xref: /linux/arch/x86/kvm/cpuid.c (revision 6c8c1406a6d6a3f2e61ac590f5c0994231bc6be7)
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_PSFD		(13*32+28) /* Predictive Store Forwarding Disable */
66 
67 #define F feature_bit
68 #define SF(name) (boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0)
69 
70 /*
71  * Magic value used by KVM when querying userspace-provided CPUID entries and
72  * doesn't care about the CPIUD index because the index of the function in
73  * question is not significant.  Note, this magic value must have at least one
74  * bit set in bits[63:32] and must be consumed as a u64 by cpuid_entry2_find()
75  * to avoid false positives when processing guest CPUID input.
76  */
77 #define KVM_CPUID_INDEX_NOT_SIGNIFICANT -1ull
78 
79 static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
80 	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u64 index)
81 {
82 	struct kvm_cpuid_entry2 *e;
83 	int i;
84 
85 	for (i = 0; i < nent; i++) {
86 		e = &entries[i];
87 
88 		if (e->function != function)
89 			continue;
90 
91 		/*
92 		 * If the index isn't significant, use the first entry with a
93 		 * matching function.  It's userspace's responsibilty to not
94 		 * provide "duplicate" entries in all cases.
95 		 */
96 		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
97 			return e;
98 
99 
100 		/*
101 		 * Similarly, use the first matching entry if KVM is doing a
102 		 * lookup (as opposed to emulating CPUID) for a function that's
103 		 * architecturally defined as not having a significant index.
104 		 */
105 		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
106 			/*
107 			 * Direct lookups from KVM should not diverge from what
108 			 * KVM defines internally (the architectural behavior).
109 			 */
110 			WARN_ON_ONCE(cpuid_function_is_indexed(function));
111 			return e;
112 		}
113 	}
114 
115 	return NULL;
116 }
117 
118 static int kvm_check_cpuid(struct kvm_vcpu *vcpu,
119 			   struct kvm_cpuid_entry2 *entries,
120 			   int nent)
121 {
122 	struct kvm_cpuid_entry2 *best;
123 	u64 xfeatures;
124 
125 	/*
126 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
127 	 * canonical address checks; exit if it is ever changed.
128 	 */
129 	best = cpuid_entry2_find(entries, nent, 0x80000008,
130 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
131 	if (best) {
132 		int vaddr_bits = (best->eax & 0xff00) >> 8;
133 
134 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
135 			return -EINVAL;
136 	}
137 
138 	/*
139 	 * Exposing dynamic xfeatures to the guest requires additional
140 	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
141 	 */
142 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
143 	if (!best)
144 		return 0;
145 
146 	xfeatures = best->eax | ((u64)best->edx << 32);
147 	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
148 	if (!xfeatures)
149 		return 0;
150 
151 	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
152 }
153 
154 /* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
155 static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
156 				 int nent)
157 {
158 	struct kvm_cpuid_entry2 *orig;
159 	int i;
160 
161 	if (nent != vcpu->arch.cpuid_nent)
162 		return -EINVAL;
163 
164 	for (i = 0; i < nent; i++) {
165 		orig = &vcpu->arch.cpuid_entries[i];
166 		if (e2[i].function != orig->function ||
167 		    e2[i].index != orig->index ||
168 		    e2[i].flags != orig->flags ||
169 		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
170 		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
171 			return -EINVAL;
172 	}
173 
174 	return 0;
175 }
176 
177 static void kvm_update_kvm_cpuid_base(struct kvm_vcpu *vcpu)
178 {
179 	u32 function;
180 	struct kvm_cpuid_entry2 *entry;
181 
182 	vcpu->arch.kvm_cpuid_base = 0;
183 
184 	for_each_possible_hypervisor_cpuid_base(function) {
185 		entry = kvm_find_cpuid_entry(vcpu, function);
186 
187 		if (entry) {
188 			u32 signature[3];
189 
190 			signature[0] = entry->ebx;
191 			signature[1] = entry->ecx;
192 			signature[2] = entry->edx;
193 
194 			BUILD_BUG_ON(sizeof(signature) > sizeof(KVM_SIGNATURE));
195 			if (!memcmp(signature, KVM_SIGNATURE, sizeof(signature))) {
196 				vcpu->arch.kvm_cpuid_base = function;
197 				break;
198 			}
199 		}
200 	}
201 }
202 
203 static struct kvm_cpuid_entry2 *__kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu,
204 					      struct kvm_cpuid_entry2 *entries, int nent)
205 {
206 	u32 base = vcpu->arch.kvm_cpuid_base;
207 
208 	if (!base)
209 		return NULL;
210 
211 	return cpuid_entry2_find(entries, nent, base | KVM_CPUID_FEATURES,
212 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
213 }
214 
215 static struct kvm_cpuid_entry2 *kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu)
216 {
217 	return __kvm_find_kvm_cpuid_features(vcpu, vcpu->arch.cpuid_entries,
218 					     vcpu->arch.cpuid_nent);
219 }
220 
221 void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
222 {
223 	struct kvm_cpuid_entry2 *best = kvm_find_kvm_cpuid_features(vcpu);
224 
225 	/*
226 	 * save the feature bitmap to avoid cpuid lookup for every PV
227 	 * operation
228 	 */
