1 #define pr_fmt(fmt) "SVM: " fmt
3 #include <linux/kvm_host.h>
7 #include "kvm_cache_regs.h"
12 #include <linux/module.h>
13 #include <linux/mod_devicetable.h>
14 #include <linux/kernel.h>
15 #include <linux/vmalloc.h>
16 #include <linux/highmem.h>
17 #include <linux/amd-iommu.h>
18 #include <linux/sched.h>
19 #include <linux/trace_events.h>
20 #include <linux/slab.h>
21 #include <linux/hashtable.h>
22 #include <linux/objtool.h>
23 #include <linux/psp-sev.h>
24 #include <linux/file.h>
25 #include <linux/pagemap.h>
26 #include <linux/swap.h>
27 #include <linux/rwsem.h>
30 #include <asm/perf_event.h>
31 #include <asm/tlbflush.h>
33 #include <asm/debugreg.h>
34 #include <asm/kvm_para.h>
35 #include <asm/irq_remapping.h>
36 #include <asm/spec-ctrl.h>
37 #include <asm/cpu_device_id.h>
38 #include <asm/traps.h>
40 #include <asm/virtext.h>
45 #define __ex(x) __kvm_handle_fault_on_reboot(x)
47 MODULE_AUTHOR("Qumranet");
48 MODULE_LICENSE("GPL");
51 static const struct x86_cpu_id svm_cpu_id[] = {
52 X86_MATCH_FEATURE(X86_FEATURE_SVM, NULL),
55 MODULE_DEVICE_TABLE(x86cpu, svm_cpu_id);
58 #define IOPM_ALLOC_ORDER 2
59 #define MSRPM_ALLOC_ORDER 1
61 #define SEG_TYPE_LDT 2
62 #define SEG_TYPE_BUSY_TSS16 3
64 #define SVM_FEATURE_LBRV (1 << 1)
65 #define SVM_FEATURE_SVML (1 << 2)
66 #define SVM_FEATURE_TSC_RATE (1 << 4)
67 #define SVM_FEATURE_VMCB_CLEAN (1 << 5)
68 #define SVM_FEATURE_FLUSH_ASID (1 << 6)
69 #define SVM_FEATURE_DECODE_ASSIST (1 << 7)
70 #define SVM_FEATURE_PAUSE_FILTER (1 << 10)
72 #define DEBUGCTL_RESERVED_BITS (~(0x3fULL))
74 #define TSC_RATIO_RSVD 0xffffff0000000000ULL
75 #define TSC_RATIO_MIN 0x0000000000000001ULL
76 #define TSC_RATIO_MAX 0x000000ffffffffffULL
78 static bool erratum_383_found __read_mostly;
80 u32 msrpm_offsets[MSRPM_OFFSETS] __read_mostly;
83 * Set osvw_len to higher value when updated Revision Guides
84 * are published and we know what the new status bits are
86 static uint64_t osvw_len = 4, osvw_status;
88 static DEFINE_PER_CPU(u64, current_tsc_ratio);
89 #define TSC_RATIO_DEFAULT 0x0100000000ULL
91 static const struct svm_direct_access_msrs {
92 u32 index; /* Index of the MSR */
93 bool always; /* True if intercept is always on */
94 } direct_access_msrs[MAX_DIRECT_ACCESS_MSRS] = {
95 { .index = MSR_STAR, .always = true },
96 { .index = MSR_IA32_SYSENTER_CS, .always = true },
98 { .index = MSR_GS_BASE, .always = true },
99 { .index = MSR_FS_BASE, .always = true },
100 { .index = MSR_KERNEL_GS_BASE, .always = true },
101 { .index = MSR_LSTAR, .always = true },
102 { .index = MSR_CSTAR, .always = true },
103 { .index = MSR_SYSCALL_MASK, .always = true },
105 { .index = MSR_IA32_SPEC_CTRL, .always = false },
106 { .index = MSR_IA32_PRED_CMD, .always = false },
107 { .index = MSR_IA32_LASTBRANCHFROMIP, .always = false },
108 { .index = MSR_IA32_LASTBRANCHTOIP, .always = false },
109 { .index = MSR_IA32_LASTINTFROMIP, .always = false },
110 { .index = MSR_IA32_LASTINTTOIP, .always = false },
111 { .index = MSR_INVALID, .always = false },
114 /* enable NPT for AMD64 and X86 with PAE */
115 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
116 bool npt_enabled = true;
122 * These 2 parameters are used to config the controls for Pause-Loop Exiting:
123 * pause_filter_count: On processors that support Pause filtering(indicated
124 * by CPUID Fn8000_000A_EDX), the VMCB provides a 16 bit pause filter
125 * count value. On VMRUN this value is loaded into an internal counter.
126 * Each time a pause instruction is executed, this counter is decremented
127 * until it reaches zero at which time a #VMEXIT is generated if pause
128 * intercept is enabled. Refer to AMD APM Vol 2 Section 15.14.4 Pause
129 * Intercept Filtering for more details.
130 * This also indicate if ple logic enabled.
132 * pause_filter_thresh: In addition, some processor families support advanced
133 * pause filtering (indicated by CPUID Fn8000_000A_EDX) upper bound on
134 * the amount of time a guest is allowed to execute in a pause loop.
135 * In this mode, a 16-bit pause filter threshold field is added in the
136 * VMCB. The threshold value is a cycle count that is used to reset the
137 * pause counter. As with simple pause filtering, VMRUN loads the pause
138 * count value from VMCB into an internal counter. Then, on each pause
139 * instruction the hardware checks the elapsed number of cycles since
140 * the most recent pause instruction against the pause filter threshold.
141 * If the elapsed cycle count is greater than the pause filter threshold,
142 * then the internal pause count is reloaded from the VMCB and execution
143 * continues. If the elapsed cycle count is less than the pause filter
144 * threshold, then the internal pause count is decremented. If the count
145 * value is less than zero and PAUSE intercept is enabled, a #VMEXIT is
146 * triggered. If advanced pause filtering is supported and pause filter
147 * threshold field is set to zero, the filter will operate in the simpler,
151 static unsigned short pause_filter_thresh = KVM_DEFAULT_PLE_GAP;
152 module_param(pause_filter_thresh, ushort, 0444);
154 static unsigned short pause_filter_count = KVM_SVM_DEFAULT_PLE_WINDOW;
155 module_param(pause_filter_count, ushort, 0444);
157 /* Default doubles per-vcpu window every exit. */
158 static unsigned short pause_filter_count_grow = KVM_DEFAULT_PLE_WINDOW_GROW;
159 module_param(pause_filter_count_grow, ushort, 0444);
161 /* Default resets per-vcpu window every exit to pause_filter_count. */
162 static unsigned short pause_filter_count_shrink = KVM_DEFAULT_PLE_WINDOW_SHRINK;
163 module_param(pause_filter_count_shrink, ushort, 0444);
165 /* Default is to compute the maximum so we can never overflow. */
166 static unsigned short pause_filter_count_max = KVM_SVM_DEFAULT_PLE_WINDOW_MAX;
167 module_param(pause_filter_count_max, ushort, 0444);
169 /* allow nested paging (virtualized MMU) for all guests */
170 static int npt = true;
171 module_param(npt, int, S_IRUGO);
173 /* allow nested virtualization in KVM/SVM */
174 static int nested = true;
175 module_param(nested, int, S_IRUGO);
177 /* enable/disable Next RIP Save */
178 static int nrips = true;
179 module_param(nrips, int, 0444);
181 /* enable/disable Virtual VMLOAD VMSAVE */
182 static int vls = true;
183 module_param(vls, int, 0444);
185 /* enable/disable Virtual GIF */
186 static int vgif = true;
187 module_param(vgif, int, 0444);
189 /* enable/disable SEV support */
190 int sev = IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT_ACTIVE_BY_DEFAULT);
191 module_param(sev, int, 0444);
193 /* enable/disable SEV-ES support */
194 int sev_es = IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT_ACTIVE_BY_DEFAULT);
195 module_param(sev_es, int, 0444);
197 bool __read_mostly dump_invalid_vmcb;
198 module_param(dump_invalid_vmcb, bool, 0644);
200 static u8 rsm_ins_bytes[] = "\x0f\xaa";
202 static void svm_complete_interrupts(struct vcpu_svm *svm);
204 static unsigned long iopm_base;
206 struct kvm_ldttss_desc {
209 unsigned base1:8, type:5, dpl:2, p:1;
210 unsigned limit1:4, zero0:3, g:1, base2:8;
213 } __attribute__((packed));
215 DEFINE_PER_CPU(struct svm_cpu_data *, svm_data);
217 static const u32 msrpm_ranges[] = {0, 0xc0000000, 0xc0010000};
219 #define NUM_MSR_MAPS ARRAY_SIZE(msrpm_ranges)
220 #define MSRS_RANGE_SIZE 2048
221 #define MSRS_IN_RANGE (MSRS_RANGE_SIZE * 8 / 2)
223 u32 svm_msrpm_offset(u32 msr)
228 for (i = 0; i < NUM_MSR_MAPS; i++) {
229 if (msr < msrpm_ranges[i] ||
230 msr >= msrpm_ranges[i] + MSRS_IN_RANGE)
233 offset = (msr - msrpm_ranges[i]) / 4; /* 4 msrs per u8 */
234 offset += (i * MSRS_RANGE_SIZE); /* add range offset */
236 /* Now we have the u8 offset - but need the u32 offset */
240 /* MSR not in any range */
244 #define MAX_INST_SIZE 15
246 static inline void clgi(void)
248 asm volatile (__ex("clgi"));
251 static inline void stgi(void)
253 asm volatile (__ex("stgi"));
256 static inline void invlpga(unsigned long addr, u32 asid)
258 asm volatile (__ex("invlpga %1, %0") : : "c"(asid), "a"(addr));
261 static int get_max_npt_level(void)
264 return PT64_ROOT_4LEVEL;
266 return PT32E_ROOT_LEVEL;
270 int svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)
272 struct vcpu_svm *svm = to_svm(vcpu);
273 u64 old_efer = vcpu->arch.efer;
274 vcpu->arch.efer = efer;
277 /* Shadow paging assumes NX to be available. */
280 if (!(efer & EFER_LMA))
284 if ((old_efer & EFER_SVME) != (efer & EFER_SVME)) {
285 if (!(efer & EFER_SVME)) {
286 svm_leave_nested(svm);
287 svm_set_gif(svm, true);
290 * Free the nested guest state, unless we are in SMM.
291 * In this case we will return to the nested guest
292 * as soon as we leave SMM.
294 if (!is_smm(&svm->vcpu))
295 svm_free_nested(svm);
298 int ret = svm_allocate_nested(svm);
301 vcpu->arch.efer = old_efer;
307 svm->vmcb->save.efer = efer | EFER_SVME;
308 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
312 static int is_external_interrupt(u32 info)
314 info &= SVM_EVTINJ_TYPE_MASK | SVM_EVTINJ_VALID;
315 return info == (SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR);
318 static u32 svm_get_interrupt_shadow(struct kvm_vcpu *vcpu)
320 struct vcpu_svm *svm = to_svm(vcpu);
323 if (svm->vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK)
324 ret = KVM_X86_SHADOW_INT_STI | KVM_X86_SHADOW_INT_MOV_SS;
328 static void svm_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
330 struct vcpu_svm *svm = to_svm(vcpu);
333 svm->vmcb->control.int_state &= ~SVM_INTERRUPT_SHADOW_MASK;
335 svm->vmcb->control.int_state |= SVM_INTERRUPT_SHADOW_MASK;
339 static int skip_emulated_instruction(struct kvm_vcpu *vcpu)
341 struct vcpu_svm *svm = to_svm(vcpu);
344 * SEV-ES does not expose the next RIP. The RIP update is controlled by
345 * the type of exit and the #VC handler in the guest.
347 if (sev_es_guest(vcpu->kvm))
350 if (nrips && svm->vmcb->control.next_rip != 0) {
351 WARN_ON_ONCE(!static_cpu_has(X86_FEATURE_NRIPS));
352 svm->next_rip = svm->vmcb->control.next_rip;
355 if (!svm->next_rip) {
356 if (!kvm_emulate_instruction(vcpu, EMULTYPE_SKIP))
359 kvm_rip_write(vcpu, svm->next_rip);
363 svm_set_interrupt_shadow(vcpu, 0);
368 static void svm_queue_exception(struct kvm_vcpu *vcpu)
370 struct vcpu_svm *svm = to_svm(vcpu);
371 unsigned nr = vcpu->arch.exception.nr;
372 bool has_error_code = vcpu->arch.exception.has_error_code;
373 u32 error_code = vcpu->arch.exception.error_code;
375 kvm_deliver_exception_payload(&svm->vcpu);
377 if (nr == BP_VECTOR && !nrips) {
378 unsigned long rip, old_rip = kvm_rip_read(&svm->vcpu);
381 * For guest debugging where we have to reinject #BP if some
382 * INT3 is guest-owned:
383 * Emulate nRIP by moving RIP forward. Will fail if injection
384 * raises a fault that is not intercepted. Still better than
385 * failing in all cases.
387 (void)skip_emulated_instruction(&svm->vcpu);
388 rip = kvm_rip_read(&svm->vcpu);
389 svm->int3_rip = rip + svm->vmcb->save.cs.base;
390 svm->int3_injected = rip - old_rip;
393 svm->vmcb->control.event_inj = nr
395 | (has_error_code ? SVM_EVTINJ_VALID_ERR : 0)
396 | SVM_EVTINJ_TYPE_EXEPT;
397 svm->vmcb->control.event_inj_err = error_code;
400 static void svm_init_erratum_383(void)
406 if (!static_cpu_has_bug(X86_BUG_AMD_TLB_MMATCH))
409 /* Use _safe variants to not break nested virtualization */
410 val = native_read_msr_safe(MSR_AMD64_DC_CFG, &err);
416 low = lower_32_bits(val);
417 high = upper_32_bits(val);
419 native_write_msr_safe(MSR_AMD64_DC_CFG, low, high);
421 erratum_383_found = true;
424 static void svm_init_osvw(struct kvm_vcpu *vcpu)
427 * Guests should see errata 400 and 415 as fixed (assuming that
428 * HLT and IO instructions are intercepted).
430 vcpu->arch.osvw.length = (osvw_len >= 3) ? (osvw_len) : 3;
431 vcpu->arch.osvw.status = osvw_status & ~(6ULL);
434 * By increasing VCPU's osvw.length to 3 we are telling the guest that
435 * all osvw.status bits inside that length, including bit 0 (which is
436 * reserved for erratum 298), are valid. However, if host processor's
437 * osvw_len is 0 then osvw_status[0] carries no information. We need to
438 * be conservative here and therefore we tell the guest that erratum 298
439 * is present (because we really don't know).
441 if (osvw_len == 0 && boot_cpu_data.x86 == 0x10)
442 vcpu->arch.osvw.status |= 1;
445 static int has_svm(void)
449 if (!cpu_has_svm(&msg)) {
450 printk(KERN_INFO "has_svm: %s\n", msg);
457 static void svm_hardware_disable(void)
459 /* Make sure we clean up behind us */
460 if (static_cpu_has(X86_FEATURE_TSCRATEMSR))
461 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
465 amd_pmu_disable_virt();
468 static int svm_hardware_enable(void)
471 struct svm_cpu_data *sd;
473 struct desc_struct *gdt;
474 int me = raw_smp_processor_id();
476 rdmsrl(MSR_EFER, efer);
477 if (efer & EFER_SVME)
481 pr_err("%s: err EOPNOTSUPP on %d\n", __func__, me);
484 sd = per_cpu(svm_data, me);
486 pr_err("%s: svm_data is NULL on %d\n", __func__, me);
490 sd->asid_generation = 1;
491 sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
492 sd->next_asid = sd->max_asid + 1;
493 sd->min_asid = max_sev_asid + 1;
495 gdt = get_current_gdt_rw();
496 sd->tss_desc = (struct kvm_ldttss_desc *)(gdt + GDT_ENTRY_TSS);
498 wrmsrl(MSR_EFER, efer | EFER_SVME);
500 wrmsrl(MSR_VM_HSAVE_PA, __sme_page_pa(sd->save_area));
502 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
503 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
504 __this_cpu_write(current_tsc_ratio, TSC_RATIO_DEFAULT);
511 * Note that it is possible to have a system with mixed processor
512 * revisions and therefore different OSVW bits. If bits are not the same
513 * on different processors then choose the worst case (i.e. if erratum
514 * is present on one processor and not on another then assume that the
515 * erratum is present everywhere).
