1 #define pr_fmt(fmt) "SVM: " fmt
3 #include <linux/kvm_host.h>
7 #include "kvm_cache_regs.h"
12 #include <linux/module.h>
13 #include <linux/mod_devicetable.h>
14 #include <linux/kernel.h>
15 #include <linux/vmalloc.h>
16 #include <linux/highmem.h>
17 #include <linux/amd-iommu.h>
18 #include <linux/sched.h>
19 #include <linux/trace_events.h>
20 #include <linux/slab.h>
21 #include <linux/hashtable.h>
22 #include <linux/objtool.h>
23 #include <linux/psp-sev.h>
24 #include <linux/file.h>
25 #include <linux/pagemap.h>
26 #include <linux/swap.h>
27 #include <linux/rwsem.h>
28 #include <linux/cc_platform.h>
31 #include <asm/perf_event.h>
32 #include <asm/tlbflush.h>
34 #include <asm/debugreg.h>
35 #include <asm/kvm_para.h>
36 #include <asm/irq_remapping.h>
37 #include <asm/spec-ctrl.h>
38 #include <asm/cpu_device_id.h>
39 #include <asm/traps.h>
40 #include <asm/fpu/api.h>
42 #include <asm/virtext.h>
48 #include "kvm_onhyperv.h"
49 #include "svm_onhyperv.h"
51 MODULE_AUTHOR("Qumranet");
52 MODULE_LICENSE("GPL");
55 static const struct x86_cpu_id svm_cpu_id[] = {
56 X86_MATCH_FEATURE(X86_FEATURE_SVM, NULL),
59 MODULE_DEVICE_TABLE(x86cpu, svm_cpu_id);
62 #define SEG_TYPE_LDT 2
63 #define SEG_TYPE_BUSY_TSS16 3
65 static bool erratum_383_found __read_mostly;
67 u32 msrpm_offsets[MSRPM_OFFSETS] __read_mostly;
70 * Set osvw_len to higher value when updated Revision Guides
71 * are published and we know what the new status bits are
73 static uint64_t osvw_len = 4, osvw_status;
75 static DEFINE_PER_CPU(u64, current_tsc_ratio);
77 static const struct svm_direct_access_msrs {
78 u32 index; /* Index of the MSR */
79 bool always; /* True if intercept is initially cleared */
80 } direct_access_msrs[MAX_DIRECT_ACCESS_MSRS] = {
81 { .index = MSR_STAR, .always = true },
82 { .index = MSR_IA32_SYSENTER_CS, .always = true },
83 { .index = MSR_IA32_SYSENTER_EIP, .always = false },
84 { .index = MSR_IA32_SYSENTER_ESP, .always = false },
86 { .index = MSR_GS_BASE, .always = true },
87 { .index = MSR_FS_BASE, .always = true },
88 { .index = MSR_KERNEL_GS_BASE, .always = true },
89 { .index = MSR_LSTAR, .always = true },
90 { .index = MSR_CSTAR, .always = true },
91 { .index = MSR_SYSCALL_MASK, .always = true },
93 { .index = MSR_IA32_SPEC_CTRL, .always = false },
94 { .index = MSR_IA32_PRED_CMD, .always = false },
95 { .index = MSR_IA32_LASTBRANCHFROMIP, .always = false },
96 { .index = MSR_IA32_LASTBRANCHTOIP, .always = false },
97 { .index = MSR_IA32_LASTINTFROMIP, .always = false },
98 { .index = MSR_IA32_LASTINTTOIP, .always = false },
99 { .index = MSR_EFER, .always = false },
100 { .index = MSR_IA32_CR_PAT, .always = false },
101 { .index = MSR_AMD64_SEV_ES_GHCB, .always = true },
102 { .index = MSR_TSC_AUX, .always = false },
103 { .index = MSR_INVALID, .always = false },
107 * These 2 parameters are used to config the controls for Pause-Loop Exiting:
108 * pause_filter_count: On processors that support Pause filtering(indicated
109 * by CPUID Fn8000_000A_EDX), the VMCB provides a 16 bit pause filter
110 * count value. On VMRUN this value is loaded into an internal counter.
111 * Each time a pause instruction is executed, this counter is decremented
112 * until it reaches zero at which time a #VMEXIT is generated if pause
113 * intercept is enabled. Refer to AMD APM Vol 2 Section 15.14.4 Pause
114 * Intercept Filtering for more details.
115 * This also indicate if ple logic enabled.
117 * pause_filter_thresh: In addition, some processor families support advanced
118 * pause filtering (indicated by CPUID Fn8000_000A_EDX) upper bound on
119 * the amount of time a guest is allowed to execute in a pause loop.
120 * In this mode, a 16-bit pause filter threshold field is added in the
121 * VMCB. The threshold value is a cycle count that is used to reset the
122 * pause counter. As with simple pause filtering, VMRUN loads the pause
123 * count value from VMCB into an internal counter. Then, on each pause
124 * instruction the hardware checks the elapsed number of cycles since
125 * the most recent pause instruction against the pause filter threshold.
126 * If the elapsed cycle count is greater than the pause filter threshold,
127 * then the internal pause count is reloaded from the VMCB and execution
128 * continues. If the elapsed cycle count is less than the pause filter
129 * threshold, then the internal pause count is decremented. If the count
130 * value is less than zero and PAUSE intercept is enabled, a #VMEXIT is
131 * triggered. If advanced pause filtering is supported and pause filter
132 * threshold field is set to zero, the filter will operate in the simpler,
136 static unsigned short pause_filter_thresh = KVM_DEFAULT_PLE_GAP;
137 module_param(pause_filter_thresh, ushort, 0444);
139 static unsigned short pause_filter_count = KVM_SVM_DEFAULT_PLE_WINDOW;
140 module_param(pause_filter_count, ushort, 0444);
142 /* Default doubles per-vcpu window every exit. */
143 static unsigned short pause_filter_count_grow = KVM_DEFAULT_PLE_WINDOW_GROW;
144 module_param(pause_filter_count_grow, ushort, 0444);
146 /* Default resets per-vcpu window every exit to pause_filter_count. */
147 static unsigned short pause_filter_count_shrink = KVM_DEFAULT_PLE_WINDOW_SHRINK;
148 module_param(pause_filter_count_shrink, ushort, 0444);
150 /* Default is to compute the maximum so we can never overflow. */
151 static unsigned short pause_filter_count_max = KVM_SVM_DEFAULT_PLE_WINDOW_MAX;
152 module_param(pause_filter_count_max, ushort, 0444);
155 * Use nested page tables by default. Note, NPT may get forced off by
156 * svm_hardware_setup() if it's unsupported by hardware or the host kernel.
158 bool npt_enabled = true;
159 module_param_named(npt, npt_enabled, bool, 0444);
161 /* allow nested virtualization in KVM/SVM */
162 static int nested = true;
163 module_param(nested, int, S_IRUGO);
165 /* enable/disable Next RIP Save */
166 static int nrips = true;
167 module_param(nrips, int, 0444);
169 /* enable/disable Virtual VMLOAD VMSAVE */
170 static int vls = true;
171 module_param(vls, int, 0444);
173 /* enable/disable Virtual GIF */
175 module_param(vgif, int, 0444);
177 /* enable/disable LBR virtualization */
178 static int lbrv = true;
179 module_param(lbrv, int, 0444);
181 static int tsc_scaling = true;
182 module_param(tsc_scaling, int, 0444);
185 * enable / disable AVIC. Because the defaults differ for APICv
186 * support between VMX and SVM we cannot use module_param_named.
189 module_param(avic, bool, 0444);
191 bool __read_mostly dump_invalid_vmcb;
192 module_param(dump_invalid_vmcb, bool, 0644);
195 bool intercept_smi = true;
196 module_param(intercept_smi, bool, 0444);
199 static bool svm_gp_erratum_intercept = true;
201 static u8 rsm_ins_bytes[] = "\x0f\xaa";
203 static unsigned long iopm_base;
205 struct kvm_ldttss_desc {
208 unsigned base1:8, type:5, dpl:2, p:1;
209 unsigned limit1:4, zero0:3, g:1, base2:8;
212 } __attribute__((packed));
214 DEFINE_PER_CPU(struct svm_cpu_data *, svm_data);
217 * Only MSR_TSC_AUX is switched via the user return hook. EFER is switched via
218 * the VMCB, and the SYSCALL/SYSENTER MSRs are handled by VMLOAD/VMSAVE.
220 * RDTSCP and RDPID are not used in the kernel, specifically to allow KVM to
221 * defer the restoration of TSC_AUX until the CPU returns to userspace.
223 static int tsc_aux_uret_slot __read_mostly = -1;
225 static const u32 msrpm_ranges[] = {0, 0xc0000000, 0xc0010000};
227 #define NUM_MSR_MAPS ARRAY_SIZE(msrpm_ranges)
228 #define MSRS_RANGE_SIZE 2048
229 #define MSRS_IN_RANGE (MSRS_RANGE_SIZE * 8 / 2)
231 u32 svm_msrpm_offset(u32 msr)
236 for (i = 0; i < NUM_MSR_MAPS; i++) {
237 if (msr < msrpm_ranges[i] ||
238 msr >= msrpm_ranges[i] + MSRS_IN_RANGE)
241 offset = (msr - msrpm_ranges[i]) / 4; /* 4 msrs per u8 */
242 offset += (i * MSRS_RANGE_SIZE); /* add range offset */
244 /* Now we have the u8 offset - but need the u32 offset */
248 /* MSR not in any range */
252 static void svm_flush_tlb_current(struct kvm_vcpu *vcpu);
254 static int get_npt_level(void)
257 return pgtable_l5_enabled() ? PT64_ROOT_5LEVEL : PT64_ROOT_4LEVEL;
259 return PT32E_ROOT_LEVEL;
263 int svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)
265 struct vcpu_svm *svm = to_svm(vcpu);
266 u64 old_efer = vcpu->arch.efer;
267 vcpu->arch.efer = efer;
270 /* Shadow paging assumes NX to be available. */
273 if (!(efer & EFER_LMA))
277 if ((old_efer & EFER_SVME) != (efer & EFER_SVME)) {
278 if (!(efer & EFER_SVME)) {
279 svm_leave_nested(vcpu);
280 svm_set_gif(svm, true);
281 /* #GP intercept is still needed for vmware backdoor */
282 if (!enable_vmware_backdoor)
283 clr_exception_intercept(svm, GP_VECTOR);
286 * Free the nested guest state, unless we are in SMM.
287 * In this case we will return to the nested guest
288 * as soon as we leave SMM.
291 svm_free_nested(svm);
294 int ret = svm_allocate_nested(svm);
297 vcpu->arch.efer = old_efer;
302 * Never intercept #GP for SEV guests, KVM can't
303 * decrypt guest memory to workaround the erratum.
305 if (svm_gp_erratum_intercept && !sev_guest(vcpu->kvm))
306 set_exception_intercept(svm, GP_VECTOR);
310 svm->vmcb->save.efer = efer | EFER_SVME;
311 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
315 static int is_external_interrupt(u32 info)
317 info &= SVM_EVTINJ_TYPE_MASK | SVM_EVTINJ_VALID;
318 return info == (SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR);
321 static u32 svm_get_interrupt_shadow(struct kvm_vcpu *vcpu)
323 struct vcpu_svm *svm = to_svm(vcpu);
326 if (svm->vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK)
327 ret = KVM_X86_SHADOW_INT_STI | KVM_X86_SHADOW_INT_MOV_SS;
331 static void svm_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
333 struct vcpu_svm *svm = to_svm(vcpu);
336 svm->vmcb->control.int_state &= ~SVM_INTERRUPT_SHADOW_MASK;
338 svm->vmcb->control.int_state |= SVM_INTERRUPT_SHADOW_MASK;
342 static int __svm_skip_emulated_instruction(struct kvm_vcpu *vcpu,
343 bool commit_side_effects)
345 struct vcpu_svm *svm = to_svm(vcpu);
346 unsigned long old_rflags;
349 * SEV-ES does not expose the next RIP. The RIP update is controlled by
350 * the type of exit and the #VC handler in the guest.
352 if (sev_es_guest(vcpu->kvm))
355 if (nrips && svm->vmcb->control.next_rip != 0) {
356 WARN_ON_ONCE(!static_cpu_has(X86_FEATURE_NRIPS));
357 svm->next_rip = svm->vmcb->control.next_rip;
360 if (!svm->next_rip) {
361 if (unlikely(!commit_side_effects))
362 old_rflags = svm->vmcb->save.rflags;
364 if (!kvm_emulate_instruction(vcpu, EMULTYPE_SKIP))
367 if (unlikely(!commit_side_effects))
368 svm->vmcb->save.rflags = old_rflags;
370 kvm_rip_write(vcpu, svm->next_rip);
374 if (likely(commit_side_effects))
375 svm_set_interrupt_shadow(vcpu, 0);
380 static int svm_skip_emulated_instruction(struct kvm_vcpu *vcpu)
382 return __svm_skip_emulated_instruction(vcpu, true);
385 static int svm_update_soft_interrupt_rip(struct kvm_vcpu *vcpu)
387 unsigned long rip, old_rip = kvm_rip_read(vcpu);
388 struct vcpu_svm *svm = to_svm(vcpu);
391 * Due to architectural shortcomings, the CPU doesn't always provide
392 * NextRIP, e.g. if KVM intercepted an exception that occurred while
393 * the CPU was vectoring an INTO/INT3 in the guest. Temporarily skip
394 * the instruction even if NextRIP is supported to acquire the next
395 * RIP so that it can be shoved into the NextRIP field, otherwise
396 * hardware will fail to advance guest RIP during event injection.
397 * Drop the exception/interrupt if emulation fails and effectively
398 * retry the instruction, it's the least awful option. If NRIPS is
399 * in use, the skip must not commit any side effects such as clearing
400 * the interrupt shadow or RFLAGS.RF.
402 if (!__svm_skip_emulated_instruction(vcpu, !nrips))
405 rip = kvm_rip_read(vcpu);
408 * Save the injection information, even when using next_rip, as the
409 * VMCB's next_rip will be lost (cleared on VM-Exit) if the injection
410 * doesn't complete due to a VM-Exit occurring while the CPU is
411 * vectoring the event. Decoding the instruction isn't guaranteed to
412 * work as there may be no backing instruction, e.g. if the event is
413 * being injected by L1 for L2, or if the guest is patching INT3 into
414 * a different instruction.
416 svm->soft_int_injected = true;
417 svm->soft_int_csbase = svm->vmcb->save.cs.base;
418 svm->soft_int_old_rip = old_rip;
419 svm->soft_int_next_rip = rip;
422 kvm_rip_write(vcpu, old_rip);
424 if (static_cpu_has(X86_FEATURE_NRIPS))
425 svm->vmcb->control.next_rip = rip;
430 static void svm_queue_exception(struct kvm_vcpu *vcpu)
432 struct vcpu_svm *svm = to_svm(vcpu);
433 unsigned nr = vcpu->arch.exception.nr;
434 bool has_error_code = vcpu->arch.exception.has_error_code;
435 u32 error_code = vcpu->arch.exception.error_code;
437 kvm_deliver_exception_payload(vcpu);
439 if (kvm_exception_is_soft(nr) &&
440 svm_update_soft_interrupt_rip(vcpu))
443 svm->vmcb->control.event_inj = nr
445 | (has_error_code ? SVM_EVTINJ_VALID_ERR : 0)
446 | SVM_EVTINJ_TYPE_EXEPT;
447 svm->vmcb->control.event_inj_err = error_code;
450 static void svm_init_erratum_383(void)
456 if (!static_cpu_has_bug(X86_BUG_AMD_TLB_MMATCH))
459 /* Use _safe variants to not break nested virtualization */
460 val = native_read_msr_safe(MSR_AMD64_DC_CFG, &err);
466 low = lower_32_bits(val);
467 high = upper_32_bits(val);
469 native_write_msr_safe(MSR_AMD64_DC_CFG, low, high);
471 erratum_383_found = true;
474 static void svm_init_osvw(struct kvm_vcpu *vcpu)
477 * Guests should see errata 400 and 415 as fixed (assuming that
478 * HLT and IO instructions are intercepted).
480 vcpu->arch.osvw.length = (osvw_len >= 3) ? (osvw_len) : 3;
481 vcpu->arch.osvw.status = osvw_status & ~(6ULL);
484 * By increasing VCPU's osvw.length to 3 we are telling the guest that
485 * all osvw.status bits inside that length, including bit 0 (which is
486 * reserved for erratum 298), are valid. However, if host processor's
487 * osvw_len is 0 then osvw_status[0] carries no information. We need to
488 * be conservative here and therefore we tell the guest that erratum 298
489 * is present (because we really don't know).
491 if (osvw_len == 0 && boot_cpu_data.x86 == 0x10)
492 vcpu->arch.osvw.status |= 1;
495 static int has_svm(void)
499 if (!cpu_has_svm(&msg)) {
500 printk(KERN_INFO "has_svm: %s\n", msg);
504 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) {
505 pr_info("KVM is unsupported when running as an SEV guest\n");
512 void __svm_write_tsc_multiplier(u64 multiplier)
516 if (multiplier == __this_cpu_read(current_tsc_ratio))
519 wrmsrl(MSR_AMD64_TSC_RATIO, multiplier);
520 __this_cpu_write(current_tsc_ratio, multiplier);
525 static void svm_hardware_disable(void)
527 /* Make sure we clean up behind us */
529 __svm_write_tsc_multiplier(SVM_TSC_RATIO_DEFAULT);
533 amd_pmu_disable_virt();
536 static int svm_hardware_enable(void)
539 struct svm_cpu_data *sd;
541 struct desc_struct *gdt;
542 int me = raw_smp_processor_id();
544 rdmsrl(MSR_EFER, efer);
545 if (efer & EFER_SVME)
549 pr_err("%s: err EOPNOTSUPP on %d\n", __func__, me);
552 sd = per_cpu(svm_data, me);
554 pr_err("%s: svm_data is NULL on %d\n", __func__, me);
558 sd->asid_generation = 1;
559 sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
560 sd->next_asid = sd->max_asid + 1;
561 sd->min_asid = max_sev_asid + 1;
563 gdt = get_current_gdt_rw();
564 sd->tss_desc = (struct kvm_ldttss_desc *)(gdt + GDT_ENTRY_TSS);
566 wrmsrl(MSR_EFER, efer | EFER_SVME);
568 wrmsrl(MSR_VM_HSAVE_PA, __sme_page_pa(sd->save_area));
570 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
572 * Set the default value, even if we don't use TSC scaling
573 * to avoid having stale value in the msr
575 __svm_write_tsc_multiplier(SVM_TSC_RATIO_DEFAULT);
582 * Note that it is possible to have a system with mixed processor
583 * revisions and therefore different OSVW bits. If bits are not the same
584 * on different processors then choose the worst case (i.e. if erratum
585 * is present on one processor and not on another then assume that the
586 * erratum is present everywhere).
588 if (cpu_has(&boot_cpu_data, X86_FEATURE_OSVW)) {
589 uint64_t len, status = 0;
592 len = native_read_msr_safe(MSR_AMD64_OSVW_ID_LENGTH, &err);
594 status = native_read_msr_safe(MSR_AMD64_OSVW_STATUS,
598 osvw_status = osvw_len = 0;
602 osvw_status |= status;
603 osvw_status &= (1ULL << osvw_len) - 1;
606 osvw_status = osvw_len = 0;
608 svm_init_erratum_383();
610 amd_pmu_enable_virt();
615 static void svm_cpu_uninit(int cpu)
617 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
622 per_cpu(svm_data, cpu) = NULL;
623 kfree(sd->sev_vmcbs);
624 __free_page(sd->save_area);
628 static int svm_cpu_init(int cpu)
630 struct svm_cpu_data *sd;
633 sd = kzalloc(sizeof(struct svm_cpu_data), GFP_KERNEL);
637 sd->save_area = alloc_page(GFP_KERNEL | __GFP_ZERO);
641 ret = sev_cpu_init(sd);
645 per_cpu(svm_data, cpu) = sd;
650 __free_page(sd->save_area);
657 static int direct_access_msr_slot(u32 msr)
661 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++)
662 if (direct_access_msrs[i].index == msr)
668 static void set_shadow_msr_intercept(struct kvm_vcpu *vcpu, u32 msr, int read,
671 struct vcpu_svm *svm = to_svm(vcpu);
672 int slot = direct_access_msr_slot(msr);
677 /* Set the shadow bitmaps to the desired intercept states */
679 set_bit(slot, svm->shadow_msr_intercept.read);
681 clear_bit(slot, svm->shadow_msr_intercept.read);
684 set_bit(slot, svm->shadow_msr_intercept.write);
686 clear_bit(slot, svm->shadow_msr_intercept.write);
689 static bool valid_msr_intercept(u32 index)
691 return direct_access_msr_slot(index) != -ENOENT;
694 static bool msr_write_intercepted(struct kvm_vcpu *vcpu, u32 msr)
701 msrpm = is_guest_mode(vcpu) ? to_svm(vcpu)->nested.msrpm:
704 offset = svm_msrpm_offset(msr);
705 bit_write = 2 * (msr & 0x0f) + 1;
708 BUG_ON(offset == MSR_INVALID);
710 return !!test_bit(bit_write, &tmp);
713 static void set_msr_interception_bitmap(struct kvm_vcpu *vcpu, u32 *msrpm,
714 u32 msr, int read, int write)
716 struct vcpu_svm *svm = to_svm(vcpu);
717 u8 bit_read, bit_write;
722 * If this warning triggers extend the direct_access_msrs list at the
723 * beginning of the file
725 WARN_ON(!valid_msr_intercept(msr));
727 /* Enforce non allowed MSRs to trap */
728 if (read && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_READ))
731 if (write && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_WRITE))
734 offset = svm_msrpm_offset(msr);
735 bit_read = 2 * (msr & 0x0f);
736 bit_write = 2 * (msr & 0x0f) + 1;
739 BUG_ON(offset == MSR_INVALID);
741 read ? clear_bit(bit_read, &tmp) : set_bit(bit_read, &tmp);
742 write ? clear_bit(bit_write, &tmp) : set_bit(bit_write, &tmp);
746 svm_hv_vmcb_dirty_nested_enlightenments(vcpu);
747 svm->nested.force_msr_bitmap_recalc = true;
750 void set_msr_interception(struct kvm_vcpu *vcpu, u32 *msrpm, u32 msr,
753 set_shadow_msr_intercept(vcpu, msr, read, write);
754 set_msr_interception_bitmap(vcpu, msrpm, msr, read, write);
757 u32 *svm_vcpu_alloc_msrpm(void)
759 unsigned int order = get_order(MSRPM_SIZE);
760 struct page *pages = alloc_pages(GFP_KERNEL_ACCOUNT, order);
766 msrpm = page_address(pages);
767 memset(msrpm, 0xff, PAGE_SIZE * (1 << order));
772 void svm_vcpu_init_msrpm(struct kvm_vcpu *vcpu, u32 *msrpm)
776 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
777 if (!direct_access_msrs[i].always)
779 set_msr_interception(vcpu, msrpm, direct_access_msrs[i].index, 1, 1);
784 void svm_vcpu_free_msrpm(u32 *msrpm)
786 __free_pages(virt_to_page(msrpm), get_order(MSRPM_SIZE));
789 static void svm_msr_filter_changed(struct kvm_vcpu *vcpu)
791 struct vcpu_svm *svm = to_svm(vcpu);
795 * Set intercept permissions for all direct access MSRs again. They
796 * will automatically get filtered through the MSR filter, so we are
797 * back in sync after this.
