcpu_svm *svm)
3072 struct kvm_vcpu *vcpu = &svm->vcpu;
3076 if (!guest_cpuid_has(vcpu, X86_FEATURE_INVPCID)) {
3077 kvm_queue_exception(vcpu, UD_VECTOR);
3082 * For an INVPCID intercept:
3083 * EXITINFO1 provides the linear address of the memory operand.
3084 * EXITINFO2 provides the contents of the register operand.
3086 type = svm->vmcb->control.exit_info_2;
3087 gva = svm->vmcb->control.exit_info_1;
3090 kvm_inject_gp(vcpu, 0);
3094 return kvm_handle_invpcid(vcpu, type, gva);
3097 static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = {
3098 [SVM_EXIT_READ_CR0] = cr_interception,
3099 [SVM_EXIT_READ_CR3] = cr_interception,
3100 [SVM_EXIT_READ_CR4] = cr_interception,
3101 [SVM_EXIT_READ_CR8] = cr_interception,
3102 [SVM_EXIT_CR0_SEL_WRITE] = cr_interception,
3103 [SVM_EXIT_WRITE_CR0] = cr_interception,
3104 [SVM_EXIT_WRITE_CR3] = cr_interception,
3105 [SVM_EXIT_WRITE_CR4] = cr_interception,
3106 [SVM_EXIT_WRITE_CR8] = cr8_write_interception,
3107 [SVM_EXIT_READ_DR0] = dr_interception,
3108 [SVM_EXIT_READ_DR1] = dr_interception,
3109 [SVM_EXIT_READ_DR2] = dr_interception,
3110 [SVM_EXIT_READ_DR3] = dr_interception,
3111 [SVM_EXIT_READ_DR4] = dr_interception,
3112 [SVM_EXIT_READ_DR5] = dr_interception,
3113 [SVM_EXIT_READ_DR6] = dr_interception,
3114 [SVM_EXIT_READ_DR7] = dr_interception,
3115 [SVM_EXIT_WRITE_DR0] = dr_interception,
3116 [SVM_EXIT_WRITE_DR1] = dr_interception,
3117 [SVM_EXIT_WRITE_DR2] = dr_interception,
3118 [SVM_EXIT_WRITE_DR3] = dr_interception,
3119 [SVM_EXIT_WRITE_DR4] = dr_interception,
3120 [SVM_EXIT_WRITE_DR5] = dr_interception,
3121 [SVM_EXIT_WRITE_DR6] = dr_interception,
3122 [SVM_EXIT_WRITE_DR7] = dr_interception,
3123 [SVM_EXIT_EXCP_BASE + DB_VECTOR] = db_interception,
3124 [SVM_EXIT_EXCP_BASE + BP_VECTOR] = bp_interception,
3125 [SVM_EXIT_EXCP_BASE + UD_VECTOR] = ud_interception,
3126 [SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
3127 [SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
3128 [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
3129 [SVM_EXIT_EXCP_BASE + GP_VECTOR] = gp_interception,
3130 [SVM_EXIT_INTR] = intr_interception,
3131 [SVM_EXIT_NMI] = nmi_interception,
3132 [SVM_EXIT_SMI] = nop_on_interception,
3133 [SVM_EXIT_INIT] = nop_on_interception,
3134 [SVM_EXIT_VINTR] = interrupt_window_interception,
3135 [SVM_EXIT_RDPMC] = rdpmc_interception,
3136 [SVM_EXIT_CPUID] = cpuid_interception,
3137 [SVM_EXIT_IRET] = iret_interception,
3138 [SVM_EXIT_INVD] = invd_interception,
3139 [SVM_EXIT_PAUSE] = pause_interception,
3140 [SVM_EXIT_HLT] = halt_interception,
3141 [SVM_EXIT_INVLPG] = invlpg_interception,
3142 [SVM_EXIT_INVLPGA] = invlpga_interception,
3143 [SVM_EXIT_IOIO] = io_interception,
3144 [SVM_EXIT_MSR] = msr_interception,
3145 [SVM_EXIT_TASK_SWITCH] = task_switch_interception,
3146 [SVM_EXIT_SHUTDOWN] = shutdown_interception,
3147 [SVM_EXIT_VMRUN] = vmrun_interception,
3148 [SVM_EXIT_VMMCALL] = vmmcall_interception,
3149 [SVM_EXIT_VMLOAD] = vmload_interception,
3150 [SVM_EXIT_VMSAVE] = vmsave_interception,
3151 [SVM_EXIT_STGI] = stgi_interception,
3152 [SVM_EXIT_CLGI] = clgi_interception,
3153 [SVM_EXIT_SKINIT] = skinit_interception,
3154 [SVM_EXIT_WBINVD] = wbinvd_interception,
3155 [SVM_EXIT_MONITOR] = monitor_interception,
3156 [SVM_EXIT_MWAIT] = mwait_interception,
3157 [SVM_EXIT_XSETBV] = xsetbv_interception,
3158 [SVM_EXIT_RDPRU] = rdpru_interception,
3159 [SVM_EXIT_EFER_WRITE_TRAP] = efer_trap,
3160 [SVM_EXIT_CR0_WRITE_TRAP] = cr_trap,
3161 [SVM_EXIT_CR4_WRITE_TRAP] = cr_trap,
3162 [SVM_EXIT_CR8_WRITE_TRAP] = cr_trap,
3163 [SVM_EXIT_INVPCID] = invpcid_interception,
3164 [SVM_EXIT_NPF] = npf_interception,
3165 [SVM_EXIT_RSM] = rsm_interception,
3166 [SVM_EXIT_AVIC_INCOMPLETE_IPI] = avic_incomplete_ipi_interception,
3167 [SVM_EXIT_AVIC_UNACCELERATED_ACCESS] = avic_unaccelerated_access_interception,
3168 [SVM_EXIT_VMGEXIT] = sev_handle_vmgexit,
3171 static void dump_vmcb(struct kvm_vcpu *vcpu)
3173 struct vcpu_svm *svm = to_svm(vcpu);
3174 struct vmcb_control_area *control = &svm->vmcb->control;
3175 struct vmcb_save_area *save = &svm->vmcb->save;
3177 if (!dump_invalid_vmcb) {
3178 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3182 pr_err("VMCB Control Area:\n");
3183 pr_err("%-20s%04x\n", "cr_read:", control->intercepts[INTERCEPT_CR] & 0xffff);
3184 pr_err("%-20s%04x\n", "cr_write:", control->intercepts[INTERCEPT_CR] >> 16);
3185 pr_err("%-20s%04x\n", "dr_read:", control->intercepts[INTERCEPT_DR] & 0xffff);
3186 pr_err("%-20s%04x\n", "dr_write:", control->intercepts[INTERCEPT_DR] >> 16);
3187 pr_err("%-20s%08x\n", "exceptions:", control->intercepts[INTERCEPT_EXCEPTION]);
3188 pr_err("%-20s%08x %08x\n", "intercepts:",
3189 control->intercepts[INTERCEPT_WORD3],
3190 control->intercepts[INTERCEPT_WORD4]);
3191 pr_err("%-20s%d\n", "pause filter count:", control->pause_filter_count);
3192 pr_err("%-20s%d\n", "pause filter threshold:",
3193 control->pause_filter_thresh);
3194 pr_err("%-20s%016llx\n", "iopm_base_pa:", control->iopm_base_pa);
3195 pr_err("%-20s%016llx\n", "msrpm_base_pa:", control->msrpm_base_pa);
3196 pr_err("%-20s%016llx\n", "tsc_offset:", control->tsc_offset);
3197 pr_err("%-20s%d\n", "asid:", control->asid);
3198 pr_err("%-20s%d\n", "tlb_ctl:", control->tlb_ctl);
3199 pr_err("%-20s%08x\n", "int_ctl:", control->int_ctl);
3200 pr_err("%-20s%08x\n", "int_vector:", control->int_vector);
3201 pr_err("%-20s%08x\n", "int_state:", control->int_state);
3202 pr_err("%-20s%08x\n", "exit_code:", control->exit_code);
3203 pr_err("%-20s%016llx\n", "exit_info1:", control->exit_info_1);
3204 pr_err("%-20s%016llx\n", "exit_info2:", control->exit_info_2);
3205 pr_err("%-20s%08x\n", "exit_int_info:", control->exit_int_info);
