xref: /linux/arch/x86/kvm/cpuid.c (revision e9adcfecf572fcfaa9f8525904cf49c709974f73)
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 <asm/cpuid.h>
23 #include "cpuid.h"
24 #include "lapic.h"
25 #include "mmu.h"
26 #include "trace.h"
27 #include "pmu.h"
28 
29 /*
30  * Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be
31  * aligned to sizeof(unsigned long) because it's not accessed via bitops.
32  */
33 u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
34 EXPORT_SYMBOL_GPL(kvm_cpu_caps);
35 
36 u32 xstate_required_size(u64 xstate_bv, bool compacted)
37 {
38 	int feature_bit = 0;
39 	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
40 
41 	xstate_bv &= XFEATURE_MASK_EXTEND;
42 	while (xstate_bv) {
43 		if (xstate_bv & 0x1) {
44 		        u32 eax, ebx, ecx, edx, offset;
45 		        cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
46 			/* ECX[1]: 64B alignment in compacted form */
47 			if (compacted)
48 				offset = (ecx & 0x2) ? ALIGN(ret, 64) : ret;
49 			else
50 				offset = ebx;
51 			ret = max(ret, offset + eax);
52 		}
53 
54 		xstate_bv >>= 1;
55 		feature_bit++;
56 	}
57 
58 	return ret;
59 }
60 
61 /*
62  * This one is tied to SSB in the user API, and not
63  * visible in /proc/cpuinfo.
64  */
65 #define KVM_X86_FEATURE_AMD_PSFD	(13*32+28) /* Predictive Store Forwarding Disable */
66 
67 #define F feature_bit
68 
69 /* Scattered Flag - For features that are scattered by cpufeatures.h. */
70 #define SF(name)						\
71 ({								\
72 	BUILD_BUG_ON(X86_FEATURE_##name >= MAX_CPU_FEATURES);	\
73 	(boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0);	\
74 })
75 
76 /*
77  * Magic value used by KVM when querying userspace-provided CPUID entries and
78  * doesn't care about the CPIUD index because the index of the function in
79  * question is not significant.  Note, this magic value must have at least one
80  * bit set in bits[63:32] and must be consumed as a u64 by cpuid_entry2_find()
81  * to avoid false positives when processing guest CPUID input.
82  */
83 #define KVM_CPUID_INDEX_NOT_SIGNIFICANT -1ull
84 
85 static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
86 	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u64 index)
87 {
88 	struct kvm_cpuid_entry2 *e;
89 	int i;
90 
91 	for (i = 0; i < nent; i++) {
92 		e = &entries[i];
93 
94 		if (e->function != function)
95 			continue;
96 
97 		/*
98 		 * If the index isn't significant, use the first entry with a
99 		 * matching function.  It's userspace's responsibilty to not
100 		 * provide "duplicate" entries in all cases.
101 		 */
102 		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
103 			return e;
104 
105 
106 		/*
107 		 * Similarly, use the first matching entry if KVM is doing a
108 		 * lookup (as opposed to emulating CPUID) for a function that's
109 		 * architecturally defined as not having a significant index.
110 		 */
111 		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
112 			/*
113 			 * Direct lookups from KVM should not diverge from what
114 			 * KVM defines internally (the architectural behavior).
115 			 */
116 			WARN_ON_ONCE(cpuid_function_is_indexed(function));
117 			return e;
118 		}
119 	}
120 
121 	return NULL;
122 }
123 
124 static int kvm_check_cpuid(struct kvm_vcpu *vcpu,
125 			   struct kvm_cpuid_entry2 *entries,
126 			   int nent)
127 {
128 	struct kvm_cpuid_entry2 *best;
129 	u64 xfeatures;
130 
131 	/*
132 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
133 	 * canonical address checks; exit if it is ever changed.
134 	 */
135 	best = cpuid_entry2_find(entries, nent, 0x80000008,
136 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
137 	if (best) {
138 		int vaddr_bits = (best->eax & 0xff00) >> 8;
139 
140 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
141 			return -EINVAL;
142 	}
143 
144 	/*
145 	 * Exposing dynamic xfeatures to the guest requires additional
146 	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
147 	 */
148 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
149 	if (!best)
150 		return 0;
151 
152 	xfeatures = best->eax | ((u64)best->edx << 32);
153 	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
154 	if (!xfeatures)
155 		return 0;
156 
157 	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
158 }
159 
160 /* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
161 static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
162 				 int nent)
163 {
164 	struct kvm_cpuid_entry2 *orig;
165 	int i;
166 
167 	if (nent != vcpu->arch.cpuid_nent)
168 		return -EINVAL;
169 
170 	for (i = 0; i < nent; i++) {
171 		orig = &vcpu->arch.cpuid_entries[i];
172 		if (e2[i].function != orig->function ||
173 		    e2[i].index != orig->index ||
174 		    e2[i].flags != orig->flags ||
175 		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
176 		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
177 			return -EINVAL;
178 	}
179 
180 	return 0;
181 }
182 
183 static void kvm_update_kvm_cpuid_base(struct kvm_vcpu *vcpu)
184 {
185 	u32 function;
186 	struct kvm_cpuid_entry2 *entry;
187 
188 	vcpu->arch.kvm_cpuid_base = 0;
189 
190 	for_each_possible_hypervisor_cpuid_base(function) {
191 		entry = kvm_find_cpuid_entry(vcpu, function);
192 
193 		if (entry) {
194 			u32 signature[3];
195 
196 			signature[0] = entry->ebx;
197 			signature[1] = entry->ecx;
198 			signature[2] = entry->edx;
199 