799 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
800 u32 msr = direct_access_msrs[i].index;
801 u32 read = test_bit(i, svm->shadow_msr_intercept.read);
802 u32 write = test_bit(i, svm->shadow_msr_intercept.write);
804 set_msr_interception_bitmap(vcpu, svm->msrpm, msr, read, write);
808 static void add_msr_offset(u32 offset)
812 for (i = 0; i < MSRPM_OFFSETS; ++i) {
814 /* Offset already in list? */
815 if (msrpm_offsets[i] == offset)
818 /* Slot used by another offset? */
819 if (msrpm_offsets[i] != MSR_INVALID)
822 /* Add offset to list */
823 msrpm_offsets[i] = offset;
829 * If this BUG triggers the msrpm_offsets table has an overflow. Just
830 * increase MSRPM_OFFSETS in this case.
835 static void init_msrpm_offsets(void)
839 memset(msrpm_offsets, 0xff, sizeof(msrpm_offsets));
841 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
844 offset = svm_msrpm_offset(direct_access_msrs[i].index);
845 BUG_ON(offset == MSR_INVALID);
847 add_msr_offset(offset);
851 void svm_copy_lbrs(struct vmcb *to_vmcb, struct vmcb *from_vmcb)
853 to_vmcb->save.dbgctl = from_vmcb->save.dbgctl;
854 to_vmcb->save.br_from = from_vmcb->save.br_from;
855 to_vmcb->save.br_to = from_vmcb->save.br_to;
856 to_vmcb->save.last_excp_from = from_vmcb->save.last_excp_from;
857 to_vmcb->save.last_excp_to = from_vmcb->save.last_excp_to;
859 vmcb_mark_dirty(to_vmcb, VMCB_LBR);
862 static void svm_enable_lbrv(struct kvm_vcpu *vcpu)
864 struct vcpu_svm *svm = to_svm(vcpu);
866 svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
867 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
868 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
869 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
870 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
872 /* Move the LBR msrs to the vmcb02 so that the guest can see them. */
873 if (is_guest_mode(vcpu))
874 svm_copy_lbrs(svm->vmcb, svm->vmcb01.ptr);
877 static void svm_disable_lbrv(struct kvm_vcpu *vcpu)
879 struct vcpu_svm *svm = to_svm(vcpu);
881 svm->vmcb->control.virt_ext &= ~LBR_CTL_ENABLE_MASK;
882 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 0, 0);
883 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 0, 0);
884 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 0, 0);
885 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 0, 0);
888 * Move the LBR msrs back to the vmcb01 to avoid copying them
889 * on nested guest entries.
891 if (is_guest_mode(vcpu))
892 svm_copy_lbrs(svm->vmcb01.ptr, svm->vmcb);
895 static int svm_get_lbr_msr(struct vcpu_svm *svm, u32 index)
898 * If the LBR virtualization is disabled, the LBR msrs are always
899 * kept in the vmcb01 to avoid copying them on nested guest entries.
901 * If nested, and the LBR virtualization is enabled/disabled, the msrs
902 * are moved between the vmcb01 and vmcb02 as needed.
905 (svm->vmcb->control.virt_ext & LBR_CTL_ENABLE_MASK) ?
906 svm->vmcb : svm->vmcb01.ptr;
909 case MSR_IA32_DEBUGCTLMSR:
910 return vmcb->save.dbgctl;
911 case MSR_IA32_LASTBRANCHFROMIP:
912 return vmcb->save.br_from;
913 case MSR_IA32_LASTBRANCHTOIP:
914 return vmcb->save.br_to;
915 case MSR_IA32_LASTINTFROMIP:
916 return vmcb->save.last_excp_from;
917 case MSR_IA32_LASTINTTOIP:
918 return vmcb->save.last_excp_to;
920 KVM_BUG(false, svm->vcpu.kvm,
921 "%s: Unknown MSR 0x%x", __func__, index);
926 void svm_update_lbrv(struct kvm_vcpu *vcpu)
928 struct vcpu_svm *svm = to_svm(vcpu);
930 bool enable_lbrv = svm_get_lbr_msr(svm, MSR_IA32_DEBUGCTLMSR) &
933 bool current_enable_lbrv = !!(svm->vmcb->control.virt_ext &
934 LBR_CTL_ENABLE_MASK);
936 if (unlikely(is_guest_mode(vcpu) && svm->lbrv_enabled))
937 if (unlikely(svm->nested.ctl.virt_ext & LBR_CTL_ENABLE_MASK))
940 if (enable_lbrv == current_enable_lbrv)
944 svm_enable_lbrv(vcpu);
946 svm_disable_lbrv(vcpu);
949 void disable_nmi_singlestep(struct vcpu_svm *svm)
951 svm->nmi_singlestep = false;
953 if (!(svm->vcpu.guest_debug & KVM_GUESTDBG_SINGLESTEP)) {
954 /* Clear our flags if they were not set by the guest */
955 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
956 svm->vmcb->save.rflags &= ~X86_EFLAGS_TF;
957 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
958 svm->vmcb->save.rflags &= ~X86_EFLAGS_RF;
962 static void grow_ple_window(struct kvm_vcpu *vcpu)
964 struct vcpu_svm *svm = to_svm(vcpu);
965 struct vmcb_control_area *control = &svm->vmcb->control;
966 int old = control->pause_filter_count;
968 if (kvm_pause_in_guest(vcpu->kvm))
971 control->pause_filter_count = __grow_ple_window(old,
973 pause_filter_count_grow,
974 pause_filter_count_max);
976 if (control->pause_filter_count != old) {
977 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
978 trace_kvm_ple_window_update(vcpu->vcpu_id,
979 control->pause_filter_count, old);
983 static void shrink_ple_window(struct kvm_vcpu *vcpu)
985 struct vcpu_svm *svm = to_svm(vcpu);
986 struct vmcb_control_area *control = &svm->vmcb->control;
987 int old = control->pause_filter_count;
989 if (kvm_pause_in_guest(vcpu->kvm))
992 control->pause_filter_count =
993 __shrink_ple_window(old,
995 pause_filter_count_shrink,
997 if (control->pause_filter_count != old) {
998 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
999 trace_kvm_ple_window_update(vcpu->vcpu_id,
1000 control->pause_filter_count, old);
1004 static void svm_hardware_unsetup(void)
1008 sev_hardware_unsetup();
1010 for_each_possible_cpu(cpu)
1011 svm_cpu_uninit(cpu);
1013 __free_pages(pfn_to_page(iopm_base >> PAGE_SHIFT),
1014 get_order(IOPM_SIZE));
1018 static void init_seg(struct vmcb_seg *seg)
1021 seg->attrib = SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK |
1022 SVM_SELECTOR_WRITE_MASK; /* Read/Write Data Segment */
1023 seg->limit = 0xffff;
1027 static void init_sys_seg(struct vmcb_seg *seg, uint32_t type)
1030 seg->attrib = SVM_SELECTOR_P_MASK | type;
1031 seg->limit = 0xffff;
1035 static u64 svm_get_l2_tsc_offset(struct kvm_vcpu *vcpu)
1037 struct vcpu_svm *svm = to_svm(vcpu);
1039 return svm->nested.ctl.tsc_offset;
1042 static u64 svm_get_l2_tsc_multiplier(struct kvm_vcpu *vcpu)
1044 struct vcpu_svm *svm = to_svm(vcpu);
1046 return svm->tsc_ratio_msr;
1049 static void svm_write_tsc_offset(struct kvm_vcpu *vcpu, u64 offset)
1051 struct vcpu_svm *svm = to_svm(vcpu);
1053 svm->vmcb01.ptr->control.tsc_offset = vcpu->arch.l1_tsc_offset;
1054 svm->vmcb->control.tsc_offset = offset;
1055 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1058 static void svm_write_tsc_multiplier(struct kvm_vcpu *vcpu, u64 multiplier)
1060 __svm_write_tsc_multiplier(multiplier);
1064 /* Evaluate instruction intercepts that depend on guest CPUID features. */
1065 static void svm_recalc_instruction_intercepts(struct kvm_vcpu *vcpu,
1066 struct vcpu_svm *svm)
1069 * Intercept INVPCID if shadow paging is enabled to sync/free shadow
1070 * roots, or if INVPCID is disabled in the guest to inject #UD.
1072 if (kvm_cpu_cap_has(X86_FEATURE_INVPCID)) {
1074 !guest_cpuid_has(&svm->vcpu, X86_FEATURE_INVPCID))
1075 svm_set_intercept(svm, INTERCEPT_INVPCID);
1077 svm_clr_intercept(svm, INTERCEPT_INVPCID);
1080 if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP)) {
1081 if (guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP))
1082 svm_clr_intercept(svm, INTERCEPT_RDTSCP);
1084 svm_set_intercept(svm, INTERCEPT_RDTSCP);
1088 static inline void init_vmcb_after_set_cpuid(struct kvm_vcpu *vcpu)
1090 struct vcpu_svm *svm = to_svm(vcpu);
1092 if (guest_cpuid_is_intel(vcpu)) {
1094 * We must intercept SYSENTER_EIP and SYSENTER_ESP
1095 * accesses because the processor only stores 32 bits.
1096 * For the same reason we cannot use virtual VMLOAD/VMSAVE.
1098 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1099 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1100 svm->vmcb->control.virt_ext &= ~VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
1102 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 0, 0);
1103 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 0, 0);
1105 svm->v_vmload_vmsave_enabled = false;
1108 * If hardware supports Virtual VMLOAD VMSAVE then enable it
1109 * in VMCB and clear intercepts to avoid #VMEXIT.
1112 svm_clr_intercept(svm, INTERCEPT_VMLOAD);
1113 svm_clr_intercept(svm, INTERCEPT_VMSAVE);
1114 svm->vmcb->control.virt_ext |= VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
1116 /* No need to intercept these MSRs */
1117 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 1, 1);
1118 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 1, 1);
1122 static void init_vmcb(struct kvm_vcpu *vcpu)
1124 struct vcpu_svm *svm = to_svm(vcpu);
1125 struct vmcb *vmcb = svm->vmcb01.ptr;
1126 struct vmcb_control_area *control = &vmcb->control;
1127 struct vmcb_save_area *save = &vmcb->save;
1129 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1130 svm_set_intercept(svm, INTERCEPT_CR3_READ);
1131 svm_set_intercept(svm, INTERCEPT_CR4_READ);
1132 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1133 svm_set_intercept(svm, INTERCEPT_CR3_WRITE);
1134 svm_set_intercept(svm, INTERCEPT_CR4_WRITE);
1135 if (!kvm_vcpu_apicv_active(vcpu))
1136 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
1138 set_dr_intercepts(svm);
1140 set_exception_intercept(svm, PF_VECTOR);
1141 set_exception_intercept(svm, UD_VECTOR);
1142 set_exception_intercept(svm, MC_VECTOR);
1143 set_exception_intercept(svm, AC_VECTOR);
1144 set_exception_intercept(svm, DB_VECTOR);
1146 * Guest access to VMware backdoor ports could legitimately
1147 * trigger #GP because of TSS I/O permission bitmap.
1148 * We intercept those #GP and allow access to them anyway
1149 * as VMware does. Don't intercept #GP for SEV guests as KVM can't
1150 * decrypt guest memory to decode the faulting instruction.
1152 if (enable_vmware_backdoor && !sev_guest(vcpu->kvm))
1153 set_exception_intercept(svm, GP_VECTOR);
1155 svm_set_intercept(svm, INTERCEPT_INTR);
1156 svm_set_intercept(svm, INTERCEPT_NMI);
1159 svm_set_intercept(svm, INTERCEPT_SMI);
1161 svm_set_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1162 svm_set_intercept(svm, INTERCEPT_RDPMC);
1163 svm_set_intercept(svm, INTERCEPT_CPUID);
1164 svm_set_intercept(svm, INTERCEPT_INVD);
1165 svm_set_intercept(svm, INTERCEPT_INVLPG);
1166 svm_set_intercept(svm, INTERCEPT_INVLPGA);
1167 svm_set_intercept(svm, INTERCEPT_IOIO_PROT);
1168 svm_set_intercept(svm, INTERCEPT_MSR_PROT);
1169 svm_set_intercept(svm, INTERCEPT_TASK_SWITCH);
1170 svm_set_intercept(svm, INTERCEPT_SHUTDOWN);
1171 svm_set_intercept(svm, INTERCEPT_VMRUN);
1172 svm_set_intercept(svm, INTERCEPT_VMMCALL);
1173 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1174 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1175 svm_set_intercept(svm, INTERCEPT_STGI);
1176 svm_set_intercept(svm, INTERCEPT_CLGI);
1177 svm_set_intercept(svm, INTERCEPT_SKINIT);
1178 svm_set_intercept(svm, INTERCEPT_WBINVD);
1179 svm_set_intercept(svm, INTERCEPT_XSETBV);
1180 svm_set_intercept(svm, INTERCEPT_RDPRU);
1181 svm_set_intercept(svm, INTERCEPT_RSM);
1183 if (!kvm_mwait_in_guest(vcpu->kvm)) {
1184 svm_set_intercept(svm, INTERCEPT_MONITOR);
1185 svm_set_intercept(svm, INTERCEPT_MWAIT);
1188 if (!kvm_hlt_in_guest(vcpu->kvm))
1189 svm_set_intercept(svm, INTERCEPT_HLT);
1191 control->iopm_base_pa = __sme_set(iopm_base);
1192 control->msrpm_base_pa = __sme_set(__pa(svm->msrpm));
1193 control->int_ctl = V_INTR_MASKING_MASK;
1195 init_seg(&save->es);
1196 init_seg(&save->ss);
1197 init_seg(&save->ds);
1198 init_seg(&save->fs);
1199 init_seg(&save->gs);
1201 save->cs.selector = 0xf000;
1202 save->cs.base = 0xffff0000;
1203 /* Executable/Readable Code Segment */
1204 save->cs.attrib = SVM_SELECTOR_READ_MASK | SVM_SELECTOR_P_MASK |
1205 SVM_SELECTOR_S_MASK | SVM_SELECTOR_CODE_MASK;
1206 save->cs.limit = 0xffff;
1208 save->gdtr.base = 0;
1209 save->gdtr.limit = 0xffff;
1210 save->idtr.base = 0;
1211 save->idtr.limit = 0xffff;
1213 init_sys_seg(&save->ldtr, SEG_TYPE_LDT);
1214 init_sys_seg(&save->tr, SEG_TYPE_BUSY_TSS16);
1217 /* Setup VMCB for Nested Paging */
1218 control->nested_ctl |= SVM_NESTED_CTL_NP_ENABLE;
1219 svm_clr_intercept(svm, INTERCEPT_INVLPG);
1220 clr_exception_intercept(svm, PF_VECTOR);
1221 svm_clr_intercept(svm, INTERCEPT_CR3_READ);
1222 svm_clr_intercept(svm, INTERCEPT_CR3_WRITE);
1223 save->g_pat = vcpu->arch.pat;
1226 svm->current_vmcb->asid_generation = 0;
1229 svm->nested.vmcb12_gpa = INVALID_GPA;
1230 svm->nested.last_vmcb12_gpa = INVALID_GPA;
1232 if (!kvm_pause_in_guest(vcpu->kvm)) {
1233 control->pause_filter_count = pause_filter_count;
1234 if (pause_filter_thresh)
1235 control->pause_filter_thresh = pause_filter_thresh;
1236 svm_set_intercept(svm, INTERCEPT_PAUSE);
1238 svm_clr_intercept(svm, INTERCEPT_PAUSE);
1241 svm_recalc_instruction_intercepts(vcpu, svm);
1244 * If the host supports V_SPEC_CTRL then disable the interception
1245 * of MSR_IA32_SPEC_CTRL.
1247 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
1248 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
1250 if (kvm_vcpu_apicv_active(vcpu))
1251 avic_init_vmcb(svm, vmcb);
1254 svm_clr_intercept(svm, INTERCEPT_STGI);
1255 svm_clr_intercept(svm, INTERCEPT_CLGI);
1256 svm->vmcb->control.int_ctl |= V_GIF_ENABLE_MASK;
1259 if (sev_guest(vcpu->kvm)) {
1260 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
1261 clr_exception_intercept(svm, UD_VECTOR);
1263 if (sev_es_guest(vcpu->kvm)) {
1264 /* Perform SEV-ES specific VMCB updates */
1265 sev_es_init_vmcb(svm);
1269 svm_hv_init_vmcb(vmcb);
1270 init_vmcb_after_set_cpuid(vcpu);
1272 vmcb_mark_all_dirty(vmcb);
1277 static void __svm_vcpu_reset(struct kvm_vcpu *vcpu)
1279 struct vcpu_svm *svm = to_svm(vcpu);
1281 svm_vcpu_init_msrpm(vcpu, svm->msrpm);
1283 svm_init_osvw(vcpu);
1284 vcpu->arch.microcode_version = 0x01000065;
1285 svm->tsc_ratio_msr = kvm_caps.default_tsc_scaling_ratio;
1287 if (sev_es_guest(vcpu->kvm))
1288 sev_es_vcpu_reset(svm);
1291 static void svm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
1293 struct vcpu_svm *svm = to_svm(vcpu);
1296 svm->virt_spec_ctrl = 0;
1301 __svm_vcpu_reset(vcpu);
1304 void svm_switch_vmcb(struct vcpu_svm *svm, struct kvm_vmcb_info *target_vmcb)
1306 svm->current_vmcb = target_vmcb;
1307 svm->vmcb = target_vmcb->ptr;
1310 static int svm_vcpu_create(struct kvm_vcpu *vcpu)
1312 struct vcpu_svm *svm;
1313 struct page *vmcb01_page;
1314 struct page *vmsa_page = NULL;
1317 BUILD_BUG_ON(offsetof(struct vcpu_svm, vcpu) != 0);
1321 vmcb01_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1325 if (sev_es_guest(vcpu->kvm)) {
1327 * SEV-ES guests require a separate VMSA page used to contain
1328 * the encrypted register state of the guest.
1330 vmsa_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1332 goto error_free_vmcb_page;
1335 * SEV-ES guests maintain an encrypted version of their FPU
1336 * state which is restored and saved on VMRUN and VMEXIT.
1337 * Mark vcpu->arch.guest_fpu->fpstate as scratch so it won't
1338 * do xsave/xrstor on it.
1340 fpstate_set_confidential(&vcpu->arch.guest_fpu);
1343 err = avic_init_vcpu(svm);
1345 goto error_free_vmsa_page;
1347 svm->msrpm = svm_vcpu_alloc_msrpm();
1350 goto error_free_vmsa_page;
1353 svm->vmcb01.ptr = page_address(vmcb01_page);
1354 svm->vmcb01.pa = __sme_set(page_to_pfn(vmcb01_page) << PAGE_SHIFT);
1355 svm_switch_vmcb(svm, &svm->vmcb01);
1358 svm->sev_es.vmsa = page_address(vmsa_page);
1360 svm->guest_state_loaded = false;
1364 error_free_vmsa_page:
1366 __free_page(vmsa_page);
1367 error_free_vmcb_page:
1368 __free_page(vmcb01_page);
1373 static void svm_clear_current_vmcb(struct vmcb *vmcb)
1377 for_each_online_cpu(i)
1378 cmpxchg(&per_cpu(svm_data, i)->current_vmcb, vmcb, NULL);
1381 static void svm_vcpu_free(struct kvm_vcpu *vcpu)
1383 struct vcpu_svm *svm = to_svm(vcpu);
1386 * The vmcb page can be recycled, causing a false negative in
1387 * svm_vcpu_load(). So, ensure that no logical CPU has this
1388 * vmcb page recorded as its current vmcb.