3206 pr_err("%-20s%08x\n", "exit_int_info_err:", control->exit_int_info_err);
3207 pr_err("%-20s%lld\n", "nested_ctl:", control->nested_ctl);
3208 pr_err("%-20s%016llx\n", "nested_cr3:", control->nested_cr3);
3209 pr_err("%-20s%016llx\n", "avic_vapic_bar:", control->avic_vapic_bar);
3210 pr_err("%-20s%016llx\n", "ghcb:", control->ghcb_gpa);
3211 pr_err("%-20s%08x\n", "event_inj:", control->event_inj);
3212 pr_err("%-20s%08x\n", "event_inj_err:", control->event_inj_err);
3213 pr_err("%-20s%lld\n", "virt_ext:", control->virt_ext);
3214 pr_err("%-20s%016llx\n", "next_rip:", control->next_rip);
3215 pr_err("%-20s%016llx\n", "avic_backing_page:", control->avic_backing_page);
3216 pr_err("%-20s%016llx\n", "avic_logical_id:", control->avic_logical_id);
3217 pr_err("%-20s%016llx\n", "avic_physical_id:", control->avic_physical_id);
3218 pr_err("%-20s%016llx\n", "vmsa_pa:", control->vmsa_pa);
3219 pr_err("VMCB State Save Area:\n");
3220 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3222 save->es.selector, save->es.attrib,
3223 save->es.limit, save->es.base);
3224 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3226 save->cs.selector, save->cs.attrib,
3227 save->cs.limit, save->cs.base);
3228 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3230 save->ss.selector, save->ss.attrib,
3231 save->ss.limit, save->ss.base);
3232 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3234 save->ds.selector, save->ds.attrib,
3235 save->ds.limit, save->ds.base);
3236 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3238 save->fs.selector, save->fs.attrib,
3239 save->fs.limit, save->fs.base);
3240 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3242 save->gs.selector, save->gs.attrib,
3243 save->gs.limit, save->gs.base);
3244 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3246 save->gdtr.selector, save->gdtr.attrib,
3247 save->gdtr.limit, save->gdtr.base);
3248 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3250 save->ldtr.selector, save->ldtr.attrib,
3251 save->ldtr.limit, save->ldtr.base);
3252 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3254 save->idtr.selector, save->idtr.attrib,
3255 save->idtr.limit, save->idtr.base);
3256 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3258 save->tr.selector, save->tr.attrib,
3259 save->tr.limit, save->tr.base);
3260 pr_err("cpl: %d efer: %016llx\n",
3261 save->cpl, save->efer);
3262 pr_err("%-15s %016llx %-13s %016llx\n",
3263 "cr0:", save->cr0, "cr2:", save->cr2);
3264 pr_err("%-15s %016llx %-13s %016llx\n",
3265 "cr3:", save->cr3, "cr4:", save->cr4);
3266 pr_err("%-15s %016llx %-13s %016llx\n",
3267 "dr6:", save->dr6, "dr7:", save->dr7);
3268 pr_err("%-15s %016llx %-13s %016llx\n",
3269 "rip:", save->rip, "rflags:", save->rflags);
3270 pr_err("%-15s %016llx %-13s %016llx\n",
3271 "rsp:", save->rsp, "rax:", save->rax);
3272 pr_err("%-15s %016llx %-13s %016llx\n",
3273 "star:", save->star, "lstar:", save->lstar);
3274 pr_err("%-15s %016llx %-13s %016llx\n",
3275 "cstar:", save->cstar, "sfmask:", save->sfmask);
3276 pr_err("%-15s %016llx %-13s %016llx\n",
3277 "kernel_gs_base:", save->kernel_gs_base,
3278 "sysenter_cs:", save->sysenter_cs);
3279 pr_err("%-15s %016llx %-13s %016llx\n",
3280 "sysenter_esp:", save->sysenter_esp,
3281 "sysenter_eip:", save->sysenter_eip);
3282 pr_err("%-15s %016llx %-13s %016llx\n",
3283 "gpat:", save->g_pat, "dbgctl:", save->dbgctl);
3284 pr_err("%-15s %016llx %-13s %016llx\n",
3285 "br_from:", save->br_from, "br_to:", save->br_to);
3286 pr_err("%-15s %016llx %-13s %016llx\n",
3287 "excp_from:", save->last_excp_from,
3288 "excp_to:", save->last_excp_to);
3291 static int svm_handle_invalid_exit(struct kvm_vcpu *vcpu, u64 exit_code)
3293 if (exit_code < ARRAY_SIZE(svm_exit_handlers) &&
3294 svm_exit_handlers[exit_code])
3297 vcpu_unimpl(vcpu, "svm: unexpected exit reason 0x%llx\n", exit_code);
3299 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3300 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
3301 vcpu->run->internal.ndata = 2;
3302 vcpu->run->internal.data[0] = exit_code;
3303 vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
3308 int svm_invoke_exit_handler(struct vcpu_svm *svm, u64 exit_code)
3310 if (svm_handle_invalid_exit(&svm->vcpu, exit_code))
3313 #ifdef CONFIG_RETPOLINE
3314 if (exit_code == SVM_EXIT_MSR)
3315 return msr_interception(svm);
3316 else if (exit_code == SVM_EXIT_VINTR)
3317 return interrupt_window_interception(svm);
3318 else if (exit_code == SVM_EXIT_INTR)
3319 return intr_interception(svm);
3320 else if (exit_code == SVM_EXIT_HLT)
3321 return halt_interception(svm);
3322 else if (exit_code == SVM_EXIT_NPF)
3323 return npf_interception(svm);
3325 return svm_exit_handlers[exit_code](svm);
3328 static void svm_get_exit_info(struct kvm_vcpu *vcpu, u64 *info1, u64 *info2,
3329 u32 *intr_info, u32 *error_code)
3331 struct vmcb_control_area *control = &to_svm(vcpu)->vmcb->control;
3333 *info1 = control->exit_info_1;
3334 *info2 = control->exit_info_2;
3335 *intr_info = control->exit_int_info;
3336 if ((*intr_info & SVM_EXITINTINFO_VALID) &&
3337 (*intr_info & SVM_EXITINTINFO_VALID_ERR))
3338 *error_code = control->exit_int_info_err;
3343 static int handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
3345 struct vcpu_svm *svm = to_svm(vcpu);
3346 struct kvm_run *kvm_run = vcpu->run;
3347 u32 exit_code = svm->vmcb->control.exit_code;
3349 trace_kvm_exit(exit_code, vcpu, KVM_ISA_SVM);
3351 /* SEV-ES guests must use the CR write traps to track CR registers. */
3352 if (!sev_es_guest(vcpu->kvm)) {
3353 if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
3354 vcpu->arch.cr0 = svm->vmcb->save.cr0;
3356 vcpu->arch.cr3 = svm->vmcb->save.cr3;
3359 if (is_guest_mode(vcpu)) {
3362 trace_kvm_nested_vmexit(exit_code, vcpu, KVM_ISA_SVM);