229 	if (best)
230 		vcpu->arch.pv_cpuid.features = best->eax;
231 }
232 
233 /*
234  * Calculate guest's supported XCR0 taking into account guest CPUID data and
235  * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
236  */
237 static u64 cpuid_get_supported_xcr0(struct kvm_cpuid_entry2 *entries, int nent)
238 {
239 	struct kvm_cpuid_entry2 *best;
240 
241 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
242 	if (!best)
243 		return 0;
244 
245 	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
246 }
247 
248 static void __kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *entries,
249 				       int nent)
250 {
251 	struct kvm_cpuid_entry2 *best;
252 	u64 guest_supported_xcr0 = cpuid_get_supported_xcr0(entries, nent);
253 
254 	best = cpuid_entry2_find(entries, nent, 1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
255 	if (best) {
256 		/* Update OSXSAVE bit */
257 		if (boot_cpu_has(X86_FEATURE_XSAVE))
258 			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
259 				   kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE));
260 
261 		cpuid_entry_change(best, X86_FEATURE_APIC,
262 			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
263 	}
264 
265 	best = cpuid_entry2_find(entries, nent, 7, 0);
266 	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
267 		cpuid_entry_change(best, X86_FEATURE_OSPKE,
268 				   kvm_read_cr4_bits(vcpu, X86_CR4_PKE));
269 
270 	best = cpuid_entry2_find(entries, nent, 0xD, 0);
271 	if (best)
272 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
273 
274 	best = cpuid_entry2_find(entries, nent, 0xD, 1);
275 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
276 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
277 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
278 
279 	best = __kvm_find_kvm_cpuid_features(vcpu, entries, nent);
280 	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
281 		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
282 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
283 
284 	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
285 		best = cpuid_entry2_find(entries, nent, 0x1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
286 		if (best)
287 			cpuid_entry_change(best, X86_FEATURE_MWAIT,
288 					   vcpu->arch.ia32_misc_enable_msr &
289 					   MSR_IA32_MISC_ENABLE_MWAIT);
290 	}
291 
292 	/*
293 	 * Bits 127:0 of the allowed SECS.ATTRIBUTES (CPUID.0x12.0x1) enumerate
294 	 * the supported XSAVE Feature Request Mask (XFRM), i.e. the enclave's
295 	 * requested XCR0 value.  The enclave's XFRM must be a subset of XCRO
296 	 * at the time of EENTER, thus adjust the allowed XFRM by the guest's
297 	 * supported XCR0.  Similar to XCR0 handling, FP and SSE are forced to
298 	 * '1' even on CPUs that don't support XSAVE.
299 	 */
300 	best = cpuid_entry2_find(entries, nent, 0x12, 0x1);
301 	if (best) {
302 		best->ecx &= guest_supported_xcr0 & 0xffffffff;
303 		best->edx &= guest_supported_xcr0 >> 32;
304 		best->ecx |= XFEATURE_MASK_FPSSE;
305 	}
306 }
307 
308 void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
309 {
310 	__kvm_update_cpuid_runtime(vcpu, vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
311 }
312 EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
313 
314 static bool kvm_cpuid_has_hyperv(struct kvm_cpuid_entry2 *entries, int nent)
315 {
316 	struct kvm_cpuid_entry2 *entry;
317 
318 	entry = cpuid_entry2_find(entries, nent, HYPERV_CPUID_INTERFACE,
319 				  KVM_CPUID_INDEX_NOT_SIGNIFICANT);
320 	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
321 }
322 
323 static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
324 {
325 	struct kvm_lapic *apic = vcpu->arch.apic;
326 	struct kvm_cpuid_entry2 *best;
327 
328 	best = kvm_find_cpuid_entry(vcpu, 1);
329 	if (best && apic) {
330 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
331 			apic->lapic_timer.timer_mode_mask = 3 << 17;
332 		else
333 			apic->lapic_timer.timer_mode_mask = 1 << 17;
334 
335 		kvm_apic_set_version(vcpu);
336 	}
337 
338 	vcpu->arch.guest_supported_xcr0 =
339 		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
340 
341 	/*
342 	 * FP+SSE can always be saved/restored via KVM_{G,S}ET_XSAVE, even if
343 	 * XSAVE/XCRO are not exposed to the guest, and even if XSAVE isn't
344 	 * supported by the host.
345 	 */
346 	vcpu->arch.guest_fpu.fpstate->user_xfeatures = vcpu->arch.guest_supported_xcr0 |
347 						       XFEATURE_MASK_FPSSE;
348 
349 	kvm_update_pv_runtime(vcpu);
350 
351 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
352 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
353 
354 	kvm_pmu_refresh(vcpu);
355 	vcpu->arch.cr4_guest_rsvd_bits =
356 	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
357 
358 	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu->arch.cpuid_entries,
359 						    vcpu->arch.cpuid_nent));
360 
361 	/* Invoke the vendor callback only after the above state is updated. */
362 	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
363 
364 	/*
365 	 * Except for the MMU, which needs to do its thing any vendor specific
366 	 * adjustments to the reserved GPA bits.