200 			BUILD_BUG_ON(sizeof(signature) > sizeof(KVM_SIGNATURE));
201 			if (!memcmp(signature, KVM_SIGNATURE, sizeof(signature))) {
202 				vcpu->arch.kvm_cpuid_base = function;
203 				break;
204 			}
205 		}
206 	}
207 }
208 
209 static struct kvm_cpuid_entry2 *__kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu,
210 					      struct kvm_cpuid_entry2 *entries, int nent)
211 {
212 	u32 base = vcpu->arch.kvm_cpuid_base;
213 
214 	if (!base)
215 		return NULL;
216 
217 	return cpuid_entry2_find(entries, nent, base | KVM_CPUID_FEATURES,
218 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
219 }
220 
221 static struct kvm_cpuid_entry2 *kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu)
222 {
223 	return __kvm_find_kvm_cpuid_features(vcpu, vcpu->arch.cpuid_entries,
224 					     vcpu->arch.cpuid_nent);
225 }
226 
227 void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
228 {
229 	struct kvm_cpuid_entry2 *best = kvm_find_kvm_cpuid_features(vcpu);
230 
231 	/*
232 	 * save the feature bitmap to avoid cpuid lookup for every PV
233 	 * operation
234 	 */
235 	if (best)
236 		vcpu->arch.pv_cpuid.features = best->eax;
237 }
238 
239 /*
240  * Calculate guest's supported XCR0 taking into account guest CPUID data and
241  * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
242  */
243 static u64 cpuid_get_supported_xcr0(struct kvm_cpuid_entry2 *entries, int nent)
244 {
245 	struct kvm_cpuid_entry2 *best;
246 
247 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
248 	if (!best)
249 		return 0;
250 
251 	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
252 }
253 
254 static void __kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *entries,
255 				       int nent)
256 {
257 	struct kvm_cpuid_entry2 *best;
258 	u64 guest_supported_xcr0 = cpuid_get_supported_xcr0(entries, nent);
259 
260 	best = cpuid_entry2_find(entries, nent, 1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
261 	if (best) {
262 		/* Update OSXSAVE bit */
263 		if (boot_cpu_has(X86_FEATURE_XSAVE))
264 			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
265 				   kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE));
266 
267 		cpuid_entry_change(best, X86_FEATURE_APIC,
268 			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
269 	}
270 
271 	best = cpuid_entry2_find(entries, nent, 7, 0);
272 	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
273 		cpuid_entry_change(best, X86_FEATURE_OSPKE,
274 				   kvm_read_cr4_bits(vcpu, X86_CR4_PKE));
275 
276 	best = cpuid_entry2_find(entries, nent, 0xD, 0);
277 	if (best)
278 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
279 
280 	best = cpuid_entry2_find(entries, nent, 0xD, 1);
281 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
282 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
283 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
284 
285 	best = __kvm_find_kvm_cpuid_features(vcpu, entries, nent);
286 	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
287 		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
288 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
289 
290 	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
291 		best = cpuid_entry2_find(entries, nent, 0x1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
292 		if (best)
293 			cpuid_entry_change(best, X86_FEATURE_MWAIT,
294 					   vcpu->arch.ia32_misc_enable_msr &
295 					   MSR_IA32_MISC_ENABLE_MWAIT);
296 	}
297 
298 	/*
299 	 * Bits 127:0 of the allowed SECS.ATTRIBUTES (CPUID.0x12.0x1) enumerate
300 	 * the supported XSAVE Feature Request Mask (XFRM), i.e. the enclave's
301 	 * requested XCR0 value.  The enclave's XFRM must be a subset of XCRO
302 	 * at the time of EENTER, thus adjust the allowed XFRM by the guest's
303 	 * supported XCR0.  Similar to XCR0 handling, FP and SSE are forced to
304 	 * '1' even on CPUs that don't support XSAVE.
305 	 */
306 	best = cpuid_entry2_find(entries, nent, 0x12, 0x1);
307 	if (best) {
308 		best->ecx &= guest_supported_xcr0 & 0xffffffff;
309 		best->edx &= guest_supported_xcr0 >> 32;
310 		best->ecx |= XFEATURE_MASK_FPSSE;
311 	}
312 }
313 
314 void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
315 {
316 	__kvm_update_cpuid_runtime(vcpu, vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
317 }
318 EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
319 
320 static bool kvm_cpuid_has_hyperv(struct kvm_cpuid_entry2 *entries, int nent)
321 {
322 	struct kvm_cpuid_entry2 *entry;
323 
324 	entry = cpuid_entry2_find(entries, nent, HYPERV_CPUID_INTERFACE,
325 				  KVM_CPUID_INDEX_NOT_SIGNIFICANT);
326 	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
327 }
328 
329 static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
330 {
331 	struct kvm_lapic *apic = vcpu->arch.apic;
332 	struct kvm_cpuid_entry2 *best;
333 
334 	best = kvm_find_cpuid_entry(vcpu, 1);
335 	if (best && apic) {
336 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
337 			apic->lapic_timer.timer_mode_mask = 3 << 17;
338 		else
339 			apic->lapic_timer.timer_mode_mask = 1 << 17;
340 
341 		kvm_apic_set_version(vcpu);
342 	}
343 
344 	vcpu->arch.guest_supported_xcr0 =
345 		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
346 
347 	/*
348 	 * FP+SSE can always be saved/restored via KVM_{G,S}ET_XSAVE, even if
349 	 * XSAVE/XCRO are not exposed to the guest, and even if XSAVE isn't
350 	 * supported by the host.