3364 vmexit = nested_svm_exit_special(svm);
3366 if (vmexit == NESTED_EXIT_CONTINUE)
3367 vmexit = nested_svm_exit_handled(svm);
3369 if (vmexit == NESTED_EXIT_DONE)
3373 if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
3374 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3375 kvm_run->fail_entry.hardware_entry_failure_reason
3376 = svm->vmcb->control.exit_code;
3377 kvm_run->fail_entry.cpu = vcpu->arch.last_vmentry_cpu;
3382 if (is_external_interrupt(svm->vmcb->control.exit_int_info) &&
3383 exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR &&
3384 exit_code != SVM_EXIT_NPF && exit_code != SVM_EXIT_TASK_SWITCH &&
3385 exit_code != SVM_EXIT_INTR && exit_code != SVM_EXIT_NMI)
3386 printk(KERN_ERR "%s: unexpected exit_int_info 0x%x "
3388 __func__, svm->vmcb->control.exit_int_info,
3391 if (exit_fastpath != EXIT_FASTPATH_NONE)
3394 return svm_invoke_exit_handler(svm, exit_code);
3397 static void reload_tss(struct kvm_vcpu *vcpu)
3399 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3401 sd->tss_desc->type = 9; /* available 32/64-bit TSS */
3405 static void pre_svm_run(struct vcpu_svm *svm)
3407 struct svm_cpu_data *sd = per_cpu(svm_data, svm->vcpu.cpu);
3409 if (sev_guest(svm->vcpu.kvm))
3410 return pre_sev_run(svm, svm->vcpu.cpu);
3412 /* FIXME: handle wraparound of asid_generation */
3413 if (svm->asid_generation != sd->asid_generation)
3417 static void svm_inject_nmi(struct kvm_vcpu *vcpu)
3419 struct vcpu_svm *svm = to_svm(vcpu);
3421 svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
3422 vcpu->arch.hflags |= HF_NMI_MASK;
3423 if (!sev_es_guest(svm->vcpu.kvm))
3424 svm_set_intercept(svm, INTERCEPT_IRET);
3425 ++vcpu->stat.nmi_injections;
3428 static void svm_set_irq(struct kvm_vcpu *vcpu)
3430 struct vcpu_svm *svm = to_svm(vcpu);
3432 BUG_ON(!(gif_set(svm)));
3434 trace_kvm_inj_virq(vcpu->arch.interrupt.nr);
3435 ++vcpu->stat.irq_injections;
3437 svm->vmcb->control.event_inj = vcpu->arch.interrupt.nr |
3438 SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR;
3441 static void svm_update_cr8_intercept(struct kvm_vcpu *vcpu, int tpr, int irr)
3443 struct vcpu_svm *svm = to_svm(vcpu);
3446 * SEV-ES guests must always keep the CR intercepts cleared. CR
3447 * tracking is done using the CR write traps.
3449 if (sev_es_guest(vcpu->kvm))
3452 if (nested_svm_virtualize_tpr(vcpu))
3455 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3461 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
3464 bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
3466 struct vcpu_svm *svm = to_svm(vcpu);
3467 struct vmcb *vmcb = svm->vmcb;
3473 if (is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3476 ret = (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK) ||
3477 (svm->vcpu.arch.hflags & HF_NMI_MASK);
3482 static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3484 struct vcpu_svm *svm = to_svm(vcpu);
3485 if (svm->nested.nested_run_pending)
3488 /* An NMI must not be injected into L2 if it's supposed to VM-Exit. */
3489 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3492 return !svm_nmi_blocked(vcpu);
3495 static bool svm_get_nmi_mask(struct kvm_vcpu *vcpu)
3497 struct vcpu_svm *svm = to_svm(vcpu);
3499 return !!(svm->vcpu.arch.hflags & HF_NMI_MASK);
3502 static void svm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
3504 struct vcpu_svm *svm = to_svm(vcpu);
3507 svm->vcpu.arch.hflags |= HF_NMI_MASK;
3508 if (!sev_es_guest(svm->vcpu.kvm))
3509 svm_set_intercept(svm, INTERCEPT_IRET);
3511 svm->vcpu.arch.hflags &= ~HF_NMI_MASK;
3512 if (!sev_es_guest(svm->vcpu.kvm))
3513 svm_clr_intercept(svm, INTERCEPT_IRET);
3517 bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
3519 struct vcpu_svm *svm = to_svm(vcpu);
3520 struct vmcb *vmcb = svm->vmcb;
3525 if (sev_es_guest(svm->vcpu.kvm)) {
3527 * SEV-ES guests to not expose RFLAGS. Use the VMCB interrupt mask
3528 * bit to determine the state of the IF flag.
3530 if (!(vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK))
3532 } else if (is_guest_mode(vcpu)) {
3533 /* As long as interrupts are being delivered... */
3534 if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
3535 ? !(svm->nested.hsave->save.rflags & X86_EFLAGS_IF)
3536 : !(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3539 /* ... vmexits aren't blocked by the interrupt shadow */
3540 if (nested_exit_on_intr(svm))
3543 if (!(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3547 return (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK);
3550 static int svm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3552 struct vcpu_svm *svm = to_svm(vcpu);
3553 if (svm->nested.nested_run_pending)
3557 * An IRQ must not be injected into L2 if it's supposed to VM-Exit,
3558 * e.g. if the IRQ arrived asynchronously after checking nested events.
3560 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_intr(svm))
3563 return !svm_interrupt_blocked(vcpu);
3566 static void svm_enable_irq_window(struct kvm_vcpu *vcpu)
3568 struct vcpu_svm *svm = to_svm(vcpu);
3571 * In case GIF=0 we can't rely on the CPU to tell us when GIF becomes
3572 * 1, because that's a separate STGI/VMRUN intercept. The next time we
3573 * get that intercept, this function will be called again though and
3574 * we'll get the vintr intercept. However, if the vGIF feature is
3575 * enabled, the STGI interception will not occur. Enable the irq
3576 * window under the assumption that the hardware will set the GIF.
3578 if (vgif_enabled(svm) || gif_set(svm)) {
3580 * IRQ window is not needed when AVIC is enabled,
3581 * unless we have pending ExtINT since it cannot be injected
3582 * via AVIC. In such case, we need to temporarily disable AVIC,
3583 * and fallback to injecting IRQ via V_IRQ.