367 	 */
368 	kvm_mmu_after_set_cpuid(vcpu);
369 }
370 
371 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
372 {
373 	struct kvm_cpuid_entry2 *best;
374 
375 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
376 	if (!best || best->eax < 0x80000008)
377 		goto not_found;
378 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
379 	if (best)
380 		return best->eax & 0xff;
381 not_found:
382 	return 36;
383 }
384 
385 /*
386  * This "raw" version returns the reserved GPA bits without any adjustments for
387  * encryption technologies that usurp bits.  The raw mask should be used if and
388  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
389  */
390 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
391 {
392 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
393 }
394 
395 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
396                         int nent)
397 {
398 	int r;
399 
400 	__kvm_update_cpuid_runtime(vcpu, e2, nent);
401 
402 	/*
403 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
404 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
405 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
406 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
407 	 * the core vCPU model on the fly. It would've been better to forbid any
408 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
409 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
410 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
411 	 * whether the supplied CPUID data is equal to what's already set.
412 	 */
413 	if (vcpu->arch.last_vmentry_cpu != -1) {
414 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
415 		if (r)
416 			return r;
417 
418 		kvfree(e2);
419 		return 0;
420 	}
421 
422 	if (kvm_cpuid_has_hyperv(e2, nent)) {
423 		r = kvm_hv_vcpu_init(vcpu);
424 		if (r)
425 			return r;
426 	}
427 
428 	r = kvm_check_cpuid(vcpu, e2, nent);
429 	if (r)
430 		return r;
431 
432 	kvfree(vcpu->arch.cpuid_entries);
433 	vcpu->arch.cpuid_entries = e2;
434 	vcpu->arch.cpuid_nent = nent;
435 
436 	kvm_update_kvm_cpuid_base(vcpu);
437 	kvm_vcpu_after_set_cpuid(vcpu);
438 
439 	return 0;
440 }
441 
442 /* when an old userspace process fills a new kernel module */
443 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
444 			     struct kvm_cpuid *cpuid,
445 			     struct kvm_cpuid_entry __user *entries)
446 {
447 	int r, i;
448 	struct kvm_cpuid_entry *e = NULL;
449 	struct kvm_cpuid_entry2 *e2 = NULL;
450 
451 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
452 		return -E2BIG;
453 
454 	if (cpuid->nent) {
455 		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
456 		if (IS_ERR(e))
457 			return PTR_ERR(e);
458 
459 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
460 		if (!e2) {
461 			r = -ENOMEM;
462 			goto out_free_cpuid;
463 		}
464 	}
465 	for (i = 0; i < cpuid->nent; i++) {
466 		e2[i].function = e[i].function;
467 		e2[i].eax = e[i].eax;
468 		e2[i].ebx = e[i].ebx;
469 		e2[i].ecx = e[i].ecx;
470 		e2[i].edx = e[i].edx;
471 		e2[i].index = 0;
472 		e2[i].flags = 0;
473 		e2[i].padding[0] = 0;
474 		e2[i].padding[1] = 0;
475 		e2[i].padding[2] = 0;
476 	}
477 
478 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
479 	if (r)
480 		kvfree(e2);
481 
482 out_free_cpuid:
483 	kvfree(e);
484 
485 	return r;
486 }
487 
488 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
489 			      struct kvm_cpuid2 *cpuid,
490 			      struct kvm_cpuid_entry2 __user *entries)
491 {
492 	struct kvm_cpuid_entry2 *e2 = NULL;
493 	int r;
494 
495 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
496 		return -E2BIG;
497 
498 	if (cpuid->nent) {
499 		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
500 		if (IS_ERR(e2))
501 			return PTR_ERR(e2);
502 	}
503 
504 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
505 	if (r)
506 		kvfree(e2);
507 
508 	return r;
509 }
510 
511 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
512 			      struct kvm_cpuid2 *cpuid,
513 			      struct kvm_cpuid_entry2 __user *entries)
514 {
515 	int r;
516 
517 	r = -E2BIG;
518 	if (cpuid->nent < vcpu->arch.cpuid_nent)
519 		goto out;
520 	r = -EFAULT;
521 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
522 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
523 		goto out;
524 	return 0;
525 
526 out:
527 	cpuid->nent = vcpu->arch.cpuid_nent;
528 	return r;
529 }
530 
531 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
532 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
533 {
534 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
535 	struct kvm_cpuid_entry2 entry;
536 
537 	reverse_cpuid_check(leaf);
538 
539 	cpuid_count(cpuid.function, cpuid.index,
540 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
541 
542 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
543 }
544 
545 static __always_inline
546 void kvm_cpu_cap_init_scattered(enum kvm_only_cpuid_leafs leaf, u32 mask)
547 {
548 	/* Use kvm_cpu_cap_mask for non-scattered leafs. */
549 	BUILD_BUG_ON(leaf < NCAPINTS);
550 
551 	kvm_cpu_caps[leaf] = mask;
552 
553 	__kvm_cpu_cap_mask(leaf);
554 }
555 
556 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
557 {
558 	/* Use kvm_cpu_cap_init_scattered for scattered leafs. */
559 	BUILD_BUG_ON(leaf >= NCAPINTS);
560 
561 	kvm_cpu_caps[leaf] &= mask;
562 
563 	__kvm_cpu_cap_mask(leaf);
564 }
565 
566 void kvm_set_cpu_caps(void)
567 {
568 #ifdef CONFIG_X86_64
569 	unsigned int f_gbpages = F(GBPAGES);
570 	unsigned int f_lm = F(LM);
571 	unsigned int f_xfd = F(XFD);
572 #else
573 	unsigned int f_gbpages = 0;
574 	unsigned int f_lm = 0;
575 	unsigned int f_xfd = 0;
576 #endif
577 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
578 
579 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
580 		     sizeof(boot_cpu_data.x86_capability));
581 
582 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
583 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
584 
585 	kvm_cpu_cap_mask(CPUID_1_ECX,
586 		/*
587 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
588 		 * advertised to guests via CPUID!
