xref: /linux/arch/x86/kvm/cpuid.c (revision eb2f932100288dbb881eadfed02e1459c6b9504c)
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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 
13 #include <linux/kvm_host.h>
14 #include "linux/lockdep.h"
15 #include <linux/export.h>
16 #include <linux/vmalloc.h>
17 #include <linux/uaccess.h>
18 #include <linux/sched/stat.h>
19 
20 #include <asm/processor.h>
21 #include <asm/user.h>
22 #include <asm/fpu/xstate.h>
23 #include <asm/sgx.h>
24 #include <asm/cpuid.h>
25 #include "cpuid.h"
26 #include "lapic.h"
27 #include "mmu.h"
28 #include "trace.h"
29 #include "pmu.h"
30 #include "xen.h"
31 
32 /*
33  * Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be
34  * aligned to sizeof(unsigned long) because it's not accessed via bitops.
35  */
36 u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
37 EXPORT_SYMBOL_GPL(kvm_cpu_caps);
38 
39 u32 xstate_required_size(u64 xstate_bv, bool compacted)
40 {
41 	int feature_bit = 0;
42 	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
43 
44 	xstate_bv &= XFEATURE_MASK_EXTEND;
45 	while (xstate_bv) {
46 		if (xstate_bv & 0x1) {
47 		        u32 eax, ebx, ecx, edx, offset;
48 		        cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
49 			/* ECX[1]: 64B alignment in compacted form */
50 			if (compacted)
51 				offset = (ecx & 0x2) ? ALIGN(ret, 64) : ret;
52 			else
53 				offset = ebx;
54 			ret = max(ret, offset + eax);
55 		}
56 
57 		xstate_bv >>= 1;
58 		feature_bit++;
59 	}
60 
61 	return ret;
62 }
63 
64 #define F feature_bit
65 
66 /* Scattered Flag - For features that are scattered by cpufeatures.h. */
67 #define SF(name)						\
68 ({								\
69 	BUILD_BUG_ON(X86_FEATURE_##name >= MAX_CPU_FEATURES);	\
70 	(boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0);	\
71 })
72 
73 /*
74  * Magic value used by KVM when querying userspace-provided CPUID entries and
75  * doesn't care about the CPIUD index because the index of the function in
76  * question is not significant.  Note, this magic value must have at least one
77  * bit set in bits[63:32] and must be consumed as a u64 by cpuid_entry2_find()
78  * to avoid false positives when processing guest CPUID input.
79  */
80 #define KVM_CPUID_INDEX_NOT_SIGNIFICANT -1ull
81 
82 static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
83 	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u64 index)
84 {
85 	struct kvm_cpuid_entry2 *e;
86 	int i;
87 
88 	/*
89 	 * KVM has a semi-arbitrary rule that querying the guest's CPUID model
90 	 * with IRQs disabled is disallowed.  The CPUID model can legitimately
91 	 * have over one hundred entries, i.e. the lookup is slow, and IRQs are
92 	 * typically disabled in KVM only when KVM is in a performance critical
93 	 * path, e.g. the core VM-Enter/VM-Exit run loop.  Nothing will break
94 	 * if this rule is violated, this assertion is purely to flag potential
95 	 * performance issues.  If this fires, consider moving the lookup out
96 	 * of the hotpath, e.g. by caching information during CPUID updates.
97 	 */
98 	lockdep_assert_irqs_enabled();
99 
100 	for (i = 0; i < nent; i++) {
101 		e = &entries[i];
102 
103 		if (e->function != function)
104 			continue;
105 
106 		/*
107 		 * If the index isn't significant, use the first entry with a
108 		 * matching function.  It's userspace's responsibility to not
109 		 * provide "duplicate" entries in all cases.
110 		 */
111 		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
112 			return e;
113 
114 
115 		/*
116 		 * Similarly, use the first matching entry if KVM is doing a
117 		 * lookup (as opposed to emulating CPUID) for a function that's
118 		 * architecturally defined as not having a significant index.
119 		 */
120 		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
121 			/*
122 			 * Direct lookups from KVM should not diverge from what
123 			 * KVM defines internally (the architectural behavior).
