1 #define pr_fmt(fmt) "SVM: " fmt
3 #include <linux/kvm_host.h>
7 #include "kvm_cache_regs.h"
12 #include <linux/module.h>
13 #include <linux/mod_devicetable.h>
14 #include <linux/kernel.h>
15 #include <linux/vmalloc.h>
16 #include <linux/highmem.h>
17 #include <linux/amd-iommu.h>
18 #include <linux/sched.h>
19 #include <linux/trace_events.h>
20 #include <linux/slab.h>
21 #include <linux/hashtable.h>
22 #include <linux/objtool.h>
23 #include <linux/psp-sev.h>
24 #include <linux/file.h>
25 #include <linux/pagemap.h>
26 #include <linux/swap.h>
27 #include <linux/rwsem.h>
30 #include <asm/perf_event.h>
31 #include <asm/tlbflush.h>
33 #include <asm/debugreg.h>
34 #include <asm/kvm_para.h>
35 #include <asm/irq_remapping.h>
36 #include <asm/spec-ctrl.h>
37 #include <asm/cpu_device_id.h>
38 #include <asm/traps.h>
40 #include <asm/virtext.h>
46 #define __ex(x) __kvm_handle_fault_on_reboot(x)
48 MODULE_AUTHOR("Qumranet");
49 MODULE_LICENSE("GPL");
52 static const struct x86_cpu_id svm_cpu_id[] = {
53 X86_MATCH_FEATURE(X86_FEATURE_SVM, NULL),
56 MODULE_DEVICE_TABLE(x86cpu, svm_cpu_id);
59 #define SEG_TYPE_LDT 2
60 #define SEG_TYPE_BUSY_TSS16 3
62 #define SVM_FEATURE_LBRV (1 << 1)
63 #define SVM_FEATURE_SVML (1 << 2)
64 #define SVM_FEATURE_TSC_RATE (1 << 4)
65 #define SVM_FEATURE_VMCB_CLEAN (1 << 5)
66 #define SVM_FEATURE_FLUSH_ASID (1 << 6)
67 #define SVM_FEATURE_DECODE_ASSIST (1 << 7)
68 #define SVM_FEATURE_PAUSE_FILTER (1 << 10)
70 #define DEBUGCTL_RESERVED_BITS (~(0x3fULL))
72 #define TSC_RATIO_RSVD 0xffffff0000000000ULL
73 #define TSC_RATIO_MIN 0x0000000000000001ULL
74 #define TSC_RATIO_MAX 0x000000ffffffffffULL
76 static bool erratum_383_found __read_mostly;
78 u32 msrpm_offsets[MSRPM_OFFSETS] __read_mostly;
81 * Set osvw_len to higher value when updated Revision Guides
82 * are published and we know what the new status bits are
84 static uint64_t osvw_len = 4, osvw_status;
86 static DEFINE_PER_CPU(u64, current_tsc_ratio);
87 #define TSC_RATIO_DEFAULT 0x0100000000ULL
89 static const struct svm_direct_access_msrs {
90 u32 index; /* Index of the MSR */
91 bool always; /* True if intercept is initially cleared */
92 } direct_access_msrs[MAX_DIRECT_ACCESS_MSRS] = {
93 { .index = MSR_STAR, .always = true },
94 { .index = MSR_IA32_SYSENTER_CS, .always = true },
95 { .index = MSR_IA32_SYSENTER_EIP, .always = false },
96 { .index = MSR_IA32_SYSENTER_ESP, .always = false },
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_EFER, .always = false },
112 { .index = MSR_IA32_CR_PAT, .always = false },
113 { .index = MSR_AMD64_SEV_ES_GHCB, .always = true },
114 { .index = MSR_INVALID, .always = false },
118 * These 2 parameters are used to config the controls for Pause-Loop Exiting:
119 * pause_filter_count: On processors that support Pause filtering(indicated
120 * by CPUID Fn8000_000A_EDX), the VMCB provides a 16 bit pause filter
121 * count value. On VMRUN this value is loaded into an internal counter.
122 * Each time a pause instruction is executed, this counter is decremented
123 * until it reaches zero at which time a #VMEXIT is generated if pause
124 * intercept is enabled. Refer to AMD APM Vol 2 Section 15.14.4 Pause
125 * Intercept Filtering for more details.
126 * This also indicate if ple logic enabled.
128 * pause_filter_thresh: In addition, some processor families support advanced
129 * pause filtering (indicated by CPUID Fn8000_000A_EDX) upper bound on
130 * the amount of time a guest is allowed to execute in a pause loop.
131 * In this mode, a 16-bit pause filter threshold field is added in the
132 * VMCB. The threshold value is a cycle count that is used to reset the
133 * pause counter. As with simple pause filtering, VMRUN loads the pause
134 * count value from VMCB into an internal counter. Then, on each pause
135 * instruction the hardware checks the elapsed number of cycles since
136 * the most recent pause instruction against the pause filter threshold.
137 * If the elapsed cycle count is greater than the pause filter threshold,
138 * then the internal pause count is reloaded from the VMCB and execution
139 * continues. If the elapsed cycle count is less than the pause filter
140 * threshold, then the internal pause count is decremented. If the count
141 * value is less than zero and PAUSE intercept is enabled, a #VMEXIT is
142 * triggered. If advanced pause filtering is supported and pause filter
143 * threshold field is set to zero, the filter will operate in the simpler,
147 static unsigned short pause_filter_thresh = KVM_DEFAULT_PLE_GAP;
148 module_param(pause_filter_thresh, ushort, 0444);
150 static unsigned short pause_filter_count = KVM_SVM_DEFAULT_PLE_WINDOW;
151 module_param(pause_filter_count, ushort, 0444);
153 /* Default doubles per-vcpu window every exit. */
154 static unsigned short pause_filter_count_grow = KVM_DEFAULT_PLE_WINDOW_GROW;
155 module_param(pause_filter_count_grow, ushort, 0444);
157 /* Default resets per-vcpu window every exit to pause_filter_count. */
158 static unsigned short pause_filter_count_shrink = KVM_DEFAULT_PLE_WINDOW_SHRINK;
159 module_param(pause_filter_count_shrink, ushort, 0444);
161 /* Default is to compute the maximum so we can never overflow. */
162 static unsigned short pause_filter_count_max = KVM_SVM_DEFAULT_PLE_WINDOW_MAX;
163 module_param(pause_filter_count_max, ushort, 0444);
166 * Use nested page tables by default. Note, NPT may get forced off by
167 * svm_hardware_setup() if it's unsupported by hardware or the host kernel.
169 bool npt_enabled = true;
170 module_param_named(npt, npt_enabled, bool, 0444);
172 /* allow nested virtualization in KVM/SVM */
173 static int nested = true;
174 module_param(nested, int, S_IRUGO);
176 /* enable/disable Next RIP Save */
177 static int nrips = true;
178 module_param(nrips, int, 0444);
180 /* enable/disable Virtual VMLOAD VMSAVE */
181 static int vls = true;
182 module_param(vls, int, 0444);
184 /* enable/disable Virtual GIF */
185 static int vgif = true;
186 module_param(vgif, int, 0444);
188 /* enable/disable SEV support */
189 int sev = IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT_ACTIVE_BY_DEFAULT);
190 module_param(sev, int, 0444);
192 /* enable/disable SEV-ES support */
193 int sev_es = IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT_ACTIVE_BY_DEFAULT);
194 module_param(sev_es, int, 0444);
196 bool __read_mostly dump_invalid_vmcb;
197 module_param(dump_invalid_vmcb, bool, 0644);
199 static bool svm_gp_erratum_intercept = true;
201 static u8 rsm_ins_bytes[] = "\x0f\xaa";
203 static unsigned long iopm_base;
205 struct kvm_ldttss_desc {
208 unsigned base1:8, type:5, dpl:2, p:1;
209 unsigned limit1:4, zero0:3, g:1, base2:8;
212 } __attribute__((packed));
214 DEFINE_PER_CPU(struct svm_cpu_data *, svm_data);
217 * Only MSR_TSC_AUX is switched via the user return hook. EFER is switched via
218 * the VMCB, and the SYSCALL/SYSENTER MSRs are handled by VMLOAD/VMSAVE.
220 * RDTSCP and RDPID are not used in the kernel, specifically to allow KVM to
221 * defer the restoration of TSC_AUX until the CPU returns to userspace.
223 #define TSC_AUX_URET_SLOT 0
225 static const u32 msrpm_ranges[] = {0, 0xc0000000, 0xc0010000};
227 #define NUM_MSR_MAPS ARRAY_SIZE(msrpm_ranges)
228 #define MSRS_RANGE_SIZE 2048
229 #define MSRS_IN_RANGE (MSRS_RANGE_SIZE * 8 / 2)
231 u32 svm_msrpm_offset(u32 msr)
236 for (i = 0; i < NUM_MSR_MAPS; i++) {
237 if (msr < msrpm_ranges[i] ||
238 msr >= msrpm_ranges[i] + MSRS_IN_RANGE)
241 offset = (msr - msrpm_ranges[i]) / 4; /* 4 msrs per u8 */
242 offset += (i * MSRS_RANGE_SIZE); /* add range offset */
244 /* Now we have the u8 offset - but need the u32 offset */
248 /* MSR not in any range */
252 #define MAX_INST_SIZE 15
254 static int get_max_npt_level(void)
257 return PT64_ROOT_4LEVEL;
259 return PT32E_ROOT_LEVEL;
263 int svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)
265 struct vcpu_svm *svm = to_svm(vcpu);
266 u64 old_efer = vcpu->arch.efer;
267 vcpu->arch.efer = efer;
270 /* Shadow paging assumes NX to be available. */
273 if (!(efer & EFER_LMA))
277 if ((old_efer & EFER_SVME) != (efer & EFER_SVME)) {
278 if (!(efer & EFER_SVME)) {
279 svm_leave_nested(svm);
280 svm_set_gif(svm, true);
281 /* #GP intercept is still needed for vmware backdoor */
282 if (!enable_vmware_backdoor)
283 clr_exception_intercept(svm, GP_VECTOR);
286 * Free the nested guest state, unless we are in SMM.
287 * In this case we will return to the nested guest
288 * as soon as we leave SMM.
291 svm_free_nested(svm);
294 int ret = svm_allocate_nested(svm);
297 vcpu->arch.efer = old_efer;
301 if (svm_gp_erratum_intercept)
302 set_exception_intercept(svm, GP_VECTOR);
306 svm->vmcb->save.efer = efer | EFER_SVME;
307 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
311 static int is_external_interrupt(u32 info)
313 info &= SVM_EVTINJ_TYPE_MASK | SVM_EVTINJ_VALID;
314 return info == (SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR);
317 static u32 svm_get_interrupt_shadow(struct kvm_vcpu *vcpu)
319 struct vcpu_svm *svm = to_svm(vcpu);
322 if (svm->vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK)
323 ret = KVM_X86_SHADOW_INT_STI | KVM_X86_SHADOW_INT_MOV_SS;
327 static void svm_set_interrupt_shadow(struct kvm_vcpu *vcpu, int mask)
329 struct vcpu_svm *svm = to_svm(vcpu);
332 svm->vmcb->control.int_state &= ~SVM_INTERRUPT_SHADOW_MASK;
334 svm->vmcb->control.int_state |= SVM_INTERRUPT_SHADOW_MASK;
338 static int skip_emulated_instruction(struct kvm_vcpu *vcpu)
340 struct vcpu_svm *svm = to_svm(vcpu);
343 * SEV-ES does not expose the next RIP. The RIP update is controlled by
344 * the type of exit and the #VC handler in the guest.
346 if (sev_es_guest(vcpu->kvm))
349 if (nrips && svm->vmcb->control.next_rip != 0) {
350 WARN_ON_ONCE(!static_cpu_has(X86_FEATURE_NRIPS));
351 svm->next_rip = svm->vmcb->control.next_rip;
354 if (!svm->next_rip) {
355 if (!kvm_emulate_instruction(vcpu, EMULTYPE_SKIP))
358 kvm_rip_write(vcpu, svm->next_rip);
362 svm_set_interrupt_shadow(vcpu, 0);
367 static void svm_queue_exception(struct kvm_vcpu *vcpu)
369 struct vcpu_svm *svm = to_svm(vcpu);
370 unsigned nr = vcpu->arch.exception.nr;
371 bool has_error_code = vcpu->arch.exception.has_error_code;
372 u32 error_code = vcpu->arch.exception.error_code;
374 kvm_deliver_exception_payload(vcpu);
376 if (nr == BP_VECTOR && !nrips) {
377 unsigned long rip, old_rip = kvm_rip_read(vcpu);
380 * For guest debugging where we have to reinject #BP if some
381 * INT3 is guest-owned:
382 * Emulate nRIP by moving RIP forward. Will fail if injection
383 * raises a fault that is not intercepted. Still better than
384 * failing in all cases.
386 (void)skip_emulated_instruction(vcpu);
387 rip = kvm_rip_read(vcpu);
388 svm->int3_rip = rip + svm->vmcb->save.cs.base;
389 svm->int3_injected = rip - old_rip;
392 svm->vmcb->control.event_inj = nr
394 | (has_error_code ? SVM_EVTINJ_VALID_ERR : 0)
395 | SVM_EVTINJ_TYPE_EXEPT;
396 svm->vmcb->control.event_inj_err = error_code;
399 static void svm_init_erratum_383(void)
405 if (!static_cpu_has_bug(X86_BUG_AMD_TLB_MMATCH))
408 /* Use _safe variants to not break nested virtualization */
409 val = native_read_msr_safe(MSR_AMD64_DC_CFG, &err);
415 low = lower_32_bits(val);
416 high = upper_32_bits(val);
418 native_write_msr_safe(MSR_AMD64_DC_CFG, low, high);
420 erratum_383_found = true;
423 static void svm_init_osvw(struct kvm_vcpu *vcpu)
426 * Guests should see errata 400 and 415 as fixed (assuming that
427 * HLT and IO instructions are intercepted).
429 vcpu->arch.osvw.length = (osvw_len >= 3) ? (osvw_len) : 3;
430 vcpu->arch.osvw.status = osvw_status & ~(6ULL);
433 * By increasing VCPU's osvw.length to 3 we are telling the guest that
434 * all osvw.status bits inside that length, including bit 0 (which is
435 * reserved for erratum 298), are valid. However, if host processor's
436 * osvw_len is 0 then osvw_status[0] carries no information. We need to
437 * be conservative here and therefore we tell the guest that erratum 298
438 * is present (because we really don't know).
440 if (osvw_len == 0 && boot_cpu_data.x86 == 0x10)
441 vcpu->arch.osvw.status |= 1;
444 static int has_svm(void)
448 if (!cpu_has_svm(&msg)) {
449 printk(KERN_INFO "has_svm: %s\n", msg);
454 pr_info("KVM is unsupported when running as an SEV guest\n");
461 static void svm_hardware_disable(void)
463 /* Make sure we clean up behind us */
464 if (static_cpu_has(X86_FEATURE_TSCRATEMSR))
465 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
469 amd_pmu_disable_virt();
472 static int svm_hardware_enable(void)
475 struct svm_cpu_data *sd;
477 struct desc_struct *gdt;
478 int me = raw_smp_processor_id();
480 rdmsrl(MSR_EFER, efer);
481 if (efer & EFER_SVME)
485 pr_err("%s: err EOPNOTSUPP on %d\n", __func__, me);
488 sd = per_cpu(svm_data, me);
490 pr_err("%s: svm_data is NULL on %d\n", __func__, me);
494 sd->asid_generation = 1;
495 sd->max_asid = cpuid_ebx(SVM_CPUID_FUNC) - 1;
496 sd->next_asid = sd->max_asid + 1;
497 sd->min_asid = max_sev_asid + 1;
499 gdt = get_current_gdt_rw();
500 sd->tss_desc = (struct kvm_ldttss_desc *)(gdt + GDT_ENTRY_TSS);
502 wrmsrl(MSR_EFER, efer | EFER_SVME);
504 wrmsrl(MSR_VM_HSAVE_PA, __sme_page_pa(sd->save_area));
506 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
507 wrmsrl(MSR_AMD64_TSC_RATIO, TSC_RATIO_DEFAULT);
508 __this_cpu_write(current_tsc_ratio, TSC_RATIO_DEFAULT);
515 * Note that it is possible to have a system with mixed processor
516 * revisions and therefore different OSVW bits. If bits are not the same
517 * on different processors then choose the worst case (i.e. if erratum
518 * is present on one processor and not on another then assume that the
519 * erratum is present everywhere).
521 if (cpu_has(&boot_cpu_data, X86_FEATURE_OSVW)) {
522 uint64_t len, status = 0;
525 len = native_read_msr_safe(MSR_AMD64_OSVW_ID_LENGTH, &err);
527 status = native_read_msr_safe(MSR_AMD64_OSVW_STATUS,
531 osvw_status = osvw_len = 0;
535 osvw_status |= status;
536 osvw_status &= (1ULL << osvw_len) - 1;
539 osvw_status = osvw_len = 0;
541 svm_init_erratum_383();
543 amd_pmu_enable_virt();
548 static void svm_cpu_uninit(int cpu)
550 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
555 per_cpu(svm_data, cpu) = NULL;
556 kfree(sd->sev_vmcbs);
557 __free_page(sd->save_area);
561 static int svm_cpu_init(int cpu)
563 struct svm_cpu_data *sd;
565 sd = kzalloc(sizeof(struct svm_cpu_data), GFP_KERNEL);
569 sd->save_area = alloc_page(GFP_KERNEL);
572 clear_page(page_address(sd->save_area));
574 if (svm_sev_enabled()) {
575 sd->sev_vmcbs = kmalloc_array(max_sev_asid + 1,
582 per_cpu(svm_data, cpu) = sd;
587 __free_page(sd->save_area);
594 static int direct_access_msr_slot(u32 msr)
598 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++)
599 if (direct_access_msrs[i].index == msr)
605 static void set_shadow_msr_intercept(struct kvm_vcpu *vcpu, u32 msr, int read,
608 struct vcpu_svm *svm = to_svm(vcpu);
609 int slot = direct_access_msr_slot(msr);
614 /* Set the shadow bitmaps to the desired intercept states */
616 set_bit(slot, svm->shadow_msr_intercept.read);
618 clear_bit(slot, svm->shadow_msr_intercept.read);
621 set_bit(slot, svm->shadow_msr_intercept.write);
623 clear_bit(slot, svm->shadow_msr_intercept.write);
626 static bool valid_msr_intercept(u32 index)
628 return direct_access_msr_slot(index) != -ENOENT;
631 static bool msr_write_intercepted(struct kvm_vcpu *vcpu, u32 msr)
638 msrpm = is_guest_mode(vcpu) ? to_svm(vcpu)->nested.msrpm:
641 offset = svm_msrpm_offset(msr);
642 bit_write = 2 * (msr & 0x0f) + 1;
645 BUG_ON(offset == MSR_INVALID);
647 return !!test_bit(bit_write, &tmp);
650 static void set_msr_interception_bitmap(struct kvm_vcpu *vcpu, u32 *msrpm,
651 u32 msr, int read, int write)
653 u8 bit_read, bit_write;
658 * If this warning triggers extend the direct_access_msrs list at the
659 * beginning of the file
661 WARN_ON(!valid_msr_intercept(msr));
663 /* Enforce non allowed MSRs to trap */
664 if (read && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_READ))
667 if (write && !kvm_msr_allowed(vcpu, msr, KVM_MSR_FILTER_WRITE))
670 offset = svm_msrpm_offset(msr);
671 bit_read = 2 * (msr & 0x0f);
672 bit_write = 2 * (msr & 0x0f) + 1;
675 BUG_ON(offset == MSR_INVALID);
677 read ? clear_bit(bit_read, &tmp) : set_bit(bit_read, &tmp);
678 write ? clear_bit(bit_write, &tmp) : set_bit(bit_write, &tmp);
683 void set_msr_interception(struct kvm_vcpu *vcpu, u32 *msrpm, u32 msr,
686 set_shadow_msr_intercept(vcpu, msr, read, write);
687 set_msr_interception_bitmap(vcpu, msrpm, msr, read, write);
690 u32 *svm_vcpu_alloc_msrpm(void)
692 unsigned int order = get_order(MSRPM_SIZE);
693 struct page *pages = alloc_pages(GFP_KERNEL_ACCOUNT, order);
699 msrpm = page_address(pages);
700 memset(msrpm, 0xff, PAGE_SIZE * (1 << order));
705 void svm_vcpu_init_msrpm(struct kvm_vcpu *vcpu, u32 *msrpm)
709 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
710 if (!direct_access_msrs[i].always)
712 set_msr_interception(vcpu, msrpm, direct_access_msrs[i].index, 1, 1);
717 void svm_vcpu_free_msrpm(u32 *msrpm)
719 __free_pages(virt_to_page(msrpm), get_order(MSRPM_SIZE));
722 static void svm_msr_filter_changed(struct kvm_vcpu *vcpu)
724 struct vcpu_svm *svm = to_svm(vcpu);
728 * Set intercept permissions for all direct access MSRs again. They
729 * will automatically get filtered through the MSR filter, so we are
730 * back in sync after this.
