xref: /openbmc/linux/arch/x86/kvm/cpuid.c (revision 9ac17575)
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  */
31 u32 kvm_cpu_caps[NCAPINTS] __read_mostly;
32 EXPORT_SYMBOL_GPL(kvm_cpu_caps);
33 
34 static 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 
57 int kvm_update_cpuid(struct kvm_vcpu *vcpu)
58 {
59 	struct kvm_cpuid_entry2 *best;
60 	struct kvm_lapic *apic = vcpu->arch.apic;
61 
62 	best = kvm_find_cpuid_entry(vcpu, 1, 0);
63 	if (!best)
64 		return 0;
65 
66 	/* Update OSXSAVE bit */
67 	if (boot_cpu_has(X86_FEATURE_XSAVE) && best->function == 0x1)
68 		cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
69 				   kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE));
70 
71 	cpuid_entry_change(best, X86_FEATURE_APIC,
72 			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
73 
74 	if (apic) {
75 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
76 			apic->lapic_timer.timer_mode_mask = 3 << 17;
77 		else
78 			apic->lapic_timer.timer_mode_mask = 1 << 17;
79 	}
80 
81 	best = kvm_find_cpuid_entry(vcpu, 7, 0);
82 	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
83 		cpuid_entry_change(best, X86_FEATURE_OSPKE,
84 				   kvm_read_cr4_bits(vcpu, X86_CR4_PKE));
85 
86 	best = kvm_find_cpuid_entry(vcpu, 0xD, 0);
87 	if (!best) {
88 		vcpu->arch.guest_supported_xcr0 = 0;
89 	} else {
90 		vcpu->arch.guest_supported_xcr0 =
91 			(best->eax | ((u64)best->edx << 32)) & supported_xcr0;
92 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
93 	}
94 
95 	best = kvm_find_cpuid_entry(vcpu, 0xD, 1);
96 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
97 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
98 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
99 
100 	/*
101 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
102 	 * canonical address checks; exit if it is ever changed.
103 	 */
104 	best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
105 	if (best) {
106 		int vaddr_bits = (best->eax & 0xff00) >> 8;
107 
108 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
109 			return -EINVAL;
110 	}
111 
112 	best = kvm_find_cpuid_entry(vcpu, KVM_CPUID_FEATURES, 0);
113 	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
114 		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
115 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
116 
117 	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
118 		best = kvm_find_cpuid_entry(vcpu, 0x1, 0);
119 		if (best)
120 			cpuid_entry_change(best, X86_FEATURE_MWAIT,
121 					   vcpu->arch.ia32_misc_enable_msr &
122 					   MSR_IA32_MISC_ENABLE_MWAIT);
123 	}
124 
125 	/* Note, maxphyaddr must be updated before tdp_level. */
126 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
127 	vcpu->arch.tdp_level = kvm_x86_ops.get_tdp_level(vcpu);
128 	kvm_mmu_reset_context(vcpu);
129 
130 	kvm_pmu_refresh(vcpu);
131 	return 0;
132 }
133 
134 static int is_efer_nx(void)
135 {
136 	return host_efer & EFER_NX;
137 }
138 
139 static void cpuid_fix_nx_cap(struct kvm_vcpu *vcpu)
140 {
141 	int i;
142 	struct kvm_cpuid_entry2 *e, *entry;
143 
144 	entry = NULL;
145 	for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
146 		e = &vcpu->arch.cpuid_entries[i];
147 		if (e->function == 0x80000001) {
148 			entry = e;
149 			break;
150 		}
151 	}
152 	if (entry && cpuid_entry_has(entry, X86_FEATURE_NX) && !is_efer_nx()) {
153 		cpuid_entry_clear(entry, X86_FEATURE_NX);
154 		printk(KERN_INFO "kvm: guest NX capability removed\n");
155 	}
156 }
157 
158 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
159 {
160 	struct kvm_cpuid_entry2 *best;
161 
162 	best = kvm_find_cpuid_entry(vcpu, 0x80000000, 0);
163 	if (!best || best->eax < 0x80000008)
164 		goto not_found;
165 	best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
166 	if (best)
