xref: /openbmc/linux/arch/x86/kvm/cpuid.c (revision fbb6b31a)
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 <asm/sgx.h>
22 #include "cpuid.h"
23 #include "lapic.h"
24 #include "mmu.h"
25 #include "trace.h"
26 #include "pmu.h"
27 
28 /*
29  * Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be
30  * aligned to sizeof(unsigned long) because it's not accessed via bitops.
31  */
32 u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
33 EXPORT_SYMBOL_GPL(kvm_cpu_caps);
34 
35 u32 xstate_required_size(u64 xstate_bv, bool compacted)
36 {
37 	int feature_bit = 0;
38 	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
39 
40 	xstate_bv &= XFEATURE_MASK_EXTEND;
41 	while (xstate_bv) {
42 		if (xstate_bv & 0x1) {
43 		        u32 eax, ebx, ecx, edx, offset;
44 		        cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
45 			/* ECX[1]: 64B alignment in compacted form */
46 			if (compacted)
47 				offset = (ecx & 0x2) ? ALIGN(ret, 64) : ret;
48 			else
49 				offset = ebx;
50 			ret = max(ret, offset + eax);
51 		}
52 
53 		xstate_bv >>= 1;
54 		feature_bit++;
55 	}
56 
57 	return ret;
58 }
59 
60 /*
61  * This one is tied to SSB in the user API, and not
62  * visible in /proc/cpuinfo.
63  */
64 #define KVM_X86_FEATURE_PSFD		(13*32+28) /* Predictive Store Forwarding Disable */
65 
66 #define F feature_bit
67 #define SF(name) (boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0)
68 
69 
70 static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
71 	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u32 index)
72 {
73 	struct kvm_cpuid_entry2 *e;
74 	int i;
75 
76 	for (i = 0; i < nent; i++) {
77 		e = &entries[i];
78 
79 		if (e->function == function &&
80 		    (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index))
81 			return e;
82 	}
83 
84 	return NULL;
85 }
86 
87 static int kvm_check_cpuid(struct kvm_vcpu *vcpu,
88 			   struct kvm_cpuid_entry2 *entries,
89 			   int nent)
90 {
91 	struct kvm_cpuid_entry2 *best;
92 	u64 xfeatures;
93 
94 	/*
95 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
96 	 * canonical address checks; exit if it is ever changed.
97 	 */
98 	best = cpuid_entry2_find(entries, nent, 0x80000008, 0);
99 	if (best) {
100 		int vaddr_bits = (best->eax & 0xff00) >> 8;
101 
102 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
103 			return -EINVAL;
104 	}
105 
106 	/*
107 	 * Exposing dynamic xfeatures to the guest requires additional
108 	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
109 	 */
110 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
111 	if (!best)
112 		return 0;
113 
114 	xfeatures = best->eax | ((u64)best->edx << 32);
115 	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
116 	if (!xfeatures)
117 		return 0;
118 
119 	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
120 }
121 
122 /* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
123 static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
124 				 int nent)
125 {
126 	struct kvm_cpuid_entry2 *orig;
127 	int i;
128 
129 	if (nent != vcpu->arch.cpuid_nent)
130 		return -EINVAL;
131 
132 	for (i = 0; i < nent; i++) {
133 		orig = &vcpu->arch.cpuid_entries[i];
134 		if (e2[i].function != orig->function ||
135 		    e2[i].index != orig->index ||
136 		    e2[i].flags != orig->flags ||
137 		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
138 		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
139 			return -EINVAL;
140 	}
141 
142 	return 0;
143 }
144 
145 static void kvm_update_kvm_cpuid_base(struct kvm_vcpu *vcpu)
146 {
147 	u32 function;
148 	struct kvm_cpuid_entry2 *entry;
149 
150 	vcpu->arch.kvm_cpuid_base = 0;
151 
152 	for_each_possible_hypervisor_cpuid_base(function) {
153 		entry = kvm_find_cpuid_entry(vcpu, function, 0);
154 
155 		if (entry) {
156 			u32 signature[3];
157 
158 			signature[0] = entry->ebx;
159 			signature[1] = entry->ecx;
160 			signature[2] = entry->edx;
161 
162 			BUILD_BUG_ON(sizeof(signature) > sizeof(KVM_SIGNATURE));
163 			if (!memcmp(signature, KVM_SIGNATURE, sizeof(signature))) {
164 				vcpu->arch.kvm_cpuid_base = function;
165 				break;
166 			}
167 		}
168 	}
169 }
170 
171 static struct kvm_cpuid_entry2 *__kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu,
172 					      struct kvm_cpuid_entry2 *entries, int nent)
173 {
174 	u32 base = vcpu->arch.kvm_cpuid_base;
175 
176 	if (!base)
177 		return NULL;
178 
179 	return cpuid_entry2_find(entries, nent, base | KVM_CPUID_FEATURES, 0);
180 }
181 
182 static struct kvm_cpuid_entry2 *kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu)
183 {
184 	return __kvm_find_kvm_cpuid_features(vcpu, vcpu->arch.cpuid_entries,
185 					     vcpu->arch.cpuid_nent);
186 }
187 
188 void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
189 {
190 	struct kvm_cpuid_entry2 *best = kvm_find_kvm_cpuid_features(vcpu);
191 
192 	/*
193 	 * save the feature bitmap to avoid cpuid lookup for every PV
194 	 * operation
195 	 */
196 	if (best)
197 		vcpu->arch.pv_cpuid.features = best->eax;
198 }
199 
200 /*
201  * Calculate guest's supported XCR0 taking into account guest CPUID data and
202  * supported_xcr0 (comprised of host configuration and KVM_SUPPORTED_XCR0).
