xref: /openbmc/linux/arch/x86/kvm/cpuid.c (revision 07d9a767)
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 	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 
187 static int is_efer_nx(void)
188 {
189 	return host_efer & EFER_NX;
190 }
191 
192 static 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 
211 int 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;
221 not_found:
222 	return 36;
223 }
224 
225 /* when an old userspace process fills a new kernel module */
226 int 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 
275 out_free_cpuid:
276 	kvfree(e);
277 
278 	return r;
279 }
280 
281 int 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 
313 int 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 
328 out:
329 	cpuid->nent = vcpu->arch.cpuid_nent;
330 	return r;
331 }
332 
333 static __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 
347 void 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 	/*
495 	 * Hide all SVM features by default, SVM will set the cap bits for
496 	 * features it emulates and/or exposes for L1.
497 	 */
498 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
499 
500 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
501 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
502 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
503 		F(PMM) | F(PMM_EN)
504 	);
505 }
506 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
507 
508 struct kvm_cpuid_array {
509 	struct kvm_cpuid_entry2 *entries;
510 	int maxnent;
511 	int nent;
512 };
513 
514 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
515 					      u32 function, u32 index)
516 {
517 	struct kvm_cpuid_entry2 *entry;
518 
519 	if (array->nent >= array->maxnent)
520 		return NULL;
521 
522 	entry = &array->entries[array->nent++];
523 
524 	entry->function = function;
525 	entry->index = index;
526 	entry->flags = 0;
527 
528 	cpuid_count(entry->function, entry->index,
529 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
530 
531 	switch (function) {
532 	case 4:
533 	case 7:
534 	case 0xb:
535 	case 0xd:
536 	case 0xf:
537 	case 0x10:
538 	case 0x12:
539 	case 0x14:
540 	case 0x17:
541 	case 0x18:
542 	case 0x1f:
543 	case 0x8000001d:
544 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
545 		break;
546 	}
547 
548 	return entry;
549 }
550 
551 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
552 {
553 	struct kvm_cpuid_entry2 *entry;
554 
555 	if (array->nent >= array->maxnent)
556 		return -E2BIG;
557 
558 	entry = &array->entries[array->nent];
559 	entry->function = func;
560 	entry->index = 0;
561 	entry->flags = 0;
562 
563 	switch (func) {
564 	case 0:
565 		entry->eax = 7;
566 		++array->nent;
567 		break;
568 	case 1:
569 		entry->ecx = F(MOVBE);
570 		++array->nent;
571 		break;
572 	case 7:
573 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
574 		entry->eax = 0;
575 		entry->ecx = F(RDPID);
576 		++array->nent;
577 	default:
578 		break;
579 	}
580 
581 	return 0;
582 }
583 
584 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
585 {
586 	struct kvm_cpuid_entry2 *entry;
587 	int r, i, max_idx;
588 
589 	/* all calls to cpuid_count() should be made on the same cpu */
590 	get_cpu();
591 
592 	r = -E2BIG;
593 
594 	entry = do_host_cpuid(array, function, 0);
595 	if (!entry)
596 		goto out;
597 
598 	switch (function) {
599 	case 0:
600 		/* Limited to the highest leaf implemented in KVM. */
601 		entry->eax = min(entry->eax, 0x1fU);
602 		break;
603 	case 1:
604 		cpuid_entry_override(entry, CPUID_1_EDX);
605 		cpuid_entry_override(entry, CPUID_1_ECX);
606 		break;
607 	case 2:
608 		/*
609 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
610 		 * CPUID(function=2, index=0) may return different results each
611 		 * time, with the least-significant byte in EAX enumerating the
612 		 * number of times software should do CPUID(2, 0).
613 		 *
614 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
615 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
616 		 * return 01H. Software should ignore this value and not
617 		 * interpret it as an informational descriptor", while AMD's
618 		 * APM states that CPUID(2) is reserved.
619 		 *
620 		 * WARN if a frankenstein CPU that supports virtualization and
621 		 * a stateful CPUID.0x2 is encountered.
