xref: /linux/arch/x86/kvm/cpuid.c (revision ed30aef3c864f99111e16d4ea5cf29488d99a278)
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 static 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 
74 static 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 
93 void 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 
107 void 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 
150 static 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 	/* Invoke the vendor callback only after the above state is updated. */
182 	kvm_x86_ops.vcpu_after_set_cpuid(vcpu);
183 }
184 
185 static int is_efer_nx(void)
186 {
187 	return host_efer & EFER_NX;
188 }
189 
190 static void cpuid_fix_nx_cap(struct kvm_vcpu *vcpu)
191 {
192 	int i;
193 	struct kvm_cpuid_entry2 *e, *entry;
194 
195 	entry = NULL;
196 	for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
197 		e = &vcpu->arch.cpuid_entries[i];
198 		if (e->function == 0x80000001) {
199 			entry = e;
200 			break;
201 		}
202 	}
203 	if (entry && cpuid_entry_has(entry, X86_FEATURE_NX) && !is_efer_nx()) {
204 		cpuid_entry_clear(entry, X86_FEATURE_NX);
205 		printk(KERN_INFO "kvm: guest NX capability removed\n");
206 	}
207 }
208 
209 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
210 {
211 	struct kvm_cpuid_entry2 *best;
212 
213 	best = kvm_find_cpuid_entry(vcpu, 0x80000000, 0);
214 	if (!best || best->eax < 0x80000008)
215 		goto not_found;
216 	best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
217 	if (best)
218 		return best->eax & 0xff;
219 not_found:
220 	return 36;
221 }
222 
223 /* when an old userspace process fills a new kernel module */
224 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
225 			     struct kvm_cpuid *cpuid,
226 			     struct kvm_cpuid_entry __user *entries)
227 {
228 	int r, i;
229 	struct kvm_cpuid_entry *e = NULL;
230 	struct kvm_cpuid_entry2 *e2 = NULL;
231 
232 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
233 		return -E2BIG;
234 
235 	if (cpuid->nent) {
236 		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
237 		if (IS_ERR(e))
238 			return PTR_ERR(e);
239 
240 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
241 		if (!e2) {
242 			r = -ENOMEM;
243 			goto out_free_cpuid;
244 		}
245 	}
246 	for (i = 0; i < cpuid->nent; i++) {
247 		e2[i].function = e[i].function;
248 		e2[i].eax = e[i].eax;
249 		e2[i].ebx = e[i].ebx;
250 		e2[i].ecx = e[i].ecx;
251 		e2[i].edx = e[i].edx;
252 		e2[i].index = 0;
253 		e2[i].flags = 0;
254 		e2[i].padding[0] = 0;
255 		e2[i].padding[1] = 0;
256 		e2[i].padding[2] = 0;
257 	}
258 
259 	r = kvm_check_cpuid(e2, cpuid->nent);
260 	if (r) {
261 		kvfree(e2);
262 		goto out_free_cpuid;
263 	}
264 
265 	kvfree(vcpu->arch.cpuid_entries);
266 	vcpu->arch.cpuid_entries = e2;
267 	vcpu->arch.cpuid_nent = cpuid->nent;
268 
269 	cpuid_fix_nx_cap(vcpu);
270 	kvm_update_cpuid_runtime(vcpu);
271 	kvm_vcpu_after_set_cpuid(vcpu);
272 
273 out_free_cpuid:
274 	kvfree(e);
275 
276 	return r;
277 }
278 
279 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
280 			      struct kvm_cpuid2 *cpuid,
281 			      struct kvm_cpuid_entry2 __user *entries)
282 {
283 	struct kvm_cpuid_entry2 *e2 = NULL;
284 	int r;
285 
286 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
287 		return -E2BIG;
288 
289 	if (cpuid->nent) {
290 		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
291 		if (IS_ERR(e2))
292 			return PTR_ERR(e2);
293 	}
294 
295 	r = kvm_check_cpuid(e2, cpuid->nent);
296 	if (r) {
297 		kvfree(e2);
298 		return r;
299 	}
300 
301 	kvfree(vcpu->arch.cpuid_entries);
302 	vcpu->arch.cpuid_entries = e2;
303 	vcpu->arch.cpuid_nent = cpuid->nent;
304 
305 	kvm_update_cpuid_runtime(vcpu);
306 	kvm_vcpu_after_set_cpuid(vcpu);
307 
308 	return 0;
309 }
310 
311 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
312 			      struct kvm_cpuid2 *cpuid,
313 			      struct kvm_cpuid_entry2 __user *entries)
314 {
315 	int r;
316 
317 	r = -E2BIG;
318 	if (cpuid->nent < vcpu->arch.cpuid_nent)
319 		goto out;
320 	r = -EFAULT;
321 	if (copy_to_user(entries, &vcpu->arch.cpuid_entries,
322 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
323 		goto out;
324 	return 0;
325 
326 out:
327 	cpuid->nent = vcpu->arch.cpuid_nent;
328 	return r;
329 }
330 
331 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
332 {
333 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
334 	struct kvm_cpuid_entry2 entry;
335 
336 	reverse_cpuid_check(leaf);
337 	kvm_cpu_caps[leaf] &= mask;
338 
339 	cpuid_count(cpuid.function, cpuid.index,
340 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
341 
342 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
343 }
344 
345 void kvm_set_cpu_caps(void)
346 {
347 	unsigned int f_nx = is_efer_nx() ? F(NX) : 0;
348 #ifdef CONFIG_X86_64
349 	unsigned int f_gbpages = F(GBPAGES);