203  */
204 static u64 cpuid_get_supported_xcr0(struct kvm_cpuid_entry2 *entries, int nent)
205 {
206 	struct kvm_cpuid_entry2 *best;
207 
208 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
209 	if (!best)
210 		return 0;
211 
212 	return (best->eax | ((u64)best->edx << 32)) & supported_xcr0;
213 }
214 
215 static void __kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *entries,
216 				       int nent)
217 {
218 	struct kvm_cpuid_entry2 *best;
219 	u64 guest_supported_xcr0 = cpuid_get_supported_xcr0(entries, nent);
220 
221 	best = cpuid_entry2_find(entries, nent, 1, 0);
222 	if (best) {
223 		/* Update OSXSAVE bit */
224 		if (boot_cpu_has(X86_FEATURE_XSAVE))
225 			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
226 				   kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE));
227 
228 		cpuid_entry_change(best, X86_FEATURE_APIC,
229 			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
230 	}
231 
232 	best = cpuid_entry2_find(entries, nent, 7, 0);
233 	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
234 		cpuid_entry_change(best, X86_FEATURE_OSPKE,
235 				   kvm_read_cr4_bits(vcpu, X86_CR4_PKE));
236 
237 	best = cpuid_entry2_find(entries, nent, 0xD, 0);
238 	if (best)
239 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
240 
241 	best = cpuid_entry2_find(entries, nent, 0xD, 1);
242 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
243 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
244 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
245 
246 	best = __kvm_find_kvm_cpuid_features(vcpu, entries, nent);
247 	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
248 		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
249 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
250 
251 	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
252 		best = cpuid_entry2_find(entries, nent, 0x1, 0);
253 		if (best)
254 			cpuid_entry_change(best, X86_FEATURE_MWAIT,
255 					   vcpu->arch.ia32_misc_enable_msr &
256 					   MSR_IA32_MISC_ENABLE_MWAIT);
257 	}
258 
259 	/*
260 	 * Bits 127:0 of the allowed SECS.ATTRIBUTES (CPUID.0x12.0x1) enumerate
261 	 * the supported XSAVE Feature Request Mask (XFRM), i.e. the enclave's
262 	 * requested XCR0 value.  The enclave's XFRM must be a subset of XCRO
263 	 * at the time of EENTER, thus adjust the allowed XFRM by the guest's
264 	 * supported XCR0.  Similar to XCR0 handling, FP and SSE are forced to
265 	 * '1' even on CPUs that don't support XSAVE.
266 	 */
267 	best = cpuid_entry2_find(entries, nent, 0x12, 0x1);
268 	if (best) {
269 		best->ecx &= guest_supported_xcr0 & 0xffffffff;
270 		best->edx &= guest_supported_xcr0 >> 32;
271 		best->ecx |= XFEATURE_MASK_FPSSE;
272 	}
273 }
274 
275 void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
276 {
277 	__kvm_update_cpuid_runtime(vcpu, vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
278 }
279 EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
280 
281 static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
282 {
283 	struct kvm_lapic *apic = vcpu->arch.apic;
284 	struct kvm_cpuid_entry2 *best;
285 	u64 guest_supported_xcr0;
286 
287 	best = kvm_find_cpuid_entry(vcpu, 1, 0);
288 	if (best && apic) {
289 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
290 			apic->lapic_timer.timer_mode_mask = 3 << 17;
291 		else
292 			apic->lapic_timer.timer_mode_mask = 1 << 17;
293 
294 		kvm_apic_set_version(vcpu);
295 	}
296 
297 	guest_supported_xcr0 =
298 		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
299 
300 	vcpu->arch.guest_fpu.fpstate->user_xfeatures = guest_supported_xcr0;
301 
302 	kvm_update_pv_runtime(vcpu);
303 
304 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
305 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
306 
307 	kvm_pmu_refresh(vcpu);
308 	vcpu->arch.cr4_guest_rsvd_bits =
309 	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
310 
311 	kvm_hv_set_cpuid(vcpu);
312 
313 	/* Invoke the vendor callback only after the above state is updated. */
314 	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
315 
316 	/*
317 	 * Except for the MMU, which needs to do its thing any vendor specific
318 	 * adjustments to the reserved GPA bits.
319 	 */
320 	kvm_mmu_after_set_cpuid(vcpu);
321 }
322 
323 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
324 {
325 	struct kvm_cpuid_entry2 *best;
326 
327 	best = kvm_find_cpuid_entry(vcpu, 0x80000000, 0);
328 	if (!best || best->eax < 0x80000008)
329 		goto not_found;
330 	best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
331 	if (best)
332 		return best->eax & 0xff;
333 not_found:
334 	return 36;
335 }
336 
337 /*
338  * This "raw" version returns the reserved GPA bits without any adjustments for
339  * encryption technologies that usurp bits.  The raw mask should be used if and
340  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
341  */
342 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
343 {
344 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
345 }
346 
347 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
348                         int nent)
349 {
350 	int r;
351 
352 	__kvm_update_cpuid_runtime(vcpu, e2, nent);
353 
354 	/*
355 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
356 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
357 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
358 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
359 	 * the core vCPU model on the fly. It would've been better to forbid any
360 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
361 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
362 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
363 	 * whether the supplied CPUID data is equal to what's already set.