1390 svm_clear_current_vmcb(svm->vmcb);
1392 svm_free_nested(svm);
1394 sev_free_vcpu(vcpu);
1396 __free_page(pfn_to_page(__sme_clr(svm->vmcb01.pa) >> PAGE_SHIFT));
1397 __free_pages(virt_to_page(svm->msrpm), get_order(MSRPM_SIZE));
1400 static void svm_prepare_switch_to_guest(struct kvm_vcpu *vcpu)
1402 struct vcpu_svm *svm = to_svm(vcpu);
1403 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
1405 if (sev_es_guest(vcpu->kvm))
1406 sev_es_unmap_ghcb(svm);
1408 if (svm->guest_state_loaded)
1412 * Save additional host state that will be restored on VMEXIT (sev-es)
1413 * or subsequent vmload of host save area.
1415 vmsave(__sme_page_pa(sd->save_area));
1416 if (sev_es_guest(vcpu->kvm)) {
1417 struct sev_es_save_area *hostsa;
1418 hostsa = (struct sev_es_save_area *)(page_address(sd->save_area) + 0x400);
1420 sev_es_prepare_switch_to_guest(hostsa);
1424 __svm_write_tsc_multiplier(vcpu->arch.tsc_scaling_ratio);
1426 if (likely(tsc_aux_uret_slot >= 0))
1427 kvm_set_user_return_msr(tsc_aux_uret_slot, svm->tsc_aux, -1ull);
1429 svm->guest_state_loaded = true;
1432 static void svm_prepare_host_switch(struct kvm_vcpu *vcpu)
1434 to_svm(vcpu)->guest_state_loaded = false;
1437 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1439 struct vcpu_svm *svm = to_svm(vcpu);
1440 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1442 if (sd->current_vmcb != svm->vmcb) {
1443 sd->current_vmcb = svm->vmcb;
1444 indirect_branch_prediction_barrier();
1446 if (kvm_vcpu_apicv_active(vcpu))
1447 avic_vcpu_load(vcpu, cpu);
1450 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
1452 if (kvm_vcpu_apicv_active(vcpu))
1453 avic_vcpu_put(vcpu);
1455 svm_prepare_host_switch(vcpu);
1457 ++vcpu->stat.host_state_reload;
1460 static unsigned long svm_get_rflags(struct kvm_vcpu *vcpu)
1462 struct vcpu_svm *svm = to_svm(vcpu);
1463 unsigned long rflags = svm->vmcb->save.rflags;
1465 if (svm->nmi_singlestep) {
1466 /* Hide our flags if they were not set by the guest */
1467 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
1468 rflags &= ~X86_EFLAGS_TF;
1469 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
1470 rflags &= ~X86_EFLAGS_RF;
1475 static void svm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
1477 if (to_svm(vcpu)->nmi_singlestep)
1478 rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
1481 * Any change of EFLAGS.VM is accompanied by a reload of SS
1482 * (caused by either a task switch or an inter-privilege IRET),
1483 * so we do not need to update the CPL here.
1485 to_svm(vcpu)->vmcb->save.rflags = rflags;
1488 static bool svm_get_if_flag(struct kvm_vcpu *vcpu)
1490 struct vmcb *vmcb = to_svm(vcpu)->vmcb;
1492 return sev_es_guest(vcpu->kvm)
1493 ? vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK
1494 : kvm_get_rflags(vcpu) & X86_EFLAGS_IF;
1497 static void svm_cache_reg(struct kvm_vcpu *vcpu, enum kvm_reg reg)
1499 kvm_register_mark_available(vcpu, reg);
1502 case VCPU_EXREG_PDPTR:
1504 * When !npt_enabled, mmu->pdptrs[] is already available since
1505 * it is always updated per SDM when moving to CRs.
1508 load_pdptrs(vcpu, kvm_read_cr3(vcpu));
1511 KVM_BUG_ON(1, vcpu->kvm);
1515 static void svm_set_vintr(struct vcpu_svm *svm)
1517 struct vmcb_control_area *control;
1520 * The following fields are ignored when AVIC is enabled
1522 WARN_ON(kvm_vcpu_apicv_activated(&svm->vcpu));
1524 svm_set_intercept(svm, INTERCEPT_VINTR);
1527 * This is just a dummy VINTR to actually cause a vmexit to happen.
1528 * Actual injection of virtual interrupts happens through EVENTINJ.
1530 control = &svm->vmcb->control;
1531 control->int_vector = 0x0;
1532 control->int_ctl &= ~V_INTR_PRIO_MASK;
1533 control->int_ctl |= V_IRQ_MASK |
1534 ((/*control->int_vector >> 4*/ 0xf) << V_INTR_PRIO_SHIFT);
1535 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1538 static void svm_clear_vintr(struct vcpu_svm *svm)
1540 svm_clr_intercept(svm, INTERCEPT_VINTR);
1542 /* Drop int_ctl fields related to VINTR injection. */
1543 svm->vmcb->control.int_ctl &= ~V_IRQ_INJECTION_BITS_MASK;
1544 if (is_guest_mode(&svm->vcpu)) {
1545 svm->vmcb01.ptr->control.int_ctl &= ~V_IRQ_INJECTION_BITS_MASK;
1547 WARN_ON((svm->vmcb->control.int_ctl & V_TPR_MASK) !=
1548 (svm->nested.ctl.int_ctl & V_TPR_MASK));
1550 svm->vmcb->control.int_ctl |= svm->nested.ctl.int_ctl &
1551 V_IRQ_INJECTION_BITS_MASK;
1553 svm->vmcb->control.int_vector = svm->nested.ctl.int_vector;
1556 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1559 static struct vmcb_seg *svm_seg(struct kvm_vcpu *vcpu, int seg)
1561 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1562 struct vmcb_save_area *save01 = &to_svm(vcpu)->vmcb01.ptr->save;
1565 case VCPU_SREG_CS: return &save->cs;
1566 case VCPU_SREG_DS: return &save->ds;
1567 case VCPU_SREG_ES: return &save->es;
1568 case VCPU_SREG_FS: return &save01->fs;
1569 case VCPU_SREG_GS: return &save01->gs;
1570 case VCPU_SREG_SS: return &save->ss;
1571 case VCPU_SREG_TR: return &save01->tr;
1572 case VCPU_SREG_LDTR: return &save01->ldtr;
1578 static u64 svm_get_segment_base(struct kvm_vcpu *vcpu, int seg)
1580 struct vmcb_seg *s = svm_seg(vcpu, seg);
1585 static void svm_get_segment(struct kvm_vcpu *vcpu,
1586 struct kvm_segment *var, int seg)
1588 struct vmcb_seg *s = svm_seg(vcpu, seg);
1590 var->base = s->base;
1591 var->limit = s->limit;
1592 var->selector = s->selector;
1593 var->type = s->attrib & SVM_SELECTOR_TYPE_MASK;
1594 var->s = (s->attrib >> SVM_SELECTOR_S_SHIFT) & 1;
1595 var->dpl = (s->attrib >> SVM_SELECTOR_DPL_SHIFT) & 3;
1596 var->present = (s->attrib >> SVM_SELECTOR_P_SHIFT) & 1;
1597 var->avl = (s->attrib >> SVM_SELECTOR_AVL_SHIFT) & 1;
1598 var->l = (s->attrib >> SVM_SELECTOR_L_SHIFT) & 1;
1599 var->db = (s->attrib >> SVM_SELECTOR_DB_SHIFT) & 1;
1602 * AMD CPUs circa 2014 track the G bit for all segments except CS.
1603 * However, the SVM spec states that the G bit is not observed by the
1604 * CPU, and some VMware virtual CPUs drop the G bit for all segments.
1605 * So let's synthesize a legal G bit for all segments, this helps
1606 * running KVM nested. It also helps cross-vendor migration, because
1607 * Intel's vmentry has a check on the 'G' bit.
1609 var->g = s->limit > 0xfffff;
1612 * AMD's VMCB does not have an explicit unusable field, so emulate it
1613 * for cross vendor migration purposes by "not present"
1615 var->unusable = !var->present;
1620 * Work around a bug where the busy flag in the tr selector
1630 * The accessed bit must always be set in the segment
1631 * descriptor cache, although it can be cleared in the
1632 * descriptor, the cached bit always remains at 1. Since
1633 * Intel has a check on this, set it here to support
1634 * cross-vendor migration.
1641 * On AMD CPUs sometimes the DB bit in the segment
1642 * descriptor is left as 1, although the whole segment has
1643 * been made unusable. Clear it here to pass an Intel VMX
1644 * entry check when cross vendor migrating.
1648 /* This is symmetric with svm_set_segment() */
1649 var->dpl = to_svm(vcpu)->vmcb->save.cpl;
1654 static int svm_get_cpl(struct kvm_vcpu *vcpu)
1656 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1661 static void svm_get_cs_db_l_bits(struct kvm_vcpu *vcpu, int *db, int *l)
1663 struct kvm_segment cs;
1665 svm_get_segment(vcpu, &cs, VCPU_SREG_CS);
1670 static void svm_get_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1672 struct vcpu_svm *svm = to_svm(vcpu);
1674 dt->size = svm->vmcb->save.idtr.limit;
1675 dt->address = svm->vmcb->save.idtr.base;
1678 static void svm_set_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1680 struct vcpu_svm *svm = to_svm(vcpu);
1682 svm->vmcb->save.idtr.limit = dt->size;
1683 svm->vmcb->save.idtr.base = dt->address ;
1684 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1687 static void svm_get_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1689 struct vcpu_svm *svm = to_svm(vcpu);
1691 dt->size = svm->vmcb->save.gdtr.limit;
1692 dt->address = svm->vmcb->save.gdtr.base;
1695 static void svm_set_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1697 struct vcpu_svm *svm = to_svm(vcpu);
1699 svm->vmcb->save.gdtr.limit = dt->size;
1700 svm->vmcb->save.gdtr.base = dt->address ;
1701 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1704 static void sev_post_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3)
1706 struct vcpu_svm *svm = to_svm(vcpu);
1709 * For guests that don't set guest_state_protected, the cr3 update is
1710 * handled via kvm_mmu_load() while entering the guest. For guests
1711 * that do (SEV-ES/SEV-SNP), the cr3 update needs to be written to
1712 * VMCB save area now, since the save area will become the initial
1713 * contents of the VMSA, and future VMCB save area updates won't be
1716 if (sev_es_guest(vcpu->kvm)) {
1717 svm->vmcb->save.cr3 = cr3;
1718 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1722 void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1724 struct vcpu_svm *svm = to_svm(vcpu);
1726 bool old_paging = is_paging(vcpu);
1728 #ifdef CONFIG_X86_64
1729 if (vcpu->arch.efer & EFER_LME && !vcpu->arch.guest_state_protected) {
1730 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
1731 vcpu->arch.efer |= EFER_LMA;
1732 svm->vmcb->save.efer |= EFER_LMA | EFER_LME;
1735 if (is_paging(vcpu) && !(cr0 & X86_CR0_PG)) {
1736 vcpu->arch.efer &= ~EFER_LMA;
1737 svm->vmcb->save.efer &= ~(EFER_LMA | EFER_LME);
1741 vcpu->arch.cr0 = cr0;
1744 hcr0 |= X86_CR0_PG | X86_CR0_WP;
1745 if (old_paging != is_paging(vcpu))
1746 svm_set_cr4(vcpu, kvm_read_cr4(vcpu));
1750 * re-enable caching here because the QEMU bios
1751 * does not do it - this results in some delay at
1754 if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
1755 hcr0 &= ~(X86_CR0_CD | X86_CR0_NW);
1757 svm->vmcb->save.cr0 = hcr0;
1758 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1761 * SEV-ES guests must always keep the CR intercepts cleared. CR
1762 * tracking is done using the CR write traps.
1764 if (sev_es_guest(vcpu->kvm))
1768 /* Selective CR0 write remains on. */
1769 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1770 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1772 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1773 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1777 static bool svm_is_valid_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1782 void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1784 unsigned long host_cr4_mce = cr4_read_shadow() & X86_CR4_MCE;
1785 unsigned long old_cr4 = vcpu->arch.cr4;
1787 if (npt_enabled && ((old_cr4 ^ cr4) & X86_CR4_PGE))
1788 svm_flush_tlb_current(vcpu);
1790 vcpu->arch.cr4 = cr4;
1794 if (!is_paging(vcpu))
1795 cr4 &= ~(X86_CR4_SMEP | X86_CR4_SMAP | X86_CR4_PKE);
1797 cr4 |= host_cr4_mce;
1798 to_svm(vcpu)->vmcb->save.cr4 = cr4;
1799 vmcb_mark_dirty(to_svm(vcpu)->vmcb, VMCB_CR);
1801 if ((cr4 ^ old_cr4) & (X86_CR4_OSXSAVE | X86_CR4_PKE))
1802 kvm_update_cpuid_runtime(vcpu);
1805 static void svm_set_segment(struct kvm_vcpu *vcpu,
1806 struct kvm_segment *var, int seg)
1808 struct vcpu_svm *svm = to_svm(vcpu);
1809 struct vmcb_seg *s = svm_seg(vcpu, seg);
1811 s->base = var->base;
1812 s->limit = var->limit;
1813 s->selector = var->selector;
1814 s->attrib = (var->type & SVM_SELECTOR_TYPE_MASK);
1815 s->attrib |= (var->s & 1) << SVM_SELECTOR_S_SHIFT;
1816 s->attrib |= (var->dpl & 3) << SVM_SELECTOR_DPL_SHIFT;
1817 s->attrib |= ((var->present & 1) && !var->unusable) << SVM_SELECTOR_P_SHIFT;
1818 s->attrib |= (var->avl & 1) << SVM_SELECTOR_AVL_SHIFT;
1819 s->attrib |= (var->l & 1) << SVM_SELECTOR_L_SHIFT;
1820 s->attrib |= (var->db & 1) << SVM_SELECTOR_DB_SHIFT;
1821 s->attrib |= (var->g & 1) << SVM_SELECTOR_G_SHIFT;
1824 * This is always accurate, except if SYSRET returned to a segment
1825 * with SS.DPL != 3. Intel does not have this quirk, and always
1826 * forces SS.DPL to 3 on sysret, so we ignore that case; fixing it
1827 * would entail passing the CPL to userspace and back.
1829 if (seg == VCPU_SREG_SS)
1830 /* This is symmetric with svm_get_segment() */
1831 svm->vmcb->save.cpl = (var->dpl & 3);
1833 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
1836 static void svm_update_exception_bitmap(struct kvm_vcpu *vcpu)
1838 struct vcpu_svm *svm = to_svm(vcpu);
1840 clr_exception_intercept(svm, BP_VECTOR);
1842 if (vcpu->guest_debug & KVM_GUESTDBG_ENABLE) {
1843 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP)
1844 set_exception_intercept(svm, BP_VECTOR);
1848 static void new_asid(struct vcpu_svm *svm, struct svm_cpu_data *sd)
1850 if (sd->next_asid > sd->max_asid) {
1851 ++sd->asid_generation;
1852 sd->next_asid = sd->min_asid;
1853 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
1854 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1857 svm->current_vmcb->asid_generation = sd->asid_generation;
1858 svm->asid = sd->next_asid++;
1861 static void svm_set_dr6(struct vcpu_svm *svm, unsigned long value)
1863 struct vmcb *vmcb = svm->vmcb;
1865 if (svm->vcpu.arch.guest_state_protected)
1868 if (unlikely(value != vmcb->save.dr6)) {
1869 vmcb->save.dr6 = value;
1870 vmcb_mark_dirty(vmcb, VMCB_DR);
1874 static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
1876 struct vcpu_svm *svm = to_svm(vcpu);
1878 if (vcpu->arch.guest_state_protected)
1881 get_debugreg(vcpu->arch.db[0], 0);
1882 get_debugreg(vcpu->arch.db[1], 1);
1883 get_debugreg(vcpu->arch.db[2], 2);
1884 get_debugreg(vcpu->arch.db[3], 3);
1886 * We cannot reset svm->vmcb->save.dr6 to DR6_ACTIVE_LOW here,
1887 * because db_interception might need it. We can do it before vmentry.
1889 vcpu->arch.dr6 = svm->vmcb->save.dr6;
1890 vcpu->arch.dr7 = svm->vmcb->save.dr7;
1891 vcpu->arch.switch_db_regs &= ~KVM_DEBUGREG_WONT_EXIT;
1892 set_dr_intercepts(svm);
1895 static void svm_set_dr7(struct kvm_vcpu *vcpu, unsigned long value)
1897 struct vcpu_svm *svm = to_svm(vcpu);
1899 if (vcpu->arch.guest_state_protected)
1902 svm->vmcb->save.dr7 = value;
1903 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
1906 static int pf_interception(struct kvm_vcpu *vcpu)
1908 struct vcpu_svm *svm = to_svm(vcpu);
1910 u64 fault_address = svm->vmcb->control.exit_info_2;
1911 u64 error_code = svm->vmcb->control.exit_info_1;
1913 return kvm_handle_page_fault(vcpu, error_code, fault_address,
1914 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1915 svm->vmcb->control.insn_bytes : NULL,
1916 svm->vmcb->control.insn_len);
1919 static int npf_interception(struct kvm_vcpu *vcpu)
1921 struct vcpu_svm *svm = to_svm(vcpu);
1923 u64 fault_address = svm->vmcb->control.exit_info_2;
1924 u64 error_code = svm->vmcb->control.exit_info_1;
1926 trace_kvm_page_fault(fault_address, error_code);
1927 return kvm_mmu_page_fault(vcpu, fault_address, error_code,
1928 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1929 svm->vmcb->control.insn_bytes : NULL,
1930 svm->vmcb->control.insn_len);
1933 static int db_interception(struct kvm_vcpu *vcpu)
1935 struct kvm_run *kvm_run = vcpu->run;
1936 struct vcpu_svm *svm = to_svm(vcpu);
1938 if (!(vcpu->guest_debug &
1939 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) &&
1940 !svm->nmi_singlestep) {
1941 u32 payload = svm->vmcb->save.dr6 ^ DR6_ACTIVE_LOW;
1942 kvm_queue_exception_p(vcpu, DB_VECTOR, payload);
1946 if (svm->nmi_singlestep) {
1947 disable_nmi_singlestep(svm);
1948 /* Make sure we check for pending NMIs upon entry */
1949 kvm_make_request(KVM_REQ_EVENT, vcpu);
1952 if (vcpu->guest_debug &
1953 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) {
1954 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1955 kvm_run->debug.arch.dr6 = svm->vmcb->save.dr6;
1956 kvm_run->debug.arch.dr7 = svm->vmcb->save.dr7;
1957 kvm_run->debug.arch.pc =
1958 svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1959 kvm_run->debug.arch.exception = DB_VECTOR;
1966 static int bp_interception(struct kvm_vcpu *vcpu)
1968 struct vcpu_svm *svm = to_svm(vcpu);
1969 struct kvm_run *kvm_run = vcpu->run;
1971 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1972 kvm_run->debug.arch.pc = svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1973 kvm_run->debug.arch.exception = BP_VECTOR;
1977 static int ud_interception(struct kvm_vcpu *vcpu)
1979 return handle_ud(vcpu);
1982 static int ac_interception(struct kvm_vcpu *vcpu)
1984 kvm_queue_exception_e(vcpu, AC_VECTOR, 0);
1988 static bool is_erratum_383(void)
1993 if (!erratum_383_found)
1996 value = native_read_msr_safe(MSR_IA32_MC0_STATUS, &err);
2000 /* Bit 62 may or may not be set for this mce */
2001 value &= ~(1ULL << 62);
2003 if (value != 0xb600000000010015ULL)
2006 /* Clear MCi_STATUS registers */
2007 for (i = 0; i < 6; ++i)
2008 native_write_msr_safe(MSR_IA32_MCx_STATUS(i), 0, 0);
2010 value = native_read_msr_safe(MSR_IA32_MCG_STATUS, &err);
2014 value &= ~(1ULL << 2);
2015 low = lower_32_bits(value);
2016 high = upper_32_bits(value);
2018 native_write_msr_safe(MSR_IA32_MCG_STATUS, low, high);
2021 /* Flush tlb to evict multi-match entries */
2027 static void svm_handle_mce(struct kvm_vcpu *vcpu)
2029 if (is_erratum_383()) {
2031 * Erratum 383 triggered. Guest state is corrupt so kill the
2034 pr_err("KVM: Guest triggered AMD Erratum 383\n");
2036 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2042 * On an #MC intercept the MCE handler is not called automatically in
2043 * the host. So do it by hand here.
2045 kvm_machine_check();
2048 static int mc_interception(struct kvm_vcpu *vcpu)
2053 static int shutdown_interception(struct kvm_vcpu *vcpu)
2055 struct kvm_run *kvm_run = vcpu->run;
2056 struct vcpu_svm *svm = to_svm(vcpu);
2059 * The VM save area has already been encrypted so it
2060 * cannot be reinitialized - just terminate.