167 		return best->eax & 0xff;
168 not_found:
169 	return 36;
170 }
171 EXPORT_SYMBOL_GPL(cpuid_query_maxphyaddr);
172 
173 /* when an old userspace process fills a new kernel module */
174 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
175 			     struct kvm_cpuid *cpuid,
176 			     struct kvm_cpuid_entry __user *entries)
177 {
178 	int r, i;
179 	struct kvm_cpuid_entry *cpuid_entries = NULL;
180 
181 	r = -E2BIG;
182 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
183 		goto out;
184 	r = -ENOMEM;
185 	if (cpuid->nent) {
186 		cpuid_entries =
187 			vmalloc(array_size(sizeof(struct kvm_cpuid_entry),
188 					   cpuid->nent));
189 		if (!cpuid_entries)
190 			goto out;
191 		r = -EFAULT;
192 		if (copy_from_user(cpuid_entries, entries,
193 				   cpuid->nent * sizeof(struct kvm_cpuid_entry)))
194 			goto out;
195 	}
196 	for (i = 0; i < cpuid->nent; i++) {
197 		vcpu->arch.cpuid_entries[i].function = cpuid_entries[i].function;
198 		vcpu->arch.cpuid_entries[i].eax = cpuid_entries[i].eax;
199 		vcpu->arch.cpuid_entries[i].ebx = cpuid_entries[i].ebx;
200 		vcpu->arch.cpuid_entries[i].ecx = cpuid_entries[i].ecx;
201 		vcpu->arch.cpuid_entries[i].edx = cpuid_entries[i].edx;
202 		vcpu->arch.cpuid_entries[i].index = 0;
203 		vcpu->arch.cpuid_entries[i].flags = 0;
204 		vcpu->arch.cpuid_entries[i].padding[0] = 0;
205 		vcpu->arch.cpuid_entries[i].padding[1] = 0;
206 		vcpu->arch.cpuid_entries[i].padding[2] = 0;
207 	}
208 	vcpu->arch.cpuid_nent = cpuid->nent;
209 	cpuid_fix_nx_cap(vcpu);
210 	kvm_apic_set_version(vcpu);
211 	kvm_x86_ops.cpuid_update(vcpu);
212 	r = kvm_update_cpuid(vcpu);
213 
214 out:
215 	vfree(cpuid_entries);
216 	return r;
217 }
218 
219 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
220 			      struct kvm_cpuid2 *cpuid,
221 			      struct kvm_cpuid_entry2 __user *entries)
222 {
223 	int r;
224 
225 	r = -E2BIG;
226 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
227 		goto out;
228 	r = -EFAULT;
229 	if (copy_from_user(&vcpu->arch.cpuid_entries, entries,
230 			   cpuid->nent * sizeof(struct kvm_cpuid_entry2)))
231 		goto out;
232 	vcpu->arch.cpuid_nent = cpuid->nent;
233 	kvm_apic_set_version(vcpu);
234 	kvm_x86_ops.cpuid_update(vcpu);
235 	r = kvm_update_cpuid(vcpu);
236 out:
237 	return r;
238 }
239 
240 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
241 			      struct kvm_cpuid2 *cpuid,
242 			      struct kvm_cpuid_entry2 __user *entries)
243 {
244 	int r;
245 
246 	r = -E2BIG;
247 	if (cpuid->nent < vcpu->arch.cpuid_nent)
248 		goto out;
249 	r = -EFAULT;
250 	if (copy_to_user(entries, &vcpu->arch.cpuid_entries,
251 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
252 		goto out;
253 	return 0;
254 
255 out:
256 	cpuid->nent = vcpu->arch.cpuid_nent;
257 	return r;
258 }
259 
260 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
261 {
262 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
263 	struct kvm_cpuid_entry2 entry;
264 
265 	reverse_cpuid_check(leaf);
266 	kvm_cpu_caps[leaf] &= mask;
267 
268 	cpuid_count(cpuid.function, cpuid.index,
269 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
270 
271 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
272 }
273 
274 void kvm_set_cpu_caps(void)
275 {
276 	unsigned int f_nx = is_efer_nx() ? F(NX) : 0;
277 #ifdef CONFIG_X86_64
278 	unsigned int f_gbpages = F(GBPAGES);
279 	unsigned int f_lm = F(LM);
280 #else
281 	unsigned int f_gbpages = 0;
282 	unsigned int f_lm = 0;
283 #endif
284 
285 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) >
286 		     sizeof(boot_cpu_data.x86_capability));
287 
288 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
289 	       sizeof(kvm_cpu_caps));
290 
291 	kvm_cpu_cap_mask(CPUID_1_ECX,
292 		/*
293 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
294 		 * advertised to guests via CPUID!