364 	 */
365 	if (vcpu->arch.last_vmentry_cpu != -1) {
366 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
367 		if (r)
368 			return r;
369 
370 		kvfree(e2);
371 		return 0;
372 	}
373 
374 	r = kvm_check_cpuid(vcpu, e2, nent);
375 	if (r)
376 		return r;
377 
378 	kvfree(vcpu->arch.cpuid_entries);
379 	vcpu->arch.cpuid_entries = e2;
380 	vcpu->arch.cpuid_nent = nent;
381 
382 	kvm_update_kvm_cpuid_base(vcpu);
383 	kvm_vcpu_after_set_cpuid(vcpu);
384 
385 	return 0;
386 }
387 
388 /* when an old userspace process fills a new kernel module */
389 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
390 			     struct kvm_cpuid *cpuid,
391 			     struct kvm_cpuid_entry __user *entries)
392 {
393 	int r, i;
394 	struct kvm_cpuid_entry *e = NULL;
395 	struct kvm_cpuid_entry2 *e2 = NULL;
396 
397 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
398 		return -E2BIG;
399 
400 	if (cpuid->nent) {
401 		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
402 		if (IS_ERR(e))
403 			return PTR_ERR(e);
404 
405 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
406 		if (!e2) {
407 			r = -ENOMEM;
408 			goto out_free_cpuid;
409 		}
410 	}
411 	for (i = 0; i < cpuid->nent; i++) {
412 		e2[i].function = e[i].function;
413 		e2[i].eax = e[i].eax;
414 		e2[i].ebx = e[i].ebx;
415 		e2[i].ecx = e[i].ecx;
416 		e2[i].edx = e[i].edx;
417 		e2[i].index = 0;
418 		e2[i].flags = 0;
419 		e2[i].padding[0] = 0;
420 		e2[i].padding[1] = 0;
421 		e2[i].padding[2] = 0;
422 	}
423 
424 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
425 	if (r)
426 		kvfree(e2);
427 
428 out_free_cpuid:
429 	kvfree(e);
430 
431 	return r;
432 }
433 
434 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
435 			      struct kvm_cpuid2 *cpuid,
436 			      struct kvm_cpuid_entry2 __user *entries)
437 {
438 	struct kvm_cpuid_entry2 *e2 = NULL;
439 	int r;
440 
441 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
442 		return -E2BIG;
443 
444 	if (cpuid->nent) {
445 		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
446 		if (IS_ERR(e2))
447 			return PTR_ERR(e2);
448 	}
449 
450 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
451 	if (r)
452 		kvfree(e2);
453 
454 	return r;
455 }
456 
457 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
458 			      struct kvm_cpuid2 *cpuid,
459 			      struct kvm_cpuid_entry2 __user *entries)
460 {
461 	int r;
462 
463 	r = -E2BIG;
464 	if (cpuid->nent < vcpu->arch.cpuid_nent)
465 		goto out;
466 	r = -EFAULT;
467 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
468 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
469 		goto out;
470 	return 0;
471 
472 out:
473 	cpuid->nent = vcpu->arch.cpuid_nent;
474 	return r;
475 }
476 
477 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
478 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
479 {
480 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
481 	struct kvm_cpuid_entry2 entry;
482 
483 	reverse_cpuid_check(leaf);
484 
485 	cpuid_count(cpuid.function, cpuid.index,
486 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
487 
488 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
489 }
490 
491 static __always_inline
492 void kvm_cpu_cap_init_scattered(enum kvm_only_cpuid_leafs leaf, u32 mask)
493 {
494 	/* Use kvm_cpu_cap_mask for non-scattered leafs. */
495 	BUILD_BUG_ON(leaf < NCAPINTS);
496 
497 	kvm_cpu_caps[leaf] = mask;
498 
499 	__kvm_cpu_cap_mask(leaf);
500 }
501 
502 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
503 {
504 	/* Use kvm_cpu_cap_init_scattered for scattered leafs. */
505 	BUILD_BUG_ON(leaf >= NCAPINTS);
506 
507 	kvm_cpu_caps[leaf] &= mask;
508 
509 	__kvm_cpu_cap_mask(leaf);
510 }
511 
512 void kvm_set_cpu_caps(void)
513 {
514 #ifdef CONFIG_X86_64
515 	unsigned int f_gbpages = F(GBPAGES);
516 	unsigned int f_lm = F(LM);
517 	unsigned int f_xfd = F(XFD);
518 #else
519 	unsigned int f_gbpages = 0;
520 	unsigned int f_lm = 0;
521 	unsigned int f_xfd = 0;
522 #endif
523 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
524 
525 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
526 		     sizeof(boot_cpu_data.x86_capability));
527 
528 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
529 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
530 
531 	kvm_cpu_cap_mask(CPUID_1_ECX,
532 		/*
533 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
534 		 * advertised to guests via CPUID!