351 	 */
352 	vcpu->arch.guest_fpu.fpstate->user_xfeatures = vcpu->arch.guest_supported_xcr0 |
353 						       XFEATURE_MASK_FPSSE;
354 
355 	kvm_update_pv_runtime(vcpu);
356 
357 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
358 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
359 
360 	kvm_pmu_refresh(vcpu);
361 	vcpu->arch.cr4_guest_rsvd_bits =
362 	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
363 
364 	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu->arch.cpuid_entries,
365 						    vcpu->arch.cpuid_nent));
366 
367 	/* Invoke the vendor callback only after the above state is updated. */
368 	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
369 
370 	/*
371 	 * Except for the MMU, which needs to do its thing any vendor specific
372 	 * adjustments to the reserved GPA bits.
373 	 */
374 	kvm_mmu_after_set_cpuid(vcpu);
375 }
376 
377 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
378 {
379 	struct kvm_cpuid_entry2 *best;
380 
381 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
382 	if (!best || best->eax < 0x80000008)
383 		goto not_found;
384 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
385 	if (best)
386 		return best->eax & 0xff;
387 not_found:
388 	return 36;
389 }
390 
391 /*
392  * This "raw" version returns the reserved GPA bits without any adjustments for
393  * encryption technologies that usurp bits.  The raw mask should be used if and
394  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
395  */
396 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
397 {
398 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
399 }
400 
401 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
402                         int nent)
403 {
404 	int r;
405 
406 	__kvm_update_cpuid_runtime(vcpu, e2, nent);
407 
408 	/*
409 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
410 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
411 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
412 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
413 	 * the core vCPU model on the fly. It would've been better to forbid any
414 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
415 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
416 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
417 	 * whether the supplied CPUID data is equal to what's already set.
418 	 */
419 	if (vcpu->arch.last_vmentry_cpu != -1) {
420 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
421 		if (r)
422 			return r;
423 
424 		kvfree(e2);
425 		return 0;
426 	}
427 
428 	if (kvm_cpuid_has_hyperv(e2, nent)) {
429 		r = kvm_hv_vcpu_init(vcpu);
430 		if (r)
431 			return r;
432 	}
433 
434 	r = kvm_check_cpuid(vcpu, e2, nent);
435 	if (r)
436 		return r;
437 
438 	kvfree(vcpu->arch.cpuid_entries);
439 	vcpu->arch.cpuid_entries = e2;
440 	vcpu->arch.cpuid_nent = nent;
441 
442 	kvm_update_kvm_cpuid_base(vcpu);
443 	kvm_vcpu_after_set_cpuid(vcpu);
444 
445 	return 0;
446 }
447 
448 /* when an old userspace process fills a new kernel module */
449 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
450 			     struct kvm_cpuid *cpuid,
451 			     struct kvm_cpuid_entry __user *entries)
452 {
453 	int r, i;
454 	struct kvm_cpuid_entry *e = NULL;
455 	struct kvm_cpuid_entry2 *e2 = NULL;
456 
457 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
458 		return -E2BIG;
459 
460 	if (cpuid->nent) {
461 		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
462 		if (IS_ERR(e))
463 			return PTR_ERR(e);
464 
465 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
466 		if (!e2) {
467 			r = -ENOMEM;
468 			goto out_free_cpuid;
469 		}
470 	}
471 	for (i = 0; i < cpuid->nent; i++) {
472 		e2[i].function = e[i].function;
473 		e2[i].eax = e[i].eax;
474 		e2[i].ebx = e[i].ebx;
475 		e2[i].ecx = e[i].ecx;
476 		e2[i].edx = e[i].edx;
477 		e2[i].index = 0;
478 		e2[i].flags = 0;
479 		e2[i].padding[0] = 0;
480 		e2[i].padding[1] = 0;
481 		e2[i].padding[2] = 0;
482 	}
483 
484 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
485 	if (r)
486 		kvfree(e2);
487 
488 out_free_cpuid:
489 	kvfree(e);
490 
491 	return r;
492 }
493 
494 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
495 			      struct kvm_cpuid2 *cpuid,
496 			      struct kvm_cpuid_entry2 __user *entries)
497 {
498 	struct kvm_cpuid_entry2 *e2 = NULL;
499 	int r;
500 
501 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
502 		return -E2BIG;
503 
504 	if (cpuid->nent) {
505 		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
506 		if (IS_ERR(e2))
507 			return PTR_ERR(e2);
508 	}
509 
510 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
511 	if (r)
512 		kvfree(e2);
513 
514 	return r;
515 }
516 
517 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
518 			      struct kvm_cpuid2 *cpuid,
519 			      struct kvm_cpuid_entry2 __user *entries)
520 {
521 	int r;
522 
523 	r = -E2BIG;
524 	if (cpuid->nent < vcpu->arch.cpuid_nent)
525 		goto out;
526 	r = -EFAULT;
527 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
528 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
529 		goto out;
530 	return 0;
531 
532 out:
533 	cpuid->nent = vcpu->arch.cpuid_nent;
534 	return r;
535 }
536 
537 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
538 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
539 {
540 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
541 	struct kvm_cpuid_entry2 entry;
542 
543 	reverse_cpuid_check(leaf);
544 
545 	cpuid_count(cpuid.function, cpuid.index,
546 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
547 
548 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
549 }
550 
551 static __always_inline
552 void kvm_cpu_cap_init_kvm_defined(enum kvm_only_cpuid_leafs leaf, u32 mask)
553 {
554 	/* Use kvm_cpu_cap_mask for leafs that aren't KVM-only. */
555 	BUILD_BUG_ON(leaf < NCAPINTS);
556 
557 	kvm_cpu_caps[leaf] = mask;
558 
559 	__kvm_cpu_cap_mask(leaf);
560 }
561 
562 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
563 {
564 	/* Use kvm_cpu_cap_init_kvm_defined for KVM-only leafs. */
565 	BUILD_BUG_ON(leaf >= NCAPINTS);
566 
567 	kvm_cpu_caps[leaf] &= mask;
568 
569 	__kvm_cpu_cap_mask(leaf);
570 }
571 
572 void kvm_set_cpu_caps(void)
573 {
574 #ifdef CONFIG_X86_64
575 	unsigned int f_gbpages = F(GBPAGES);
576 	unsigned int f_lm = F(LM);
577 	unsigned int f_xfd = F(XFD);
578 #else
579 	unsigned int f_gbpages = 0;
580 	unsigned int f_lm = 0;
581 	unsigned int f_xfd = 0;
582 #endif
583 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
584 
585 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
586 		     sizeof(boot_cpu_data.x86_capability));
587 
588 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
589 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
590 
591 	kvm_cpu_cap_mask(CPUID_1_ECX,
592 		/*
593 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
594 		 * advertised to guests via CPUID!