3585 svm_toggle_avic_for_irq_window(vcpu, false);
3590 static void svm_enable_nmi_window(struct kvm_vcpu *vcpu)
3592 struct vcpu_svm *svm = to_svm(vcpu);
3594 if ((svm->vcpu.arch.hflags & (HF_NMI_MASK | HF_IRET_MASK))
3596 return; /* IRET will cause a vm exit */
3598 if (!gif_set(svm)) {
3599 if (vgif_enabled(svm))
3600 svm_set_intercept(svm, INTERCEPT_STGI);
3601 return; /* STGI will cause a vm exit */
3605 * Something prevents NMI from been injected. Single step over possible
3606 * problem (IRET or exception injection or interrupt shadow)
3608 svm->nmi_singlestep_guest_rflags = svm_get_rflags(vcpu);
3609 svm->nmi_singlestep = true;
3610 svm->vmcb->save.rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
3613 static int svm_set_tss_addr(struct kvm *kvm, unsigned int addr)
3618 static int svm_set_identity_map_addr(struct kvm *kvm, u64 ident_addr)
3623 void svm_flush_tlb(struct kvm_vcpu *vcpu)
3625 struct vcpu_svm *svm = to_svm(vcpu);
3628 * Flush only the current ASID even if the TLB flush was invoked via
3629 * kvm_flush_remote_tlbs(). Although flushing remote TLBs requires all
3630 * ASIDs to be flushed, KVM uses a single ASID for L1 and L2, and
3631 * unconditionally does a TLB flush on both nested VM-Enter and nested
3632 * VM-Exit (via kvm_mmu_reset_context()).
3634 if (static_cpu_has(X86_FEATURE_FLUSHBYASID))
3635 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3637 svm->asid_generation--;
3640 static void svm_flush_tlb_gva(struct kvm_vcpu *vcpu, gva_t gva)
3642 struct vcpu_svm *svm = to_svm(vcpu);
3644 invlpga(gva, svm->vmcb->control.asid);
3647 static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)
3649 struct vcpu_svm *svm = to_svm(vcpu);
3651 if (nested_svm_virtualize_tpr(vcpu))
3654 if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
3655 int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
3656 kvm_set_cr8(vcpu, cr8);
3660 static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
3662 struct vcpu_svm *svm = to_svm(vcpu);
3665 if (nested_svm_virtualize_tpr(vcpu) ||
3666 kvm_vcpu_apicv_active(vcpu))
3669 cr8 = kvm_get_cr8(vcpu);
3670 svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
3671 svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
3674 static void svm_complete_interrupts(struct vcpu_svm *svm)
3678 u32 exitintinfo = svm->vmcb->control.exit_int_info;
3679 unsigned int3_injected = svm->int3_injected;
3681 svm->int3_injected = 0;
3684 * If we've made progress since setting HF_IRET_MASK, we've
3685 * executed an IRET and can allow NMI injection.
3687 if ((svm->vcpu.arch.hflags & HF_IRET_MASK) &&
3688 (sev_es_guest(svm->vcpu.kvm) ||
3689 kvm_rip_read(&svm->vcpu) != svm->nmi_iret_rip)) {
3690 svm->vcpu.arch.hflags &= ~(HF_NMI_MASK | HF_IRET_MASK);
3691 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
3694 svm->vcpu.arch.nmi_injected = false;
3695 kvm_clear_exception_queue(&svm->vcpu);
3696 kvm_clear_interrupt_queue(&svm->vcpu);
3698 if (!(exitintinfo & SVM_EXITINTINFO_VALID))
3701 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
3703 vector = exitintinfo & SVM_EXITINTINFO_VEC_MASK;
3704 type = exitintinfo & SVM_EXITINTINFO_TYPE_MASK;
3707 case SVM_EXITINTINFO_TYPE_NMI:
3708 svm->vcpu.arch.nmi_injected = true;
3710 case SVM_EXITINTINFO_TYPE_EXEPT:
3712 * Never re-inject a #VC exception.
3714 if (vector == X86_TRAP_VC)
3718 * In case of software exceptions, do not reinject the vector,
3719 * but re-execute the instruction instead. Rewind RIP first
3720 * if we emulated INT3 before.
3722 if (kvm_exception_is_soft(vector)) {
3723 if (vector == BP_VECTOR && int3_injected &&
3724 kvm_is_linear_rip(&svm->vcpu, svm->int3_rip))
3725 kvm_rip_write(&svm->vcpu,
3726 kvm_rip_read(&svm->vcpu) -
3730 if (exitintinfo & SVM_EXITINTINFO_VALID_ERR) {
3731 u32 err = svm->vmcb->control.exit_int_info_err;
3732 kvm_requeue_exception_e(&svm->vcpu, vector, err);
3735 kvm_requeue_exception(&svm->vcpu, vector);
3737 case SVM_EXITINTINFO_TYPE_INTR:
3738 kvm_queue_interrupt(&svm->vcpu, vector, false);
3745 static void svm_cancel_injection(struct kvm_vcpu *vcpu)
3747 struct vcpu_svm *svm = to_svm(vcpu);
3748 struct vmcb_control_area *control = &svm->vmcb->control;
3750 control->exit_int_info = control->event_inj;
3751 control->exit_int_info_err = control->event_inj_err;
3752 control->event_inj = 0;
3753 svm_complete_interrupts(svm);
3756 static fastpath_t svm_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
3758 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_MSR &&
3759 to_svm(vcpu)->vmcb->control.exit_info_1)
3760 return handle_fastpath_set_msr_irqoff(vcpu);
3762 return EXIT_FASTPATH_NONE;
3765 static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu,
3766 struct vcpu_svm *svm)
3769 * VMENTER enables interrupts (host state), but the kernel state is
3770 * interrupts disabled when this is invoked. Also tell RCU about
3771 * it. This is the same logic as for exit_to_user_mode().
3773 * This ensures that e.g. latency analysis on the host observes
3774 * guest mode as interrupt enabled.
3776 * guest_enter_irqoff() informs context tracking about the
3777 * transition to guest mode and if enabled adjusts RCU state
3780 instrumentation_begin();
3781 trace_hardirqs_on_prepare();
3782 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
3783 instrumentation_end();
3785 guest_enter_irqoff();
3786 lockdep_hardirqs_on(CALLER_ADDR0);
3788 if (sev_es_guest(svm->vcpu.kvm)) {
3789 __svm_sev_es_vcpu_run(svm->vmcb_pa);
3791 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3793 __svm_vcpu_run(svm->vmcb_pa, (unsigned long *)&svm->vcpu.arch.regs);
3795 vmload(__sme_page_pa(sd->save_area));
3799 * VMEXIT disables interrupts (host state), but tracing and lockdep
3800 * have them in state 'on' as recorded before entering guest mode.
3801 * Same as enter_from_user_mode().
3803 * guest_exit_irqoff() restores host context and reinstates RCU if
3804 * enabled and required.
3806 * This needs to be done before the below as native_read_msr()
3807 * contains a tracepoint and x86_spec_ctrl_restore_host() calls
3808 * into world and some more.
3810 lockdep_hardirqs_off(CALLER_ADDR0);
3811 guest_exit_irqoff();
3813 instrumentation_begin();
3814 trace_hardirqs_off_finish();
3815 instrumentation_end();
3818 static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
3820 struct vcpu_svm *svm = to_svm(vcpu);
3822 trace_kvm_entry(vcpu);
3824 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
3825 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
3826 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
3829 * Disable singlestep if we're injecting an interrupt/exception.
3830 * We don't want our modified rflags to be pushed on the stack where
3831 * we might not be able to easily reset them if we disabled NMI
3834 if (svm->nmi_singlestep && svm->vmcb->control.event_inj) {
3836 * Event injection happens before external interrupts cause a
3837 * vmexit and interrupts are disabled here, so smp_send_reschedule
3838 * is enough to force an immediate vmexit.