535 		 */
536 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
537 		0 /* DS-CPL, VMX, SMX, EST */ |
538 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
539 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
540 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
541 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
542 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
543 		F(F16C) | F(RDRAND)
544 	);
545 	/* KVM emulates x2apic in software irrespective of host support. */
546 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
547 
548 	kvm_cpu_cap_mask(CPUID_1_EDX,
549 		F(FPU) | F(VME) | F(DE) | F(PSE) |
550 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
551 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
552 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
553 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
554 		0 /* Reserved, DS, ACPI */ | F(MMX) |
555 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
556 		0 /* HTT, TM, Reserved, PBE */
557 	);
558 
559 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
560 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
561 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
562 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
563 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
564 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
565 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
566 		F(AVX512VL));
567 
568 	kvm_cpu_cap_mask(CPUID_7_ECX,
569 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
570 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
571 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
572 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
573 		F(SGX_LC) | F(BUS_LOCK_DETECT)
574 	);
575 	/* Set LA57 based on hardware capability. */
576 	if (cpuid_ecx(7) & F(LA57))
577 		kvm_cpu_cap_set(X86_FEATURE_LA57);
578 
579 	/*
580 	 * PKU not yet implemented for shadow paging and requires OSPKE
581 	 * to be set on the host. Clear it if that is not the case
582 	 */
583 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
584 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
585 
586 	kvm_cpu_cap_mask(CPUID_7_EDX,
587 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
588 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
589 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
590 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
591 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16)
592 	);
593 
594 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
595 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
596 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
597 
598 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
599 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
600 	if (boot_cpu_has(X86_FEATURE_STIBP))
601 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
602 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
603 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
604 
605 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
606 		F(AVX_VNNI) | F(AVX512_BF16)
607 	);
608 
609 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
610 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
611 	);
612 
613 	kvm_cpu_cap_init_scattered(CPUID_12_EAX,
614 		SF(SGX1) | SF(SGX2)
615 	);
616 
617 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
618 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
619 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
620 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
621 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
622 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
623 	);
624 
625 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
626 		F(FPU) | F(VME) | F(DE) | F(PSE) |
627 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
628 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
629 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
630 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
631 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
632 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
633 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
634 	);
635 
636 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
637 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
638 
639 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
640 		F(CLZERO) | F(XSAVEERPTR) |
641 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
642 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
643 		__feature_bit(KVM_X86_FEATURE_PSFD)
644 	);
645 
646 	/*
647 	 * AMD has separate bits for each SPEC_CTRL bit.
648 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
649 	 * record that in cpufeatures so use them.
650 	 */
651 	if (boot_cpu_has(X86_FEATURE_IBPB))
652 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
653 	if (boot_cpu_has(X86_FEATURE_IBRS))
654 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
655 	if (boot_cpu_has(X86_FEATURE_STIBP))
656 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
657 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
658 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
659 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
660 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
661 	/*
662 	 * The preference is to use SPEC CTRL MSR instead of the
663 	 * VIRT_SPEC MSR.
664 	 */
665 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
666 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
667 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
668 
669 	/*
670 	 * Hide all SVM features by default, SVM will set the cap bits for
671 	 * features it emulates and/or exposes for L1.
672 	 */
673 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
674 
675 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
676 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
677 		F(SME_COHERENT));
678 
679 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
680 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
681 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
682 		F(PMM) | F(PMM_EN)
683 	);
684 
685 	/*
686 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
687 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
688 	 * should never happen, but the guest will likely crash if RDTSCP or
689 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
690 	 * the past.  For example, the sanity check may fire if this instance of
691 	 * KVM is running as L1 on top of an older, broken KVM.
692 	 */
693 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
694 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
695 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
696 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
697 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
698 	}
699 }
700 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
701 
702 struct kvm_cpuid_array {
703 	struct kvm_cpuid_entry2 *entries;
704 	int maxnent;
705 	int nent;
706 };
707 
708 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
709 					      u32 function, u32 index)
710 {
711 	struct kvm_cpuid_entry2 *entry;
712 
713 	if (array->nent >= array->maxnent)
714 		return NULL;
715 
716 	entry = &array->entries[array->nent++];
717 
718 	memset(entry, 0, sizeof(*entry));
719 	entry->function = function;
720 	entry->index = index;
721 	switch (function & 0xC0000000) {
722 	case 0x40000000:
723 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
724 		return entry;
725 
726 	case 0x80000000:
727 		/*
728 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
729 		 * would result in out-of-bounds calls to do_host_cpuid.
730 		 */
731 		{
732 			static int max_cpuid_80000000;
733 			if (!READ_ONCE(max_cpuid_80000000))
734 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
735 			if (function > READ_ONCE(max_cpuid_80000000))
736 				return entry;
737 		}
738 		break;
739 
740 	default:
741 		break;
742 	}
743 
744 	cpuid_count(entry->function, entry->index,
745 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
746 
747 	switch (function) {
748 	case 4:
749 	case 7:
750 	case 0xb:
751 	case 0xd:
752 	case 0xf:
753 	case 0x10:
754 	case 0x12:
755 	case 0x14:
756 	case 0x17:
757 	case 0x18:
758 	case 0x1d:
759 	case 0x1e:
760 	case 0x1f:
761 	case 0x8000001d:
762 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
763 		break;
764 	}
765 
766 	return entry;
767 }
768 
769 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
770 {
771 	struct kvm_cpuid_entry2 *entry;
772 
773 	if (array->nent >= array->maxnent)
774 		return -E2BIG;
775 
776 	entry = &array->entries[array->nent];
777 	entry->function = func;
778 	entry->index = 0;
779 	entry->flags = 0;
780 
781 	switch (func) {
782 	case 0:
783 		entry->eax = 7;
784 		++array->nent;
785 		break;
786 	case 1:
787 		entry->ecx = F(MOVBE);
788 		++array->nent;
789 		break;
790 	case 7:
791 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
792 		entry->eax = 0;
793 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
794 			entry->ecx = F(RDPID);
795 		++array->nent;
796 		break;
797 	default:
798 		break;
799 	}
800 
801 	return 0;
802 }
803 
804 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
805 {
806 	struct kvm_cpuid_entry2 *entry;
807 	int r, i, max_idx;
808 
809 	/* all calls to cpuid_count() should be made on the same cpu */
810 	get_cpu();
811 
812 	r = -E2BIG;
813 
814 	entry = do_host_cpuid(array, function, 0);
815 	if (!entry)
816 		goto out;
817 
818 	switch (function) {
819 	case 0:
820 		/* Limited to the highest leaf implemented in KVM. */
821 		entry->eax = min(entry->eax, 0x1fU);
822 		break;
823 	case 1:
824 		cpuid_entry_override(entry, CPUID_1_EDX);
825 		cpuid_entry_override(entry, CPUID_1_ECX);
826 		break;
827 	case 2:
828 		/*
829 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
830 		 * CPUID(function=2, index=0) may return different results each
831 		 * time, with the least-significant byte in EAX enumerating the
832 		 * number of times software should do CPUID(2, 0).