622 		 */
623 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
624 		break;
625 	/* functions 4 and 0x8000001d have additional index. */
626 	case 4:
627 	case 0x8000001d:
628 		/*
629 		 * Read entries until the cache type in the previous entry is
630 		 * zero, i.e. indicates an invalid entry.
631 		 */
632 		for (i = 1; entry->eax & 0x1f; ++i) {
633 			entry = do_host_cpuid(array, function, i);
634 			if (!entry)
635 				goto out;
636 		}
637 		break;
638 	case 6: /* Thermal management */
639 		entry->eax = 0x4; /* allow ARAT */
640 		entry->ebx = 0;
641 		entry->ecx = 0;
642 		entry->edx = 0;
643 		break;
644 	/* function 7 has additional index. */
645 	case 7:
646 		entry->eax = min(entry->eax, 1u);
647 		cpuid_entry_override(entry, CPUID_7_0_EBX);
648 		cpuid_entry_override(entry, CPUID_7_ECX);
649 		cpuid_entry_override(entry, CPUID_7_EDX);
650 
651 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
652 		if (entry->eax == 1) {
653 			entry = do_host_cpuid(array, function, 1);
654 			if (!entry)
655 				goto out;
656 
657 			cpuid_entry_override(entry, CPUID_7_1_EAX);
658 			entry->ebx = 0;
659 			entry->ecx = 0;
660 			entry->edx = 0;
661 		}
662 		break;
663 	case 9:
664 		break;
665 	case 0xa: { /* Architectural Performance Monitoring */
666 		struct x86_pmu_capability cap;
667 		union cpuid10_eax eax;
668 		union cpuid10_edx edx;
669 
670 		perf_get_x86_pmu_capability(&cap);
671 
672 		/*
673 		 * Only support guest architectural pmu on a host
674 		 * with architectural pmu.
675 		 */
676 		if (!cap.version)
677 			memset(&cap, 0, sizeof(cap));
678 
679 		eax.split.version_id = min(cap.version, 2);
680 		eax.split.num_counters = cap.num_counters_gp;
681 		eax.split.bit_width = cap.bit_width_gp;
682 		eax.split.mask_length = cap.events_mask_len;
683 
684 		edx.split.num_counters_fixed = min(cap.num_counters_fixed, MAX_FIXED_COUNTERS);
685 		edx.split.bit_width_fixed = cap.bit_width_fixed;
686 		edx.split.anythread_deprecated = 1;
687 		edx.split.reserved1 = 0;
688 		edx.split.reserved2 = 0;
689 
690 		entry->eax = eax.full;
691 		entry->ebx = cap.events_mask;
692 		entry->ecx = 0;
693 		entry->edx = edx.full;
694 		break;
695 	}
696 	/*
697 	 * Per Intel's SDM, the 0x1f is a superset of 0xb,
698 	 * thus they can be handled by common code.
699 	 */
700 	case 0x1f:
701 	case 0xb:
702 		/*
703 		 * Populate entries until the level type (ECX[15:8]) of the
704 		 * previous entry is zero.  Note, CPUID EAX.{0x1f,0xb}.0 is
705 		 * the starting entry, filled by the primary do_host_cpuid().