589 		 */
590 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
591 		0 /* DS-CPL, VMX, SMX, EST */ |
592 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
593 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
594 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
595 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
596 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
597 		F(F16C) | F(RDRAND)
598 	);
599 	/* KVM emulates x2apic in software irrespective of host support. */
600 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
601 
602 	kvm_cpu_cap_mask(CPUID_1_EDX,
603 		F(FPU) | F(VME) | F(DE) | F(PSE) |
604 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
605 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
606 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
607 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
608 		0 /* Reserved, DS, ACPI */ | F(MMX) |
609 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
610 		0 /* HTT, TM, Reserved, PBE */
611 	);
612 
613 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
614 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
615 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
616 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
617 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
618 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
619 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
620 		F(AVX512VL));
621 
622 	kvm_cpu_cap_mask(CPUID_7_ECX,
623 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
624 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
625 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
626 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
627 		F(SGX_LC) | F(BUS_LOCK_DETECT)
628 	);
629 	/* Set LA57 based on hardware capability. */
630 	if (cpuid_ecx(7) & F(LA57))
631 		kvm_cpu_cap_set(X86_FEATURE_LA57);
632 
633 	/*
634 	 * PKU not yet implemented for shadow paging and requires OSPKE
635 	 * to be set on the host. Clear it if that is not the case
636 	 */
637 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
638 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
639 
640 	kvm_cpu_cap_mask(CPUID_7_EDX,
641 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
642 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
643 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
644 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
645 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16)
646 	);
647 
648 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
649 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
650 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
651 
652 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
653 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
654 	if (boot_cpu_has(X86_FEATURE_STIBP))
655 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
656 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
657 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
658 
659 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
660 		F(AVX_VNNI) | F(AVX512_BF16)
661 	);
662 
663 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
664 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
665 	);
666 
667 	kvm_cpu_cap_init_scattered(CPUID_12_EAX,
668 		SF(SGX1) | SF(SGX2)
669 	);
670 
671 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
672 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
673 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
674 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
675 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
676 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
677 	);
678 
679 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
680 		F(FPU) | F(VME) | F(DE) | F(PSE) |
681 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
682 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
683 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
684 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
685 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
686 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
687 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
688 	);
689 
690 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
691 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
692 
693 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
694 		F(CLZERO) | F(XSAVEERPTR) |
695 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
696 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
697 		__feature_bit(KVM_X86_FEATURE_PSFD)
698 	);
699 
700 	/*
701 	 * AMD has separate bits for each SPEC_CTRL bit.
702 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
703 	 * record that in cpufeatures so use them.
704 	 */
705 	if (boot_cpu_has(X86_FEATURE_IBPB))
706 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
707 	if (boot_cpu_has(X86_FEATURE_IBRS))
708 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
709 	if (boot_cpu_has(X86_FEATURE_STIBP))
710 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
711 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
712 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
713 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
714 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
715 	/*
716 	 * The preference is to use SPEC CTRL MSR instead of the
717 	 * VIRT_SPEC MSR.
718 	 */
719 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
720 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
721 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
722 
723 	/*
724 	 * Hide all SVM features by default, SVM will set the cap bits for
725 	 * features it emulates and/or exposes for L1.
726 	 */
727 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
728 
729 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
730 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
731 		F(SME_COHERENT));
732 
733 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
734 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
735 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
736 		F(PMM) | F(PMM_EN)
737 	);
738 
739 	/*
740 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
741 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
742 	 * should never happen, but the guest will likely crash if RDTSCP or
743 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
744 	 * the past.  For example, the sanity check may fire if this instance of
745 	 * KVM is running as L1 on top of an older, broken KVM.
746 	 */
747 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
748 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
749 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
750 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
751 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
752 	}
753 }
754 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
755 
756 struct kvm_cpuid_array {
757 	struct kvm_cpuid_entry2 *entries;
758 	int maxnent;
759 	int nent;
760 };
761 
762 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
763 					      u32 function, u32 index)
764 {
765 	struct kvm_cpuid_entry2 *entry;
766 
767 	if (array->nent >= array->maxnent)
768 		return NULL;
769 
770 	entry = &array->entries[array->nent++];
771 
772 	memset(entry, 0, sizeof(*entry));
773 	entry->function = function;
774 	entry->index = index;
775 	switch (function & 0xC0000000) {
776 	case 0x40000000:
777 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
778 		return entry;
779 
780 	case 0x80000000:
781 		/*
782 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
783 		 * would result in out-of-bounds calls to do_host_cpuid.