833 		 *
834 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
835 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
836 		 * return 01H. Software should ignore this value and not
837 		 * interpret it as an informational descriptor", while AMD's
838 		 * APM states that CPUID(2) is reserved.
839 		 *
840 		 * WARN if a frankenstein CPU that supports virtualization and
841 		 * a stateful CPUID.0x2 is encountered.
842 		 */
843 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
844 		break;
845 	/* functions 4 and 0x8000001d have additional index. */
846 	case 4:
847 	case 0x8000001d:
848 		/*
849 		 * Read entries until the cache type in the previous entry is
850 		 * zero, i.e. indicates an invalid entry.
851 		 */
852 		for (i = 1; entry->eax & 0x1f; ++i) {
853 			entry = do_host_cpuid(array, function, i);
854 			if (!entry)
855 				goto out;
856 		}
857 		break;
858 	case 6: /* Thermal management */
859 		entry->eax = 0x4; /* allow ARAT */
860 		entry->ebx = 0;
861 		entry->ecx = 0;
862 		entry->edx = 0;
863 		break;
864 	/* function 7 has additional index. */
865 	case 7:
866 		entry->eax = min(entry->eax, 1u);
867 		cpuid_entry_override(entry, CPUID_7_0_EBX);
868 		cpuid_entry_override(entry, CPUID_7_ECX);
869 		cpuid_entry_override(entry, CPUID_7_EDX);
870 
871 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
872 		if (entry->eax == 1) {
873 			entry = do_host_cpuid(array, function, 1);
874 			if (!entry)
875 				goto out;
876 
877 			cpuid_entry_override(entry, CPUID_7_1_EAX);
878 			entry->ebx = 0;
879 			entry->ecx = 0;
880 			entry->edx = 0;
881 		}
882 		break;
883 	case 9:
884 		break;
885 	case 0xa: { /* Architectural Performance Monitoring */
886 		struct x86_pmu_capability cap;
887 		union cpuid10_eax eax;
888 		union cpuid10_edx edx;
889 
890 		if (!static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
891 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
892 			break;
893 		}
894 
895 		perf_get_x86_pmu_capability(&cap);
896 
897 		/*
898 		 * The guest architecture pmu is only supported if the architecture
899 		 * pmu exists on the host and the module parameters allow it.
900 		 */
901 		if (!cap.version || !enable_pmu)
902 			memset(&cap, 0, sizeof(cap));
903 
904 		eax.split.version_id = min(cap.version, 2);
905 		eax.split.num_counters = cap.num_counters_gp;
906 		eax.split.bit_width = cap.bit_width_gp;
907 		eax.split.mask_length = cap.events_mask_len;
908 
909 		edx.split.num_counters_fixed =
910 			min(cap.num_counters_fixed, KVM_PMC_MAX_FIXED);
911 		edx.split.bit_width_fixed = cap.bit_width_fixed;
912 		if (cap.version)
913 			edx.split.anythread_deprecated = 1;
914 		edx.split.reserved1 = 0;
915 		edx.split.reserved2 = 0;
916 
917 		entry->eax = eax.full;
918 		entry->ebx = cap.events_mask;
919 		entry->ecx = 0;
920 		entry->edx = edx.full;
921 		break;
922 	}
923 	/*
924 	 * Per Intel's SDM, the 0x1f is a superset of 0xb,
925 	 * thus they can be handled by common code.
926 	 */
927 	case 0x1f:
928 	case 0xb:
929 		/*
930 		 * Populate entries until the level type (ECX[15:8]) of the
931 		 * previous entry is zero.  Note, CPUID EAX.{0x1f,0xb}.0 is
932 		 * the starting entry, filled by the primary do_host_cpuid().