3840 disable_nmi_singlestep(svm);
3841 smp_send_reschedule(vcpu->cpu);
3846 sync_lapic_to_cr8(vcpu);
3848 if (unlikely(svm->asid != svm->vmcb->control.asid)) {
3849 svm->vmcb->control.asid = svm->asid;
3850 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
3852 svm->vmcb->save.cr2 = vcpu->arch.cr2;
3855 * Run with all-zero DR6 unless needed, so that we can get the exact cause
3858 if (unlikely(svm->vcpu.arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT))
3859 svm_set_dr6(svm, vcpu->arch.dr6);
3861 svm_set_dr6(svm, DR6_ACTIVE_LOW);
3864 kvm_load_guest_xsave_state(vcpu);
3866 kvm_wait_lapic_expire(vcpu);
3869 * If this vCPU has touched SPEC_CTRL, restore the guest's value if
3870 * it's non-zero. Since vmentry is serialising on affected CPUs, there
3871 * is no need to worry about the conditional branch over the wrmsr
3872 * being speculatively taken.
3874 x86_spec_ctrl_set_guest(svm->spec_ctrl, svm->virt_spec_ctrl);
3876 svm_vcpu_enter_exit(vcpu, svm);
3879 * We do not use IBRS in the kernel. If this vCPU has used the
3880 * SPEC_CTRL MSR it may have left it on; save the value and
3881 * turn it off. This is much more efficient than blindly adding
3882 * it to the atomic save/restore list. Especially as the former
3883 * (Saving guest MSRs on vmexit) doesn't even exist in KVM.
3885 * For non-nested case:
3886 * If the L01 MSR bitmap does not intercept the MSR, then we need to
3890 * If the L02 MSR bitmap does not intercept the MSR, then we need to
3893 if (unlikely(!msr_write_intercepted(vcpu, MSR_IA32_SPEC_CTRL)))
3894 svm->spec_ctrl = native_read_msr(MSR_IA32_SPEC_CTRL);
3896 if (!sev_es_guest(svm->vcpu.kvm))
3899 x86_spec_ctrl_restore_host(svm->spec_ctrl, svm->virt_spec_ctrl);
3901 if (!sev_es_guest(svm->vcpu.kvm)) {
3902 vcpu->arch.cr2 = svm->vmcb->save.cr2;
3903 vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
3904 vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
3905 vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
3908 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3909 kvm_before_interrupt(&svm->vcpu);
3911 kvm_load_host_xsave_state(vcpu);
3914 /* Any pending NMI will happen here */
3916 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3917 kvm_after_interrupt(&svm->vcpu);
3919 sync_cr8_to_lapic(vcpu);
3922 if (is_guest_mode(&svm->vcpu)) {
3923 sync_nested_vmcb_control(svm);
3924 svm->nested.nested_run_pending = 0;
3927 svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
3928 vmcb_mark_all_clean(svm->vmcb);
3930 /* if exit due to PF check for async PF */
3931 if (svm->vmcb->control.exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR)
3932 svm->vcpu.arch.apf.host_apf_flags =
3933 kvm_read_and_reset_apf_flags();
3936 vcpu->arch.regs_avail &= ~(1 << VCPU_EXREG_PDPTR);
3937 vcpu->arch.regs_dirty &= ~(1 << VCPU_EXREG_PDPTR);
3941 * We need to handle MC intercepts here before the vcpu has a chance to
3942 * change the physical cpu
3944 if (unlikely(svm->vmcb->control.exit_code ==
3945 SVM_EXIT_EXCP_BASE + MC_VECTOR))
3946 svm_handle_mce(svm);
3948 svm_complete_interrupts(svm);
3950 if (is_guest_mode(vcpu))
3951 return EXIT_FASTPATH_NONE;
3953 return svm_exit_handlers_fastpath(vcpu);
3956 static void svm_load_mmu_pgd(struct kvm_vcpu *vcpu, unsigned long root,
3959 struct vcpu_svm *svm = to_svm(vcpu);
3962 cr3 = __sme_set(root);
3964 svm->vmcb->control.nested_cr3 = cr3;
3965 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
3967 /* Loading L2's CR3 is handled by enter_svm_guest_mode. */
3968 if (!test_bit(VCPU_EXREG_CR3, (ulong *)&vcpu->arch.regs_avail))
3970 cr3 = vcpu->arch.cr3;
3973 svm->vmcb->save.cr3 = cr3;
3974 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
3977 static int is_disabled(void)
3981 rdmsrl(MSR_VM_CR, vm_cr);
3982 if (vm_cr & (1 << SVM_VM_CR_SVM_DISABLE))
3989 svm_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
3992 * Patch in the VMMCALL instruction:
3994 hypercall[0] = 0x0f;
3995 hypercall[1] = 0x01;
3996 hypercall[2] = 0xd9;
3999 static int __init svm_check_processor_compat(void)
4004 static bool svm_cpu_has_accelerated_tpr(void)
4010 * The kvm parameter can be NULL (module initialization, or invocation before
4011 * VM creation). Be sure to check the kvm parameter before using it.
4013 static bool svm_has_emulated_msr(struct kvm *kvm, u32 index)
4016 case MSR_IA32_MCG_EXT_CTL:
4017 case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:
4019 case MSR_IA32_SMBASE:
4020 /* SEV-ES guests do not support SMM, so report false */
4021 if (kvm && sev_es_guest(kvm))
4031 static u64 svm_get_mt_mask(struct kvm_vcpu *vcpu, gfn_t gfn, bool is_mmio)
4036 static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
4038 struct vcpu_svm *svm = to_svm(vcpu);
4039 struct kvm_cpuid_entry2 *best;
4041 vcpu->arch.xsaves_enabled = guest_cpuid_has(vcpu, X86_FEATURE_XSAVE) &&
4042 boot_cpu_has(X86_FEATURE_XSAVE) &&
4043 boot_cpu_has(X86_FEATURE_XSAVES);
4045 /* Update nrips enabled cache */
4046 svm->nrips_enabled = kvm_cpu_cap_has(X86_FEATURE_NRIPS) &&
4047 guest_cpuid_has(&svm->vcpu, X86_FEATURE_NRIPS);
4049 /* Check again if INVPCID interception if required */
4050 svm_check_invpcid(svm);
4052 /* For sev guests, the memory encryption bit is not reserved in CR3. */
4053 if (sev_guest(vcpu->kvm)) {
4054 best = kvm_find_cpuid_entry(vcpu, 0x8000001F, 0);
4056 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
4059 if (!kvm_vcpu_apicv_active(vcpu))
4063 * AVIC does not work with an x2APIC mode guest. If the X2APIC feature
4064 * is exposed to the guest, disable AVIC.
4066 if (guest_cpuid_has(vcpu, X86_FEATURE_X2APIC))
4067 kvm_request_apicv_update(vcpu->kvm, false,
4068 APICV_INHIBIT_REASON_X2APIC);
4071 * Currently, AVIC does not work with nested virtualization.
4072 * So, we disable AVIC when cpuid for SVM is set in the L1 guest.
