xref: /kernel/linux/linux-5.10/arch/x86/kvm/cpuid.c (revision 8c2ecf20)
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Kernel-based Virtual Machine driver for Linux
4 * cpuid support routines
5 *
6 * derived from arch/x86/kvm/x86.c
7 *
8 * Copyright 2011 Red Hat, Inc. and/or its affiliates.
9 * Copyright IBM Corporation, 2008
10 */
11
12#include <linux/kvm_host.h>
13#include <linux/export.h>
14#include <linux/vmalloc.h>
15#include <linux/uaccess.h>
16#include <linux/sched/stat.h>
17
18#include <asm/processor.h>
19#include <asm/user.h>
20#include <asm/fpu/xstate.h>
21#include "cpuid.h"
22#include "lapic.h"
23#include "mmu.h"
24#include "trace.h"
25#include "pmu.h"
26
27/*
28 * Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be
29 * aligned to sizeof(unsigned long) because it's not accessed via bitops.
30 */
31u32 kvm_cpu_caps[NCAPINTS] __read_mostly;
32EXPORT_SYMBOL_GPL(kvm_cpu_caps);
33
34static u32 xstate_required_size(u64 xstate_bv, bool compacted)
35{
36	int feature_bit = 0;
37	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
38
39	xstate_bv &= XFEATURE_MASK_EXTEND;
40	while (xstate_bv) {
41		if (xstate_bv & 0x1) {
42		        u32 eax, ebx, ecx, edx, offset;
43		        cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
44			offset = compacted ? ret : ebx;
45			ret = max(ret, offset + eax);
46		}
47
48		xstate_bv >>= 1;
49		feature_bit++;
50	}
51
52	return ret;
53}
54
55#define F feature_bit
56
57static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
58	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u32 index)
59{
60	struct kvm_cpuid_entry2 *e;
61	int i;
62
63	for (i = 0; i < nent; i++) {
64		e = &entries[i];
65
66		if (e->function == function && (e->index == index ||
67		    !(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX)))
68			return e;
69	}
70
71	return NULL;
72}
73
74static int kvm_check_cpuid(struct kvm_cpuid_entry2 *entries, int nent)
75{
76	struct kvm_cpuid_entry2 *best;
77
78	/*
79	 * The existing code assumes virtual address is 48-bit or 57-bit in the
80	 * canonical address checks; exit if it is ever changed.
81	 */
82	best = cpuid_entry2_find(entries, nent, 0x80000008, 0);
83	if (best) {
84		int vaddr_bits = (best->eax & 0xff00) >> 8;
85
86		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
87			return -EINVAL;
88	}
89
90	return 0;
91}
92
93void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
94{
95	struct kvm_cpuid_entry2 *best;
96
97	best = kvm_find_cpuid_entry(vcpu, KVM_CPUID_FEATURES, 0);
98
99	/*
100	 * save the feature bitmap to avoid cpuid lookup for every PV
101	 * operation
102	 */
103	if (best)
104		vcpu->arch.pv_cpuid.features = best->eax;
105}
106
107void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
108{
109	struct kvm_cpuid_entry2 *best;
110
111	best = kvm_find_cpuid_entry(vcpu, 1, 0);
112	if (best) {
113		/* Update OSXSAVE bit */
114		if (boot_cpu_has(X86_FEATURE_XSAVE))
115			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
116				   kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE));
117
118		cpuid_entry_change(best, X86_FEATURE_APIC,
119			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
120	}
121
122	best = kvm_find_cpuid_entry(vcpu, 7, 0);
123	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
124		cpuid_entry_change(best, X86_FEATURE_OSPKE,
125				   kvm_read_cr4_bits(vcpu, X86_CR4_PKE));
126
127	best = kvm_find_cpuid_entry(vcpu, 0xD, 0);
128	if (best)
129		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
130
131	best = kvm_find_cpuid_entry(vcpu, 0xD, 1);
132	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
133		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
134		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
135
136	best = kvm_find_cpuid_entry(vcpu, KVM_CPUID_FEATURES, 0);
137	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
138		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
139		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
140
141	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
142		best = kvm_find_cpuid_entry(vcpu, 0x1, 0);
143		if (best)
144			cpuid_entry_change(best, X86_FEATURE_MWAIT,
145					   vcpu->arch.ia32_misc_enable_msr &
146					   MSR_IA32_MISC_ENABLE_MWAIT);
147	}
148}
149
150static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
151{
152	struct kvm_lapic *apic = vcpu->arch.apic;
153	struct kvm_cpuid_entry2 *best;
154
155	best = kvm_find_cpuid_entry(vcpu, 1, 0);
156	if (best && apic) {
157		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
158			apic->lapic_timer.timer_mode_mask = 3 << 17;