732 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
733 u32 msr = direct_access_msrs[i].index;
734 u32 read = test_bit(i, svm->shadow_msr_intercept.read);
735 u32 write = test_bit(i, svm->shadow_msr_intercept.write);
737 set_msr_interception_bitmap(vcpu, svm->msrpm, msr, read, write);
741 static void add_msr_offset(u32 offset)
745 for (i = 0; i < MSRPM_OFFSETS; ++i) {
747 /* Offset already in list? */
748 if (msrpm_offsets[i] == offset)
751 /* Slot used by another offset? */
752 if (msrpm_offsets[i] != MSR_INVALID)
755 /* Add offset to list */
756 msrpm_offsets[i] = offset;
762 * If this BUG triggers the msrpm_offsets table has an overflow. Just
763 * increase MSRPM_OFFSETS in this case.
768 static void init_msrpm_offsets(void)
772 memset(msrpm_offsets, 0xff, sizeof(msrpm_offsets));
774 for (i = 0; direct_access_msrs[i].index != MSR_INVALID; i++) {
777 offset = svm_msrpm_offset(direct_access_msrs[i].index);
778 BUG_ON(offset == MSR_INVALID);
780 add_msr_offset(offset);
784 static void svm_enable_lbrv(struct kvm_vcpu *vcpu)
786 struct vcpu_svm *svm = to_svm(vcpu);
788 svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
789 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
790 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
791 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
792 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
795 static void svm_disable_lbrv(struct kvm_vcpu *vcpu)
797 struct vcpu_svm *svm = to_svm(vcpu);
799 svm->vmcb->control.virt_ext &= ~LBR_CTL_ENABLE_MASK;
800 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 0, 0);
801 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 0, 0);
802 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 0, 0);
803 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 0, 0);
806 void disable_nmi_singlestep(struct vcpu_svm *svm)
808 svm->nmi_singlestep = false;
810 if (!(svm->vcpu.guest_debug & KVM_GUESTDBG_SINGLESTEP)) {
811 /* Clear our flags if they were not set by the guest */
812 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
813 svm->vmcb->save.rflags &= ~X86_EFLAGS_TF;
814 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
815 svm->vmcb->save.rflags &= ~X86_EFLAGS_RF;
819 static void grow_ple_window(struct kvm_vcpu *vcpu)
821 struct vcpu_svm *svm = to_svm(vcpu);
822 struct vmcb_control_area *control = &svm->vmcb->control;
823 int old = control->pause_filter_count;
825 control->pause_filter_count = __grow_ple_window(old,
827 pause_filter_count_grow,
828 pause_filter_count_max);
830 if (control->pause_filter_count != old) {
831 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
832 trace_kvm_ple_window_update(vcpu->vcpu_id,
833 control->pause_filter_count, old);
837 static void shrink_ple_window(struct kvm_vcpu *vcpu)
839 struct vcpu_svm *svm = to_svm(vcpu);
840 struct vmcb_control_area *control = &svm->vmcb->control;
841 int old = control->pause_filter_count;
843 control->pause_filter_count =
844 __shrink_ple_window(old,
846 pause_filter_count_shrink,
848 if (control->pause_filter_count != old) {
849 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
850 trace_kvm_ple_window_update(vcpu->vcpu_id,
851 control->pause_filter_count, old);
856 * The default MMIO mask is a single bit (excluding the present bit),
857 * which could conflict with the memory encryption bit. Check for
858 * memory encryption support and override the default MMIO mask if
859 * memory encryption is enabled.
861 static __init void svm_adjust_mmio_mask(void)
863 unsigned int enc_bit, mask_bit;
866 /* If there is no memory encryption support, use existing mask */
867 if (cpuid_eax(0x80000000) < 0x8000001f)
870 /* If memory encryption is not enabled, use existing mask */
871 rdmsrl(MSR_K8_SYSCFG, msr);
872 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
875 enc_bit = cpuid_ebx(0x8000001f) & 0x3f;
876 mask_bit = boot_cpu_data.x86_phys_bits;
878 /* Increment the mask bit if it is the same as the encryption bit */
879 if (enc_bit == mask_bit)
883 * If the mask bit location is below 52, then some bits above the
884 * physical addressing limit will always be reserved, so use the
885 * rsvd_bits() function to generate the mask. This mask, along with
886 * the present bit, will be used to generate a page fault with
889 * If the mask bit location is 52 (or above), then clear the mask.
891 mask = (mask_bit < 52) ? rsvd_bits(mask_bit, 51) | PT_PRESENT_MASK : 0;
893 kvm_mmu_set_mmio_spte_mask(mask, mask, PT_WRITABLE_MASK | PT_USER_MASK);
896 static void svm_hardware_teardown(void)
900 if (svm_sev_enabled())
901 sev_hardware_teardown();
903 for_each_possible_cpu(cpu)
906 __free_pages(pfn_to_page(iopm_base >> PAGE_SHIFT),
907 get_order(IOPM_SIZE));
911 static __init void svm_set_cpu_caps(void)
917 /* CPUID 0x80000001 and 0x8000000A (SVM features) */
919 kvm_cpu_cap_set(X86_FEATURE_SVM);
922 kvm_cpu_cap_set(X86_FEATURE_NRIPS);
925 kvm_cpu_cap_set(X86_FEATURE_NPT);
927 /* Nested VM can receive #VMEXIT instead of triggering #GP */
928 kvm_cpu_cap_set(X86_FEATURE_SVME_ADDR_CHK);
931 /* CPUID 0x80000008 */
932 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) ||
933 boot_cpu_has(X86_FEATURE_AMD_SSBD))
934 kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
937 static __init int svm_hardware_setup(void)
940 struct page *iopm_pages;
943 unsigned int order = get_order(IOPM_SIZE);
945 iopm_pages = alloc_pages(GFP_KERNEL, order);
950 iopm_va = page_address(iopm_pages);
951 memset(iopm_va, 0xff, PAGE_SIZE * (1 << order));
952 iopm_base = page_to_pfn(iopm_pages) << PAGE_SHIFT;
954 init_msrpm_offsets();
956 supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
958 if (boot_cpu_has(X86_FEATURE_NX))
959 kvm_enable_efer_bits(EFER_NX);
961 if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
962 kvm_enable_efer_bits(EFER_FFXSR);
964 if (boot_cpu_has(X86_FEATURE_TSCRATEMSR)) {
965 kvm_has_tsc_control = true;
966 kvm_max_tsc_scaling_ratio = TSC_RATIO_MAX;
967 kvm_tsc_scaling_ratio_frac_bits = 32;
970 if (boot_cpu_has(X86_FEATURE_RDTSCP))
971 kvm_define_user_return_msr(TSC_AUX_URET_SLOT, MSR_TSC_AUX);
973 /* Check for pause filtering support */
974 if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
975 pause_filter_count = 0;
976 pause_filter_thresh = 0;
977 } else if (!boot_cpu_has(X86_FEATURE_PFTHRESHOLD)) {
978 pause_filter_thresh = 0;
982 printk(KERN_INFO "kvm: Nested Virtualization enabled\n");
983 kvm_enable_efer_bits(EFER_SVME | EFER_LMSLE);
986 if (IS_ENABLED(CONFIG_KVM_AMD_SEV) && sev) {
987 sev_hardware_setup();
993 svm_adjust_mmio_mask();
995 for_each_possible_cpu(cpu) {
996 r = svm_cpu_init(cpu);
1002 * KVM's MMU doesn't support using 2-level paging for itself, and thus
1003 * NPT isn't supported if the host is using 2-level paging since host
1004 * CR4 is unchanged on VMRUN.
1006 if (!IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_X86_PAE))
1007 npt_enabled = false;
1009 if (!boot_cpu_has(X86_FEATURE_NPT))
1010 npt_enabled = false;
1012 kvm_configure_mmu(npt_enabled, get_max_npt_level(), PG_LEVEL_1G);
1013 pr_info("kvm: Nested Paging %sabled\n", npt_enabled ? "en" : "dis");
1016 if (!boot_cpu_has(X86_FEATURE_NRIPS))
1022 !boot_cpu_has(X86_FEATURE_AVIC) ||
1023 !IS_ENABLED(CONFIG_X86_LOCAL_APIC)) {
1026 pr_info("AVIC enabled\n");
1028 amd_iommu_register_ga_log_notifier(&avic_ga_log_notifier);
1034 !boot_cpu_has(X86_FEATURE_V_VMSAVE_VMLOAD) ||
1035 !IS_ENABLED(CONFIG_X86_64)) {
1038 pr_info("Virtual VMLOAD VMSAVE supported\n");
1042 if (boot_cpu_has(X86_FEATURE_SVME_ADDR_CHK))
1043 svm_gp_erratum_intercept = false;
1046 if (!boot_cpu_has(X86_FEATURE_VGIF))
1049 pr_info("Virtual GIF supported\n");
1055 * It seems that on AMD processors PTE's accessed bit is
1056 * being set by the CPU hardware before the NPF vmexit.
1057 * This is not expected behaviour and our tests fail because
1059 * A workaround here is to disable support for
1060 * GUEST_MAXPHYADDR < HOST_MAXPHYADDR if NPT is enabled.
1061 * In this case userspace can know if there is support using
1062 * KVM_CAP_SMALLER_MAXPHYADDR extension and decide how to handle
1064 * If future AMD CPU models change the behaviour described above,
1065 * this variable can be changed accordingly
1067 allow_smaller_maxphyaddr = !npt_enabled;
1072 svm_hardware_teardown();
1076 static void init_seg(struct vmcb_seg *seg)
1079 seg->attrib = SVM_SELECTOR_P_MASK | SVM_SELECTOR_S_MASK |
1080 SVM_SELECTOR_WRITE_MASK; /* Read/Write Data Segment */
1081 seg->limit = 0xffff;
1085 static void init_sys_seg(struct vmcb_seg *seg, uint32_t type)
1088 seg->attrib = SVM_SELECTOR_P_MASK | type;
1089 seg->limit = 0xffff;
1093 static u64 svm_write_l1_tsc_offset(struct kvm_vcpu *vcpu, u64 offset)
1095 struct vcpu_svm *svm = to_svm(vcpu);
1096 u64 g_tsc_offset = 0;
1098 if (is_guest_mode(vcpu)) {
1099 /* Write L1's TSC offset. */
1100 g_tsc_offset = svm->vmcb->control.tsc_offset -
1101 svm->vmcb01.ptr->control.tsc_offset;
1102 svm->vmcb01.ptr->control.tsc_offset = offset;
1105 trace_kvm_write_tsc_offset(vcpu->vcpu_id,
1106 svm->vmcb->control.tsc_offset - g_tsc_offset,
1109 svm->vmcb->control.tsc_offset = offset + g_tsc_offset;
1111 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
1112 return svm->vmcb->control.tsc_offset;
1115 static void svm_check_invpcid(struct vcpu_svm *svm)
1118 * Intercept INVPCID if shadow paging is enabled to sync/free shadow
1119 * roots, or if INVPCID is disabled in the guest to inject #UD.
1121 if (kvm_cpu_cap_has(X86_FEATURE_INVPCID)) {
1123 !guest_cpuid_has(&svm->vcpu, X86_FEATURE_INVPCID))
1124 svm_set_intercept(svm, INTERCEPT_INVPCID);
1126 svm_clr_intercept(svm, INTERCEPT_INVPCID);
1130 static void init_vmcb(struct kvm_vcpu *vcpu)
1132 struct vcpu_svm *svm = to_svm(vcpu);
1133 struct vmcb_control_area *control = &svm->vmcb->control;
1134 struct vmcb_save_area *save = &svm->vmcb->save;
1136 vcpu->arch.hflags = 0;
1138 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1139 svm_set_intercept(svm, INTERCEPT_CR3_READ);
1140 svm_set_intercept(svm, INTERCEPT_CR4_READ);
1141 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1142 svm_set_intercept(svm, INTERCEPT_CR3_WRITE);
1143 svm_set_intercept(svm, INTERCEPT_CR4_WRITE);
1144 if (!kvm_vcpu_apicv_active(vcpu))
1145 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
1147 set_dr_intercepts(svm);
1149 set_exception_intercept(svm, PF_VECTOR);
1150 set_exception_intercept(svm, UD_VECTOR);
1151 set_exception_intercept(svm, MC_VECTOR);
1152 set_exception_intercept(svm, AC_VECTOR);
1153 set_exception_intercept(svm, DB_VECTOR);
1155 * Guest access to VMware backdoor ports could legitimately
1156 * trigger #GP because of TSS I/O permission bitmap.
1157 * We intercept those #GP and allow access to them anyway
1160 if (enable_vmware_backdoor)
1161 set_exception_intercept(svm, GP_VECTOR);
1163 svm_set_intercept(svm, INTERCEPT_INTR);
1164 svm_set_intercept(svm, INTERCEPT_NMI);
1165 svm_set_intercept(svm, INTERCEPT_SMI);
1166 svm_set_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1167 svm_set_intercept(svm, INTERCEPT_RDPMC);
1168 svm_set_intercept(svm, INTERCEPT_CPUID);
1169 svm_set_intercept(svm, INTERCEPT_INVD);
1170 svm_set_intercept(svm, INTERCEPT_INVLPG);
1171 svm_set_intercept(svm, INTERCEPT_INVLPGA);
1172 svm_set_intercept(svm, INTERCEPT_IOIO_PROT);
1173 svm_set_intercept(svm, INTERCEPT_MSR_PROT);
1174 svm_set_intercept(svm, INTERCEPT_TASK_SWITCH);
1175 svm_set_intercept(svm, INTERCEPT_SHUTDOWN);
1176 svm_set_intercept(svm, INTERCEPT_VMRUN);
1177 svm_set_intercept(svm, INTERCEPT_VMMCALL);
1178 svm_set_intercept(svm, INTERCEPT_VMLOAD);
1179 svm_set_intercept(svm, INTERCEPT_VMSAVE);
1180 svm_set_intercept(svm, INTERCEPT_STGI);
1181 svm_set_intercept(svm, INTERCEPT_CLGI);
1182 svm_set_intercept(svm, INTERCEPT_SKINIT);
1183 svm_set_intercept(svm, INTERCEPT_WBINVD);
1184 svm_set_intercept(svm, INTERCEPT_XSETBV);
1185 svm_set_intercept(svm, INTERCEPT_RDPRU);
1186 svm_set_intercept(svm, INTERCEPT_RSM);
1188 if (!kvm_mwait_in_guest(vcpu->kvm)) {
1189 svm_set_intercept(svm, INTERCEPT_MONITOR);
1190 svm_set_intercept(svm, INTERCEPT_MWAIT);
1193 if (!kvm_hlt_in_guest(vcpu->kvm))
1194 svm_set_intercept(svm, INTERCEPT_HLT);
1196 control->iopm_base_pa = __sme_set(iopm_base);
1197 control->msrpm_base_pa = __sme_set(__pa(svm->msrpm));
1198 control->int_ctl = V_INTR_MASKING_MASK;
1200 init_seg(&save->es);
1201 init_seg(&save->ss);
1202 init_seg(&save->ds);
1203 init_seg(&save->fs);
1204 init_seg(&save->gs);
1206 save->cs.selector = 0xf000;
1207 save->cs.base = 0xffff0000;
1208 /* Executable/Readable Code Segment */
1209 save->cs.attrib = SVM_SELECTOR_READ_MASK | SVM_SELECTOR_P_MASK |
1210 SVM_SELECTOR_S_MASK | SVM_SELECTOR_CODE_MASK;
1211 save->cs.limit = 0xffff;
1213 save->gdtr.limit = 0xffff;
1214 save->idtr.limit = 0xffff;
1216 init_sys_seg(&save->ldtr, SEG_TYPE_LDT);
1217 init_sys_seg(&save->tr, SEG_TYPE_BUSY_TSS16);
1219 svm_set_cr4(vcpu, 0);
1220 svm_set_efer(vcpu, 0);
1221 save->dr6 = 0xffff0ff0;
1222 kvm_set_rflags(vcpu, X86_EFLAGS_FIXED);
1223 save->rip = 0x0000fff0;
1224 vcpu->arch.regs[VCPU_REGS_RIP] = save->rip;
1227 * svm_set_cr0() sets PG and WP and clears NW and CD on save->cr0.
1228 * It also updates the guest-visible cr0 value.
1230 svm_set_cr0(vcpu, X86_CR0_NW | X86_CR0_CD | X86_CR0_ET);
1231 kvm_mmu_reset_context(vcpu);
1233 save->cr4 = X86_CR4_PAE;
1237 /* Setup VMCB for Nested Paging */
1238 control->nested_ctl |= SVM_NESTED_CTL_NP_ENABLE;
1239 svm_clr_intercept(svm, INTERCEPT_INVLPG);
1240 clr_exception_intercept(svm, PF_VECTOR);
1241 svm_clr_intercept(svm, INTERCEPT_CR3_READ);
1242 svm_clr_intercept(svm, INTERCEPT_CR3_WRITE);
1243 save->g_pat = vcpu->arch.pat;
1247 svm->current_vmcb->asid_generation = 0;
1250 svm->nested.vmcb12_gpa = 0;
1251 svm->nested.last_vmcb12_gpa = 0;
1252 vcpu->arch.hflags = 0;
1254 if (!kvm_pause_in_guest(vcpu->kvm)) {
1255 control->pause_filter_count = pause_filter_count;
1256 if (pause_filter_thresh)
1257 control->pause_filter_thresh = pause_filter_thresh;
1258 svm_set_intercept(svm, INTERCEPT_PAUSE);
1260 svm_clr_intercept(svm, INTERCEPT_PAUSE);
1263 svm_check_invpcid(svm);
1266 * If the host supports V_SPEC_CTRL then disable the interception
1267 * of MSR_IA32_SPEC_CTRL.