124 			 */
125 			WARN_ON_ONCE(cpuid_function_is_indexed(function));
126 			return e;
127 		}
128 	}
129 
130 	return NULL;
131 }
132 
133 static int kvm_check_cpuid(struct kvm_vcpu *vcpu,
134 			   struct kvm_cpuid_entry2 *entries,
135 			   int nent)
136 {
137 	struct kvm_cpuid_entry2 *best;
138 	u64 xfeatures;
139 
140 	/*
141 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
142 	 * canonical address checks; exit if it is ever changed.
143 	 */
144 	best = cpuid_entry2_find(entries, nent, 0x80000008,
145 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
146 	if (best) {
147 		int vaddr_bits = (best->eax & 0xff00) >> 8;
148 
149 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
150 			return -EINVAL;
151 	}
152 
153 	/*
154 	 * Exposing dynamic xfeatures to the guest requires additional
155 	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
156 	 */
157 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
158 	if (!best)
159 		return 0;
160 
161 	xfeatures = best->eax | ((u64)best->edx << 32);
162 	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
163 	if (!xfeatures)
164 		return 0;
165 
166 	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
167 }
168 
169 /* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
170 static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
171 				 int nent)
172 {
173 	struct kvm_cpuid_entry2 *orig;
174 	int i;
175 
176 	if (nent != vcpu->arch.cpuid_nent)
177 		return -EINVAL;
178 
179 	for (i = 0; i < nent; i++) {
180 		orig = &vcpu->arch.cpuid_entries[i];
181 		if (e2[i].function != orig->function ||
182 		    e2[i].index != orig->index ||
183 		    e2[i].flags != orig->flags ||
184 		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
185 		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
186 			return -EINVAL;
187 	}
188 
189 	return 0;
190 }
191 
192 static struct kvm_hypervisor_cpuid kvm_get_hypervisor_cpuid(struct kvm_vcpu *vcpu,
193 							    const char *sig)
194 {
195 	struct kvm_hypervisor_cpuid cpuid = {};
196 	struct kvm_cpuid_entry2 *entry;
197 	u32 base;
198 
199 	for_each_possible_hypervisor_cpuid_base(base) {
200 		entry = kvm_find_cpuid_entry(vcpu, base);
201 
202 		if (entry) {
203 			u32 signature[3];
204 
205 			signature[0] = entry->ebx;
206 			signature[1] = entry->ecx;
207 			signature[2] = entry->edx;
208 
209 			if (!memcmp(signature, sig, sizeof(signature))) {
210 				cpuid.base = base;
211 				cpuid.limit = entry->eax;
212 				break;
213 			}
214 		}
215 	}
216 
217 	return cpuid;
218 }
219 
220 static struct kvm_cpuid_entry2 *__kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu,
221 					      struct kvm_cpuid_entry2 *entries, int nent)
222 {
223 	u32 base = vcpu->arch.kvm_cpuid.base;
224 
225 	if (!base)
226 		return NULL;
227 
228 	return cpuid_entry2_find(entries, nent, base | KVM_CPUID_FEATURES,
229 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
230 }
231 
232 static struct kvm_cpuid_entry2 *kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu)
233 {
234 	return __kvm_find_kvm_cpuid_features(vcpu, vcpu->arch.cpuid_entries,
235 					     vcpu->arch.cpuid_nent);
236 }
237 
238 void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
239 {
240 	struct kvm_cpuid_entry2 *best = kvm_find_kvm_cpuid_features(vcpu);
241 
242 	/*
243 	 * save the feature bitmap to avoid cpuid lookup for every PV
244 	 * operation
245 	 */
246 	if (best)
247 		vcpu->arch.pv_cpuid.features = best->eax;
248 }
249 
250 /*
251  * Calculate guest's supported XCR0 taking into account guest CPUID data and
252  * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
253  */
254 static u64 cpuid_get_supported_xcr0(struct kvm_cpuid_entry2 *entries, int nent)
255 {
256 	struct kvm_cpuid_entry2 *best;
257 
258 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
259 	if (!best)
260 		return 0;
261 
262 	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
263 }
264 
265 static void __kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *entries,
266 				       int nent)
267 {
268 	struct kvm_cpuid_entry2 *best;
269 
270 	best = cpuid_entry2_find(entries, nent, 1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
271 	if (best) {
272 		/* Update OSXSAVE bit */
273 		if (boot_cpu_has(X86_FEATURE_XSAVE))
274 			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