350 	unsigned int f_lm = F(LM);
351 #else
352 	unsigned int f_gbpages = 0;
353 	unsigned int f_lm = 0;
354 #endif
355 
356 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) >
357 		     sizeof(boot_cpu_data.x86_capability));
358 
359 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
360 	       sizeof(kvm_cpu_caps));
361 
362 	kvm_cpu_cap_mask(CPUID_1_ECX,
363 		/*
364 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
365 		 * advertised to guests via CPUID!
366 		 */
367 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
368 		0 /* DS-CPL, VMX, SMX, EST */ |
369 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
370 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
371 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
372 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
373 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
374 		F(F16C) | F(RDRAND)
375 	);
376 	/* KVM emulates x2apic in software irrespective of host support. */
377 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
378 
379 	kvm_cpu_cap_mask(CPUID_1_EDX,
380 		F(FPU) | F(VME) | F(DE) | F(PSE) |
381 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
382 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
383 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
384 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
385 		0 /* Reserved, DS, ACPI */ | F(MMX) |
386 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
387 		0 /* HTT, TM, Reserved, PBE */
388 	);
389 
390 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
391 		F(FSGSBASE) | F(BMI1) | F(HLE) | F(AVX2) | F(SMEP) |
392 		F(BMI2) | F(ERMS) | 0 /*INVPCID*/ | F(RTM) | 0 /*MPX*/ | F(RDSEED) |
393 		F(ADX) | F(SMAP) | F(AVX512IFMA) | F(AVX512F) | F(AVX512PF) |
394 		F(AVX512ER) | F(AVX512CD) | F(CLFLUSHOPT) | F(CLWB) | F(AVX512DQ) |
395 		F(SHA_NI) | F(AVX512BW) | F(AVX512VL) | 0 /*INTEL_PT*/
396 	);
397 
398 	kvm_cpu_cap_mask(CPUID_7_ECX,
399 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
400 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
401 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
402 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/
403 	);
404 	/* Set LA57 based on hardware capability. */
405 	if (cpuid_ecx(7) & F(LA57))
406 		kvm_cpu_cap_set(X86_FEATURE_LA57);
407 
408 	/*
409 	 * PKU not yet implemented for shadow paging and requires OSPKE
410 	 * to be set on the host. Clear it if that is not the case
411 	 */
412 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
413 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
414 
415 	kvm_cpu_cap_mask(CPUID_7_EDX,
416 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
417 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
418 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
419 		F(SERIALIZE) | F(TSXLDTRK)
420 	);
421 
422 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
423 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
424 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
425 
426 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
427 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
428 	if (boot_cpu_has(X86_FEATURE_STIBP))
429 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
430 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
431 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
432 
433 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
434 		F(AVX512_BF16)
435 	);
436 
437 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
438 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES)
439 	);
440 
441 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
442 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
443 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
444 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
445 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
446 		F(TOPOEXT) | F(PERFCTR_CORE)
447 	);
448 
449 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
450 		F(FPU) | F(VME) | F(DE) | F(PSE) |
451 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
452 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
453 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
454 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
455 		f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
456 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
457 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
458 	);
459 
460 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
461 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
462 
463 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
464 		F(CLZERO) | F(XSAVEERPTR) |
465 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
466 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON)
467 	);
468 
469 	/*
470 	 * AMD has separate bits for each SPEC_CTRL bit.