784 		 */
785 		{
786 			static int max_cpuid_80000000;
787 			if (!READ_ONCE(max_cpuid_80000000))
788 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
789 			if (function > READ_ONCE(max_cpuid_80000000))
790 				return entry;
791 		}
792 		break;
793 
794 	default:
795 		break;
796 	}
797 
798 	cpuid_count(entry->function, entry->index,
799 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
800 
801 	if (cpuid_function_is_indexed(function))
802 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
803 
804 	return entry;
805 }
806 
807 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
808 {
809 	struct kvm_cpuid_entry2 *entry;
810 
811 	if (array->nent >= array->maxnent)
812 		return -E2BIG;
813 
814 	entry = &array->entries[array->nent];
815 	entry->function = func;
816 	entry->index = 0;
817 	entry->flags = 0;
818 
819 	switch (func) {
820 	case 0:
821 		entry->eax = 7;
822 		++array->nent;
823 		break;
824 	case 1:
825 		entry->ecx = F(MOVBE);
826 		++array->nent;
827 		break;
828 	case 7:
829 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
830 		entry->eax = 0;
831 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
832 			entry->ecx = F(RDPID);
833 		++array->nent;
834 		break;
835 	default:
836 		break;
837 	}
838 
839 	return 0;
840 }
841 
842 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
843 {
844 	struct kvm_cpuid_entry2 *entry;
845 	int r, i, max_idx;
846 
847 	/* all calls to cpuid_count() should be made on the same cpu */
848 	get_cpu();
849 
850 	r = -E2BIG;
851 
852 	entry = do_host_cpuid(array, function, 0);
853 	if (!entry)
854 		goto out;
855 
856 	switch (function) {
857 	case 0:
858 		/* Limited to the highest leaf implemented in KVM. */
859 		entry->eax = min(entry->eax, 0x1fU);
860 		break;
861 	case 1:
862 		cpuid_entry_override(entry, CPUID_1_EDX);
863 		cpuid_entry_override(entry, CPUID_1_ECX);
864 		break;
865 	case 2:
866 		/*
867 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
868 		 * CPUID(function=2, index=0) may return different results each
869 		 * time, with the least-significant byte in EAX enumerating the
870 		 * number of times software should do CPUID(2, 0).
871 		 *
872 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
873 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
874 		 * return 01H. Software should ignore this value and not
875 		 * interpret it as an informational descriptor", while AMD's
876 		 * APM states that CPUID(2) is reserved.
877 		 *
878 		 * WARN if a frankenstein CPU that supports virtualization and
879 		 * a stateful CPUID.0x2 is encountered.
880 		 */
881 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
882 		break;
883 	/* functions 4 and 0x8000001d have additional index. */
884 	case 4:
885 	case 0x8000001d:
886 		/*
887 		 * Read entries until the cache type in the previous entry is
888 		 * zero, i.e. indicates an invalid entry.
889 		 */
890 		for (i = 1; entry->eax & 0x1f; ++i) {
891 			entry = do_host_cpuid(array, function, i);
892 			if (!entry)
893 				goto out;
894 		}
895 		break;
896 	case 6: /* Thermal management */
897 		entry->eax = 0x4; /* allow ARAT */
898 		entry->ebx = 0;
899 		entry->ecx = 0;
900 		entry->edx = 0;
901 		break;
902 	/* function 7 has additional index. */
903 	case 7:
904 		entry->eax = min(entry->eax, 1u);
905 		cpuid_entry_override(entry, CPUID_7_0_EBX);
906 		cpuid_entry_override(entry, CPUID_7_ECX);
907 		cpuid_entry_override(entry, CPUID_7_EDX);
908 
909 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
910 		if (entry->eax == 1) {
911 			entry = do_host_cpuid(array, function, 1);
912 			if (!entry)
913 				goto out;
914 
915 			cpuid_entry_override(entry, CPUID_7_1_EAX);
916 			entry->ebx = 0;
917 			entry->ecx = 0;
918 			entry->edx = 0;
919 		}
920 		break;
921 	case 0xa: { /* Architectural Performance Monitoring */
922 		union cpuid10_eax eax;
923 		union cpuid10_edx edx;
924 
925 		if (!static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
926 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
927 			break;
928 		}
929 
930 		eax.split.version_id = kvm_pmu_cap.version;
931 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
932 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
933 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
934 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
935 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
936 
937 		if (kvm_pmu_cap.version)
938 			edx.split.anythread_deprecated = 1;
939 		edx.split.reserved1 = 0;
940 		edx.split.reserved2 = 0;
941 
942 		entry->eax = eax.full;
943 		entry->ebx = kvm_pmu_cap.events_mask;
944 		entry->ecx = 0;
945 		entry->edx = edx.full;
946 		break;
947 	}
948 	/*
949 	 * Per Intel's SDM, the 0x1f is a superset of 0xb,
950 	 * thus they can be handled by common code.
951 	 */
952 	case 0x1f:
953 	case 0xb:
954 		/*
955 		 * Populate entries until the level type (ECX[15:8]) of the
956 		 * previous entry is zero.  Note, CPUID EAX.{0x1f,0xb}.0 is
957 		 * the starting entry, filled by the primary do_host_cpuid().