295 		 */
296 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
297 		0 /* DS-CPL, VMX, SMX, EST */ |
298 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
299 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
300 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
301 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
302 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
303 		F(F16C) | F(RDRAND)
304 	);
305 	/* KVM emulates x2apic in software irrespective of host support. */
306 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
307 
308 	kvm_cpu_cap_mask(CPUID_1_EDX,
309 		F(FPU) | F(VME) | F(DE) | F(PSE) |
310 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
311 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
312 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
313 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
314 		0 /* Reserved, DS, ACPI */ | F(MMX) |
315 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
316 		0 /* HTT, TM, Reserved, PBE */
317 	);
318 
319 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
320 		F(FSGSBASE) | F(BMI1) | F(HLE) | F(AVX2) | F(SMEP) |
321 		F(BMI2) | F(ERMS) | 0 /*INVPCID*/ | F(RTM) | 0 /*MPX*/ | F(RDSEED) |
322 		F(ADX) | F(SMAP) | F(AVX512IFMA) | F(AVX512F) | F(AVX512PF) |
323 		F(AVX512ER) | F(AVX512CD) | F(CLFLUSHOPT) | F(CLWB) | F(AVX512DQ) |
324 		F(SHA_NI) | F(AVX512BW) | F(AVX512VL) | 0 /*INTEL_PT*/
325 	);
326 
327 	kvm_cpu_cap_mask(CPUID_7_ECX,
328 		F(AVX512VBMI) | F(LA57) | 0 /*PKU*/ | 0 /*OSPKE*/ | F(RDPID) |
329 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
330 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
331 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/
332 	);
333 	/* Set LA57 based on hardware capability. */
334 	if (cpuid_ecx(7) & F(LA57))
335 		kvm_cpu_cap_set(X86_FEATURE_LA57);
336 
337 	kvm_cpu_cap_mask(CPUID_7_EDX,
338 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
339 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
340 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM)
341 	);
342 
343 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
344 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
345 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
346 
347 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
348 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
349 	if (boot_cpu_has(X86_FEATURE_STIBP))
350 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
351 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
352 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
353 
354 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
355 		F(AVX512_BF16)
356 	);
357 
358 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
359 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES)
360 	);
361 
362 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
363 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
364 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
365 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
366 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
367 		F(TOPOEXT) | F(PERFCTR_CORE)
368 	);
369 
370 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
371 		F(FPU) | F(VME) | F(DE) | F(PSE) |
372 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
373 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
374 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
375 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
376 		f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
377 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
378 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
379 	);
380 
381 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
382 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
383 
384 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
385 		F(CLZERO) | F(XSAVEERPTR) |
386 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
387 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON)
388 	);
389 
390 	/*
391 	 * AMD has separate bits for each SPEC_CTRL bit.
392 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
393 	 * record that in cpufeatures so use them.