471 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
472 	 * record that in cpufeatures so use them.
473 	 */
474 	if (boot_cpu_has(X86_FEATURE_IBPB))
475 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
476 	if (boot_cpu_has(X86_FEATURE_IBRS))
477 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
478 	if (boot_cpu_has(X86_FEATURE_STIBP))
479 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
480 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
481 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
482 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
483 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
484 	/*
485 	 * The preference is to use SPEC CTRL MSR instead of the
486 	 * VIRT_SPEC MSR.
487 	 */
488 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
489 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
490 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
491 
492 	/*
493 	 * Hide all SVM features by default, SVM will set the cap bits for
494 	 * features it emulates and/or exposes for L1.
495 	 */
496 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
497 
498 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
499 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
500 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
501 		F(PMM) | F(PMM_EN)
502 	);
503 }
504 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
505 
506 struct kvm_cpuid_array {
507 	struct kvm_cpuid_entry2 *entries;
508 	int maxnent;
509 	int nent;
510 };
511 
512 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
513 					      u32 function, u32 index)
514 {
515 	struct kvm_cpuid_entry2 *entry;
516 
517 	if (array->nent >= array->maxnent)
518 		return NULL;
519 
520 	entry = &array->entries[array->nent++];
521 
522 	entry->function = function;
523 	entry->index = index;
524 	entry->flags = 0;
525 
526 	cpuid_count(entry->function, entry->index,
527 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
528 
529 	switch (function) {
530 	case 4:
531 	case 7:
532 	case 0xb:
533 	case 0xd:
534 	case 0xf:
535 	case 0x10:
536 	case 0x12:
537 	case 0x14:
538 	case 0x17:
539 	case 0x18:
540 	case 0x1f:
541 	case 0x8000001d:
542 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
543 		break;
544 	}
545 
546 	return entry;
547 }
548 
549 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
550 {
551 	struct kvm_cpuid_entry2 *entry;
552 
553 	if (array->nent >= array->maxnent)
554 		return -E2BIG;
555 
556 	entry = &array->entries[array->nent];
557 	entry->function = func;
558 	entry->index = 0;
559 	entry->flags = 0;
560 
561 	switch (func) {
562 	case 0:
563 		entry->eax = 7;
564 		++array->nent;
565 		break;
566 	case 1:
567 		entry->ecx = F(MOVBE);
568 		++array->nent;
569 		break;
570 	case 7:
571 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
572 		entry->eax = 0;
573 		entry->ecx = F(RDPID);
574 		++array->nent;
575 	default:
576 		break;
577 	}
578 
579 	return 0;
580 }
581 
582 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
583 {
584 	struct kvm_cpuid_entry2 *entry;
585 	int r, i, max_idx;
586 
587 	/* all calls to cpuid_count() should be made on the same cpu */
588 	get_cpu();
589 
590 	r = -E2BIG;
591 
592 	entry = do_host_cpuid(array, function, 0);
593 	if (!entry)
594 		goto out;
595 
596 	switch (function) {
597 	case 0:
598 		/* Limited to the highest leaf implemented in KVM. */
599 		entry->eax = min(entry->eax, 0x1fU);
600 		break;
601 	case 1:
602 		cpuid_entry_override(entry, CPUID_1_EDX);
603 		cpuid_entry_override(entry, CPUID_1_ECX);
604 		break;
605 	case 2:
606 		/*
607 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
608 		 * CPUID(function=2, index=0) may return different results each
609 		 * time, with the least-significant byte in EAX enumerating the
610 		 * number of times software should do CPUID(2, 0).