933 		 */
934 		for (i = 1; entry->ecx & 0xff00; ++i) {
935 			entry = do_host_cpuid(array, function, i);
936 			if (!entry)
937 				goto out;
938 		}
939 		break;
940 	case 0xd: {
941 		u64 permitted_xcr0 = supported_xcr0 & xstate_get_guest_group_perm();
942 		u64 permitted_xss = supported_xss;
943 
944 		entry->eax &= permitted_xcr0;
945 		entry->ebx = xstate_required_size(permitted_xcr0, false);
946 		entry->ecx = entry->ebx;
947 		entry->edx &= permitted_xcr0 >> 32;
948 		if (!permitted_xcr0)
949 			break;
950 
951 		entry = do_host_cpuid(array, function, 1);
952 		if (!entry)
953 			goto out;
954 
955 		cpuid_entry_override(entry, CPUID_D_1_EAX);
956 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
957 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
958 							  true);
959 		else {
960 			WARN_ON_ONCE(permitted_xss != 0);
961 			entry->ebx = 0;
962 		}
963 		entry->ecx &= permitted_xss;
964 		entry->edx &= permitted_xss >> 32;
965 
966 		for (i = 2; i < 64; ++i) {
967 			bool s_state;
968 			if (permitted_xcr0 & BIT_ULL(i))
969 				s_state = false;
970 			else if (permitted_xss & BIT_ULL(i))
971 				s_state = true;
972 			else
973 				continue;
974 
975 			entry = do_host_cpuid(array, function, i);
976 			if (!entry)
977 				goto out;
978 
979 			/*
980 			 * The supported check above should have filtered out
981 			 * invalid sub-leafs.  Only valid sub-leafs should
982 			 * reach this point, and they should have a non-zero
983 			 * save state size.  Furthermore, check whether the
984 			 * processor agrees with permitted_xcr0/permitted_xss
985 			 * on whether this is an XCR0- or IA32_XSS-managed area.
986 			 */
987 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
988 				--array->nent;
989 				continue;
990 			}
991 
992 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
993 				entry->ecx &= ~BIT_ULL(2);
994 			entry->edx = 0;
995 		}
996 		break;
997 	}
998 	case 0x12:
999 		/* Intel SGX */
1000 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1001 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1002 			break;
1003 		}
1004 
1005 		/*
1006 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1007 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1008 		 * are restricted by kernel and KVM capabilities (like most
1009 		 * feature flags), while enclave size is unrestricted.
1010 		 */
1011 		cpuid_entry_override(entry, CPUID_12_EAX);
1012 		entry->ebx &= SGX_MISC_EXINFO;
1013 
1014 		entry = do_host_cpuid(array, function, 1);
1015 		if (!entry)
1016 			goto out;
1017 
1018 		/*
1019 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1020 		 * feature flags.  Advertise all supported flags, including
1021 		 * privileged attributes that require explicit opt-in from
1022 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1023 		 * expected to derive it from supported XCR0.
1024 		 */
1025 		entry->eax &= SGX_ATTR_DEBUG | SGX_ATTR_MODE64BIT |
1026 			      SGX_ATTR_PROVISIONKEY | SGX_ATTR_EINITTOKENKEY |
1027 			      SGX_ATTR_KSS;
1028 		entry->ebx &= 0;
1029 		break;
1030 	/* Intel PT */
1031 	case 0x14:
1032 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1033 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1034 			break;
1035 		}
1036 
1037 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1038 			if (!do_host_cpuid(array, function, i))
1039 				goto out;
1040 		}
1041 		break;
1042 	/* Intel AMX TILE */
1043 	case 0x1d:
1044 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1045 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1046 			break;
1047 		}
1048 
1049 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1050 			if (!do_host_cpuid(array, function, i))
1051 				goto out;
1052 		}
1053 		break;
1054 	case 0x1e: /* TMUL information */
1055 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1056 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1057 			break;
1058 		}
1059 		break;
1060 	case KVM_CPUID_SIGNATURE: {
1061 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1062 		entry->eax = KVM_CPUID_FEATURES;
1063 		entry->ebx = sigptr[0];
1064 		entry->ecx = sigptr[1];
1065 		entry->edx = sigptr[2];
1066 		break;
1067 	}
1068 	case KVM_CPUID_FEATURES:
1069 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1070 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1071 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1072 			     (1 << KVM_FEATURE_ASYNC_PF) |
1073 			     (1 << KVM_FEATURE_PV_EOI) |
1074 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1075 			     (1 << KVM_FEATURE_PV_UNHALT) |
1076 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1077 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1078 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1079 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1080 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1081 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1082 
1083 		if (sched_info_on())
1084 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1085 
1086 		entry->ebx = 0;
1087 		entry->ecx = 0;
1088 		entry->edx = 0;
1089 		break;
1090 	case 0x80000000:
1091 		entry->eax = min(entry->eax, 0x80000021);
1092 		/*
1093 		 * Serializing LFENCE is reported in a multitude of ways, and
1094 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1095 		 * userspace by providing the CPUID leaf ourselves.
1096 		 *
1097 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1098 		 * QEMU thinks that it can query the host blindly for that
1099 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1100 		 * above.  The processor merrily returns values from the
1101 		 * highest Intel leaf which QEMU tries to use as the guest's
1102 		 * 0x8000001d.  Even worse, this can result in an infinite
1103 		 * loop if said highest leaf has no subleaves indexed by ECX.
1104 		 */
1105 		if (entry->eax >= 0x8000001d &&
1106 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1107 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1108 			entry->eax = max(entry->eax, 0x80000021);
1109 		break;
1110 	case 0x80000001:
1111 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1112 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1113 		break;
1114 	case 0x80000006:
1115 		/* L2 cache and TLB: pass through host info. */
1116 		break;
1117 	case 0x80000007: /* Advanced power management */
1118 		/* invariant TSC is CPUID.80000007H:EDX[8] */
1119 		entry->edx &= (1 << 8);
1120 		/* mask against host */
1121 		entry->edx &= boot_cpu_data.x86_power;
1122 		entry->eax = entry->ebx = entry->ecx = 0;
1123 		break;
1124 	case 0x80000008: {
1125 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1126 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1127 		unsigned phys_as = entry->eax & 0xff;
1128 
1129 		/*
1130 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1131 		 * the guest operates in the same PA space as the host, i.e.