394 	 */
395 	if (boot_cpu_has(X86_FEATURE_IBPB))
396 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
397 	if (boot_cpu_has(X86_FEATURE_IBRS))
398 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
399 	if (boot_cpu_has(X86_FEATURE_STIBP))
400 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
401 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
402 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
403 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
404 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
405 	/*
406 	 * The preference is to use SPEC CTRL MSR instead of the
407 	 * VIRT_SPEC MSR.
408 	 */
409 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
410 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
411 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
412 
413 	/*
414 	 * Hide all SVM features by default, SVM will set the cap bits for
415 	 * features it emulates and/or exposes for L1.
416 	 */
417 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
418 
419 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
420 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
421 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
422 		F(PMM) | F(PMM_EN)
423 	);
424 }
425 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
426 
427 struct kvm_cpuid_array {
428 	struct kvm_cpuid_entry2 *entries;
429 	const int maxnent;
430 	int nent;
431 };
432 
433 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
434 					      u32 function, u32 index)
435 {
436 	struct kvm_cpuid_entry2 *entry;
437 
438 	if (array->nent >= array->maxnent)
439 		return NULL;
440 
441 	entry = &array->entries[array->nent++];
442 
443 	entry->function = function;
444 	entry->index = index;
445 	entry->flags = 0;
446 
447 	cpuid_count(entry->function, entry->index,
448 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
449 
450 	switch (function) {
451 	case 4:
452 	case 7:
453 	case 0xb:
454 	case 0xd:
455 	case 0xf:
456 	case 0x10:
457 	case 0x12:
458 	case 0x14:
459 	case 0x17:
460 	case 0x18:
461 	case 0x1f:
462 	case 0x8000001d:
463 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
464 		break;
465 	}
466 
467 	return entry;
468 }
469 
470 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
471 {
472 	struct kvm_cpuid_entry2 *entry;
473 
474 	if (array->nent >= array->maxnent)
475 		return -E2BIG;
476 
477 	entry = &array->entries[array->nent];
478 	entry->function = func;
479 	entry->index = 0;
480 	entry->flags = 0;
481 
482 	switch (func) {
483 	case 0:
484 		entry->eax = 7;
485 		++array->nent;
486 		break;
487 	case 1:
488 		entry->ecx = F(MOVBE);
489 		++array->nent;
490 		break;
491 	case 7:
492 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
493 		entry->eax = 0;
494 		entry->ecx = F(RDPID);
495 		++array->nent;
496 	default:
497 		break;
498 	}
499 
500 	return 0;
501 }
502 
503 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
504 {
505 	struct kvm_cpuid_entry2 *entry;
506 	int r, i, max_idx;
507 
508 	/* all calls to cpuid_count() should be made on the same cpu */
509 	get_cpu();
510 
511 	r = -E2BIG;
512 
513 	entry = do_host_cpuid(array, function, 0);
514 	if (!entry)
515 		goto out;
516 
517 	switch (function) {
518 	case 0:
519 		/* Limited to the highest leaf implemented in KVM. */
520 		entry->eax = min(entry->eax, 0x1fU);
521 		break;
522 	case 1:
523 		cpuid_entry_override(entry, CPUID_1_EDX);
524 		cpuid_entry_override(entry, CPUID_1_ECX);
525 		break;
526 	case 2:
527 		/*
528 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
529 		 * CPUID(function=2, index=0) may return different results each
530 		 * time, with the least-significant byte in EAX enumerating the
531 		 * number of times software should do CPUID(2, 0).
532 		 *
533 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
534 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
535 		 * return 01H. Software should ignore this value and not
536 		 * interpret it as an informational descriptor", while AMD's
537 		 * APM states that CPUID(2) is reserved.
538 		 *
539 		 * WARN if a frankenstein CPU that supports virtualization and
540 		 * a stateful CPUID.0x2 is encountered.
541 		 */
542 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
543 		break;
544 	/* functions 4 and 0x8000001d have additional index. */
545 	case 4:
546 	case 0x8000001d:
547 		/*
548 		 * Read entries until the cache type in the previous entry is
549 		 * zero, i.e. indicates an invalid entry.