1269 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
1270 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
1272 if (kvm_vcpu_apicv_active(vcpu))
1273 avic_init_vmcb(svm);
1276 svm_clr_intercept(svm, INTERCEPT_STGI);
1277 svm_clr_intercept(svm, INTERCEPT_CLGI);
1278 svm->vmcb->control.int_ctl |= V_GIF_ENABLE_MASK;
1281 if (sev_guest(vcpu->kvm)) {
1282 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
1283 clr_exception_intercept(svm, UD_VECTOR);
1285 if (sev_es_guest(vcpu->kvm)) {
1286 /* Perform SEV-ES specific VMCB updates */
1287 sev_es_init_vmcb(svm);
1291 vmcb_mark_all_dirty(svm->vmcb);
1297 static void svm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
1299 struct vcpu_svm *svm = to_svm(vcpu);
1304 svm->virt_spec_ctrl = 0;
1307 vcpu->arch.apic_base = APIC_DEFAULT_PHYS_BASE |
1308 MSR_IA32_APICBASE_ENABLE;
1309 if (kvm_vcpu_is_reset_bsp(vcpu))
1310 vcpu->arch.apic_base |= MSR_IA32_APICBASE_BSP;
1314 kvm_cpuid(vcpu, &eax, &dummy, &dummy, &dummy, false);
1315 kvm_rdx_write(vcpu, eax);
1317 if (kvm_vcpu_apicv_active(vcpu) && !init_event)
1318 avic_update_vapic_bar(svm, APIC_DEFAULT_PHYS_BASE);
1321 void svm_switch_vmcb(struct vcpu_svm *svm, struct kvm_vmcb_info *target_vmcb)
1323 svm->current_vmcb = target_vmcb;
1324 svm->vmcb = target_vmcb->ptr;
1327 static int svm_create_vcpu(struct kvm_vcpu *vcpu)
1329 struct vcpu_svm *svm;
1330 struct page *vmcb01_page;
1331 struct page *vmsa_page = NULL;
1334 BUILD_BUG_ON(offsetof(struct vcpu_svm, vcpu) != 0);
1338 vmcb01_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1342 if (sev_es_guest(vcpu->kvm)) {
1344 * SEV-ES guests require a separate VMSA page used to contain
1345 * the encrypted register state of the guest.
1347 vmsa_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1349 goto error_free_vmcb_page;
1352 * SEV-ES guests maintain an encrypted version of their FPU
1353 * state which is restored and saved on VMRUN and VMEXIT.
1354 * Free the fpu structure to prevent KVM from attempting to
1355 * access the FPU state.
1357 kvm_free_guest_fpu(vcpu);
1360 err = avic_init_vcpu(svm);
1362 goto error_free_vmsa_page;
1364 /* We initialize this flag to true to make sure that the is_running
1365 * bit would be set the first time the vcpu is loaded.
1367 if (irqchip_in_kernel(vcpu->kvm) && kvm_apicv_activated(vcpu->kvm))
1368 svm->avic_is_running = true;
1370 svm->msrpm = svm_vcpu_alloc_msrpm();
1373 goto error_free_vmsa_page;
1376 svm_vcpu_init_msrpm(vcpu, svm->msrpm);
1378 svm->vmcb01.ptr = page_address(vmcb01_page);
1379 svm->vmcb01.pa = __sme_set(page_to_pfn(vmcb01_page) << PAGE_SHIFT);
1382 svm->vmsa = page_address(vmsa_page);
1384 svm->guest_state_loaded = false;
1386 svm_switch_vmcb(svm, &svm->vmcb01);
1389 svm_init_osvw(vcpu);
1390 vcpu->arch.microcode_version = 0x01000065;
1392 if (sev_es_guest(vcpu->kvm))
1393 /* Perform SEV-ES specific VMCB creation updates */
1394 sev_es_create_vcpu(svm);
1398 error_free_vmsa_page:
1400 __free_page(vmsa_page);
1401 error_free_vmcb_page:
1402 __free_page(vmcb01_page);
1407 static void svm_clear_current_vmcb(struct vmcb *vmcb)
1411 for_each_online_cpu(i)
1412 cmpxchg(&per_cpu(svm_data, i)->current_vmcb, vmcb, NULL);
1415 static void svm_free_vcpu(struct kvm_vcpu *vcpu)
1417 struct vcpu_svm *svm = to_svm(vcpu);
1420 * The vmcb page can be recycled, causing a false negative in
1421 * svm_vcpu_load(). So, ensure that no logical CPU has this
1422 * vmcb page recorded as its current vmcb.
1424 svm_clear_current_vmcb(svm->vmcb);
1426 svm_free_nested(svm);
1428 sev_free_vcpu(vcpu);
1430 __free_page(pfn_to_page(__sme_clr(svm->vmcb01.pa) >> PAGE_SHIFT));
1431 __free_pages(virt_to_page(svm->msrpm), get_order(MSRPM_SIZE));
1434 static void svm_prepare_guest_switch(struct kvm_vcpu *vcpu)
1436 struct vcpu_svm *svm = to_svm(vcpu);
1437 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
1439 if (svm->guest_state_loaded)
1443 * Save additional host state that will be restored on VMEXIT (sev-es)
1444 * or subsequent vmload of host save area.
1446 if (sev_es_guest(vcpu->kvm)) {
1447 sev_es_prepare_guest_switch(svm, vcpu->cpu);
1449 vmsave(__sme_page_pa(sd->save_area));
1452 if (static_cpu_has(X86_FEATURE_TSCRATEMSR)) {
1453 u64 tsc_ratio = vcpu->arch.tsc_scaling_ratio;
1454 if (tsc_ratio != __this_cpu_read(current_tsc_ratio)) {
1455 __this_cpu_write(current_tsc_ratio, tsc_ratio);
1456 wrmsrl(MSR_AMD64_TSC_RATIO, tsc_ratio);
1460 if (static_cpu_has(X86_FEATURE_RDTSCP))
1461 kvm_set_user_return_msr(TSC_AUX_URET_SLOT, svm->tsc_aux, -1ull);
1463 svm->guest_state_loaded = true;
1466 static void svm_prepare_host_switch(struct kvm_vcpu *vcpu)
1468 to_svm(vcpu)->guest_state_loaded = false;
1471 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1473 struct vcpu_svm *svm = to_svm(vcpu);
1474 struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1476 if (sd->current_vmcb != svm->vmcb) {
1477 sd->current_vmcb = svm->vmcb;
1478 indirect_branch_prediction_barrier();
1480 avic_vcpu_load(vcpu, cpu);
1483 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
1485 avic_vcpu_put(vcpu);
1486 svm_prepare_host_switch(vcpu);
1488 ++vcpu->stat.host_state_reload;
1491 static unsigned long svm_get_rflags(struct kvm_vcpu *vcpu)
1493 struct vcpu_svm *svm = to_svm(vcpu);
1494 unsigned long rflags = svm->vmcb->save.rflags;
1496 if (svm->nmi_singlestep) {
1497 /* Hide our flags if they were not set by the guest */
1498 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_TF))
1499 rflags &= ~X86_EFLAGS_TF;
1500 if (!(svm->nmi_singlestep_guest_rflags & X86_EFLAGS_RF))
1501 rflags &= ~X86_EFLAGS_RF;
1506 static void svm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
1508 if (to_svm(vcpu)->nmi_singlestep)
1509 rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
1512 * Any change of EFLAGS.VM is accompanied by a reload of SS
1513 * (caused by either a task switch or an inter-privilege IRET),
1514 * so we do not need to update the CPL here.
1516 to_svm(vcpu)->vmcb->save.rflags = rflags;
1519 static void svm_cache_reg(struct kvm_vcpu *vcpu, enum kvm_reg reg)
1522 case VCPU_EXREG_PDPTR:
1523 BUG_ON(!npt_enabled);
1524 load_pdptrs(vcpu, vcpu->arch.walk_mmu, kvm_read_cr3(vcpu));
1531 static void svm_set_vintr(struct vcpu_svm *svm)
1533 struct vmcb_control_area *control;
1535 /* The following fields are ignored when AVIC is enabled */
1536 WARN_ON(kvm_vcpu_apicv_active(&svm->vcpu));
1537 svm_set_intercept(svm, INTERCEPT_VINTR);
1540 * This is just a dummy VINTR to actually cause a vmexit to happen.
1541 * Actual injection of virtual interrupts happens through EVENTINJ.
1543 control = &svm->vmcb->control;
1544 control->int_vector = 0x0;
1545 control->int_ctl &= ~V_INTR_PRIO_MASK;
1546 control->int_ctl |= V_IRQ_MASK |
1547 ((/*control->int_vector >> 4*/ 0xf) << V_INTR_PRIO_SHIFT);
1548 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1551 static void svm_clear_vintr(struct vcpu_svm *svm)
1553 const u32 mask = V_TPR_MASK | V_GIF_ENABLE_MASK | V_GIF_MASK | V_INTR_MASKING_MASK;
1554 svm_clr_intercept(svm, INTERCEPT_VINTR);
1556 /* Drop int_ctl fields related to VINTR injection. */
1557 svm->vmcb->control.int_ctl &= mask;
1558 if (is_guest_mode(&svm->vcpu)) {
1559 svm->vmcb01.ptr->control.int_ctl &= mask;
1561 WARN_ON((svm->vmcb->control.int_ctl & V_TPR_MASK) !=
1562 (svm->nested.ctl.int_ctl & V_TPR_MASK));
1563 svm->vmcb->control.int_ctl |= svm->nested.ctl.int_ctl & ~mask;
1566 vmcb_mark_dirty(svm->vmcb, VMCB_INTR);
1569 static struct vmcb_seg *svm_seg(struct kvm_vcpu *vcpu, int seg)
1571 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1572 struct vmcb_save_area *save01 = &to_svm(vcpu)->vmcb01.ptr->save;
1575 case VCPU_SREG_CS: return &save->cs;
1576 case VCPU_SREG_DS: return &save->ds;
1577 case VCPU_SREG_ES: return &save->es;
1578 case VCPU_SREG_FS: return &save01->fs;
1579 case VCPU_SREG_GS: return &save01->gs;
1580 case VCPU_SREG_SS: return &save->ss;
1581 case VCPU_SREG_TR: return &save01->tr;
1582 case VCPU_SREG_LDTR: return &save01->ldtr;
1588 static u64 svm_get_segment_base(struct kvm_vcpu *vcpu, int seg)
1590 struct vmcb_seg *s = svm_seg(vcpu, seg);
1595 static void svm_get_segment(struct kvm_vcpu *vcpu,
1596 struct kvm_segment *var, int seg)
1598 struct vmcb_seg *s = svm_seg(vcpu, seg);
1600 var->base = s->base;
1601 var->limit = s->limit;
1602 var->selector = s->selector;
1603 var->type = s->attrib & SVM_SELECTOR_TYPE_MASK;
1604 var->s = (s->attrib >> SVM_SELECTOR_S_SHIFT) & 1;
1605 var->dpl = (s->attrib >> SVM_SELECTOR_DPL_SHIFT) & 3;
1606 var->present = (s->attrib >> SVM_SELECTOR_P_SHIFT) & 1;
1607 var->avl = (s->attrib >> SVM_SELECTOR_AVL_SHIFT) & 1;
1608 var->l = (s->attrib >> SVM_SELECTOR_L_SHIFT) & 1;
1609 var->db = (s->attrib >> SVM_SELECTOR_DB_SHIFT) & 1;
1612 * AMD CPUs circa 2014 track the G bit for all segments except CS.
1613 * However, the SVM spec states that the G bit is not observed by the
1614 * CPU, and some VMware virtual CPUs drop the G bit for all segments.
1615 * So let's synthesize a legal G bit for all segments, this helps
1616 * running KVM nested. It also helps cross-vendor migration, because
1617 * Intel's vmentry has a check on the 'G' bit.
1619 var->g = s->limit > 0xfffff;
1622 * AMD's VMCB does not have an explicit unusable field, so emulate it
1623 * for cross vendor migration purposes by "not present"
1625 var->unusable = !var->present;
1630 * Work around a bug where the busy flag in the tr selector
1640 * The accessed bit must always be set in the segment
1641 * descriptor cache, although it can be cleared in the
1642 * descriptor, the cached bit always remains at 1. Since
1643 * Intel has a check on this, set it here to support
1644 * cross-vendor migration.
1651 * On AMD CPUs sometimes the DB bit in the segment
1652 * descriptor is left as 1, although the whole segment has
1653 * been made unusable. Clear it here to pass an Intel VMX
1654 * entry check when cross vendor migrating.
1658 /* This is symmetric with svm_set_segment() */
1659 var->dpl = to_svm(vcpu)->vmcb->save.cpl;
1664 static int svm_get_cpl(struct kvm_vcpu *vcpu)
1666 struct vmcb_save_area *save = &to_svm(vcpu)->vmcb->save;
1671 static void svm_get_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1673 struct vcpu_svm *svm = to_svm(vcpu);
1675 dt->size = svm->vmcb->save.idtr.limit;
1676 dt->address = svm->vmcb->save.idtr.base;
1679 static void svm_set_idt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1681 struct vcpu_svm *svm = to_svm(vcpu);
1683 svm->vmcb->save.idtr.limit = dt->size;
1684 svm->vmcb->save.idtr.base = dt->address ;
1685 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1688 static void svm_get_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1690 struct vcpu_svm *svm = to_svm(vcpu);
1692 dt->size = svm->vmcb->save.gdtr.limit;
1693 dt->address = svm->vmcb->save.gdtr.base;
1696 static void svm_set_gdt(struct kvm_vcpu *vcpu, struct desc_ptr *dt)
1698 struct vcpu_svm *svm = to_svm(vcpu);
1700 svm->vmcb->save.gdtr.limit = dt->size;
1701 svm->vmcb->save.gdtr.base = dt->address ;
1702 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
1705 void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
1707 struct vcpu_svm *svm = to_svm(vcpu);
1710 #ifdef CONFIG_X86_64
1711 if (vcpu->arch.efer & EFER_LME && !vcpu->arch.guest_state_protected) {
1712 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
1713 vcpu->arch.efer |= EFER_LMA;
1714 svm->vmcb->save.efer |= EFER_LMA | EFER_LME;
1717 if (is_paging(vcpu) && !(cr0 & X86_CR0_PG)) {
1718 vcpu->arch.efer &= ~EFER_LMA;
1719 svm->vmcb->save.efer &= ~(EFER_LMA | EFER_LME);
1723 vcpu->arch.cr0 = cr0;
1726 hcr0 |= X86_CR0_PG | X86_CR0_WP;
1729 * re-enable caching here because the QEMU bios
1730 * does not do it - this results in some delay at
1733 if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))
1734 hcr0 &= ~(X86_CR0_CD | X86_CR0_NW);
1736 svm->vmcb->save.cr0 = hcr0;
1737 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
1740 * SEV-ES guests must always keep the CR intercepts cleared. CR
1741 * tracking is done using the CR write traps.
1743 if (sev_es_guest(vcpu->kvm))
1747 /* Selective CR0 write remains on. */
1748 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1749 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1751 svm_set_intercept(svm, INTERCEPT_CR0_READ);
1752 svm_set_intercept(svm, INTERCEPT_CR0_WRITE);
1756 static bool svm_is_valid_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1761 void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
1763 unsigned long host_cr4_mce = cr4_read_shadow() & X86_CR4_MCE;
1764 unsigned long old_cr4 = vcpu->arch.cr4;
1766 if (npt_enabled && ((old_cr4 ^ cr4) & X86_CR4_PGE))
1767 svm_flush_tlb(vcpu);
1769 vcpu->arch.cr4 = cr4;
1772 cr4 |= host_cr4_mce;
1773 to_svm(vcpu)->vmcb->save.cr4 = cr4;
1774 vmcb_mark_dirty(to_svm(vcpu)->vmcb, VMCB_CR);
1776 if ((cr4 ^ old_cr4) & (X86_CR4_OSXSAVE | X86_CR4_PKE))
1777 kvm_update_cpuid_runtime(vcpu);
1780 static void svm_set_segment(struct kvm_vcpu *vcpu,
1781 struct kvm_segment *var, int seg)
1783 struct vcpu_svm *svm = to_svm(vcpu);
1784 struct vmcb_seg *s = svm_seg(vcpu, seg);
1786 s->base = var->base;
1787 s->limit = var->limit;
1788 s->selector = var->selector;
1789 s->attrib = (var->type & SVM_SELECTOR_TYPE_MASK);
1790 s->attrib |= (var->s & 1) << SVM_SELECTOR_S_SHIFT;
1791 s->attrib |= (var->dpl & 3) << SVM_SELECTOR_DPL_SHIFT;
1792 s->attrib |= ((var->present & 1) && !var->unusable) << SVM_SELECTOR_P_SHIFT;
1793 s->attrib |= (var->avl & 1) << SVM_SELECTOR_AVL_SHIFT;
1794 s->attrib |= (var->l & 1) << SVM_SELECTOR_L_SHIFT;
1795 s->attrib |= (var->db & 1) << SVM_SELECTOR_DB_SHIFT;
1796 s->attrib |= (var->g & 1) << SVM_SELECTOR_G_SHIFT;
1799 * This is always accurate, except if SYSRET returned to a segment
1800 * with SS.DPL != 3. Intel does not have this quirk, and always
1801 * forces SS.DPL to 3 on sysret, so we ignore that case; fixing it
1802 * would entail passing the CPL to userspace and back.
1804 if (seg == VCPU_SREG_SS)
1805 /* This is symmetric with svm_get_segment() */
1806 svm->vmcb->save.cpl = (var->dpl & 3);
1808 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
1811 static void svm_update_exception_bitmap(struct kvm_vcpu *vcpu)
1813 struct vcpu_svm *svm = to_svm(vcpu);
1815 clr_exception_intercept(svm, BP_VECTOR);
1817 if (vcpu->guest_debug & KVM_GUESTDBG_ENABLE) {
1818 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP)
1819 set_exception_intercept(svm, BP_VECTOR);
1823 static void new_asid(struct vcpu_svm *svm, struct svm_cpu_data *sd)
1825 if (sd->next_asid > sd->max_asid) {
1826 ++sd->asid_generation;
1827 sd->next_asid = sd->min_asid;
1828 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
1829 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1832 svm->current_vmcb->asid_generation = sd->asid_generation;
1833 svm->asid = sd->next_asid++;
1836 static void svm_set_dr6(struct vcpu_svm *svm, unsigned long value)
1838 struct vmcb *vmcb = svm->vmcb;
1840 if (svm->vcpu.arch.guest_state_protected)
1843 if (unlikely(value != vmcb->save.dr6)) {
1844 vmcb->save.dr6 = value;
1845 vmcb_mark_dirty(vmcb, VMCB_DR);
1849 static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
1851 struct vcpu_svm *svm = to_svm(vcpu);
1853 if (vcpu->arch.guest_state_protected)
1856 get_debugreg(vcpu->arch.db[0], 0);
1857 get_debugreg(vcpu->arch.db[1], 1);
1858 get_debugreg(vcpu->arch.db[2], 2);
1859 get_debugreg(vcpu->arch.db[3], 3);
1861 * We cannot reset svm->vmcb->save.dr6 to DR6_ACTIVE_LOW here,
1862 * because db_interception might need it. We can do it before vmentry.