595 		 */
596 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
597 		0 /* DS-CPL, VMX, SMX, EST */ |
598 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
599 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
600 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
601 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
602 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
603 		F(F16C) | F(RDRAND)
604 	);
605 	/* KVM emulates x2apic in software irrespective of host support. */
606 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
607 
608 	kvm_cpu_cap_mask(CPUID_1_EDX,
609 		F(FPU) | F(VME) | F(DE) | F(PSE) |
610 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
611 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
612 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
613 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
614 		0 /* Reserved, DS, ACPI */ | F(MMX) |
615 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
616 		0 /* HTT, TM, Reserved, PBE */
617 	);
618 
619 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
620 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
621 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
622 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
623 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
624 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
625 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
626 		F(AVX512VL));
627 
628 	kvm_cpu_cap_mask(CPUID_7_ECX,
629 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
630 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
631 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
632 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
633 		F(SGX_LC) | F(BUS_LOCK_DETECT)
634 	);
635 	/* Set LA57 based on hardware capability. */
636 	if (cpuid_ecx(7) & F(LA57))
637 		kvm_cpu_cap_set(X86_FEATURE_LA57);
638 
639 	/*
640 	 * PKU not yet implemented for shadow paging and requires OSPKE
641 	 * to be set on the host. Clear it if that is not the case
642 	 */
643 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
644 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
645 
646 	kvm_cpu_cap_mask(CPUID_7_EDX,
647 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
648 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
649 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
650 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
651 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16)
652 	);
653 
654 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
655 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
656 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
657 
658 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
659 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
660 	if (boot_cpu_has(X86_FEATURE_STIBP))
661 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
662 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
663 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
664 
665 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
666 		F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) | F(AMX_FP16) |
667 		F(AVX_IFMA)
668 	);
669 
670 	kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
671 		F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(PREFETCHITI)
672 	);
673 
674 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
675 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
676 	);
677 
678 	kvm_cpu_cap_init_kvm_defined(CPUID_12_EAX,
679 		SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
680 	);
681 
682 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
683 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
684 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
685 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
686 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
687 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
688 	);
689 
690 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
691 		F(FPU) | F(VME) | F(DE) | F(PSE) |
692 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
693 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
694 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
695 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
696 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
697 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
698 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
699 	);
700 
701 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
702 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
703 
704 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
705 		F(CLZERO) | F(XSAVEERPTR) |
706 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
707 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
708 		__feature_bit(KVM_X86_FEATURE_AMD_PSFD)
709 	);
710 
711 	/*
712 	 * AMD has separate bits for each SPEC_CTRL bit.
713 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
714 	 * record that in cpufeatures so use them.
715 	 */
716 	if (boot_cpu_has(X86_FEATURE_IBPB))
717 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
718 	if (boot_cpu_has(X86_FEATURE_IBRS))
719 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
720 	if (boot_cpu_has(X86_FEATURE_STIBP))
721 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
722 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
723 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
724 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
725 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
726 	/*
727 	 * The preference is to use SPEC CTRL MSR instead of the
728 	 * VIRT_SPEC MSR.
729 	 */
730 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
731 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
732 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
733 
734 	/*
735 	 * Hide all SVM features by default, SVM will set the cap bits for
736 	 * features it emulates and/or exposes for L1.
737 	 */
738 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
739 
740 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
741 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
742 		F(SME_COHERENT));
743 
744 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
745 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
746 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
747 		F(PMM) | F(PMM_EN)
748 	);
749 
750 	/*
751 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
752 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
753 	 * should never happen, but the guest will likely crash if RDTSCP or
754 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
755 	 * the past.  For example, the sanity check may fire if this instance of
756 	 * KVM is running as L1 on top of an older, broken KVM.
757 	 */
758 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
759 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
760 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
761 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
762 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
763 	}
764 }
765 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
766 
767 struct kvm_cpuid_array {
768 	struct kvm_cpuid_entry2 *entries;
769 	int maxnent;
770 	int nent;
771 };
772 
773 static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
774 {
775 	if (array->nent >= array->maxnent)
776 		return NULL;
777 
778 	return &array->entries[array->nent++];
779 }
780 
781 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
782 					      u32 function, u32 index)
783 {
784 	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
785 
786 	if (!entry)
787 		return NULL;
788 
789 	memset(entry, 0, sizeof(*entry));
790 	entry->function = function;
791 	entry->index = index;
792 	switch (function & 0xC0000000) {
793 	case 0x40000000:
794 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
795 		return entry;
796 
797 	case 0x80000000:
798 		/*
799 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
800 		 * would result in out-of-bounds calls to do_host_cpuid.