1132 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1133 		 * paging, too.
1134 		 *
1135 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1136 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1137 		 * the HPAs do not affect GPAs.
1138 		 */
1139 		if (!tdp_enabled)
1140 			g_phys_as = boot_cpu_data.x86_phys_bits;
1141 		else if (!g_phys_as)
1142 			g_phys_as = phys_as;
1143 
1144 		entry->eax = g_phys_as | (virt_as << 8);
1145 		entry->edx = 0;
1146 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1147 		break;
1148 	}
1149 	case 0x8000000A:
1150 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1151 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1152 			break;
1153 		}
1154 		entry->eax = 1; /* SVM revision 1 */
1155 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1156 				   ASID emulation to nested SVM */
1157 		entry->ecx = 0; /* Reserved */
1158 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1159 		break;
1160 	case 0x80000019:
1161 		entry->ecx = entry->edx = 0;
1162 		break;
1163 	case 0x8000001a:
1164 	case 0x8000001e:
1165 		break;
1166 	case 0x8000001F:
1167 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1168 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1169 		} else {
1170 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1171 
1172 			/*
1173 			 * Enumerate '0' for "PA bits reduction", the adjusted
1174 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1175 			 */
1176 			entry->ebx &= ~GENMASK(11, 6);
1177 		}
1178 		break;
1179 	case 0x80000020:
1180 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1181 		break;
1182 	case 0x80000021:
1183 		entry->ebx = entry->ecx = entry->edx = 0;
1184 		/*
1185 		 * Pass down these bits:
1186 		 *    EAX      0      NNDBP, Processor ignores nested data breakpoints
1187 		 *    EAX      2      LAS, LFENCE always serializing
1188 		 *    EAX      6      NSCB, Null selector clear base
1189 		 *
1190 		 * Other defined bits are for MSRs that KVM does not expose:
1191 		 *   EAX      3      SPCL, SMM page configuration lock
1192 		 *   EAX      13     PCMSR, Prefetch control MSR
1193 		 */
1194 		entry->eax &= BIT(0) | BIT(2) | BIT(6);
1195 		if (static_cpu_has(X86_FEATURE_LFENCE_RDTSC))
1196 			entry->eax |= BIT(2);
1197 		if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
1198 			entry->eax |= BIT(6);
1199 		break;
1200 	/*Add support for Centaur's CPUID instruction*/
1201 	case 0xC0000000:
1202 		/*Just support up to 0xC0000004 now*/
1203 		entry->eax = min(entry->eax, 0xC0000004);
1204 		break;
1205 	case 0xC0000001:
1206 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1207 		break;
1208 	case 3: /* Processor serial number */
1209 	case 5: /* MONITOR/MWAIT */
1210 	case 0xC0000002:
1211 	case 0xC0000003:
1212 	case 0xC0000004:
1213 	default:
1214 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1215 		break;
1216 	}
1217 
1218 	r = 0;
1219 
1220 out:
1221 	put_cpu();
1222 
1223 	return r;
1224 }
1225 
1226 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1227 			 unsigned int type)
1228 {
1229 	if (type == KVM_GET_EMULATED_CPUID)
1230 		return __do_cpuid_func_emulated(array, func);
1231 
1232 	return __do_cpuid_func(array, func);
1233 }
1234 
1235 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1236 
1237 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1238 			  unsigned int type)
1239 {
1240 	u32 limit;
1241 	int r;
1242 
1243 	if (func == CENTAUR_CPUID_SIGNATURE &&
1244 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1245 		return 0;
1246 
1247 	r = do_cpuid_func(array, func, type);
1248 	if (r)
1249 		return r;
1250 
1251 	limit = array->entries[array->nent - 1].eax;
1252 	for (func = func + 1; func <= limit; ++func) {
1253 		r = do_cpuid_func(array, func, type);
1254 		if (r)
1255 			break;
1256 	}
1257 
1258 	return r;
1259 }
1260 
1261 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1262 				 __u32 num_entries, unsigned int ioctl_type)
1263 {
1264 	int i;
1265 	__u32 pad[3];
1266 
1267 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1268 		return false;
1269 
1270 	/*
1271 	 * We want to make sure that ->padding is being passed clean from
1272 	 * userspace in case we want to use it for something in the future.
1273 	 *
1274 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1275 	 * have to give ourselves satisfied only with the emulated side. /me
1276 	 * sheds a tear.