4074 if (nested && guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4075 kvm_request_apicv_update(vcpu->kvm, false,
4076 APICV_INHIBIT_REASON_NESTED);
4079 static bool svm_has_wbinvd_exit(void)
4084 #define PRE_EX(exit) { .exit_code = (exit), \
4085 .stage = X86_ICPT_PRE_EXCEPT, }
4086 #define POST_EX(exit) { .exit_code = (exit), \
4087 .stage = X86_ICPT_POST_EXCEPT, }
4088 #define POST_MEM(exit) { .exit_code = (exit), \
4089 .stage = X86_ICPT_POST_MEMACCESS, }
4091 static const struct __x86_intercept {
4093 enum x86_intercept_stage stage;
4094 } x86_intercept_map[] = {
4095 [x86_intercept_cr_read] = POST_EX(SVM_EXIT_READ_CR0),
4096 [x86_intercept_cr_write] = POST_EX(SVM_EXIT_WRITE_CR0),
4097 [x86_intercept_clts] = POST_EX(SVM_EXIT_WRITE_CR0),
4098 [x86_intercept_lmsw] = POST_EX(SVM_EXIT_WRITE_CR0),
4099 [x86_intercept_smsw] = POST_EX(SVM_EXIT_READ_CR0),
4100 [x86_intercept_dr_read] = POST_EX(SVM_EXIT_READ_DR0),
4101 [x86_intercept_dr_write] = POST_EX(SVM_EXIT_WRITE_DR0),
4102 [x86_intercept_sldt] = POST_EX(SVM_EXIT_LDTR_READ),
4103 [x86_intercept_str] = POST_EX(SVM_EXIT_TR_READ),
4104 [x86_intercept_lldt] = POST_EX(SVM_EXIT_LDTR_WRITE),
4105 [x86_intercept_ltr] = POST_EX(SVM_EXIT_TR_WRITE),
4106 [x86_intercept_sgdt] = POST_EX(SVM_EXIT_GDTR_READ),
4107 [x86_intercept_sidt] = POST_EX(SVM_EXIT_IDTR_READ),
4108 [x86_intercept_lgdt] = POST_EX(SVM_EXIT_GDTR_WRITE),
4109 [x86_intercept_lidt] = POST_EX(SVM_EXIT_IDTR_WRITE),
4110 [x86_intercept_vmrun] = POST_EX(SVM_EXIT_VMRUN),
4111 [x86_intercept_vmmcall] = POST_EX(SVM_EXIT_VMMCALL),
4112 [x86_intercept_vmload] = POST_EX(SVM_EXIT_VMLOAD),
4113 [x86_intercept_vmsave] = POST_EX(SVM_EXIT_VMSAVE),
4114 [x86_intercept_stgi] = POST_EX(SVM_EXIT_STGI),
4115 [x86_intercept_clgi] = POST_EX(SVM_EXIT_CLGI),
4116 [x86_intercept_skinit] = POST_EX(SVM_EXIT_SKINIT),
4117 [x86_intercept_invlpga] = POST_EX(SVM_EXIT_INVLPGA),
4118 [x86_intercept_rdtscp] = POST_EX(SVM_EXIT_RDTSCP),
4119 [x86_intercept_monitor] = POST_MEM(SVM_EXIT_MONITOR),
4120 [x86_intercept_mwait] = POST_EX(SVM_EXIT_MWAIT),
4121 [x86_intercept_invlpg] = POST_EX(SVM_EXIT_INVLPG),
4122 [x86_intercept_invd] = POST_EX(SVM_EXIT_INVD),
4123 [x86_intercept_wbinvd] = POST_EX(SVM_EXIT_WBINVD),
4124 [x86_intercept_wrmsr] = POST_EX(SVM_EXIT_MSR),
4125 [x86_intercept_rdtsc] = POST_EX(SVM_EXIT_RDTSC),
4126 [x86_intercept_rdmsr] = POST_EX(SVM_EXIT_MSR),
4127 [x86_intercept_rdpmc] = POST_EX(SVM_EXIT_RDPMC),
4128 [x86_intercept_cpuid] = PRE_EX(SVM_EXIT_CPUID),
4129 [x86_intercept_rsm] = PRE_EX(SVM_EXIT_RSM),
4130 [x86_intercept_pause] = PRE_EX(SVM_EXIT_PAUSE),
4131 [x86_intercept_pushf] = PRE_EX(SVM_EXIT_PUSHF),
4132 [x86_intercept_popf] = PRE_EX(SVM_EXIT_POPF),
4133 [x86_intercept_intn] = PRE_EX(SVM_EXIT_SWINT),
4134 [x86_intercept_iret] = PRE_EX(SVM_EXIT_IRET),
4135 [x86_intercept_icebp] = PRE_EX(SVM_EXIT_ICEBP),
4136 [x86_intercept_hlt] = POST_EX(SVM_EXIT_HLT),
4137 [x86_intercept_in] = POST_EX(SVM_EXIT_IOIO),
4138 [x86_intercept_ins] = POST_EX(SVM_EXIT_IOIO),
4139 [x86_intercept_out] = POST_EX(SVM_EXIT_IOIO),
4140 [x86_intercept_outs] = POST_EX(SVM_EXIT_IOIO),
4141 [x86_intercept_xsetbv] = PRE_EX(SVM_EXIT_XSETBV),
4148 static int svm_check_intercept(struct kvm_vcpu *vcpu,
4149 struct x86_instruction_info *info,
4150 enum x86_intercept_stage stage,
4151 struct x86_exception *exception)
4153 struct vcpu_svm *svm = to_svm(vcpu);
4154 int vmexit, ret = X86EMUL_CONTINUE;
4155 struct __x86_intercept icpt_info;
4156 struct vmcb *vmcb = svm->vmcb;
4158 if (info->intercept >= ARRAY_SIZE(x86_intercept_map))
4161 icpt_info = x86_intercept_map[info->intercept];
4163 if (stage != icpt_info.stage)
4166 switch (icpt_info.exit_code) {
4167 case SVM_EXIT_READ_CR0:
4168 if (info->intercept == x86_intercept_cr_read)
4169 icpt_info.exit_code += info->modrm_reg;
4171 case SVM_EXIT_WRITE_CR0: {
4172 unsigned long cr0, val;
4174 if (info->intercept == x86_intercept_cr_write)
4175 icpt_info.exit_code += info->modrm_reg;
4177 if (icpt_info.exit_code != SVM_EXIT_WRITE_CR0 ||
4178 info->intercept == x86_intercept_clts)
4181 if (!(vmcb_is_intercept(&svm->nested.ctl,
4182 INTERCEPT_SELECTIVE_CR0)))
4185 cr0 = vcpu->arch.cr0 & ~SVM_CR0_SELECTIVE_MASK;
4186 val = info->src_val & ~SVM_CR0_SELECTIVE_MASK;
4188 if (info->intercept == x86_intercept_lmsw) {
4191 /* lmsw can't clear PE - catch this here */
4192 if (cr0 & X86_CR0_PE)
4197 icpt_info.exit_code = SVM_EXIT_CR0_SEL_WRITE;
4201 case SVM_EXIT_READ_DR0:
4202 case SVM_EXIT_WRITE_DR0:
4203 icpt_info.exit_code += info->modrm_reg;
4206 if (info->intercept == x86_intercept_wrmsr)
4207 vmcb->control.exit_info_1 = 1;
4209 vmcb->control.exit_info_1 = 0;
4211 case SVM_EXIT_PAUSE:
4213 * We get this for NOP only, but pause
4214 * is rep not, check this here
4216 if (info->rep_prefix != REPE_PREFIX)
4219 case SVM_EXIT_IOIO: {
4223 if (info->intercept == x86_intercept_in ||
4224 info->intercept == x86_intercept_ins) {
4225 exit_info = ((info->src_val & 0xffff) << 16) |
4227 bytes = info->dst_bytes;
4229 exit_info = (info->dst_val & 0xffff) << 16;
4230 bytes = info->src_bytes;
4233 if (info->intercept == x86_intercept_outs ||
4234 info->intercept == x86_intercept_ins)
4235 exit_info |= SVM_IOIO_STR_MASK;
4237 if (info->rep_prefix)
4238 exit_info |= SVM_IOIO_REP_MASK;
4240 bytes = min(bytes, 4u);
4242 exit_info |= bytes << SVM_IOIO_SIZE_SHIFT;
4244 exit_info |= (u32)info->ad_bytes << (SVM_IOIO_ASIZE_SHIFT - 1);
4246 vmcb->control.exit_info_1 = exit_info;
4247 vmcb->control.exit_info_2 = info->next_rip;
4255 /* TODO: Advertise NRIPS to guest hypervisor unconditionally */
4256 if (static_cpu_has(X86_FEATURE_NRIPS))