801 		 */
802 		{
803 			static int max_cpuid_80000000;
804 			if (!READ_ONCE(max_cpuid_80000000))
805 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
806 			if (function > READ_ONCE(max_cpuid_80000000))
807 				return entry;
808 		}
809 		break;
810 
811 	default:
812 		break;
813 	}
814 
815 	cpuid_count(entry->function, entry->index,
816 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
817 
818 	if (cpuid_function_is_indexed(function))
819 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
820 
821 	return entry;
822 }
823 
824 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
825 {
826 	struct kvm_cpuid_entry2 *entry;
827 
828 	if (array->nent >= array->maxnent)
829 		return -E2BIG;
830 
831 	entry = &array->entries[array->nent];
832 	entry->function = func;
833 	entry->index = 0;
834 	entry->flags = 0;
835 
836 	switch (func) {
837 	case 0:
838 		entry->eax = 7;
839 		++array->nent;
840 		break;
841 	case 1:
842 		entry->ecx = F(MOVBE);
843 		++array->nent;
844 		break;
845 	case 7:
846 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
847 		entry->eax = 0;
848 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
849 			entry->ecx = F(RDPID);
850 		++array->nent;
851 		break;
852 	default:
853 		break;
854 	}
855 
856 	return 0;
857 }
858 
859 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
860 {
861 	struct kvm_cpuid_entry2 *entry;
862 	int r, i, max_idx;
863 
864 	/* all calls to cpuid_count() should be made on the same cpu */
865 	get_cpu();
866 
867 	r = -E2BIG;
868 
869 	entry = do_host_cpuid(array, function, 0);
870 	if (!entry)
871 		goto out;
872 
873 	switch (function) {
874 	case 0:
875 		/* Limited to the highest leaf implemented in KVM. */
876 		entry->eax = min(entry->eax, 0x1fU);
877 		break;
878 	case 1:
879 		cpuid_entry_override(entry, CPUID_1_EDX);
880 		cpuid_entry_override(entry, CPUID_1_ECX);
881 		break;
882 	case 2:
883 		/*
884 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
885 		 * CPUID(function=2, index=0) may return different results each
886 		 * time, with the least-significant byte in EAX enumerating the
887 		 * number of times software should do CPUID(2, 0).
888 		 *
889 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
890 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
891 		 * return 01H. Software should ignore this value and not
892 		 * interpret it as an informational descriptor", while AMD's
893 		 * APM states that CPUID(2) is reserved.
894 		 *
895 		 * WARN if a frankenstein CPU that supports virtualization and
896 		 * a stateful CPUID.0x2 is encountered.
897 		 */
898 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
899 		break;
900 	/* functions 4 and 0x8000001d have additional index. */
901 	case 4:
902 	case 0x8000001d:
903 		/*
904 		 * Read entries until the cache type in the previous entry is
905 		 * zero, i.e. indicates an invalid entry.
906 		 */
907 		for (i = 1; entry->eax & 0x1f; ++i) {
908 			entry = do_host_cpuid(array, function, i);
909 			if (!entry)
910 				goto out;
911 		}
912 		break;
913 	case 6: /* Thermal management */
914 		entry->eax = 0x4; /* allow ARAT */
915 		entry->ebx = 0;
916 		entry->ecx = 0;
917 		entry->edx = 0;
918 		break;
919 	/* function 7 has additional index. */
920 	case 7:
921 		entry->eax = min(entry->eax, 1u);
922 		cpuid_entry_override(entry, CPUID_7_0_EBX);
923 		cpuid_entry_override(entry, CPUID_7_ECX);
924 		cpuid_entry_override(entry, CPUID_7_EDX);
925 
926 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
927 		if (entry->eax == 1) {
928 			entry = do_host_cpuid(array, function, 1);
929 			if (!entry)
930 				goto out;
931 
932 			cpuid_entry_override(entry, CPUID_7_1_EAX);
933 			cpuid_entry_override(entry, CPUID_7_1_EDX);
934 			entry->ebx = 0;
935 			entry->ecx = 0;
936 		}
937 		break;
938 	case 0xa: { /* Architectural Performance Monitoring */
939 		union cpuid10_eax eax;
940 		union cpuid10_edx edx;
941 
942 		if (!static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
943 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
944 			break;
945 		}
946 
947 		eax.split.version_id = kvm_pmu_cap.version;
948 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
949 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
950 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
951 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
952 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
953 
954 		if (kvm_pmu_cap.version)
955 			edx.split.anythread_deprecated = 1;
956 		edx.split.reserved1 = 0;
957 		edx.split.reserved2 = 0;
958 
959 		entry->eax = eax.full;
960 		entry->ebx = kvm_pmu_cap.events_mask;
961 		entry->ecx = 0;
962 		entry->edx = edx.full;
963 		break;
964 	}
965 	case 0x1f:
966 	case 0xb:
967 		/*
968 		 * No topology; a valid topology is indicated by the presence
969 		 * of subleaf 1.