2062 if (sev_es_guest(vcpu->kvm))
2066 * VMCB is undefined after a SHUTDOWN intercept. INIT the vCPU to put
2067 * the VMCB in a known good state. Unfortuately, KVM doesn't have
2068 * KVM_MP_STATE_SHUTDOWN and can't add it without potentially breaking
2069 * userspace. At a platform view, INIT is acceptable behavior as
2070 * there exist bare metal platforms that automatically INIT the CPU
2071 * in response to shutdown.
2073 clear_page(svm->vmcb);
2074 kvm_vcpu_reset(vcpu, true);
2076 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
2080 static int io_interception(struct kvm_vcpu *vcpu)
2082 struct vcpu_svm *svm = to_svm(vcpu);
2083 u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
2084 int size, in, string;
2087 ++vcpu->stat.io_exits;
2088 string = (io_info & SVM_IOIO_STR_MASK) != 0;
2089 in = (io_info & SVM_IOIO_TYPE_MASK) != 0;
2090 port = io_info >> 16;
2091 size = (io_info & SVM_IOIO_SIZE_MASK) >> SVM_IOIO_SIZE_SHIFT;
2094 if (sev_es_guest(vcpu->kvm))
2095 return sev_es_string_io(svm, size, port, in);
2097 return kvm_emulate_instruction(vcpu, 0);
2100 svm->next_rip = svm->vmcb->control.exit_info_2;
2102 return kvm_fast_pio(vcpu, size, port, in);
2105 static int nmi_interception(struct kvm_vcpu *vcpu)
2110 static int smi_interception(struct kvm_vcpu *vcpu)
2115 static int intr_interception(struct kvm_vcpu *vcpu)
2117 ++vcpu->stat.irq_exits;
2121 static int vmload_vmsave_interception(struct kvm_vcpu *vcpu, bool vmload)
2123 struct vcpu_svm *svm = to_svm(vcpu);
2124 struct vmcb *vmcb12;
2125 struct kvm_host_map map;
2128 if (nested_svm_check_permissions(vcpu))
2131 ret = kvm_vcpu_map(vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2134 kvm_inject_gp(vcpu, 0);
2140 ret = kvm_skip_emulated_instruction(vcpu);
2143 svm_copy_vmloadsave_state(svm->vmcb, vmcb12);
2144 svm->sysenter_eip_hi = 0;
2145 svm->sysenter_esp_hi = 0;
2147 svm_copy_vmloadsave_state(vmcb12, svm->vmcb);
2150 kvm_vcpu_unmap(vcpu, &map, true);
2155 static int vmload_interception(struct kvm_vcpu *vcpu)
2157 return vmload_vmsave_interception(vcpu, true);
2160 static int vmsave_interception(struct kvm_vcpu *vcpu)
2162 return vmload_vmsave_interception(vcpu, false);
2165 static int vmrun_interception(struct kvm_vcpu *vcpu)
2167 if (nested_svm_check_permissions(vcpu))
2170 return nested_svm_vmrun(vcpu);
2180 /* Return NONE_SVM_INSTR if not SVM instrs, otherwise return decode result */
2181 static int svm_instr_opcode(struct kvm_vcpu *vcpu)
2183 struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt;
2185 if (ctxt->b != 0x1 || ctxt->opcode_len != 2)
2186 return NONE_SVM_INSTR;
2188 switch (ctxt->modrm) {
2189 case 0xd8: /* VMRUN */
2190 return SVM_INSTR_VMRUN;
2191 case 0xda: /* VMLOAD */
2192 return SVM_INSTR_VMLOAD;
2193 case 0xdb: /* VMSAVE */
2194 return SVM_INSTR_VMSAVE;
2199 return NONE_SVM_INSTR;
2202 static int emulate_svm_instr(struct kvm_vcpu *vcpu, int opcode)
2204 const int guest_mode_exit_codes[] = {
2205 [SVM_INSTR_VMRUN] = SVM_EXIT_VMRUN,
2206 [SVM_INSTR_VMLOAD] = SVM_EXIT_VMLOAD,
2207 [SVM_INSTR_VMSAVE] = SVM_EXIT_VMSAVE,
2209 int (*const svm_instr_handlers[])(struct kvm_vcpu *vcpu) = {
2210 [SVM_INSTR_VMRUN] = vmrun_interception,
2211 [SVM_INSTR_VMLOAD] = vmload_interception,
2212 [SVM_INSTR_VMSAVE] = vmsave_interception,
2214 struct vcpu_svm *svm = to_svm(vcpu);
2217 if (is_guest_mode(vcpu)) {
2218 /* Returns '1' or -errno on failure, '0' on success. */
2219 ret = nested_svm_simple_vmexit(svm, guest_mode_exit_codes[opcode]);
2224 return svm_instr_handlers[opcode](vcpu);
2228 * #GP handling code. Note that #GP can be triggered under the following two
2230 * 1) SVM VM-related instructions (VMRUN/VMSAVE/VMLOAD) that trigger #GP on
2231 * some AMD CPUs when EAX of these instructions are in the reserved memory
2232 * regions (e.g. SMM memory on host).
2233 * 2) VMware backdoor
2235 static int gp_interception(struct kvm_vcpu *vcpu)
2237 struct vcpu_svm *svm = to_svm(vcpu);
2238 u32 error_code = svm->vmcb->control.exit_info_1;
2241 /* Both #GP cases have zero error_code */
2245 /* Decode the instruction for usage later */
2246 if (x86_decode_emulated_instruction(vcpu, 0, NULL, 0) != EMULATION_OK)
2249 opcode = svm_instr_opcode(vcpu);
2251 if (opcode == NONE_SVM_INSTR) {
2252 if (!enable_vmware_backdoor)
2256 * VMware backdoor emulation on #GP interception only handles
2257 * IN{S}, OUT{S}, and RDPMC.
2259 if (!is_guest_mode(vcpu))
2260 return kvm_emulate_instruction(vcpu,
2261 EMULTYPE_VMWARE_GP | EMULTYPE_NO_DECODE);
2263 /* All SVM instructions expect page aligned RAX */
2264 if (svm->vmcb->save.rax & ~PAGE_MASK)
2267 return emulate_svm_instr(vcpu, opcode);
2271 kvm_queue_exception_e(vcpu, GP_VECTOR, error_code);
2275 void svm_set_gif(struct vcpu_svm *svm, bool value)
2279 * If VGIF is enabled, the STGI intercept is only added to
2280 * detect the opening of the SMI/NMI window; remove it now.
2281 * Likewise, clear the VINTR intercept, we will set it
2282 * again while processing KVM_REQ_EVENT if needed.
2285 svm_clr_intercept(svm, INTERCEPT_STGI);
2286 if (svm_is_intercept(svm, INTERCEPT_VINTR))
2287 svm_clear_vintr(svm);
2290 if (svm->vcpu.arch.smi_pending ||
2291 svm->vcpu.arch.nmi_pending ||
2292 kvm_cpu_has_injectable_intr(&svm->vcpu))
2293 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2298 * After a CLGI no interrupts should come. But if vGIF is
2299 * in use, we still rely on the VINTR intercept (rather than
2300 * STGI) to detect an open interrupt window.
2303 svm_clear_vintr(svm);
2307 static int stgi_interception(struct kvm_vcpu *vcpu)
2311 if (nested_svm_check_permissions(vcpu))
2314 ret = kvm_skip_emulated_instruction(vcpu);
2315 svm_set_gif(to_svm(vcpu), true);
2319 static int clgi_interception(struct kvm_vcpu *vcpu)
2323 if (nested_svm_check_permissions(vcpu))
2326 ret = kvm_skip_emulated_instruction(vcpu);
2327 svm_set_gif(to_svm(vcpu), false);
2331 static int invlpga_interception(struct kvm_vcpu *vcpu)
2333 gva_t gva = kvm_rax_read(vcpu);
2334 u32 asid = kvm_rcx_read(vcpu);
2336 /* FIXME: Handle an address size prefix. */
2337 if (!is_long_mode(vcpu))
2340 trace_kvm_invlpga(to_svm(vcpu)->vmcb->save.rip, asid, gva);
2342 /* Let's treat INVLPGA the same as INVLPG (can be optimized!) */
2343 kvm_mmu_invlpg(vcpu, gva);
2345 return kvm_skip_emulated_instruction(vcpu);
2348 static int skinit_interception(struct kvm_vcpu *vcpu)
2350 trace_kvm_skinit(to_svm(vcpu)->vmcb->save.rip, kvm_rax_read(vcpu));
2352 kvm_queue_exception(vcpu, UD_VECTOR);
2356 static int task_switch_interception(struct kvm_vcpu *vcpu)
2358 struct vcpu_svm *svm = to_svm(vcpu);
2361 int int_type = svm->vmcb->control.exit_int_info &
2362 SVM_EXITINTINFO_TYPE_MASK;
2363 int int_vec = svm->vmcb->control.exit_int_info & SVM_EVTINJ_VEC_MASK;
2365 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_TYPE_MASK;
2367 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID;
2368 bool has_error_code = false;
2371 tss_selector = (u16)svm->vmcb->control.exit_info_1;
2373 if (svm->vmcb->control.exit_info_2 &
2374 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
2375 reason = TASK_SWITCH_IRET;
2376 else if (svm->vmcb->control.exit_info_2 &
2377 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
2378 reason = TASK_SWITCH_JMP;
2380 reason = TASK_SWITCH_GATE;
2382 reason = TASK_SWITCH_CALL;
2384 if (reason == TASK_SWITCH_GATE) {
2386 case SVM_EXITINTINFO_TYPE_NMI:
2387 vcpu->arch.nmi_injected = false;
2389 case SVM_EXITINTINFO_TYPE_EXEPT:
2390 if (svm->vmcb->control.exit_info_2 &
2391 (1ULL << SVM_EXITINFOSHIFT_TS_HAS_ERROR_CODE)) {
2392 has_error_code = true;
2394 (u32)svm->vmcb->control.exit_info_2;
2396 kvm_clear_exception_queue(vcpu);
2398 case SVM_EXITINTINFO_TYPE_INTR:
2399 kvm_clear_interrupt_queue(vcpu);
2406 if (reason != TASK_SWITCH_GATE ||
2407 int_type == SVM_EXITINTINFO_TYPE_SOFT ||
2408 (int_type == SVM_EXITINTINFO_TYPE_EXEPT &&
2409 (int_vec == OF_VECTOR || int_vec == BP_VECTOR))) {
2410 if (!svm_skip_emulated_instruction(vcpu))
2414 if (int_type != SVM_EXITINTINFO_TYPE_SOFT)
2417 return kvm_task_switch(vcpu, tss_selector, int_vec, reason,
2418 has_error_code, error_code);
2421 static int iret_interception(struct kvm_vcpu *vcpu)
2423 struct vcpu_svm *svm = to_svm(vcpu);
2425 ++vcpu->stat.nmi_window_exits;
2426 vcpu->arch.hflags |= HF_IRET_MASK;
2427 if (!sev_es_guest(vcpu->kvm)) {
2428 svm_clr_intercept(svm, INTERCEPT_IRET);
2429 svm->nmi_iret_rip = kvm_rip_read(vcpu);
2431 kvm_make_request(KVM_REQ_EVENT, vcpu);
2435 static int invlpg_interception(struct kvm_vcpu *vcpu)
2437 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2438 return kvm_emulate_instruction(vcpu, 0);
2440 kvm_mmu_invlpg(vcpu, to_svm(vcpu)->vmcb->control.exit_info_1);
2441 return kvm_skip_emulated_instruction(vcpu);
2444 static int emulate_on_interception(struct kvm_vcpu *vcpu)
2446 return kvm_emulate_instruction(vcpu, 0);
2449 static int rsm_interception(struct kvm_vcpu *vcpu)
2451 return kvm_emulate_instruction_from_buffer(vcpu, rsm_ins_bytes, 2);
2454 static bool check_selective_cr0_intercepted(struct kvm_vcpu *vcpu,
2457 struct vcpu_svm *svm = to_svm(vcpu);
2458 unsigned long cr0 = vcpu->arch.cr0;
2461 if (!is_guest_mode(vcpu) ||
2462 (!(vmcb12_is_intercept(&svm->nested.ctl, INTERCEPT_SELECTIVE_CR0))))
2465 cr0 &= ~SVM_CR0_SELECTIVE_MASK;
2466 val &= ~SVM_CR0_SELECTIVE_MASK;
2469 svm->vmcb->control.exit_code = SVM_EXIT_CR0_SEL_WRITE;
2470 ret = (nested_svm_exit_handled(svm) == NESTED_EXIT_DONE);
2476 #define CR_VALID (1ULL << 63)
2478 static int cr_interception(struct kvm_vcpu *vcpu)
2480 struct vcpu_svm *svm = to_svm(vcpu);
2485 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2486 return emulate_on_interception(vcpu);
2488 if (unlikely((svm->vmcb->control.exit_info_1 & CR_VALID) == 0))
2489 return emulate_on_interception(vcpu);
2491 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2492 if (svm->vmcb->control.exit_code == SVM_EXIT_CR0_SEL_WRITE)
2493 cr = SVM_EXIT_WRITE_CR0 - SVM_EXIT_READ_CR0;
2495 cr = svm->vmcb->control.exit_code - SVM_EXIT_READ_CR0;
2498 if (cr >= 16) { /* mov to cr */
2500 val = kvm_register_read(vcpu, reg);
2501 trace_kvm_cr_write(cr, val);
2504 if (!check_selective_cr0_intercepted(vcpu, val))
2505 err = kvm_set_cr0(vcpu, val);
2511 err = kvm_set_cr3(vcpu, val);
2514 err = kvm_set_cr4(vcpu, val);
2517 err = kvm_set_cr8(vcpu, val);
2520 WARN(1, "unhandled write to CR%d", cr);
2521 kvm_queue_exception(vcpu, UD_VECTOR);
2524 } else { /* mov from cr */
2527 val = kvm_read_cr0(vcpu);
2530 val = vcpu->arch.cr2;
2533 val = kvm_read_cr3(vcpu);
2536 val = kvm_read_cr4(vcpu);
2539 val = kvm_get_cr8(vcpu);
2542 WARN(1, "unhandled read from CR%d", cr);
2543 kvm_queue_exception(vcpu, UD_VECTOR);
2546 kvm_register_write(vcpu, reg, val);
2547 trace_kvm_cr_read(cr, val);
2549 return kvm_complete_insn_gp(vcpu, err);
2552 static int cr_trap(struct kvm_vcpu *vcpu)
2554 struct vcpu_svm *svm = to_svm(vcpu);
2555 unsigned long old_value, new_value;
2559 new_value = (unsigned long)svm->vmcb->control.exit_info_1;
2561 cr = svm->vmcb->control.exit_code - SVM_EXIT_CR0_WRITE_TRAP;
2564 old_value = kvm_read_cr0(vcpu);
2565 svm_set_cr0(vcpu, new_value);
2567 kvm_post_set_cr0(vcpu, old_value, new_value);
2570 old_value = kvm_read_cr4(vcpu);
2571 svm_set_cr4(vcpu, new_value);
2573 kvm_post_set_cr4(vcpu, old_value, new_value);
2576 ret = kvm_set_cr8(vcpu, new_value);
2579 WARN(1, "unhandled CR%d write trap", cr);
2580 kvm_queue_exception(vcpu, UD_VECTOR);
2584 return kvm_complete_insn_gp(vcpu, ret);
2587 static int dr_interception(struct kvm_vcpu *vcpu)
2589 struct vcpu_svm *svm = to_svm(vcpu);
2594 if (vcpu->guest_debug == 0) {
2596 * No more DR vmexits; force a reload of the debug registers
2597 * and reenter on this instruction. The next vmexit will
2598 * retrieve the full state of the debug registers.
2600 clr_dr_intercepts(svm);
2601 vcpu->arch.switch_db_regs |= KVM_DEBUGREG_WONT_EXIT;
2605 if (!boot_cpu_has(X86_FEATURE_DECODEASSISTS))
2606 return emulate_on_interception(vcpu);
2608 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2609 dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
2610 if (dr >= 16) { /* mov to DRn */
2612 val = kvm_register_read(vcpu, reg);
2613 err = kvm_set_dr(vcpu, dr, val);
2615 kvm_get_dr(vcpu, dr, &val);
2616 kvm_register_write(vcpu, reg, val);
2619 return kvm_complete_insn_gp(vcpu, err);
2622 static int cr8_write_interception(struct kvm_vcpu *vcpu)
2626 u8 cr8_prev = kvm_get_cr8(vcpu);
2627 /* instruction emulation calls kvm_set_cr8() */
2628 r = cr_interception(vcpu);
2629 if (lapic_in_kernel(vcpu))
2631 if (cr8_prev <= kvm_get_cr8(vcpu))
2633 vcpu->run->exit_reason = KVM_EXIT_SET_TPR;
2637 static int efer_trap(struct kvm_vcpu *vcpu)
2639 struct msr_data msr_info;
2643 * Clear the EFER_SVME bit from EFER. The SVM code always sets this
2644 * bit in svm_set_efer(), but __kvm_valid_efer() checks it against
2645 * whether the guest has X86_FEATURE_SVM - this avoids a failure if
2646 * the guest doesn't have X86_FEATURE_SVM.
2648 msr_info.host_initiated = false;
2649 msr_info.index = MSR_EFER;
2650 msr_info.data = to_svm(vcpu)->vmcb->control.exit_info_1 & ~EFER_SVME;
2651 ret = kvm_set_msr_common(vcpu, &msr_info);
2653 return kvm_complete_insn_gp(vcpu, ret);
2656 static int svm_get_msr_feature(struct kvm_msr_entry *msr)
2660 switch (msr->index) {
2661 case MSR_F10H_DECFG:
2662 if (boot_cpu_has(X86_FEATURE_LFENCE_RDTSC))
2663 msr->data |= MSR_F10H_DECFG_LFENCE_SERIALIZE;
2665 case MSR_IA32_PERF_CAPABILITIES:
2668 return KVM_MSR_RET_INVALID;
2674 static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
2676 struct vcpu_svm *svm = to_svm(vcpu);
2678 switch (msr_info->index) {
2679 case MSR_AMD64_TSC_RATIO:
2680 if (!msr_info->host_initiated && !svm->tsc_scaling_enabled)
2682 msr_info->data = svm->tsc_ratio_msr;
2685 msr_info->data = svm->vmcb01.ptr->save.star;
2687 #ifdef CONFIG_X86_64
2689 msr_info->data = svm->vmcb01.ptr->save.lstar;
2692 msr_info->data = svm->vmcb01.ptr->save.cstar;
2694 case MSR_KERNEL_GS_BASE:
2695 msr_info->data = svm->vmcb01.ptr->save.kernel_gs_base;
2697 case MSR_SYSCALL_MASK:
2698 msr_info->data = svm->vmcb01.ptr->save.sfmask;
2701 case MSR_IA32_SYSENTER_CS:
2702 msr_info->data = svm->vmcb01.ptr->save.sysenter_cs;
2704 case MSR_IA32_SYSENTER_EIP:
2705 msr_info->data = (u32)svm->vmcb01.ptr->save.sysenter_eip;
2706 if (guest_cpuid_is_intel(vcpu))
2707 msr_info->data |= (u64)svm->sysenter_eip_hi << 32;
2709 case MSR_IA32_SYSENTER_ESP:
2710 msr_info->data = svm->vmcb01.ptr->save.sysenter_esp;
2711 if (guest_cpuid_is_intel(vcpu))
2712 msr_info->data |= (u64)svm->sysenter_esp_hi << 32;
2715 msr_info->data = svm->tsc_aux;
2717 case MSR_IA32_DEBUGCTLMSR:
2718 case MSR_IA32_LASTBRANCHFROMIP:
2719 case MSR_IA32_LASTBRANCHTOIP:
2720 case MSR_IA32_LASTINTFROMIP:
2721 case MSR_IA32_LASTINTTOIP:
2722 msr_info->data = svm_get_lbr_msr(svm, msr_info->index);
2724 case MSR_VM_HSAVE_PA:
2725 msr_info->data = svm->nested.hsave_msr;
2728 msr_info->data = svm->nested.vm_cr_msr;
2730 case MSR_IA32_SPEC_CTRL:
2731 if (!msr_info->host_initiated &&
2732 !guest_has_spec_ctrl_msr(vcpu))
2735 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2736 msr_info->data = svm->vmcb->save.spec_ctrl;
2738 msr_info->data = svm->spec_ctrl;
2740 case MSR_AMD64_VIRT_SPEC_CTRL:
2741 if (!msr_info->host_initiated &&
2742 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2745 msr_info->data = svm->virt_spec_ctrl;
2747 case MSR_F15H_IC_CFG: {
2751 family = guest_cpuid_family(vcpu);
2752 model = guest_cpuid_model(vcpu);
2754 if (family < 0 || model < 0)
2755 return kvm_get_msr_common(vcpu, msr_info);
2759 if (family == 0x15 &&
2760 (model >= 0x2 && model < 0x20))
2761 msr_info->data = 0x1E;
2764 case MSR_F10H_DECFG:
2765 msr_info->data = svm->msr_decfg;
2768 return kvm_get_msr_common(vcpu, msr_info);
2773 static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
2775 struct vcpu_svm *svm = to_svm(vcpu);
2776 if (!err || !sev_es_guest(vcpu->kvm) || WARN_ON_ONCE(!svm->sev_es.ghcb))
2777 return kvm_complete_insn_gp(vcpu, err);
2779 ghcb_set_sw_exit_info_1(svm->sev_es.ghcb, 1);
2780 ghcb_set_sw_exit_info_2(svm->sev_es.ghcb,
2782 SVM_EVTINJ_TYPE_EXEPT |
2787 static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
2789 struct vcpu_svm *svm = to_svm(vcpu);
2790 int svm_dis, chg_mask;
2792 if (data & ~SVM_VM_CR_VALID_MASK)
2795 chg_mask = SVM_VM_CR_VALID_MASK;
2797 if (svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK)
2798 chg_mask &= ~(SVM_VM_CR_SVM_LOCK_MASK | SVM_VM_CR_SVM_DIS_MASK);
2800 svm->nested.vm_cr_msr &= ~chg_mask;
2801 svm->nested.vm_cr_msr |= (data & chg_mask);
2803 svm_dis = svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK;
2805 /* check for svm_disable while efer.svme is set */
2806 if (svm_dis && (vcpu->arch.efer & EFER_SVME))
2812 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
2814 struct vcpu_svm *svm = to_svm(vcpu);
2817 u32 ecx = msr->index;
2818 u64 data = msr->data;
2820 case MSR_AMD64_TSC_RATIO:
2822 if (!svm->tsc_scaling_enabled) {
2824 if (!msr->host_initiated)
2827 * In case TSC scaling is not enabled, always
2828 * leave this MSR at the default value.