706 		 */
707 		for (i = 1; entry->ecx & 0xff00; ++i) {
708 			entry = do_host_cpuid(array, function, i);
709 			if (!entry)
710 				goto out;
711 		}
712 		break;
713 	case 0xd:
714 		entry->eax &= supported_xcr0;
715 		entry->ebx = xstate_required_size(supported_xcr0, false);
716 		entry->ecx = entry->ebx;
717 		entry->edx &= supported_xcr0 >> 32;
718 		if (!supported_xcr0)
719 			break;
720 
721 		entry = do_host_cpuid(array, function, 1);
722 		if (!entry)
723 			goto out;
724 
725 		cpuid_entry_override(entry, CPUID_D_1_EAX);
726 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
727 			entry->ebx = xstate_required_size(supported_xcr0 | supported_xss,
728 							  true);
729 		else {
730 			WARN_ON_ONCE(supported_xss != 0);
731 			entry->ebx = 0;
732 		}
733 		entry->ecx &= supported_xss;
734 		entry->edx &= supported_xss >> 32;
735 
736 		for (i = 2; i < 64; ++i) {
737 			bool s_state;
738 			if (supported_xcr0 & BIT_ULL(i))
739 				s_state = false;
740 			else if (supported_xss & BIT_ULL(i))
741 				s_state = true;
742 			else
743 				continue;
744 
745 			entry = do_host_cpuid(array, function, i);
746 			if (!entry)
747 				goto out;
748 
749 			/*
750 			 * The supported check above should have filtered out
751 			 * invalid sub-leafs.  Only valid sub-leafs should
752 			 * reach this point, and they should have a non-zero
753 			 * save state size.  Furthermore, check whether the
754 			 * processor agrees with supported_xcr0/supported_xss
755 			 * on whether this is an XCR0- or IA32_XSS-managed area.
756 			 */
757 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
758 				--array->nent;
759 				continue;
760 			}
761 			entry->edx = 0;
762 		}
763 		break;
764 	/* Intel PT */
765 	case 0x14:
766 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
767 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
768 			break;
769 		}
770 
771 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
772 			if (!do_host_cpuid(array, function, i))
773 				goto out;
774 		}
775 		break;
776 	case KVM_CPUID_SIGNATURE: {
777 		static const char signature[12] = "KVMKVMKVM\0\0";
778 		const u32 *sigptr = (const u32 *)signature;
779 		entry->eax = KVM_CPUID_FEATURES;
780 		entry->ebx = sigptr[0];
781 		entry->ecx = sigptr[1];
782 		entry->edx = sigptr[2];
783 		break;
784 	}
785 	case KVM_CPUID_FEATURES:
786 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
787 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
788 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
789 			     (1 << KVM_FEATURE_ASYNC_PF) |
790 			     (1 << KVM_FEATURE_PV_EOI) |
791 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
792 			     (1 << KVM_FEATURE_PV_UNHALT) |
793 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
794 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
795 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
796 			     (1 << KVM_FEATURE_POLL_CONTROL) |
797 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
798 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
799 
800 		if (sched_info_on())
801 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
802 
803 		entry->ebx = 0;
804 		entry->ecx = 0;
805 		entry->edx = 0;
806 		break;
807 	case 0x80000000:
808 		entry->eax = min(entry->eax, 0x8000001f);
809 		break;
810 	case 0x80000001:
811 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
812 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
813 		break;
814 	case 0x80000006:
815 		/* L2 cache and TLB: pass through host info. */
816 		break;
817 	case 0x80000007: /* Advanced power management */
818 		/* invariant TSC is CPUID.80000007H:EDX[8] */
819 		entry->edx &= (1 << 8);
820 		/* mask against host */
821 		entry->edx &= boot_cpu_data.x86_power;
822 		entry->eax = entry->ebx = entry->ecx = 0;
823 		break;
824 	case 0x80000008: {
825 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
826 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
827 		unsigned phys_as = entry->eax & 0xff;
828 
829 		if (!g_phys_as)
830 			g_phys_as = phys_as;
831 		entry->eax = g_phys_as | (virt_as << 8);
832 		entry->edx = 0;
833 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
834 		break;
835 	}
836 	case 0x8000000A:
837 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
838 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
839 			break;
840 		}
841 		entry->eax = 1; /* SVM revision 1 */
842 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
843 				   ASID emulation to nested SVM */