159		else
160			apic->lapic_timer.timer_mode_mask = 1 << 17;
161
162		kvm_apic_set_version(vcpu);
163	}
164
165	best = kvm_find_cpuid_entry(vcpu, 0xD, 0);
166	if (!best)
167		vcpu->arch.guest_supported_xcr0 = 0;
168	else
169		vcpu->arch.guest_supported_xcr0 =
170			(best->eax | ((u64)best->edx << 32)) & supported_xcr0;
171
172	kvm_update_pv_runtime(vcpu);
173
174	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
175	kvm_mmu_reset_context(vcpu);
176
177	kvm_pmu_refresh(vcpu);
178	vcpu->arch.cr4_guest_rsvd_bits =
179	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
180
181	vcpu->arch.cr3_lm_rsvd_bits = rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
182
183	/* Invoke the vendor callback only after the above state is updated. */
184	kvm_x86_ops.vcpu_after_set_cpuid(vcpu);
185}
186
187static int is_efer_nx(void)
188{
189	return host_efer & EFER_NX;
190}
191
192static void cpuid_fix_nx_cap(struct kvm_vcpu *vcpu)
193{
194	int i;
195	struct kvm_cpuid_entry2 *e, *entry;
196
197	entry = NULL;
198	for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
199		e = &vcpu->arch.cpuid_entries[i];
200		if (e->function == 0x80000001) {
201			entry = e;
202			break;
203		}
204	}
205	if (entry && cpuid_entry_has(entry, X86_FEATURE_NX) && !is_efer_nx()) {
206		cpuid_entry_clear(entry, X86_FEATURE_NX);
207		printk(KERN_INFO "kvm: guest NX capability removed\n");
208	}
209}
210
211int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
212{
213	struct kvm_cpuid_entry2 *best;
214
215	best = kvm_find_cpuid_entry(vcpu, 0x80000000, 0);
216	if (!best || best->eax < 0x80000008)
217		goto not_found;
218	best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
219	if (best)
220		return best->eax & 0xff;
221not_found:
222	return 36;
223}
224
225/* when an old userspace process fills a new kernel module */
226int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
227			     struct kvm_cpuid *cpuid,
228			     struct kvm_cpuid_entry __user *entries)
229{
230	int r, i;
231	struct kvm_cpuid_entry *e = NULL;
232	struct kvm_cpuid_entry2 *e2 = NULL;
233
234	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
235		return -E2BIG;
236
237	if (cpuid->nent) {
238		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
239		if (IS_ERR(e))
240			return PTR_ERR(e);
241
242		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
243		if (!e2) {
244			r = -ENOMEM;
245			goto out_free_cpuid;
246		}
247	}
248	for (i = 0; i < cpuid->nent; i++) {
249		e2[i].function = e[i].function;
250		e2[i].eax = e[i].eax;
251		e2[i].ebx = e[i].ebx;
252		e2[i].ecx = e[i].ecx;
253		e2[i].edx = e[i].edx;
254		e2[i].index = 0;
255		e2[i].flags = 0;
256		e2[i].padding[0] = 0;
257		e2[i].padding[1] = 0;
258		e2[i].padding[2] = 0;
259	}
260
261	r = kvm_check_cpuid(e2, cpuid->nent);
262	if (r) {
263		kvfree(e2);
264		goto out_free_cpuid;
265	}
266
267	kvfree(vcpu->arch.cpuid_entries);
268	vcpu->arch.cpuid_entries = e2;
269	vcpu->arch.cpuid_nent = cpuid->nent;
270
271	cpuid_fix_nx_cap(vcpu);
272	kvm_update_cpuid_runtime(vcpu);
273	kvm_vcpu_after_set_cpuid(vcpu);
274
275out_free_cpuid:
276	kvfree(e);
277
278	return r;
279}
280
281int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
282			      struct kvm_cpuid2 *cpuid,
283			      struct kvm_cpuid_entry2 __user *entries)
284{
285	struct kvm_cpuid_entry2 *e2 = NULL;
286	int r;
287
288	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
289		return -E2BIG;
290
291	if (cpuid->nent) {
292		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
293		if (IS_ERR(e2))
294			return PTR_ERR(e2);
295	}
296
297	r = kvm_check_cpuid(e2, cpuid->nent);
298	if (r) {
299		kvfree(e2);
300		return r;
301	}
302
303	kvfree(vcpu->arch.cpuid_entries);
304	vcpu->arch.cpuid_entries = e2;
305	vcpu->arch.cpuid_nent = cpuid->nent;
306
307	kvm_update_cpuid_runtime(vcpu);
308	kvm_vcpu_after_set_cpuid(vcpu);
309
310	return 0;
311}
312
313int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
314			      struct kvm_cpuid2 *cpuid,
315			      struct kvm_cpuid_entry2 __user *entries)
316{
317	int r;
318
319	r = -E2BIG;
320	if (cpuid->nent < vcpu->arch.cpuid_nent)
321		goto out;
322	r = -EFAULT;
323	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
324			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
325		goto out;
326	return 0;
327
328out:
329	cpuid->nent = vcpu->arch.cpuid_nent;
330	return r;
331}
332
333static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
334{
335	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
336	struct kvm_cpuid_entry2 entry;
337
338	reverse_cpuid_check(leaf);