517 if (cpu_has(&boot_cpu_data, X86_FEATURE_OSVW)) {
518 uint64_t len, status = 0;
521 len = native_read_msr_safe(MSR_AMD64_OSVW_ID_LENGTH, &err);
523 status = native_read_msr_safe(MSR_AMD64_OSVW_STATUS,
527 osvw_status = osvw_len = 0;
531 osvw_status |= status;
532 osvw_status &= (1ULL << osvw_len) - 1;
535 osvw_status = osvw_len = 0;
537 svm_init_erratum_383();
539 amd_pmu_enable_virt();
544 static void svm_cpu_uninit(int cpu)
546 struct svm_cpu_data *sd = per_cpu(svm_data, raw_smp_processor_id());
551 per_cpu(svm_data, raw_smp_processor_id()) = NULL;
552 kfree(sd->sev_vmcbs);
553 __free_page(sd->save_area);
557 static int svm_cpu_init(int cpu)
559 struct svm_cpu_data *sd;
561 sd = kzalloc(sizeof(struct svm_cpu_data), GFP_KERNEL);
565 sd->save_area = alloc_page(GFP_KERNEL);
568 clear_page(page_address(sd->save_area));
570 if (svm_sev_enabled()) {
571 sd->sev_vmcbs = kmalloc_array(max_sev_asid + 1,
578 per_cpu(svm_data, cpu) = sd;
583 __free_page(sd->save_area);
590 static int direct_access_msr_slot(u32 msr)
594 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++)
595 if (direct_access_msrs[i].index == msr)
601 static void set_shadow_msr_intercept(struct kvm_vcpu *vcpu, u32 msr, int read,
604 struct vcpu_svm *svm = to_svm(vcpu);
605 int slot = direct_access_msr_slot(msr);
610 /* Set the shadow bitmaps to the desired intercept states */
612 set_bit(slot, svm->shadow_msr_intercept.read);
614 clear_bit(slot, svm->shadow_msr_intercept.read);
617 set_bit(slot, svm->shadow_msr_intercept.write);
619 clear_bit(slot, svm->shadow_msr_intercept.write);
622 static bool valid_msr_intercept(u32 index)
624 return direct_access_msr_slot(index) != -ENOENT;
627 static bool msr_write_intercepted(struct kvm_vcpu *vcpu, u32 msr)
634 msrpm = is_guest_mode(vcpu) ? to_svm(vcpu)->nested.msrpm:
637 offset = svm_msrpm_offset(msr);
638 bit_write = 2 * (msr & 0x0f) + 1;
641 BUG_ON(offset == MSR_INVALID);
643 return !!test_bit(bit_write, &tmp);
646 static void set_msr_interception_bitmap(struct kvm_vcpu *vcpu, u32 *msrpm,
647 u32 msr, int read, int write)
649 u8 bit_read, bit_write;
654 * If this warning triggers extend the direct_access_msrs list at the
655 * beginning of the file
657 WARN_ON(!valid_msr_intercept(msr));
659 /* Enforce non allowed MSRs to trap */
660 if (read && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_READ))
663 if (write && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_WRITE))
666 offset = svm_msrpm_offset(msr);
667 bit_read = 2 * (msr & 0x0f);
668 bit_write = 2 * (msr & 0x0f) + 1;
671 BUG_ON(offset == MSR_INVALID);
673 read ? clear_bit(bit_read, &tmp) : set_bit(bit_read, &tmp);
674 write ? clear_bit(bit_write, &tmp) : set_bit(bit_write, &tmp);
679 static void set_msr_interception(struct kvm_vcpu *vcpu, u32 *msrpm, u32 msr,
682 set_shadow_msr_intercept(vcpu, msr, read, write);
683 set_msr_interception_bitmap(vcpu, msrpm, msr, read, write);
686 u32 *svm_vcpu_alloc_msrpm(void)
688 struct page *pages = alloc_pages(GFP_KERNEL_ACCOUNT, MSRPM_ALLOC_ORDER);
694 msrpm = page_address(pages);
695 memset(msrpm, 0xff, PAGE_SIZE * (1 << MSRPM_ALLOC_ORDER));
700 void svm_vcpu_init_msrpm(struct kvm_vcpu *vcpu, u32 *msrpm)
704 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
705 if (!direct_access_msrs[i].always)
707 set_msr_interception(vcpu, msrpm, direct_access_msrs[i].index, 1, 1);
712 void svm_vcpu_free_msrpm(u32 *msrpm)
714 __free_pages(virt_to_page(msrpm), MSRPM_ALLOC_ORDER);
717 static void svm_msr_filter_changed(struct kvm_vcpu *vcpu)
719 struct vcpu_svm *svm = to_svm(vcpu);
723 * Set intercept permissions for all direct access MSRs again. They
724 * will automatically get filtered through the MSR filter, so we are
725 * back in sync after this.
727 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
728 u32 msr = direct_access_msrs[i].index;
729 u32 read = test_bit(i, svm->shadow_msr_intercept.read);
730 u32 write = test_bit(i, svm->shadow_msr_intercept.write);
732 set_msr_interception_bitmap(vcpu, svm->msrpm, msr, read, write);
736 static void add_msr_offset(u32 offset)
740 for (i = 0; i < MSRPM_OFFSETS; ++i) {
742 /* Offset already in list? */
743 if (msrpm_offsets[i] == offset)
746 /* Slot used by another offset? */
747 if (msrpm_offsets[i] != MSR_INVALID)
750 /* Add offset to list */
751 msrpm_offsets[i] = offset;
757 * If this BUG triggers the msrpm_offsets table has an overflow. Just
758 * increase MSRPM_OFFSETS in this case.
763 static void init_msrpm_offsets(void)
767 memset(msrpm_offsets, 0xff, sizeof(msrpm_offsets));
769 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
772 offset = svm_msrpm_offset(direct_access_msrs[i].index);
773 BUG_ON(offset == MSR_INVALID);
775 add_msr_offset(offset);
779 static void svm_enable_lbrv(struct kvm_vcpu *vcpu)
781 struct vcpu_svm *svm = to_svm(vcpu);
783 svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
784 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
785 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
786 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
787 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
790 static void svm_disable_lbrv(struct kvm_vcpu *vcpu)
792 struct vcpu_svm *svm = to_svm(vcpu);
794 svm->vmcb->control.virt_ext &= ~LBR_CTL_ENABLE_MASK;
795 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 0, 0);
796 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 0, 0);
797 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 0, 0);
798 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 0, 0);
801 void disable_nmi_singlestep(struct vcpu_svm *svm)
803 svm->nmi_singlestep = false;
805 if (!(svm->vcpu.guest_debug & KVM_GUESTDBG_SINGLESTEP)) {
806 /* Clear our flags if they were not set by the guest */
807 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
808 svm->vmcb->save.rflags &= ~X86_EFLAGS_TF;
809 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
810 svm->vmcb->save.rflags &= ~X86_EFLAGS_RF;
814 static void grow_ple_window(struct kvm_vcpu *vcpu)
816 struct vcpu_svm *svm = to_svm(vcpu);
817 struct vmcb_control_area *control = &svm->vmcb->control;
818 int old = control->pause_filter_count;
820 control->pause_filter_count = __grow_ple_window(old,
822 pause_filter_count_grow,
823 pause_filter_count_max);
825 if (control->pause_filter_count != old) {
826 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
827 trace_kvm_ple_window_update(vcpu->vcpu_id,
828 control->pause_filter_count, old);
832 static void shrink_ple_window(struct kvm_vcpu *vcpu)
834 struct vcpu_svm *svm = to_svm(vcpu);
835 struct vmcb_control_area *control = &svm->vmcb->control;
836 int old = control->pause_filter_count;
838 control->pause_filter_count =
839 __shrink_ple_window(old,
841 pause_filter_count_shrink,
843 if (control->pause_filter_count != old) {
844 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
845 trace_kvm_ple_window_update(vcpu->vcpu_id,
846 control->pause_filter_count, old);
851 * The default MMIO mask is a single bit (excluding the present bit),
852 * which could conflict with the memory encryption bit. Check for
853 * memory encryption support and override the default MMIO mask if
854 * memory encryption is enabled.
856 static __init void svm_adjust_mmio_mask(void)
858 unsigned int enc_bit, mask_bit;
861 /* If there is no memory encryption support, use existing mask */
862 if (cpuid_eax(0x80000000) < 0x8000001f)
865 /* If memory encryption is not enabled, use existing mask */
866 rdmsrl(MSR_K8_SYSCFG, msr);
867 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
870 enc_bit = cpuid_ebx(0x8000001f) & 0x3f;
871 mask_bit = boot_cpu_data.x86_phys_bits;
873 /* Increment the mask bit if it is the same as the encryption bit */
874 if (enc_bit == mask_bit)
878 * If the mask bit location is below 52, then some bits above the
879 * physical addressing limit will always be reserved, so use the
880 * rsvd_bits() function to generate the mask. This mask, along with
881 * the present bit, will be used to generate a page fault with
884 * If the mask bit location is 52 (or above), then clear the mask.
886 mask = (mask_bit < 52) ? rsvd_bits(mask_bit, 51) | PT_PRESENT_MASK : 0;
888 kvm_mmu_set_mmio_spte_mask(mask, PT_WRITABLE_MASK | PT_USER_MASK);
891 static void svm_hardware_teardown(void)
895 if (svm_sev_enabled())
896 sev_hardware_teardown();
898 for_each_possible_cpu(cpu)
901 __free_pages(pfn_to_page(iopm_base >> PAGE_SHIFT), IOPM_ALLOC_ORDER);
905 static __init void svm_set_cpu_caps(void)
911 /* CPUID 0x80000001 and 0x8000000A (SVM features) */
913 kvm_cpu_cap_set(X86_FEATURE_SVM);
916 kvm_cpu_cap_set(X86_FEATURE_NRIPS);
919 kvm_cpu_cap_set(X86_FEATURE_NPT);
922 /* CPUID 0x80000008 */
923 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) ||
924 boot_cpu_has(X86_FEATURE_AMD_SSBD))
925 kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
927 /* Enable INVPCID feature */
928 kvm_cpu_cap_check_and_set(X86_FEATURE_INVPCID);
931 static __init int svm_hardware_setup(void)
934 struct page *iopm_pages;
938 iopm_pages = alloc_pages(GFP_KERNEL, IOPM_ALLOC_ORDER);
943 iopm_va = page_address(iopm_pages);
944 memset(iopm_va, 0xff, PAGE_SIZE * (1 << IOPM_ALLOC_ORDER));
945 iopm_base = page_to_pfn(iopm_pages) << PAGE_SHIFT;
947 init_msrpm_offsets();
949 supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
951 if (boot_cpu_has(X86_FEATURE_NX))
952 kvm_enable_efer_bits(EFER_NX);
954 if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
955 kvm_enable_efer_bits(EFER_FFXSR);
957 if (boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
958 kvm_has_tsc_control = true;
959 kvm_max_tsc_scaling_ratio = TSC_RATIO_MAX;
960 kvm_tsc_scaling_ratio_frac_bits = 32;
963 /* Check for pause filtering support */
964 if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
965 pause_filter_count = 0;
966 pause_filter_thresh = 0;
967 } else if (!boot_cpu_has(X86_FEATURE_PFTHRESHOLD)) {
968 pause_filter_thresh = 0;
972 printk(KERN_INFO "kvm: Nested Virtualization enabled\n");
973 kvm_enable_efer_bits(EFER_SVME | EFER_LMSLE);
976 if (IS_ENABLED(CONFIG_KVM_AMD_SEV) && sev) {
977 sev_hardware_setup();
983 svm_adjust_mmio_mask();
985 for_each_possible_cpu(cpu) {
986 r = svm_cpu_init(cpu);
991 if (!boot_cpu_has(X86_FEATURE_NPT))
994 if (npt_enabled && !npt)
997 kvm_configure_mmu(npt_enabled, get_max_npt_level(), PG_LEVEL_1G);
998 pr_info("kvm: Nested Paging %sabled\n", npt_enabled ? "en" : "dis");
1001 if (!boot_cpu_has(X86_FEATURE_NRIPS))
1007 !boot_cpu_has(X86_FEATURE_AVIC) ||
1008 !IS_ENABLED(CONFIG_X86_LOCAL_APIC)) {
1011 pr_info("AVIC enabled\n");
1013 amd_iommu_register_ga_log_notifier(&avic_ga_log_notifier);
1019 !boot_cpu_has(X86_FEATURE_V_VMSAVE_VMLOAD) ||
1020 !IS_ENABLED(CONFIG_X86_64)) {
1023 pr_info("Virtual VMLOAD VMSAVE supported\n");
1028 if (!boot_cpu_has(X86_FEATURE_VGIF))
1031 pr_info("Virtual GIF supported\n");
1037 * It seems that on AMD processors PTE's accessed bit is
1038 * being set by the CPU hardware before the NPF vmexit.
1039 * This is not expected behaviour and our tests fail because
1041 * A workaround here is to disable support for
1042 * GUEST_MAXPHYADDR < HOST_MAXPHYADDR if NPT is enabled.
1043 * In this case userspace can know if there is support using
1044 * KVM_CAP_SMALLER_MAXPHYADDR extension and decide how to handle
1046 * If future AMD CPU models change the behaviour described above,
1047 * this variable can be changed accordingly
1049 allow_smaller_maxphyaddr = !npt_enabled;
1054 svm_hardware_teardown();
1058 static void init_seg(struct vmcb_seg *seg)
1061 seg->attrib = SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK |
1062 SVM_SELECTOR_WRITE_MASK; /* Read/Write Data Segment */
1063 seg->limit = 0xffff;
1067 static void init_sys_seg(struct vmcb_seg *seg, uint32_t type)
1070 seg->attrib = SVM_SELECTOR_P_MASK | type;
1071 seg->limit = 0xffff;
1075 static u64 svm_write_l1_tsc_offset(struct kvm_vcpu *vcpu, u64 offset)
1077 struct vcpu_svm *svm = to_svm(vcpu);
1078 u64 g_tsc_offset = 0;
1080 if (is_guest_mode(vcpu)) {
1081 /* Write L1's TSC offset. */
1082 g_tsc_offset = svm->vmcb->control.tsc_offset -
1083 svm->nested.hsave->control.tsc_offset;
1084 svm->nested.hsave->control.tsc_offset = offset;
1087 trace_kvm_write_tsc_offset(vcpu->vcpu_id,
1088 svm->vmcb->control.tsc_offset - g_tsc_offset,
1091 svm->vmcb->control.tsc_offset = offset + g_tsc_offset;
1093 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1094 return svm->vmcb->control.tsc_offset;
1097 static void svm_check_invpcid(struct vcpu_svm *svm)
1100 * Intercept INVPCID instruction only if shadow page table is
1101 * enabled. Interception is not required with nested page table
1104 if (kvm_cpu_cap_has(X86_FEATURE_INVPCID)) {
1106 svm_set_intercept(svm, INTERCEPT_INVPCID);
1108 svm_clr_intercept(svm, INTERCEPT_INVPCID);
1112 static void init_vmcb(struct vcpu_svm *svm)
1114 struct vmcb_control_area *control = &svm->vmcb->control;
1115 struct vmcb_save_area *save = &svm->vmcb->save;
1117 svm->vcpu.arch.hflags = 0;
1119 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1120 svm_set_intercept(svm, INTERCEPT_CR3_READ);
1121 svm_set_intercept(svm, INTERCEPT_CR4_READ);
1122 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1123 svm_set_intercept(svm, INTERCEPT_CR3_WRITE);
1124 svm_set_intercept(svm, INTERCEPT_CR4_WRITE);
1125 if (!kvm_vcpu_apicv_active(&svm->vcpu))
1126 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
1128 set_dr_intercepts(svm);
1130 set_exception_intercept(svm, PF_VECTOR);
1131 set_exception_intercept(svm, UD_VECTOR);
1132 set_exception_intercept(svm, MC_VECTOR);
1133 set_exception_intercept(svm, AC_VECTOR);
1134 set_exception_intercept(svm, DB_VECTOR);
1136 * Guest access to VMware backdoor ports could legitimately
1137 * trigger #GP because of TSS I/O permission bitmap.
1138 * We intercept those #GP and allow access to them anyway
1141 if (enable_vmware_backdoor)
1142 set_exception_intercept(svm, GP_VECTOR);
1144 svm_set_intercept(svm, INTERCEPT_INTR);
1145 svm_set_intercept(svm, INTERCEPT_NMI);
1146 svm_set_intercept(svm, INTERCEPT_SMI);
1147 svm_set_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1148 svm_set_intercept(svm, INTERCEPT_RDPMC);
1149 svm_set_intercept(svm, INTERCEPT_CPUID);
1150 svm_set_intercept(svm, INTERCEPT_INVD);
1151 svm_set_intercept(svm, INTERCEPT_INVLPG);
1152 svm_set_intercept(svm, INTERCEPT_INVLPGA);
1153 svm_set_intercept(svm, INTERCEPT_IOIO_PROT);
1154 svm_set_intercept(svm, INTERCEPT_MSR_PROT);
1155 svm_set_intercept(svm, INTERCEPT_TASK_SWITCH);
1156 svm_set_intercept(svm, INTERCEPT_SHUTDOWN);
1157 svm_set_intercept(svm, INTERCEPT_VMRUN);
1158 svm_set_intercept(svm, INTERCEPT_VMMCALL);
1159 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1160 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1161 svm_set_intercept(svm, INTERCEPT_STGI);
1162 svm_set_intercept(svm, INTERCEPT_CLGI);
1163 svm_set_intercept(svm, INTERCEPT_SKINIT);
1164 svm_set_intercept(svm, INTERCEPT_WBINVD);
1165 svm_set_intercept(svm, INTERCEPT_XSETBV);
1166 svm_set_intercept(svm, INTERCEPT_RDPRU);
1167 svm_set_intercept(svm, INTERCEPT_RSM);
1169 if (!kvm_mwait_in_guest(svm->vcpu.kvm)) {
1170 svm_set_intercept(svm, INTERCEPT_MONITOR);
1171 svm_set_intercept(svm, INTERCEPT_MWAIT);
1174 if (!kvm_hlt_in_guest(svm->vcpu.kvm))
1175 svm_set_intercept(svm, INTERCEPT_HLT);
1177 control->iopm_base_pa = __sme_set(iopm_base);
1178 control->msrpm_base_pa = __sme_set(__pa(svm->msrpm));
1179 control->int_ctl = V_INTR_MASKING_MASK;
1181 init_seg(&save->es);
1182 init_seg(&save->ss);
1183 init_seg(&save->ds);
1184 init_seg(&save->fs);
1185 init_seg(&save->gs);
1187 save->cs.selector = 0xf000;
1188 save->cs.base = 0xffff0000;
1189 /* Executable/Readable Code Segment */
1190 save->cs.attrib = SVM_SELECTOR_READ_MASK | SVM_SELECTOR_P_MASK |
1191 SVM_SELECTOR_S_MASK | SVM_SELECTOR_CODE_MASK;
1192 save->cs.limit = 0xffff;
1194 save->gdtr.limit = 0xffff;
1195 save->idtr.limit = 0xffff;
1197 init_sys_seg(&save->ldtr, SEG_TYPE_LDT);
1198 init_sys_seg(&save->tr, SEG_TYPE_BUSY_TSS16);
1200 svm_set_efer(&svm->vcpu, 0);
1201 save->dr6 = 0xffff0ff0;
1202 kvm_set_rflags(&svm->vcpu, 2);
1203 save->rip = 0x0000fff0;
1204 svm->vcpu.arch.regs[VCPU_REGS_RIP] = save->rip;
1207 * svm_set_cr0() sets PG and WP and clears NW and CD on save->cr0.
1208 * It also updates the guest-visible cr0 value.