1864 vcpu->arch.dr6 = svm->vmcb->save.dr6;
1865 vcpu->arch.dr7 = svm->vmcb->save.dr7;
1866 vcpu->arch.switch_db_regs &= ~KVM_DEBUGREG_WONT_EXIT;
1867 set_dr_intercepts(svm);
1870 static void svm_set_dr7(struct kvm_vcpu *vcpu, unsigned long value)
1872 struct vcpu_svm *svm = to_svm(vcpu);
1874 if (vcpu->arch.guest_state_protected)
1877 svm->vmcb->save.dr7 = value;
1878 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
1881 static int pf_interception(struct kvm_vcpu *vcpu)
1883 struct vcpu_svm *svm = to_svm(vcpu);
1885 u64 fault_address = svm->vmcb->control.exit_info_2;
1886 u64 error_code = svm->vmcb->control.exit_info_1;
1888 return kvm_handle_page_fault(vcpu, error_code, fault_address,
1889 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1890 svm->vmcb->control.insn_bytes : NULL,
1891 svm->vmcb->control.insn_len);
1894 static int npf_interception(struct kvm_vcpu *vcpu)
1896 struct vcpu_svm *svm = to_svm(vcpu);
1898 u64 fault_address = __sme_clr(svm->vmcb->control.exit_info_2);
1899 u64 error_code = svm->vmcb->control.exit_info_1;
1901 trace_kvm_page_fault(fault_address, error_code);
1902 return kvm_mmu_page_fault(vcpu, fault_address, error_code,
1903 static_cpu_has(X86_FEATURE_DECODEASSISTS) ?
1904 svm->vmcb->control.insn_bytes : NULL,
1905 svm->vmcb->control.insn_len);
1908 static int db_interception(struct kvm_vcpu *vcpu)
1910 struct kvm_run *kvm_run = vcpu->run;
1911 struct vcpu_svm *svm = to_svm(vcpu);
1913 if (!(vcpu->guest_debug &
1914 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) &&
1915 !svm->nmi_singlestep) {
1916 u32 payload = svm->vmcb->save.dr6 ^ DR6_ACTIVE_LOW;
1917 kvm_queue_exception_p(vcpu, DB_VECTOR, payload);
1921 if (svm->nmi_singlestep) {
1922 disable_nmi_singlestep(svm);
1923 /* Make sure we check for pending NMIs upon entry */
1924 kvm_make_request(KVM_REQ_EVENT, vcpu);
1927 if (vcpu->guest_debug &
1928 (KVM_GUESTDBG_SINGLESTEP | KVM_GUESTDBG_USE_HW_BP)) {
1929 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1930 kvm_run->debug.arch.dr6 = svm->vmcb->save.dr6;
1931 kvm_run->debug.arch.dr7 = svm->vmcb->save.dr7;
1932 kvm_run->debug.arch.pc =
1933 svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1934 kvm_run->debug.arch.exception = DB_VECTOR;
1941 static int bp_interception(struct kvm_vcpu *vcpu)
1943 struct vcpu_svm *svm = to_svm(vcpu);
1944 struct kvm_run *kvm_run = vcpu->run;
1946 kvm_run->exit_reason = KVM_EXIT_DEBUG;
1947 kvm_run->debug.arch.pc = svm->vmcb->save.cs.base + svm->vmcb->save.rip;
1948 kvm_run->debug.arch.exception = BP_VECTOR;
1952 static int ud_interception(struct kvm_vcpu *vcpu)
1954 return handle_ud(vcpu);
1957 static int ac_interception(struct kvm_vcpu *vcpu)
1959 kvm_queue_exception_e(vcpu, AC_VECTOR, 0);
1963 static bool is_erratum_383(void)
1968 if (!erratum_383_found)
1971 value = native_read_msr_safe(MSR_IA32_MC0_STATUS, &err);
1975 /* Bit 62 may or may not be set for this mce */
1976 value &= ~(1ULL << 62);
1978 if (value != 0xb600000000010015ULL)
1981 /* Clear MCi_STATUS registers */
1982 for (i = 0; i < 6; ++i)
1983 native_write_msr_safe(MSR_IA32_MCx_STATUS(i), 0, 0);
1985 value = native_read_msr_safe(MSR_IA32_MCG_STATUS, &err);
1989 value &= ~(1ULL << 2);
1990 low = lower_32_bits(value);
1991 high = upper_32_bits(value);
1993 native_write_msr_safe(MSR_IA32_MCG_STATUS, low, high);
1996 /* Flush tlb to evict multi-match entries */
2002 static void svm_handle_mce(struct kvm_vcpu *vcpu)
2004 if (is_erratum_383()) {
2006 * Erratum 383 triggered. Guest state is corrupt so kill the
2009 pr_err("KVM: Guest triggered AMD Erratum 383\n");
2011 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
2017 * On an #MC intercept the MCE handler is not called automatically in
2018 * the host. So do it by hand here.
2020 kvm_machine_check();
2023 static int mc_interception(struct kvm_vcpu *vcpu)
2028 static int shutdown_interception(struct kvm_vcpu *vcpu)
2030 struct kvm_run *kvm_run = vcpu->run;
2031 struct vcpu_svm *svm = to_svm(vcpu);
2034 * The VM save area has already been encrypted so it
2035 * cannot be reinitialized - just terminate.
2037 if (sev_es_guest(vcpu->kvm))
2041 * VMCB is undefined after a SHUTDOWN intercept
2042 * so reinitialize it.
2044 clear_page(svm->vmcb);
2047 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
2051 static int io_interception(struct kvm_vcpu *vcpu)
2053 struct vcpu_svm *svm = to_svm(vcpu);
2054 u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
2055 int size, in, string;
2058 ++vcpu->stat.io_exits;
2059 string = (io_info & SVM_IOIO_STR_MASK) != 0;
2060 in = (io_info & SVM_IOIO_TYPE_MASK) != 0;
2061 port = io_info >> 16;
2062 size = (io_info & SVM_IOIO_SIZE_MASK) >> SVM_IOIO_SIZE_SHIFT;
2065 if (sev_es_guest(vcpu->kvm))
2066 return sev_es_string_io(svm, size, port, in);
2068 return kvm_emulate_instruction(vcpu, 0);
2071 svm->next_rip = svm->vmcb->control.exit_info_2;
2073 return kvm_fast_pio(vcpu, size, port, in);
2076 static int nmi_interception(struct kvm_vcpu *vcpu)
2081 static int intr_interception(struct kvm_vcpu *vcpu)
2083 ++vcpu->stat.irq_exits;
2087 static int vmload_vmsave_interception(struct kvm_vcpu *vcpu, bool vmload)
2089 struct vcpu_svm *svm = to_svm(vcpu);
2090 struct vmcb *vmcb12;
2091 struct kvm_host_map map;
2094 if (nested_svm_check_permissions(vcpu))
2097 ret = kvm_vcpu_map(vcpu, gpa_to_gfn(svm->vmcb->save.rax), &map);
2100 kvm_inject_gp(vcpu, 0);
2106 ret = kvm_skip_emulated_instruction(vcpu);
2109 nested_svm_vmloadsave(vmcb12, svm->vmcb);
2110 svm->sysenter_eip_hi = 0;
2111 svm->sysenter_esp_hi = 0;
2113 nested_svm_vmloadsave(svm->vmcb, vmcb12);
2115 kvm_vcpu_unmap(vcpu, &map, true);
2120 static int vmload_interception(struct kvm_vcpu *vcpu)
2122 return vmload_vmsave_interception(vcpu, true);
2125 static int vmsave_interception(struct kvm_vcpu *vcpu)
2127 return vmload_vmsave_interception(vcpu, false);
2130 static int vmrun_interception(struct kvm_vcpu *vcpu)
2132 if (nested_svm_check_permissions(vcpu))
2135 return nested_svm_vmrun(vcpu);
2145 /* Return NONE_SVM_INSTR if not SVM instrs, otherwise return decode result */
2146 static int svm_instr_opcode(struct kvm_vcpu *vcpu)
2148 struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt;
2150 if (ctxt->b != 0x1 || ctxt->opcode_len != 2)
2151 return NONE_SVM_INSTR;
2153 switch (ctxt->modrm) {
2154 case 0xd8: /* VMRUN */
2155 return SVM_INSTR_VMRUN;
2156 case 0xda: /* VMLOAD */
2157 return SVM_INSTR_VMLOAD;
2158 case 0xdb: /* VMSAVE */
2159 return SVM_INSTR_VMSAVE;
2164 return NONE_SVM_INSTR;
2167 static int emulate_svm_instr(struct kvm_vcpu *vcpu, int opcode)
2169 const int guest_mode_exit_codes[] = {
2170 [SVM_INSTR_VMRUN] = SVM_EXIT_VMRUN,
2171 [SVM_INSTR_VMLOAD] = SVM_EXIT_VMLOAD,
2172 [SVM_INSTR_VMSAVE] = SVM_EXIT_VMSAVE,
2174 int (*const svm_instr_handlers[])(struct kvm_vcpu *vcpu) = {
2175 [SVM_INSTR_VMRUN] = vmrun_interception,
2176 [SVM_INSTR_VMLOAD] = vmload_interception,
2177 [SVM_INSTR_VMSAVE] = vmsave_interception,
2179 struct vcpu_svm *svm = to_svm(vcpu);
2182 if (is_guest_mode(vcpu)) {
2183 /* Returns '1' or -errno on failure, '0' on success. */
2184 ret = nested_svm_simple_vmexit(svm, guest_mode_exit_codes[opcode]);
2189 return svm_instr_handlers[opcode](vcpu);
2193 * #GP handling code. Note that #GP can be triggered under the following two
2195 * 1) SVM VM-related instructions (VMRUN/VMSAVE/VMLOAD) that trigger #GP on
2196 * some AMD CPUs when EAX of these instructions are in the reserved memory
2197 * regions (e.g. SMM memory on host).
2198 * 2) VMware backdoor
2200 static int gp_interception(struct kvm_vcpu *vcpu)
2202 struct vcpu_svm *svm = to_svm(vcpu);
2203 u32 error_code = svm->vmcb->control.exit_info_1;
2206 /* Both #GP cases have zero error_code */
2210 /* Decode the instruction for usage later */
2211 if (x86_decode_emulated_instruction(vcpu, 0, NULL, 0) != EMULATION_OK)
2214 opcode = svm_instr_opcode(vcpu);
2216 if (opcode == NONE_SVM_INSTR) {
2217 if (!enable_vmware_backdoor)
2221 * VMware backdoor emulation on #GP interception only handles
2222 * IN{S}, OUT{S}, and RDPMC.
2224 if (!is_guest_mode(vcpu))
2225 return kvm_emulate_instruction(vcpu,
2226 EMULTYPE_VMWARE_GP | EMULTYPE_NO_DECODE);
2228 return emulate_svm_instr(vcpu, opcode);
2231 kvm_queue_exception_e(vcpu, GP_VECTOR, error_code);
2235 void svm_set_gif(struct vcpu_svm *svm, bool value)
2239 * If VGIF is enabled, the STGI intercept is only added to
2240 * detect the opening of the SMI/NMI window; remove it now.
2241 * Likewise, clear the VINTR intercept, we will set it
2242 * again while processing KVM_REQ_EVENT if needed.
2244 if (vgif_enabled(svm))
2245 svm_clr_intercept(svm, INTERCEPT_STGI);
2246 if (svm_is_intercept(svm, INTERCEPT_VINTR))
2247 svm_clear_vintr(svm);
2250 if (svm->vcpu.arch.smi_pending ||
2251 svm->vcpu.arch.nmi_pending ||
2252 kvm_cpu_has_injectable_intr(&svm->vcpu))
2253 kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
2258 * After a CLGI no interrupts should come. But if vGIF is
2259 * in use, we still rely on the VINTR intercept (rather than
2260 * STGI) to detect an open interrupt window.
2262 if (!vgif_enabled(svm))
2263 svm_clear_vintr(svm);
2267 static int stgi_interception(struct kvm_vcpu *vcpu)
2271 if (nested_svm_check_permissions(vcpu))
2274 ret = kvm_skip_emulated_instruction(vcpu);
2275 svm_set_gif(to_svm(vcpu), true);
2279 static int clgi_interception(struct kvm_vcpu *vcpu)
2283 if (nested_svm_check_permissions(vcpu))
2286 ret = kvm_skip_emulated_instruction(vcpu);
2287 svm_set_gif(to_svm(vcpu), false);
2291 static int invlpga_interception(struct kvm_vcpu *vcpu)
2293 trace_kvm_invlpga(to_svm(vcpu)->vmcb->save.rip, kvm_rcx_read(vcpu),
2294 kvm_rax_read(vcpu));
2296 /* Let's treat INVLPGA the same as INVLPG (can be optimized!) */
2297 kvm_mmu_invlpg(vcpu, kvm_rax_read(vcpu));
2299 return kvm_skip_emulated_instruction(vcpu);
2302 static int skinit_interception(struct kvm_vcpu *vcpu)
2304 trace_kvm_skinit(to_svm(vcpu)->vmcb->save.rip, kvm_rax_read(vcpu));
2306 kvm_queue_exception(vcpu, UD_VECTOR);
2310 static int task_switch_interception(struct kvm_vcpu *vcpu)
2312 struct vcpu_svm *svm = to_svm(vcpu);
2315 int int_type = svm->vmcb->control.exit_int_info &
2316 SVM_EXITINTINFO_TYPE_MASK;
2317 int int_vec = svm->vmcb->control.exit_int_info & SVM_EVTINJ_VEC_MASK;
2319 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_TYPE_MASK;
2321 svm->vmcb->control.exit_int_info & SVM_EXITINTINFO_VALID;
2322 bool has_error_code = false;
2325 tss_selector = (u16)svm->vmcb->control.exit_info_1;
2327 if (svm->vmcb->control.exit_info_2 &
2328 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_IRET))
2329 reason = TASK_SWITCH_IRET;
2330 else if (svm->vmcb->control.exit_info_2 &
2331 (1ULL << SVM_EXITINFOSHIFT_TS_REASON_JMP))
2332 reason = TASK_SWITCH_JMP;
2334 reason = TASK_SWITCH_GATE;
2336 reason = TASK_SWITCH_CALL;
2338 if (reason == TASK_SWITCH_GATE) {
2340 case SVM_EXITINTINFO_TYPE_NMI:
2341 vcpu->arch.nmi_injected = false;
2343 case SVM_EXITINTINFO_TYPE_EXEPT:
2344 if (svm->vmcb->control.exit_info_2 &
2345 (1ULL << SVM_EXITINFOSHIFT_TS_HAS_ERROR_CODE)) {
2346 has_error_code = true;
2348 (u32)svm->vmcb->control.exit_info_2;
2350 kvm_clear_exception_queue(vcpu);
2352 case SVM_EXITINTINFO_TYPE_INTR:
2353 kvm_clear_interrupt_queue(vcpu);
2360 if (reason != TASK_SWITCH_GATE ||
2361 int_type == SVM_EXITINTINFO_TYPE_SOFT ||
2362 (int_type == SVM_EXITINTINFO_TYPE_EXEPT &&
2363 (int_vec == OF_VECTOR || int_vec == BP_VECTOR))) {
2364 if (!skip_emulated_instruction(vcpu))
2368 if (int_type != SVM_EXITINTINFO_TYPE_SOFT)
2371 return kvm_task_switch(vcpu, tss_selector, int_vec, reason,
2372 has_error_code, error_code);
2375 static int iret_interception(struct kvm_vcpu *vcpu)
2377 struct vcpu_svm *svm = to_svm(vcpu);
2379 ++vcpu->stat.nmi_window_exits;
2380 vcpu->arch.hflags |= HF_IRET_MASK;
2381 if (!sev_es_guest(vcpu->kvm)) {
2382 svm_clr_intercept(svm, INTERCEPT_IRET);
2383 svm->nmi_iret_rip = kvm_rip_read(vcpu);
2385 kvm_make_request(KVM_REQ_EVENT, vcpu);
2389 static int invlpg_interception(struct kvm_vcpu *vcpu)
2391 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2392 return kvm_emulate_instruction(vcpu, 0);
2394 kvm_mmu_invlpg(vcpu, to_svm(vcpu)->vmcb->control.exit_info_1);
2395 return kvm_skip_emulated_instruction(vcpu);
2398 static int emulate_on_interception(struct kvm_vcpu *vcpu)
2400 return kvm_emulate_instruction(vcpu, 0);
2403 static int rsm_interception(struct kvm_vcpu *vcpu)
2405 return kvm_emulate_instruction_from_buffer(vcpu, rsm_ins_bytes, 2);
2408 static bool check_selective_cr0_intercepted(struct kvm_vcpu *vcpu,
2411 struct vcpu_svm *svm = to_svm(vcpu);
2412 unsigned long cr0 = vcpu->arch.cr0;
2415 if (!is_guest_mode(vcpu) ||
2416 (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_SELECTIVE_CR0))))
2419 cr0 &= ~SVM_CR0_SELECTIVE_MASK;
2420 val &= ~SVM_CR0_SELECTIVE_MASK;
2423 svm->vmcb->control.exit_code = SVM_EXIT_CR0_SEL_WRITE;
2424 ret = (nested_svm_exit_handled(svm) == NESTED_EXIT_DONE);
2430 #define CR_VALID (1ULL << 63)
2432 static int cr_interception(struct kvm_vcpu *vcpu)
2434 struct vcpu_svm *svm = to_svm(vcpu);
2439 if (!static_cpu_has(X86_FEATURE_DECODEASSISTS))
2440 return emulate_on_interception(vcpu);
2442 if (unlikely((svm->vmcb->control.exit_info_1 & CR_VALID) == 0))
2443 return emulate_on_interception(vcpu);
2445 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2446 if (svm->vmcb->control.exit_code == SVM_EXIT_CR0_SEL_WRITE)
2447 cr = SVM_EXIT_WRITE_CR0 - SVM_EXIT_READ_CR0;
2449 cr = svm->vmcb->control.exit_code - SVM_EXIT_READ_CR0;
2452 if (cr >= 16) { /* mov to cr */
2454 val = kvm_register_read(vcpu, reg);
2455 trace_kvm_cr_write(cr, val);
2458 if (!check_selective_cr0_intercepted(vcpu, val))
2459 err = kvm_set_cr0(vcpu, val);
2465 err = kvm_set_cr3(vcpu, val);
2468 err = kvm_set_cr4(vcpu, val);
2471 err = kvm_set_cr8(vcpu, val);
2474 WARN(1, "unhandled write to CR%d", cr);
2475 kvm_queue_exception(vcpu, UD_VECTOR);
2478 } else { /* mov from cr */
2481 val = kvm_read_cr0(vcpu);
2484 val = vcpu->arch.cr2;
2487 val = kvm_read_cr3(vcpu);
2490 val = kvm_read_cr4(vcpu);
2493 val = kvm_get_cr8(vcpu);
2496 WARN(1, "unhandled read from CR%d", cr);
2497 kvm_queue_exception(vcpu, UD_VECTOR);
2500 kvm_register_write(vcpu, reg, val);
2501 trace_kvm_cr_read(cr, val);
2503 return kvm_complete_insn_gp(vcpu, err);
2506 static int cr_trap(struct kvm_vcpu *vcpu)
2508 struct vcpu_svm *svm = to_svm(vcpu);
2509 unsigned long old_value, new_value;
2513 new_value = (unsigned long)svm->vmcb->control.exit_info_1;
2515 cr = svm->vmcb->control.exit_code - SVM_EXIT_CR0_WRITE_TRAP;
2518 old_value = kvm_read_cr0(vcpu);
2519 svm_set_cr0(vcpu, new_value);
2521 kvm_post_set_cr0(vcpu, old_value, new_value);
2524 old_value = kvm_read_cr4(vcpu);
2525 svm_set_cr4(vcpu, new_value);
2527 kvm_post_set_cr4(vcpu, old_value, new_value);
2530 ret = kvm_set_cr8(vcpu, new_value);
2533 WARN(1, "unhandled CR%d write trap", cr);
2534 kvm_queue_exception(vcpu, UD_VECTOR);
2538 return kvm_complete_insn_gp(vcpu, ret);
2541 static int dr_interception(struct kvm_vcpu *vcpu)
2543 struct vcpu_svm *svm = to_svm(vcpu);
2548 if (vcpu->guest_debug == 0) {
2550 * No more DR vmexits; force a reload of the debug registers
2551 * and reenter on this instruction. The next vmexit will
2552 * retrieve the full state of the debug registers.