970 		 */
971 		entry->eax = entry->ebx = entry->ecx = 0;
972 		break;
973 	case 0xd: {
974 		u64 permitted_xcr0 = kvm_caps.supported_xcr0 & xstate_get_guest_group_perm();
975 		u64 permitted_xss = kvm_caps.supported_xss;
976 
977 		entry->eax &= permitted_xcr0;
978 		entry->ebx = xstate_required_size(permitted_xcr0, false);
979 		entry->ecx = entry->ebx;
980 		entry->edx &= permitted_xcr0 >> 32;
981 		if (!permitted_xcr0)
982 			break;
983 
984 		entry = do_host_cpuid(array, function, 1);
985 		if (!entry)
986 			goto out;
987 
988 		cpuid_entry_override(entry, CPUID_D_1_EAX);
989 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
990 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
991 							  true);
992 		else {
993 			WARN_ON_ONCE(permitted_xss != 0);
994 			entry->ebx = 0;
995 		}
996 		entry->ecx &= permitted_xss;
997 		entry->edx &= permitted_xss >> 32;
998 
999 		for (i = 2; i < 64; ++i) {
1000 			bool s_state;
1001 			if (permitted_xcr0 & BIT_ULL(i))
1002 				s_state = false;
1003 			else if (permitted_xss & BIT_ULL(i))
1004 				s_state = true;
1005 			else
1006 				continue;
1007 
1008 			entry = do_host_cpuid(array, function, i);
1009 			if (!entry)
1010 				goto out;
1011 
1012 			/*
1013 			 * The supported check above should have filtered out
1014 			 * invalid sub-leafs.  Only valid sub-leafs should
1015 			 * reach this point, and they should have a non-zero
1016 			 * save state size.  Furthermore, check whether the
1017 			 * processor agrees with permitted_xcr0/permitted_xss
1018 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1019 			 */
1020 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1021 				--array->nent;
1022 				continue;
1023 			}
1024 
1025 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1026 				entry->ecx &= ~BIT_ULL(2);
1027 			entry->edx = 0;
1028 		}
1029 		break;
1030 	}
1031 	case 0x12:
1032 		/* Intel SGX */
1033 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1034 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1035 			break;
1036 		}
1037 
1038 		/*
1039 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1040 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1041 		 * are restricted by kernel and KVM capabilities (like most
1042 		 * feature flags), while enclave size is unrestricted.
1043 		 */
1044 		cpuid_entry_override(entry, CPUID_12_EAX);
1045 		entry->ebx &= SGX_MISC_EXINFO;
1046 
1047 		entry = do_host_cpuid(array, function, 1);
1048 		if (!entry)
1049 			goto out;
1050 
1051 		/*
1052 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1053 		 * feature flags.  Advertise all supported flags, including
1054 		 * privileged attributes that require explicit opt-in from
1055 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1056 		 * expected to derive it from supported XCR0.
1057 		 */
1058 		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1059 		entry->ebx &= 0;
1060 		break;
1061 	/* Intel PT */
1062 	case 0x14:
1063 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1064 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1065 			break;
1066 		}
1067 
1068 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1069 			if (!do_host_cpuid(array, function, i))
1070 				goto out;
1071 		}
1072 		break;
1073 	/* Intel AMX TILE */
1074 	case 0x1d:
1075 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1076 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1077 			break;
1078 		}
1079 
1080 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1081 			if (!do_host_cpuid(array, function, i))
1082 				goto out;
1083 		}
1084 		break;
1085 	case 0x1e: /* TMUL information */
1086 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1087 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1088 			break;
1089 		}
1090 		break;
1091 	case KVM_CPUID_SIGNATURE: {
1092 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1093 		entry->eax = KVM_CPUID_FEATURES;
1094 		entry->ebx = sigptr[0];
1095 		entry->ecx = sigptr[1];
1096 		entry->edx = sigptr[2];
1097 		break;
1098 	}
1099 	case KVM_CPUID_FEATURES:
1100 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1101 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1102 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1103 			     (1 << KVM_FEATURE_ASYNC_PF) |
1104 			     (1 << KVM_FEATURE_PV_EOI) |
1105 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1106 			     (1 << KVM_FEATURE_PV_UNHALT) |
1107 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1108 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1109 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1110 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1111 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1112 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1113 
1114 		if (sched_info_on())
1115 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1116 
1117 		entry->ebx = 0;
1118 		entry->ecx = 0;
1119 		entry->edx = 0;
1120 		break;
1121 	case 0x80000000:
1122 		entry->eax = min(entry->eax, 0x80000021);
1123 		/*
1124 		 * Serializing LFENCE is reported in a multitude of ways, and
1125 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1126 		 * userspace by providing the CPUID leaf ourselves.
1127 		 *
1128 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1129 		 * QEMU thinks that it can query the host blindly for that
1130 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1131 		 * above.  The processor merrily returns values from the
1132 		 * highest Intel leaf which QEMU tries to use as the guest's
1133 		 * 0x8000001d.  Even worse, this can result in an infinite
1134 		 * loop if said highest leaf has no subleaves indexed by ECX.
1135 		 */
1136 		if (entry->eax >= 0x8000001d &&
1137 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1138 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1139 			entry->eax = max(entry->eax, 0x80000021);
1140 		break;
1141 	case 0x80000001:
1142 		entry->ebx &= ~GENMASK(27, 16);
1143 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1144 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1145 		break;
1146 	case 0x80000006:
1147 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1148 		entry->edx &= ~GENMASK(17, 16);
1149 		break;
1150 	case 0x80000007: /* Advanced power management */
1151 		/* invariant TSC is CPUID.80000007H:EDX[8] */
1152 		entry->edx &= (1 << 8);
1153 		/* mask against host */
1154 		entry->edx &= boot_cpu_data.x86_power;
1155 		entry->eax = entry->ebx = entry->ecx = 0;
1156 		break;
1157 	case 0x80000008: {
1158 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1159 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1160 		unsigned phys_as = entry->eax & 0xff;
1161 
1162 		/*
1163 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1164 		 * the guest operates in the same PA space as the host, i.e.