1210 svm_set_cr0(&svm->vcpu, X86_CR0_NW | X86_CR0_CD | X86_CR0_ET);
1211 kvm_mmu_reset_context(&svm->vcpu);
1213 save->cr4 = X86_CR4_PAE;
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 = svm->vcpu.arch.pat;
1227 svm->asid_generation = 0;
1230 svm->nested.vmcb12_gpa = 0;
1231 svm->vcpu.arch.hflags = 0;
1233 if (!kvm_pause_in_guest(svm->vcpu.kvm)) {
1234 control->pause_filter_count = pause_filter_count;
1235 if (pause_filter_thresh)
1236 control->pause_filter_thresh = pause_filter_thresh;
1237 svm_set_intercept(svm, INTERCEPT_PAUSE);
1239 svm_clr_intercept(svm, INTERCEPT_PAUSE);
1242 svm_check_invpcid(svm);
1244 if (kvm_vcpu_apicv_active(&svm->vcpu))
1245 avic_init_vmcb(svm);
1248 * If hardware supports Virtual VMLOAD VMSAVE then enable it
1249 * in VMCB and clear intercepts to avoid #VMEXIT.
1252 svm_clr_intercept(svm, INTERCEPT_VMLOAD);
1253 svm_clr_intercept(svm, INTERCEPT_VMSAVE);
1254 svm->vmcb->control.virt_ext |= VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
1258 svm_clr_intercept(svm, INTERCEPT_STGI);
1259 svm_clr_intercept(svm, INTERCEPT_CLGI);
1260 svm->vmcb->control.int_ctl |= V_GIF_ENABLE_MASK;
1263 if (sev_guest(svm->vcpu.kvm)) {
1264 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
1265 clr_exception_intercept(svm, UD_VECTOR);
1268 vmcb_mark_all_dirty(svm->vmcb);
1274 static void svm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
1276 struct vcpu_svm *svm = to_svm(vcpu);
1281 svm->virt_spec_ctrl = 0;
1284 svm->vcpu.arch.apic_base = APIC_DEFAULT_PHYS_BASE |
1285 MSR_IA32_APICBASE_ENABLE;
1286 if (kvm_vcpu_is_reset_bsp(&svm->vcpu))
1287 svm->vcpu.arch.apic_base |= MSR_IA32_APICBASE_BSP;
1291 kvm_cpuid(vcpu, &eax, &dummy, &dummy, &dummy, false);
1292 kvm_rdx_write(vcpu, eax);
1294 if (kvm_vcpu_apicv_active(vcpu) && !init_event)
1295 avic_update_vapic_bar(svm, APIC_DEFAULT_PHYS_BASE);
1298 static int svm_create_vcpu(struct kvm_vcpu *vcpu)
1300 struct vcpu_svm *svm;
1301 struct page *vmcb_page;
1302 struct page *vmsa_page = NULL;
1305 BUILD_BUG_ON(offsetof(struct vcpu_svm, vcpu) != 0);
1309 vmcb_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1313 if (sev_es_guest(svm->vcpu.kvm)) {
1315 * SEV-ES guests require a separate VMSA page used to contain
1316 * the encrypted register state of the guest.
1318 vmsa_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1320 goto error_free_vmcb_page;
1323 * SEV-ES guests maintain an encrypted version of their FPU
1324 * state which is restored and saved on VMRUN and VMEXIT.
1325 * Free the fpu structure to prevent KVM from attempting to
1326 * access the FPU state.
1328 kvm_free_guest_fpu(vcpu);
1331 err = avic_init_vcpu(svm);
1333 goto error_free_vmsa_page;
1335 /* We initialize this flag to true to make sure that the is_running
1336 * bit would be set the first time the vcpu is loaded.
1338 if (irqchip_in_kernel(vcpu->kvm) && kvm_apicv_activated(vcpu->kvm))
1339 svm->avic_is_running = true;
1341 svm->msrpm = svm_vcpu_alloc_msrpm();
1343 goto error_free_vmsa_page;
1345 svm_vcpu_init_msrpm(vcpu, svm->msrpm);
1347 svm->vmcb = page_address(vmcb_page);
1348 svm->vmcb_pa = __sme_set(page_to_pfn(vmcb_page) << PAGE_SHIFT);
1351 svm->vmsa = page_address(vmsa_page);
1353 svm->asid_generation = 0;
1356 svm_init_osvw(vcpu);
1357 vcpu->arch.microcode_version = 0x01000065;
1361 error_free_vmsa_page:
1363 __free_page(vmsa_page);
1364 error_free_vmcb_page:
1365 __free_page(vmcb_page);
1370 static void svm_clear_current_vmcb(struct vmcb *vmcb)
1374 for_each_online_cpu(i)
1375 cmpxchg(&per_cpu(svm_data, i)->current_vmcb, vmcb, NULL);
1378 static void svm_free_vcpu(struct kvm_vcpu *vcpu)
1380 struct vcpu_svm *svm = to_svm(vcpu);
1383 * The vmcb page can be recycled, causing a false negative in
1384 * svm_vcpu_load(). So, ensure that no logical CPU has this
1385 * vmcb page recorded as its current vmcb.
1387 svm_clear_current_vmcb(svm->vmcb);
1389 svm_free_nested(svm);
1391 sev_free_vcpu(vcpu);
1393 __free_page(pfn_to_page(__sme_clr(svm->vmcb_pa) >> PAGE_SHIFT));
1394 __free_pages(virt_to_page(svm->msrpm), MSRPM_ALLOC_ORDER);
1397 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1399 struct vcpu_svm *svm = to_svm(vcpu);
1400 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1403 if (unlikely(cpu != vcpu->cpu)) {
1404 svm->asid_generation = 0;
1405 vmcb_mark_all_dirty(svm->vmcb);
1408 #ifdef CONFIG_X86_64
1409 rdmsrl(MSR_GS_BASE, to_svm(vcpu)->host.gs_base);
1411 savesegment(fs, svm->host.fs);
1412 savesegment(gs, svm->host.gs);
1413 svm->host.ldt = kvm_read_ldt();
1415 for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
1416 rdmsrl(host_save_user_msrs[i], svm->host_user_msrs[i]);
1418 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
1419 u64 tsc_ratio = vcpu->arch.tsc_scaling_ratio;
1420 if (tsc_ratio != __this_cpu_read(current_tsc_ratio)) {
1421 __this_cpu_write(current_tsc_ratio, tsc_ratio);
1422 wrmsrl(MSR_AMD64_TSC_RATIO, tsc_ratio);
1425 /* This assumes that the kernel never uses MSR_TSC_AUX */
1426 if (static_cpu_has(X86_FEATURE_RDTSCP))
1427 wrmsrl(MSR_TSC_AUX, svm->tsc_aux);
1429 if (sd->current_vmcb != svm->vmcb) {
1430 sd->current_vmcb = svm->vmcb;
1431 indirect_branch_prediction_barrier();
1433 avic_vcpu_load(vcpu, cpu);
1436 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
1438 struct vcpu_svm *svm = to_svm(vcpu);
1441 avic_vcpu_put(vcpu);
1443 ++vcpu->stat.host_state_reload;
1444 kvm_load_ldt(svm->host.ldt);
1445 #ifdef CONFIG_X86_64
1446 loadsegment(fs, svm->host.fs);
1447 wrmsrl(MSR_KERNEL_GS_BASE, current->thread.gsbase);
1448 load_gs_index(svm->host.gs);
1450 #ifdef CONFIG_X86_32_LAZY_GS
1451 loadsegment(gs, svm->host.gs);
1454 for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
1455 wrmsrl(host_save_user_msrs[i], svm->host_user_msrs[i]);
1458 static unsigned long svm_get_rflags(struct kvm_vcpu *vcpu)
1460 struct vcpu_svm *svm = to_svm(vcpu);
1461 unsigned long rflags = svm->vmcb->save.rflags;
1463 if (svm->nmi_singlestep) {
1464 /* Hide our flags if they were not set by the guest */
1465 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
1466 rflags &= ~X86_EFLAGS_TF;
1467 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
1468 rflags &= ~X86_EFLAGS_RF;
1473 static void svm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
1475 if (to_svm(vcpu)->nmi_singlestep)
1476 rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
1479 * Any change of EFLAGS.VM is accompanied by a reload of SS
1480 * (caused by either a task switch or an inter-privilege IRET),
1481 * so we do not need to update the CPL here.
1483 to_svm(vcpu)->vmcb->save.rflags = rflags;
1486 static void svm_cache_reg(struct kvm_vcpu *vcpu, enum kvm_reg reg)
1489 case VCPU_EXREG_PDPTR:
1490 BUG_ON(!npt_enabled);
1491 load_pdptrs(vcpu, vcpu->arch.walk_mmu, kvm_read_cr3(vcpu));
1498 static void svm_set_vintr(struct vcpu_svm *svm)
1500 struct vmcb_control_area *control;
1502 /* The following fields are ignored when AVIC is enabled */
1503 WARN_ON(kvm_vcpu_apicv_active(&svm->vcpu));
1504 svm_set_intercept(svm, INTERCEPT_VINTR);
1507 * This is just a dummy VINTR to actually cause a vmexit to happen.
1508 * Actual injection of virtual interrupts happens through EVENTINJ.
1510 control = &svm->vmcb->control;
1511 control->int_vector = 0x0;
1512 control->int_ctl &= ~V_INTR_PRIO_MASK;
1513 control->int_ctl |= V_IRQ_MASK |
1514 ((/*control->int_vector >> 4*/ 0xf) << V_INTR_PRIO_SHIFT);
1515 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1518 static void svm_clear_vintr(struct vcpu_svm *svm)
1520 const u32 mask = V_TPR_MASK | V_GIF_ENABLE_MASK | V_GIF_MASK | V_INTR_MASKING_MASK;
1521 svm_clr_intercept(svm, INTERCEPT_VINTR);
1523 /* Drop int_ctl fields related to VINTR injection. */
1524 svm->vmcb->control.int_ctl &= mask;
1525 if (is_guest_mode(&svm->vcpu)) {
1526 svm->nested.hsave->control.int_ctl &= mask;
1528 WARN_ON((svm->vmcb->control.int_ctl & V_TPR_MASK) !=
1529 (svm->nested.ctl.int_ctl & V_TPR_MASK));
1530 svm->vmcb->control.int_ctl |= svm->nested.ctl.int_ctl & ~mask;
1533 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1536 static struct vmcb_seg *svm_seg(struct kvm_vcpu *vcpu, int seg)
1538 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1541 case VCPU_SREG_CS: return &save->cs;
1542 case VCPU_SREG_DS: return &save->ds;
1543 case VCPU_SREG_ES: return &save->es;
1544 case VCPU_SREG_FS: return &save->fs;
1545 case VCPU_SREG_GS: return &save->gs;
1546 case VCPU_SREG_SS: return &save->ss;
1547 case VCPU_SREG_TR: return &save->tr;
1548 case VCPU_SREG_LDTR: return &save->ldtr;
1554 static u64 svm_get_segment_base(struct kvm_vcpu *vcpu, int seg)
1556 struct vmcb_seg *s = svm_seg(vcpu, seg);
1561 static void svm_get_segment(struct kvm_vcpu *vcpu,
1562 struct kvm_segment *var, int seg)
1564 struct vmcb_seg *s = svm_seg(vcpu, seg);
1566 var->base = s->base;
1567 var->limit = s->limit;
1568 var->selector = s->selector;
1569 var->type = s->attrib & SVM_SELECTOR_TYPE_MASK;
1570 var->s = (s->attrib >> SVM_SELECTOR_S_SHIFT) & 1;
1571 var->dpl = (s->attrib >> SVM_SELECTOR_DPL_SHIFT) & 3;
1572 var->present = (s->attrib >> SVM_SELECTOR_P_SHIFT) & 1;
1573 var->avl = (s->attrib >> SVM_SELECTOR_AVL_SHIFT) & 1;
1574 var->l = (s->attrib >> SVM_SELECTOR_L_SHIFT) & 1;
1575 var->db = (s->attrib >> SVM_SELECTOR_DB_SHIFT) & 1;
1578 * AMD CPUs circa 2014 track the G bit for all segments except CS.
1579 * However, the SVM spec states that the G bit is not observed by the
1580 * CPU, and some VMware virtual CPUs drop the G bit for all segments.
1581 * So let's synthesize a legal G bit for all segments, this helps
1582 * running KVM nested. It also helps cross-vendor migration, because
1583 * Intel's vmentry has a check on the 'G' bit.
1585 var->g = s->limit > 0xfffff;
1588 * AMD's VMCB does not have an explicit unusable field, so emulate it
1589 * for cross vendor migration purposes by "not present"
1591 var->unusable = !var->present;
1596 * Work around a bug where the busy flag in the tr selector
1606 * The accessed bit must always be set in the segment
1607 * descriptor cache, although it can be cleared in the
1608 * descriptor, the cached bit always remains at 1. Since
1609 * Intel has a check on this, set it here to support
1610 * cross-vendor migration.
1617 * On AMD CPUs sometimes the DB bit in the segment
1618 * descriptor is left as 1, although the whole segment has
1619 * been made unusable. Clear it here to pass an Intel VMX
1620 * entry check when cross vendor migrating.
1624 /* This is symmetric with svm_set_segment() */
1625 var->dpl = to_svm(vcpu)->vmcb->save.cpl;
1630 static int svm_get_cpl(struct kvm_vcpu *vcpu)
1632 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1637 static void svm_get_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1639 struct vcpu_svm *svm = to_svm(vcpu);
1641 dt->size = svm->vmcb->save.idtr.limit;
1642 dt->address = svm->vmcb->save.idtr.base;
1645 static void svm_set_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1647 struct vcpu_svm *svm = to_svm(vcpu);
1649 svm->vmcb->save.idtr.limit = dt->size;
1650 svm->vmcb->save.idtr.base = dt->address ;
1651 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1654 static void svm_get_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1656 struct vcpu_svm *svm = to_svm(vcpu);
1658 dt->size = svm->vmcb->save.gdtr.limit;
1659 dt->address = svm->vmcb->save.gdtr.base;
1662 static void svm_set_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1664 struct vcpu_svm *svm = to_svm(vcpu);
1666 svm->vmcb->save.gdtr.limit = dt->size;
1667 svm->vmcb->save.gdtr.base = dt->address ;
1668 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1671 static void update_cr0_intercept(struct vcpu_svm *svm)
1677 * SEV-ES guests must always keep the CR intercepts cleared. CR
1678 * tracking is done using the CR write traps.
1680 if (sev_es_guest(svm->vcpu.kvm))
1683 gcr0 = svm->vcpu.arch.cr0;
1684 hcr0 = &svm->vmcb->save.cr0;
1685 *hcr0 = (*hcr0 & ~SVM_CR0_SELECTIVE_MASK)
1686 | (gcr0 & SVM_CR0_SELECTIVE_MASK);
1688 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1690 if (gcr0 == *hcr0) {
1691 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1692 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1694 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1695 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1699 void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1701 struct vcpu_svm *svm = to_svm(vcpu);
1703 #ifdef CONFIG_X86_64
1704 if (vcpu->arch.efer & EFER_LME && !vcpu->arch.guest_state_protected) {
1705 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
1706 vcpu->arch.efer |= EFER_LMA;
1707 svm->vmcb->save.efer |= EFER_LMA | EFER_LME;
1710 if (is_paging(vcpu) && !(cr0 & X86_CR0_PG)) {
1711 vcpu->arch.efer &= ~EFER_LMA;
1712 svm->vmcb->save.efer &= ~(EFER_LMA | EFER_LME);
1716 vcpu->arch.cr0 = cr0;
1719 cr0 |= X86_CR0_PG | X86_CR0_WP;
1722 * re-enable caching here because the QEMU bios
1723 * does not do it - this results in some delay at
1726 if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
1727 cr0 &= ~(X86_CR0_CD | X86_CR0_NW);
1728 svm->vmcb->save.cr0 = cr0;
1729 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1730 update_cr0_intercept(svm);
1733 static bool svm_is_valid_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1738 void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1740 unsigned long host_cr4_mce = cr4_read_shadow() & X86_CR4_MCE;
1741 unsigned long old_cr4 = vcpu->arch.cr4;
1743 if (npt_enabled && ((old_cr4 ^ cr4) & X86_CR4_PGE))
1744 svm_flush_tlb(vcpu);
1746 vcpu->arch.cr4 = cr4;
1749 cr4 |= host_cr4_mce;
1750 to_svm(vcpu)->vmcb->save.cr4 = cr4;
1751 vmcb_mark_dirty(to_svm(vcpu)->vmcb, VMCB_CR);
1753 if ((cr4 ^ old_cr4) & (X86_CR4_OSXSAVE | X86_CR4_PKE))
1754 kvm_update_cpuid_runtime(vcpu);
1757 static void svm_set_segment(struct kvm_vcpu *vcpu,
1758 struct kvm_segment *var, int seg)
1760 struct vcpu_svm *svm = to_svm(vcpu);
1761 struct vmcb_seg *s = svm_seg(vcpu, seg);
1763 s->base = var->base;
1764 s->limit = var->limit;
1765 s->selector = var->selector;
1766 s->attrib = (var->type & SVM_SELECTOR_TYPE_MASK);
1767 s->attrib |= (var->s & 1) << SVM_SELECTOR_S_SHIFT;
1768 s->attrib |= (var->dpl & 3) << SVM_SELECTOR_DPL_SHIFT;
1769 s->attrib |= ((var->present & 1) && !var->unusable) << SVM_SELECTOR_P_SHIFT;
1770 s->attrib |= (var->avl & 1) << SVM_SELECTOR_AVL_SHIFT;
1771 s->attrib |= (var->l & 1) << SVM_SELECTOR_L_SHIFT;
1772 s->attrib |= (var->db & 1) << SVM_SELECTOR_DB_SHIFT;
1773 s->attrib |= (var->g & 1) << SVM_SELECTOR_G_SHIFT;
1776 * This is always accurate, except if SYSRET returned to a segment
1777 * with SS.DPL != 3. Intel does not have this quirk, and always
1778 * forces SS.DPL to 3 on sysret, so we ignore that case; fixing it
1779 * would entail passing the CPL to userspace and back.