550 		 */
551 		for (i = 1; entry->eax & 0x1f; ++i) {
552 			entry = do_host_cpuid(array, function, i);
553 			if (!entry)
554 				goto out;
555 		}
556 		break;
557 	case 6: /* Thermal management */
558 		entry->eax = 0x4; /* allow ARAT */
559 		entry->ebx = 0;
560 		entry->ecx = 0;
561 		entry->edx = 0;
562 		break;
563 	/* function 7 has additional index. */
564 	case 7:
565 		entry->eax = min(entry->eax, 1u);
566 		cpuid_entry_override(entry, CPUID_7_0_EBX);
567 		cpuid_entry_override(entry, CPUID_7_ECX);
568 		cpuid_entry_override(entry, CPUID_7_EDX);
569 
570 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
571 		if (entry->eax == 1) {
572 			entry = do_host_cpuid(array, function, 1);
573 			if (!entry)
574 				goto out;
575 
576 			cpuid_entry_override(entry, CPUID_7_1_EAX);
577 			entry->ebx = 0;
578 			entry->ecx = 0;
579 			entry->edx = 0;
580 		}
581 		break;
582 	case 9:
583 		break;
584 	case 0xa: { /* Architectural Performance Monitoring */
585 		struct x86_pmu_capability cap;
586 		union cpuid10_eax eax;
587 		union cpuid10_edx edx;
588 
589 		perf_get_x86_pmu_capability(&cap);
590 
591 		/*
592 		 * Only support guest architectural pmu on a host
593 		 * with architectural pmu.
594 		 */
595 		if (!cap.version)
596 			memset(&cap, 0, sizeof(cap));
597 
598 		eax.split.version_id = min(cap.version, 2);
599 		eax.split.num_counters = cap.num_counters_gp;
600 		eax.split.bit_width = cap.bit_width_gp;
601 		eax.split.mask_length = cap.events_mask_len;
602 
603 		edx.split.num_counters_fixed = cap.num_counters_fixed;
604 		edx.split.bit_width_fixed = cap.bit_width_fixed;
605 		edx.split.reserved = 0;
606 
607 		entry->eax = eax.full;
608 		entry->ebx = cap.events_mask;
609 		entry->ecx = 0;
610 		entry->edx = edx.full;
611 		break;
612 	}
613 	/*
614 	 * Per Intel's SDM, the 0x1f is a superset of 0xb,
615 	 * thus they can be handled by common code.
616 	 */
617 	case 0x1f:
618 	case 0xb:
619 		/*
620 		 * Populate entries until the level type (ECX[15:8]) of the
621 		 * previous entry is zero.  Note, CPUID EAX.{0x1f,0xb}.0 is
622 		 * the starting entry, filled by the primary do_host_cpuid().
623 		 */
624 		for (i = 1; entry->ecx & 0xff00; ++i) {
625 			entry = do_host_cpuid(array, function, i);
626 			if (!entry)
627 				goto out;
628 		}
629 		break;
630 	case 0xd:
631 		entry->eax &= supported_xcr0;
632 		entry->ebx = xstate_required_size(supported_xcr0, false);
633 		entry->ecx = entry->ebx;
634 		entry->edx &= supported_xcr0 >> 32;
635 		if (!supported_xcr0)
636 			break;
637 
638 		entry = do_host_cpuid(array, function, 1);
639 		if (!entry)
640 			goto out;
641 
642 		cpuid_entry_override(entry, CPUID_D_1_EAX);
643 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
644 			entry->ebx = xstate_required_size(supported_xcr0 | supported_xss,
645 							  true);
646 		else {
647 			WARN_ON_ONCE(supported_xss != 0);
648 			entry->ebx = 0;
649 		}
650 		entry->ecx &= supported_xss;
651 		entry->edx &= supported_xss >> 32;
652 
653 		for (i = 2; i < 64; ++i) {
654 			bool s_state;
655 			if (supported_xcr0 & BIT_ULL(i))
656 				s_state = false;
657 			else if (supported_xss & BIT_ULL(i))
658 				s_state = true;
659 			else
660 				continue;
661 
662 			entry = do_host_cpuid(array, function, i);
663 			if (!entry)
664 				goto out;
665 
666 			/*
667 			 * The supported check above should have filtered out
668 			 * invalid sub-leafs.  Only valid sub-leafs should
669 			 * reach this point, and they should have a non-zero
670 			 * save state size.  Furthermore, check whether the
671 			 * processor agrees with supported_xcr0/supported_xss
672 			 * on whether this is an XCR0- or IA32_XSS-managed area.