611 		 *
612 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
613 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
614 		 * return 01H. Software should ignore this value and not
615 		 * interpret it as an informational descriptor", while AMD's
616 		 * APM states that CPUID(2) is reserved.
617 		 *
618 		 * WARN if a frankenstein CPU that supports virtualization and
619 		 * a stateful CPUID.0x2 is encountered.
620 		 */
621 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
622 		break;
623 	/* functions 4 and 0x8000001d have additional index. */
624 	case 4:
625 	case 0x8000001d:
626 		/*
627 		 * Read entries until the cache type in the previous entry is
628 		 * zero, i.e. indicates an invalid entry.
629 		 */
630 		for (i = 1; entry->eax & 0x1f; ++i) {
631 			entry = do_host_cpuid(array, function, i);
632 			if (!entry)
633 				goto out;
634 		}
635 		break;
636 	case 6: /* Thermal management */
637 		entry->eax = 0x4; /* allow ARAT */
638 		entry->ebx = 0;
639 		entry->ecx = 0;
640 		entry->edx = 0;
641 		break;
642 	/* function 7 has additional index. */
643 	case 7:
644 		entry->eax = min(entry->eax, 1u);
645 		cpuid_entry_override(entry, CPUID_7_0_EBX);
646 		cpuid_entry_override(entry, CPUID_7_ECX);
647 		cpuid_entry_override(entry, CPUID_7_EDX);
648 
649 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
650 		if (entry->eax == 1) {
651 			entry = do_host_cpuid(array, function, 1);
652 			if (!entry)
653 				goto out;
654 
655 			cpuid_entry_override(entry, CPUID_7_1_EAX);
656 			entry->ebx = 0;
657 			entry->ecx = 0;
658 			entry->edx = 0;
659 		}
660 		break;
661 	case 9:
662 		break;
663 	case 0xa: { /* Architectural Performance Monitoring */
664 		struct x86_pmu_capability cap;
665 		union cpuid10_eax eax;
666 		union cpuid10_edx edx;
667 
668 		perf_get_x86_pmu_capability(&cap);
669 
670 		/*
671 		 * Only support guest architectural pmu on a host
672 		 * with architectural pmu.
673 		 */
674 		if (!cap.version)
675 			memset(&cap, 0, sizeof(cap));
676 
677 		eax.split.version_id = min(cap.version, 2);
678 		eax.split.num_counters = cap.num_counters_gp;
679 		eax.split.bit_width = cap.bit_width_gp;
680 		eax.split.mask_length = cap.events_mask_len;
681 
682 		edx.split.num_counters_fixed = min(cap.num_counters_fixed, MAX_FIXED_COUNTERS);
683 		edx.split.bit_width_fixed = cap.bit_width_fixed;
684 		edx.split.reserved = 0;
685 
686 		entry->eax = eax.full;
687 		entry->ebx = cap.events_mask;
688 		entry->ecx = 0;
689 		entry->edx = edx.full;
690 		break;
691 	}
692 	/*
693 	 * Per Intel's SDM, the 0x1f is a superset of 0xb,
694 	 * thus they can be handled by common code.
695 	 */
696 	case 0x1f:
697 	case 0xb:
698 		/*
699 		 * Populate entries until the level type (ECX[15:8]) of the
700 		 * previous entry is zero.  Note, CPUID EAX.{0x1f,0xb}.0 is
701 		 * the starting entry, filled by the primary do_host_cpuid().