2554 clr_dr_intercepts(svm);
2555 vcpu->arch.switch_db_regs |= KVM_DEBUGREG_WONT_EXIT;
2559 if (!boot_cpu_has(X86_FEATURE_DECODEASSISTS))
2560 return emulate_on_interception(vcpu);
2562 reg = svm->vmcb->control.exit_info_1 & SVM_EXITINFO_REG_MASK;
2563 dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
2564 if (dr >= 16) { /* mov to DRn */
2566 val = kvm_register_read(vcpu, reg);
2567 err = kvm_set_dr(vcpu, dr, val);
2569 kvm_get_dr(vcpu, dr, &val);
2570 kvm_register_write(vcpu, reg, val);
2573 return kvm_complete_insn_gp(vcpu, err);
2576 static int cr8_write_interception(struct kvm_vcpu *vcpu)
2580 u8 cr8_prev = kvm_get_cr8(vcpu);
2581 /* instruction emulation calls kvm_set_cr8() */
2582 r = cr_interception(vcpu);
2583 if (lapic_in_kernel(vcpu))
2585 if (cr8_prev <= kvm_get_cr8(vcpu))
2587 vcpu->run->exit_reason = KVM_EXIT_SET_TPR;
2591 static int efer_trap(struct kvm_vcpu *vcpu)
2593 struct msr_data msr_info;
2597 * Clear the EFER_SVME bit from EFER. The SVM code always sets this
2598 * bit in svm_set_efer(), but __kvm_valid_efer() checks it against
2599 * whether the guest has X86_FEATURE_SVM - this avoids a failure if
2600 * the guest doesn't have X86_FEATURE_SVM.
2602 msr_info.host_initiated = false;
2603 msr_info.index = MSR_EFER;
2604 msr_info.data = to_svm(vcpu)->vmcb->control.exit_info_1 & ~EFER_SVME;
2605 ret = kvm_set_msr_common(vcpu, &msr_info);
2607 return kvm_complete_insn_gp(vcpu, ret);
2610 static int svm_get_msr_feature(struct kvm_msr_entry *msr)
2614 switch (msr->index) {
2615 case MSR_F10H_DECFG:
2616 if (boot_cpu_has(X86_FEATURE_LFENCE_RDTSC))
2617 msr->data |= MSR_F10H_DECFG_LFENCE_SERIALIZE;
2619 case MSR_IA32_PERF_CAPABILITIES:
2622 return KVM_MSR_RET_INVALID;
2628 static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
2630 struct vcpu_svm *svm = to_svm(vcpu);
2632 switch (msr_info->index) {
2634 msr_info->data = svm->vmcb01.ptr->save.star;
2636 #ifdef CONFIG_X86_64
2638 msr_info->data = svm->vmcb01.ptr->save.lstar;
2641 msr_info->data = svm->vmcb01.ptr->save.cstar;
2643 case MSR_KERNEL_GS_BASE:
2644 msr_info->data = svm->vmcb01.ptr->save.kernel_gs_base;
2646 case MSR_SYSCALL_MASK:
2647 msr_info->data = svm->vmcb01.ptr->save.sfmask;
2650 case MSR_IA32_SYSENTER_CS:
2651 msr_info->data = svm->vmcb01.ptr->save.sysenter_cs;
2653 case MSR_IA32_SYSENTER_EIP:
2654 msr_info->data = (u32)svm->vmcb01.ptr->save.sysenter_eip;
2655 if (guest_cpuid_is_intel(vcpu))
2656 msr_info->data |= (u64)svm->sysenter_eip_hi << 32;
2658 case MSR_IA32_SYSENTER_ESP:
2659 msr_info->data = svm->vmcb01.ptr->save.sysenter_esp;
2660 if (guest_cpuid_is_intel(vcpu))
2661 msr_info->data |= (u64)svm->sysenter_esp_hi << 32;
2664 if (!boot_cpu_has(X86_FEATURE_RDTSCP))
2666 if (!msr_info->host_initiated &&
2667 !guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP))
2669 msr_info->data = svm->tsc_aux;
2672 * Nobody will change the following 5 values in the VMCB so we can
2673 * safely return them on rdmsr. They will always be 0 until LBRV is
2676 case MSR_IA32_DEBUGCTLMSR:
2677 msr_info->data = svm->vmcb->save.dbgctl;
2679 case MSR_IA32_LASTBRANCHFROMIP:
2680 msr_info->data = svm->vmcb->save.br_from;
2682 case MSR_IA32_LASTBRANCHTOIP:
2683 msr_info->data = svm->vmcb->save.br_to;
2685 case MSR_IA32_LASTINTFROMIP:
2686 msr_info->data = svm->vmcb->save.last_excp_from;
2688 case MSR_IA32_LASTINTTOIP:
2689 msr_info->data = svm->vmcb->save.last_excp_to;
2691 case MSR_VM_HSAVE_PA:
2692 msr_info->data = svm->nested.hsave_msr;
2695 msr_info->data = svm->nested.vm_cr_msr;
2697 case MSR_IA32_SPEC_CTRL:
2698 if (!msr_info->host_initiated &&
2699 !guest_has_spec_ctrl_msr(vcpu))
2702 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2703 msr_info->data = svm->vmcb->save.spec_ctrl;
2705 msr_info->data = svm->spec_ctrl;
2707 case MSR_AMD64_VIRT_SPEC_CTRL:
2708 if (!msr_info->host_initiated &&
2709 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2712 msr_info->data = svm->virt_spec_ctrl;
2714 case MSR_F15H_IC_CFG: {
2718 family = guest_cpuid_family(vcpu);
2719 model = guest_cpuid_model(vcpu);
2721 if (family < 0 || model < 0)
2722 return kvm_get_msr_common(vcpu, msr_info);
2726 if (family == 0x15 &&
2727 (model >= 0x2 && model < 0x20))
2728 msr_info->data = 0x1E;
2731 case MSR_F10H_DECFG:
2732 msr_info->data = svm->msr_decfg;
2735 return kvm_get_msr_common(vcpu, msr_info);
2740 static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
2742 struct vcpu_svm *svm = to_svm(vcpu);
2743 if (!err || !sev_es_guest(vcpu->kvm) || WARN_ON_ONCE(!svm->ghcb))
2744 return kvm_complete_insn_gp(vcpu, err);
2746 ghcb_set_sw_exit_info_1(svm->ghcb, 1);
2747 ghcb_set_sw_exit_info_2(svm->ghcb,
2749 SVM_EVTINJ_TYPE_EXEPT |
2754 static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
2756 struct vcpu_svm *svm = to_svm(vcpu);
2757 int svm_dis, chg_mask;
2759 if (data & ~SVM_VM_CR_VALID_MASK)
2762 chg_mask = SVM_VM_CR_VALID_MASK;
2764 if (svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK)
2765 chg_mask &= ~(SVM_VM_CR_SVM_LOCK_MASK | SVM_VM_CR_SVM_DIS_MASK);
2767 svm->nested.vm_cr_msr &= ~chg_mask;
2768 svm->nested.vm_cr_msr |= (data & chg_mask);
2770 svm_dis = svm->nested.vm_cr_msr & SVM_VM_CR_SVM_DIS_MASK;
2772 /* check for svm_disable while efer.svme is set */
2773 if (svm_dis && (vcpu->arch.efer & EFER_SVME))
2779 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
2781 struct vcpu_svm *svm = to_svm(vcpu);
2784 u32 ecx = msr->index;
2785 u64 data = msr->data;
2787 case MSR_IA32_CR_PAT:
2788 if (!kvm_mtrr_valid(vcpu, MSR_IA32_CR_PAT, data))
2790 vcpu->arch.pat = data;
2791 svm->vmcb01.ptr->save.g_pat = data;
2792 if (is_guest_mode(vcpu))
2793 nested_vmcb02_compute_g_pat(svm);
2794 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
2796 case MSR_IA32_SPEC_CTRL:
2797 if (!msr->host_initiated &&
2798 !guest_has_spec_ctrl_msr(vcpu))
2801 if (kvm_spec_ctrl_test_value(data))
2804 if (boot_cpu_has(X86_FEATURE_V_SPEC_CTRL))
2805 svm->vmcb->save.spec_ctrl = data;
2807 svm->spec_ctrl = data;
2813 * When it's written (to non-zero) for the first time, pass
2817 * The handling of the MSR bitmap for L2 guests is done in
2818 * nested_svm_vmrun_msrpm.
2819 * We update the L1 MSR bit as well since it will end up
2820 * touching the MSR anyway now.
2822 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SPEC_CTRL, 1, 1);
2824 case MSR_IA32_PRED_CMD:
2825 if (!msr->host_initiated &&
2826 !guest_has_pred_cmd_msr(vcpu))
2829 if (data & ~PRED_CMD_IBPB)
2831 if (!boot_cpu_has(X86_FEATURE_IBPB))
2836 wrmsrl(MSR_IA32_PRED_CMD, PRED_CMD_IBPB);
2837 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_PRED_CMD, 0, 1);
2839 case MSR_AMD64_VIRT_SPEC_CTRL:
2840 if (!msr->host_initiated &&
2841 !guest_cpuid_has(vcpu, X86_FEATURE_VIRT_SSBD))
2844 if (data & ~SPEC_CTRL_SSBD)
2847 svm->virt_spec_ctrl = data;
2850 svm->vmcb01.ptr->save.star = data;
2852 #ifdef CONFIG_X86_64
2854 svm->vmcb01.ptr->save.lstar = data;
2857 svm->vmcb01.ptr->save.cstar = data;
2859 case MSR_KERNEL_GS_BASE:
2860 svm->vmcb01.ptr->save.kernel_gs_base = data;
2862 case MSR_SYSCALL_MASK:
2863 svm->vmcb01.ptr->save.sfmask = data;
2866 case MSR_IA32_SYSENTER_CS:
2867 svm->vmcb01.ptr->save.sysenter_cs = data;
2869 case MSR_IA32_SYSENTER_EIP:
2870 svm->vmcb01.ptr->save.sysenter_eip = (u32)data;
2872 * We only intercept the MSR_IA32_SYSENTER_{EIP|ESP} msrs
2873 * when we spoof an Intel vendor ID (for cross vendor migration).
2874 * In this case we use this intercept to track the high
2875 * 32 bit part of these msrs to support Intel's
2876 * implementation of SYSENTER/SYSEXIT.
2878 svm->sysenter_eip_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2880 case MSR_IA32_SYSENTER_ESP:
2881 svm->vmcb01.ptr->save.sysenter_esp = (u32)data;
2882 svm->sysenter_esp_hi = guest_cpuid_is_intel(vcpu) ? (data >> 32) : 0;
2885 if (!boot_cpu_has(X86_FEATURE_RDTSCP))
2888 if (!msr->host_initiated &&
2889 !guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP))
2893 * Per Intel's SDM, bits 63:32 are reserved, but AMD's APM has
2894 * incomplete and conflicting architectural behavior. Current
2895 * AMD CPUs completely ignore bits 63:32, i.e. they aren't
2896 * reserved and always read as zeros. Emulate AMD CPU behavior
2897 * to avoid explosions if the vCPU is migrated from an AMD host
2903 * TSC_AUX is usually changed only during boot and never read
2904 * directly. Intercept TSC_AUX instead of exposing it to the
2905 * guest via direct_access_msrs, and switch it via user return.
2908 r = kvm_set_user_return_msr(TSC_AUX_URET_SLOT, data, -1ull);
2913 svm->tsc_aux = data;
2915 case MSR_IA32_DEBUGCTLMSR:
2916 if (!boot_cpu_has(X86_FEATURE_LBRV)) {
2917 vcpu_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTL 0x%llx, nop\n",
2921 if (data & DEBUGCTL_RESERVED_BITS)
2924 svm->vmcb->save.dbgctl = data;
2925 vmcb_mark_dirty(svm->vmcb, VMCB_LBR);
2926 if (data & (1ULL<<0))
2927 svm_enable_lbrv(vcpu);
2929 svm_disable_lbrv(vcpu);
2931 case MSR_VM_HSAVE_PA:
2932 svm->nested.hsave_msr = data;
2935 return svm_set_vm_cr(vcpu, data);
2937 vcpu_unimpl(vcpu, "unimplemented wrmsr: 0x%x data 0x%llx\n", ecx, data);
2939 case MSR_F10H_DECFG: {
2940 struct kvm_msr_entry msr_entry;
2942 msr_entry.index = msr->index;
2943 if (svm_get_msr_feature(&msr_entry))
2946 /* Check the supported bits */
2947 if (data & ~msr_entry.data)
2950 /* Don't allow the guest to change a bit, #GP */
2951 if (!msr->host_initiated && (data ^ msr_entry.data))
2954 svm->msr_decfg = data;
2957 case MSR_IA32_APICBASE:
2958 if (kvm_vcpu_apicv_active(vcpu))
2959 avic_update_vapic_bar(to_svm(vcpu), data);
2962 return kvm_set_msr_common(vcpu, msr);
2967 static int msr_interception(struct kvm_vcpu *vcpu)
2969 if (to_svm(vcpu)->vmcb->control.exit_info_1)
2970 return kvm_emulate_wrmsr(vcpu);
2972 return kvm_emulate_rdmsr(vcpu);
2975 static int interrupt_window_interception(struct kvm_vcpu *vcpu)
2977 kvm_make_request(KVM_REQ_EVENT, vcpu);
2978 svm_clear_vintr(to_svm(vcpu));
2981 * For AVIC, the only reason to end up here is ExtINTs.
2982 * In this case AVIC was temporarily disabled for
2983 * requesting the IRQ window and we have to re-enable it.
2985 svm_toggle_avic_for_irq_window(vcpu, true);
2987 ++vcpu->stat.irq_window_exits;
2991 static int pause_interception(struct kvm_vcpu *vcpu)
2996 * CPL is not made available for an SEV-ES guest, therefore
2997 * vcpu->arch.preempted_in_kernel can never be true. Just
2998 * set in_kernel to false as well.
3000 in_kernel = !sev_es_guest(vcpu->kvm) && svm_get_cpl(vcpu) == 0;
3002 if (!kvm_pause_in_guest(vcpu->kvm))
3003 grow_ple_window(vcpu);
3005 kvm_vcpu_on_spin(vcpu, in_kernel);
3006 return kvm_skip_emulated_instruction(vcpu);
3009 static int invpcid_interception(struct kvm_vcpu *vcpu)
3011 struct vcpu_svm *svm = to_svm(vcpu);
3015 if (!guest_cpuid_has(vcpu, X86_FEATURE_INVPCID)) {
3016 kvm_queue_exception(vcpu, UD_VECTOR);
3021 * For an INVPCID intercept:
3022 * EXITINFO1 provides the linear address of the memory operand.
3023 * EXITINFO2 provides the contents of the register operand.