1781 if (seg == VCPU_SREG_SS)
1782 /* This is symmetric with svm_get_segment() */
1783 svm->vmcb->save.cpl = (var->dpl & 3);
1785 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
1788 static void update_exception_bitmap(struct kvm_vcpu *vcpu)
1790 struct vcpu_svm *svm = to_svm(vcpu);
1792 clr_exception_intercept(svm, BP_VECTOR);
1794 if (vcpu->guest_debug & KVM_GUESTDBG_ENABLE) {
1795 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP)
1796 set_exception_intercept(svm, BP_VECTOR);
1800 static void new_asid(struct vcpu_svm *svm, struct svm_cpu_data *sd)
1802 if (sd->next_asid > sd->max_asid) {
1803 ++sd->asid_generation;
1804 sd->next_asid = sd->min_asid;
1805 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
1806 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1809 svm->asid_generation = sd->asid_generation;
1810 svm->asid = sd->next_asid++;
1813 static void svm_set_dr6(struct vcpu_svm *svm, unsigned long value)
1815 struct vmcb *vmcb = svm->vmcb;
1817 if (svm->vcpu.arch.guest_state_protected)
1820 if (unlikely(value != vmcb->save.dr6)) {
1821 vmcb->save.dr6 = value;
1822 vmcb_mark_dirty(vmcb, VMCB_DR);
1826 static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
1828 struct vcpu_svm *svm = to_svm(vcpu);
1830 if (vcpu->arch.guest_state_protected)
1833 get_debugreg(vcpu->arch.db[0], 0);
1834 get_debugreg(vcpu->arch.db[1], 1);
1835 get_debugreg(vcpu->arch.db[2], 2);
1836 get_debugreg(vcpu->arch.db[3], 3);
1838 * We cannot reset svm->vmcb->save.dr6 to DR6_FIXED_1|DR6_RTM here,
1839 * because db_interception might need it. We can do it before vmentry.
1841 vcpu->arch.dr6 = svm->vmcb->save.dr6;
1842 vcpu->arch.dr7 = svm->vmcb->save.dr7;
1843 vcpu->arch.switch_db_regs &= ~KVM_DEBUGREG_WONT_EXIT;
1844 set_dr_intercepts(svm);
1847 static void svm_set_dr7(struct kvm_vcpu *vcpu, unsigned long value)
1849 struct vcpu_svm *svm = to_svm(vcpu);
1851 if (vcpu->arch.guest_state_protected)
1854 svm->vmcb->save.dr7 = value;
1855 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
1858 static int pf_interception(struct vcpu_svm *svm)
1860 u64 fault_address = __sme_clr(svm->vmcb->control.exit_info_2);
1861 u64 error_code = svm->vmcb->control.exit_info_1;
1863 return kvm_handle_page_fault(&svm->vcpu, error_code, fault_address,
1864 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1865 svm->vmcb->control.insn_bytes : NULL,
1866 svm->vmcb->control.insn_len);
1869 static int npf_interception(struct vcpu_svm *svm)
1871 u64 fault_address = __sme_clr(svm->vmcb->control.exit_info_2);
1872 u64 error_code = svm->vmcb->control.exit_info_1;
1874 trace_kvm_page_fault(fault_address, error_code);
1875 return kvm_mmu_page_fault(&svm->vcpu, fault_address, error_code,
1876 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1877 svm->vmcb->control.insn_bytes : NULL,
1878 svm->vmcb->control.insn_len);
1881 static int db_interception(struct vcpu_svm *svm)
1883 struct kvm_run *kvm_run = svm->vcpu.run;
1884 struct kvm_vcpu *vcpu = &svm->vcpu;
1886 if (!(svm->vcpu.guest_debug &
1887 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) &&
1888 !svm->nmi_singlestep) {
1889 u32 payload = (svm->vmcb->save.dr6 ^ DR6_RTM) & ~DR6_FIXED_1;
1890 kvm_queue_exception_p(&svm->vcpu, DB_VECTOR, payload);
1894 if (svm->nmi_singlestep) {
1895 disable_nmi_singlestep(svm);
1896 /* Make sure we check for pending NMIs upon entry */
1897 kvm_make_request(KVM_REQ_EVENT, vcpu);
1900 if (svm->vcpu.guest_debug &
1901 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) {
1902 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1903 kvm_run->debug.arch.dr6 = svm->vmcb->save.dr6;
1904 kvm_run->debug.arch.dr7 = svm->vmcb->save.dr7;
1905 kvm_run->debug.arch.pc =
1906 svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1907 kvm_run->debug.arch.exception = DB_VECTOR;
1914 static int bp_interception(struct vcpu_svm *svm)
1916 struct kvm_run *kvm_run = svm->vcpu.run;
1918 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1919 kvm_run->debug.arch.pc = svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1920 kvm_run->debug.arch.exception = BP_VECTOR;
1924 static int ud_interception(struct vcpu_svm *svm)
1926 return handle_ud(&svm->vcpu);
1929 static int ac_interception(struct vcpu_svm *svm)
1931 kvm_queue_exception_e(&svm->vcpu, AC_VECTOR, 0);
1935 static int gp_interception(struct vcpu_svm *svm)
1937 struct kvm_vcpu *vcpu = &svm->vcpu;
1938 u32 error_code = svm->vmcb->control.exit_info_1;
1940 WARN_ON_ONCE(!enable_vmware_backdoor);
1943 * VMware backdoor emulation on #GP interception only handles IN{S},
1944 * OUT{S}, and RDPMC, none of which generate a non-zero error code.
1947 kvm_queue_exception_e(vcpu, GP_VECTOR, error_code);
1950 return kvm_emulate_instruction(vcpu, EMULTYPE_VMWARE_GP);
1953 static bool is_erratum_383(void)
1958 if (!erratum_383_found)
1961 value = native_read_msr_safe(MSR_IA32_MC0_STATUS, &err);
1965 /* Bit 62 may or may not be set for this mce */
1966 value &= ~(1ULL << 62);
1968 if (value != 0xb600000000010015ULL)
1971 /* Clear MCi_STATUS registers */
1972 for (i = 0; i < 6; ++i)
1973 native_write_msr_safe(MSR_IA32_MCx_STATUS(i), 0, 0);
1975 value = native_read_msr_safe(MSR_IA32_MCG_STATUS, &err);
1979 value &= ~(1ULL << 2);
1980 low = lower_32_bits(value);
1981 high = upper_32_bits(value);
1983 native_write_msr_safe(MSR_IA32_MCG_STATUS, low, high);
1986 /* Flush tlb to evict multi-match entries */
1992 static void svm_handle_mce(struct vcpu_svm *svm)
1994 if (is_erratum_383()) {
1996 * Erratum 383 triggered. Guest state is corrupt so kill the
1999 pr_err("KVM: Guest triggered AMD Erratum 383\n");
2001 kvm_make_request(KVM_REQ_TRIPLE_FAULT, &svm->vcpu);
2007 * On an #MC intercept the MCE handler is not called automatically in
2008 * the host. So do it by hand here.
2010 kvm_machine_check();
2013 static int mc_interception(struct vcpu_svm *svm)
2018 static int shutdown_interception(struct vcpu_svm *svm)
2020 struct kvm_run *kvm_run = svm->vcpu.run;
2023 * The VM save area has already been encrypted so it
2024 * cannot be reinitialized - just terminate.
2026 if (sev_es_guest(svm->vcpu.kvm))
2030 * VMCB is undefined after a SHUTDOWN intercept
2031 * so reinitialize it.
2033 clear_page(svm->vmcb);
2036 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
2040 static int io_interception(struct vcpu_svm *svm)
2042 struct kvm_vcpu *vcpu = &svm->vcpu;
2043 u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
2044 int size, in, string;
2047 ++svm->vcpu.stat.io_exits;
2048 string = (io_info & SVM_IOIO_STR_MASK) != 0;
2049 in = (io_info & SVM_IOIO_TYPE_MASK) != 0;
2050 port = io_info >> 16;
2051 size = (io_info & SVM_IOIO_SIZE_MASK) >> SVM_IOIO_SIZE_SHIFT;
2054 if (sev_es_guest(vcpu->kvm))
2055 return sev_es_string_io(svm, size, port, in);
2057 return kvm_emulate_instruction(vcpu, 0);
2060 svm->next_rip = svm->vmcb->control.exit_info_2;
2062 return kvm_fast_pio(&svm->vcpu, size, port, in);
2065 static int nmi_interception(struct vcpu_svm *svm)
2070 static int intr_interception(struct vcpu_svm *svm)
2072 ++svm->vcpu.stat.irq_exits;
2076 static int nop_on_interception(struct vcpu_svm *svm)
2081 static int halt_interception(struct vcpu_svm *svm)
2083 return kvm_emulate_halt(&svm->vcpu);
2086 static int vmmcall_interception(struct vcpu_svm *svm)
2088 return kvm_emulate_hypercall(&svm->vcpu);
2091 static int vmload_interception(struct vcpu_svm *svm)
2093 struct vmcb *nested_vmcb;
2094 struct kvm_host_map map;
2097 if (nested_svm_check_permissions(svm))
2100 ret = kvm_vcpu_map(&svm->vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2103 kvm_inject_gp(&svm->vcpu, 0);
2107 nested_vmcb = map.hva;
2109 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2111 nested_svm_vmloadsave(nested_vmcb, svm->vmcb);
2112 kvm_vcpu_unmap(&svm->vcpu, &map, true);
2117 static int vmsave_interception(struct vcpu_svm *svm)
2119 struct vmcb *nested_vmcb;
2120 struct kvm_host_map map;
2123 if (nested_svm_check_permissions(svm))
2126 ret = kvm_vcpu_map(&svm->vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2129 kvm_inject_gp(&svm->vcpu, 0);
2133 nested_vmcb = map.hva;
2135 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2137 nested_svm_vmloadsave(svm->vmcb, nested_vmcb);
2138 kvm_vcpu_unmap(&svm->vcpu, &map, true);
2143 static int vmrun_interception(struct vcpu_svm *svm)
2145 if (nested_svm_check_permissions(svm))
2148 return nested_svm_vmrun(svm);
2151 void svm_set_gif(struct vcpu_svm *svm, bool value)
2155 * If VGIF is enabled, the STGI intercept is only added to
2156 * detect the opening of the SMI/NMI window; remove it now.
2157 * Likewise, clear the VINTR intercept, we will set it
2158 * again while processing KVM_REQ_EVENT if needed.
2160 if (vgif_enabled(svm))
2161 svm_clr_intercept(svm, INTERCEPT_STGI);
2162 if (svm_is_intercept(svm, INTERCEPT_VINTR))
2163 svm_clear_vintr(svm);
2166 if (svm->vcpu.arch.smi_pending ||
2167 svm->vcpu.arch.nmi_pending ||
2168 kvm_cpu_has_injectable_intr(&svm->vcpu))
2169 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2174 * After a CLGI no interrupts should come. But if vGIF is
2175 * in use, we still rely on the VINTR intercept (rather than
2176 * STGI) to detect an open interrupt window.
2178 if (!vgif_enabled(svm))
2179 svm_clear_vintr(svm);
2183 static int stgi_interception(struct vcpu_svm *svm)
2187 if (nested_svm_check_permissions(svm))
2190 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2191 svm_set_gif(svm, true);
2195 static int clgi_interception(struct vcpu_svm *svm)
2199 if (nested_svm_check_permissions(svm))
2202 ret = kvm_skip_emulated_instruction(&svm->vcpu);
2203 svm_set_gif(svm, false);
2207 static int invlpga_interception(struct vcpu_svm *svm)
2209 struct kvm_vcpu *vcpu = &svm->vcpu;
2211 trace_kvm_invlpga(svm->vmcb->save.rip, kvm_rcx_read(&svm->vcpu),
2212 kvm_rax_read(&svm->vcpu));
2214 /* Let's treat INVLPGA the same as INVLPG (can be optimized!) */
2215 kvm_mmu_invlpg(vcpu, kvm_rax_read(&svm->vcpu));
2217 return kvm_skip_emulated_instruction(&svm->vcpu);
2220 static int skinit_interception(struct vcpu_svm *svm)
2222 trace_kvm_skinit(svm->vmcb->save.rip, kvm_rax_read(&svm->vcpu));
2224 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2228 static int wbinvd_interception(struct vcpu_svm *svm)
2230 return kvm_emulate_wbinvd(&svm->vcpu);
2233 static int xsetbv_interception(struct vcpu_svm *svm)
2235 u64 new_bv = kvm_read_edx_eax(&svm->vcpu);
2236 u32 index = kvm_rcx_read(&svm->vcpu);
2238 if (kvm_set_xcr(&svm->vcpu, index, new_bv) == 0) {
2239 return kvm_skip_emulated_instruction(&svm->vcpu);
2245 static int rdpru_interception(struct vcpu_svm *svm)
2247 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2251 static int task_switch_interception(struct vcpu_svm *svm)
2255 int int_type = svm->vmcb->control.exit_int_info &
2256 SVM_EXITINTINFO_TYPE_MASK;
2257 int int_vec = svm->vmcb->control.exit_int_info & SVM_EVTINJ_VEC_MASK;
2259 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_TYPE_MASK;
2261 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID;
2262 bool has_error_code = false;
2265 tss_selector = (u16)svm->vmcb->control.exit_info_1;
2267 if (svm->vmcb->control.exit_info_2 &
2268 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
2269 reason = TASK_SWITCH_IRET;
2270 else if (svm->vmcb->control.exit_info_2 &
2271 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
2272 reason = TASK_SWITCH_JMP;
2274 reason = TASK_SWITCH_GATE;
2276 reason = TASK_SWITCH_CALL;
2278 if (reason == TASK_SWITCH_GATE) {
2280 case SVM_EXITINTINFO_TYPE_NMI:
2281 svm->vcpu.arch.nmi_injected = false;
2283 case SVM_EXITINTINFO_TYPE_EXEPT:
2284 if (svm->vmcb->control.exit_info_2 &
2285 (1ULL << SVM_EXITINFOSHIFT_TS_HAS_ERROR_CODE)) {
2286 has_error_code = true;
2288 (u32)svm->vmcb->control.exit_info_2;
2290 kvm_clear_exception_queue(&svm->vcpu);
2292 case SVM_EXITINTINFO_TYPE_INTR:
2293 kvm_clear_interrupt_queue(&svm->vcpu);
2300 if (reason != TASK_SWITCH_GATE ||
2301 int_type == SVM_EXITINTINFO_TYPE_SOFT ||
2302 (int_type == SVM_EXITINTINFO_TYPE_EXEPT &&
2303 (int_vec == OF_VECTOR || int_vec == BP_VECTOR))) {
2304 if (!skip_emulated_instruction(&svm->vcpu))
2308 if (int_type != SVM_EXITINTINFO_TYPE_SOFT)
2311 return kvm_task_switch(&svm->vcpu, tss_selector, int_vec, reason,
2312 has_error_code, error_code);
2315 static int cpuid_interception(struct vcpu_svm *svm)
2317 return kvm_emulate_cpuid(&svm->vcpu);
2320 static int iret_interception(struct vcpu_svm *svm)
2322 ++svm->vcpu.stat.nmi_window_exits;
2323 svm->vcpu.arch.hflags |= HF_IRET_MASK;
2324 if (!sev_es_guest(svm->vcpu.kvm)) {
2325 svm_clr_intercept(svm, INTERCEPT_IRET);
2326 svm->nmi_iret_rip = kvm_rip_read(&svm->vcpu);
2328 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2332 static int invd_interception(struct vcpu_svm *svm)
2334 /* Treat an INVD instruction as a NOP and just skip it. */
2335 return kvm_skip_emulated_instruction(&svm->vcpu);
2338 static int invlpg_interception(struct vcpu_svm *svm)
2340 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2341 return kvm_emulate_instruction(&svm->vcpu, 0);
2343 kvm_mmu_invlpg(&svm->vcpu, svm->vmcb->control.exit_info_1);
2344 return kvm_skip_emulated_instruction(&svm->vcpu);
2347 static int emulate_on_interception(struct vcpu_svm *svm)
2349 return kvm_emulate_instruction(&svm->vcpu, 0);
2352 static int rsm_interception(struct vcpu_svm *svm)
2354 return kvm_emulate_instruction_from_buffer(&svm->vcpu, rsm_ins_bytes, 2);
2357 static int rdpmc_interception(struct vcpu_svm *svm)
2362 return emulate_on_interception(svm);
2364 err = kvm_rdpmc(&svm->vcpu);
2365 return kvm_complete_insn_gp(&svm->vcpu, err);
2368 static bool check_selective_cr0_intercepted(struct vcpu_svm *svm,
2371 unsigned long cr0 = svm->vcpu.arch.cr0;
2374 if (!is_guest_mode(&svm->vcpu) ||
2375 (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_SELECTIVE_CR0))))
2378 cr0 &= ~SVM_CR0_SELECTIVE_MASK;
2379 val &= ~SVM_CR0_SELECTIVE_MASK;
2382 svm->vmcb->control.exit_code = SVM_EXIT_CR0_SEL_WRITE;
2383 ret = (nested_svm_exit_handled(svm) == NESTED_EXIT_DONE);
2389 #define CR_VALID (1ULL << 63)
2391 static int cr_interception(struct vcpu_svm *svm)
2397 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2398 return emulate_on_interception(svm);
2400 if (unlikely((svm->vmcb->control.exit_info_1 & CR_VALID) == 0))
2401 return emulate_on_interception(svm);
2403 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2404 if (svm->vmcb->control.exit_code == SVM_EXIT_CR0_SEL_WRITE)
2405 cr = SVM_EXIT_WRITE_CR0 - SVM_EXIT_READ_CR0;
2407 cr = svm->vmcb->control.exit_code - SVM_EXIT_READ_CR0;
2410 if (cr >= 16) { /* mov to cr */
2412 val = kvm_register_read(&svm->vcpu, reg);
2413 trace_kvm_cr_write(cr, val);
2416 if (!check_selective_cr0_intercepted(svm, val))
2417 err = kvm_set_cr0(&svm->vcpu, val);
2423 err = kvm_set_cr3(&svm->vcpu, val);
2426 err = kvm_set_cr4(&svm->vcpu, val);
2429 err = kvm_set_cr8(&svm->vcpu, val);
2432 WARN(1, "unhandled write to CR%d", cr);
2433 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2436 } else { /* mov from cr */
2439 val = kvm_read_cr0(&svm->vcpu);
2442 val = svm->vcpu.arch.cr2;
2445 val = kvm_read_cr3(&svm->vcpu);
2448 val = kvm_read_cr4(&svm->vcpu);
2451 val = kvm_get_cr8(&svm->vcpu);
2454 WARN(1, "unhandled read from CR%d", cr);
2455 kvm_queue_exception(&svm->vcpu, UD_VECTOR);
2458 kvm_register_write(&svm->vcpu, reg, val);
2459 trace_kvm_cr_read(cr, val);
2461 return kvm_complete_insn_gp(&svm->vcpu, err);
2464 static int cr_trap(struct vcpu_svm *svm)
2466 struct kvm_vcpu *vcpu = &svm->vcpu;
2467 unsigned long old_value, new_value;
2471 new_value = (unsigned long)svm->vmcb->control.exit_info_1;
2473 cr = svm->vmcb->control.exit_code - SVM_EXIT_CR0_WRITE_TRAP;
2476 old_value = kvm_read_cr0(vcpu);
2477 svm_set_cr0(vcpu, new_value);
2479 kvm_post_set_cr0(vcpu, old_value, new_value);
2482 old_value = kvm_read_cr4(vcpu);
2483 svm_set_cr4(vcpu, new_value);
2485 kvm_post_set_cr4(vcpu, old_value, new_value);
2488 ret = kvm_set_cr8(&svm->vcpu, new_value);
2491 WARN(1, "unhandled CR%d write trap", cr);
2492 kvm_queue_exception(vcpu, UD_VECTOR);
2496 return kvm_complete_insn_gp(vcpu, ret);
2499 static int dr_interception(struct vcpu_svm *svm)
2504 if (svm->vcpu.guest_debug == 0) {
2506 * No more DR vmexits; force a reload of the debug registers
2507 * and reenter on this instruction. The next vmexit will
2508 * retrieve the full state of the debug registers.