702 		 */
703 		for (i = 1; entry->ecx & 0xff00; ++i) {
704 			entry = do_host_cpuid(array, function, i);
705 			if (!entry)
706 				goto out;
707 		}
708 		break;
709 	case 0xd:
710 		entry->eax &= supported_xcr0;
711 		entry->ebx = xstate_required_size(supported_xcr0, false);
712 		entry->ecx = entry->ebx;
713 		entry->edx &= supported_xcr0 >> 32;
714 		if (!supported_xcr0)
715 			break;
716 
717 		entry = do_host_cpuid(array, function, 1);
718 		if (!entry)
719 			goto out;
720 
721 		cpuid_entry_override(entry, CPUID_D_1_EAX);
722 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
723 			entry->ebx = xstate_required_size(supported_xcr0 | supported_xss,
724 							  true);
725 		else {
726 			WARN_ON_ONCE(supported_xss != 0);
727 			entry->ebx = 0;
728 		}
729 		entry->ecx &= supported_xss;
730 		entry->edx &= supported_xss >> 32;
731 
732 		for (i = 2; i < 64; ++i) {
733 			bool s_state;
734 			if (supported_xcr0 & BIT_ULL(i))
735 				s_state = false;
736 			else if (supported_xss & BIT_ULL(i))
737 				s_state = true;
738 			else
739 				continue;
740 
741 			entry = do_host_cpuid(array, function, i);
742 			if (!entry)
743 				goto out;
744 
745 			/*
746 			 * The supported check above should have filtered out
747 			 * invalid sub-leafs.  Only valid sub-leafs should
748 			 * reach this point, and they should have a non-zero
749 			 * save state size.  Furthermore, check whether the
750 			 * processor agrees with supported_xcr0/supported_xss
751 			 * on whether this is an XCR0- or IA32_XSS-managed area.
752 			 */
753 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
754 				--array->nent;
755 				continue;
756 			}
757 			entry->edx = 0;
758 		}
759 		break;
760 	/* Intel PT */
761 	case 0x14:
762 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
763 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
764 			break;
765 		}
766 
767 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
768 			if (!do_host_cpuid(array, function, i))
769 				goto out;
770 		}
771 		break;
772 	case KVM_CPUID_SIGNATURE: {
773 		static const char signature[12] = "KVMKVMKVM\0\0";
774 		const u32 *sigptr = (const u32 *)signature;
775 		entry->eax = KVM_CPUID_FEATURES;
776 		entry->ebx = sigptr[0];
777 		entry->ecx = sigptr[1];
778 		entry->edx = sigptr[2];
779 		break;
780 	}
781 	case KVM_CPUID_FEATURES:
782 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
783 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
784 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
785 			     (1 << KVM_FEATURE_ASYNC_PF) |
786 			     (1 << KVM_FEATURE_PV_EOI) |
787 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
788 			     (1 << KVM_FEATURE_PV_UNHALT) |
789 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
790 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
791 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
792 			     (1 << KVM_FEATURE_POLL_CONTROL) |
793 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
794 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
795 
796 		if (sched_info_on())
797 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
798 
799 		entry->ebx = 0;
800 		entry->ecx = 0;
801 		entry->edx = 0;
802 		break;
803 	case 0x80000000:
804 		entry->eax = min(entry->eax, 0x8000001f);
805 		break;
806 	case 0x80000001:
807 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
808 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
809 		break;
810 	case 0x80000006:
811 		/* L2 cache and TLB: pass through host info. */
812 		break;
813 	case 0x80000007: /* Advanced power management */
814 		/* invariant TSC is CPUID.80000007H:EDX[8] */
815 		entry->edx &= (1 << 8);
816 		/* mask against host */
817 		entry->edx &= boot_cpu_data.x86_power;
818 		entry->eax = entry->ebx = entry->ecx = 0;
819 		break;
820 	case 0x80000008: {
821 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
822 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
823 		unsigned phys_as = entry->eax & 0xff;
824 
825 		if (!g_phys_as)
826 			g_phys_as = phys_as;
827 		entry->eax = g_phys_as | (virt_as << 8);
828 		entry->edx = 0;
829 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
830 		break;
831 	}
832 	case 0x8000000A:
833 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
834 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
835 			break;
836 		}
837 		entry->eax = 1; /* SVM revision 1 */
838 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
839 				   ASID emulation to nested SVM */