339	kvm_cpu_caps[leaf] &= mask;
340
341	cpuid_count(cpuid.function, cpuid.index,
342		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
343
344	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
345}
346
347void kvm_set_cpu_caps(void)
348{
349	unsigned int f_nx = is_efer_nx() ? F(NX) : 0;
350#ifdef CONFIG_X86_64
351	unsigned int f_gbpages = F(GBPAGES);
352	unsigned int f_lm = F(LM);
353#else
354	unsigned int f_gbpages = 0;
355	unsigned int f_lm = 0;
356#endif
357
358	BUILD_BUG_ON(sizeof(kvm_cpu_caps) >
359		     sizeof(boot_cpu_data.x86_capability));
360
361	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
362	       sizeof(kvm_cpu_caps));
363
364	kvm_cpu_cap_mask(CPUID_1_ECX,
365		/*
366		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
367		 * advertised to guests via CPUID!
368		 */
369		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
370		0 /* DS-CPL, VMX, SMX, EST */ |
371		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
372		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
373		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
374		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
375		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
376		F(F16C) | F(RDRAND)
377	);
378	/* KVM emulates x2apic in software irrespective of host support. */
379	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
380
381	kvm_cpu_cap_mask(CPUID_1_EDX,
382		F(FPU) | F(VME) | F(DE) | F(PSE) |
383		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
384		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
385		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
386		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
387		0 /* Reserved, DS, ACPI */ | F(MMX) |
388		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
389		0 /* HTT, TM, Reserved, PBE */
390	);
391
392	kvm_cpu_cap_mask(CPUID_7_0_EBX,
393		F(FSGSBASE) | F(BMI1) | F(HLE) | F(AVX2) | F(SMEP) |
394		F(BMI2) | F(ERMS) | 0 /*INVPCID*/ | F(RTM) | 0 /*MPX*/ | F(RDSEED) |
395		F(ADX) | F(SMAP) | F(AVX512IFMA) | F(AVX512F) | F(AVX512PF) |
396		F(AVX512ER) | F(AVX512CD) | F(CLFLUSHOPT) | F(CLWB) | F(AVX512DQ) |
397		F(SHA_NI) | F(AVX512BW) | F(AVX512VL) | 0 /*INTEL_PT*/
398	);
399
400	kvm_cpu_cap_mask(CPUID_7_ECX,
401		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
402		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
403		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
404		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/
405	);
406	/* Set LA57 based on hardware capability. */
407	if (cpuid_ecx(7) & F(LA57))
408		kvm_cpu_cap_set(X86_FEATURE_LA57);
409
410	/*
411	 * PKU not yet implemented for shadow paging and requires OSPKE
412	 * to be set on the host. Clear it if that is not the case
413	 */
414	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
415		kvm_cpu_cap_clear(X86_FEATURE_PKU);
416
417	kvm_cpu_cap_mask(CPUID_7_EDX,
418		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
419		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
420		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
421		F(SERIALIZE) | F(TSXLDTRK)
422	);
423
424	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
425	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
426	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
427
428	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
429		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
430	if (boot_cpu_has(X86_FEATURE_STIBP))
431		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
432	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
433		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
434
435	kvm_cpu_cap_mask(CPUID_7_1_EAX,
436		F(AVX512_BF16)
437	);
438
439	kvm_cpu_cap_mask(CPUID_D_1_EAX,
440		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES)
441	);
442
443	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
444		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
445		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
446		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
447		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
448		F(TOPOEXT) | F(PERFCTR_CORE)
449	);
450
451	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
452		F(FPU) | F(VME) | F(DE) | F(PSE) |
453		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
454		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
455		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
456		F(PAT) | F(PSE36) | 0 /* Reserved */ |
457		f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