3025 type = svm->vmcb->control.exit_info_2;
3026 gva = svm->vmcb->control.exit_info_1;
3029 kvm_inject_gp(vcpu, 0);
3033 return kvm_handle_invpcid(vcpu, type, gva);
3036 static int (*const svm_exit_handlers[])(struct kvm_vcpu *vcpu) = {
3037 [SVM_EXIT_READ_CR0] = cr_interception,
3038 [SVM_EXIT_READ_CR3] = cr_interception,
3039 [SVM_EXIT_READ_CR4] = cr_interception,
3040 [SVM_EXIT_READ_CR8] = cr_interception,
3041 [SVM_EXIT_CR0_SEL_WRITE] = cr_interception,
3042 [SVM_EXIT_WRITE_CR0] = cr_interception,
3043 [SVM_EXIT_WRITE_CR3] = cr_interception,
3044 [SVM_EXIT_WRITE_CR4] = cr_interception,
3045 [SVM_EXIT_WRITE_CR8] = cr8_write_interception,
3046 [SVM_EXIT_READ_DR0] = dr_interception,
3047 [SVM_EXIT_READ_DR1] = dr_interception,
3048 [SVM_EXIT_READ_DR2] = dr_interception,
3049 [SVM_EXIT_READ_DR3] = dr_interception,
3050 [SVM_EXIT_READ_DR4] = dr_interception,
3051 [SVM_EXIT_READ_DR5] = dr_interception,
3052 [SVM_EXIT_READ_DR6] = dr_interception,
3053 [SVM_EXIT_READ_DR7] = dr_interception,
3054 [SVM_EXIT_WRITE_DR0] = dr_interception,
3055 [SVM_EXIT_WRITE_DR1] = dr_interception,
3056 [SVM_EXIT_WRITE_DR2] = dr_interception,
3057 [SVM_EXIT_WRITE_DR3] = dr_interception,
3058 [SVM_EXIT_WRITE_DR4] = dr_interception,
3059 [SVM_EXIT_WRITE_DR5] = dr_interception,
3060 [SVM_EXIT_WRITE_DR6] = dr_interception,
3061 [SVM_EXIT_WRITE_DR7] = dr_interception,
3062 [SVM_EXIT_EXCP_BASE + DB_VECTOR] = db_interception,
3063 [SVM_EXIT_EXCP_BASE + BP_VECTOR] = bp_interception,
3064 [SVM_EXIT_EXCP_BASE + UD_VECTOR] = ud_interception,
3065 [SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
3066 [SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
3067 [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
3068 [SVM_EXIT_EXCP_BASE + GP_VECTOR] = gp_interception,
3069 [SVM_EXIT_INTR] = intr_interception,
3070 [SVM_EXIT_NMI] = nmi_interception,
3071 [SVM_EXIT_SMI] = kvm_emulate_as_nop,
3072 [SVM_EXIT_INIT] = kvm_emulate_as_nop,
3073 [SVM_EXIT_VINTR] = interrupt_window_interception,
3074 [SVM_EXIT_RDPMC] = kvm_emulate_rdpmc,
3075 [SVM_EXIT_CPUID] = kvm_emulate_cpuid,
3076 [SVM_EXIT_IRET] = iret_interception,
3077 [SVM_EXIT_INVD] = kvm_emulate_invd,
3078 [SVM_EXIT_PAUSE] = pause_interception,
3079 [SVM_EXIT_HLT] = kvm_emulate_halt,
3080 [SVM_EXIT_INVLPG] = invlpg_interception,
3081 [SVM_EXIT_INVLPGA] = invlpga_interception,
3082 [SVM_EXIT_IOIO] = io_interception,
3083 [SVM_EXIT_MSR] = msr_interception,
3084 [SVM_EXIT_TASK_SWITCH] = task_switch_interception,
3085 [SVM_EXIT_SHUTDOWN] = shutdown_interception,
3086 [SVM_EXIT_VMRUN] = vmrun_interception,
3087 [SVM_EXIT_VMMCALL] = kvm_emulate_hypercall,
3088 [SVM_EXIT_VMLOAD] = vmload_interception,
3089 [SVM_EXIT_VMSAVE] = vmsave_interception,
3090 [SVM_EXIT_STGI] = stgi_interception,
3091 [SVM_EXIT_CLGI] = clgi_interception,
3092 [SVM_EXIT_SKINIT] = skinit_interception,
3093 [SVM_EXIT_WBINVD] = kvm_emulate_wbinvd,
3094 [SVM_EXIT_MONITOR] = kvm_emulate_monitor,
3095 [SVM_EXIT_MWAIT] = kvm_emulate_mwait,
3096 [SVM_EXIT_XSETBV] = kvm_emulate_xsetbv,
3097 [SVM_EXIT_RDPRU] = kvm_handle_invalid_op,
3098 [SVM_EXIT_EFER_WRITE_TRAP] = efer_trap,
3099 [SVM_EXIT_CR0_WRITE_TRAP] = cr_trap,
3100 [SVM_EXIT_CR4_WRITE_TRAP] = cr_trap,
3101 [SVM_EXIT_CR8_WRITE_TRAP] = cr_trap,
3102 [SVM_EXIT_INVPCID] = invpcid_interception,
3103 [SVM_EXIT_NPF] = npf_interception,
3104 [SVM_EXIT_RSM] = rsm_interception,
3105 [SVM_EXIT_AVIC_INCOMPLETE_IPI] = avic_incomplete_ipi_interception,
3106 [SVM_EXIT_AVIC_UNACCELERATED_ACCESS] = avic_unaccelerated_access_interception,
3107 [SVM_EXIT_VMGEXIT] = sev_handle_vmgexit,
3110 static void dump_vmcb(struct kvm_vcpu *vcpu)
3112 struct vcpu_svm *svm = to_svm(vcpu);
3113 struct vmcb_control_area *control = &svm->vmcb->control;
3114 struct vmcb_save_area *save = &svm->vmcb->save;
3115 struct vmcb_save_area *save01 = &svm->vmcb01.ptr->save;
3117 if (!dump_invalid_vmcb) {
3118 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3122 pr_err("VMCB Control Area:\n");
3123 pr_err("%-20s%04x\n", "cr_read:", control->intercepts[INTERCEPT_CR] & 0xffff);
3124 pr_err("%-20s%04x\n", "cr_write:", control->intercepts[INTERCEPT_CR] >> 16);
3125 pr_err("%-20s%04x\n", "dr_read:", control->intercepts[INTERCEPT_DR] & 0xffff);
3126 pr_err("%-20s%04x\n", "dr_write:", control->intercepts[INTERCEPT_DR] >> 16);
3127 pr_err("%-20s%08x\n", "exceptions:", control->intercepts[INTERCEPT_EXCEPTION]);
3128 pr_err("%-20s%08x %08x\n", "intercepts:",
3129 control->intercepts[INTERCEPT_WORD3],
3130 control->intercepts[INTERCEPT_WORD4]);
3131 pr_err("%-20s%d\n", "pause filter count:", control->pause_filter_count);
3132 pr_err("%-20s%d\n", "pause filter threshold:",
3133 control->pause_filter_thresh);
3134 pr_err("%-20s%016llx\n", "iopm_base_pa:", control->iopm_base_pa);
3135 pr_err("%-20s%016llx\n", "msrpm_base_pa:", control->msrpm_base_pa);
3136 pr_err("%-20s%016llx\n", "tsc_offset:", control->tsc_offset);
3137 pr_err("%-20s%d\n", "asid:", control->asid);
3138 pr_err("%-20s%d\n", "tlb_ctl:", control->tlb_ctl);
3139 pr_err("%-20s%08x\n", "int_ctl:", control->int_ctl);
3140 pr_err("%-20s%08x\n", "int_vector:", control->int_vector);
3141 pr_err("%-20s%08x\n", "int_state:", control->int_state);
3142 pr_err("%-20s%08x\n", "exit_code:", control->exit_code);
3143 pr_err("%-20s%016llx\n", "exit_info1:", control->exit_info_1);
3144 pr_err("%-20s%016llx\n", "exit_info2:", control->exit_info_2);
3145 pr_err("%-20s%08x\n", "exit_int_info:", control->exit_int_info);
3146 pr_err("%-20s%08x\n", "exit_int_info_err:", control->exit_int_info_err);
3147 pr_err("%-20s%lld\n", "nested_ctl:", control->nested_ctl);
3148 pr_err("%-20s%016llx\n", "nested_cr3:", control->nested_cr3);
3149 pr_err("%-20s%016llx\n", "avic_vapic_bar:", control->avic_vapic_bar);
3150 pr_err("%-20s%016llx\n", "ghcb:", control->ghcb_gpa);
3151 pr_err("%-20s%08x\n", "event_inj:", control->event_inj);
3152 pr_err("%-20s%08x\n", "event_inj_err:", control->event_inj_err);
3153 pr_err("%-20s%lld\n", "virt_ext:", control->virt_ext);
3154 pr_err("%-20s%016llx\n", "next_rip:", control->next_rip);
3155 pr_err("%-20s%016llx\n", "avic_backing_page:", control->avic_backing_page);
3156 pr_err("%-20s%016llx\n", "avic_logical_id:", control->avic_logical_id);
3157 pr_err("%-20s%016llx\n", "avic_physical_id:", control->avic_physical_id);
3158 pr_err("%-20s%016llx\n", "vmsa_pa:", control->vmsa_pa);
3159 pr_err("VMCB State Save Area:\n");
3160 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3162 save->es.selector, save->es.attrib,
3163 save->es.limit, save->es.base);
3164 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3166 save->cs.selector, save->cs.attrib,
3167 save->cs.limit, save->cs.base);
3168 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3170 save->ss.selector, save->ss.attrib,
3171 save->ss.limit, save->ss.base);
3172 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3174 save->ds.selector, save->ds.attrib,
3175 save->ds.limit, save->ds.base);
3176 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3178 save01->fs.selector, save01->fs.attrib,
3179 save01->fs.limit, save01->fs.base);
3180 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3182 save01->gs.selector, save01->gs.attrib,
3183 save01->gs.limit, save01->gs.base);
3184 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3186 save->gdtr.selector, save->gdtr.attrib,
3187 save->gdtr.limit, save->gdtr.base);
3188 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3190 save01->ldtr.selector, save01->ldtr.attrib,
3191 save01->ldtr.limit, save01->ldtr.base);
3192 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3194 save->idtr.selector, save->idtr.attrib,
3195 save->idtr.limit, save->idtr.base);
3196 pr_err("%-5s s: %04x a: %04x l: %08x b: %016llx\n",
3198 save01->tr.selector, save01->tr.attrib,
3199 save01->tr.limit, save01->tr.base);
3200 pr_err("cpl: %d efer: %016llx\n",
3201 save->cpl, save->efer);
3202 pr_err("%-15s %016llx %-13s %016llx\n",
3203 "cr0:", save->cr0, "cr2:", save->cr2);
3204 pr_err("%-15s %016llx %-13s %016llx\n",
3205 "cr3:", save->cr3, "cr4:", save->cr4);
3206 pr_err("%-15s %016llx %-13s %016llx\n",
3207 "dr6:", save->dr6, "dr7:", save->dr7);
3208 pr_err("%-15s %016llx %-13s %016llx\n",
3209 "rip:", save->rip, "rflags:", save->rflags);
3210 pr_err("%-15s %016llx %-13s %016llx\n",
3211 "rsp:", save->rsp, "rax:", save->rax);
3212 pr_err("%-15s %016llx %-13s %016llx\n",
3213 "star:", save01->star, "lstar:", save01->lstar);
3214 pr_err("%-15s %016llx %-13s %016llx\n",
3215 "cstar:", save01->cstar, "sfmask:", save01->sfmask);
3216 pr_err("%-15s %016llx %-13s %016llx\n",
3217 "kernel_gs_base:", save01->kernel_gs_base,
3218 "sysenter_cs:", save01->sysenter_cs);
3219 pr_err("%-15s %016llx %-13s %016llx\n",
3220 "sysenter_esp:", save01->sysenter_esp,
3221 "sysenter_eip:", save01->sysenter_eip);
3222 pr_err("%-15s %016llx %-13s %016llx\n",
3223 "gpat:", save->g_pat, "dbgctl:", save->dbgctl);
3224 pr_err("%-15s %016llx %-13s %016llx\n",
3225 "br_from:", save->br_from, "br_to:", save->br_to);
3226 pr_err("%-15s %016llx %-13s %016llx\n",
3227 "excp_from:", save->last_excp_from,
3228 "excp_to:", save->last_excp_to);
3231 static int svm_handle_invalid_exit(struct kvm_vcpu *vcpu, u64 exit_code)
3233 if (exit_code < ARRAY_SIZE(svm_exit_handlers) &&
3234 svm_exit_handlers[exit_code])
3237 vcpu_unimpl(vcpu, "svm: unexpected exit reason 0x%llx\n", exit_code);
3239 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3240 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
3241 vcpu->run->internal.ndata = 2;
3242 vcpu->run->internal.data[0] = exit_code;
3243 vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
3248 int svm_invoke_exit_handler(struct kvm_vcpu *vcpu, u64 exit_code)
3250 if (svm_handle_invalid_exit(vcpu, exit_code))
3253 #ifdef CONFIG_RETPOLINE
3254 if (exit_code == SVM_EXIT_MSR)
3255 return msr_interception(vcpu);
3256 else if (exit_code == SVM_EXIT_VINTR)
3257 return interrupt_window_interception(vcpu);
3258 else if (exit_code == SVM_EXIT_INTR)
3259 return intr_interception(vcpu);
3260 else if (exit_code == SVM_EXIT_HLT)
3261 return kvm_emulate_halt(vcpu);
3262 else if (exit_code == SVM_EXIT_NPF)
3263 return npf_interception(vcpu);
3265 return svm_exit_handlers[exit_code](vcpu);
3268 static void svm_get_exit_info(struct kvm_vcpu *vcpu, u64 *info1, u64 *info2,
3269 u32 *intr_info, u32 *error_code)
3271 struct vmcb_control_area *control = &to_svm(vcpu)->vmcb->control;
3273 *info1 = control->exit_info_1;
3274 *info2 = control->exit_info_2;
3275 *intr_info = control->exit_int_info;
3276 if ((*intr_info & SVM_EXITINTINFO_VALID) &&
3277 (*intr_info & SVM_EXITINTINFO_VALID_ERR))
3278 *error_code = control->exit_int_info_err;
3283 static int handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)
3285 struct vcpu_svm *svm = to_svm(vcpu);
3286 struct kvm_run *kvm_run = vcpu->run;
3287 u32 exit_code = svm->vmcb->control.exit_code;
3289 trace_kvm_exit(exit_code, vcpu, KVM_ISA_SVM);
3291 /* SEV-ES guests must use the CR write traps to track CR registers. */
3292 if (!sev_es_guest(vcpu->kvm)) {
3293 if (!svm_is_intercept(svm, INTERCEPT_CR0_WRITE))
3294 vcpu->arch.cr0 = svm->vmcb->save.cr0;
3296 vcpu->arch.cr3 = svm->vmcb->save.cr3;
3299 if (is_guest_mode(vcpu)) {
3302 trace_kvm_nested_vmexit(exit_code, vcpu, KVM_ISA_SVM);
3304 vmexit = nested_svm_exit_special(svm);
3306 if (vmexit == NESTED_EXIT_CONTINUE)
3307 vmexit = nested_svm_exit_handled(svm);
3309 if (vmexit == NESTED_EXIT_DONE)
3313 if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
3314 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3315 kvm_run->fail_entry.hardware_entry_failure_reason
3316 = svm->vmcb->control.exit_code;
3317 kvm_run->fail_entry.cpu = vcpu->arch.last_vmentry_cpu;
3322 if (is_external_interrupt(svm->vmcb->control.exit_int_info) &&
3323 exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR &&
3324 exit_code != SVM_EXIT_NPF && exit_code != SVM_EXIT_TASK_SWITCH &&
3325 exit_code != SVM_EXIT_INTR && exit_code != SVM_EXIT_NMI)
3326 printk(KERN_ERR "%s: unexpected exit_int_info 0x%x "
3328 __func__, svm->vmcb->control.exit_int_info,
3331 if (exit_fastpath != EXIT_FASTPATH_NONE)
3334 return svm_invoke_exit_handler(vcpu, exit_code);
3337 static void reload_tss(struct kvm_vcpu *vcpu)
3339 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3341 sd->tss_desc->type = 9; /* available 32/64-bit TSS */
3345 static void pre_svm_run(struct kvm_vcpu *vcpu)
3347 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3348 struct vcpu_svm *svm = to_svm(vcpu);
3351 * If the previous vmrun of the vmcb occurred on a different physical
3352 * cpu, then mark the vmcb dirty and assign a new asid. Hardware's
3353 * vmcb clean bits are per logical CPU, as are KVM's asid assignments.
3355 if (unlikely(svm->current_vmcb->cpu != vcpu->cpu)) {
3356 svm->current_vmcb->asid_generation = 0;
3357 vmcb_mark_all_dirty(svm->vmcb);
3358 svm->current_vmcb->cpu = vcpu->cpu;
3361 if (sev_guest(vcpu->kvm))
3362 return pre_sev_run(svm, vcpu->cpu);
3364 /* FIXME: handle wraparound of asid_generation */
3365 if (svm->current_vmcb->asid_generation != sd->asid_generation)
3369 static void svm_inject_nmi(struct kvm_vcpu *vcpu)
3371 struct vcpu_svm *svm = to_svm(vcpu);
3373 svm->vmcb->control.event_inj = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
3374 vcpu->arch.hflags |= HF_NMI_MASK;
3375 if (!sev_es_guest(vcpu->kvm))
3376 svm_set_intercept(svm, INTERCEPT_IRET);
3377 ++vcpu->stat.nmi_injections;
3380 static void svm_set_irq(struct kvm_vcpu *vcpu)
3382 struct vcpu_svm *svm = to_svm(vcpu);
3384 BUG_ON(!(gif_set(svm)));
3386 trace_kvm_inj_virq(vcpu->arch.interrupt.nr);
3387 ++vcpu->stat.irq_injections;
3389 svm->vmcb->control.event_inj = vcpu->arch.interrupt.nr |
3390 SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_INTR;
3393 static void svm_update_cr8_intercept(struct kvm_vcpu *vcpu, int tpr, int irr)
3395 struct vcpu_svm *svm = to_svm(vcpu);
3398 * SEV-ES guests must always keep the CR intercepts cleared. CR
3399 * tracking is done using the CR write traps.
3401 if (sev_es_guest(vcpu->kvm))
3404 if (nested_svm_virtualize_tpr(vcpu))
3407 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3413 svm_set_intercept(svm, INTERCEPT_CR8_WRITE);
3416 bool svm_nmi_blocked(struct kvm_vcpu *vcpu)
3418 struct vcpu_svm *svm = to_svm(vcpu);
3419 struct vmcb *vmcb = svm->vmcb;
3425 if (is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3428 ret = (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK) ||
3429 (vcpu->arch.hflags & HF_NMI_MASK);
3434 static int svm_nmi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3436 struct vcpu_svm *svm = to_svm(vcpu);
3437 if (svm->nested.nested_run_pending)
3440 /* An NMI must not be injected into L2 if it's supposed to VM-Exit. */
3441 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_nmi(svm))
3444 return !svm_nmi_blocked(vcpu);
3447 static bool svm_get_nmi_mask(struct kvm_vcpu *vcpu)
3449 return !!(vcpu->arch.hflags & HF_NMI_MASK);
3452 static void svm_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked)
3454 struct vcpu_svm *svm = to_svm(vcpu);
3457 vcpu->arch.hflags |= HF_NMI_MASK;
3458 if (!sev_es_guest(vcpu->kvm))
3459 svm_set_intercept(svm, INTERCEPT_IRET);
3461 vcpu->arch.hflags &= ~HF_NMI_MASK;
3462 if (!sev_es_guest(vcpu->kvm))
3463 svm_clr_intercept(svm, INTERCEPT_IRET);
3467 bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
3469 struct vcpu_svm *svm = to_svm(vcpu);
3470 struct vmcb *vmcb = svm->vmcb;
3475 if (sev_es_guest(vcpu->kvm)) {
3477 * SEV-ES guests to not expose RFLAGS. Use the VMCB interrupt mask
3478 * bit to determine the state of the IF flag.
3480 if (!(vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK))
3482 } else if (is_guest_mode(vcpu)) {
3483 /* As long as interrupts are being delivered... */
3484 if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
3485 ? !(svm->vmcb01.ptr->save.rflags & X86_EFLAGS_IF)
3486 : !(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3489 /* ... vmexits aren't blocked by the interrupt shadow */
3490 if (nested_exit_on_intr(svm))
3493 if (!(kvm_get_rflags(vcpu) & X86_EFLAGS_IF))
3497 return (vmcb->control.int_state & SVM_INTERRUPT_SHADOW_MASK);
3500 static int svm_interrupt_allowed(struct kvm_vcpu *vcpu, bool for_injection)
3502 struct vcpu_svm *svm = to_svm(vcpu);
3503 if (svm->nested.nested_run_pending)
3507 * An IRQ must not be injected into L2 if it's supposed to VM-Exit,
3508 * e.g. if the IRQ arrived asynchronously after checking nested events.
3510 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_intr(svm))
3513 return !svm_interrupt_blocked(vcpu);
3516 static void svm_enable_irq_window(struct kvm_vcpu *vcpu)
3518 struct vcpu_svm *svm = to_svm(vcpu);
3521 * In case GIF=0 we can't rely on the CPU to tell us when GIF becomes
3522 * 1, because that's a separate STGI/VMRUN intercept. The next time we
3523 * get that intercept, this function will be called again though and
3524 * we'll get the vintr intercept. However, if the vGIF feature is
3525 * enabled, the STGI interception will not occur. Enable the irq
3526 * window under the assumption that the hardware will set the GIF.
3528 if (vgif_enabled(svm) || gif_set(svm)) {
3530 * IRQ window is not needed when AVIC is enabled,
3531 * unless we have pending ExtINT since it cannot be injected
3532 * via AVIC. In such case, we need to temporarily disable AVIC,
3533 * and fallback to injecting IRQ via V_IRQ.
3535 svm_toggle_avic_for_irq_window(vcpu, false);
3540 static void svm_enable_nmi_window(struct kvm_vcpu *vcpu)
3542 struct vcpu_svm *svm = to_svm(vcpu);
3544 if ((vcpu->arch.hflags & (HF_NMI_MASK | HF_IRET_MASK)) == HF_NMI_MASK)
3545 return; /* IRET will cause a vm exit */
3547 if (!gif_set(svm)) {
3548 if (vgif_enabled(svm))
3549 svm_set_intercept(svm, INTERCEPT_STGI);
3550 return; /* STGI will cause a vm exit */
3554 * Something prevents NMI from been injected. Single step over possible
3555 * problem (IRET or exception injection or interrupt shadow)
3557 svm->nmi_singlestep_guest_rflags = svm_get_rflags(vcpu);
3558 svm->nmi_singlestep = true;
3559 svm->vmcb->save.rflags |= (X86_EFLAGS_TF | X86_EFLAGS_RF);
3562 static int svm_set_tss_addr(struct kvm *kvm, unsigned int addr)
3567 static int svm_set_identity_map_addr(struct kvm *kvm, u64 ident_addr)
3572 void svm_flush_tlb(struct kvm_vcpu *vcpu)
3574 struct vcpu_svm *svm = to_svm(vcpu);
3577 * Flush only the current ASID even if the TLB flush was invoked via
3578 * kvm_flush_remote_tlbs(). Although flushing remote TLBs requires all
3579 * ASIDs to be flushed, KVM uses a single ASID for L1 and L2, and
3580 * unconditionally does a TLB flush on both nested VM-Enter and nested
3581 * VM-Exit (via kvm_mmu_reset_context()).