840 		entry->ecx = 0; /* Reserved */
841 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
842 		break;
843 	case 0x80000019:
844 		entry->ecx = entry->edx = 0;
845 		break;
846 	case 0x8000001a:
847 	case 0x8000001e:
848 		break;
849 	/* Support memory encryption cpuid if host supports it */
850 	case 0x8000001F:
851 		if (!boot_cpu_has(X86_FEATURE_SEV))
852 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
853 		break;
854 	/*Add support for Centaur's CPUID instruction*/
855 	case 0xC0000000:
856 		/*Just support up to 0xC0000004 now*/
857 		entry->eax = min(entry->eax, 0xC0000004);
858 		break;
859 	case 0xC0000001:
860 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
861 		break;
862 	case 3: /* Processor serial number */
863 	case 5: /* MONITOR/MWAIT */
864 	case 0xC0000002:
865 	case 0xC0000003:
866 	case 0xC0000004:
867 	default:
868 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
869 		break;
870 	}
871 
872 	r = 0;
873 
874 out:
875 	put_cpu();
876 
877 	return r;
878 }
879 
880 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
881 			 unsigned int type)
882 {
883 	if (type == KVM_GET_EMULATED_CPUID)
884 		return __do_cpuid_func_emulated(array, func);
885 
886 	return __do_cpuid_func(array, func);
887 }
888 
889 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
890 
891 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
892 			  unsigned int type)
893 {
894 	u32 limit;
895 	int r;
896 
897 	if (func == CENTAUR_CPUID_SIGNATURE &&
898 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
899 		return 0;
900 
901 	r = do_cpuid_func(array, func, type);
902 	if (r)
903 		return r;
904 
905 	limit = array->entries[array->nent - 1].eax;
906 	for (func = func + 1; func <= limit; ++func) {
907 		r = do_cpuid_func(array, func, type);
908 		if (r)
909 			break;
910 	}
911 
912 	return r;
913 }
914 
915 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
916 				 __u32 num_entries, unsigned int ioctl_type)
917 {
918 	int i;
919 	__u32 pad[3];
920 
921 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
922 		return false;
923 
924 	/*
925 	 * We want to make sure that ->padding is being passed clean from
926 	 * userspace in case we want to use it for something in the future.
927 	 *
928 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
929 	 * have to give ourselves satisfied only with the emulated side. /me
930 	 * sheds a tear.
931 	 */
932 	for (i = 0; i < num_entries; i++) {
933 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
934 			return true;
935 
936 		if (pad[0] || pad[1] || pad[2])
937 			return true;
938 	}
939 	return false;
940 }
941 
942 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
943 			    struct kvm_cpuid_entry2 __user *entries,
944 			    unsigned int type)
945 {
946 	static const u32 funcs[] = {
947 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
948 	};
949 
950 	struct kvm_cpuid_array array = {
951 		.nent = 0,
952 	};
953 	int r, i;
954 
955 	if (cpuid->nent < 1)
956 		return -E2BIG;
957 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
958 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
959 
960 	if (sanity_check_entries(entries, cpuid->nent, type))
961 		return -EINVAL;
962 
963 	array.entries = vzalloc(array_size(sizeof(struct kvm_cpuid_entry2),
964 					   cpuid->nent));
965 	if (!array.entries)
966 		return -ENOMEM;
967 
968 	array.maxnent = cpuid->nent;
969 
970 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
971 		r = get_cpuid_func(&array, funcs[i], type);
972 		if (r)
973 			goto out_free;
974 	}
975 	cpuid->nent = array.nent;
976 
977 	if (copy_to_user(entries, array.entries,
978 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
979 		r = -EFAULT;
980 
981 out_free:
982 	vfree(array.entries);
983 	return r;
984 }
985 
986 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
987 					      u32 function, u32 index)
988 {
989 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
990 				 function, index);
991 }
992 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
993 
994 /*
995  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
996  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
997  * returns all zeroes for any undefined leaf, whether or not the leaf is in
998  * range.  Centaur/VIA follows Intel semantics.
999  *
1000  * A leaf is considered out-of-range if its function is higher than the maximum
1001  * supported leaf of its associated class or if its associated class does not
1002  * exist.