458		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
459		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
460	);
461
462	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
463		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
464
465	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
466		F(CLZERO) | F(XSAVEERPTR) |
467		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
468		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON)
469	);
470
471	/*
472	 * AMD has separate bits for each SPEC_CTRL bit.
473	 * arch/x86/kernel/cpu/bugs.c is kind enough to
474	 * record that in cpufeatures so use them.
475	 */
476	if (boot_cpu_has(X86_FEATURE_IBPB))
477		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
478	if (boot_cpu_has(X86_FEATURE_IBRS))
479		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
480	if (boot_cpu_has(X86_FEATURE_STIBP))
481		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
482	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
483		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
484	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
485		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
486	/*
487	 * The preference is to use SPEC CTRL MSR instead of the
488	 * VIRT_SPEC MSR.
489	 */
490	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
491	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
492		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
493
494	if (cpu_feature_enabled(X86_FEATURE_SRSO_NO))
495		kvm_cpu_cap_set(X86_FEATURE_SRSO_NO);
496
497	/*
498	 * Hide all SVM features by default, SVM will set the cap bits for
499	 * features it emulates and/or exposes for L1.
500	 */
501	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
502
503	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
504		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
505		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
506		F(PMM) | F(PMM_EN)
507	);
508}
509EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
510
511struct kvm_cpuid_array {
512	struct kvm_cpuid_entry2 *entries;
513	int maxnent;
514	int nent;
515};
516
517static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
518{
519	if (array->nent >= array->maxnent)
520		return NULL;
521
522	return &array->entries[array->nent++];
523}
524
525static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
526					      u32 function, u32 index)
527{
528	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
529
530	if (!entry)
531		return NULL;
532
533	entry->function = function;
534	entry->index = index;
535	entry->flags = 0;
536
537	cpuid_count(entry->function, entry->index,
538		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
539
540	switch (function) {
541	case 4:
542	case 7:
543	case 0xb:
544	case 0xd:
545	case 0xf:
546	case 0x10:
547	case 0x12:
548	case 0x14:
549	case 0x17:
550	case 0x18:
551	case 0x1f:
552	case 0x8000001d:
553		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
554		break;
555	}
556
557	return entry;
558}
559
560static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
561{
562	struct kvm_cpuid_entry2 *entry;
563
564	if (array->nent >= array->maxnent)
565		return -E2BIG;
566
567	entry = &array->entries[array->nent];
568	entry->function = func;
569	entry->index = 0;
570	entry->flags = 0;
571
572	switch (func) {
573	case 0:
574		entry->eax = 7;
575		++array->nent;
576		break;
577	case 1:
578		entry->ecx = F(MOVBE);
579		++array->nent;
580		break;
581	case 7:
582		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
583		entry->eax = 0;
584		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
585			entry->ecx = F(RDPID);
586		++array->nent;
587	default:
588		break;
589	}
590
591	return 0;
592}
593
594static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
595{
596	struct kvm_cpuid_entry2 *entry;
597	int r, i, max_idx;
598
599	/* all calls to cpuid_count() should be made on the same cpu */
600	get_cpu();
601
602	r = -E2BIG;
603
604	entry = do_host_cpuid(array, function, 0);
605	if (!entry)
606		goto out;
607
608	switch (function) {
609	case 0:
610		/* Limited to the highest leaf implemented in KVM. */
611		entry->eax = min(entry->eax, 0x1fU);
612		break;
613	case 1:
614		cpuid_entry_override(entry, CPUID_1_EDX);
615		cpuid_entry_override(entry, CPUID_1_ECX);
616		break;
617	case 2:
618		/*
619		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
620		 * CPUID(function=2, index=0) may return different results each
621		 * time, with the least-significant byte in EAX enumerating the
622		 * number of times software should do CPUID(2, 0).