4257 vmcb->control.next_rip = info->next_rip;
4258 vmcb->control.exit_code = icpt_info.exit_code;
4259 vmexit = nested_svm_exit_handled(svm);
4261 ret = (vmexit == NESTED_EXIT_DONE) ? X86EMUL_INTERCEPTED
4268 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
4272 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
4274 if (!kvm_pause_in_guest(vcpu->kvm))
4275 shrink_ple_window(vcpu);
4278 static void svm_setup_mce(struct kvm_vcpu *vcpu)
4280 /* [63:9] are reserved. */
4281 vcpu->arch.mcg_cap &= 0x1ff;
4284 bool svm_smi_blocked(struct kvm_vcpu *vcpu)
4286 struct vcpu_svm *svm = to_svm(vcpu);
4288 /* Per APM Vol.2 15.22.2 "Response to SMI" */
4292 return is_smm(vcpu);
4295 static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
4297 struct vcpu_svm *svm = to_svm(vcpu);
4298 if (svm->nested.nested_run_pending)
4301 /* An SMI must not be injected into L2 if it's supposed to VM-Exit. */
4302 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_smi(svm))
4305 return !svm_smi_blocked(vcpu);
4308 static int svm_pre_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
4310 struct vcpu_svm *svm = to_svm(vcpu);
4313 if (is_guest_mode(vcpu)) {
4314 /* FED8h - SVM Guest */
4315 put_smstate(u64, smstate, 0x7ed8, 1);
4316 /* FEE0h - SVM Guest VMCB Physical Address */
4317 put_smstate(u64, smstate, 0x7ee0, svm->nested.vmcb12_gpa);
4319 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4320 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4321 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4323 ret = nested_svm_vmexit(svm);
4330 static int svm_pre_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
4332 struct vcpu_svm *svm = to_svm(vcpu);
4333 struct kvm_host_map map;
4336 if (guest_cpuid_has(vcpu, X86_FEATURE_LM)) {
4337 u64 saved_efer = GET_SMSTATE(u64, smstate, 0x7ed0);
4338 u64 guest = GET_SMSTATE(u64, smstate, 0x7ed8);
4339 u64 vmcb12_gpa = GET_SMSTATE(u64, smstate, 0x7ee0);
4342 if (!guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4345 if (!(saved_efer & EFER_SVME))
4348 if (kvm_vcpu_map(&svm->vcpu,
4349 gpa_to_gfn(vmcb12_gpa), &map) == -EINVAL)
4352 if (svm_allocate_nested(svm))
4355 ret = enter_svm_guest_mode(svm, vmcb12_gpa, map.hva);
4356 kvm_vcpu_unmap(&svm->vcpu, &map, true);
4363 static void svm_enable_smi_window(struct kvm_vcpu *vcpu)
4365 struct vcpu_svm *svm = to_svm(vcpu);
4367 if (!gif_set(svm)) {
4368 if (vgif_enabled(svm))
4369 svm_set_intercept(svm, INTERCEPT_STGI);
4370 /* STGI will cause a vm exit */
4372 /* We must be in SMM; RSM will cause a vmexit anyway. */
4376 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, void *insn, int insn_len)
4378 bool smep, smap, is_user;
4382 * When the guest is an SEV-ES guest, emulation is not possible.
4384 if (sev_es_guest(vcpu->kvm))
4388 * Detect and workaround Errata 1096 Fam_17h_00_0Fh.
4391 * When CPU raise #NPF on guest data access and vCPU CR4.SMAP=1, it is
4392 * possible that CPU microcode implementing DecodeAssist will fail
4393 * to read bytes of instruction which caused #NPF. In this case,
4394 * GuestIntrBytes field of the VMCB on a VMEXIT will incorrectly
4395 * return 0 instead of the correct guest instruction bytes.
4397 * This happens because CPU microcode reading instruction bytes
4398 * uses a special opcode which attempts to read data using CPL=0
4399 * priviledges. The microcode reads CS:RIP and if it hits a SMAP
4400 * fault, it gives up and returns no instruction bytes.
4403 * We reach here in case CPU supports DecodeAssist, raised #NPF and
4404 * returned 0 in GuestIntrBytes field of the VMCB.
4405 * First, errata can only be triggered in case vCPU CR4.SMAP=1.
4406 * Second, if vCPU CR4.SMEP=1, errata could only be triggered
4407 * in case vCPU CPL==3 (Because otherwise guest would have triggered
4408 * a SMEP fault instead of #NPF).
4409 * Otherwise, vCPU CR4.SMEP=0, errata could be triggered by any vCPU CPL.
4410 * As most guests enable SMAP if they have also enabled SMEP, use above
4411 * logic in order to attempt minimize false-positive of detecting errata
4412 * while still preserving all cases semantic correctness.
4415 * To determine what instruction the guest was executing, the hypervisor
4416 * will have to decode the instruction at the instruction pointer.
4418 * In non SEV guest, hypervisor will be able to read the guest
4419 * memory to decode the instruction pointer when insn_len is zero
4420 * so we return true to indicate that decoding is possible.
4422 * But in the SEV guest, the guest memory is encrypted with the
4423 * guest specific key and hypervisor will not be able to decode the
4424 * instruction pointer so we will not able to workaround it. Lets
4425 * print the error and request to kill the guest.
4427 if (likely(!insn || insn_len))
4431 * If RIP is invalid, go ahead with emulation which will cause an
4432 * internal error exit.
4434 if (!kvm_vcpu_gfn_to_memslot(vcpu, kvm_rip_read(vcpu) >> PAGE_SHIFT))
4437 cr4 = kvm_read_cr4(vcpu);
4438 smep = cr4 & X86_CR4_SMEP;
4439 smap = cr4 & X86_CR4_SMAP;
4440 is_user = svm_get_cpl(vcpu) == 3;
4441 if (smap && (!smep || is_user)) {
4442 if (!sev_guest(vcpu->kvm))
4445 pr_err_ratelimited("KVM: SEV Guest triggered AMD Erratum 1096\n");
4446 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
4452 static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
4454 struct vcpu_svm *svm = to_svm(vcpu);
4457 * TODO: Last condition latch INIT signals on vCPU when
4458 * vCPU is in guest-mode and vmcb12 defines intercept on INIT.
4459 * To properly emulate the INIT intercept,
4460 * svm_check_nested_events() should call nested_svm_vmexit()
4461 * if an INIT signal is pending.