1003  *
1004  * There are three primary classes to be considered, with their respective
1005  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1006  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1007  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1008  *
1009  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1010  *  - Hypervisor: 0x40000000 - 0x4fffffff
1011  *  - Extended:   0x80000000 - 0xbfffffff
1012  *  - Centaur:    0xc0000000 - 0xcfffffff
1013  *
1014  * The Hypervisor class is further subdivided into sub-classes that each act as
1015  * their own indepdent class associated with a 0x100 byte range.  E.g. if Qemu
1016  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1017  * CPUID sub-classes are:
1018  *
1019  *  - HyperV:     0x40000000 - 0x400000ff
1020  *  - KVM:        0x40000100 - 0x400001ff
1021  */
1022 static struct kvm_cpuid_entry2 *
1023 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1024 {
1025 	struct kvm_cpuid_entry2 *basic, *class;
1026 	u32 function = *fn_ptr;
1027 
1028 	basic = kvm_find_cpuid_entry(vcpu, 0, 0);
1029 	if (!basic)
1030 		return NULL;
1031 
1032 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1033 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1034 		return NULL;
1035 
1036 	if (function >= 0x40000000 && function <= 0x4fffffff)
1037 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00, 0);
1038 	else if (function >= 0xc0000000)
1039 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000, 0);
1040 	else
1041 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
1042 
1043 	if (class && function <= class->eax)
1044 		return NULL;
1045 
1046 	/*
1047 	 * Leaf specific adjustments are also applied when redirecting to the
1048 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1049 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1050 	 * needs to be pulled from CPUID.0xb.1.
1051 	 */
1052 	*fn_ptr = basic->eax;
1053 
1054 	/*
1055 	 * The class does not exist or the requested function is out of range;
1056 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1057 	 * the original requested leaf is observed!
1058 	 */
1059 	return kvm_find_cpuid_entry(vcpu, basic->eax, index);
1060 }
1061 
1062 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1063 	       u32 *ecx, u32 *edx, bool exact_only)
1064 {
1065 	u32 orig_function = *eax, function = *eax, index = *ecx;
1066 	struct kvm_cpuid_entry2 *entry;
1067 	bool exact, used_max_basic = false;
1068 
1069 	entry = kvm_find_cpuid_entry(vcpu, function, index);
1070 	exact = !!entry;
1071 
1072 	if (!entry && !exact_only) {
1073 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1074 		used_max_basic = !!entry;
1075 	}
1076 
1077 	if (entry) {
1078 		*eax = entry->eax;
1079 		*ebx = entry->ebx;
1080 		*ecx = entry->ecx;
1081 		*edx = entry->edx;
1082 		if (function == 7 && index == 0) {
1083 			u64 data;
1084 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1085 			    (data & TSX_CTRL_CPUID_CLEAR))
1086 				*ebx &= ~(F(RTM) | F(HLE));
1087 		}
1088 	} else {
1089 		*eax = *ebx = *ecx = *edx = 0;
1090 		/*
1091 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1092 		 * and EDX is always the x2APIC ID, even for undefined
1093 		 * subleaves. Index 1 will exist iff the leaf is
1094 		 * implemented, so we pass through CL iff leaf 1
1095 		 * exists. EDX can be copied from any existing index.
1096 		 */
1097 		if (function == 0xb || function == 0x1f) {
1098 			entry = kvm_find_cpuid_entry(vcpu, function, 1);
1099 			if (entry) {
1100 				*ecx = index & 0xff;
1101 				*edx = entry->edx;
1102 			}
1103 		}
1104 	}
1105 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1106 			used_max_basic);
1107 	return exact;
1108 }
1109 EXPORT_SYMBOL_GPL(kvm_cpuid);
1110 
1111 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1112 {
1113 	u32 eax, ebx, ecx, edx;
1114 
1115 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1116 		return 1;
1117 
1118 	eax = kvm_rax_read(vcpu);
1119 	ecx = kvm_rcx_read(vcpu);
1120 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1121 	kvm_rax_write(vcpu, eax);
1122 	kvm_rbx_write(vcpu, ebx);
1123 	kvm_rcx_write(vcpu, ecx);
1124 	kvm_rdx_write(vcpu, edx);
1125 	return kvm_skip_emulated_instruction(vcpu);
1126 }
1127 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1128