2510 clr_dr_intercepts(svm);
2511 svm->vcpu.arch.switch_db_regs |= KVM_DEBUGREG_WONT_EXIT;
2515 if (!boot_cpu_has(X86_FEATURE_DECODEASSISTS))
2516 return emulate_on_interception(svm);
2518 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2519 dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
2521 if (dr >= 16) { /* mov to DRn */
2522 if (!kvm_require_dr(&svm->vcpu, dr - 16))
2524 val = kvm_register_read(&svm->vcpu, reg);
2525 kvm_set_dr(&svm->vcpu, dr - 16, val);
2527 if (!kvm_require_dr(&svm->vcpu, dr))
2529 kvm_get_dr(&svm->vcpu, dr, &val);
2530 kvm_register_write(&svm->vcpu, reg, val);
2533 return kvm_skip_emulated_instruction(&svm->vcpu);
2536 static int cr8_write_interception(struct vcpu_svm *svm)
2538 struct kvm_run *kvm_run = svm->vcpu.run;
2541 u8 cr8_prev = kvm_get_cr8(&svm->vcpu);
2542 /* instruction emulation calls kvm_set_cr8() */
2543 r = cr_interception(svm);
2544 if (lapic_in_kernel(&svm->vcpu))
2546 if (cr8_prev <= kvm_get_cr8(&svm->vcpu))
2548 kvm_run->exit_reason = KVM_EXIT_SET_TPR;
2552 static int efer_trap(struct vcpu_svm *svm)
2554 struct msr_data msr_info;
2558 * Clear the EFER_SVME bit from EFER. The SVM code always sets this
2559 * bit in svm_set_efer(), but __kvm_valid_efer() checks it against
2560 * whether the guest has X86_FEATURE_SVM - this avoids a failure if
2561 * the guest doesn't have X86_FEATURE_SVM.
2563 msr_info.host_initiated = false;
2564 msr_info.index = MSR_EFER;
2565 msr_info.data = svm->vmcb->control.exit_info_1 & ~EFER_SVME;
2566 ret = kvm_set_msr_common(&svm->vcpu, &msr_info);
2568 return kvm_complete_insn_gp(&svm->vcpu, ret);
2571 static int svm_get_msr_feature(struct kvm_msr_entry *msr)
2575 switch (msr->index) {
2576 case MSR_F10H_DECFG:
2577 if (boot_cpu_has(X86_FEATURE_LFENCE_RDTSC))
2578 msr->data |= MSR_F10H_DECFG_LFENCE_SERIALIZE;
2580 case MSR_IA32_PERF_CAPABILITIES:
2583 return KVM_MSR_RET_INVALID;
2589 static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
2591 struct vcpu_svm *svm = to_svm(vcpu);
2593 switch (msr_info->index) {
2595 msr_info->data = svm->vmcb->save.star;
2597 #ifdef CONFIG_X86_64
2599 msr_info->data = svm->vmcb->save.lstar;
2602 msr_info->data = svm->vmcb->save.cstar;
2604 case MSR_KERNEL_GS_BASE:
2605 msr_info->data = svm->vmcb->save.kernel_gs_base;
2607 case MSR_SYSCALL_MASK:
2608 msr_info->data = svm->vmcb->save.sfmask;
2611 case MSR_IA32_SYSENTER_CS:
2612 msr_info->data = svm->vmcb->save.sysenter_cs;
2614 case MSR_IA32_SYSENTER_EIP:
2615 msr_info->data = svm->sysenter_eip;
2617 case MSR_IA32_SYSENTER_ESP:
2618 msr_info->data = svm->sysenter_esp;
2621 if (!boot_cpu_has(X86_FEATURE_RDTSCP))
2623 msr_info->data = svm->tsc_aux;
2626 * Nobody will change the following 5 values in the VMCB so we can
2627 * safely return them on rdmsr. They will always be 0 until LBRV is
2630 case MSR_IA32_DEBUGCTLMSR:
2631 msr_info->data = svm->vmcb->save.dbgctl;
2633 case MSR_IA32_LASTBRANCHFROMIP:
2634 msr_info->data = svm->vmcb->save.br_from;
2636 case MSR_IA32_LASTBRANCHTOIP:
2637 msr_info->data = svm->vmcb->save.br_to;
2639 case MSR_IA32_LASTINTFROMIP:
2640 msr_info->data = svm->vmcb->save.last_excp_from;
2642 case MSR_IA32_LASTINTTOIP:
2643 msr_info->data = svm->vmcb->save.last_excp_to;
2645 case MSR_VM_HSAVE_PA:
2646 msr_info->data = svm->nested.hsave_msr;
2649 msr_info->data = svm->nested.vm_cr_msr;
2651 case MSR_IA32_SPEC_CTRL:
2652 if (!msr_info->host_initiated &&
2653 !guest_has_spec_ctrl_msr(vcpu))
2656 msr_info->data = svm->spec_ctrl;
2658 case MSR_AMD64_VIRT_SPEC_CTRL:
2659 if (!msr_info->host_initiated &&
2660 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2663 msr_info->data = svm->virt_spec_ctrl;
2665 case MSR_F15H_IC_CFG: {
2669 family = guest_cpuid_family(vcpu);
2670 model = guest_cpuid_model(vcpu);
2672 if (family < 0 || model < 0)
2673 return kvm_get_msr_common(vcpu, msr_info);
2677 if (family == 0x15 &&
2678 (model >= 0x2 && model < 0x20))
2679 msr_info->data = 0x1E;
2682 case MSR_F10H_DECFG:
2683 msr_info->data = svm->msr_decfg;
2686 return kvm_get_msr_common(vcpu, msr_info);
2691 static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
2693 struct vcpu_svm *svm = to_svm(vcpu);
2694 if (!sev_es_guest(svm->vcpu.kvm) || !err)
2695 return kvm_complete_insn_gp(&svm->vcpu, err);
2697 ghcb_set_sw_exit_info_1(svm->ghcb, 1);
2698 ghcb_set_sw_exit_info_2(svm->ghcb,
2700 SVM_EVTINJ_TYPE_EXEPT |
2705 static int rdmsr_interception(struct vcpu_svm *svm)
2707 return kvm_emulate_rdmsr(&svm->vcpu);
2710 static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
2712 struct vcpu_svm *svm = to_svm(vcpu);
2713 int svm_dis, chg_mask;
2715 if (data & ~SVM_VM_CR_VALID_MASK)
2718 chg_mask = SVM_VM_CR_VALID_MASK;
2720 if (svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK)
2721 chg_mask &= ~(SVM_VM_CR_SVM_LOCK_MASK | SVM_VM_CR_SVM_DIS_MASK);
2723 svm->nested.vm_cr_msr &= ~chg_mask;
2724 svm->nested.vm_cr_msr |= (data & chg_mask);
2726 svm_dis = svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK;
2728 /* check for svm_disable while efer.svme is set */
2729 if (svm_dis && (vcpu->arch.efer & EFER_SVME))
2735 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
2737 struct vcpu_svm *svm = to_svm(vcpu);
2739 u32 ecx = msr->index;
2740 u64 data = msr->data;
2742 case MSR_IA32_CR_PAT:
2743 if (!kvm_mtrr_valid(vcpu, MSR_IA32_CR_PAT, data))
2745 vcpu->arch.pat = data;
2746 svm->vmcb->save.g_pat = data;
2747 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
2749 case MSR_IA32_SPEC_CTRL:
2750 if (!msr->host_initiated &&
2751 !guest_has_spec_ctrl_msr(vcpu))
2754 if (kvm_spec_ctrl_test_value(data))
2757 svm->spec_ctrl = data;
2763 * When it's written (to non-zero) for the first time, pass
2767 * The handling of the MSR bitmap for L2 guests is done in
2768 * nested_svm_vmrun_msrpm.
2769 * We update the L1 MSR bit as well since it will end up
2770 * touching the MSR anyway now.
2772 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
2774 case MSR_IA32_PRED_CMD:
2775 if (!msr->host_initiated &&
2776 !guest_has_pred_cmd_msr(vcpu))
2779 if (data & ~PRED_CMD_IBPB)
2781 if (!boot_cpu_has(X86_FEATURE_IBPB))
2786 wrmsrl(MSR_IA32_PRED_CMD, PRED_CMD_IBPB);
2787 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_PRED_CMD, 0, 1);
2789 case MSR_AMD64_VIRT_SPEC_CTRL:
2790 if (!msr->host_initiated &&
2791 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2794 if (data & ~SPEC_CTRL_SSBD)
2797 svm->virt_spec_ctrl = data;
2800 svm->vmcb->save.star = data;
2802 #ifdef CONFIG_X86_64
2804 svm->vmcb->save.lstar = data;
2807 svm->vmcb->save.cstar = data;
2809 case MSR_KERNEL_GS_BASE:
2810 svm->vmcb->save.kernel_gs_base = data;
2812 case MSR_SYSCALL_MASK:
2813 svm->vmcb->save.sfmask = data;
2816 case MSR_IA32_SYSENTER_CS:
2817 svm->vmcb->save.sysenter_cs = data;
2819 case MSR_IA32_SYSENTER_EIP:
2820 svm->sysenter_eip = data;
2821 svm->vmcb->save.sysenter_eip = data;
2823 case MSR_IA32_SYSENTER_ESP:
2824 svm->sysenter_esp = data;
2825 svm->vmcb->save.sysenter_esp = data;
2828 if (!boot_cpu_has(X86_FEATURE_RDTSCP))
2832 * This is rare, so we update the MSR here instead of using
2833 * direct_access_msrs. Doing that would require a rdmsr in
2836 svm->tsc_aux = data;
2837 wrmsrl(MSR_TSC_AUX, svm->tsc_aux);
2839 case MSR_IA32_DEBUGCTLMSR:
2840 if (!boot_cpu_has(X86_FEATURE_LBRV)) {
2841 vcpu_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTL 0x%llx, nop\n",
2845 if (data & DEBUGCTL_RESERVED_BITS)
2848 svm->vmcb->save.dbgctl = data;
2849 vmcb_mark_dirty(svm->vmcb, VMCB_LBR);
2850 if (data & (1ULL<<0))
2851 svm_enable_lbrv(vcpu);
2853 svm_disable_lbrv(vcpu);
2855 case MSR_VM_HSAVE_PA:
2856 svm->nested.hsave_msr = data;
2859 return svm_set_vm_cr(vcpu, data);
2861 vcpu_unimpl(vcpu, "unimplemented wrmsr: 0x%x data 0x%llx\n", ecx, data);
2863 case MSR_F10H_DECFG: {
2864 struct kvm_msr_entry msr_entry;
2866 msr_entry.index = msr->index;
2867 if (svm_get_msr_feature(&msr_entry))
2870 /* Check the supported bits */
2871 if (data & ~msr_entry.data)
2874 /* Don't allow the guest to change a bit, #GP */
2875 if (!msr->host_initiated && (data ^ msr_entry.data))
2878 svm->msr_decfg = data;
2881 case MSR_IA32_APICBASE:
2882 if (kvm_vcpu_apicv_active(vcpu))
2883 avic_update_vapic_bar(to_svm(vcpu), data);
2886 return kvm_set_msr_common(vcpu, msr);
2891 static int wrmsr_interception(struct vcpu_svm *svm)
2893 return kvm_emulate_wrmsr(&svm->vcpu);
2896 static int msr_interception(struct vcpu_svm *svm)
2898 if (svm->vmcb->control.exit_info_1)
2899 return wrmsr_interception(svm);
2901 return rdmsr_interception(svm);
2904 static int interrupt_window_interception(struct vcpu_svm *svm)
2906 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2907 svm_clear_vintr(svm);
2910 * For AVIC, the only reason to end up here is ExtINTs.
2911 * In this case AVIC was temporarily disabled for
2912 * requesting the IRQ window and we have to re-enable it.
2914 svm_toggle_avic_for_irq_window(&svm->vcpu, true);
2916 ++svm->vcpu.stat.irq_window_exits;
2920 static int pause_interception(struct vcpu_svm *svm)
2922 struct kvm_vcpu *vcpu = &svm->vcpu;
2926 * CPL is not made available for an SEV-ES guest, therefore
2927 * vcpu->arch.preempted_in_kernel can never be true. Just
2928 * set in_kernel to false as well.
2930 in_kernel = !sev_es_guest(svm->vcpu.kvm) && svm_get_cpl(vcpu) == 0;
2932 if (!kvm_pause_in_guest(vcpu->kvm))
2933 grow_ple_window(vcpu);
2935 kvm_vcpu_on_spin(vcpu, in_kernel);
2939 static int nop_interception(struct vcpu_svm *svm)
2941 return kvm_skip_emulated_instruction(&(svm->vcpu));
2944 static int monitor_interception(struct vcpu_svm *svm)
2946 printk_once(KERN_WARNING "kvm: MONITOR instruction emulated as NOP!\n");
2947 return nop_interception(svm);
2950 static int mwait_interception(struct vcpu_svm *svm)
2952 printk_once(KERN_WARNING "kvm: MWAIT instruction emulated as NOP!\n");
2953 return nop_interception(svm);
2956 static int invpcid_interception(struct vcpu_svm *svm)
2958 struct kvm_vcpu *vcpu = &svm->vcpu;
2962 if (!guest_cpuid_has(vcpu, X86_FEATURE_INVPCID)) {
2963 kvm_queue_exception(vcpu, UD_VECTOR);
2968 * For an INVPCID intercept:
2969 * EXITINFO1 provides the linear address of the memory operand.
2970 * EXITINFO2 provides the contents of the register operand.