1165 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1166 		 * paging, too.
1167 		 *
1168 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1169 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1170 		 * the HPAs do not affect GPAs.
1171 		 */
1172 		if (!tdp_enabled)
1173 			g_phys_as = boot_cpu_data.x86_phys_bits;
1174 		else if (!g_phys_as)
1175 			g_phys_as = phys_as;
1176 
1177 		entry->eax = g_phys_as | (virt_as << 8);
1178 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1179 		entry->edx = 0;
1180 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1181 		break;
1182 	}
1183 	case 0x8000000A:
1184 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1185 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1186 			break;
1187 		}
1188 		entry->eax = 1; /* SVM revision 1 */
1189 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1190 				   ASID emulation to nested SVM */
1191 		entry->ecx = 0; /* Reserved */
1192 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1193 		break;
1194 	case 0x80000019:
1195 		entry->ecx = entry->edx = 0;
1196 		break;
1197 	case 0x8000001a:
1198 		entry->eax &= GENMASK(2, 0);
1199 		entry->ebx = entry->ecx = entry->edx = 0;
1200 		break;
1201 	case 0x8000001e:
1202 		/* Do not return host topology information.  */
1203 		entry->eax = entry->ebx = entry->ecx = 0;
1204 		entry->edx = 0; /* reserved */
1205 		break;
1206 	case 0x8000001F:
1207 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1208 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1209 		} else {
1210 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1211 			/* Clear NumVMPL since KVM does not support VMPL.  */
1212 			entry->ebx &= ~GENMASK(31, 12);
1213 			/*
1214 			 * Enumerate '0' for "PA bits reduction", the adjusted
1215 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1216 			 */
1217 			entry->ebx &= ~GENMASK(11, 6);
1218 		}
1219 		break;
1220 	case 0x80000020:
1221 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1222 		break;
1223 	case 0x80000021:
1224 		entry->ebx = entry->ecx = entry->edx = 0;
1225 		/*
1226 		 * Pass down these bits:
1227 		 *    EAX      0      NNDBP, Processor ignores nested data breakpoints
1228 		 *    EAX      2      LAS, LFENCE always serializing
1229 		 *    EAX      6      NSCB, Null selector clear base
1230 		 *
1231 		 * Other defined bits are for MSRs that KVM does not expose:
1232 		 *   EAX      3      SPCL, SMM page configuration lock
1233 		 *   EAX      13     PCMSR, Prefetch control MSR
1234 		 *
1235 		 * KVM doesn't support SMM_CTL.
1236 		 *   EAX       9     SMM_CTL MSR is not supported
1237 		 */
1238 		entry->eax &= BIT(0) | BIT(2) | BIT(6);
1239 		entry->eax |= BIT(9);
1240 		if (static_cpu_has(X86_FEATURE_LFENCE_RDTSC))
1241 			entry->eax |= BIT(2);
1242 		if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
1243 			entry->eax |= BIT(6);
1244 		break;
1245 	/*Add support for Centaur's CPUID instruction*/
1246 	case 0xC0000000:
1247 		/*Just support up to 0xC0000004 now*/
1248 		entry->eax = min(entry->eax, 0xC0000004);
1249 		break;
1250 	case 0xC0000001:
1251 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1252 		break;
1253 	case 3: /* Processor serial number */
1254 	case 5: /* MONITOR/MWAIT */
1255 	case 0xC0000002:
1256 	case 0xC0000003:
1257 	case 0xC0000004:
1258 	default:
1259 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1260 		break;
1261 	}
1262 
1263 	r = 0;
1264 
1265 out:
1266 	put_cpu();
1267 
1268 	return r;
1269 }
1270 
1271 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1272 			 unsigned int type)
1273 {
1274 	if (type == KVM_GET_EMULATED_CPUID)
1275 		return __do_cpuid_func_emulated(array, func);
1276 
1277 	return __do_cpuid_func(array, func);
1278 }
1279 
1280 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1281 
1282 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1283 			  unsigned int type)
1284 {
1285 	u32 limit;
1286 	int r;
1287 
1288 	if (func == CENTAUR_CPUID_SIGNATURE &&
1289 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1290 		return 0;
1291 
1292 	r = do_cpuid_func(array, func, type);
1293 	if (r)
1294 		return r;
1295 
1296 	limit = array->entries[array->nent - 1].eax;
1297 	for (func = func + 1; func <= limit; ++func) {
1298 		r = do_cpuid_func(array, func, type);
1299 		if (r)
1300 			break;
1301 	}
1302 
1303 	return r;
1304 }
1305 
1306 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1307 				 __u32 num_entries, unsigned int ioctl_type)
1308 {
1309 	int i;
1310 	__u32 pad[3];
1311 
1312 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1313 		return false;
1314 
1315 	/*
1316 	 * We want to make sure that ->padding is being passed clean from
1317 	 * userspace in case we want to use it for something in the future.
1318 	 *
1319 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1320 	 * have to give ourselves satisfied only with the emulated side. /me
1321 	 * sheds a tear.