958 		 */
959 		for (i = 1; entry->ecx & 0xff00; ++i) {
960 			entry = do_host_cpuid(array, function, i);
961 			if (!entry)
962 				goto out;
963 		}
964 		break;
965 	case 0xd: {
966 		u64 permitted_xcr0 = kvm_caps.supported_xcr0 & xstate_get_guest_group_perm();
967 		u64 permitted_xss = kvm_caps.supported_xss;
968 
969 		entry->eax &= permitted_xcr0;
970 		entry->ebx = xstate_required_size(permitted_xcr0, false);
971 		entry->ecx = entry->ebx;
972 		entry->edx &= permitted_xcr0 >> 32;
973 		if (!permitted_xcr0)
974 			break;
975 
976 		entry = do_host_cpuid(array, function, 1);
977 		if (!entry)
978 			goto out;
979 
980 		cpuid_entry_override(entry, CPUID_D_1_EAX);
981 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
982 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
983 							  true);
984 		else {
985 			WARN_ON_ONCE(permitted_xss != 0);
986 			entry->ebx = 0;
987 		}
988 		entry->ecx &= permitted_xss;
989 		entry->edx &= permitted_xss >> 32;
990 
991 		for (i = 2; i < 64; ++i) {
992 			bool s_state;
993 			if (permitted_xcr0 & BIT_ULL(i))
994 				s_state = false;
995 			else if (permitted_xss & BIT_ULL(i))
996 				s_state = true;
997 			else
998 				continue;
999 
1000 			entry = do_host_cpuid(array, function, i);
1001 			if (!entry)
1002 				goto out;
1003 
1004 			/*
1005 			 * The supported check above should have filtered out
1006 			 * invalid sub-leafs.  Only valid sub-leafs should
1007 			 * reach this point, and they should have a non-zero
1008 			 * save state size.  Furthermore, check whether the
1009 			 * processor agrees with permitted_xcr0/permitted_xss
1010 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1011 			 */
1012 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1013 				--array->nent;
1014 				continue;
1015 			}
1016 
1017 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1018 				entry->ecx &= ~BIT_ULL(2);
1019 			entry->edx = 0;
1020 		}
1021 		break;
1022 	}
1023 	case 0x12:
1024 		/* Intel SGX */
1025 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1026 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1027 			break;
1028 		}
1029 
1030 		/*
1031 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1032 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1033 		 * are restricted by kernel and KVM capabilities (like most
1034 		 * feature flags), while enclave size is unrestricted.
1035 		 */
1036 		cpuid_entry_override(entry, CPUID_12_EAX);
1037 		entry->ebx &= SGX_MISC_EXINFO;
1038 
1039 		entry = do_host_cpuid(array, function, 1);
1040 		if (!entry)
1041 			goto out;
1042 
1043 		/*
1044 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1045 		 * feature flags.  Advertise all supported flags, including
1046 		 * privileged attributes that require explicit opt-in from
1047 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1048 		 * expected to derive it from supported XCR0.
1049 		 */
1050 		entry->eax &= SGX_ATTR_DEBUG | SGX_ATTR_MODE64BIT |
1051 			      SGX_ATTR_PROVISIONKEY | SGX_ATTR_EINITTOKENKEY |
1052 			      SGX_ATTR_KSS;
1053 		entry->ebx &= 0;
1054 		break;
1055 	/* Intel PT */
1056 	case 0x14:
1057 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1058 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1059 			break;
1060 		}
1061 
1062 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1063 			if (!do_host_cpuid(array, function, i))
1064 				goto out;
1065 		}
1066 		break;
1067 	/* Intel AMX TILE */
1068 	case 0x1d:
1069 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1070 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1071 			break;
1072 		}
1073 
1074 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1075 			if (!do_host_cpuid(array, function, i))
1076 				goto out;
1077 		}
1078 		break;
1079 	case 0x1e: /* TMUL information */
1080 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1081 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1082 			break;
1083 		}
1084 		break;
1085 	case KVM_CPUID_SIGNATURE: {
1086 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1087 		entry->eax = KVM_CPUID_FEATURES;
1088 		entry->ebx = sigptr[0];
1089 		entry->ecx = sigptr[1];
1090 		entry->edx = sigptr[2];
1091 		break;
1092 	}
1093 	case KVM_CPUID_FEATURES:
1094 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1095 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1096 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1097 			     (1 << KVM_FEATURE_ASYNC_PF) |
1098 			     (1 << KVM_FEATURE_PV_EOI) |
1099 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1100 			     (1 << KVM_FEATURE_PV_UNHALT) |
1101 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1102 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1103 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1104 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1105 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1106 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1107 
1108 		if (sched_info_on())
1109 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1110 
1111 		entry->ebx = 0;
1112 		entry->ecx = 0;
1113 		entry->edx = 0;
1114 		break;
1115 	case 0x80000000:
1116 		entry->eax = min(entry->eax, 0x80000021);
1117 		/*
1118 		 * Serializing LFENCE is reported in a multitude of ways, and
1119 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1120 		 * userspace by providing the CPUID leaf ourselves.
1121 		 *
1122 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1123 		 * QEMU thinks that it can query the host blindly for that
1124 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1125 		 * above.  The processor merrily returns values from the
1126 		 * highest Intel leaf which QEMU tries to use as the guest's
1127 		 * 0x8000001d.  Even worse, this can result in an infinite
1128 		 * loop if said highest leaf has no subleaves indexed by ECX.
1129 		 */
1130 		if (entry->eax >= 0x8000001d &&
1131 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1132 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1133 			entry->eax = max(entry->eax, 0x80000021);
1134 		break;
1135 	case 0x80000001:
1136 		entry->ebx &= ~GENMASK(27, 16);
1137 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1138 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1139 		break;
1140 	case 0x80000006:
1141 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1142 		entry->edx &= ~GENMASK(17, 16);
1143 		break;
1144 	case 0x80000007: /* Advanced power management */
1145 		/* invariant TSC is CPUID.80000007H:EDX[8] */
1146 		entry->edx &= (1 << 8);
1147 		/* mask against host */
1148 		entry->edx &= boot_cpu_data.x86_power;
1149 		entry->eax = entry->ebx = entry->ecx = 0;
1150 		break;
1151 	case 0x80000008: {
1152 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1153 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1154 		unsigned phys_as = entry->eax & 0xff;
1155 
1156 		/*
1157 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1158 		 * the guest operates in the same PA space as the host, i.e.