2972 type = svm->vmcb->control.exit_info_2;
2973 gva = svm->vmcb->control.exit_info_1;
2976 kvm_inject_gp(vcpu, 0);
2980 return kvm_handle_invpcid(vcpu, type, gva);
2983 static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = {
2984 [SVM_EXIT_READ_CR0] = cr_interception,
2985 [SVM_EXIT_READ_CR3] = cr_interception,
2986 [SVM_EXIT_READ_CR4] = cr_interception,
2987 [SVM_EXIT_READ_CR8] = cr_interception,
2988 [SVM_EXIT_CR0_SEL_WRITE] = cr_interception,
2989 [SVM_EXIT_WRITE_CR0] = cr_interception,
2990 [SVM_EXIT_WRITE_CR3] = cr_interception,
2991 [SVM_EXIT_WRITE_CR4] = cr_interception,
2992 [SVM_EXIT_WRITE_CR8] = cr8_write_interception,
2993 [SVM_EXIT_READ_DR0] = dr_interception,
2994 [SVM_EXIT_READ_DR1] = dr_interception,
2995 [SVM_EXIT_READ_DR2] = dr_interception,
2996 [SVM_EXIT_READ_DR3] = dr_interception,
2997 [SVM_EXIT_READ_DR4] = dr_interception,
2998 [SVM_EXIT_READ_DR5] = dr_interception,
2999 [SVM_EXIT_READ_DR6] = dr_interception,
3000 [SVM_EXIT_READ_DR7] = dr_interception,
3001 [SVM_EXIT_WRITE_DR0] = dr_interception,
3002 [SVM_EXIT_WRITE_DR1] = dr_interception,
3003 [SVM_EXIT_WRITE_DR2] = dr_interception,
3004 [SVM_EXIT_WRITE_DR3] = dr_interception,
3005 [SVM_EXIT_WRITE_DR4] = dr_interception,
3006 [SVM_EXIT_WRITE_DR5] = dr_interception,
3007 [SVM_EXIT_WRITE_DR6] = dr_interception,
3008 [SVM_EXIT_WRITE_DR7] = dr_interception,
3009 [SVM_EXIT_EXCP_BASE + DB_VECTOR] = db_interception,
3010 [SVM_EXIT_EXCP_BASE + BP_VECTOR] = bp_interception,
3011 [SVM_EXIT_EXCP_BASE + UD_VECTOR] = ud_interception,
3012 [SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
3013 [SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
3014 [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
3015 [SVM_EXIT_EXCP_BASE + GP_VECTOR] = gp_interception,
3016 [SVM_EXIT_INTR] = intr_interception,
3017 [SVM_EXIT_NMI] = nmi_interception,
3018 [SVM_EXIT_SMI] = nop_on_interception,
3019 [SVM_EXIT_INIT] = nop_on_interception,
3020 [SVM_EXIT_VINTR] = interrupt_window_interception,
3021 [SVM_EXIT_RDPMC] = rdpmc_interception,
3022 [SVM_EXIT_CPUID] = cpuid_interception,
3023 [SVM_EXIT_IRET] = iret_interception,
3024 [SVM_EXIT_INVD] = invd_interception,
3025 [SVM_EXIT_PAUSE] = pause_interception,
3026 [SVM_EXIT_HLT] = halt_interception,
3027 [SVM_EXIT_INVLPG] = invlpg_interception,
3028 [SVM_EXIT_INVLPGA] = invlpga_interception,
3029 [SVM_EXIT_IOIO] = io_interception,
3030 [SVM_EXIT_MSR] = msr_interception,
3031 [SVM_EXIT_TASK_SWITCH] = task_switch_interception,
3032 [SVM_EXIT_SHUTDOWN] = shutdown_interception,
3033 [SVM_EXIT_VMRUN] = vmrun_interception,
3034 [SVM_EXIT_VMMCALL] = vmmcall_interception,
3035 [SVM_EXIT_VMLOAD] = vmload_interception,
3036 [SVM_EXIT_VMSAVE] = vmsave_interception,
3037 [SVM_EXIT_STGI] = stgi_interception,
3038 [SVM_EXIT_CLGI] = clgi_interception,
3039 [SVM_EXIT_SKINIT] = skinit_interception,
3040 [SVM_EXIT_WBINVD] = wbinvd_interception,
3041 [SVM_EXIT_MONITOR] = monitor_interception,
3042 [SVM_EXIT_MWAIT] = mwait_interception,
3043 [SVM_EXIT_XSETBV] = xsetbv_interception,
3044 [SVM_EXIT_RDPRU] = rdpru_interception,
3045 [SVM_EXIT_EFER_WRITE_TRAP] = efer_trap,
3046 [SVM_EXIT_CR0_WRITE_TRAP] = cr_trap,
3047 [SVM_EXIT_CR4_WRITE_TRAP] = cr_trap,
3048 [SVM_EXIT_CR8_WRITE_TRAP] = cr_trap,
3049 [SVM_EXIT_INVPCID] = invpcid_interception,
3050 [SVM_EXIT_NPF] = npf_interception,
3051 [SVM_EXIT_RSM] = rsm_interception,
3052 [SVM_EXIT_AVIC_INCOMPLETE_IPI] = avic_incomplete_ipi_interception,
3053 [SVM_EXIT_AVIC_UNACCELERATED_ACCESS] = avic_unaccelerated_access_interception,
3054 [SVM_EXIT_VMGEXIT] = sev_handle_vmgexit,
3057 static void dump_vmcb(struct kvm_vcpu *vcpu)
3059 struct vcpu_svm *svm = to_svm(vcpu);
3060 struct vmcb_control_area *control = &svm->vmcb->control;
3061 struct vmcb_save_area *save = &svm->vmcb->save;
3063 if (!dump_invalid_vmcb) {
3064 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3068 pr_err("VMCB Control Area:\n");
3069 pr_err("%-20s%04x\n", "cr_read:", control->intercepts[INTERCEPT_CR] & 0xffff);
3070 pr_err("%-20s%04x\n", "cr_write:", control->intercepts[INTERCEPT_CR] >> 16);
3071 pr_err("%-20s%04x\n", "dr_read:", control->intercepts[INTERCEPT_DR] & 0xffff);
3072 pr_err("%-20s%04x\n", "dr_write:", control->intercepts[INTERCEPT_DR] >> 16);
3073 pr_err("%-20s%08x\n", "exceptions:", control->intercepts[INTERCEPT_EXCEPTION]);
3074 pr_err("%-20s%08x %08x\n", "intercepts:",
3075 control->intercepts[INTERCEPT_WORD3],
3076 control->intercepts[INTERCEPT_WORD4]);
3077 pr_err("%-20s%d\n", "pause filter count:", control->pause_filter_count);
3078 pr_err("%-20s%d\n", "pause filter threshold:",
3079 control->pause_filter_thresh);
3080 pr_err("%-20s%016llx\n", "iopm_base_pa:", control->iopm_base_pa);
3081 pr_err("%-20s%016llx\n", "msrpm_base_pa:", control->msrpm_base_pa);
3082 pr_err("%-20s%016llx\n", "tsc_offset:", control->tsc_offset);
3083 pr_err("%-20s%d\n", "asid:", control->asid);
3084 pr_err("%-20s%d\n", "tlb_ctl:", control->tlb_ctl);
3085 pr_err("%-20s%08x\n", "int_ctl:", control->int_ctl);
3086 pr_err("%-20s%08x\n", "int_vector:", control->int_vector);
3087 pr_err("%-20s%08x\n", "int_state:", control->int_state);
3088 pr_err("%-20s%08x\n", "exit_code:", control->exit_code);
3089 pr_err("%-20s%016llx\n", "exit_info1:", control->exit_info_1);
3090 pr_err("%-20s%016llx\n", "exit_info2:", control->exit_info_2);
3091 pr_err("%-20s%08x\n", "exit_int_info:", control->exit_int_info);
3092 pr_err("%-20s%08x\n", "exit_int_info_err:", control->exit_int_info_err);
3093 pr_err("%-20s%lld\n", "nested_ctl:", control->nested_ctl);
3094 pr_err("%-20s%016llx\n", "nested_cr3:", control->nested_cr3);
3095 pr_err("%-20s%016llx\n", "avic_vapic_bar:", control->avic_vapic_bar);
3096 pr_err("%-20s%016llx\n", "ghcb:", control->ghcb_gpa);
3097 pr_err("%-20s%08x\n", "event_inj:", control->event_inj);
3098 pr_err("%-20s%08x\n", "event_inj_err:", control->event_inj_err);
3099 pr_err("%-20s%lld\n", "virt_ext:", control->virt_ext);
3100 pr_err("%-20s%016llx\n", "next_rip:", control->next_rip);
3101 pr_err("%-20s%016llx\n", "avic_backing_page:", control->avic_backing_page);
3102 pr_err("%-20s%016llx\n", "avic_logical_id:", control->avic_logical_id);
3103 pr_err("%-20s%016llx\n", "avic_physical_id:", control->avic_physical_id);
3104 pr_err("VMCB State Save Area:\n");
3105 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3107 save->es.selector, save->es.attrib,
3108 save->es.limit, save->es.base);
3109 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3111 save->cs.selector, save->cs.attrib,
3112 save->cs.limit, save->cs.base);
3113 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3115 save->ss.selector, save->ss.attrib,
3116 save->ss.limit, save->ss.base);
3117 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3119 save->ds.selector, save->ds.attrib,
3120 save->ds.limit, save->ds.base);
3121 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3123 save->fs.selector, save->fs.attrib,
3124 save->fs.limit, save->fs.base);
3125 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3127 save->gs.selector, save->gs.attrib,
3128 save->gs.limit, save->gs.base);
3129 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3131 save->gdtr.selector, save->gdtr.attrib,
3132 save->gdtr.limit, save->gdtr.base);
3133 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3135 save->ldtr.selector, save->ldtr.attrib,
3136 save->ldtr.limit, save->ldtr.base);
3137 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3139 save->idtr.selector, save->idtr.attrib,
3140 save->idtr.limit, save->idtr.base);
3141 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3143 save->tr.selector, save->tr.attrib,
3144 save->tr.limit, save->tr.base);
3145 pr_err("cpl: %d efer: %016llx\n",
3146 save->cpl, save->efer);
3147 pr_err("%-15s %016llx %-13s %016llx\n",
3148 "cr0:", save->cr0, "cr2:", save->cr2);
3149 pr_err("%-15s %016llx %-13s %016llx\n",
3150 "cr3:", save->cr3, "cr4:", save->cr4);
3151 pr_err("%-15s %016llx %-13s %016llx\n",
3152 "dr6:", save->dr6, "dr7:", save->dr7);
3153 pr_err("%-15s %016llx %-13s %016llx\n",
3154 "rip:", save->rip, "rflags:", save->rflags);
3155 pr_err("%-15s %016llx %-13s %016llx\n",
3156 "rsp:", save->rsp, "rax:", save->rax);
3157 pr_err("%-15s %016llx %-13s %016llx\n",
3158 "star:", save->star, "lstar:", save->lstar);
3159 pr_err("%-15s %016llx %-13s %016llx\n",
3160 "cstar:", save->cstar, "sfmask:", save->sfmask);
3161 pr_err("%-15s %016llx %-13s %016llx\n",
3162 "kernel_gs_base:", save->kernel_gs_base,
3163 "sysenter_cs:", save->sysenter_cs);
3164 pr_err("%-15s %016llx %-13s %016llx\n",
3165 "sysenter_esp:", save->sysenter_esp,
3166 "sysenter_eip:", save->sysenter_eip);
3167 pr_err("%-15s %016llx %-13s %016llx\n",
3168 "gpat:", save->g_pat, "dbgctl:", save->dbgctl);
3169 pr_err("%-15s %016llx %-13s %016llx\n",
3170 "br_from:", save->br_from, "br_to:", save->br_to);
3171 pr_err("%-15s %016llx %-13s %016llx\n",
3172 "excp_from:", save->last_excp_from,
3173 "excp_to:", save->last_excp_to);
3176 static int svm_handle_invalid_exit(struct kvm_vcpu *vcpu, u64 exit_code)
3178 if (exit_code < ARRAY_SIZE(svm_exit_handlers) &&
3179 svm_exit_handlers[exit_code])
3182 vcpu_unimpl(vcpu, "svm: unexpected exit reason 0x%llx\n", exit_code);
3184 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3185 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
3186 vcpu->run->internal.ndata = 2;
3187 vcpu->run->internal.data[0] = exit_code;
3188 vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
3193 int svm_invoke_exit_handler(struct vcpu_svm *svm, u64 exit_code)
3195 if (svm_handle_invalid_exit(&svm->vcpu, exit_code))
3198 #ifdef CONFIG_RETPOLINE
3199 if (exit_code == SVM_EXIT_MSR)
3200 return msr_interception(svm);
3201 else if (exit_code == SVM_EXIT_VINTR)
3202 return interrupt_window_interception(svm);
3203 else if (exit_code == SVM_EXIT_INTR)
3204 return intr_interception(svm);
3205 else if (exit_code == SVM_EXIT_HLT)
3206 return halt_interception(svm);
3207 else if (exit_code == SVM_EXIT_NPF)
3208 return npf_interception(svm);
3210 return svm_exit_handlers[exit_code](svm);
3213 static void svm_get_exit_info(struct kvm_vcpu *vcpu, u64 *info1, u64 *info2,
3214 u32 *intr_info, u32 *error_code)
3216 struct vmcb_control_area *control = &to_svm(vcpu)->vmcb->control;
3218 *info1 = control->exit_info_1;
3219 *info2 = control->exit_info_2;
3220 *intr_info = control->exit_int_info;
3221 if ((*intr_info & SVM_EXITINTINFO_VALID) &&
3222 (*intr_info & SVM_EXITINTINFO_VALID_ERR))
3223 *error_code = control->exit_int_info_err;
3228 static int handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
3230 struct vcpu_svm *svm = to_svm(vcpu);
3231 struct kvm_run *kvm_run = vcpu->run;
3232 u32 exit_code = svm->vmcb->control.exit_code;
3234 trace_kvm_exit(exit_code, vcpu, KVM_ISA_SVM);
3236 /* SEV-ES guests must use the CR write traps to track CR registers. */
3237 if (!sev_es_guest(vcpu->kvm)) {
3238 if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
3239 vcpu->arch.cr0 = svm->vmcb->save.cr0;
3241 vcpu->arch.cr3 = svm->vmcb->save.cr3;
3244 if (is_guest_mode(vcpu)) {
3247 trace_kvm_nested_vmexit(exit_code, vcpu, KVM_ISA_SVM);
3249 vmexit = nested_svm_exit_special(svm);
3251 if (vmexit == NESTED_EXIT_CONTINUE)
3252 vmexit = nested_svm_exit_handled(svm);
3254 if (vmexit == NESTED_EXIT_DONE)
3258 if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
3259 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3260 kvm_run->fail_entry.hardware_entry_failure_reason
3261 = svm->vmcb->control.exit_code;
3262 kvm_run->fail_entry.cpu = vcpu->arch.last_vmentry_cpu;
3267 if (is_external_interrupt(svm->vmcb->control.exit_int_info) &&
3268 exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR &&
3269 exit_code != SVM_EXIT_NPF && exit_code != SVM_EXIT_TASK_SWITCH &&
3270 exit_code != SVM_EXIT_INTR && exit_code != SVM_EXIT_NMI)
3271 printk(KERN_ERR "%s: unexpected exit_int_info 0x%x "
3273 __func__, svm->vmcb->control.exit_int_info,
3276 if (exit_fastpath != EXIT_FASTPATH_NONE)
3279 return svm_invoke_exit_handler(svm, exit_code);
3282 static void reload_tss(struct kvm_vcpu *vcpu)
3284 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3286 sd->tss_desc->type = 9; /* available 32/64-bit TSS */
3290 static void pre_svm_run(struct vcpu_svm *svm)
3292 struct svm_cpu_data *sd = per_cpu(svm_data, svm->vcpu.cpu);
3294 if (sev_guest(svm->vcpu.kvm))
3295 return pre_sev_run(svm, svm->vcpu.cpu);
3297 /* FIXME: handle wraparound of asid_generation */
3298 if (svm->asid_generation != sd->asid_generation)
3302 static void svm_inject_nmi(struct kvm_vcpu *vcpu)
3304 struct vcpu_svm *svm = to_svm(vcpu);
3306 svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
3307 vcpu->arch.hflags |= HF_NMI_MASK;
3308 if (!sev_es_guest(svm->vcpu.kvm))
3309 svm_set_intercept(svm, INTERCEPT_IRET);
3310 ++vcpu->stat.nmi_injections;
3313 static void svm_set_irq(struct kvm_vcpu *vcpu)
3315 struct vcpu_svm *svm = to_svm(vcpu);
3317 BUG_ON(!(gif_set(svm)));
3319 trace_kvm_inj_virq(vcpu->arch.interrupt.nr);
3320 ++vcpu->stat.irq_injections;
3322 svm->vmcb->control.event_inj = vcpu->arch.interrupt.nr |
3323 SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR;
3326 static void update_cr8_intercept(struct kvm_vcpu *vcpu, int tpr, int irr)
3328 struct vcpu_svm *svm = to_svm(vcpu);
3331 * SEV-ES guests must always keep the CR intercepts cleared. CR
3332 * tracking is done using the CR write traps.
3334 if (sev_es_guest(vcpu->kvm))
3337 if (nested_svm_virtualize_tpr(vcpu))
3340 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3346 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
3349 bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
3351 struct vcpu_svm *svm = to_svm(vcpu);
3352 struct vmcb *vmcb = svm->vmcb;
3358 if (is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3361 ret = (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK) ||
3362 (svm->vcpu.arch.hflags & HF_NMI_MASK);
3367 static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3369 struct vcpu_svm *svm = to_svm(vcpu);
3370 if (svm->nested.nested_run_pending)
3373 /* An NMI must not be injected into L2 if it's supposed to VM-Exit. */
3374 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3377 return !svm_nmi_blocked(vcpu);
3380 static bool svm_get_nmi_mask(struct kvm_vcpu *vcpu)
3382 struct vcpu_svm *svm = to_svm(vcpu);
3384 return !!(svm->vcpu.arch.hflags & HF_NMI_MASK);
3387 static void svm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
3389 struct vcpu_svm *svm = to_svm(vcpu);
3392 svm->vcpu.arch.hflags |= HF_NMI_MASK;
3393 if (!sev_es_guest(svm->vcpu.kvm))
3394 svm_set_intercept(svm, INTERCEPT_IRET);
3396 svm->vcpu.arch.hflags &= ~HF_NMI_MASK;
3397 if (!sev_es_guest(svm->vcpu.kvm))
3398 svm_clr_intercept(svm, INTERCEPT_IRET);
3402 bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
3404 struct vcpu_svm *svm = to_svm(vcpu);
3405 struct vmcb *vmcb = svm->vmcb;
3410 if (sev_es_guest(svm->vcpu.kvm)) {
3412 * SEV-ES guests to not expose RFLAGS. Use the VMCB interrupt mask
3413 * bit to determine the state of the IF flag.
3415 if (!(vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK))
3417 } else if (is_guest_mode(vcpu)) {
3418 /* As long as interrupts are being delivered... */
3419 if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
3420 ? !(svm->nested.hsave->save.rflags & X86_EFLAGS_IF)
3421 : !(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3424 /* ... vmexits aren't blocked by the interrupt shadow */
3425 if (nested_exit_on_intr(svm))
3428 if (!(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3432 return (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK);
3435 static int svm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3437 struct vcpu_svm *svm = to_svm(vcpu);
3438 if (svm->nested.nested_run_pending)
3442 * An IRQ must not be injected into L2 if it's supposed to VM-Exit,
3443 * e.g. if the IRQ arrived asynchronously after checking nested events.
3445 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_intr(svm))
3448 return !svm_interrupt_blocked(vcpu);
3451 static void enable_irq_window(struct kvm_vcpu *vcpu)
3453 struct vcpu_svm *svm = to_svm(vcpu);
3456 * In case GIF=0 we can't rely on the CPU to tell us when GIF becomes
3457 * 1, because that's a separate STGI/VMRUN intercept. The next time we
3458 * get that intercept, this function will be called again though and
3459 * we'll get the vintr intercept. However, if the vGIF feature is
3460 * enabled, the STGI interception will not occur. Enable the irq
3461 * window under the assumption that the hardware will set the GIF.
3463 if (vgif_enabled(svm) || gif_set(svm)) {
3465 * IRQ window is not needed when AVIC is enabled,
3466 * unless we have pending ExtINT since it cannot be injected
3467 * via AVIC. In such case, we need to temporarily disable AVIC,
3468 * and fallback to injecting IRQ via V_IRQ.