2830 * Due to bug in qemu 6.2.0, it would try to set
2831 * this msr to 0 if tsc scaling is not enabled.
2832 * Ignore this value as well.
2834 if (data != 0 && data != svm->tsc_ratio_msr)
2839 if (data & SVM_TSC_RATIO_RSVD)
2842 svm->tsc_ratio_msr = data;
2844 if (svm->tsc_scaling_enabled && is_guest_mode(vcpu))
2845 nested_svm_update_tsc_ratio_msr(vcpu);
2848 case MSR_IA32_CR_PAT:
2849 if (!kvm_mtrr_valid(vcpu, MSR_IA32_CR_PAT, data))
2851 vcpu->arch.pat = data;
2852 svm->vmcb01.ptr->save.g_pat = data;
2853 if (is_guest_mode(vcpu))
2854 nested_vmcb02_compute_g_pat(svm);
2855 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
2857 case MSR_IA32_SPEC_CTRL:
2858 if (!msr->host_initiated &&
2859 !guest_has_spec_ctrl_msr(vcpu))
2862 if (kvm_spec_ctrl_test_value(data))
2865 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2866 svm->vmcb->save.spec_ctrl = data;
2868 svm->spec_ctrl = data;
2874 * When it's written (to non-zero) for the first time, pass
2878 * The handling of the MSR bitmap for L2 guests is done in
2879 * nested_svm_vmrun_msrpm.
2880 * We update the L1 MSR bit as well since it will end up
2881 * touching the MSR anyway now.
2883 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
2885 case MSR_IA32_PRED_CMD:
2886 if (!msr->host_initiated &&
2887 !guest_has_pred_cmd_msr(vcpu))
2890 if (data & ~PRED_CMD_IBPB)
2892 if (!boot_cpu_has(X86_FEATURE_IBPB))
2897 wrmsrl(MSR_IA32_PRED_CMD, PRED_CMD_IBPB);
2898 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_PRED_CMD, 0, 1);
2900 case MSR_AMD64_VIRT_SPEC_CTRL:
2901 if (!msr->host_initiated &&
2902 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2905 if (data & ~SPEC_CTRL_SSBD)
2908 svm->virt_spec_ctrl = data;
2911 svm->vmcb01.ptr->save.star = data;
2913 #ifdef CONFIG_X86_64
2915 svm->vmcb01.ptr->save.lstar = data;
2918 svm->vmcb01.ptr->save.cstar = data;
2920 case MSR_KERNEL_GS_BASE:
2921 svm->vmcb01.ptr->save.kernel_gs_base = data;
2923 case MSR_SYSCALL_MASK:
2924 svm->vmcb01.ptr->save.sfmask = data;
2927 case MSR_IA32_SYSENTER_CS:
2928 svm->vmcb01.ptr->save.sysenter_cs = data;
2930 case MSR_IA32_SYSENTER_EIP:
2931 svm->vmcb01.ptr->save.sysenter_eip = (u32)data;
2933 * We only intercept the MSR_IA32_SYSENTER_{EIP|ESP} msrs
2934 * when we spoof an Intel vendor ID (for cross vendor migration).
2935 * In this case we use this intercept to track the high
2936 * 32 bit part of these msrs to support Intel's
2937 * implementation of SYSENTER/SYSEXIT.
2939 svm->sysenter_eip_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2941 case MSR_IA32_SYSENTER_ESP:
2942 svm->vmcb01.ptr->save.sysenter_esp = (u32)data;
2943 svm->sysenter_esp_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2947 * TSC_AUX is usually changed only during boot and never read
2948 * directly. Intercept TSC_AUX instead of exposing it to the
2949 * guest via direct_access_msrs, and switch it via user return.
2952 r = kvm_set_user_return_msr(tsc_aux_uret_slot, data, -1ull);
2957 svm->tsc_aux = data;
2959 case MSR_IA32_DEBUGCTLMSR:
2961 vcpu_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTL 0x%llx, nop\n",
2965 if (data & DEBUGCTL_RESERVED_BITS)
2968 if (svm->vmcb->control.virt_ext & LBR_CTL_ENABLE_MASK)
2969 svm->vmcb->save.dbgctl = data;
2971 svm->vmcb01.ptr->save.dbgctl = data;
2973 svm_update_lbrv(vcpu);
2976 case MSR_VM_HSAVE_PA:
2978 * Old kernels did not validate the value written to
2979 * MSR_VM_HSAVE_PA. Allow KVM_SET_MSR to set an invalid
2980 * value to allow live migrating buggy or malicious guests
2981 * originating from those kernels.
2983 if (!msr->host_initiated && !page_address_valid(vcpu, data))
2986 svm->nested.hsave_msr = data & PAGE_MASK;
2989 return svm_set_vm_cr(vcpu, data);
2991 vcpu_unimpl(vcpu, "unimplemented wrmsr: 0x%x data 0x%llx\n", ecx, data);
2993 case MSR_F10H_DECFG: {
2994 struct kvm_msr_entry msr_entry;
2996 msr_entry.index = msr->index;
2997 if (svm_get_msr_feature(&msr_entry))
3000 /* Check the supported bits */
3001 if (data & ~msr_entry.data)
3004 /* Don't allow the guest to change a bit, #GP */
3005 if (!msr->host_initiated && (data ^ msr_entry.data))
3008 svm->msr_decfg = data;
3012 return kvm_set_msr_common(vcpu, msr);
3017 static int msr_interception(struct kvm_vcpu *vcpu)
3019 if (to_svm(vcpu)->vmcb->control.exit_info_1)
3020 return kvm_emulate_wrmsr(vcpu);
3022 return kvm_emulate_rdmsr(vcpu);
3025 static int interrupt_window_interception(struct kvm_vcpu *vcpu)
3027 kvm_make_request(KVM_REQ_EVENT, vcpu);
3028 svm_clear_vintr(to_svm(vcpu));
3031 * If not running nested, for AVIC, the only reason to end up here is ExtINTs.
3032 * In this case AVIC was temporarily disabled for
3033 * requesting the IRQ window and we have to re-enable it.
3035 * If running nested, still remove the VM wide AVIC inhibit to
3036 * support case in which the interrupt window was requested when the
3037 * vCPU was not running nested.
3039 * All vCPUs which run still run nested, will remain to have their
3040 * AVIC still inhibited due to per-cpu AVIC inhibition.
3042 kvm_clear_apicv_inhibit(vcpu->kvm, APICV_INHIBIT_REASON_IRQWIN);
3044 ++vcpu->stat.irq_window_exits;
3048 static int pause_interception(struct kvm_vcpu *vcpu)
3052 * CPL is not made available for an SEV-ES guest, therefore
3053 * vcpu->arch.preempted_in_kernel can never be true. Just
3054 * set in_kernel to false as well.
3056 in_kernel = !sev_es_guest(vcpu->kvm) && svm_get_cpl(vcpu) == 0;
3058 grow_ple_window(vcpu);
3060 kvm_vcpu_on_spin(vcpu, in_kernel);
3061 return kvm_skip_emulated_instruction(vcpu);
3064 static int invpcid_interception(struct kvm_vcpu *vcpu)
3066 struct vcpu_svm *svm = to_svm(vcpu);
3070 if (!guest_cpuid_has(vcpu, X86_FEATURE_INVPCID)) {
3071 kvm_queue_exception(vcpu, UD_VECTOR);
3076 * For an INVPCID intercept:
3077 * EXITINFO1 provides the linear address of the memory operand.
3078 * EXITINFO2 provides the contents of the register operand.
3080 type = svm->vmcb->control.exit_info_2;
3081 gva = svm->vmcb->control.exit_info_1;
3083 return kvm_handle_invpcid(vcpu, type, gva);
3086 static int (*const svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {
3087 [SVM_EXIT_READ_CR0] = cr_interception,
3088 [SVM_EXIT_READ_CR3] = cr_interception,
3089 [SVM_EXIT_READ_CR4] = cr_interception,
3090 [SVM_EXIT_READ_CR8] = cr_interception,
3091 [SVM_EXIT_CR0_SEL_WRITE] = cr_interception,
3092 [SVM_EXIT_WRITE_CR0] = cr_interception,
3093 [SVM_EXIT_WRITE_CR3] = cr_interception,
3094 [SVM_EXIT_WRITE_CR4] = cr_interception,
3095 [SVM_EXIT_WRITE_CR8] = cr8_write_interception,
3096 [SVM_EXIT_READ_DR0] = dr_interception,
3097 [SVM_EXIT_READ_DR1] = dr_interception,
3098 [SVM_EXIT_READ_DR2] = dr_interception,
3099 [SVM_EXIT_READ_DR3] = dr_interception,
3100 [SVM_EXIT_READ_DR4] = dr_interception,
3101 [SVM_EXIT_READ_DR5] = dr_interception,
3102 [SVM_EXIT_READ_DR6] = dr_interception,
3103 [SVM_EXIT_READ_DR7] = dr_interception,
3104 [SVM_EXIT_WRITE_DR0] = dr_interception,
3105 [SVM_EXIT_WRITE_DR1] = dr_interception,
3106 [SVM_EXIT_WRITE_DR2] = dr_interception,
3107 [SVM_EXIT_WRITE_DR3] = dr_interception,
3108 [SVM_EXIT_WRITE_DR4] = dr_interception,
3109 [SVM_EXIT_WRITE_DR5] = dr_interception,
3110 [SVM_EXIT_WRITE_DR6] = dr_interception,
3111 [SVM_EXIT_WRITE_DR7] = dr_interception,
3112 [SVM_EXIT_EXCP_BASE + DB_VECTOR] = db_interception,
3113 [SVM_EXIT_EXCP_BASE + BP_VECTOR] = bp_interception,
3114 [SVM_EXIT_EXCP_BASE + UD_VECTOR] = ud_interception,
3115 [SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
3116 [SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
3117 [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
3118 [SVM_EXIT_EXCP_BASE + GP_VECTOR] = gp_interception,
3119 [SVM_EXIT_INTR] = intr_interception,
3120 [SVM_EXIT_NMI] = nmi_interception,
3121 [SVM_EXIT_SMI] = smi_interception,
3122 [SVM_EXIT_VINTR] = interrupt_window_interception,
3123 [SVM_EXIT_RDPMC] = kvm_emulate_rdpmc,
3124 [SVM_EXIT_CPUID] = kvm_emulate_cpuid,
3125 [SVM_EXIT_IRET] = iret_interception,
3126 [SVM_EXIT_INVD] = kvm_emulate_invd,
3127 [SVM_EXIT_PAUSE] = pause_interception,
3128 [SVM_EXIT_HLT] = kvm_emulate_halt,
3129 [SVM_EXIT_INVLPG] = invlpg_interception,
3130 [SVM_EXIT_INVLPGA] = invlpga_interception,
3131 [SVM_EXIT_IOIO] = io_interception,
3132 [SVM_EXIT_MSR] = msr_interception,
3133 [SVM_EXIT_TASK_SWITCH] = task_switch_interception,
3134 [SVM_EXIT_SHUTDOWN] = shutdown_interception,
3135 [SVM_EXIT_VMRUN] = vmrun_interception,
3136 [SVM_EXIT_VMMCALL] = kvm_emulate_hypercall,
3137 [SVM_EXIT_VMLOAD] = vmload_interception,
3138 [SVM_EXIT_VMSAVE] = vmsave_interception,
3139 [SVM_EXIT_STGI] = stgi_interception,
3140 [SVM_EXIT_CLGI] = clgi_interception,
3141 [SVM_EXIT_SKINIT] = skinit_interception,
3142 [SVM_EXIT_RDTSCP] = kvm_handle_invalid_op,
3143 [SVM_EXIT_WBINVD] = kvm_emulate_wbinvd,
3144 [SVM_EXIT_MONITOR] = kvm_emulate_monitor,
3145 [SVM_EXIT_MWAIT] = kvm_emulate_mwait,
3146 [SVM_EXIT_XSETBV] = kvm_emulate_xsetbv,
3147 [SVM_EXIT_RDPRU] = kvm_handle_invalid_op,
3148 [SVM_EXIT_EFER_WRITE_TRAP] = efer_trap,
3149 [SVM_EXIT_CR0_WRITE_TRAP] = cr_trap,
3150 [SVM_EXIT_CR4_WRITE_TRAP] = cr_trap,
3151 [SVM_EXIT_CR8_WRITE_TRAP] = cr_trap,
3152 [SVM_EXIT_INVPCID] = invpcid_interception,
3153 [SVM_EXIT_NPF] = npf_interception,
3154 [SVM_EXIT_RSM] = rsm_interception,
3155 [SVM_EXIT_AVIC_INCOMPLETE_IPI] = avic_incomplete_ipi_interception,
3156 [SVM_EXIT_AVIC_UNACCELERATED_ACCESS] = avic_unaccelerated_access_interception,
3157 [SVM_EXIT_VMGEXIT] = sev_handle_vmgexit,
3160 static void dump_vmcb(struct kvm_vcpu *vcpu)
3162 struct vcpu_svm *svm = to_svm(vcpu);
3163 struct vmcb_control_area *control = &svm->vmcb->control;
3164 struct vmcb_save_area *save = &svm->vmcb->save;
3165 struct vmcb_save_area *save01 = &svm->vmcb01.ptr->save;
3167 if (!dump_invalid_vmcb) {
3168 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3172 pr_err("VMCB %p, last attempted VMRUN on CPU %d\n",
3173 svm->current_vmcb->ptr, vcpu->arch.last_vmentry_cpu);
3174 pr_err("VMCB Control Area:\n");
3175 pr_err("%-20s%04x\n", "cr_read:", control->intercepts[INTERCEPT_CR] & 0xffff);
3176 pr_err("%-20s%04x\n", "cr_write:", control->intercepts[INTERCEPT_CR] >> 16);
3177 pr_err("%-20s%04x\n", "dr_read:", control->intercepts[INTERCEPT_DR] & 0xffff);
3178 pr_err("%-20s%04x\n", "dr_write:", control->intercepts[INTERCEPT_DR] >> 16);
3179 pr_err("%-20s%08x\n", "exceptions:", control->intercepts[INTERCEPT_EXCEPTION]);
3180 pr_err("%-20s%08x %08x\n", "intercepts:",
3181 control->intercepts[INTERCEPT_WORD3],
3182 control->intercepts[INTERCEPT_WORD4]);
3183 pr_err("%-20s%d\n", "pause filter count:", control->pause_filter_count);
3184 pr_err("%-20s%d\n", "pause filter threshold:",
3185 control->pause_filter_thresh);
3186 pr_err("%-20s%016llx\n", "iopm_base_pa:", control->iopm_base_pa);
3187 pr_err("%-20s%016llx\n", "msrpm_base_pa:", control->msrpm_base_pa);
3188 pr_err("%-20s%016llx\n", "tsc_offset:", control->tsc_offset);
3189 pr_err("%-20s%d\n", "asid:", control->asid);
3190 pr_err("%-20s%d\n", "tlb_ctl:", control->tlb_ctl);
3191 pr_err("%-20s%08x\n", "int_ctl:", control->int_ctl);
3192 pr_err("%-20s%08x\n", "int_vector:", control->int_vector);
3193 pr_err("%-20s%08x\n", "int_state:", control->int_state);
3194 pr_err("%-20s%08x\n", "exit_code:", control->exit_code);
3195 pr_err("%-20s%016llx\n", "exit_info1:", control->exit_info_1);
3196 pr_err("%-20s%016llx\n", "exit_info2:", control->exit_info_2);
3197 pr_err("%-20s%08x\n", "exit_int_info:", control->exit_int_info);
3198 pr_err("%-20s%08x\n", "exit_int_info_err:", control->exit_int_info_err);
3199 pr_err("%-20s%lld\n", "nested_ctl:", control->nested_ctl);
3200 pr_err("%-20s%016llx\n", "nested_cr3:", control->nested_cr3);
3201 pr_err("%-20s%016llx\n", "avic_vapic_bar:", control->avic_vapic_bar);
3202 pr_err("%-20s%016llx\n", "ghcb:", control->ghcb_gpa);
3203 pr_err("%-20s%08x\n", "event_inj:", control->event_inj);
3204 pr_err("%-20s%08x\n", "event_inj_err:", control->event_inj_err);
3205 pr_err("%-20s%lld\n", "virt_ext:", control->virt_ext);
3206 pr_err("%-20s%016llx\n", "next_rip:", control->next_rip);
3207 pr_err("%-20s%016llx\n", "avic_backing_page:", control->avic_backing_page);
3208 pr_err("%-20s%016llx\n", "avic_logical_id:", control->avic_logical_id);
3209 pr_err("%-20s%016llx\n", "avic_physical_id:", control->avic_physical_id);
3210 pr_err("%-20s%016llx\n", "vmsa_pa:", control->vmsa_pa);
3211 pr_err("VMCB State Save Area:\n");
3212 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3214 save->es.selector, save->es.attrib,
3215 save->es.limit, save->es.base);
3216 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3218 save->cs.selector, save->cs.attrib,
3219 save->cs.limit, save->cs.base);
3220 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3222 save->ss.selector, save->ss.attrib,
3223 save->ss.limit, save->ss.base);
3224 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3226 save->ds.selector, save->ds.attrib,
3227 save->ds.limit, save->ds.base);
3228 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3230 save01->fs.selector, save01->fs.attrib,
3231 save01->fs.limit, save01->fs.base);
3232 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3234 save01->gs.selector, save01->gs.attrib,
3235 save01->gs.limit, save01->gs.base);
3236 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3238 save->gdtr.selector, save->gdtr.attrib,
3239 save->gdtr.limit, save->gdtr.base);
3240 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3242 save01->ldtr.selector, save01->ldtr.attrib,
3243 save01->ldtr.limit, save01->ldtr.base);
3244 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3246 save->idtr.selector, save->idtr.attrib,
3247 save->idtr.limit, save->idtr.base);
3248 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3250 save01->tr.selector, save01->tr.attrib,
3251 save01->tr.limit, save01->tr.base);
3252 pr_err("vmpl: %d cpl: %d efer: %016llx\n",
3253 save->vmpl, save->cpl, save->efer);
3254 pr_err("%-15s %016llx %-13s %016llx\n",
3255 "cr0:", save->cr0, "cr2:", save->cr2);
3256 pr_err("%-15s %016llx %-13s %016llx\n",
3257 "cr3:", save->cr3, "cr4:", save->cr4);
3258 pr_err("%-15s %016llx %-13s %016llx\n",
3259 "dr6:", save->dr6, "dr7:", save->dr7);
3260 pr_err("%-15s %016llx %-13s %016llx\n",
3261 "rip:", save->rip, "rflags:", save->rflags);
3262 pr_err("%-15s %016llx %-13s %016llx\n",
3263 "rsp:", save->rsp, "rax:", save->rax);
3264 pr_err("%-15s %016llx %-13s %016llx\n",
3265 "star:", save01->star, "lstar:", save01->lstar);
3266 pr_err("%-15s %016llx %-13s %016llx\n",
3267 "cstar:", save01->cstar, "sfmask:", save01->sfmask);
3268 pr_err("%-15s %016llx %-13s %016llx\n",
3269 "kernel_gs_base:", save01->kernel_gs_base,
3270 "sysenter_cs:", save01->sysenter_cs);
3271 pr_err("%-15s %016llx %-13s %016llx\n",
3272 "sysenter_esp:", save01->sysenter_esp,
3273 "sysenter_eip:", save01->sysenter_eip);
3274 pr_err("%-15s %016llx %-13s %016llx\n",
3275 "gpat:", save->g_pat, "dbgctl:", save->dbgctl);
3276 pr_err("%-15s %016llx %-13s %016llx\n",
3277 "br_from:", save->br_from, "br_to:", save->br_to);
3278 pr_err("%-15s %016llx %-13s %016llx\n",
3279 "excp_from:", save->last_excp_from,
3280 "excp_to:", save->last_excp_to);
3283 static bool svm_check_exit_valid(u64 exit_code)
3285 return (exit_code < ARRAY_SIZE(svm_exit_handlers) &&
3286 svm_exit_handlers[exit_code]);
3289 static int svm_handle_invalid_exit(struct kvm_vcpu *vcpu, u64 exit_code)
3291 vcpu_unimpl(vcpu, "svm: unexpected exit reason 0x%llx\n", exit_code);
3293 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3294 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
3295 vcpu->run->internal.ndata = 2;
3296 vcpu->run->internal.data[0] = exit_code;
3297 vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
3301 int svm_invoke_exit_handler(struct kvm_vcpu *vcpu, u64 exit_code)
3303 if (!svm_check_exit_valid(exit_code))
3304 return svm_handle_invalid_exit(vcpu, exit_code);
3306 #ifdef CONFIG_RETPOLINE
3307 if (exit_code == SVM_EXIT_MSR)
3308 return msr_interception(vcpu);
3309 else if (exit_code == SVM_EXIT_VINTR)
3310 return interrupt_window_interception(vcpu);
3311 else if (exit_code == SVM_EXIT_INTR)
3312 return intr_interception(vcpu);
3313 else if (exit_code == SVM_EXIT_HLT)
3314 return kvm_emulate_halt(vcpu);
3315 else if (exit_code == SVM_EXIT_NPF)
3316 return npf_interception(vcpu);
3318 return svm_exit_handlers[exit_code](vcpu);
3321 static void svm_get_exit_info(struct kvm_vcpu *vcpu, u32 *reason,
3322 u64 *info1, u64 *info2,
3323 u32 *intr_info, u32 *error_code)
3325 struct vmcb_control_area *control = &to_svm(vcpu)->vmcb->control;
3327 *reason = control->exit_code;
3328 *info1 = control->exit_info_1;
3329 *info2 = control->exit_info_2;
3330 *intr_info = control->exit_int_info;
3331 if ((*intr_info & SVM_EXITINTINFO_VALID) &&
3332 (*intr_info & SVM_EXITINTINFO_VALID_ERR))
3333 *error_code = control->exit_int_info_err;
3338 static int svm_handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
3340 struct vcpu_svm *svm = to_svm(vcpu);
3341 struct kvm_run *kvm_run = vcpu->run;
3342 u32 exit_code = svm->vmcb->control.exit_code;
3344 trace_kvm_exit(vcpu, KVM_ISA_SVM);
3346 /* SEV-ES guests must use the CR write traps to track CR registers. */
3347 if (!sev_es_guest(vcpu->kvm)) {
3348 if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
3349 vcpu->arch.cr0 = svm->vmcb->save.cr0;
3351 vcpu->arch.cr3 = svm->vmcb->save.cr3;
3354 if (is_guest_mode(vcpu)) {
3357 trace_kvm_nested_vmexit(vcpu, KVM_ISA_SVM);
3359 vmexit = nested_svm_exit_special(svm);
3361 if (vmexit == NESTED_EXIT_CONTINUE)
3362 vmexit = nested_svm_exit_handled(svm);
3364 if (vmexit == NESTED_EXIT_DONE)
3368 if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
3369 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3370 kvm_run->fail_entry.hardware_entry_failure_reason
3371 = svm->vmcb->control.exit_code;
3372 kvm_run->fail_entry.cpu = vcpu->arch.last_vmentry_cpu;
3377 if (is_external_interrupt(svm->vmcb->control.exit_int_info) &&
3378 exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR &&
3379 exit_code != SVM_EXIT_NPF && exit_code != SVM_EXIT_TASK_SWITCH &&
3380 exit_code != SVM_EXIT_INTR && exit_code != SVM_EXIT_NMI)
3381 printk(KERN_ERR "%s: unexpected exit_int_info 0x%x "
3383 __func__, svm->vmcb->control.exit_int_info,
3386 if (exit_fastpath != EXIT_FASTPATH_NONE)
3389 return svm_invoke_exit_handler(vcpu, exit_code);
3392 static void reload_tss(struct kvm_vcpu *vcpu)
3394 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3396 sd->tss_desc->type = 9; /* available 32/64-bit TSS */
3400 static void pre_svm_run(struct kvm_vcpu *vcpu)
3402 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3403 struct vcpu_svm *svm = to_svm(vcpu);
3406 * If the previous vmrun of the vmcb occurred on a different physical
3407 * cpu, then mark the vmcb dirty and assign a new asid. Hardware's
3408 * vmcb clean bits are per logical CPU, as are KVM's asid assignments.