1159 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1160 		 * paging, too.
1161 		 *
1162 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1163 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1164 		 * the HPAs do not affect GPAs.
1165 		 */
1166 		if (!tdp_enabled)
1167 			g_phys_as = boot_cpu_data.x86_phys_bits;
1168 		else if (!g_phys_as)
1169 			g_phys_as = phys_as;
1170 
1171 		entry->eax = g_phys_as | (virt_as << 8);
1172 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1173 		entry->edx = 0;
1174 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1175 		break;
1176 	}
1177 	case 0x8000000A:
1178 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1179 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1180 			break;
1181 		}
1182 		entry->eax = 1; /* SVM revision 1 */
1183 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1184 				   ASID emulation to nested SVM */
1185 		entry->ecx = 0; /* Reserved */
1186 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1187 		break;
1188 	case 0x80000019:
1189 		entry->ecx = entry->edx = 0;
1190 		break;
1191 	case 0x8000001a:
1192 		entry->eax &= GENMASK(2, 0);
1193 		entry->ebx = entry->ecx = entry->edx = 0;
1194 		break;
1195 	case 0x8000001e:
1196 		break;
1197 	case 0x8000001F:
1198 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1199 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1200 		} else {
1201 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1202 			/* Clear NumVMPL since KVM does not support VMPL.  */
1203 			entry->ebx &= ~GENMASK(31, 12);
1204 			/*
1205 			 * Enumerate '0' for "PA bits reduction", the adjusted
1206 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1207 			 */
1208 			entry->ebx &= ~GENMASK(11, 6);
1209 		}
1210 		break;
1211 	case 0x80000020:
1212 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1213 		break;
1214 	case 0x80000021:
1215 		entry->ebx = entry->ecx = entry->edx = 0;
1216 		/*
1217 		 * Pass down these bits:
1218 		 *    EAX      0      NNDBP, Processor ignores nested data breakpoints
1219 		 *    EAX      2      LAS, LFENCE always serializing
1220 		 *    EAX      6      NSCB, Null selector clear base
1221 		 *
1222 		 * Other defined bits are for MSRs that KVM does not expose:
1223 		 *   EAX      3      SPCL, SMM page configuration lock
1224 		 *   EAX      13     PCMSR, Prefetch control MSR
1225 		 */
1226 		entry->eax &= BIT(0) | BIT(2) | BIT(6);
1227 		if (static_cpu_has(X86_FEATURE_LFENCE_RDTSC))
1228 			entry->eax |= BIT(2);
1229 		if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
1230 			entry->eax |= BIT(6);
1231 		break;
1232 	/*Add support for Centaur's CPUID instruction*/
1233 	case 0xC0000000:
1234 		/*Just support up to 0xC0000004 now*/
1235 		entry->eax = min(entry->eax, 0xC0000004);
1236 		break;
1237 	case 0xC0000001:
1238 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1239 		break;
1240 	case 3: /* Processor serial number */
1241 	case 5: /* MONITOR/MWAIT */
1242 	case 0xC0000002:
1243 	case 0xC0000003:
1244 	case 0xC0000004:
1245 	default:
1246 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1247 		break;
1248 	}
1249 
1250 	r = 0;
1251 
1252 out:
1253 	put_cpu();
1254 
1255 	return r;
1256 }
1257 
1258 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1259 			 unsigned int type)
1260 {
1261 	if (type == KVM_GET_EMULATED_CPUID)
1262 		return __do_cpuid_func_emulated(array, func);
1263 
1264 	return __do_cpuid_func(array, func);
1265 }
1266 
1267 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1268 
1269 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1270 			  unsigned int type)
1271 {
1272 	u32 limit;
1273 	int r;
1274 
1275 	if (func == CENTAUR_CPUID_SIGNATURE &&
1276 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1277 		return 0;
1278 
1279 	r = do_cpuid_func(array, func, type);
1280 	if (r)
1281 		return r;
1282 
1283 	limit = array->entries[array->nent - 1].eax;
1284 	for (func = func + 1; func <= limit; ++func) {
1285 		r = do_cpuid_func(array, func, type);
1286 		if (r)
1287 			break;
1288 	}
1289 
1290 	return r;
1291 }
1292 
1293 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1294 				 __u32 num_entries, unsigned int ioctl_type)
1295 {
1296 	int i;
1297 	__u32 pad[3];
1298 
1299 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1300 		return false;
1301 
1302 	/*
1303 	 * We want to make sure that ->padding is being passed clean from
1304 	 * userspace in case we want to use it for something in the future.
1305 	 *
1306 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1307 	 * have to give ourselves satisfied only with the emulated side. /me
1308 	 * sheds a tear.