673 			 */
674 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
675 				--array->nent;
676 				continue;
677 			}
678 			entry->edx = 0;
679 		}
680 		break;
681 	/* Intel PT */
682 	case 0x14:
683 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
684 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
685 			break;
686 		}
687 
688 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
689 			if (!do_host_cpuid(array, function, i))
690 				goto out;
691 		}
692 		break;
693 	case KVM_CPUID_SIGNATURE: {
694 		static const char signature[12] = "KVMKVMKVM\0\0";
695 		const u32 *sigptr = (const u32 *)signature;
696 		entry->eax = KVM_CPUID_FEATURES;
697 		entry->ebx = sigptr[0];
698 		entry->ecx = sigptr[1];
699 		entry->edx = sigptr[2];
700 		break;
701 	}
702 	case KVM_CPUID_FEATURES:
703 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
704 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
705 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
706 			     (1 << KVM_FEATURE_ASYNC_PF) |
707 			     (1 << KVM_FEATURE_PV_EOI) |
708 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
709 			     (1 << KVM_FEATURE_PV_UNHALT) |
710 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
711 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
712 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
713 			     (1 << KVM_FEATURE_POLL_CONTROL) |
714 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
715 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
716 
717 		if (sched_info_on())
718 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
719 
720 		entry->ebx = 0;
721 		entry->ecx = 0;
722 		entry->edx = 0;
723 		break;
724 	case 0x80000000:
725 		entry->eax = min(entry->eax, 0x8000001f);
726 		break;
727 	case 0x80000001:
728 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
729 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
730 		break;
731 	case 0x80000006:
732 		/* L2 cache and TLB: pass through host info. */
733 		break;
734 	case 0x80000007: /* Advanced power management */
735 		/* invariant TSC is CPUID.80000007H:EDX[8] */
736 		entry->edx &= (1 << 8);
737 		/* mask against host */
738 		entry->edx &= boot_cpu_data.x86_power;
739 		entry->eax = entry->ebx = entry->ecx = 0;
740 		break;
741 	case 0x80000008: {
742 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
743 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
744 		unsigned phys_as = entry->eax & 0xff;
745 
746 		if (!g_phys_as)
747 			g_phys_as = phys_as;
748 		entry->eax = g_phys_as | (virt_as << 8);
749 		entry->edx = 0;
750 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
751 		break;
752 	}
753 	case 0x8000000A:
754 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
755 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
756 			break;
757 		}
758 		entry->eax = 1; /* SVM revision 1 */
759 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
760 				   ASID emulation to nested SVM */
761 		entry->ecx = 0; /* Reserved */
762 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
763 		break;
764 	case 0x80000019:
765 		entry->ecx = entry->edx = 0;
766 		break;
767 	case 0x8000001a:
768 	case 0x8000001e:
769 		break;
770 	/* Support memory encryption cpuid if host supports it */
771 	case 0x8000001F:
772 		if (!boot_cpu_has(X86_FEATURE_SEV))
773 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
774 		break;
775 	/*Add support for Centaur's CPUID instruction*/
776 	case 0xC0000000:
777 		/*Just support up to 0xC0000004 now*/
778 		entry->eax = min(entry->eax, 0xC0000004);
779 		break;
780 	case 0xC0000001:
781 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
782 		break;
783 	case 3: /* Processor serial number */
784 	case 5: /* MONITOR/MWAIT */
785 	case 0xC0000002:
786 	case 0xC0000003:
787 	case 0xC0000004:
788 	default:
789 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
790 		break;
791 	}
792 
793 	r = 0;
794 
795 out:
796 	put_cpu();
797 
798 	return r;
799 }
800 
801 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
802 			 unsigned int type)
803 {
804 	if (type == KVM_GET_EMULATED_CPUID)
805 		return __do_cpuid_func_emulated(array, func);
806 
807 	return __do_cpuid_func(array, func);
808 }
809 
810 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
811 
812 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
813 			  unsigned int type)
814 {
815 	u32 limit;
816 	int r;
817 
818 	if (func == CENTAUR_CPUID_SIGNATURE &&
819 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
820 		return 0;
821 
822 	r = do_cpuid_func(array, func, type);
823 	if (r)
824 		return r;
825 
826 	limit = array->entries[array->nent - 1].eax;
827 	for (func = func + 1; func <= limit; ++func) {
828 		r = do_cpuid_func(array, func, type);
829 		if (r)
830 			break;
831 	}
832 
833 	return r;
834 }
835 
836 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
837 				 __u32 num_entries, unsigned int ioctl_type)
838 {
839 	int i;
840 	__u32 pad[3];
841 
842 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
843 		return false;
844 
845 	/*
846 	 * We want to make sure that ->padding is being passed clean from
847 	 * userspace in case we want to use it for something in the future.
848 	 *
849 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
850 	 * have to give ourselves satisfied only with the emulated side. /me
851 	 * sheds a tear.
852 	 */
853 	for (i = 0; i < num_entries; i++) {
854 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
855 			return true;
856 
857 		if (pad[0] || pad[1] || pad[2])
858 			return true;
859 	}
860 	return false;
861 }
862 
863 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
864 			    struct kvm_cpuid_entry2 __user *entries,
865 			    unsigned int type)
866 {
867 	static const u32 funcs[] = {
868 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
869 	};
870 
871 	struct kvm_cpuid_array array = {
872 		.nent = 0,
873 		.maxnent = cpuid->nent,
874 	};
875 	int r, i;
876 
877 	if (cpuid->nent < 1)
878 		return -E2BIG;
879 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
880 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
881 
882 	if (sanity_check_entries(entries, cpuid->nent, type))
883 		return -EINVAL;
884 
885 	array.entries = vzalloc(array_size(sizeof(struct kvm_cpuid_entry2),
886 					   cpuid->nent));
887 	if (!array.entries)
888 		return -ENOMEM;
889 
890 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
891 		r = get_cpuid_func(&array, funcs[i], type);
892 		if (r)
893 			goto out_free;
894 	}
895 	cpuid->nent = array.nent;
896 
897 	if (copy_to_user(entries, array.entries,
898 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
899 		r = -EFAULT;
900 
901 out_free:
902 	vfree(array.entries);
903 	return r;
904 }
905 
906 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
907 					      u32 function, u32 index)
908 {
909 	struct kvm_cpuid_entry2 *e;
910 	int i;
911 
912 	for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
913 		e = &vcpu->arch.cpuid_entries[i];
914 
915 		if (e->function == function && (e->index == index ||
916 		    !(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX)))
917 			return e;
918 	}
919 	return NULL;
920 }
921 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
922 
923 /*
924  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
925  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
926  * returns all zeroes for any undefined leaf, whether or not the leaf is in
927  * range.  Centaur/VIA follows Intel semantics.
928  *
929  * A leaf is considered out-of-range if its function is higher than the maximum
930  * supported leaf of its associated class or if its associated class does not
931  * exist.