3583 if (static_cpu_has(X86_FEATURE_FLUSHBYASID))
3584 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3586 svm->current_vmcb->asid_generation--;
3589 static void svm_flush_tlb_gva(struct kvm_vcpu *vcpu, gva_t gva)
3591 struct vcpu_svm *svm = to_svm(vcpu);
3593 invlpga(gva, svm->vmcb->control.asid);
3596 static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)
3598 struct vcpu_svm *svm = to_svm(vcpu);
3600 if (nested_svm_virtualize_tpr(vcpu))
3603 if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
3604 int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
3605 kvm_set_cr8(vcpu, cr8);
3609 static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
3611 struct vcpu_svm *svm = to_svm(vcpu);
3614 if (nested_svm_virtualize_tpr(vcpu) ||
3615 kvm_vcpu_apicv_active(vcpu))
3618 cr8 = kvm_get_cr8(vcpu);
3619 svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
3620 svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
3623 static void svm_complete_interrupts(struct kvm_vcpu *vcpu)
3625 struct vcpu_svm *svm = to_svm(vcpu);
3628 u32 exitintinfo = svm->vmcb->control.exit_int_info;
3629 unsigned int3_injected = svm->int3_injected;
3631 svm->int3_injected = 0;
3634 * If we've made progress since setting HF_IRET_MASK, we've
3635 * executed an IRET and can allow NMI injection.
3637 if ((vcpu->arch.hflags & HF_IRET_MASK) &&
3638 (sev_es_guest(vcpu->kvm) ||
3639 kvm_rip_read(vcpu) != svm->nmi_iret_rip)) {
3640 vcpu->arch.hflags &= ~(HF_NMI_MASK | HF_IRET_MASK);
3641 kvm_make_request(KVM_REQ_EVENT, vcpu);
3644 vcpu->arch.nmi_injected = false;
3645 kvm_clear_exception_queue(vcpu);
3646 kvm_clear_interrupt_queue(vcpu);
3648 if (!(exitintinfo & SVM_EXITINTINFO_VALID))
3651 kvm_make_request(KVM_REQ_EVENT, vcpu);
3653 vector = exitintinfo & SVM_EXITINTINFO_VEC_MASK;
3654 type = exitintinfo & SVM_EXITINTINFO_TYPE_MASK;
3657 case SVM_EXITINTINFO_TYPE_NMI:
3658 vcpu->arch.nmi_injected = true;
3660 case SVM_EXITINTINFO_TYPE_EXEPT:
3662 * Never re-inject a #VC exception.
3664 if (vector == X86_TRAP_VC)
3668 * In case of software exceptions, do not reinject the vector,
3669 * but re-execute the instruction instead. Rewind RIP first
3670 * if we emulated INT3 before.
3672 if (kvm_exception_is_soft(vector)) {
3673 if (vector == BP_VECTOR && int3_injected &&
3674 kvm_is_linear_rip(vcpu, svm->int3_rip))
3676 kvm_rip_read(vcpu) - int3_injected);
3679 if (exitintinfo & SVM_EXITINTINFO_VALID_ERR) {
3680 u32 err = svm->vmcb->control.exit_int_info_err;
3681 kvm_requeue_exception_e(vcpu, vector, err);
3684 kvm_requeue_exception(vcpu, vector);
3686 case SVM_EXITINTINFO_TYPE_INTR:
3687 kvm_queue_interrupt(vcpu, vector, false);
3694 static void svm_cancel_injection(struct kvm_vcpu *vcpu)
3696 struct vcpu_svm *svm = to_svm(vcpu);
3697 struct vmcb_control_area *control = &svm->vmcb->control;
3699 control->exit_int_info = control->event_inj;
3700 control->exit_int_info_err = control->event_inj_err;
3701 control->event_inj = 0;
3702 svm_complete_interrupts(vcpu);
3705 static fastpath_t svm_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
3707 if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_MSR &&
3708 to_svm(vcpu)->vmcb->control.exit_info_1)
3709 return handle_fastpath_set_msr_irqoff(vcpu);
3711 return EXIT_FASTPATH_NONE;
3714 static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
3716 struct vcpu_svm *svm = to_svm(vcpu);
3717 unsigned long vmcb_pa = svm->current_vmcb->pa;
3720 * VMENTER enables interrupts (host state), but the kernel state is
3721 * interrupts disabled when this is invoked. Also tell RCU about
3722 * it. This is the same logic as for exit_to_user_mode().
3724 * This ensures that e.g. latency analysis on the host observes
3725 * guest mode as interrupt enabled.
3727 * guest_enter_irqoff() informs context tracking about the
3728 * transition to guest mode and if enabled adjusts RCU state
3731 instrumentation_begin();
3732 trace_hardirqs_on_prepare();
3733 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
3734 instrumentation_end();
3736 guest_enter_irqoff();
3737 lockdep_hardirqs_on(CALLER_ADDR0);
3739 if (sev_es_guest(vcpu->kvm)) {
3740 __svm_sev_es_vcpu_run(vmcb_pa);
3742 struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
3745 * Use a single vmcb (vmcb01 because it's always valid) for
3746 * context switching guest state via VMLOAD/VMSAVE, that way
3747 * the state doesn't need to be copied between vmcb01 and
3748 * vmcb02 when switching vmcbs for nested virtualization.
3750 vmload(svm->vmcb01.pa);
3751 __svm_vcpu_run(vmcb_pa, (unsigned long *)&vcpu->arch.regs);
3752 vmsave(svm->vmcb01.pa);
3754 vmload(__sme_page_pa(sd->save_area));
3758 * VMEXIT disables interrupts (host state), but tracing and lockdep
3759 * have them in state 'on' as recorded before entering guest mode.
3760 * Same as enter_from_user_mode().
3762 * guest_exit_irqoff() restores host context and reinstates RCU if
3763 * enabled and required.
3765 * This needs to be done before the below as native_read_msr()
3766 * contains a tracepoint and x86_spec_ctrl_restore_host() calls
3767 * into world and some more.
3769 lockdep_hardirqs_off(CALLER_ADDR0);
3770 guest_exit_irqoff();
3772 instrumentation_begin();
3773 trace_hardirqs_off_finish();
3774 instrumentation_end();
3777 static __no_kcsan fastpath_t svm_vcpu_run(struct kvm_vcpu *vcpu)
3779 struct vcpu_svm *svm = to_svm(vcpu);
3781 trace_kvm_entry(vcpu);
3783 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
3784 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
3785 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
3788 * Disable singlestep if we're injecting an interrupt/exception.
3789 * We don't want our modified rflags to be pushed on the stack where
3790 * we might not be able to easily reset them if we disabled NMI
3793 if (svm->nmi_singlestep && svm->vmcb->control.event_inj) {
3795 * Event injection happens before external interrupts cause a
3796 * vmexit and interrupts are disabled here, so smp_send_reschedule
3797 * is enough to force an immediate vmexit.
3799 disable_nmi_singlestep(svm);
3800 smp_send_reschedule(vcpu->cpu);
3805 sync_lapic_to_cr8(vcpu);
3807 if (unlikely(svm->asid != svm->vmcb->control.asid)) {
3808 svm->vmcb->control.asid = svm->asid;
3809 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
3811 svm->vmcb->save.cr2 = vcpu->arch.cr2;
3814 * Run with all-zero DR6 unless needed, so that we can get the exact cause
3817 if (unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT))
3818 svm_set_dr6(svm, vcpu->arch.dr6);
3820 svm_set_dr6(svm, DR6_ACTIVE_LOW);
3823 kvm_load_guest_xsave_state(vcpu);
3825 kvm_wait_lapic_expire(vcpu);
3828 * If this vCPU has touched SPEC_CTRL, restore the guest's value if
3829 * it's non-zero. Since vmentry is serialising on affected CPUs, there
3830 * is no need to worry about the conditional branch over the wrmsr
3831 * being speculatively taken.
3833 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3834 x86_spec_ctrl_set_guest(svm->spec_ctrl, svm->virt_spec_ctrl);
3836 svm_vcpu_enter_exit(vcpu);
3839 * We do not use IBRS in the kernel. If this vCPU has used the
3840 * SPEC_CTRL MSR it may have left it on; save the value and
3841 * turn it off. This is much more efficient than blindly adding
3842 * it to the atomic save/restore list. Especially as the former
3843 * (Saving guest MSRs on vmexit) doesn't even exist in KVM.
3845 * For non-nested case:
3846 * If the L01 MSR bitmap does not intercept the MSR, then we need to
3850 * If the L02 MSR bitmap does not intercept the MSR, then we need to
3853 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL) &&
3854 unlikely(!msr_write_intercepted(vcpu, MSR_IA32_SPEC_CTRL)))
3855 svm->spec_ctrl = native_read_msr(MSR_IA32_SPEC_CTRL);
3857 if (!sev_es_guest(vcpu->kvm))
3860 if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
3861 x86_spec_ctrl_restore_host(svm->spec_ctrl, svm->virt_spec_ctrl);
3863 if (!sev_es_guest(vcpu->kvm)) {
3864 vcpu->arch.cr2 = svm->vmcb->save.cr2;
3865 vcpu->arch.regs[VCPU_REGS_RAX] = svm->vmcb->save.rax;
3866 vcpu->arch.regs[VCPU_REGS_RSP] = svm->vmcb->save.rsp;
3867 vcpu->arch.regs[VCPU_REGS_RIP] = svm->vmcb->save.rip;
3870 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3871 kvm_before_interrupt(vcpu);
3873 kvm_load_host_xsave_state(vcpu);
3876 /* Any pending NMI will happen here */
3878 if (unlikely(svm->vmcb->control.exit_code == SVM_EXIT_NMI))
3879 kvm_after_interrupt(vcpu);
3881 sync_cr8_to_lapic(vcpu);
3884 if (is_guest_mode(vcpu)) {
3885 nested_sync_control_from_vmcb02(svm);
3886 svm->nested.nested_run_pending = 0;
3889 svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
3890 vmcb_mark_all_clean(svm->vmcb);
3892 /* if exit due to PF check for async PF */
3893 if (svm->vmcb->control.exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR)
3894 vcpu->arch.apf.host_apf_flags =
3895 kvm_read_and_reset_apf_flags();
3898 vcpu->arch.regs_avail &= ~(1 << VCPU_EXREG_PDPTR);
3899 vcpu->arch.regs_dirty &= ~(1 << VCPU_EXREG_PDPTR);
3903 * We need to handle MC intercepts here before the vcpu has a chance to
3904 * change the physical cpu
3906 if (unlikely(svm->vmcb->control.exit_code ==
3907 SVM_EXIT_EXCP_BASE + MC_VECTOR))
3908 svm_handle_mce(vcpu);
3910 svm_complete_interrupts(vcpu);
3912 if (is_guest_mode(vcpu))
3913 return EXIT_FASTPATH_NONE;
3915 return svm_exit_handlers_fastpath(vcpu);
3918 static void svm_load_mmu_pgd(struct kvm_vcpu *vcpu, hpa_t root_hpa,
3921 struct vcpu_svm *svm = to_svm(vcpu);
3925 svm->vmcb->control.nested_cr3 = __sme_set(root_hpa);
3926 vmcb_mark_dirty(svm->vmcb, VMCB_NPT);
3928 /* Loading L2's CR3 is handled by enter_svm_guest_mode. */
3929 if (!test_bit(VCPU_EXREG_CR3, (ulong *)&vcpu->arch.regs_avail))
3931 cr3 = vcpu->arch.cr3;
3932 } else if (vcpu->arch.mmu->shadow_root_level >= PT64_ROOT_4LEVEL) {
3933 cr3 = __sme_set(root_hpa) | kvm_get_active_pcid(vcpu);
3935 /* PCID in the guest should be impossible with a 32-bit MMU. */
3936 WARN_ON_ONCE(kvm_get_active_pcid(vcpu));
3940 svm->vmcb->save.cr3 = cr3;
3941 vmcb_mark_dirty(svm->vmcb, VMCB_CR);
3944 static int is_disabled(void)
3948 rdmsrl(MSR_VM_CR, vm_cr);
3949 if (vm_cr & (1 << SVM_VM_CR_SVM_DISABLE))
3956 svm_patch_hypercall(struct kvm_vcpu *vcpu, unsigned char *hypercall)
3959 * Patch in the VMMCALL instruction:
3961 hypercall[0] = 0x0f;
3962 hypercall[1] = 0x01;
3963 hypercall[2] = 0xd9;
3966 static int __init svm_check_processor_compat(void)
3971 static bool svm_cpu_has_accelerated_tpr(void)
3977 * The kvm parameter can be NULL (module initialization, or invocation before
3978 * VM creation). Be sure to check the kvm parameter before using it.
3980 static bool svm_has_emulated_msr(struct kvm *kvm, u32 index)
3983 case MSR_IA32_MCG_EXT_CTL:
3984 case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:
3986 case MSR_IA32_SMBASE:
3987 /* SEV-ES guests do not support SMM, so report false */
3988 if (kvm && sev_es_guest(kvm))
3998 static u64 svm_get_mt_mask(struct kvm_vcpu *vcpu, gfn_t gfn, bool is_mmio)
4003 static void svm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
4005 struct vcpu_svm *svm = to_svm(vcpu);
4006 struct kvm_cpuid_entry2 *best;
4008 vcpu->arch.xsaves_enabled = guest_cpuid_has(vcpu, X86_FEATURE_XSAVE) &&
4009 boot_cpu_has(X86_FEATURE_XSAVE) &&
4010 boot_cpu_has(X86_FEATURE_XSAVES);
4012 /* Update nrips enabled cache */
4013 svm->nrips_enabled = kvm_cpu_cap_has(X86_FEATURE_NRIPS) &&
4014 guest_cpuid_has(vcpu, X86_FEATURE_NRIPS);
4016 /* Check again if INVPCID interception if required */
4017 svm_check_invpcid(svm);
4019 /* For sev guests, the memory encryption bit is not reserved in CR3. */
4020 if (sev_guest(vcpu->kvm)) {
4021 best = kvm_find_cpuid_entry(vcpu, 0x8000001F, 0);
4023 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
4026 if (kvm_vcpu_apicv_active(vcpu)) {
4028 * AVIC does not work with an x2APIC mode guest. If the X2APIC feature
4029 * is exposed to the guest, disable AVIC.
4031 if (guest_cpuid_has(vcpu, X86_FEATURE_X2APIC))
4032 kvm_request_apicv_update(vcpu->kvm, false,
4033 APICV_INHIBIT_REASON_X2APIC);
4036 * Currently, AVIC does not work with nested virtualization.
4037 * So, we disable AVIC when cpuid for SVM is set in the L1 guest.
4039 if (nested && guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4040 kvm_request_apicv_update(vcpu->kvm, false,
4041 APICV_INHIBIT_REASON_NESTED);
4044 if (guest_cpuid_is_intel(vcpu)) {
4046 * We must intercept SYSENTER_EIP and SYSENTER_ESP
4047 * accesses because the processor only stores 32 bits.
4048 * For the same reason we cannot use virtual VMLOAD/VMSAVE.
4050 svm_set_intercept(svm, INTERCEPT_VMLOAD);
4051 svm_set_intercept(svm, INTERCEPT_VMSAVE);
4052 svm->vmcb->control.virt_ext &= ~VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
4054 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 0, 0);
4055 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 0, 0);
4058 * If hardware supports Virtual VMLOAD VMSAVE then enable it
4059 * in VMCB and clear intercepts to avoid #VMEXIT.