1309 	 */
1310 	for (i = 0; i < num_entries; i++) {
1311 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1312 			return true;
1313 
1314 		if (pad[0] || pad[1] || pad[2])
1315 			return true;
1316 	}
1317 	return false;
1318 }
1319 
1320 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1321 			    struct kvm_cpuid_entry2 __user *entries,
1322 			    unsigned int type)
1323 {
1324 	static const u32 funcs[] = {
1325 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1326 	};
1327 
1328 	struct kvm_cpuid_array array = {
1329 		.nent = 0,
1330 	};
1331 	int r, i;
1332 
1333 	if (cpuid->nent < 1)
1334 		return -E2BIG;
1335 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1336 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1337 
1338 	if (sanity_check_entries(entries, cpuid->nent, type))
1339 		return -EINVAL;
1340 
1341 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1342 	if (!array.entries)
1343 		return -ENOMEM;
1344 
1345 	array.maxnent = cpuid->nent;
1346 
1347 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1348 		r = get_cpuid_func(&array, funcs[i], type);
1349 		if (r)
1350 			goto out_free;
1351 	}
1352 	cpuid->nent = array.nent;
1353 
1354 	if (copy_to_user(entries, array.entries,
1355 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1356 		r = -EFAULT;
1357 
1358 out_free:
1359 	kvfree(array.entries);
1360 	return r;
1361 }
1362 
1363 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1364 						    u32 function, u32 index)
1365 {
1366 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1367 				 function, index);
1368 }
1369 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1370 
1371 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1372 					      u32 function)
1373 {
1374 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1375 				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1376 }
1377 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1378 
1379 /*
1380  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1381  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1382  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1383  * range.  Centaur/VIA follows Intel semantics.
1384  *
1385  * A leaf is considered out-of-range if its function is higher than the maximum
1386  * supported leaf of its associated class or if its associated class does not
1387  * exist.
1388  *
1389  * There are three primary classes to be considered, with their respective
1390  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1391  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1392  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1393  *
1394  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1395  *  - Hypervisor: 0x40000000 - 0x4fffffff
1396  *  - Extended:   0x80000000 - 0xbfffffff
1397  *  - Centaur:    0xc0000000 - 0xcfffffff
1398  *
1399  * The Hypervisor class is further subdivided into sub-classes that each act as
1400  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1401  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1402  * CPUID sub-classes are:
1403  *
1404  *  - HyperV:     0x40000000 - 0x400000ff
1405  *  - KVM:        0x40000100 - 0x400001ff
1406  */
1407 static struct kvm_cpuid_entry2 *
1408 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1409 {
1410 	struct kvm_cpuid_entry2 *basic, *class;
1411 	u32 function = *fn_ptr;
1412 
1413 	basic = kvm_find_cpuid_entry(vcpu, 0);
1414 	if (!basic)
1415 		return NULL;
1416 
1417 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1418 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1419 		return NULL;
1420 
1421 	if (function >= 0x40000000 && function <= 0x4fffffff)
1422 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1423 	else if (function >= 0xc0000000)
1424 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1425 	else
1426 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1427 
1428 	if (class && function <= class->eax)
1429 		return NULL;
1430 
1431 	/*
1432 	 * Leaf specific adjustments are also applied when redirecting to the
1433 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1434 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1435 	 * needs to be pulled from CPUID.0xb.1.
1436 	 */
1437 	*fn_ptr = basic->eax;
1438 
1439 	/*
1440 	 * The class does not exist or the requested function is out of range;
1441 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1442 	 * the original requested leaf is observed!
1443 	 */
1444 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1445 }
1446 
1447 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1448 	       u32 *ecx, u32 *edx, bool exact_only)
1449 {
1450 	u32 orig_function = *eax, function = *eax, index = *ecx;
1451 	struct kvm_cpuid_entry2 *entry;
1452 	bool exact, used_max_basic = false;
1453 
1454 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1455 	exact = !!entry;
1456 
1457 	if (!entry && !exact_only) {
1458 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1459 		used_max_basic = !!entry;
1460 	}
1461 
1462 	if (entry) {
1463 		*eax = entry->eax;
1464 		*ebx = entry->ebx;
1465 		*ecx = entry->ecx;
1466 		*edx = entry->edx;
1467 		if (function == 7 && index == 0) {
1468 			u64 data;
1469 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1470 			    (data & TSX_CTRL_CPUID_CLEAR))
1471 				*ebx &= ~(F(RTM) | F(HLE));
1472 		}
1473 	} else {
1474 		*eax = *ebx = *ecx = *edx = 0;
1475 		/*
1476 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1477 		 * and EDX is always the x2APIC ID, even for undefined
1478 		 * subleaves. Index 1 will exist iff the leaf is
1479 		 * implemented, so we pass through CL iff leaf 1
1480 		 * exists. EDX can be copied from any existing index.
1481 		 */
1482 		if (function == 0xb || function == 0x1f) {
1483 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1484 			if (entry) {
1485 				*ecx = index & 0xff;
1486 				*edx = entry->edx;
1487 			}
1488 		}
1489 	}
1490 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1491 			used_max_basic);
1492 	return exact;
1493 }
1494 EXPORT_SYMBOL_GPL(kvm_cpuid);
1495 
1496 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1497 {
1498 	u32 eax, ebx, ecx, edx;
1499 
1500 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1501 		return 1;
1502 
1503 	eax = kvm_rax_read(vcpu);
1504 	ecx = kvm_rcx_read(vcpu);
1505 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1506 	kvm_rax_write(vcpu, eax);
1507 	kvm_rbx_write(vcpu, ebx);
1508 	kvm_rcx_write(vcpu, ecx);
1509 	kvm_rdx_write(vcpu, edx);
1510 	return kvm_skip_emulated_instruction(vcpu);
1511 }
1512 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1513