3470 svm_toggle_avic_for_irq_window(vcpu, false);
3475 static void enable_nmi_window(struct kvm_vcpu *vcpu)
3477 struct vcpu_svm *svm = to_svm(vcpu);
3479 if ((svm->vcpu.arch.hflags & (HF_NMI_MASK | HF_IRET_MASK))
3481 return; /* IRET will cause a vm exit */
3483 if (!gif_set(svm)) {
3484 if (vgif_enabled(svm))
3485 svm_set_intercept(svm, INTERCEPT_STGI);
3486 return; /* STGI will cause a vm exit */
3490 * Something prevents NMI from been injected. Single step over possible
3491 * problem (IRET or exception injection or interrupt shadow)
3493 svm->nmi_singlestep_guest_rflags = svm_get_rflags(vcpu);
3494 svm->nmi_singlestep = true;
3495 svm->vmcb->save.rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
3498 static int svm_set_tss_addr(struct kvm *kvm, unsigned int addr)
3503 static int svm_set_identity_map_addr(struct kvm *kvm, u64 ident_addr)
3508 void svm_flush_tlb(struct kvm_vcpu *vcpu)
3510 struct vcpu_svm *svm = to_svm(vcpu);
3513 * Flush only the current ASID even if the TLB flush was invoked via
3514 * kvm_flush_remote_tlbs(). Although flushing remote TLBs requires all
3515 * ASIDs to be flushed, KVM uses a single ASID for L1 and L2, and
3516 * unconditionally does a TLB flush on both nested VM-Enter and nested
3517 * VM-Exit (via kvm_mmu_reset_context()).
3519 if (static_cpu_has(X86_FEATURE_FLUSHBYASID))
3520 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3522 svm->asid_generation--;
3525 static void svm_flush_tlb_gva(struct kvm_vcpu *vcpu, gva_t gva)
3527 struct vcpu_svm *svm = to_svm(vcpu);
3529 invlpga(gva, svm->vmcb->control.asid);
3532 static void svm_prepare_guest_switch(struct kvm_vcpu *vcpu)
3536 static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)
3538 struct vcpu_svm *svm = to_svm(vcpu);
3540 if (nested_svm_virtualize_tpr(vcpu))
3543 if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
3544 int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
3545 kvm_set_cr8(vcpu, cr8);
3549 static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
3551 struct vcpu_svm *svm = to_svm(vcpu);
3554 if (nested_svm_virtualize_tpr(vcpu) ||
3555 kvm_vcpu_apicv_active(vcpu))
3558 cr8 = kvm_get_cr8(vcpu);
3559 svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
3560 svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
3563 static void svm_complete_interrupts(struct vcpu_svm *svm)
3567 u32 exitintinfo = svm->vmcb->control.exit_int_info;
3568 unsigned int3_injected = svm->int3_injected;
3570 svm->int3_injected = 0;
3573 * If we've made progress since setting HF_IRET_MASK, we've
3574 * executed an IRET and can allow NMI injection.
3576 if ((svm->vcpu.arch.hflags & HF_IRET_MASK) &&
3577 (sev_es_guest(svm->vcpu.kvm) ||
3578 kvm_rip_read(&svm->vcpu) != svm->nmi_iret_rip)) {
3579 svm->vcpu.arch.hflags &= ~(HF_NMI_MASK | HF_IRET_MASK);
3580 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
3583 svm->vcpu.arch.nmi_injected = false;
3584 kvm_clear_exception_queue(&svm->vcpu);
3585 kvm_clear_interrupt_queue(&svm->vcpu);
3587 if (!(exitintinfo & SVM_EXITINTINFO_VALID))
3590 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
3592 vector = exitintinfo & SVM_EXITINTINFO_VEC_MASK;
3593 type = exitintinfo & SVM_EXITINTINFO_TYPE_MASK;
3596 case SVM_EXITINTINFO_TYPE_NMI:
3597 svm->vcpu.arch.nmi_injected = true;
3599 case SVM_EXITINTINFO_TYPE_EXEPT:
3601 * Never re-inject a #VC exception.
3603 if (vector == X86_TRAP_VC)
3607 * In case of software exceptions, do not reinject the vector,
3608 * but re-execute the instruction instead. Rewind RIP first
3609 * if we emulated INT3 before.
3611 if (kvm_exception_is_soft(vector)) {
3612 if (vector == BP_VECTOR && int3_injected &&
3613 kvm_is_linear_rip(&svm->vcpu, svm->int3_rip))
3614 kvm_rip_write(&svm->vcpu,
3615 kvm_rip_read(&svm->vcpu) -
3619 if (exitintinfo & SVM_EXITINTINFO_VALID_ERR) {
3620 u32 err = svm->vmcb->control.exit_int_info_err;
3621 kvm_requeue_exception_e(&svm->vcpu, vector, err);
3624 kvm_requeue_exception(&svm->vcpu, vector);
3626 case SVM_EXITINTINFO_TYPE_INTR:
3627 kvm_queue_interrupt(&svm->vcpu, vector, false);
3634 static void svm_cancel_injection(struct kvm_vcpu *vcpu)
3636 struct vcpu_svm *svm = to_svm(vcpu);
3637 struct vmcb_control_area *control = &svm->vmcb->control;
3639 control->exit_int_info = control->event_inj;
3640 control->exit_int_info_err = control->event_inj_err;
3641 control->event_inj = 0;
3642 svm_complete_interrupts(svm);
3645 static fastpath_t svm_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
3647 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_MSR &&
3648 to_svm(vcpu)->vmcb->control.exit_info_1)
3649 return handle_fastpath_set_msr_irqoff(vcpu);
3651 return EXIT_FASTPATH_NONE;
3654 void __svm_vcpu_run(unsigned long vmcb_pa, unsigned long *regs);
3656 static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu,
3657 struct vcpu_svm *svm)
3660 * VMENTER enables interrupts (host state), but the kernel state is
3661 * interrupts disabled when this is invoked. Also tell RCU about
3662 * it. This is the same logic as for exit_to_user_mode().
3664 * This ensures that e.g. latency analysis on the host observes
3665 * guest mode as interrupt enabled.
3667 * guest_enter_irqoff() informs context tracking about the
3668 * transition to guest mode and if enabled adjusts RCU state
3671 instrumentation_begin();
3672 trace_hardirqs_on_prepare();
3673 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
3674 instrumentation_end();
3676 guest_enter_irqoff();
3677 lockdep_hardirqs_on(CALLER_ADDR0);
3679 __svm_vcpu_run(svm->vmcb_pa, (unsigned long *)&svm->vcpu.arch.regs);
3681 #ifdef CONFIG_X86_64
3682 native_wrmsrl(MSR_GS_BASE, svm->host.gs_base);
3684 loadsegment(fs, svm->host.fs);
3685 #ifndef CONFIG_X86_32_LAZY_GS
3686 loadsegment(gs, svm->host.gs);
3691 * VMEXIT disables interrupts (host state), but tracing and lockdep
3692 * have them in state 'on' as recorded before entering guest mode.
3693 * Same as enter_from_user_mode().
3695 * guest_exit_irqoff() restores host context and reinstates RCU if
3696 * enabled and required.
3698 * This needs to be done before the below as native_read_msr()
3699 * contains a tracepoint and x86_spec_ctrl_restore_host() calls
3700 * into world and some more.
3702 lockdep_hardirqs_off(CALLER_ADDR0);
3703 guest_exit_irqoff();
3705 instrumentation_begin();
3706 trace_hardirqs_off_finish();
3707 instrumentation_end();
3710 static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
3712 struct vcpu_svm *svm = to_svm(vcpu);
3714 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
3715 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
3716 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
3719 * Disable singlestep if we're injecting an interrupt/exception.
3720 * We don't want our modified rflags to be pushed on the stack where
3721 * we might not be able to easily reset them if we disabled NMI
3724 if (svm->nmi_singlestep && svm->vmcb->control.event_inj) {
3726 * Event injection happens before external interrupts cause a
3727 * vmexit and interrupts are disabled here, so smp_send_reschedule
3728 * is enough to force an immediate vmexit.
3730 disable_nmi_singlestep(svm);
3731 smp_send_reschedule(vcpu->cpu);
3736 sync_lapic_to_cr8(vcpu);
3738 if (unlikely(svm->asid != svm->vmcb->control.asid)) {
3739 svm->vmcb->control.asid = svm->asid;
3740 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
3742 svm->vmcb->save.cr2 = vcpu->arch.cr2;
3745 * Run with all-zero DR6 unless needed, so that we can get the exact cause
3748 if (unlikely(svm->vcpu.arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT))
3749 svm_set_dr6(svm, vcpu->arch.dr6);
3751 svm_set_dr6(svm, DR6_FIXED_1 | DR6_RTM);
3754 kvm_load_guest_xsave_state(vcpu);
3756 kvm_wait_lapic_expire(vcpu);
3759 * If this vCPU has touched SPEC_CTRL, restore the guest's value if
3760 * it's non-zero. Since vmentry is serialising on affected CPUs, there
3761 * is no need to worry about the conditional branch over the wrmsr
3762 * being speculatively taken.
3764 x86_spec_ctrl_set_guest(svm->spec_ctrl, svm->virt_spec_ctrl);
3766 svm_vcpu_enter_exit(vcpu, svm);
3769 * We do not use IBRS in the kernel. If this vCPU has used the
3770 * SPEC_CTRL MSR it may have left it on; save the value and
3771 * turn it off. This is much more efficient than blindly adding
3772 * it to the atomic save/restore list. Especially as the former
3773 * (Saving guest MSRs on vmexit) doesn't even exist in KVM.
3775 * For non-nested case:
3776 * If the L01 MSR bitmap does not intercept the MSR, then we need to
3780 * If the L02 MSR bitmap does not intercept the MSR, then we need to
3783 if (unlikely(!msr_write_intercepted(vcpu, MSR_IA32_SPEC_CTRL)))
3784 svm->spec_ctrl = native_read_msr(MSR_IA32_SPEC_CTRL);
3788 x86_spec_ctrl_restore_host(svm->spec_ctrl, svm->virt_spec_ctrl);
3790 vcpu->arch.cr2 = svm->vmcb->save.cr2;
3791 vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
3792 vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
3793 vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
3795 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3796 kvm_before_interrupt(&svm->vcpu);
3798 kvm_load_host_xsave_state(vcpu);
3801 /* Any pending NMI will happen here */
3803 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3804 kvm_after_interrupt(&svm->vcpu);
3806 sync_cr8_to_lapic(vcpu);
3809 if (is_guest_mode(&svm->vcpu)) {
3810 sync_nested_vmcb_control(svm);
3811 svm->nested.nested_run_pending = 0;
3814 svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
3815 vmcb_mark_all_clean(svm->vmcb);
3817 /* if exit due to PF check for async PF */
3818 if (svm->vmcb->control.exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR)
3819 svm->vcpu.arch.apf.host_apf_flags =
3820 kvm_read_and_reset_apf_flags();
3823 vcpu->arch.regs_avail &= ~(1 << VCPU_EXREG_PDPTR);
3824 vcpu->arch.regs_dirty &= ~(1 << VCPU_EXREG_PDPTR);
3828 * We need to handle MC intercepts here before the vcpu has a chance to
3829 * change the physical cpu
3831 if (unlikely(svm->vmcb->control.exit_code ==
3832 SVM_EXIT_EXCP_BASE + MC_VECTOR))
3833 svm_handle_mce(svm);
3835 svm_complete_interrupts(svm);
3837 if (is_guest_mode(vcpu))
3838 return EXIT_FASTPATH_NONE;
3840 return svm_exit_handlers_fastpath(vcpu);
3843 static void svm_load_mmu_pgd(struct kvm_vcpu *vcpu, unsigned long root,
3846 struct vcpu_svm *svm = to_svm(vcpu);
3849 cr3 = __sme_set(root);
3851 svm->vmcb->control.nested_cr3 = cr3;
3852 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
3854 /* Loading L2's CR3 is handled by enter_svm_guest_mode. */
3855 if (!test_bit(VCPU_EXREG_CR3, (ulong *)&vcpu->arch.regs_avail))
3857 cr3 = vcpu->arch.cr3;
3860 svm->vmcb->save.cr3 = cr3;
3861 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
3864 static int is_disabled(void)
3868 rdmsrl(MSR_VM_CR, vm_cr);
3869 if (vm_cr & (1 << SVM_VM_CR_SVM_DISABLE))
3876 svm_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
3879 * Patch in the VMMCALL instruction:
3881 hypercall[0] = 0x0f;
3882 hypercall[1] = 0x01;
3883 hypercall[2] = 0xd9;
3886 static int __init svm_check_processor_compat(void)
3891 static bool svm_cpu_has_accelerated_tpr(void)
3897 * The kvm parameter can be NULL (module initialization, or invocation before
3898 * VM creation). Be sure to check the kvm parameter before using it.
3900 static bool svm_has_emulated_msr(struct kvm *kvm, u32 index)
3903 case MSR_IA32_MCG_EXT_CTL:
3904 case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:
3906 case MSR_IA32_SMBASE:
3907 /* SEV-ES guests do not support SMM, so report false */
3908 if (kvm && sev_es_guest(kvm))
3918 static u64 svm_get_mt_mask(struct kvm_vcpu *vcpu, gfn_t gfn, bool is_mmio)
3923 static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
3925 struct vcpu_svm *svm = to_svm(vcpu);
3926 struct kvm_cpuid_entry2 *best;
3928 vcpu->arch.xsaves_enabled = guest_cpuid_has(vcpu, X86_FEATURE_XSAVE) &&
3929 boot_cpu_has(X86_FEATURE_XSAVE) &&
3930 boot_cpu_has(X86_FEATURE_XSAVES);
3932 /* Update nrips enabled cache */
3933 svm->nrips_enabled = kvm_cpu_cap_has(X86_FEATURE_NRIPS) &&
3934 guest_cpuid_has(&svm->vcpu, X86_FEATURE_NRIPS);
3936 /* Check again if INVPCID interception if required */
3937 svm_check_invpcid(svm);
3939 /* For sev guests, the memory encryption bit is not reserved in CR3. */
3940 if (sev_guest(vcpu->kvm)) {
3941 best = kvm_find_cpuid_entry(vcpu, 0x8000001F, 0);
3943 vcpu->arch.cr3_lm_rsvd_bits &= ~(1UL << (best->ebx & 0x3f));
3946 if (!kvm_vcpu_apicv_active(vcpu))
3950 * AVIC does not work with an x2APIC mode guest. If the X2APIC feature
3951 * is exposed to the guest, disable AVIC.
3953 if (guest_cpuid_has(vcpu, X86_FEATURE_X2APIC))
3954 kvm_request_apicv_update(vcpu->kvm, false,
3955 APICV_INHIBIT_REASON_X2APIC);
3958 * Currently, AVIC does not work with nested virtualization.
3959 * So, we disable AVIC when cpuid for SVM is set in the L1 guest.