3410 if (unlikely(svm->current_vmcb->cpu != vcpu->cpu)) {
3411 svm->current_vmcb->asid_generation = 0;
3412 vmcb_mark_all_dirty(svm->vmcb);
3413 svm->current_vmcb->cpu = vcpu->cpu;
3416 if (sev_guest(vcpu->kvm))
3417 return pre_sev_run(svm, vcpu->cpu);
3419 /* FIXME: handle wraparound of asid_generation */
3420 if (svm->current_vmcb->asid_generation != sd->asid_generation)
3424 static void svm_inject_nmi(struct kvm_vcpu *vcpu)
3426 struct vcpu_svm *svm = to_svm(vcpu);
3428 svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
3430 if (svm->nmi_l1_to_l2)
3433 vcpu->arch.hflags |= HF_NMI_MASK;
3434 if (!sev_es_guest(vcpu->kvm))
3435 svm_set_intercept(svm, INTERCEPT_IRET);
3436 ++vcpu->stat.nmi_injections;
3439 static void svm_inject_irq(struct kvm_vcpu *vcpu, bool reinjected)
3441 struct vcpu_svm *svm = to_svm(vcpu);
3444 if (vcpu->arch.interrupt.soft) {
3445 if (svm_update_soft_interrupt_rip(vcpu))
3448 type = SVM_EVTINJ_TYPE_SOFT;
3450 type = SVM_EVTINJ_TYPE_INTR;
3453 trace_kvm_inj_virq(vcpu->arch.interrupt.nr,
3454 vcpu->arch.interrupt.soft, reinjected);
3455 ++vcpu->stat.irq_injections;
3457 svm->vmcb->control.event_inj = vcpu->arch.interrupt.nr |
3458 SVM_EVTINJ_VALID | type;
3461 void svm_complete_interrupt_delivery(struct kvm_vcpu *vcpu, int delivery_mode,
3462 int trig_mode, int vector)
3465 * apic->apicv_active must be read after vcpu->mode.
3466 * Pairs with smp_store_release in vcpu_enter_guest.
3468 bool in_guest_mode = (smp_load_acquire(&vcpu->mode) == IN_GUEST_MODE);
3470 /* Note, this is called iff the local APIC is in-kernel. */
3471 if (!READ_ONCE(vcpu->arch.apic->apicv_active)) {
3472 /* Process the interrupt via inject_pending_event */
3473 kvm_make_request(KVM_REQ_EVENT, vcpu);
3474 kvm_vcpu_kick(vcpu);
3478 trace_kvm_apicv_accept_irq(vcpu->vcpu_id, delivery_mode, trig_mode, vector);
3479 if (in_guest_mode) {
3481 * Signal the doorbell to tell hardware to inject the IRQ. If
3482 * the vCPU exits the guest before the doorbell chimes, hardware
3483 * will automatically process AVIC interrupts at the next VMRUN.
3485 avic_ring_doorbell(vcpu);
3488 * Wake the vCPU if it was blocking. KVM will then detect the
3489 * pending IRQ when checking if the vCPU has a wake event.
3491 kvm_vcpu_wake_up(vcpu);
3495 static void svm_deliver_interrupt(struct kvm_lapic *apic, int delivery_mode,
3496 int trig_mode, int vector)
3498 kvm_lapic_set_irr(vector, apic);
3501 * Pairs with the smp_mb_*() after setting vcpu->guest_mode in
3502 * vcpu_enter_guest() to ensure the write to the vIRR is ordered before
3503 * the read of guest_mode. This guarantees that either VMRUN will see
3504 * and process the new vIRR entry, or that svm_complete_interrupt_delivery
3505 * will signal the doorbell if the CPU has already entered the guest.
3507 smp_mb__after_atomic();
3508 svm_complete_interrupt_delivery(apic->vcpu, delivery_mode, trig_mode, vector);
3511 static void svm_update_cr8_intercept(struct kvm_vcpu *vcpu, int tpr, int irr)
3513 struct vcpu_svm *svm = to_svm(vcpu);
3516 * SEV-ES guests must always keep the CR intercepts cleared. CR
3517 * tracking is done using the CR write traps.
3519 if (sev_es_guest(vcpu->kvm))
3522 if (nested_svm_virtualize_tpr(vcpu))
3525 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3531 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
3534 bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
3536 struct vcpu_svm *svm = to_svm(vcpu);
3537 struct vmcb *vmcb = svm->vmcb;
3543 if (is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3546 ret = (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK) ||
3547 (vcpu->arch.hflags & HF_NMI_MASK);
3552 static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3554 struct vcpu_svm *svm = to_svm(vcpu);
3555 if (svm->nested.nested_run_pending)
3558 if (svm_nmi_blocked(vcpu))
3561 /* An NMI must not be injected into L2 if it's supposed to VM-Exit. */
3562 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3567 static bool svm_get_nmi_mask(struct kvm_vcpu *vcpu)
3569 return !!(vcpu->arch.hflags & HF_NMI_MASK);
3572 static void svm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
3574 struct vcpu_svm *svm = to_svm(vcpu);
3577 vcpu->arch.hflags |= HF_NMI_MASK;
3578 if (!sev_es_guest(vcpu->kvm))
3579 svm_set_intercept(svm, INTERCEPT_IRET);
3581 vcpu->arch.hflags &= ~HF_NMI_MASK;
3582 if (!sev_es_guest(vcpu->kvm))
3583 svm_clr_intercept(svm, INTERCEPT_IRET);
3587 bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
3589 struct vcpu_svm *svm = to_svm(vcpu);
3590 struct vmcb *vmcb = svm->vmcb;
3595 if (is_guest_mode(vcpu)) {
3596 /* As long as interrupts are being delivered... */
3597 if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
3598 ? !(svm->vmcb01.ptr->save.rflags & X86_EFLAGS_IF)
3599 : !(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3602 /* ... vmexits aren't blocked by the interrupt shadow */
3603 if (nested_exit_on_intr(svm))
3606 if (!svm_get_if_flag(vcpu))
3610 return (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK);
3613 static int svm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3615 struct vcpu_svm *svm = to_svm(vcpu);
3617 if (svm->nested.nested_run_pending)
3620 if (svm_interrupt_blocked(vcpu))
3624 * An IRQ must not be injected into L2 if it's supposed to VM-Exit,
3625 * e.g. if the IRQ arrived asynchronously after checking nested events.
3627 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_intr(svm))
3633 static void svm_enable_irq_window(struct kvm_vcpu *vcpu)
3635 struct vcpu_svm *svm = to_svm(vcpu);
3638 * In case GIF=0 we can't rely on the CPU to tell us when GIF becomes
3639 * 1, because that's a separate STGI/VMRUN intercept. The next time we
3640 * get that intercept, this function will be called again though and
3641 * we'll get the vintr intercept. However, if the vGIF feature is
3642 * enabled, the STGI interception will not occur. Enable the irq
3643 * window under the assumption that the hardware will set the GIF.
3645 if (vgif || gif_set(svm)) {
3647 * IRQ window is not needed when AVIC is enabled,
3648 * unless we have pending ExtINT since it cannot be injected
3649 * via AVIC. In such case, KVM needs to temporarily disable AVIC,
3650 * and fallback to injecting IRQ via V_IRQ.
3652 * If running nested, AVIC is already locally inhibited
3653 * on this vCPU, therefore there is no need to request
3654 * the VM wide AVIC inhibition.
3656 if (!is_guest_mode(vcpu))
3657 kvm_set_apicv_inhibit(vcpu->kvm, APICV_INHIBIT_REASON_IRQWIN);
3663 static void svm_enable_nmi_window(struct kvm_vcpu *vcpu)
3665 struct vcpu_svm *svm = to_svm(vcpu);
3667 if ((vcpu->arch.hflags & (HF_NMI_MASK | HF_IRET_MASK)) == HF_NMI_MASK)
3668 return; /* IRET will cause a vm exit */
3670 if (!gif_set(svm)) {
3672 svm_set_intercept(svm, INTERCEPT_STGI);
3673 return; /* STGI will cause a vm exit */
3677 * Something prevents NMI from been injected. Single step over possible
3678 * problem (IRET or exception injection or interrupt shadow)
3680 svm->nmi_singlestep_guest_rflags = svm_get_rflags(vcpu);
3681 svm->nmi_singlestep = true;
3682 svm->vmcb->save.rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
3685 static void svm_flush_tlb_current(struct kvm_vcpu *vcpu)
3687 struct vcpu_svm *svm = to_svm(vcpu);
3690 * Flush only the current ASID even if the TLB flush was invoked via
3691 * kvm_flush_remote_tlbs(). Although flushing remote TLBs requires all
3692 * ASIDs to be flushed, KVM uses a single ASID for L1 and L2, and
3693 * unconditionally does a TLB flush on both nested VM-Enter and nested
3694 * VM-Exit (via kvm_mmu_reset_context()).
3696 if (static_cpu_has(X86_FEATURE_FLUSHBYASID))
3697 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3699 svm->current_vmcb->asid_generation--;
3702 static void svm_flush_tlb_gva(struct kvm_vcpu *vcpu, gva_t gva)
3704 struct vcpu_svm *svm = to_svm(vcpu);
3706 invlpga(gva, svm->vmcb->control.asid);
3709 static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)
3711 struct vcpu_svm *svm = to_svm(vcpu);
3713 if (nested_svm_virtualize_tpr(vcpu))
3716 if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
3717 int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
3718 kvm_set_cr8(vcpu, cr8);
3722 static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
3724 struct vcpu_svm *svm = to_svm(vcpu);
3727 if (nested_svm_virtualize_tpr(vcpu) ||
3728 kvm_vcpu_apicv_active(vcpu))
3731 cr8 = kvm_get_cr8(vcpu);
3732 svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
3733 svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
3736 static void svm_complete_soft_interrupt(struct kvm_vcpu *vcpu, u8 vector,
3739 bool is_exception = (type == SVM_EXITINTINFO_TYPE_EXEPT);
3740 bool is_soft = (type == SVM_EXITINTINFO_TYPE_SOFT);
3741 struct vcpu_svm *svm = to_svm(vcpu);
3744 * If NRIPS is enabled, KVM must snapshot the pre-VMRUN next_rip that's
3745 * associated with the original soft exception/interrupt. next_rip is
3746 * cleared on all exits that can occur while vectoring an event, so KVM
3747 * needs to manually set next_rip for re-injection. Unlike the !nrips
3748 * case below, this needs to be done if and only if KVM is re-injecting
3749 * the same event, i.e. if the event is a soft exception/interrupt,
3750 * otherwise next_rip is unused on VMRUN.
3752 if (nrips && (is_soft || (is_exception && kvm_exception_is_soft(vector))) &&
3753 kvm_is_linear_rip(vcpu, svm->soft_int_old_rip + svm->soft_int_csbase))
3754 svm->vmcb->control.next_rip = svm->soft_int_next_rip;
3756 * If NRIPS isn't enabled, KVM must manually advance RIP prior to
3757 * injecting the soft exception/interrupt. That advancement needs to
3758 * be unwound if vectoring didn't complete. Note, the new event may
3759 * not be the injected event, e.g. if KVM injected an INTn, the INTn
3760 * hit a #NP in the guest, and the #NP encountered a #PF, the #NP will
3761 * be the reported vectored event, but RIP still needs to be unwound.
3763 else if (!nrips && (is_soft || is_exception) &&
3764 kvm_is_linear_rip(vcpu, svm->soft_int_next_rip + svm->soft_int_csbase))
3765 kvm_rip_write(vcpu, svm->soft_int_old_rip);
3768 static void svm_complete_interrupts(struct kvm_vcpu *vcpu)
3770 struct vcpu_svm *svm = to_svm(vcpu);
3773 u32 exitintinfo = svm->vmcb->control.exit_int_info;
3774 bool nmi_l1_to_l2 = svm->nmi_l1_to_l2;
3775 bool soft_int_injected = svm->soft_int_injected;
3777 svm->nmi_l1_to_l2 = false;
3778 svm->soft_int_injected = false;
3781 * If we've made progress since setting HF_IRET_MASK, we've
3782 * executed an IRET and can allow NMI injection.
3784 if ((vcpu->arch.hflags & HF_IRET_MASK) &&
3785 (sev_es_guest(vcpu->kvm) ||
3786 kvm_rip_read(vcpu) != svm->nmi_iret_rip)) {
3787 vcpu->arch.hflags &= ~(HF_NMI_MASK | HF_IRET_MASK);
3788 kvm_make_request(KVM_REQ_EVENT, vcpu);
3791 vcpu->arch.nmi_injected = false;
3792 kvm_clear_exception_queue(vcpu);
3793 kvm_clear_interrupt_queue(vcpu);
3795 if (!(exitintinfo & SVM_EXITINTINFO_VALID))
3798 kvm_make_request(KVM_REQ_EVENT, vcpu);
3800 vector = exitintinfo & SVM_EXITINTINFO_VEC_MASK;
3801 type = exitintinfo & SVM_EXITINTINFO_TYPE_MASK;
3803 if (soft_int_injected)
3804 svm_complete_soft_interrupt(vcpu, vector, type);
3807 case SVM_EXITINTINFO_TYPE_NMI:
3808 vcpu->arch.nmi_injected = true;
3809 svm->nmi_l1_to_l2 = nmi_l1_to_l2;
3811 case SVM_EXITINTINFO_TYPE_EXEPT:
3813 * Never re-inject a #VC exception.
3815 if (vector == X86_TRAP_VC)
3818 if (exitintinfo & SVM_EXITINTINFO_VALID_ERR) {
3819 u32 err = svm->vmcb->control.exit_int_info_err;
3820 kvm_requeue_exception_e(vcpu, vector, err);
3823 kvm_requeue_exception(vcpu, vector);
3825 case SVM_EXITINTINFO_TYPE_INTR:
3826 kvm_queue_interrupt(vcpu, vector, false);
3828 case SVM_EXITINTINFO_TYPE_SOFT:
3829 kvm_queue_interrupt(vcpu, vector, true);
3837 static void svm_cancel_injection(struct kvm_vcpu *vcpu)
3839 struct vcpu_svm *svm = to_svm(vcpu);
3840 struct vmcb_control_area *control = &svm->vmcb->control;
3842 control->exit_int_info = control->event_inj;
3843 control->exit_int_info_err = control->event_inj_err;
3844 control->event_inj = 0;
3845 svm_complete_interrupts(vcpu);
3848 static int svm_vcpu_pre_run(struct kvm_vcpu *vcpu)
3853 static fastpath_t svm_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
3855 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_MSR &&
3856 to_svm(vcpu)->vmcb->control.exit_info_1)
3857 return handle_fastpath_set_msr_irqoff(vcpu);
3859 return EXIT_FASTPATH_NONE;
3862 static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
3864 struct vcpu_svm *svm = to_svm(vcpu);
3865 unsigned long vmcb_pa = svm->current_vmcb->pa;
3867 guest_state_enter_irqoff();
3869 if (sev_es_guest(vcpu->kvm)) {
3870 __svm_sev_es_vcpu_run(vmcb_pa);
3872 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3875 * Use a single vmcb (vmcb01 because it's always valid) for
3876 * context switching guest state via VMLOAD/VMSAVE, that way
3877 * the state doesn't need to be copied between vmcb01 and
3878 * vmcb02 when switching vmcbs for nested virtualization.
3880 vmload(svm->vmcb01.pa);
3881 __svm_vcpu_run(vmcb_pa, (unsigned long *)&vcpu->arch.regs);
3882 vmsave(svm->vmcb01.pa);
3884 vmload(__sme_page_pa(sd->save_area));
3887 guest_state_exit_irqoff();
3890 static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
3892 struct vcpu_svm *svm = to_svm(vcpu);
3894 trace_kvm_entry(vcpu);
3896 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
3897 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
3898 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
3901 * Disable singlestep if we're injecting an interrupt/exception.
3902 * We don't want our modified rflags to be pushed on the stack where
3903 * we might not be able to easily reset them if we disabled NMI
3906 if (svm->nmi_singlestep && svm->vmcb->control.event_inj) {
3908 * Event injection happens before external interrupts cause a
3909 * vmexit and interrupts are disabled here, so smp_send_reschedule
3910 * is enough to force an immediate vmexit.