275 					   kvm_is_cr4_bit_set(vcpu, X86_CR4_OSXSAVE));
276 
277 		cpuid_entry_change(best, X86_FEATURE_APIC,
278 			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
279 	}
280 
281 	best = cpuid_entry2_find(entries, nent, 7, 0);
282 	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
283 		cpuid_entry_change(best, X86_FEATURE_OSPKE,
284 				   kvm_is_cr4_bit_set(vcpu, X86_CR4_PKE));
285 
286 	best = cpuid_entry2_find(entries, nent, 0xD, 0);
287 	if (best)
288 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
289 
290 	best = cpuid_entry2_find(entries, nent, 0xD, 1);
291 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
292 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
293 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
294 
295 	best = __kvm_find_kvm_cpuid_features(vcpu, entries, nent);
296 	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
297 		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
298 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
299 
300 	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
301 		best = cpuid_entry2_find(entries, nent, 0x1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
302 		if (best)
303 			cpuid_entry_change(best, X86_FEATURE_MWAIT,
304 					   vcpu->arch.ia32_misc_enable_msr &
305 					   MSR_IA32_MISC_ENABLE_MWAIT);
306 	}
307 }
308 
309 void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
310 {
311 	__kvm_update_cpuid_runtime(vcpu, vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
312 }
313 EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
314 
315 static bool kvm_cpuid_has_hyperv(struct kvm_cpuid_entry2 *entries, int nent)
316 {
317 #ifdef CONFIG_KVM_HYPERV
318 	struct kvm_cpuid_entry2 *entry;
319 
320 	entry = cpuid_entry2_find(entries, nent, HYPERV_CPUID_INTERFACE,
321 				  KVM_CPUID_INDEX_NOT_SIGNIFICANT);
322 	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
323 #else
324 	return false;
325 #endif
326 }
327 
328 static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
329 {
330 	struct kvm_lapic *apic = vcpu->arch.apic;
331 	struct kvm_cpuid_entry2 *best;
332 	bool allow_gbpages;
333 
334 	BUILD_BUG_ON(KVM_NR_GOVERNED_FEATURES > KVM_MAX_NR_GOVERNED_FEATURES);
335 	bitmap_zero(vcpu->arch.governed_features.enabled,
336 		    KVM_MAX_NR_GOVERNED_FEATURES);
337 
338 	/*
339 	 * If TDP is enabled, let the guest use GBPAGES if they're supported in
340 	 * hardware.  The hardware page walker doesn't let KVM disable GBPAGES,
341 	 * i.e. won't treat them as reserved, and KVM doesn't redo the GVA->GPA
342 	 * walk for performance and complexity reasons.  Not to mention KVM
343 	 * _can't_ solve the problem because GVA->GPA walks aren't visible to
344 	 * KVM once a TDP translation is installed.  Mimic hardware behavior so
345 	 * that KVM's is at least consistent, i.e. doesn't randomly inject #PF.
346 	 * If TDP is disabled, honor *only* guest CPUID as KVM has full control
347 	 * and can install smaller shadow pages if the host lacks 1GiB support.
348 	 */
349 	allow_gbpages = tdp_enabled ? boot_cpu_has(X86_FEATURE_GBPAGES) :
350 				      guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES);
351 	if (allow_gbpages)
352 		kvm_governed_feature_set(vcpu, X86_FEATURE_GBPAGES);
353 
354 	best = kvm_find_cpuid_entry(vcpu, 1);
355 	if (best && apic) {
356 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
357 			apic->lapic_timer.timer_mode_mask = 3 << 17;
358 		else
359 			apic->lapic_timer.timer_mode_mask = 1 << 17;
360 
361 		kvm_apic_set_version(vcpu);
362 	}
363 
364 	vcpu->arch.guest_supported_xcr0 =
365 		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
366 
367 	kvm_update_pv_runtime(vcpu);
368 
369 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
370 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
371 
372 	kvm_pmu_refresh(vcpu);
373 	vcpu->arch.cr4_guest_rsvd_bits =
374 	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
375 
376 	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu->arch.cpuid_entries,
377 						    vcpu->arch.cpuid_nent));
378 
379 	/* Invoke the vendor callback only after the above state is updated. */
380 	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
381 
382 	/*
383 	 * Except for the MMU, which needs to do its thing any vendor specific
384 	 * adjustments to the reserved GPA bits.