623		 *
624		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
625		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
626		 * return 01H. Software should ignore this value and not
627		 * interpret it as an informational descriptor", while AMD's
628		 * APM states that CPUID(2) is reserved.
629		 *
630		 * WARN if a frankenstein CPU that supports virtualization and
631		 * a stateful CPUID.0x2 is encountered.
632		 */
633		WARN_ON_ONCE((entry->eax & 0xff) > 1);
634		break;
635	/* functions 4 and 0x8000001d have additional index. */
636	case 4:
637	case 0x8000001d:
638		/*
639		 * Read entries until the cache type in the previous entry is
640		 * zero, i.e. indicates an invalid entry.
641		 */
642		for (i = 1; entry->eax & 0x1f; ++i) {
643			entry = do_host_cpuid(array, function, i);
644			if (!entry)
645				goto out;
646		}
647		break;
648	case 6: /* Thermal management */
649		entry->eax = 0x4; /* allow ARAT */
650		entry->ebx = 0;
651		entry->ecx = 0;
652		entry->edx = 0;
653		break;
654	/* function 7 has additional index. */
655	case 7:
656		entry->eax = min(entry->eax, 1u);
657		cpuid_entry_override(entry, CPUID_7_0_EBX);
658		cpuid_entry_override(entry, CPUID_7_ECX);
659		cpuid_entry_override(entry, CPUID_7_EDX);
660
661		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
662		if (entry->eax == 1) {
663			entry = do_host_cpuid(array, function, 1);
664			if (!entry)
665				goto out;
666
667			cpuid_entry_override(entry, CPUID_7_1_EAX);
668			entry->ebx = 0;
669			entry->ecx = 0;
670			entry->edx = 0;
671		}
672		break;
673	case 0xa: { /* Architectural Performance Monitoring */
674		struct x86_pmu_capability cap;
675		union cpuid10_eax eax;
676		union cpuid10_edx edx;
677
678		if (!static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
679			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
680			break;
681		}
682
683		perf_get_x86_pmu_capability(&cap);
684
685		/*
686		 * Only support guest architectural pmu on a host
687		 * with architectural pmu.
688		 */
689		if (!cap.version)
690			memset(&cap, 0, sizeof(cap));
691
692		eax.split.version_id = min(cap.version, 2);
693		eax.split.num_counters = cap.num_counters_gp;
694		eax.split.bit_width = cap.bit_width_gp;
695		eax.split.mask_length = cap.events_mask_len;
696
697		edx.split.num_counters_fixed = min(cap.num_counters_fixed, MAX_FIXED_COUNTERS);
698		edx.split.bit_width_fixed = cap.bit_width_fixed;
699		if (cap.version)
700			edx.split.anythread_deprecated = 1;
701		edx.split.reserved1 = 0;
702		edx.split.reserved2 = 0;
703
704		entry->eax = eax.full;
705		entry->ebx = cap.events_mask;
706		entry->ecx = 0;
707		entry->edx = edx.full;
708		break;
709	}
710	case 0x1f:
711	case 0xb:
712		/*
713		 * No topology; a valid topology is indicated by the presence
714		 * of subleaf 1.