3961 if (nested && guest_cpuid_has(vcpu, X86_FEATURE_SVM))
3962 kvm_request_apicv_update(vcpu->kvm, false,
3963 APICV_INHIBIT_REASON_NESTED);
3966 static bool svm_has_wbinvd_exit(void)
3971 #define PRE_EX(exit) { .exit_code = (exit), \
3972 .stage = X86_ICPT_PRE_EXCEPT, }
3973 #define POST_EX(exit) { .exit_code = (exit), \
3974 .stage = X86_ICPT_POST_EXCEPT, }
3975 #define POST_MEM(exit) { .exit_code = (exit), \
3976 .stage = X86_ICPT_POST_MEMACCESS, }
3978 static const struct __x86_intercept {
3980 enum x86_intercept_stage stage;
3981 } x86_intercept_map[] = {
3982 [x86_intercept_cr_read] = POST_EX(SVM_EXIT_READ_CR0),
3983 [x86_intercept_cr_write] = POST_EX(SVM_EXIT_WRITE_CR0),
3984 [x86_intercept_clts] = POST_EX(SVM_EXIT_WRITE_CR0),
3985 [x86_intercept_lmsw] = POST_EX(SVM_EXIT_WRITE_CR0),
3986 [x86_intercept_smsw] = POST_EX(SVM_EXIT_READ_CR0),
3987 [x86_intercept_dr_read] = POST_EX(SVM_EXIT_READ_DR0),
3988 [x86_intercept_dr_write] = POST_EX(SVM_EXIT_WRITE_DR0),
3989 [x86_intercept_sldt] = POST_EX(SVM_EXIT_LDTR_READ),
3990 [x86_intercept_str] = POST_EX(SVM_EXIT_TR_READ),
3991 [x86_intercept_lldt] = POST_EX(SVM_EXIT_LDTR_WRITE),
3992 [x86_intercept_ltr] = POST_EX(SVM_EXIT_TR_WRITE),
3993 [x86_intercept_sgdt] = POST_EX(SVM_EXIT_GDTR_READ),
3994 [x86_intercept_sidt] = POST_EX(SVM_EXIT_IDTR_READ),
3995 [x86_intercept_lgdt] = POST_EX(SVM_EXIT_GDTR_WRITE),
3996 [x86_intercept_lidt] = POST_EX(SVM_EXIT_IDTR_WRITE),
3997 [x86_intercept_vmrun] = POST_EX(SVM_EXIT_VMRUN),
3998 [x86_intercept_vmmcall] = POST_EX(SVM_EXIT_VMMCALL),
3999 [x86_intercept_vmload] = POST_EX(SVM_EXIT_VMLOAD),
4000 [x86_intercept_vmsave] = POST_EX(SVM_EXIT_VMSAVE),
4001 [x86_intercept_stgi] = POST_EX(SVM_EXIT_STGI),
4002 [x86_intercept_clgi] = POST_EX(SVM_EXIT_CLGI),
4003 [x86_intercept_skinit] = POST_EX(SVM_EXIT_SKINIT),
4004 [x86_intercept_invlpga] = POST_EX(SVM_EXIT_INVLPGA),
4005 [x86_intercept_rdtscp] = POST_EX(SVM_EXIT_RDTSCP),
4006 [x86_intercept_monitor] = POST_MEM(SVM_EXIT_MONITOR),
4007 [x86_intercept_mwait] = POST_EX(SVM_EXIT_MWAIT),
4008 [x86_intercept_invlpg] = POST_EX(SVM_EXIT_INVLPG),
4009 [x86_intercept_invd] = POST_EX(SVM_EXIT_INVD),
4010 [x86_intercept_wbinvd] = POST_EX(SVM_EXIT_WBINVD),
4011 [x86_intercept_wrmsr] = POST_EX(SVM_EXIT_MSR),
4012 [x86_intercept_rdtsc] = POST_EX(SVM_EXIT_RDTSC),
4013 [x86_intercept_rdmsr] = POST_EX(SVM_EXIT_MSR),
4014 [x86_intercept_rdpmc] = POST_EX(SVM_EXIT_RDPMC),
4015 [x86_intercept_cpuid] = PRE_EX(SVM_EXIT_CPUID),
4016 [x86_intercept_rsm] = PRE_EX(SVM_EXIT_RSM),
4017 [x86_intercept_pause] = PRE_EX(SVM_EXIT_PAUSE),
4018 [x86_intercept_pushf] = PRE_EX(SVM_EXIT_PUSHF),
4019 [x86_intercept_popf] = PRE_EX(SVM_EXIT_POPF),
4020 [x86_intercept_intn] = PRE_EX(SVM_EXIT_SWINT),
4021 [x86_intercept_iret] = PRE_EX(SVM_EXIT_IRET),
4022 [x86_intercept_icebp] = PRE_EX(SVM_EXIT_ICEBP),
4023 [x86_intercept_hlt] = POST_EX(SVM_EXIT_HLT),
4024 [x86_intercept_in] = POST_EX(SVM_EXIT_IOIO),
4025 [x86_intercept_ins] = POST_EX(SVM_EXIT_IOIO),
4026 [x86_intercept_out] = POST_EX(SVM_EXIT_IOIO),
4027 [x86_intercept_outs] = POST_EX(SVM_EXIT_IOIO),
4028 [x86_intercept_xsetbv] = PRE_EX(SVM_EXIT_XSETBV),
4035 static int svm_check_intercept(struct kvm_vcpu *vcpu,
4036 struct x86_instruction_info *info,
4037 enum x86_intercept_stage stage,
4038 struct x86_exception *exception)
4040 struct vcpu_svm *svm = to_svm(vcpu);
4041 int vmexit, ret = X86EMUL_CONTINUE;
4042 struct __x86_intercept icpt_info;
4043 struct vmcb *vmcb = svm->vmcb;
4045 if (info->intercept >= ARRAY_SIZE(x86_intercept_map))
4048 icpt_info = x86_intercept_map[info->intercept];
4050 if (stage != icpt_info.stage)
4053 switch (icpt_info.exit_code) {
4054 case SVM_EXIT_READ_CR0:
4055 if (info->intercept == x86_intercept_cr_read)
4056 icpt_info.exit_code += info->modrm_reg;
4058 case SVM_EXIT_WRITE_CR0: {
4059 unsigned long cr0, val;
4061 if (info->intercept == x86_intercept_cr_write)
4062 icpt_info.exit_code += info->modrm_reg;
4064 if (icpt_info.exit_code != SVM_EXIT_WRITE_CR0 ||
4065 info->intercept == x86_intercept_clts)
4068 if (!(vmcb_is_intercept(&svm->nested.ctl,
4069 INTERCEPT_SELECTIVE_CR0)))
4072 cr0 = vcpu->arch.cr0 & ~SVM_CR0_SELECTIVE_MASK;
4073 val = info->src_val & ~SVM_CR0_SELECTIVE_MASK;
4075 if (info->intercept == x86_intercept_lmsw) {
4078 /* lmsw can't clear PE - catch this here */
4079 if (cr0 & X86_CR0_PE)
4084 icpt_info.exit_code = SVM_EXIT_CR0_SEL_WRITE;
4088 case SVM_EXIT_READ_DR0:
4089 case SVM_EXIT_WRITE_DR0:
4090 icpt_info.exit_code += info->modrm_reg;
4093 if (info->intercept == x86_intercept_wrmsr)
4094 vmcb->control.exit_info_1 = 1;
4096 vmcb->control.exit_info_1 = 0;
4098 case SVM_EXIT_PAUSE:
4100 * We get this for NOP only, but pause
4101 * is rep not, check this here
4103 if (info->rep_prefix != REPE_PREFIX)
4106 case SVM_EXIT_IOIO: {
4110 if (info->intercept == x86_intercept_in ||
4111 info->intercept == x86_intercept_ins) {
4112 exit_info = ((info->src_val & 0xffff) << 16) |
4114 bytes = info->dst_bytes;
4116 exit_info = (info->dst_val & 0xffff) << 16;
4117 bytes = info->src_bytes;
4120 if (info->intercept == x86_intercept_outs ||
4121 info->intercept == x86_intercept_ins)
4122 exit_info |= SVM_IOIO_STR_MASK;
4124 if (info->rep_prefix)
4125 exit_info |= SVM_IOIO_REP_MASK;
4127 bytes = min(bytes, 4u);
4129 exit_info |= bytes << SVM_IOIO_SIZE_SHIFT;
4131 exit_info |= (u32)info->ad_bytes << (SVM_IOIO_ASIZE_SHIFT - 1);
4133 vmcb->control.exit_info_1 = exit_info;
4134 vmcb->control.exit_info_2 = info->next_rip;
4142 /* TODO: Advertise NRIPS to guest hypervisor unconditionally */
4143 if (static_cpu_has(X86_FEATURE_NRIPS))
4144 vmcb->control.next_rip = info->next_rip;
4145 vmcb->control.exit_code = icpt_info.exit_code;
4146 vmexit = nested_svm_exit_handled(svm);
4148 ret = (vmexit == NESTED_EXIT_DONE) ? X86EMUL_INTERCEPTED
4155 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
4159 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
4161 if (!kvm_pause_in_guest(vcpu->kvm))
4162 shrink_ple_window(vcpu);
4165 static void svm_setup_mce(struct kvm_vcpu *vcpu)
4167 /* [63:9] are reserved. */
4168 vcpu->arch.mcg_cap &= 0x1ff;
4171 bool svm_smi_blocked(struct kvm_vcpu *vcpu)
4173 struct vcpu_svm *svm = to_svm(vcpu);
4175 /* Per APM Vol.2 15.22.2 "Response to SMI" */
4179 return is_smm(vcpu);
4182 static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
4184 struct vcpu_svm *svm = to_svm(vcpu);
4185 if (svm->nested.nested_run_pending)
4188 /* An SMI must not be injected into L2 if it's supposed to VM-Exit. */
4189 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_smi(svm))
4192 return !svm_smi_blocked(vcpu);
4195 static int svm_pre_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
4197 struct vcpu_svm *svm = to_svm(vcpu);
4200 if (is_guest_mode(vcpu)) {
4201 /* FED8h - SVM Guest */
4202 put_smstate(u64, smstate, 0x7ed8, 1);
4203 /* FEE0h - SVM Guest VMCB Physical Address */
4204 put_smstate(u64, smstate, 0x7ee0, svm->nested.vmcb12_gpa);
4206 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4207 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4208 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4210 ret = nested_svm_vmexit(svm);
4217 static int svm_pre_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
4219 struct vcpu_svm *svm = to_svm(vcpu);
4220 struct kvm_host_map map;
4223 if (guest_cpuid_has(vcpu, X86_FEATURE_LM)) {
4224 u64 saved_efer = GET_SMSTATE(u64, smstate, 0x7ed0);
4225 u64 guest = GET_SMSTATE(u64, smstate, 0x7ed8);
4226 u64 vmcb12_gpa = GET_SMSTATE(u64, smstate, 0x7ee0);
4229 if (!guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4232 if (!(saved_efer & EFER_SVME))
4235 if (kvm_vcpu_map(&svm->vcpu,
4236 gpa_to_gfn(vmcb12_gpa), &map) == -EINVAL)
4239 if (svm_allocate_nested(svm))
4242 ret = enter_svm_guest_mode(svm, vmcb12_gpa, map.hva);
4243 kvm_vcpu_unmap(&svm->vcpu, &map, true);
4250 static void enable_smi_window(struct kvm_vcpu *vcpu)
4252 struct vcpu_svm *svm = to_svm(vcpu);
4254 if (!gif_set(svm)) {
4255 if (vgif_enabled(svm))
4256 svm_set_intercept(svm, INTERCEPT_STGI);
4257 /* STGI will cause a vm exit */
4259 /* We must be in SMM; RSM will cause a vmexit anyway. */
4263 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, void *insn, int insn_len)
4265 bool smep, smap, is_user;
4269 * When the guest is an SEV-ES guest, emulation is not possible.
4271 if (sev_es_guest(vcpu->kvm))
4275 * Detect and workaround Errata 1096 Fam_17h_00_0Fh.
4278 * When CPU raise #NPF on guest data access and vCPU CR4.SMAP=1, it is
4279 * possible that CPU microcode implementing DecodeAssist will fail
4280 * to read bytes of instruction which caused #NPF. In this case,
4281 * GuestIntrBytes field of the VMCB on a VMEXIT will incorrectly
4282 * return 0 instead of the correct guest instruction bytes.
4284 * This happens because CPU microcode reading instruction bytes
4285 * uses a special opcode which attempts to read data using CPL=0
4286 * priviledges. The microcode reads CS:RIP and if it hits a SMAP
4287 * fault, it gives up and returns no instruction bytes.
4290 * We reach here in case CPU supports DecodeAssist, raised #NPF and
4291 * returned 0 in GuestIntrBytes field of the VMCB.
4292 * First, errata can only be triggered in case vCPU CR4.SMAP=1.
4293 * Second, if vCPU CR4.SMEP=1, errata could only be triggered
4294 * in case vCPU CPL==3 (Because otherwise guest would have triggered
4295 * a SMEP fault instead of #NPF).
4296 * Otherwise, vCPU CR4.SMEP=0, errata could be triggered by any vCPU CPL.
4297 * As most guests enable SMAP if they have also enabled SMEP, use above
4298 * logic in order to attempt minimize false-positive of detecting errata
4299 * while still preserving all cases semantic correctness.
4302 * To determine what instruction the guest was executing, the hypervisor
4303 * will have to decode the instruction at the instruction pointer.
4305 * In non SEV guest, hypervisor will be able to read the guest
4306 * memory to decode the instruction pointer when insn_len is zero
4307 * so we return true to indicate that decoding is possible.
4309 * But in the SEV guest, the guest memory is encrypted with the
4310 * guest specific key and hypervisor will not be able to decode the
4311 * instruction pointer so we will not able to workaround it. Lets
4312 * print the error and request to kill the guest.
4314 if (likely(!insn || insn_len))
4318 * If RIP is invalid, go ahead with emulation which will cause an
4319 * internal error exit.
4321 if (!kvm_vcpu_gfn_to_memslot(vcpu, kvm_rip_read(vcpu) >> PAGE_SHIFT))
4324 cr4 = kvm_read_cr4(vcpu);
4325 smep = cr4 & X86_CR4_SMEP;
4326 smap = cr4 & X86_CR4_SMAP;
4327 is_user = svm_get_cpl(vcpu) == 3;
4328 if (smap && (!smep || is_user)) {
4329 if (!sev_guest(vcpu->kvm))
4332 pr_err_ratelimited("KVM: SEV Guest triggered AMD Erratum 1096\n");
4333 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
4339 static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
4341 struct vcpu_svm *svm = to_svm(vcpu);
4344 * TODO: Last condition latch INIT signals on vCPU when
4345 * vCPU is in guest-mode and vmcb12 defines intercept on INIT.
4346 * To properly emulate the INIT intercept,
4347 * svm_check_nested_events() should call nested_svm_vmexit()
4348 * if an INIT signal is pending.
4350 return !gif_set(svm) ||
4351 (vmcb_is_intercept(&svm->vmcb->control, INTERCEPT_INIT));
4354 static void svm_vm_destroy(struct kvm *kvm)
4356 avic_vm_destroy(kvm);
4357 sev_vm_destroy(kvm);
4360 static int svm_vm_init(struct kvm *kvm)
4362 if (!pause_filter_count || !pause_filter_thresh)
4363 kvm->arch.pause_in_guest = true;
4366 int ret = avic_vm_init(kvm);
4371 kvm_apicv_init(kvm, avic);
4375 static struct kvm_x86_ops svm_x86_ops __initdata = {
4376 .hardware_unsetup = svm_hardware_teardown,
4377 .hardware_enable = svm_hardware_enable,
4378 .hardware_disable = svm_hardware_disable,
4379 .cpu_has_accelerated_tpr = svm_cpu_has_accelerated_tpr,
4380 .has_emulated_msr = svm_has_emulated_msr,
4382 .vcpu_create = svm_create_vcpu,
4383 .vcpu_free = svm_free_vcpu,
4384 .vcpu_reset = svm_vcpu_reset,
4386 .vm_size = sizeof(struct kvm_svm),
4387 .vm_init = svm_vm_init,
4388 .vm_destroy = svm_vm_destroy,
4390 .prepare_guest_switch = svm_prepare_guest_switch,
4391 .vcpu_load = svm_vcpu_load,
4392 .vcpu_put = svm_vcpu_put,
4393 .vcpu_blocking = svm_vcpu_blocking,
4394 .vcpu_unblocking = svm_vcpu_unblocking,
4396 .update_exception_bitmap = update_exception_bitmap,
4397 .get_msr_feature = svm_get_msr_feature,
4398 .get_msr = svm_get_msr,
4399 .set_msr = svm_set_msr,
4400 .get_segment_base = svm_get_segment_base,
4401 .get_segment = svm_get_segment,
4402 .set_segment = svm_set_segment,
4403 .get_cpl = svm_get_cpl,
4404 .get_cs_db_l_bits = kvm_get_cs_db_l_bits,
4405 .set_cr0 = svm_set_cr0,
4406 .is_valid_cr4 = svm_is_valid_cr4,
4407 .set_cr4 = svm_set_cr4,
4408 .set_efer = svm_set_efer,
4409 .get_idt = svm_get_idt,
4410 .set_idt = svm_set_idt,
4411 .get_gdt = svm_get_gdt,
4412 .set_gdt = svm_set_gdt,
4413 .set_dr7 = svm_set_dr7,
4414 .sync_dirty_debug_regs = svm_sync_dirty_debug_regs,
4415 .cache_reg = svm_cache_reg,
4416 .get_rflags = svm_get_rflags,
4417 .set_rflags = svm_set_rflags,
4419 .tlb_flush_all = svm_flush_tlb,
4420 .tlb_flush_current = svm_flush_tlb,
4421 .tlb_flush_gva = svm_flush_tlb_gva,
4422 .tlb_flush_guest = svm_flush_tlb,
4424 .run = svm_vcpu_run,
4425 .handle_exit = handle_exit,
4426 .skip_emulated_instruction = skip_emulated_instruction,
4427 .update_emulated_instruction = NULL,
4428 .set_interrupt_shadow = svm_set_interrupt_shadow,
4429 .get_interrupt_shadow = svm_get_interrupt_shadow,
4430 .patch_hypercall = svm_patch_hypercall,
4431 .set_irq = svm_set_irq,
4432 .set_nmi = svm_inject_nmi,
4433 .queue_exception = svm_queue_exception,
4434 .cancel_injection = svm_cancel_injection,
4435 .interrupt_allowed = svm_interrupt_allowed,
4436 .nmi_allowed = svm_nmi_allowed,
4437 .get_nmi_mask = svm_get_nmi_mask,
4438 .set_nmi_mask = svm_set_nmi_mask,
4439 .enable_nmi_window = enable_nmi_window,
4440 .enable_irq_window = enable_irq_window,
4441 .update_cr8_intercept = update_cr8_intercept,
4442 .set_virtual_apic_mode = svm_set_virtual_apic_mode,
4443 .refresh_apicv_exec_ctrl = svm_refresh_apicv_exec_ctrl,
4444 .check_apicv_inhibit_reasons = svm_check_apicv_inhibit_reasons,
4445 .pre_update_apicv_exec_ctrl = svm_pre_update_apicv_exec_ctrl,
4446 .load_eoi_exitmap = svm_load_eoi_exitmap,
4447 .hwapic_irr_update = svm_hwapic_irr_update,
4448 .hwapic_isr_update = svm_hwapic_isr_update,
4449 .sync_pir_to_irr = kvm_lapic_find_highest_irr,
4450 .apicv_post_state_restore = avic_post_state_restore,
4452 .set_tss_addr = svm_set_tss_addr,
4453 .set_identity_map_addr = svm_set_identity_map_addr,
4454 .get_mt_mask = svm_get_mt_mask,
4456 .get_exit_info = svm_get_exit_info,
4458 .vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
4460 .has_wbinvd_exit = svm_has_wbinvd_exit,
4462 .write_l1_tsc_offset = svm_write_l1_tsc_offset,
4464 .load_mmu_pgd = svm_load_mmu_pgd,
4466 .check_intercept = svm_check_intercept,
4467 .handle_exit_irqoff = svm_handle_exit_irqoff,
4469 .request_immediate_exit = __kvm_request_immediate_exit,
4471 .sched_in = svm_sched_in,
4473 .pmu_ops = &amd_pmu_ops,
4474 .nested_ops = &svm_nested_ops,
4476 .deliver_posted_interrupt = svm_deliver_avic_intr,
4477 .dy_apicv_has_pending_interrupt = svm_dy_apicv_has_pending_interrupt,
4478 .update_pi_irte = svm_update_pi_irte,
4479 .setup_mce = svm_setup_mce,
4481 .smi_allowed = svm_smi_allowed,
4482 .pre_enter_smm = svm_pre_enter_smm,
4483 .pre_leave_smm = svm_pre_leave_smm,
4484 .enable_smi_window = enable_smi_window,
4486 .mem_enc_op = svm_mem_enc_op,
4487 .mem_enc_reg_region = svm_register_enc_region,
4488 .mem_enc_unreg_region = svm_unregister_enc_region,
4490 .can_emulate_instruction = svm_can_emulate_instruction,
4492 .apic_init_signal_blocked = svm_apic_init_signal_blocked,
4494 .msr_filter_changed = svm_msr_filter_changed,
4495 .complete_emulated_msr = svm_complete_emulated_msr,
4498 static struct kvm_x86_init_ops svm_init_ops __initdata = {
4499 .cpu_has_kvm_support = has_svm,
4500 .disabled_by_bios = is_disabled,
4501 .hardware_setup = svm_hardware_setup,
4502 .check_processor_compatibility = svm_check_processor_compat,
4504 .runtime_ops = &svm_x86_ops,
4507 static int __init svm_init(void)
4509 __unused_size_checks();
4511 return kvm_init(&svm_init_ops, sizeof(struct vcpu_svm),
4512 __alignof__(struct vcpu_svm), THIS_MODULE);
4515 static void __exit svm_exit(void)
4520 module_init(svm_init)
4521 module_exit(svm_exit)