1322 	 */
1323 	for (i = 0; i < num_entries; i++) {
1324 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1325 			return true;
1326 
1327 		if (pad[0] || pad[1] || pad[2])
1328 			return true;
1329 	}
1330 	return false;
1331 }
1332 
1333 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1334 			    struct kvm_cpuid_entry2 __user *entries,
1335 			    unsigned int type)
1336 {
1337 	static const u32 funcs[] = {
1338 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1339 	};
1340 
1341 	struct kvm_cpuid_array array = {
1342 		.nent = 0,
1343 	};
1344 	int r, i;
1345 
1346 	if (cpuid->nent < 1)
1347 		return -E2BIG;
1348 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1349 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1350 
1351 	if (sanity_check_entries(entries, cpuid->nent, type))
1352 		return -EINVAL;
1353 
1354 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1355 	if (!array.entries)
1356 		return -ENOMEM;
1357 
1358 	array.maxnent = cpuid->nent;
1359 
1360 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1361 		r = get_cpuid_func(&array, funcs[i], type);
1362 		if (r)
1363 			goto out_free;
1364 	}
1365 	cpuid->nent = array.nent;
1366 
1367 	if (copy_to_user(entries, array.entries,
1368 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1369 		r = -EFAULT;
1370 
1371 out_free:
1372 	kvfree(array.entries);
1373 	return r;
1374 }
1375 
1376 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1377 						    u32 function, u32 index)
1378 {
1379 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1380 				 function, index);
1381 }
1382 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1383 
1384 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1385 					      u32 function)
1386 {
1387 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1388 				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1389 }
1390 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1391 
1392 /*
1393  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1394  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1395  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1396  * range.  Centaur/VIA follows Intel semantics.
1397  *
1398  * A leaf is considered out-of-range if its function is higher than the maximum
1399  * supported leaf of its associated class or if its associated class does not
1400  * exist.
1401  *
1402  * There are three primary classes to be considered, with their respective
1403  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1404  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1405  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1406  *
1407  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1408  *  - Hypervisor: 0x40000000 - 0x4fffffff
1409  *  - Extended:   0x80000000 - 0xbfffffff
1410  *  - Centaur:    0xc0000000 - 0xcfffffff
1411  *
1412  * The Hypervisor class is further subdivided into sub-classes that each act as
1413  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1414  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1415  * CPUID sub-classes are:
1416  *
1417  *  - HyperV:     0x40000000 - 0x400000ff
1418  *  - KVM:        0x40000100 - 0x400001ff
1419  */
1420 static struct kvm_cpuid_entry2 *
1421 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1422 {
1423 	struct kvm_cpuid_entry2 *basic, *class;
1424 	u32 function = *fn_ptr;
1425 
1426 	basic = kvm_find_cpuid_entry(vcpu, 0);
1427 	if (!basic)
1428 		return NULL;
1429 
1430 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1431 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1432 		return NULL;
1433 
1434 	if (function >= 0x40000000 && function <= 0x4fffffff)
1435 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1436 	else if (function >= 0xc0000000)
1437 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1438 	else
1439 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1440 
1441 	if (class && function <= class->eax)
1442 		return NULL;
1443 
1444 	/*
1445 	 * Leaf specific adjustments are also applied when redirecting to the
1446 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1447 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1448 	 * needs to be pulled from CPUID.0xb.1.
1449 	 */
1450 	*fn_ptr = basic->eax;
1451 
1452 	/*
1453 	 * The class does not exist or the requested function is out of range;
1454 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1455 	 * the original requested leaf is observed!
1456 	 */
1457 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1458 }
1459 
1460 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1461 	       u32 *ecx, u32 *edx, bool exact_only)
1462 {
1463 	u32 orig_function = *eax, function = *eax, index = *ecx;
1464 	struct kvm_cpuid_entry2 *entry;
1465 	bool exact, used_max_basic = false;
1466 
1467 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1468 	exact = !!entry;
1469 
1470 	if (!entry && !exact_only) {
1471 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1472 		used_max_basic = !!entry;
1473 	}
1474 
1475 	if (entry) {
1476 		*eax = entry->eax;
1477 		*ebx = entry->ebx;
1478 		*ecx = entry->ecx;
1479 		*edx = entry->edx;
1480 		if (function == 7 && index == 0) {
1481 			u64 data;
1482 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1483 			    (data & TSX_CTRL_CPUID_CLEAR))
1484 				*ebx &= ~(F(RTM) | F(HLE));
1485 		}
1486 	} else {
1487 		*eax = *ebx = *ecx = *edx = 0;
1488 		/*
1489 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1490 		 * and EDX is always the x2APIC ID, even for undefined
1491 		 * subleaves. Index 1 will exist iff the leaf is
1492 		 * implemented, so we pass through CL iff leaf 1
1493 		 * exists. EDX can be copied from any existing index.
1494 		 */
1495 		if (function == 0xb || function == 0x1f) {
1496 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1497 			if (entry) {
1498 				*ecx = index & 0xff;
1499 				*edx = entry->edx;
1500 			}
1501 		}
1502 	}
1503 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1504 			used_max_basic);
1505 	return exact;
1506 }
1507 EXPORT_SYMBOL_GPL(kvm_cpuid);
1508 
1509 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1510 {
1511 	u32 eax, ebx, ecx, edx;
1512 
1513 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1514 		return 1;
1515 
1516 	eax = kvm_rax_read(vcpu);
1517 	ecx = kvm_rcx_read(vcpu);
1518 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1519 	kvm_rax_write(vcpu, eax);
1520 	kvm_rbx_write(vcpu, ebx);
1521 	kvm_rcx_write(vcpu, ecx);
1522 	kvm_rdx_write(vcpu, edx);
1523 	return kvm_skip_emulated_instruction(vcpu);
1524 }
1525 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1526