715		 */
716		entry->eax = entry->ebx = entry->ecx = 0;
717		break;
718	case 0xd:
719		entry->eax &= supported_xcr0;
720		entry->ebx = xstate_required_size(supported_xcr0, false);
721		entry->ecx = entry->ebx;
722		entry->edx &= supported_xcr0 >> 32;
723		if (!supported_xcr0)
724			break;
725
726		entry = do_host_cpuid(array, function, 1);
727		if (!entry)
728			goto out;
729
730		cpuid_entry_override(entry, CPUID_D_1_EAX);
731		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
732			entry->ebx = xstate_required_size(supported_xcr0 | supported_xss,
733							  true);
734		else {
735			WARN_ON_ONCE(supported_xss != 0);
736			entry->ebx = 0;
737		}
738		entry->ecx &= supported_xss;
739		entry->edx &= supported_xss >> 32;
740
741		for (i = 2; i < 64; ++i) {
742			bool s_state;
743			if (supported_xcr0 & BIT_ULL(i))
744				s_state = false;
745			else if (supported_xss & BIT_ULL(i))
746				s_state = true;
747			else
748				continue;
749
750			entry = do_host_cpuid(array, function, i);
751			if (!entry)
752				goto out;
753
754			/*
755			 * The supported check above should have filtered out
756			 * invalid sub-leafs.  Only valid sub-leafs should
757			 * reach this point, and they should have a non-zero
758			 * save state size.  Furthermore, check whether the
759			 * processor agrees with supported_xcr0/supported_xss
760			 * on whether this is an XCR0- or IA32_XSS-managed area.
761			 */
762			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
763				--array->nent;
764				continue;
765			}
766			entry->edx = 0;
767		}
768		break;
769	/* Intel PT */
770	case 0x14:
771		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
772			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
773			break;
774		}
775
776		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
777			if (!do_host_cpuid(array, function, i))
778				goto out;
779		}
780		break;
781	case KVM_CPUID_SIGNATURE: {
782		static const char signature[12] = "KVMKVMKVM\0\0";
783		const u32 *sigptr = (const u32 *)signature;
784		entry->eax = KVM_CPUID_FEATURES;
785		entry->ebx = sigptr[0];
786		entry->ecx = sigptr[1];
787		entry->edx = sigptr[2];
788		break;
789	}
790	case KVM_CPUID_FEATURES:
791		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
792			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
793			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
794			     (1 << KVM_FEATURE_ASYNC_PF) |
795			     (1 << KVM_FEATURE_PV_EOI) |
796			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
797			     (1 << KVM_FEATURE_PV_UNHALT) |
798			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
799			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
800			     (1 << KVM_FEATURE_PV_SEND_IPI) |
801			     (1 << KVM_FEATURE_POLL_CONTROL) |
802			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
803			     (1 << KVM_FEATURE_ASYNC_PF_INT);
804
805		if (sched_info_on())
806			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
807
808		entry->ebx = 0;
809		entry->ecx = 0;
810		entry->edx = 0;
811		break;
812	case 0x80000000:
813		entry->eax = min(entry->eax, 0x8000001f);
814		break;
815	case 0x80000001:
816		entry->ebx &= ~GENMASK(27, 16);
817		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
818		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
819		break;
820	case 0x80000006:
821		/* Drop reserved bits, pass host L2 cache and TLB info. */
822		entry->edx &= ~GENMASK(17, 16);
823		break;
824	case 0x80000007: /* Advanced power management */
825		/* invariant TSC is CPUID.80000007H:EDX[8] */
826		entry->edx &= (1 << 8);
827		/* mask against host */
828		entry->edx &= boot_cpu_data.x86_power;
829		entry->eax = entry->ebx = entry->ecx = 0;
830		break;
831	case 0x80000008: {
832		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
833		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
834		unsigned phys_as = entry->eax & 0xff;
835
836		/*
837		 * Use bare metal's MAXPHADDR if the CPU doesn't report guest
838		 * MAXPHYADDR separately, or if TDP (NPT) is disabled, as the
839		 * guest version "applies only to guests using nested paging".