932  *
933  * There are three primary classes to be considered, with their respective
934  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
935  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
936  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
937  *
938  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
939  *  - Hypervisor: 0x40000000 - 0x4fffffff
940  *  - Extended:   0x80000000 - 0xbfffffff
941  *  - Centaur:    0xc0000000 - 0xcfffffff
942  *
943  * The Hypervisor class is further subdivided into sub-classes that each act as
944  * their own indepdent class associated with a 0x100 byte range.  E.g. if Qemu
945  * is advertising support for both HyperV and KVM, the resulting Hypervisor
946  * CPUID sub-classes are:
947  *
948  *  - HyperV:     0x40000000 - 0x400000ff
949  *  - KVM:        0x40000100 - 0x400001ff
950  */
951 static struct kvm_cpuid_entry2 *
952 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
953 {
954 	struct kvm_cpuid_entry2 *basic, *class;
955 	u32 function = *fn_ptr;
956 
957 	basic = kvm_find_cpuid_entry(vcpu, 0, 0);
958 	if (!basic)
959 		return NULL;
960 
961 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
962 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
963 		return NULL;
964 
965 	if (function >= 0x40000000 && function <= 0x4fffffff)
966 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00, 0);
967 	else if (function >= 0xc0000000)
968 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000, 0);
969 	else
970 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
971 
972 	if (class && function <= class->eax)
973 		return NULL;
974 
975 	/*
976 	 * Leaf specific adjustments are also applied when redirecting to the
977 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
978 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
979 	 * needs to be pulled from CPUID.0xb.1.
980 	 */
981 	*fn_ptr = basic->eax;
982 
983 	/*
984 	 * The class does not exist or the requested function is out of range;
985 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
986 	 * the original requested leaf is observed!
987 	 */
988 	return kvm_find_cpuid_entry(vcpu, basic->eax, index);
989 }
990 
991 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
992 	       u32 *ecx, u32 *edx, bool exact_only)
993 {
994 	u32 orig_function = *eax, function = *eax, index = *ecx;
995 	struct kvm_cpuid_entry2 *entry;
996 	bool exact, used_max_basic = false;
997 
998 	entry = kvm_find_cpuid_entry(vcpu, function, index);
999 	exact = !!entry;
1000 
1001 	if (!entry && !exact_only) {
1002 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1003 		used_max_basic = !!entry;
1004 	}
1005 
1006 	if (entry) {
1007 		*eax = entry->eax;
1008 		*ebx = entry->ebx;
1009 		*ecx = entry->ecx;
1010 		*edx = entry->edx;
1011 		if (function == 7 && index == 0) {
1012 			u64 data;
1013 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1014 			    (data & TSX_CTRL_CPUID_CLEAR))
1015 				*ebx &= ~(F(RTM) | F(HLE));
1016 		}
1017 	} else {
1018 		*eax = *ebx = *ecx = *edx = 0;
1019 		/*
1020 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1021 		 * and EDX is always the x2APIC ID, even for undefined
1022 		 * subleaves. Index 1 will exist iff the leaf is
1023 		 * implemented, so we pass through CL iff leaf 1
1024 		 * exists. EDX can be copied from any existing index.
1025 		 */
1026 		if (function == 0xb || function == 0x1f) {
1027 			entry = kvm_find_cpuid_entry(vcpu, function, 1);
1028 			if (entry) {
1029 				*ecx = index & 0xff;
1030 				*edx = entry->edx;
1031 			}
1032 		}
1033 	}
1034 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1035 			used_max_basic);
1036 	return exact;
1037 }
1038 EXPORT_SYMBOL_GPL(kvm_cpuid);
1039 
1040 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1041 {
1042 	u32 eax, ebx, ecx, edx;
1043 
1044 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1045 		return 1;
1046 
1047 	eax = kvm_rax_read(vcpu);
1048 	ecx = kvm_rcx_read(vcpu);
1049 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1050 	kvm_rax_write(vcpu, eax);
1051 	kvm_rbx_write(vcpu, ebx);
1052 	kvm_rcx_write(vcpu, ecx);
1053 	kvm_rdx_write(vcpu, edx);
1054 	return kvm_skip_emulated_instruction(vcpu);
1055 }
1056 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1057