4463 return !gif_set(svm) ||
4464 (vmcb_is_intercept(&svm->vmcb->control, INTERCEPT_INIT));
4467 static void svm_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
4469 if (!sev_es_guest(vcpu->kvm))
4470 return kvm_vcpu_deliver_sipi_vector(vcpu, vector);
4472 sev_vcpu_deliver_sipi_vector(vcpu, vector);
4475 static void svm_vm_destroy(struct kvm *kvm)
4477 avic_vm_destroy(kvm);
4478 sev_vm_destroy(kvm);
4481 static int svm_vm_init(struct kvm *kvm)
4483 if (!pause_filter_count || !pause_filter_thresh)
4484 kvm->arch.pause_in_guest = true;
4487 int ret = avic_vm_init(kvm);
4492 kvm_apicv_init(kvm, avic);
4496 static struct kvm_x86_ops svm_x86_ops __initdata = {
4497 .hardware_unsetup = svm_hardware_teardown,
4498 .hardware_enable = svm_hardware_enable,
4499 .hardware_disable = svm_hardware_disable,
4500 .cpu_has_accelerated_tpr = svm_cpu_has_accelerated_tpr,
4501 .has_emulated_msr = svm_has_emulated_msr,
4503 .vcpu_create = svm_create_vcpu,
4504 .vcpu_free = svm_free_vcpu,
4505 .vcpu_reset = svm_vcpu_reset,
4507 .vm_size = sizeof(struct kvm_svm),
4508 .vm_init = svm_vm_init,
4509 .vm_destroy = svm_vm_destroy,
4511 .prepare_guest_switch = svm_prepare_guest_switch,
4512 .vcpu_load = svm_vcpu_load,
4513 .vcpu_put = svm_vcpu_put,
4514 .vcpu_blocking = svm_vcpu_blocking,
4515 .vcpu_unblocking = svm_vcpu_unblocking,
4517 .update_exception_bitmap = svm_update_exception_bitmap,
4518 .get_msr_feature = svm_get_msr_feature,
4519 .get_msr = svm_get_msr,
4520 .set_msr = svm_set_msr,
4521 .get_segment_base = svm_get_segment_base,
4522 .get_segment = svm_get_segment,
4523 .set_segment = svm_set_segment,
4524 .get_cpl = svm_get_cpl,
4525 .get_cs_db_l_bits = kvm_get_cs_db_l_bits,
4526 .set_cr0 = svm_set_cr0,
4527 .is_valid_cr4 = svm_is_valid_cr4,
4528 .set_cr4 = svm_set_cr4,
4529 .set_efer = svm_set_efer,
4530 .get_idt = svm_get_idt,
4531 .set_idt = svm_set_idt,
4532 .get_gdt = svm_get_gdt,
4533 .set_gdt = svm_set_gdt,
4534 .set_dr7 = svm_set_dr7,
4535 .sync_dirty_debug_regs = svm_sync_dirty_debug_regs,
4536 .cache_reg = svm_cache_reg,
4537 .get_rflags = svm_get_rflags,
4538 .set_rflags = svm_set_rflags,
4540 .tlb_flush_all = svm_flush_tlb,
4541 .tlb_flush_current = svm_flush_tlb,
4542 .tlb_flush_gva = svm_flush_tlb_gva,
4543 .tlb_flush_guest = svm_flush_tlb,
4545 .run = svm_vcpu_run,
4546 .handle_exit = handle_exit,
4547 .skip_emulated_instruction = skip_emulated_instruction,
4548 .update_emulated_instruction = NULL,
4549 .set_interrupt_shadow = svm_set_interrupt_shadow,
4550 .get_interrupt_shadow = svm_get_interrupt_shadow,
4551 .patch_hypercall = svm_patch_hypercall,
4552 .set_irq = svm_set_irq,
4553 .set_nmi = svm_inject_nmi,
4554 .queue_exception = svm_queue_exception,
4555 .cancel_injection = svm_cancel_injection,
4556 .interrupt_allowed = svm_interrupt_allowed,
4557 .nmi_allowed = svm_nmi_allowed,
4558 .get_nmi_mask = svm_get_nmi_mask,
4559 .set_nmi_mask = svm_set_nmi_mask,
4560 .enable_nmi_window = svm_enable_nmi_window,
4561 .enable_irq_window = svm_enable_irq_window,
4562 .update_cr8_intercept = svm_update_cr8_intercept,
4563 .set_virtual_apic_mode = svm_set_virtual_apic_mode,
4564 .refresh_apicv_exec_ctrl = svm_refresh_apicv_exec_ctrl,
4565 .check_apicv_inhibit_reasons = svm_check_apicv_inhibit_reasons,
4566 .pre_update_apicv_exec_ctrl = svm_pre_update_apicv_exec_ctrl,
4567 .load_eoi_exitmap = svm_load_eoi_exitmap,
4568 .hwapic_irr_update = svm_hwapic_irr_update,
4569 .hwapic_isr_update = svm_hwapic_isr_update,
4570 .sync_pir_to_irr = kvm_lapic_find_highest_irr,
4571 .apicv_post_state_restore = avic_post_state_restore,
4573 .set_tss_addr = svm_set_tss_addr,
4574 .set_identity_map_addr = svm_set_identity_map_addr,
4575 .get_mt_mask = svm_get_mt_mask,
4577 .get_exit_info = svm_get_exit_info,
4579 .vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
4581 .has_wbinvd_exit = svm_has_wbinvd_exit,
4583 .write_l1_tsc_offset = svm_write_l1_tsc_offset,
4585 .load_mmu_pgd = svm_load_mmu_pgd,
4587 .check_intercept = svm_check_intercept,
4588 .handle_exit_irqoff = svm_handle_exit_irqoff,
4590 .request_immediate_exit = __kvm_request_immediate_exit,
4592 .sched_in = svm_sched_in,
4594 .pmu_ops = &amd_pmu_ops,
4595 .nested_ops = &svm_nested_ops,
4597 .deliver_posted_interrupt = svm_deliver_avic_intr,
4598 .dy_apicv_has_pending_interrupt = svm_dy_apicv_has_pending_interrupt,
4599 .update_pi_irte = svm_update_pi_irte,
4600 .setup_mce = svm_setup_mce,
4602 .smi_allowed = svm_smi_allowed,
4603 .pre_enter_smm = svm_pre_enter_smm,
4604 .pre_leave_smm = svm_pre_leave_smm,
4605 .enable_smi_window = svm_enable_smi_window,
4607 .mem_enc_op = svm_mem_enc_op,
4608 .mem_enc_reg_region = svm_register_enc_region,
4609 .mem_enc_unreg_region = svm_unregister_enc_region,
4611 .can_emulate_instruction = svm_can_emulate_instruction,
4613 .apic_init_signal_blocked = svm_apic_init_signal_blocked,
4615 .msr_filter_changed = svm_msr_filter_changed,
4616 .complete_emulated_msr = svm_complete_emulated_msr,
4618 .vcpu_deliver_sipi_vector = svm_vcpu_deliver_sipi_vector,
4621 static struct kvm_x86_init_ops svm_init_ops __initdata = {
4622 .cpu_has_kvm_support = has_svm,
4623 .disabled_by_bios = is_disabled,
4624 .hardware_setup = svm_hardware_setup,
4625 .check_processor_compatibility = svm_check_processor_compat,
4627 .runtime_ops = &svm_x86_ops,
4630 static int __init svm_init(void)
4632 __unused_size_checks();
4634 return kvm_init(&svm_init_ops, sizeof(struct vcpu_svm),
4635 __alignof__(struct vcpu_svm), THIS_MODULE);
4638 static void __exit svm_exit(void)
4643 module_init(svm_init)
4644 module_exit(svm_exit)