3912 disable_nmi_singlestep(svm);
3913 smp_send_reschedule(vcpu->cpu);
3918 sync_lapic_to_cr8(vcpu);
3920 if (unlikely(svm->asid != svm->vmcb->control.asid)) {
3921 svm->vmcb->control.asid = svm->asid;
3922 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
3924 svm->vmcb->save.cr2 = vcpu->arch.cr2;
3926 svm_hv_update_vp_id(svm->vmcb, vcpu);
3929 * Run with all-zero DR6 unless needed, so that we can get the exact cause
3932 if (unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT))
3933 svm_set_dr6(svm, vcpu->arch.dr6);
3935 svm_set_dr6(svm, DR6_ACTIVE_LOW);
3938 kvm_load_guest_xsave_state(vcpu);
3940 kvm_wait_lapic_expire(vcpu);
3943 * If this vCPU has touched SPEC_CTRL, restore the guest's value if
3944 * it's non-zero. Since vmentry is serialising on affected CPUs, there
3945 * is no need to worry about the conditional branch over the wrmsr
3946 * being speculatively taken.
3948 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3949 x86_spec_ctrl_set_guest(svm->spec_ctrl, svm->virt_spec_ctrl);
3951 svm_vcpu_enter_exit(vcpu);
3954 * We do not use IBRS in the kernel. If this vCPU has used the
3955 * SPEC_CTRL MSR it may have left it on; save the value and
3956 * turn it off. This is much more efficient than blindly adding
3957 * it to the atomic save/restore list. Especially as the former
3958 * (Saving guest MSRs on vmexit) doesn't even exist in KVM.
3960 * For non-nested case:
3961 * If the L01 MSR bitmap does not intercept the MSR, then we need to
3965 * If the L02 MSR bitmap does not intercept the MSR, then we need to
3968 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL) &&
3969 unlikely(!msr_write_intercepted(vcpu, MSR_IA32_SPEC_CTRL)))
3970 svm->spec_ctrl = native_read_msr(MSR_IA32_SPEC_CTRL);
3972 if (!sev_es_guest(vcpu->kvm))
3975 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3976 x86_spec_ctrl_restore_host(svm->spec_ctrl, svm->virt_spec_ctrl);
3978 if (!sev_es_guest(vcpu->kvm)) {
3979 vcpu->arch.cr2 = svm->vmcb->save.cr2;
3980 vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
3981 vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
3982 vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
3984 vcpu->arch.regs_dirty = 0;
3986 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3987 kvm_before_interrupt(vcpu, KVM_HANDLING_NMI);
3989 kvm_load_host_xsave_state(vcpu);
3992 /* Any pending NMI will happen here */
3994 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3995 kvm_after_interrupt(vcpu);
3997 sync_cr8_to_lapic(vcpu);
4000 if (is_guest_mode(vcpu)) {
4001 nested_sync_control_from_vmcb02(svm);
4003 /* Track VMRUNs that have made past consistency checking */
4004 if (svm->nested.nested_run_pending &&
4005 svm->vmcb->control.exit_code != SVM_EXIT_ERR)
4006 ++vcpu->stat.nested_run;
4008 svm->nested.nested_run_pending = 0;
4011 svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
4012 vmcb_mark_all_clean(svm->vmcb);
4014 /* if exit due to PF check for async PF */
4015 if (svm->vmcb->control.exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR)
4016 vcpu->arch.apf.host_apf_flags =
4017 kvm_read_and_reset_apf_flags();
4019 vcpu->arch.regs_avail &= ~SVM_REGS_LAZY_LOAD_SET;
4022 * We need to handle MC intercepts here before the vcpu has a chance to
4023 * change the physical cpu
4025 if (unlikely(svm->vmcb->control.exit_code ==
4026 SVM_EXIT_EXCP_BASE + MC_VECTOR))
4027 svm_handle_mce(vcpu);
4029 svm_complete_interrupts(vcpu);
4031 if (is_guest_mode(vcpu))
4032 return EXIT_FASTPATH_NONE;
4034 return svm_exit_handlers_fastpath(vcpu);
4037 static void svm_load_mmu_pgd(struct kvm_vcpu *vcpu, hpa_t root_hpa,
4040 struct vcpu_svm *svm = to_svm(vcpu);
4044 svm->vmcb->control.nested_cr3 = __sme_set(root_hpa);
4045 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
4047 hv_track_root_tdp(vcpu, root_hpa);
4049 cr3 = vcpu->arch.cr3;
4050 } else if (root_level >= PT64_ROOT_4LEVEL) {
4051 cr3 = __sme_set(root_hpa) | kvm_get_active_pcid(vcpu);
4053 /* PCID in the guest should be impossible with a 32-bit MMU. */
4054 WARN_ON_ONCE(kvm_get_active_pcid(vcpu));
4058 svm->vmcb->save.cr3 = cr3;
4059 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
4062 static int is_disabled(void)
4066 rdmsrl(MSR_VM_CR, vm_cr);
4067 if (vm_cr & (1 << SVM_VM_CR_SVM_DISABLE))
4074 svm_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
4077 * Patch in the VMMCALL instruction:
4079 hypercall[0] = 0x0f;
4080 hypercall[1] = 0x01;
4081 hypercall[2] = 0xd9;
4084 static int __init svm_check_processor_compat(void)
4090 * The kvm parameter can be NULL (module initialization, or invocation before
4091 * VM creation). Be sure to check the kvm parameter before using it.
4093 static bool svm_has_emulated_msr(struct kvm *kvm, u32 index)
4096 case MSR_IA32_MCG_EXT_CTL:
4097 case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:
4099 case MSR_IA32_SMBASE:
4100 /* SEV-ES guests do not support SMM, so report false */
4101 if (kvm && sev_es_guest(kvm))
4111 static u64 svm_get_mt_mask(struct kvm_vcpu *vcpu, gfn_t gfn, bool is_mmio)
4116 static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
4118 struct vcpu_svm *svm = to_svm(vcpu);
4119 struct kvm_cpuid_entry2 *best;
4120 struct kvm *kvm = vcpu->kvm;
4122 vcpu->arch.xsaves_enabled = guest_cpuid_has(vcpu, X86_FEATURE_XSAVE) &&
4123 boot_cpu_has(X86_FEATURE_XSAVE) &&
4124 boot_cpu_has(X86_FEATURE_XSAVES);
4126 /* Update nrips enabled cache */
4127 svm->nrips_enabled = kvm_cpu_cap_has(X86_FEATURE_NRIPS) &&
4128 guest_cpuid_has(vcpu, X86_FEATURE_NRIPS);
4130 svm->tsc_scaling_enabled = tsc_scaling && guest_cpuid_has(vcpu, X86_FEATURE_TSCRATEMSR);
4131 svm->lbrv_enabled = lbrv && guest_cpuid_has(vcpu, X86_FEATURE_LBRV);
4133 svm->v_vmload_vmsave_enabled = vls && guest_cpuid_has(vcpu, X86_FEATURE_V_VMSAVE_VMLOAD);
4135 svm->pause_filter_enabled = kvm_cpu_cap_has(X86_FEATURE_PAUSEFILTER) &&
4136 guest_cpuid_has(vcpu, X86_FEATURE_PAUSEFILTER);
4138 svm->pause_threshold_enabled = kvm_cpu_cap_has(X86_FEATURE_PFTHRESHOLD) &&
4139 guest_cpuid_has(vcpu, X86_FEATURE_PFTHRESHOLD);
4141 svm->vgif_enabled = vgif && guest_cpuid_has(vcpu, X86_FEATURE_VGIF);
4143 svm_recalc_instruction_intercepts(vcpu, svm);
4145 /* For sev guests, the memory encryption bit is not reserved in CR3. */
4146 if (sev_guest(vcpu->kvm)) {
4147 best = kvm_find_cpuid_entry(vcpu, 0x8000001F, 0);
4149 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
4152 if (kvm_vcpu_apicv_active(vcpu)) {
4154 * AVIC does not work with an x2APIC mode guest. If the X2APIC feature
4155 * is exposed to the guest, disable AVIC.
4157 if (guest_cpuid_has(vcpu, X86_FEATURE_X2APIC))
4158 kvm_set_apicv_inhibit(kvm, APICV_INHIBIT_REASON_X2APIC);
4160 init_vmcb_after_set_cpuid(vcpu);
4163 static bool svm_has_wbinvd_exit(void)
4168 #define PRE_EX(exit) { .exit_code = (exit), \
4169 .stage = X86_ICPT_PRE_EXCEPT, }
4170 #define POST_EX(exit) { .exit_code = (exit), \
4171 .stage = X86_ICPT_POST_EXCEPT, }
4172 #define POST_MEM(exit) { .exit_code = (exit), \
4173 .stage = X86_ICPT_POST_MEMACCESS, }
4175 static const struct __x86_intercept {
4177 enum x86_intercept_stage stage;
4178 } x86_intercept_map[] = {
4179 [x86_intercept_cr_read] = POST_EX(SVM_EXIT_READ_CR0),
4180 [x86_intercept_cr_write] = POST_EX(SVM_EXIT_WRITE_CR0),
4181 [x86_intercept_clts] = POST_EX(SVM_EXIT_WRITE_CR0),
4182 [x86_intercept_lmsw] = POST_EX(SVM_EXIT_WRITE_CR0),
4183 [x86_intercept_smsw] = POST_EX(SVM_EXIT_READ_CR0),
4184 [x86_intercept_dr_read] = POST_EX(SVM_EXIT_READ_DR0),
4185 [x86_intercept_dr_write] = POST_EX(SVM_EXIT_WRITE_DR0),
4186 [x86_intercept_sldt] = POST_EX(SVM_EXIT_LDTR_READ),
4187 [x86_intercept_str] = POST_EX(SVM_EXIT_TR_READ),
4188 [x86_intercept_lldt] = POST_EX(SVM_EXIT_LDTR_WRITE),
4189 [x86_intercept_ltr] = POST_EX(SVM_EXIT_TR_WRITE),
4190 [x86_intercept_sgdt] = POST_EX(SVM_EXIT_GDTR_READ),
4191 [x86_intercept_sidt] = POST_EX(SVM_EXIT_IDTR_READ),
4192 [x86_intercept_lgdt] = POST_EX(SVM_EXIT_GDTR_WRITE),
4193 [x86_intercept_lidt] = POST_EX(SVM_EXIT_IDTR_WRITE),
4194 [x86_intercept_vmrun] = POST_EX(SVM_EXIT_VMRUN),
4195 [x86_intercept_vmmcall] = POST_EX(SVM_EXIT_VMMCALL),
4196 [x86_intercept_vmload] = POST_EX(SVM_EXIT_VMLOAD),
4197 [x86_intercept_vmsave] = POST_EX(SVM_EXIT_VMSAVE),
4198 [x86_intercept_stgi] = POST_EX(SVM_EXIT_STGI),
4199 [x86_intercept_clgi] = POST_EX(SVM_EXIT_CLGI),
4200 [x86_intercept_skinit] = POST_EX(SVM_EXIT_SKINIT),
4201 [x86_intercept_invlpga] = POST_EX(SVM_EXIT_INVLPGA),
4202 [x86_intercept_rdtscp] = POST_EX(SVM_EXIT_RDTSCP),
4203 [x86_intercept_monitor] = POST_MEM(SVM_EXIT_MONITOR),
4204 [x86_intercept_mwait] = POST_EX(SVM_EXIT_MWAIT),
4205 [x86_intercept_invlpg] = POST_EX(SVM_EXIT_INVLPG),
4206 [x86_intercept_invd] = POST_EX(SVM_EXIT_INVD),
4207 [x86_intercept_wbinvd] = POST_EX(SVM_EXIT_WBINVD),
4208 [x86_intercept_wrmsr] = POST_EX(SVM_EXIT_MSR),
4209 [x86_intercept_rdtsc] = POST_EX(SVM_EXIT_RDTSC),
4210 [x86_intercept_rdmsr] = POST_EX(SVM_EXIT_MSR),
4211 [x86_intercept_rdpmc] = POST_EX(SVM_EXIT_RDPMC),
4212 [x86_intercept_cpuid] = PRE_EX(SVM_EXIT_CPUID),
4213 [x86_intercept_rsm] = PRE_EX(SVM_EXIT_RSM),
4214 [x86_intercept_pause] = PRE_EX(SVM_EXIT_PAUSE),
4215 [x86_intercept_pushf] = PRE_EX(SVM_EXIT_PUSHF),
4216 [x86_intercept_popf] = PRE_EX(SVM_EXIT_POPF),
4217 [x86_intercept_intn] = PRE_EX(SVM_EXIT_SWINT),
4218 [x86_intercept_iret] = PRE_EX(SVM_EXIT_IRET),
4219 [x86_intercept_icebp] = PRE_EX(SVM_EXIT_ICEBP),
4220 [x86_intercept_hlt] = POST_EX(SVM_EXIT_HLT),
4221 [x86_intercept_in] = POST_EX(SVM_EXIT_IOIO),
4222 [x86_intercept_ins] = POST_EX(SVM_EXIT_IOIO),
4223 [x86_intercept_out] = POST_EX(SVM_EXIT_IOIO),
4224 [x86_intercept_outs] = POST_EX(SVM_EXIT_IOIO),
4225 [x86_intercept_xsetbv] = PRE_EX(SVM_EXIT_XSETBV),
4232 static int svm_check_intercept(struct kvm_vcpu *vcpu,
4233 struct x86_instruction_info *info,
4234 enum x86_intercept_stage stage,
4235 struct x86_exception *exception)
4237 struct vcpu_svm *svm = to_svm(vcpu);
4238 int vmexit, ret = X86EMUL_CONTINUE;
4239 struct __x86_intercept icpt_info;
4240 struct vmcb *vmcb = svm->vmcb;
4242 if (info->intercept >= ARRAY_SIZE(x86_intercept_map))
4245 icpt_info = x86_intercept_map[info->intercept];
4247 if (stage != icpt_info.stage)
4250 switch (icpt_info.exit_code) {
4251 case SVM_EXIT_READ_CR0:
4252 if (info->intercept == x86_intercept_cr_read)
4253 icpt_info.exit_code += info->modrm_reg;
4255 case SVM_EXIT_WRITE_CR0: {
4256 unsigned long cr0, val;
4258 if (info->intercept == x86_intercept_cr_write)
4259 icpt_info.exit_code += info->modrm_reg;
4261 if (icpt_info.exit_code != SVM_EXIT_WRITE_CR0 ||
4262 info->intercept == x86_intercept_clts)
4265 if (!(vmcb12_is_intercept(&svm->nested.ctl,
4266 INTERCEPT_SELECTIVE_CR0)))
4269 cr0 = vcpu->arch.cr0 & ~SVM_CR0_SELECTIVE_MASK;
4270 val = info->src_val & ~SVM_CR0_SELECTIVE_MASK;
4272 if (info->intercept == x86_intercept_lmsw) {
4275 /* lmsw can't clear PE - catch this here */
4276 if (cr0 & X86_CR0_PE)
4281 icpt_info.exit_code = SVM_EXIT_CR0_SEL_WRITE;
4285 case SVM_EXIT_READ_DR0:
4286 case SVM_EXIT_WRITE_DR0:
4287 icpt_info.exit_code += info->modrm_reg;
4290 if (info->intercept == x86_intercept_wrmsr)
4291 vmcb->control.exit_info_1 = 1;
4293 vmcb->control.exit_info_1 = 0;
4295 case SVM_EXIT_PAUSE:
4297 * We get this for NOP only, but pause
4298 * is rep not, check this here
4300 if (info->rep_prefix != REPE_PREFIX)
4303 case SVM_EXIT_IOIO: {
4307 if (info->intercept == x86_intercept_in ||
4308 info->intercept == x86_intercept_ins) {
4309 exit_info = ((info->src_val & 0xffff) << 16) |
4311 bytes = info->dst_bytes;
4313 exit_info = (info->dst_val & 0xffff) << 16;
4314 bytes = info->src_bytes;
4317 if (info->intercept == x86_intercept_outs ||
4318 info->intercept == x86_intercept_ins)
4319 exit_info |= SVM_IOIO_STR_MASK;
4321 if (info->rep_prefix)
4322 exit_info |= SVM_IOIO_REP_MASK;
4324 bytes = min(bytes, 4u);
4326 exit_info |= bytes << SVM_IOIO_SIZE_SHIFT;
4328 exit_info |= (u32)info->ad_bytes << (SVM_IOIO_ASIZE_SHIFT - 1);
4330 vmcb->control.exit_info_1 = exit_info;
4331 vmcb->control.exit_info_2 = info->next_rip;
4339 /* TODO: Advertise NRIPS to guest hypervisor unconditionally */
4340 if (static_cpu_has(X86_FEATURE_NRIPS))
4341 vmcb->control.next_rip = info->next_rip;
4342 vmcb->control.exit_code = icpt_info.exit_code;
4343 vmexit = nested_svm_exit_handled(svm);
4345 ret = (vmexit == NESTED_EXIT_DONE) ? X86EMUL_INTERCEPTED
4352 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
4354 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_INTR)
4355 vcpu->arch.at_instruction_boundary = true;
4358 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
4360 if (!kvm_pause_in_guest(vcpu->kvm))
4361 shrink_ple_window(vcpu);
4364 static void svm_setup_mce(struct kvm_vcpu *vcpu)
4366 /* [63:9] are reserved. */
4367 vcpu->arch.mcg_cap &= 0x1ff;
4370 bool svm_smi_blocked(struct kvm_vcpu *vcpu)
4372 struct vcpu_svm *svm = to_svm(vcpu);
4374 /* Per APM Vol.2 15.22.2 "Response to SMI" */
4378 return is_smm(vcpu);
4381 static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
4383 struct vcpu_svm *svm = to_svm(vcpu);
4384 if (svm->nested.nested_run_pending)
4387 if (svm_smi_blocked(vcpu))
4390 /* An SMI must not be injected into L2 if it's supposed to VM-Exit. */
4391 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_smi(svm))
4397 static int svm_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
4399 struct vcpu_svm *svm = to_svm(vcpu);
4400 struct kvm_host_map map_save;
4403 if (!is_guest_mode(vcpu))
4406 /* FED8h - SVM Guest */
4407 put_smstate(u64, smstate, 0x7ed8, 1);
4408 /* FEE0h - SVM Guest VMCB Physical Address */
4409 put_smstate(u64, smstate, 0x7ee0, svm->nested.vmcb12_gpa);
4411 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4412 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4413 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4415 ret = nested_svm_simple_vmexit(svm, SVM_EXIT_SW);
4420 * KVM uses VMCB01 to store L1 host state while L2 runs but
4421 * VMCB01 is going to be used during SMM and thus the state will
4422 * be lost. Temporary save non-VMLOAD/VMSAVE state to the host save
4423 * area pointed to by MSR_VM_HSAVE_PA. APM guarantees that the
4424 * format of the area is identical to guest save area offsetted
4425 * by 0x400 (matches the offset of 'struct vmcb_save_area'
4426 * within 'struct vmcb'). Note: HSAVE area may also be used by
4427 * L1 hypervisor to save additional host context (e.g. KVM does
4428 * that, see svm_prepare_switch_to_guest()) which must be
4431 if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr),
4432 &map_save) == -EINVAL)
4435 BUILD_BUG_ON(offsetof(struct vmcb, save) != 0x400);
4437 svm_copy_vmrun_state(map_save.hva + 0x400,
4438 &svm->vmcb01.ptr->save);
4440 kvm_vcpu_unmap(vcpu, &map_save, true);
4444 static int svm_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
4446 struct vcpu_svm *svm = to_svm(vcpu);
4447 struct kvm_host_map map, map_save;
4448 u64 saved_efer, vmcb12_gpa;
4449 struct vmcb *vmcb12;
4452 if (!guest_cpuid_has(vcpu, X86_FEATURE_LM))
4455 /* Non-zero if SMI arrived while vCPU was in guest mode. */
4456 if (!GET_SMSTATE(u64, smstate, 0x7ed8))
4459 if (!guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4462 saved_efer = GET_SMSTATE(u64, smstate, 0x7ed0);
4463 if (!(saved_efer & EFER_SVME))
4466 vmcb12_gpa = GET_SMSTATE(u64, smstate, 0x7ee0);
4467 if (kvm_vcpu_map(vcpu, gpa_to_gfn(vmcb12_gpa), &map) == -EINVAL)
4471 if (kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.hsave_msr), &map_save) == -EINVAL)
4474 if (svm_allocate_nested(svm))
4478 * Restore L1 host state from L1 HSAVE area as VMCB01 was
4479 * used during SMM (see svm_enter_smm())
4482 svm_copy_vmrun_state(&svm->vmcb01.ptr->save, map_save.hva + 0x400);
4485 * Enter the nested guest now
4488 vmcb_mark_all_dirty(svm->vmcb01.ptr);
4491 nested_copy_vmcb_control_to_cache(svm, &vmcb12->control);
4492 nested_copy_vmcb_save_to_cache(svm, &vmcb12->save);
4493 ret = enter_svm_guest_mode(vcpu, vmcb12_gpa, vmcb12, false);
4498 svm->nested.nested_run_pending = 1;
4501 kvm_vcpu_unmap(vcpu, &map_save, true);
4503 kvm_vcpu_unmap(vcpu, &map, true);
4507 static void svm_enable_smi_window(struct kvm_vcpu *vcpu)
4509 struct vcpu_svm *svm = to_svm(vcpu);
4511 if (!gif_set(svm)) {
4513 svm_set_intercept(svm, INTERCEPT_STGI);
4514 /* STGI will cause a vm exit */
4516 /* We must be in SMM; RSM will cause a vmexit anyway. */
4520 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, int emul_type,
4521 void *insn, int insn_len)
4523 bool smep, smap, is_user;
4527 /* Emulation is always possible when KVM has access to all guest state. */
4528 if (!sev_guest(vcpu->kvm))
4531 /* #UD and #GP should never be intercepted for SEV guests. */
4532 WARN_ON_ONCE(emul_type & (EMULTYPE_TRAP_UD |
4533 EMULTYPE_TRAP_UD_FORCED |
4534 EMULTYPE_VMWARE_GP));
4537 * Emulation is impossible for SEV-ES guests as KVM doesn't have access
4538 * to guest register state.