844 		entry->ecx = 0; /* Reserved */
845 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
846 		break;
847 	case 0x80000019:
848 		entry->ecx = entry->edx = 0;
849 		break;
850 	case 0x8000001a:
851 	case 0x8000001e:
852 		break;
853 	/* Support memory encryption cpuid if host supports it */
854 	case 0x8000001F:
855 		if (!boot_cpu_has(X86_FEATURE_SEV))
856 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
857 		break;
858 	/*Add support for Centaur's CPUID instruction*/
859 	case 0xC0000000:
860 		/*Just support up to 0xC0000004 now*/
861 		entry->eax = min(entry->eax, 0xC0000004);
862 		break;
863 	case 0xC0000001:
864 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
865 		break;
866 	case 3: /* Processor serial number */
867 	case 5: /* MONITOR/MWAIT */
868 	case 0xC0000002:
869 	case 0xC0000003:
870 	case 0xC0000004:
871 	default:
872 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
873 		break;
874 	}
875 
876 	r = 0;
877 
878 out:
879 	put_cpu();
880 
881 	return r;
882 }
883 
884 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
885 			 unsigned int type)
886 {
887 	if (type == KVM_GET_EMULATED_CPUID)
888 		return __do_cpuid_func_emulated(array, func);
889 
890 	return __do_cpuid_func(array, func);
891 }
892 
893 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
894 
895 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
896 			  unsigned int type)
897 {
898 	u32 limit;
899 	int r;
900 
901 	if (func == CENTAUR_CPUID_SIGNATURE &&
902 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
903 		return 0;
904 
905 	r = do_cpuid_func(array, func, type);
906 	if (r)
907 		return r;
908 
909 	limit = array->entries[array->nent - 1].eax;
910 	for (func = func + 1; func <= limit; ++func) {
911 		r = do_cpuid_func(array, func, type);
912 		if (r)
913 			break;
914 	}
915 
916 	return r;
917 }
918 
919 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
920 				 __u32 num_entries, unsigned int ioctl_type)
921 {
922 	int i;
923 	__u32 pad[3];
924 
925 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
926 		return false;
927 
928 	/*
929 	 * We want to make sure that ->padding is being passed clean from
930 	 * userspace in case we want to use it for something in the future.
931 	 *
932 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
933 	 * have to give ourselves satisfied only with the emulated side. /me
934 	 * sheds a tear.
935 	 */
936 	for (i = 0; i < num_entries; i++) {
937 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
938 			return true;
939 
940 		if (pad[0] || pad[1] || pad[2])
941 			return true;
942 	}
943 	return false;
944 }
945 
946 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
947 			    struct kvm_cpuid_entry2 __user *entries,
948 			    unsigned int type)
949 {
950 	static const u32 funcs[] = {
951 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
952 	};
953 
954 	struct kvm_cpuid_array array = {
955 		.nent = 0,
956 	};
957 	int r, i;
958 
959 	if (cpuid->nent < 1)
960 		return -E2BIG;
961 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
962 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
963 
964 	if (sanity_check_entries(entries, cpuid->nent, type))
965 		return -EINVAL;
966 
967 	array.entries = vzalloc(array_size(sizeof(struct kvm_cpuid_entry2),
968 					   cpuid->nent));
969 	if (!array.entries)
970 		return -ENOMEM;
971 
972 	array.maxnent = cpuid->nent;
973 
974 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
975 		r = get_cpuid_func(&array, funcs[i], type);
976 		if (r)
977 			goto out_free;
978 	}
979 	cpuid->nent = array.nent;
980 
981 	if (copy_to_user(entries, array.entries,
982 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
983 		r = -EFAULT;
984 
985 out_free:
986 	vfree(array.entries);
987 	return r;
988 }
989 
990 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
991 					      u32 function, u32 index)
992 {
993 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
994 				 function, index);
995 }
996 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
997 
998 /*
999  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1000  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1001  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1002  * range.  Centaur/VIA follows Intel semantics.
1003  *
1004  * A leaf is considered out-of-range if its function is higher than the maximum
1005  * supported leaf of its associated class or if its associated class does not
1006  * exist.