840		 */
841		if (!g_phys_as || !tdp_enabled)
842			g_phys_as = phys_as;
843
844		entry->eax = g_phys_as | (virt_as << 8);
845		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
846		entry->edx = 0;
847		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
848		break;
849	}
850	case 0x8000000A:
851		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
852			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
853			break;
854		}
855		entry->eax = 1; /* SVM revision 1 */
856		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
857				   ASID emulation to nested SVM */
858		entry->ecx = 0; /* Reserved */
859		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
860		break;
861	case 0x80000019:
862		entry->ecx = entry->edx = 0;
863		break;
864	case 0x8000001a:
865		entry->eax &= GENMASK(2, 0);
866		entry->ebx = entry->ecx = entry->edx = 0;
867		break;
868	case 0x8000001e:
869		/* Do not return host topology information.  */
870		entry->eax = entry->ebx = entry->ecx = 0;
871		entry->edx = 0; /* reserved */
872		break;
873	/* Support memory encryption cpuid if host supports it */
874	case 0x8000001F:
875		if (!boot_cpu_has(X86_FEATURE_SEV))
876			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
877		break;
878	/*Add support for Centaur's CPUID instruction*/
879	case 0xC0000000:
880		/*Just support up to 0xC0000004 now*/
881		entry->eax = min(entry->eax, 0xC0000004);
882		break;
883	case 0xC0000001:
884		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
885		break;
886	case 3: /* Processor serial number */
887	case 5: /* MONITOR/MWAIT */
888	case 0xC0000002:
889	case 0xC0000003:
890	case 0xC0000004:
891	default:
892		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
893		break;
894	}
895
896	r = 0;
897
898out:
899	put_cpu();
900
901	return r;
902}
903
904static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
905			 unsigned int type)
906{
907	if (type == KVM_GET_EMULATED_CPUID)
908		return __do_cpuid_func_emulated(array, func);
909
910	return __do_cpuid_func(array, func);
911}
912
913#define CENTAUR_CPUID_SIGNATURE 0xC0000000
914
915static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
916			  unsigned int type)
917{
918	u32 limit;
919	int r;
920
921	if (func == CENTAUR_CPUID_SIGNATURE &&
922	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
923		return 0;
924
925	r = do_cpuid_func(array, func, type);
926	if (r)
927		return r;
928
929	limit = array->entries[array->nent - 1].eax;
930	for (func = func + 1; func <= limit; ++func) {
931		r = do_cpuid_func(array, func, type);
932		if (r)
933			break;
934	}
935
936	return r;
937}
938
939static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
940				 __u32 num_entries, unsigned int ioctl_type)
941{
942	int i;
943	__u32 pad[3];
944
945	if (ioctl_type != KVM_GET_EMULATED_CPUID)
946		return false;
947
948	/*
949	 * We want to make sure that ->padding is being passed clean from
950	 * userspace in case we want to use it for something in the future.
951	 *
952	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
953	 * have to give ourselves satisfied only with the emulated side. /me
954	 * sheds a tear.
955	 */
956	for (i = 0; i < num_entries; i++) {
957		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
958			return true;
959
960		if (pad[0] || pad[1] || pad[2])
961			return true;
962	}
963	return false;
964}
965
966int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
967			    struct kvm_cpuid_entry2 __user *entries,
968			    unsigned int type)
969{
970	static const u32 funcs[] = {
971		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
972	};
973
974	struct kvm_cpuid_array array = {
975		.nent = 0,
976	};
977	int r, i;
978
979	if (cpuid->nent < 1)
980		return -E2BIG;
981	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
982		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
983
984	if (sanity_check_entries(entries, cpuid->nent, type))
985		return -EINVAL;
986
987	array.entries = vzalloc(array_size(sizeof(struct kvm_cpuid_entry2),
988					   cpuid->nent));
989	if (!array.entries)
990		return -ENOMEM;
991
992	array.maxnent = cpuid->nent;
993
994	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
995		r = get_cpuid_func(&array, funcs[i], type);
996		if (r)
997			goto out_free;
998	}
999	cpuid->nent = array.nent;
1000
1001	if (copy_to_user(entries, array.entries,
1002			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1003		r = -EFAULT;
1004
1005out_free:
1006	vfree(array.entries);
1007	return r;
1008}
1009
1010struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1011					      u32 function, u32 index)
1012{
1013	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1014				 function, index);
1015}
1016EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1017
1018/*
1019 * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1020 * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1021 * returns all zeroes for any undefined leaf, whether or not the leaf is in
1022 * range.  Centaur/VIA follows Intel semantics.
1023 *
1024 * A leaf is considered out-of-range if its function is higher than the maximum
1025 * supported leaf of its associated class or if its associated class does not
1026 * exist.