4540 if (sev_es_guest(vcpu->kvm))
4544 * Emulation is possible if the instruction is already decoded, e.g.
4545 * when completing I/O after returning from userspace.
4547 if (emul_type & EMULTYPE_NO_DECODE)
4551 * Emulation is possible for SEV guests if and only if a prefilled
4552 * buffer containing the bytes of the intercepted instruction is
4553 * available. SEV guest memory is encrypted with a guest specific key
4554 * and cannot be decrypted by KVM, i.e. KVM would read cyphertext and
4557 * Inject #UD if KVM reached this point without an instruction buffer.
4558 * In practice, this path should never be hit by a well-behaved guest,
4559 * e.g. KVM doesn't intercept #UD or #GP for SEV guests, but this path
4560 * is still theoretically reachable, e.g. via unaccelerated fault-like
4561 * AVIC access, and needs to be handled by KVM to avoid putting the
4562 * guest into an infinite loop. Injecting #UD is somewhat arbitrary,
4563 * but its the least awful option given lack of insight into the guest.
4565 if (unlikely(!insn)) {
4566 kvm_queue_exception(vcpu, UD_VECTOR);
4571 * Emulate for SEV guests if the insn buffer is not empty. The buffer
4572 * will be empty if the DecodeAssist microcode cannot fetch bytes for
4573 * the faulting instruction because the code fetch itself faulted, e.g.
4574 * the guest attempted to fetch from emulated MMIO or a guest page
4575 * table used to translate CS:RIP resides in emulated MMIO.
4577 if (likely(insn_len))
4581 * Detect and workaround Errata 1096 Fam_17h_00_0Fh.
4584 * When CPU raises #NPF on guest data access and vCPU CR4.SMAP=1, it is
4585 * possible that CPU microcode implementing DecodeAssist will fail to
4586 * read guest memory at CS:RIP and vmcb.GuestIntrBytes will incorrectly
4587 * be '0'. This happens because microcode reads CS:RIP using a _data_
4588 * loap uop with CPL=0 privileges. If the load hits a SMAP #PF, ucode
4589 * gives up and does not fill the instruction bytes buffer.
4591 * As above, KVM reaches this point iff the VM is an SEV guest, the CPU
4592 * supports DecodeAssist, a #NPF was raised, KVM's page fault handler
4593 * triggered emulation (e.g. for MMIO), and the CPU returned 0 in the
4594 * GuestIntrBytes field of the VMCB.
4596 * This does _not_ mean that the erratum has been encountered, as the
4597 * DecodeAssist will also fail if the load for CS:RIP hits a legitimate
4598 * #PF, e.g. if the guest attempt to execute from emulated MMIO and
4599 * encountered a reserved/not-present #PF.
4601 * To hit the erratum, the following conditions must be true:
4602 * 1. CR4.SMAP=1 (obviously).
4603 * 2. CR4.SMEP=0 || CPL=3. If SMEP=1 and CPL<3, the erratum cannot
4604 * have been hit as the guest would have encountered a SMEP
4605 * violation #PF, not a #NPF.
4606 * 3. The #NPF is not due to a code fetch, in which case failure to
4607 * retrieve the instruction bytes is legitimate (see abvoe).
4609 * In addition, don't apply the erratum workaround if the #NPF occurred
4610 * while translating guest page tables (see below).
4612 error_code = to_svm(vcpu)->vmcb->control.exit_info_1;
4613 if (error_code & (PFERR_GUEST_PAGE_MASK | PFERR_FETCH_MASK))
4616 cr4 = kvm_read_cr4(vcpu);
4617 smep = cr4 & X86_CR4_SMEP;
4618 smap = cr4 & X86_CR4_SMAP;
4619 is_user = svm_get_cpl(vcpu) == 3;
4620 if (smap && (!smep || is_user)) {
4621 pr_err_ratelimited("KVM: SEV Guest triggered AMD Erratum 1096\n");
4624 * If the fault occurred in userspace, arbitrarily inject #GP
4625 * to avoid killing the guest and to hopefully avoid confusing
4626 * the guest kernel too much, e.g. injecting #PF would not be
4627 * coherent with respect to the guest's page tables. Request
4628 * triple fault if the fault occurred in the kernel as there's
4629 * no fault that KVM can inject without confusing the guest.
4630 * In practice, the triple fault is moot as no sane SEV kernel
4631 * will execute from user memory while also running with SMAP=1.
4634 kvm_inject_gp(vcpu, 0);
4636 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
4641 * If the erratum was not hit, simply resume the guest and let it fault
4642 * again. While awful, e.g. the vCPU may get stuck in an infinite loop
4643 * if the fault is at CPL=0, it's the lesser of all evils. Exiting to
4644 * userspace will kill the guest, and letting the emulator read garbage
4645 * will yield random behavior and potentially corrupt the guest.
4647 * Simply resuming the guest is technically not a violation of the SEV
4648 * architecture. AMD's APM states that all code fetches and page table
4649 * accesses for SEV guest are encrypted, regardless of the C-Bit. The
4650 * APM also states that encrypted accesses to MMIO are "ignored", but
4651 * doesn't explicitly define "ignored", i.e. doing nothing and letting
4652 * the guest spin is technically "ignoring" the access.
4657 static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
4659 struct vcpu_svm *svm = to_svm(vcpu);
4662 * TODO: Last condition latch INIT signals on vCPU when
4663 * vCPU is in guest-mode and vmcb12 defines intercept on INIT.
4664 * To properly emulate the INIT intercept,
4665 * svm_check_nested_events() should call nested_svm_vmexit()
4666 * if an INIT signal is pending.
4668 return !gif_set(svm) ||
4669 (vmcb_is_intercept(&svm->vmcb->control, INTERCEPT_INIT));
4672 static void svm_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
4674 if (!sev_es_guest(vcpu->kvm))
4675 return kvm_vcpu_deliver_sipi_vector(vcpu, vector);
4677 sev_vcpu_deliver_sipi_vector(vcpu, vector);
4680 static void svm_vm_destroy(struct kvm *kvm)
4682 avic_vm_destroy(kvm);
4683 sev_vm_destroy(kvm);
4686 static int svm_vm_init(struct kvm *kvm)
4688 if (!pause_filter_count || !pause_filter_thresh)
4689 kvm->arch.pause_in_guest = true;
4692 int ret = avic_vm_init(kvm);
4700 static struct kvm_x86_ops svm_x86_ops __initdata = {
4703 .hardware_unsetup = svm_hardware_unsetup,
4704 .hardware_enable = svm_hardware_enable,
4705 .hardware_disable = svm_hardware_disable,
4706 .has_emulated_msr = svm_has_emulated_msr,
4708 .vcpu_create = svm_vcpu_create,
4709 .vcpu_free = svm_vcpu_free,
4710 .vcpu_reset = svm_vcpu_reset,
4712 .vm_size = sizeof(struct kvm_svm),
4713 .vm_init = svm_vm_init,
4714 .vm_destroy = svm_vm_destroy,
4716 .prepare_switch_to_guest = svm_prepare_switch_to_guest,
4717 .vcpu_load = svm_vcpu_load,
4718 .vcpu_put = svm_vcpu_put,
4719 .vcpu_blocking = avic_vcpu_blocking,
4720 .vcpu_unblocking = avic_vcpu_unblocking,
4722 .update_exception_bitmap = svm_update_exception_bitmap,
4723 .get_msr_feature = svm_get_msr_feature,
4724 .get_msr = svm_get_msr,
4725 .set_msr = svm_set_msr,
4726 .get_segment_base = svm_get_segment_base,
4727 .get_segment = svm_get_segment,
4728 .set_segment = svm_set_segment,
4729 .get_cpl = svm_get_cpl,
4730 .get_cs_db_l_bits = svm_get_cs_db_l_bits,
4731 .set_cr0 = svm_set_cr0,
4732 .post_set_cr3 = sev_post_set_cr3,
4733 .is_valid_cr4 = svm_is_valid_cr4,
4734 .set_cr4 = svm_set_cr4,
4735 .set_efer = svm_set_efer,
4736 .get_idt = svm_get_idt,
4737 .set_idt = svm_set_idt,
4738 .get_gdt = svm_get_gdt,
4739 .set_gdt = svm_set_gdt,
4740 .set_dr7 = svm_set_dr7,
4741 .sync_dirty_debug_regs = svm_sync_dirty_debug_regs,
4742 .cache_reg = svm_cache_reg,
4743 .get_rflags = svm_get_rflags,
4744 .set_rflags = svm_set_rflags,
4745 .get_if_flag = svm_get_if_flag,
4747 .flush_tlb_all = svm_flush_tlb_current,
4748 .flush_tlb_current = svm_flush_tlb_current,
4749 .flush_tlb_gva = svm_flush_tlb_gva,
4750 .flush_tlb_guest = svm_flush_tlb_current,
4752 .vcpu_pre_run = svm_vcpu_pre_run,
4753 .vcpu_run = svm_vcpu_run,
4754 .handle_exit = svm_handle_exit,
4755 .skip_emulated_instruction = svm_skip_emulated_instruction,
4756 .update_emulated_instruction = NULL,
4757 .set_interrupt_shadow = svm_set_interrupt_shadow,
4758 .get_interrupt_shadow = svm_get_interrupt_shadow,
4759 .patch_hypercall = svm_patch_hypercall,
4760 .inject_irq = svm_inject_irq,
4761 .inject_nmi = svm_inject_nmi,
4762 .queue_exception = svm_queue_exception,
4763 .cancel_injection = svm_cancel_injection,
4764 .interrupt_allowed = svm_interrupt_allowed,
4765 .nmi_allowed = svm_nmi_allowed,
4766 .get_nmi_mask = svm_get_nmi_mask,
4767 .set_nmi_mask = svm_set_nmi_mask,
4768 .enable_nmi_window = svm_enable_nmi_window,
4769 .enable_irq_window = svm_enable_irq_window,
4770 .update_cr8_intercept = svm_update_cr8_intercept,
4771 .refresh_apicv_exec_ctrl = avic_refresh_apicv_exec_ctrl,
4772 .check_apicv_inhibit_reasons = avic_check_apicv_inhibit_reasons,
4773 .apicv_post_state_restore = avic_apicv_post_state_restore,
4775 .get_mt_mask = svm_get_mt_mask,
4776 .get_exit_info = svm_get_exit_info,
4778 .vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
4780 .has_wbinvd_exit = svm_has_wbinvd_exit,
4782 .get_l2_tsc_offset = svm_get_l2_tsc_offset,
4783 .get_l2_tsc_multiplier = svm_get_l2_tsc_multiplier,
4784 .write_tsc_offset = svm_write_tsc_offset,
4785 .write_tsc_multiplier = svm_write_tsc_multiplier,
4787 .load_mmu_pgd = svm_load_mmu_pgd,
4789 .check_intercept = svm_check_intercept,
4790 .handle_exit_irqoff = svm_handle_exit_irqoff,
4792 .request_immediate_exit = __kvm_request_immediate_exit,
4794 .sched_in = svm_sched_in,
4796 .nested_ops = &svm_nested_ops,
4798 .deliver_interrupt = svm_deliver_interrupt,
4799 .pi_update_irte = avic_pi_update_irte,
4800 .setup_mce = svm_setup_mce,
4802 .smi_allowed = svm_smi_allowed,
4803 .enter_smm = svm_enter_smm,
4804 .leave_smm = svm_leave_smm,
4805 .enable_smi_window = svm_enable_smi_window,
4807 .mem_enc_ioctl = sev_mem_enc_ioctl,
4808 .mem_enc_register_region = sev_mem_enc_register_region,
4809 .mem_enc_unregister_region = sev_mem_enc_unregister_region,
4810 .guest_memory_reclaimed = sev_guest_memory_reclaimed,
4812 .vm_copy_enc_context_from = sev_vm_copy_enc_context_from,
4813 .vm_move_enc_context_from = sev_vm_move_enc_context_from,
4815 .can_emulate_instruction = svm_can_emulate_instruction,
4817 .apic_init_signal_blocked = svm_apic_init_signal_blocked,
4819 .msr_filter_changed = svm_msr_filter_changed,
4820 .complete_emulated_msr = svm_complete_emulated_msr,
4822 .vcpu_deliver_sipi_vector = svm_vcpu_deliver_sipi_vector,
4823 .vcpu_get_apicv_inhibit_reasons = avic_vcpu_get_apicv_inhibit_reasons,
4827 * The default MMIO mask is a single bit (excluding the present bit),
4828 * which could conflict with the memory encryption bit. Check for
4829 * memory encryption support and override the default MMIO mask if
4830 * memory encryption is enabled.
4832 static __init void svm_adjust_mmio_mask(void)
4834 unsigned int enc_bit, mask_bit;
4837 /* If there is no memory encryption support, use existing mask */
4838 if (cpuid_eax(0x80000000) < 0x8000001f)
4841 /* If memory encryption is not enabled, use existing mask */
4842 rdmsrl(MSR_AMD64_SYSCFG, msr);
4843 if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
4846 enc_bit = cpuid_ebx(0x8000001f) & 0x3f;
4847 mask_bit = boot_cpu_data.x86_phys_bits;
4849 /* Increment the mask bit if it is the same as the encryption bit */
4850 if (enc_bit == mask_bit)
4854 * If the mask bit location is below 52, then some bits above the
4855 * physical addressing limit will always be reserved, so use the
4856 * rsvd_bits() function to generate the mask. This mask, along with
4857 * the present bit, will be used to generate a page fault with
4860 * If the mask bit location is 52 (or above), then clear the mask.
4862 mask = (mask_bit < 52) ? rsvd_bits(mask_bit, 51) | PT_PRESENT_MASK : 0;
4864 kvm_mmu_set_mmio_spte_mask(mask, mask, PT_WRITABLE_MASK | PT_USER_MASK);
4867 static __init void svm_set_cpu_caps(void)
4871 kvm_caps.supported_xss = 0;
4873 /* CPUID 0x80000001 and 0x8000000A (SVM features) */
4875 kvm_cpu_cap_set(X86_FEATURE_SVM);
4876 kvm_cpu_cap_set(X86_FEATURE_VMCBCLEAN);
4879 kvm_cpu_cap_set(X86_FEATURE_NRIPS);
4882 kvm_cpu_cap_set(X86_FEATURE_NPT);
4885 kvm_cpu_cap_set(X86_FEATURE_TSCRATEMSR);
4888 kvm_cpu_cap_set(X86_FEATURE_V_VMSAVE_VMLOAD);
4890 kvm_cpu_cap_set(X86_FEATURE_LBRV);
4892 if (boot_cpu_has(X86_FEATURE_PAUSEFILTER))
4893 kvm_cpu_cap_set(X86_FEATURE_PAUSEFILTER);
4895 if (boot_cpu_has(X86_FEATURE_PFTHRESHOLD))
4896 kvm_cpu_cap_set(X86_FEATURE_PFTHRESHOLD);
4899 kvm_cpu_cap_set(X86_FEATURE_VGIF);
4901 /* Nested VM can receive #VMEXIT instead of triggering #GP */
4902 kvm_cpu_cap_set(X86_FEATURE_SVME_ADDR_CHK);
4905 /* CPUID 0x80000008 */
4906 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) ||
4907 boot_cpu_has(X86_FEATURE_AMD_SSBD))
4908 kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
4910 /* AMD PMU PERFCTR_CORE CPUID */
4911 if (enable_pmu && boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
4912 kvm_cpu_cap_set(X86_FEATURE_PERFCTR_CORE);
4914 /* CPUID 0x8000001F (SME/SEV features) */
4918 static __init int svm_hardware_setup(void)
4921 struct page *iopm_pages;
4924 unsigned int order = get_order(IOPM_SIZE);
4927 * NX is required for shadow paging and for NPT if the NX huge pages
4928 * mitigation is enabled.
4930 if (!boot_cpu_has(X86_FEATURE_NX)) {
4931 pr_err_ratelimited("NX (Execute Disable) not supported\n");
4934 kvm_enable_efer_bits(EFER_NX);
4936 iopm_pages = alloc_pages(GFP_KERNEL, order);
4941 iopm_va = page_address(iopm_pages);
4942 memset(iopm_va, 0xff, PAGE_SIZE * (1 << order));
4943 iopm_base = page_to_pfn(iopm_pages) << PAGE_SHIFT;
4945 init_msrpm_offsets();
4947 kvm_caps.supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS |
4948 XFEATURE_MASK_BNDCSR);
4950 if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
4951 kvm_enable_efer_bits(EFER_FFXSR);
4954 if (!boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
4955 tsc_scaling = false;
4957 pr_info("TSC scaling supported\n");
4958 kvm_caps.has_tsc_control = true;
4961 kvm_caps.max_tsc_scaling_ratio = SVM_TSC_RATIO_MAX;
4962 kvm_caps.tsc_scaling_ratio_frac_bits = 32;
4964 tsc_aux_uret_slot = kvm_add_user_return_msr(MSR_TSC_AUX);
4966 /* Check for pause filtering support */
4967 if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
4968 pause_filter_count = 0;
4969 pause_filter_thresh = 0;
4970 } else if (!boot_cpu_has(X86_FEATURE_PFTHRESHOLD)) {
4971 pause_filter_thresh = 0;
4975 printk(KERN_INFO "kvm: Nested Virtualization enabled\n");
4976 kvm_enable_efer_bits(EFER_SVME | EFER_LMSLE);
4980 * KVM's MMU doesn't support using 2-level paging for itself, and thus
4981 * NPT isn't supported if the host is using 2-level paging since host
4982 * CR4 is unchanged on VMRUN.
4984 if (!IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_X86_PAE))
4985 npt_enabled = false;
4987 if (!boot_cpu_has(X86_FEATURE_NPT))
4988 npt_enabled = false;
4990 /* Force VM NPT level equal to the host's paging level */
4991 kvm_configure_mmu(npt_enabled, get_npt_level(),
4992 get_npt_level(), PG_LEVEL_1G);
4993 pr_info("kvm: Nested Paging %sabled\n", npt_enabled ? "en" : "dis");
4995 /* Setup shadow_me_value and shadow_me_mask */
4996 kvm_mmu_set_me_spte_mask(sme_me_mask, sme_me_mask);
4998 /* Note, SEV setup consumes npt_enabled. */
4999 sev_hardware_setup();
5001 svm_hv_hardware_setup();
5003 svm_adjust_mmio_mask();
5005 for_each_possible_cpu(cpu) {
5006 r = svm_cpu_init(cpu);
5012 if (!boot_cpu_has(X86_FEATURE_NRIPS))
5016 enable_apicv = avic = avic && avic_hardware_setup(&svm_x86_ops);
5018 if (!enable_apicv) {
5019 svm_x86_ops.vcpu_blocking = NULL;
5020 svm_x86_ops.vcpu_unblocking = NULL;
5021 svm_x86_ops.vcpu_get_apicv_inhibit_reasons = NULL;
5026 !boot_cpu_has(X86_FEATURE_V_VMSAVE_VMLOAD) ||
5027 !IS_ENABLED(CONFIG_X86_64)) {
5030 pr_info("Virtual VMLOAD VMSAVE supported\n");
5034 if (boot_cpu_has(X86_FEATURE_SVME_ADDR_CHK))
5035 svm_gp_erratum_intercept = false;
5038 if (!boot_cpu_has(X86_FEATURE_VGIF))
5041 pr_info("Virtual GIF supported\n");
5045 if (!boot_cpu_has(X86_FEATURE_LBRV))
5048 pr_info("LBR virtualization supported\n");
5052 pr_info("PMU virtualization is disabled\n");
5057 * It seems that on AMD processors PTE's accessed bit is
5058 * being set by the CPU hardware before the NPF vmexit.
5059 * This is not expected behaviour and our tests fail because
5061 * A workaround here is to disable support for
5062 * GUEST_MAXPHYADDR < HOST_MAXPHYADDR if NPT is enabled.
5063 * In this case userspace can know if there is support using
5064 * KVM_CAP_SMALLER_MAXPHYADDR extension and decide how to handle
5066 * If future AMD CPU models change the behaviour described above,
5067 * this variable can be changed accordingly
5069 allow_smaller_maxphyaddr = !npt_enabled;
5074 svm_hardware_unsetup();
5079 static struct kvm_x86_init_ops svm_init_ops __initdata = {
5080 .cpu_has_kvm_support = has_svm,
5081 .disabled_by_bios = is_disabled,
5082 .hardware_setup = svm_hardware_setup,
5083 .check_processor_compatibility = svm_check_processor_compat,
5085 .runtime_ops = &svm_x86_ops,
5086 .pmu_ops = &amd_pmu_ops,
5089 static int __init svm_init(void)
5091 __unused_size_checks();
5093 return kvm_init(&svm_init_ops, sizeof(struct vcpu_svm),
5094 __alignof__(struct vcpu_svm), THIS_MODULE);
5097 static void __exit svm_exit(void)
5102 module_init(svm_init)
5103 module_exit(svm_exit)