1007  *
1008  * There are three primary classes to be considered, with their respective
1009  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1010  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1011  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1012  *
1013  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1014  *  - Hypervisor: 0x40000000 - 0x4fffffff
1015  *  - Extended:   0x80000000 - 0xbfffffff
1016  *  - Centaur:    0xc0000000 - 0xcfffffff
1017  *
1018  * The Hypervisor class is further subdivided into sub-classes that each act as
1019  * their own indepdent class associated with a 0x100 byte range.  E.g. if Qemu
1020  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1021  * CPUID sub-classes are:
1022  *
1023  *  - HyperV:     0x40000000 - 0x400000ff
1024  *  - KVM:        0x40000100 - 0x400001ff
1025  */
1026 static struct kvm_cpuid_entry2 *
1027 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1028 {
1029 	struct kvm_cpuid_entry2 *basic, *class;
1030 	u32 function = *fn_ptr;
1031 
1032 	basic = kvm_find_cpuid_entry(vcpu, 0, 0);
1033 	if (!basic)
1034 		return NULL;
1035 
1036 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1037 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1038 		return NULL;
1039 
1040 	if (function >= 0x40000000 && function <= 0x4fffffff)
1041 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00, 0);
1042 	else if (function >= 0xc0000000)
1043 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000, 0);
1044 	else
1045 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
1046 
1047 	if (class && function <= class->eax)
1048 		return NULL;
1049 
1050 	/*
1051 	 * Leaf specific adjustments are also applied when redirecting to the
1052 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1053 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1054 	 * needs to be pulled from CPUID.0xb.1.
1055 	 */
1056 	*fn_ptr = basic->eax;
1057 
1058 	/*
1059 	 * The class does not exist or the requested function is out of range;
1060 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1061 	 * the original requested leaf is observed!
1062 	 */
1063 	return kvm_find_cpuid_entry(vcpu, basic->eax, index);
1064 }
1065 
1066 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1067 	       u32 *ecx, u32 *edx, bool exact_only)
1068 {
1069 	u32 orig_function = *eax, function = *eax, index = *ecx;
1070 	struct kvm_cpuid_entry2 *entry;
1071 	bool exact, used_max_basic = false;
1072 
1073 	entry = kvm_find_cpuid_entry(vcpu, function, index);
1074 	exact = !!entry;
1075 
1076 	if (!entry && !exact_only) {
1077 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1078 		used_max_basic = !!entry;
1079 	}
1080 
1081 	if (entry) {
1082 		*eax = entry->eax;
1083 		*ebx = entry->ebx;
1084 		*ecx = entry->ecx;
1085 		*edx = entry->edx;
1086 		if (function == 7 && index == 0) {
1087 			u64 data;
1088 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1089 			    (data & TSX_CTRL_CPUID_CLEAR))
1090 				*ebx &= ~(F(RTM) | F(HLE));
1091 		}
1092 	} else {
1093 		*eax = *ebx = *ecx = *edx = 0;
1094 		/*
1095 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1096 		 * and EDX is always the x2APIC ID, even for undefined
1097 		 * subleaves. Index 1 will exist iff the leaf is
1098 		 * implemented, so we pass through CL iff leaf 1
1099 		 * exists. EDX can be copied from any existing index.
1100 		 */
1101 		if (function == 0xb || function == 0x1f) {
1102 			entry = kvm_find_cpuid_entry(vcpu, function, 1);
1103 			if (entry) {
1104 				*ecx = index & 0xff;
1105 				*edx = entry->edx;
1106 			}
1107 		}
1108 	}
1109 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1110 			used_max_basic);
1111 	return exact;
1112 }
1113 EXPORT_SYMBOL_GPL(kvm_cpuid);
1114 
1115 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1116 {
1117 	u32 eax, ebx, ecx, edx;
1118 
1119 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1120 		return 1;
1121 
1122 	eax = kvm_rax_read(vcpu);
1123 	ecx = kvm_rcx_read(vcpu);
1124 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1125 	kvm_rax_write(vcpu, eax);
1126 	kvm_rbx_write(vcpu, ebx);
1127 	kvm_rcx_write(vcpu, ecx);
1128 	kvm_rdx_write(vcpu, edx);
1129 	return kvm_skip_emulated_instruction(vcpu);
1130 }
1131 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1132