1277 	 */
1278 	for (i = 0; i < num_entries; i++) {
1279 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1280 			return true;
1281 
1282 		if (pad[0] || pad[1] || pad[2])
1283 			return true;
1284 	}
1285 	return false;
1286 }
1287 
1288 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1289 			    struct kvm_cpuid_entry2 __user *entries,
1290 			    unsigned int type)
1291 {
1292 	static const u32 funcs[] = {
1293 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1294 	};
1295 
1296 	struct kvm_cpuid_array array = {
1297 		.nent = 0,
1298 	};
1299 	int r, i;
1300 
1301 	if (cpuid->nent < 1)
1302 		return -E2BIG;
1303 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1304 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1305 
1306 	if (sanity_check_entries(entries, cpuid->nent, type))
1307 		return -EINVAL;
1308 
1309 	array.entries = kvcalloc(sizeof(struct kvm_cpuid_entry2), cpuid->nent, GFP_KERNEL);
1310 	if (!array.entries)
1311 		return -ENOMEM;
1312 
1313 	array.maxnent = cpuid->nent;
1314 
1315 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1316 		r = get_cpuid_func(&array, funcs[i], type);
1317 		if (r)
1318 			goto out_free;
1319 	}
1320 	cpuid->nent = array.nent;
1321 
1322 	if (copy_to_user(entries, array.entries,
1323 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1324 		r = -EFAULT;
1325 
1326 out_free:
1327 	kvfree(array.entries);
1328 	return r;
1329 }
1330 
1331 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1332 					      u32 function, u32 index)
1333 {
1334 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1335 				 function, index);
1336 }
1337 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1338 
1339 /*
1340  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1341  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1342  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1343  * range.  Centaur/VIA follows Intel semantics.
1344  *
1345  * A leaf is considered out-of-range if its function is higher than the maximum
1346  * supported leaf of its associated class or if its associated class does not
1347  * exist.
1348  *
1349  * There are three primary classes to be considered, with their respective
1350  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1351  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1352  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1353  *
1354  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1355  *  - Hypervisor: 0x40000000 - 0x4fffffff
1356  *  - Extended:   0x80000000 - 0xbfffffff
1357  *  - Centaur:    0xc0000000 - 0xcfffffff
1358  *
1359  * The Hypervisor class is further subdivided into sub-classes that each act as
1360  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1361  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1362  * CPUID sub-classes are:
1363  *
1364  *  - HyperV:     0x40000000 - 0x400000ff
1365  *  - KVM:        0x40000100 - 0x400001ff
1366  */
1367 static struct kvm_cpuid_entry2 *
1368 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1369 {
1370 	struct kvm_cpuid_entry2 *basic, *class;
1371 	u32 function = *fn_ptr;
1372 
1373 	basic = kvm_find_cpuid_entry(vcpu, 0, 0);
1374 	if (!basic)
1375 		return NULL;
1376 
1377 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1378 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1379 		return NULL;
1380 
1381 	if (function >= 0x40000000 && function <= 0x4fffffff)
1382 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00, 0);
1383 	else if (function >= 0xc0000000)
1384 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000, 0);
1385 	else
1386 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
1387 
1388 	if (class && function <= class->eax)
1389 		return NULL;
1390 
1391 	/*
1392 	 * Leaf specific adjustments are also applied when redirecting to the
1393 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1394 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1395 	 * needs to be pulled from CPUID.0xb.1.
1396 	 */
1397 	*fn_ptr = basic->eax;
1398 
1399 	/*
1400 	 * The class does not exist or the requested function is out of range;
1401 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1402 	 * the original requested leaf is observed!
1403 	 */
1404 	return kvm_find_cpuid_entry(vcpu, basic->eax, index);
1405 }
1406 
1407 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1408 	       u32 *ecx, u32 *edx, bool exact_only)
1409 {
1410 	u32 orig_function = *eax, function = *eax, index = *ecx;
1411 	struct kvm_cpuid_entry2 *entry;
1412 	bool exact, used_max_basic = false;
1413 
1414 	entry = kvm_find_cpuid_entry(vcpu, function, index);
1415 	exact = !!entry;
1416 
1417 	if (!entry && !exact_only) {
1418 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1419 		used_max_basic = !!entry;
1420 	}
1421 
1422 	if (entry) {
1423 		*eax = entry->eax;
1424 		*ebx = entry->ebx;
1425 		*ecx = entry->ecx;
1426 		*edx = entry->edx;
1427 		if (function == 7 && index == 0) {
1428 			u64 data;
1429 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1430 			    (data & TSX_CTRL_CPUID_CLEAR))
1431 				*ebx &= ~(F(RTM) | F(HLE));
1432 		}
1433 	} else {
1434 		*eax = *ebx = *ecx = *edx = 0;
1435 		/*
1436 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1437 		 * and EDX is always the x2APIC ID, even for undefined
1438 		 * subleaves. Index 1 will exist iff the leaf is
1439 		 * implemented, so we pass through CL iff leaf 1
1440 		 * exists. EDX can be copied from any existing index.
1441 		 */
1442 		if (function == 0xb || function == 0x1f) {
1443 			entry = kvm_find_cpuid_entry(vcpu, function, 1);
1444 			if (entry) {
1445 				*ecx = index & 0xff;
1446 				*edx = entry->edx;
1447 			}
1448 		}
1449 	}
1450 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1451 			used_max_basic);
1452 	return exact;
1453 }
1454 EXPORT_SYMBOL_GPL(kvm_cpuid);
1455 
1456 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1457 {
1458 	u32 eax, ebx, ecx, edx;
1459 
1460 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1461 		return 1;
1462 
1463 	eax = kvm_rax_read(vcpu);
1464 	ecx = kvm_rcx_read(vcpu);
1465 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1466 	kvm_rax_write(vcpu, eax);
1467 	kvm_rbx_write(vcpu, ebx);
1468 	kvm_rcx_write(vcpu, ecx);
1469 	kvm_rdx_write(vcpu, edx);
1470 	return kvm_skip_emulated_instruction(vcpu);
1471 }
1472 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1473