1027 *
1028 * There are three primary classes to be considered, with their respective
1029 * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1030 * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1031 * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1032 *
1033 *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1034 *  - Hypervisor: 0x40000000 - 0x4fffffff
1035 *  - Extended:   0x80000000 - 0xbfffffff
1036 *  - Centaur:    0xc0000000 - 0xcfffffff
1037 *
1038 * The Hypervisor class is further subdivided into sub-classes that each act as
1039 * their own indepdent class associated with a 0x100 byte range.  E.g. if Qemu
1040 * is advertising support for both HyperV and KVM, the resulting Hypervisor
1041 * CPUID sub-classes are:
1042 *
1043 *  - HyperV:     0x40000000 - 0x400000ff
1044 *  - KVM:        0x40000100 - 0x400001ff
1045 */
1046static struct kvm_cpuid_entry2 *
1047get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1048{
1049	struct kvm_cpuid_entry2 *basic, *class;
1050	u32 function = *fn_ptr;
1051
1052	basic = kvm_find_cpuid_entry(vcpu, 0, 0);
1053	if (!basic)
1054		return NULL;
1055
1056	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1057	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1058		return NULL;
1059
1060	if (function >= 0x40000000 && function <= 0x4fffffff)
1061		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00, 0);
1062	else if (function >= 0xc0000000)
1063		class = kvm_find_cpuid_entry(vcpu, 0xc0000000, 0);
1064	else
1065		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
1066
1067	if (class && function <= class->eax)
1068		return NULL;
1069
1070	/*
1071	 * Leaf specific adjustments are also applied when redirecting to the
1072	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1073	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1074	 * needs to be pulled from CPUID.0xb.1.
1075	 */
1076	*fn_ptr = basic->eax;
1077
1078	/*
1079	 * The class does not exist or the requested function is out of range;
1080	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1081	 * the original requested leaf is observed!
1082	 */
1083	return kvm_find_cpuid_entry(vcpu, basic->eax, index);
1084}
1085
1086bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1087	       u32 *ecx, u32 *edx, bool exact_only)
1088{
1089	u32 orig_function = *eax, function = *eax, index = *ecx;
1090	struct kvm_cpuid_entry2 *entry;
1091	bool exact, used_max_basic = false;
1092
1093	entry = kvm_find_cpuid_entry(vcpu, function, index);
1094	exact = !!entry;
1095
1096	if (!entry && !exact_only) {
1097		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1098		used_max_basic = !!entry;
1099	}
1100
1101	if (entry) {
1102		*eax = entry->eax;
1103		*ebx = entry->ebx;
1104		*ecx = entry->ecx;
1105		*edx = entry->edx;
1106		if (function == 7 && index == 0) {
1107			u64 data;
1108		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1109			    (data & TSX_CTRL_CPUID_CLEAR))
1110				*ebx &= ~(F(RTM) | F(HLE));
1111		}
1112	} else {
1113		*eax = *ebx = *ecx = *edx = 0;
1114		/*
1115		 * When leaf 0BH or 1FH is defined, CL is pass-through
1116		 * and EDX is always the x2APIC ID, even for undefined
1117		 * subleaves. Index 1 will exist iff the leaf is
1118		 * implemented, so we pass through CL iff leaf 1
1119		 * exists. EDX can be copied from any existing index.
1120		 */
1121		if (function == 0xb || function == 0x1f) {
1122			entry = kvm_find_cpuid_entry(vcpu, function, 1);
1123			if (entry) {
1124				*ecx = index & 0xff;
1125				*edx = entry->edx;
1126			}
1127		}
1128	}
1129	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1130			used_max_basic);
1131	return exact;
1132}
1133EXPORT_SYMBOL_GPL(kvm_cpuid);
1134
1135int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1136{
1137	u32 eax, ebx, ecx, edx;
1138
1139	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1140		return 1;
1141
1142	eax = kvm_rax_read(vcpu);
1143	ecx = kvm_rcx_read(vcpu);
1144	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1145	kvm_rax_write(vcpu, eax);
1146	kvm_rbx_write(vcpu, ebx);
1147	kvm_rcx_write(vcpu, ecx);
1148	kvm_rdx_write(vcpu, edx);
1149	return kvm_skip_emulated_instruction(vcpu);
1150}
1151EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1152