4062 svm_clr_intercept(svm, INTERCEPT_VMLOAD);
4063 svm_clr_intercept(svm, INTERCEPT_VMSAVE);
4064 svm->vmcb->control.virt_ext |= VIRTUAL_VMLOAD_VMSAVE_ENABLE_MASK;
4066 /* No need to intercept these MSRs */
4067 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_EIP, 1, 1);
4068 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_SYSENTER_ESP, 1, 1);
4072 static bool svm_has_wbinvd_exit(void)
4077 #define PRE_EX(exit) { .exit_code = (exit), \
4078 .stage = X86_ICPT_PRE_EXCEPT, }
4079 #define POST_EX(exit) { .exit_code = (exit), \
4080 .stage = X86_ICPT_POST_EXCEPT, }
4081 #define POST_MEM(exit) { .exit_code = (exit), \
4082 .stage = X86_ICPT_POST_MEMACCESS, }
4084 static const struct __x86_intercept {
4086 enum x86_intercept_stage stage;
4087 } x86_intercept_map[] = {
4088 [x86_intercept_cr_read] = POST_EX(SVM_EXIT_READ_CR0),
4089 [x86_intercept_cr_write] = POST_EX(SVM_EXIT_WRITE_CR0),
4090 [x86_intercept_clts] = POST_EX(SVM_EXIT_WRITE_CR0),
4091 [x86_intercept_lmsw] = POST_EX(SVM_EXIT_WRITE_CR0),
4092 [x86_intercept_smsw] = POST_EX(SVM_EXIT_READ_CR0),
4093 [x86_intercept_dr_read] = POST_EX(SVM_EXIT_READ_DR0),
4094 [x86_intercept_dr_write] = POST_EX(SVM_EXIT_WRITE_DR0),
4095 [x86_intercept_sldt] = POST_EX(SVM_EXIT_LDTR_READ),
4096 [x86_intercept_str] = POST_EX(SVM_EXIT_TR_READ),
4097 [x86_intercept_lldt] = POST_EX(SVM_EXIT_LDTR_WRITE),
4098 [x86_intercept_ltr] = POST_EX(SVM_EXIT_TR_WRITE),
4099 [x86_intercept_sgdt] = POST_EX(SVM_EXIT_GDTR_READ),
4100 [x86_intercept_sidt] = POST_EX(SVM_EXIT_IDTR_READ),
4101 [x86_intercept_lgdt] = POST_EX(SVM_EXIT_GDTR_WRITE),
4102 [x86_intercept_lidt] = POST_EX(SVM_EXIT_IDTR_WRITE),
4103 [x86_intercept_vmrun] = POST_EX(SVM_EXIT_VMRUN),
4104 [x86_intercept_vmmcall] = POST_EX(SVM_EXIT_VMMCALL),
4105 [x86_intercept_vmload] = POST_EX(SVM_EXIT_VMLOAD),
4106 [x86_intercept_vmsave] = POST_EX(SVM_EXIT_VMSAVE),
4107 [x86_intercept_stgi] = POST_EX(SVM_EXIT_STGI),
4108 [x86_intercept_clgi] = POST_EX(SVM_EXIT_CLGI),
4109 [x86_intercept_skinit] = POST_EX(SVM_EXIT_SKINIT),
4110 [x86_intercept_invlpga] = POST_EX(SVM_EXIT_INVLPGA),
4111 [x86_intercept_rdtscp] = POST_EX(SVM_EXIT_RDTSCP),
4112 [x86_intercept_monitor] = POST_MEM(SVM_EXIT_MONITOR),
4113 [x86_intercept_mwait] = POST_EX(SVM_EXIT_MWAIT),
4114 [x86_intercept_invlpg] = POST_EX(SVM_EXIT_INVLPG),
4115 [x86_intercept_invd] = POST_EX(SVM_EXIT_INVD),
4116 [x86_intercept_wbinvd] = POST_EX(SVM_EXIT_WBINVD),
4117 [x86_intercept_wrmsr] = POST_EX(SVM_EXIT_MSR),
4118 [x86_intercept_rdtsc] = POST_EX(SVM_EXIT_RDTSC),
4119 [x86_intercept_rdmsr] = POST_EX(SVM_EXIT_MSR),
4120 [x86_intercept_rdpmc] = POST_EX(SVM_EXIT_RDPMC),
4121 [x86_intercept_cpuid] = PRE_EX(SVM_EXIT_CPUID),
4122 [x86_intercept_rsm] = PRE_EX(SVM_EXIT_RSM),
4123 [x86_intercept_pause] = PRE_EX(SVM_EXIT_PAUSE),
4124 [x86_intercept_pushf] = PRE_EX(SVM_EXIT_PUSHF),
4125 [x86_intercept_popf] = PRE_EX(SVM_EXIT_POPF),
4126 [x86_intercept_intn] = PRE_EX(SVM_EXIT_SWINT),
4127 [x86_intercept_iret] = PRE_EX(SVM_EXIT_IRET),
4128 [x86_intercept_icebp] = PRE_EX(SVM_EXIT_ICEBP),
4129 [x86_intercept_hlt] = POST_EX(SVM_EXIT_HLT),
4130 [x86_intercept_in] = POST_EX(SVM_EXIT_IOIO),
4131 [x86_intercept_ins] = POST_EX(SVM_EXIT_IOIO),
4132 [x86_intercept_out] = POST_EX(SVM_EXIT_IOIO),
4133 [x86_intercept_outs] = POST_EX(SVM_EXIT_IOIO),
4134 [x86_intercept_xsetbv] = PRE_EX(SVM_EXIT_XSETBV),
4141 static int svm_check_intercept(struct kvm_vcpu *vcpu,
4142 struct x86_instruction_info *info,
4143 enum x86_intercept_stage stage,
4144 struct x86_exception *exception)
4146 struct vcpu_svm *svm = to_svm(vcpu);
4147 int vmexit, ret = X86EMUL_CONTINUE;
4148 struct __x86_intercept icpt_info;
4149 struct vmcb *vmcb = svm->vmcb;
4151 if (info->intercept >= ARRAY_SIZE(x86_intercept_map))
4154 icpt_info = x86_intercept_map[info->intercept];
4156 if (stage != icpt_info.stage)
4159 switch (icpt_info.exit_code) {
4160 case SVM_EXIT_READ_CR0:
4161 if (info->intercept == x86_intercept_cr_read)
4162 icpt_info.exit_code += info->modrm_reg;
4164 case SVM_EXIT_WRITE_CR0: {
4165 unsigned long cr0, val;
4167 if (info->intercept == x86_intercept_cr_write)
4168 icpt_info.exit_code += info->modrm_reg;
4170 if (icpt_info.exit_code != SVM_EXIT_WRITE_CR0 ||
4171 info->intercept == x86_intercept_clts)
4174 if (!(vmcb_is_intercept(&svm->nested.ctl,
4175 INTERCEPT_SELECTIVE_CR0)))
4178 cr0 = vcpu->arch.cr0 & ~SVM_CR0_SELECTIVE_MASK;
4179 val = info->src_val & ~SVM_CR0_SELECTIVE_MASK;
4181 if (info->intercept == x86_intercept_lmsw) {
4184 /* lmsw can't clear PE - catch this here */
4185 if (cr0 & X86_CR0_PE)
4190 icpt_info.exit_code = SVM_EXIT_CR0_SEL_WRITE;
4194 case SVM_EXIT_READ_DR0:
4195 case SVM_EXIT_WRITE_DR0:
4196 icpt_info.exit_code += info->modrm_reg;
4199 if (info->intercept == x86_intercept_wrmsr)
4200 vmcb->control.exit_info_1 = 1;
4202 vmcb->control.exit_info_1 = 0;
4204 case SVM_EXIT_PAUSE:
4206 * We get this for NOP only, but pause
4207 * is rep not, check this here
4209 if (info->rep_prefix != REPE_PREFIX)
4212 case SVM_EXIT_IOIO: {
4216 if (info->intercept == x86_intercept_in ||
4217 info->intercept == x86_intercept_ins) {
4218 exit_info = ((info->src_val & 0xffff) << 16) |
4220 bytes = info->dst_bytes;
4222 exit_info = (info->dst_val & 0xffff) << 16;
4223 bytes = info->src_bytes;
4226 if (info->intercept == x86_intercept_outs ||
4227 info->intercept == x86_intercept_ins)
4228 exit_info |= SVM_IOIO_STR_MASK;
4230 if (info->rep_prefix)
4231 exit_info |= SVM_IOIO_REP_MASK;
4233 bytes = min(bytes, 4u);
4235 exit_info |= bytes << SVM_IOIO_SIZE_SHIFT;
4237 exit_info |= (u32)info->ad_bytes << (SVM_IOIO_ASIZE_SHIFT - 1);
4239 vmcb->control.exit_info_1 = exit_info;
4240 vmcb->control.exit_info_2 = info->next_rip;
4248 /* TODO: Advertise NRIPS to guest hypervisor unconditionally */
4249 if (static_cpu_has(X86_FEATURE_NRIPS))
4250 vmcb->control.next_rip = info->next_rip;
4251 vmcb->control.exit_code = icpt_info.exit_code;
4252 vmexit = nested_svm_exit_handled(svm);
4254 ret = (vmexit == NESTED_EXIT_DONE) ? X86EMUL_INTERCEPTED
4261 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
4265 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
4267 if (!kvm_pause_in_guest(vcpu->kvm))
4268 shrink_ple_window(vcpu);
4271 static void svm_setup_mce(struct kvm_vcpu *vcpu)
4273 /* [63:9] are reserved. */
4274 vcpu->arch.mcg_cap &= 0x1ff;
4277 bool svm_smi_blocked(struct kvm_vcpu *vcpu)
4279 struct vcpu_svm *svm = to_svm(vcpu);
4281 /* Per APM Vol.2 15.22.2 "Response to SMI" */
4285 return is_smm(vcpu);
4288 static int svm_smi_allowed(struct kvm_vcpu *vcpu, bool for_injection)
4290 struct vcpu_svm *svm = to_svm(vcpu);
4291 if (svm->nested.nested_run_pending)
4294 /* An SMI must not be injected into L2 if it's supposed to VM-Exit. */
4295 if (for_injection && is_guest_mode(vcpu) && nested_exit_on_smi(svm))
4298 return !svm_smi_blocked(vcpu);
4301 static int svm_pre_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
4303 struct vcpu_svm *svm = to_svm(vcpu);
4306 if (is_guest_mode(vcpu)) {
4307 /* FED8h - SVM Guest */
4308 put_smstate(u64, smstate, 0x7ed8, 1);
4309 /* FEE0h - SVM Guest VMCB Physical Address */
4310 put_smstate(u64, smstate, 0x7ee0, svm->nested.vmcb12_gpa);
4312 svm->vmcb->save.rax = vcpu->arch.regs[VCPU_REGS_RAX];
4313 svm->vmcb->save.rsp = vcpu->arch.regs[VCPU_REGS_RSP];
4314 svm->vmcb->save.rip = vcpu->arch.regs[VCPU_REGS_RIP];
4316 ret = nested_svm_vmexit(svm);
4323 static int svm_pre_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
4325 struct vcpu_svm *svm = to_svm(vcpu);
4326 struct kvm_host_map map;
4329 if (guest_cpuid_has(vcpu, X86_FEATURE_LM)) {
4330 u64 saved_efer = GET_SMSTATE(u64, smstate, 0x7ed0);
4331 u64 guest = GET_SMSTATE(u64, smstate, 0x7ed8);
4332 u64 vmcb12_gpa = GET_SMSTATE(u64, smstate, 0x7ee0);
4335 if (!guest_cpuid_has(vcpu, X86_FEATURE_SVM))
4338 if (!(saved_efer & EFER_SVME))
4341 if (kvm_vcpu_map(vcpu,
4342 gpa_to_gfn(vmcb12_gpa), &map) == -EINVAL)
4345 if (svm_allocate_nested(svm))
4348 ret = enter_svm_guest_mode(vcpu, vmcb12_gpa, map.hva);
4349 kvm_vcpu_unmap(vcpu, &map, true);
4356 static void svm_enable_smi_window(struct kvm_vcpu *vcpu)
4358 struct vcpu_svm *svm = to_svm(vcpu);
4360 if (!gif_set(svm)) {
4361 if (vgif_enabled(svm))
4362 svm_set_intercept(svm, INTERCEPT_STGI);
4363 /* STGI will cause a vm exit */
4365 /* We must be in SMM; RSM will cause a vmexit anyway. */
4369 static bool svm_can_emulate_instruction(struct kvm_vcpu *vcpu, void *insn, int insn_len)
4371 bool smep, smap, is_user;
4375 * When the guest is an SEV-ES guest, emulation is not possible.
4377 if (sev_es_guest(vcpu->kvm))
4381 * Detect and workaround Errata 1096 Fam_17h_00_0Fh.
4384 * When CPU raise #NPF on guest data access and vCPU CR4.SMAP=1, it is
4385 * possible that CPU microcode implementing DecodeAssist will fail
4386 * to read bytes of instruction which caused #NPF. In this case,
4387 * GuestIntrBytes field of the VMCB on a VMEXIT will incorrectly
4388 * return 0 instead of the correct guest instruction bytes.
4390 * This happens because CPU microcode reading instruction bytes
4391 * uses a special opcode which attempts to read data using CPL=0
4392 * priviledges. The microcode reads CS:RIP and if it hits a SMAP
4393 * fault, it gives up and returns no instruction bytes.
4396 * We reach here in case CPU supports DecodeAssist, raised #NPF and
4397 * returned 0 in GuestIntrBytes field of the VMCB.
4398 * First, errata can only be triggered in case vCPU CR4.SMAP=1.
4399 * Second, if vCPU CR4.SMEP=1, errata could only be triggered
4400 * in case vCPU CPL==3 (Because otherwise guest would have triggered
4401 * a SMEP fault instead of #NPF).
4402 * Otherwise, vCPU CR4.SMEP=0, errata could be triggered by any vCPU CPL.
4403 * As most guests enable SMAP if they have also enabled SMEP, use above
4404 * logic in order to attempt minimize false-positive of detecting errata
4405 * while still preserving all cases semantic correctness.
4408 * To determine what instruction the guest was executing, the hypervisor
4409 * will have to decode the instruction at the instruction pointer.
4411 * In non SEV guest, hypervisor will be able to read the guest
4412 * memory to decode the instruction pointer when insn_len is zero
4413 * so we return true to indicate that decoding is possible.
4415 * But in the SEV guest, the guest memory is encrypted with the
4416 * guest specific key and hypervisor will not be able to decode the
4417 * instruction pointer so we will not able to workaround it. Lets
4418 * print the error and request to kill the guest.
4420 if (likely(!insn || insn_len))
4424 * If RIP is invalid, go ahead with emulation which will cause an
4425 * internal error exit.
4427 if (!kvm_vcpu_gfn_to_memslot(vcpu, kvm_rip_read(vcpu) >> PAGE_SHIFT))
4430 cr4 = kvm_read_cr4(vcpu);
4431 smep = cr4 & X86_CR4_SMEP;
4432 smap = cr4 & X86_CR4_SMAP;
4433 is_user = svm_get_cpl(vcpu) == 3;
4434 if (smap && (!smep || is_user)) {
4435 if (!sev_guest(vcpu->kvm))
4438 pr_err_ratelimited("KVM: SEV Guest triggered AMD Erratum 1096\n");
4439 kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
4445 static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
4447 struct vcpu_svm *svm = to_svm(vcpu);
4450 * TODO: Last condition latch INIT signals on vCPU when
4451 * vCPU is in guest-mode and vmcb12 defines intercept on INIT.
4452 * To properly emulate the INIT intercept,
4453 * svm_check_nested_events() should call nested_svm_vmexit()
4454 * if an INIT signal is pending.
4456 return !gif_set(svm) ||
4457 (vmcb_is_intercept(&svm->vmcb->control, INTERCEPT_INIT));
4460 static void svm_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
4462 if (!sev_es_guest(vcpu->kvm))
4463 return kvm_vcpu_deliver_sipi_vector(vcpu, vector);
4465 sev_vcpu_deliver_sipi_vector(vcpu, vector);
4468 static void svm_vm_destroy(struct kvm *kvm)
4470 avic_vm_destroy(kvm);
4471 sev_vm_destroy(kvm);
4474 static int svm_vm_init(struct kvm *kvm)
4476 if (!pause_filter_count || !pause_filter_thresh)
4477 kvm->arch.pause_in_guest = true;
4480 int ret = avic_vm_init(kvm);
4485 kvm_apicv_init(kvm, avic);
4489 static struct kvm_x86_ops svm_x86_ops __initdata = {
4490 .hardware_unsetup = svm_hardware_teardown,
4491 .hardware_enable = svm_hardware_enable,
4492 .hardware_disable = svm_hardware_disable,
4493 .cpu_has_accelerated_tpr = svm_cpu_has_accelerated_tpr,
4494 .has_emulated_msr = svm_has_emulated_msr,
4496 .vcpu_create = svm_create_vcpu,
4497 .vcpu_free = svm_free_vcpu,
4498 .vcpu_reset = svm_vcpu_reset,
4500 .vm_size = sizeof(struct kvm_svm),
4501 .vm_init = svm_vm_init,
4502 .vm_destroy = svm_vm_destroy,
4504 .prepare_guest_switch = svm_prepare_guest_switch,
4505 .vcpu_load = svm_vcpu_load,
4506 .vcpu_put = svm_vcpu_put,
4507 .vcpu_blocking = svm_vcpu_blocking,
4508 .vcpu_unblocking = svm_vcpu_unblocking,
4510 .update_exception_bitmap = svm_update_exception_bitmap,
4511 .get_msr_feature = svm_get_msr_feature,
4512 .get_msr = svm_get_msr,
4513 .set_msr = svm_set_msr,
4514 .get_segment_base = svm_get_segment_base,
4515 .get_segment = svm_get_segment,
4516 .set_segment = svm_set_segment,
4517 .get_cpl = svm_get_cpl,
4518 .get_cs_db_l_bits = kvm_get_cs_db_l_bits,
4519 .set_cr0 = svm_set_cr0,
4520 .is_valid_cr4 = svm_is_valid_cr4,
4521 .set_cr4 = svm_set_cr4,
4522 .set_efer = svm_set_efer,
4523 .get_idt = svm_get_idt,
4524 .set_idt = svm_set_idt,
4525 .get_gdt = svm_get_gdt,
4526 .set_gdt = svm_set_gdt,
4527 .set_dr7 = svm_set_dr7,
4528 .sync_dirty_debug_regs = svm_sync_dirty_debug_regs,
4529 .cache_reg = svm_cache_reg,
4530 .get_rflags = svm_get_rflags,
4531 .set_rflags = svm_set_rflags,
4533 .tlb_flush_all = svm_flush_tlb,
4534 .tlb_flush_current = svm_flush_tlb,
4535 .tlb_flush_gva = svm_flush_tlb_gva,
4536 .tlb_flush_guest = svm_flush_tlb,
4538 .run = svm_vcpu_run,
4539 .handle_exit = handle_exit,
4540 .skip_emulated_instruction = skip_emulated_instruction,
4541 .update_emulated_instruction = NULL,
4542 .set_interrupt_shadow = svm_set_interrupt_shadow,
4543 .get_interrupt_shadow = svm_get_interrupt_shadow,
4544 .patch_hypercall = svm_patch_hypercall,
4545 .set_irq = svm_set_irq,
4546 .set_nmi = svm_inject_nmi,
4547 .queue_exception = svm_queue_exception,
4548 .cancel_injection = svm_cancel_injection,
4549 .interrupt_allowed = svm_interrupt_allowed,
4550 .nmi_allowed = svm_nmi_allowed,
4551 .get_nmi_mask = svm_get_nmi_mask,
4552 .set_nmi_mask = svm_set_nmi_mask,
4553 .enable_nmi_window = svm_enable_nmi_window,
4554 .enable_irq_window = svm_enable_irq_window,
4555 .update_cr8_intercept = svm_update_cr8_intercept,
4556 .set_virtual_apic_mode = svm_set_virtual_apic_mode,
4557 .refresh_apicv_exec_ctrl = svm_refresh_apicv_exec_ctrl,
4558 .check_apicv_inhibit_reasons = svm_check_apicv_inhibit_reasons,
4559 .pre_update_apicv_exec_ctrl = svm_pre_update_apicv_exec_ctrl,
4560 .load_eoi_exitmap = svm_load_eoi_exitmap,
4561 .hwapic_irr_update = svm_hwapic_irr_update,
4562 .hwapic_isr_update = svm_hwapic_isr_update,
4563 .sync_pir_to_irr = kvm_lapic_find_highest_irr,
4564 .apicv_post_state_restore = avic_post_state_restore,
4566 .set_tss_addr = svm_set_tss_addr,
4567 .set_identity_map_addr = svm_set_identity_map_addr,
4568 .get_mt_mask = svm_get_mt_mask,
4570 .get_exit_info = svm_get_exit_info,
4572 .vcpu_after_set_cpuid = svm_vcpu_after_set_cpuid,
4574 .has_wbinvd_exit = svm_has_wbinvd_exit,
4576 .write_l1_tsc_offset = svm_write_l1_tsc_offset,
4578 .load_mmu_pgd = svm_load_mmu_pgd,
4580 .check_intercept = svm_check_intercept,
4581 .handle_exit_irqoff = svm_handle_exit_irqoff,
4583 .request_immediate_exit = __kvm_request_immediate_exit,
4585 .sched_in = svm_sched_in,
4587 .pmu_ops = &amd_pmu_ops,
4588 .nested_ops = &svm_nested_ops,
4590 .deliver_posted_interrupt = svm_deliver_avic_intr,
4591 .dy_apicv_has_pending_interrupt = svm_dy_apicv_has_pending_interrupt,
4592 .update_pi_irte = svm_update_pi_irte,
4593 .setup_mce = svm_setup_mce,
4595 .smi_allowed = svm_smi_allowed,
4596 .pre_enter_smm = svm_pre_enter_smm,
4597 .pre_leave_smm = svm_pre_leave_smm,
4598 .enable_smi_window = svm_enable_smi_window,
4600 .mem_enc_op = svm_mem_enc_op,
4601 .mem_enc_reg_region = svm_register_enc_region,
4602 .mem_enc_unreg_region = svm_unregister_enc_region,
4604 .vm_copy_enc_context_from = svm_vm_copy_asid_from,
4606 .can_emulate_instruction = svm_can_emulate_instruction,
4608 .apic_init_signal_blocked = svm_apic_init_signal_blocked,
4610 .msr_filter_changed = svm_msr_filter_changed,
4611 .complete_emulated_msr = svm_complete_emulated_msr,
4613 .vcpu_deliver_sipi_vector = svm_vcpu_deliver_sipi_vector,
4616 static struct kvm_x86_init_ops svm_init_ops __initdata = {
4617 .cpu_has_kvm_support = has_svm,
4618 .disabled_by_bios = is_disabled,
4619 .hardware_setup = svm_hardware_setup,
4620 .check_processor_compatibility = svm_check_processor_compat,
4622 .runtime_ops = &svm_x86_ops,
4625 static int __init svm_init(void)
4627 __unused_size_checks();
4629 return kvm_init(&svm_init_ops, sizeof(struct vcpu_svm),
4630 __alignof__(struct vcpu_svm), THIS_MODULE);
4633 static void __exit svm_exit(void)
4638 module_init(svm_init)
4639 module_exit(svm_exit)