385 	 */
386 	kvm_mmu_after_set_cpuid(vcpu);
387 }
388 
389 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
390 {
391 	struct kvm_cpuid_entry2 *best;
392 
393 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
394 	if (!best || best->eax < 0x80000008)
395 		goto not_found;
396 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
397 	if (best)
398 		return best->eax & 0xff;
399 not_found:
400 	return 36;
401 }
402 
403 /*
404  * This "raw" version returns the reserved GPA bits without any adjustments for
405  * encryption technologies that usurp bits.  The raw mask should be used if and
406  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
407  */
408 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
409 {
410 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
411 }
412 
413 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
414                         int nent)
415 {
416 	int r;
417 
418 	__kvm_update_cpuid_runtime(vcpu, e2, nent);
419 
420 	/*
421 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
422 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
423 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
424 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
425 	 * the core vCPU model on the fly. It would've been better to forbid any
426 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
427 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
428 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
429 	 * whether the supplied CPUID data is equal to what's already set.
430 	 */
431 	if (kvm_vcpu_has_run(vcpu)) {
432 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
433 		if (r)
434 			return r;
435 
436 		kvfree(e2);
437 		return 0;
438 	}
439 
440 #ifdef CONFIG_KVM_HYPERV
441 	if (kvm_cpuid_has_hyperv(e2, nent)) {
442 		r = kvm_hv_vcpu_init(vcpu);
443 		if (r)
444 			return r;
445 	}
446 #endif
447 
448 	r = kvm_check_cpuid(vcpu, e2, nent);
449 	if (r)
450 		return r;
451 
452 	kvfree(vcpu->arch.cpuid_entries);
453 	vcpu->arch.cpuid_entries = e2;
454 	vcpu->arch.cpuid_nent = nent;
455 
456 	vcpu->arch.kvm_cpuid = kvm_get_hypervisor_cpuid(vcpu, KVM_SIGNATURE);
457 #ifdef CONFIG_KVM_XEN
458 	vcpu->arch.xen.cpuid = kvm_get_hypervisor_cpuid(vcpu, XEN_SIGNATURE);
459 #endif
460 	kvm_vcpu_after_set_cpuid(vcpu);
461 
462 	return 0;
463 }
464 
465 /* when an old userspace process fills a new kernel module */
466 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
467 			     struct kvm_cpuid *cpuid,
468 			     struct kvm_cpuid_entry __user *entries)
469 {
470 	int r, i;
471 	struct kvm_cpuid_entry *e = NULL;
472 	struct kvm_cpuid_entry2 *e2 = NULL;
473 
474 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
475 		return -E2BIG;
476 
477 	if (cpuid->nent) {
478 		e = vmemdup_array_user(entries, cpuid->nent, sizeof(*e));
479 		if (IS_ERR(e))
480 			return PTR_ERR(e);
481 
482 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
483 		if (!e2) {
484 			r = -ENOMEM;
485 			goto out_free_cpuid;
486 		}
487 	}
488 	for (i = 0; i < cpuid->nent; i++) {
489 		e2[i].function = e[i].function;
490 		e2[i].eax = e[i].eax;
491 		e2[i].ebx = e[i].ebx;
492 		e2[i].ecx = e[i].ecx;
493 		e2[i].edx = e[i].edx;
494 		e2[i].index = 0;
495 		e2[i].flags = 0;
496 		e2[i].padding[0] = 0;
497 		e2[i].padding[1] = 0;
498 		e2[i].padding[2] = 0;
499 	}
500 
501 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
502 	if (r)
503 		kvfree(e2);
504 
505 out_free_cpuid:
506 	kvfree(e);
507 
508 	return r;
509 }
510 
511 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
512 			      struct kvm_cpuid2 *cpuid,
513 			      struct kvm_cpuid_entry2 __user *entries)
514 {
515 	struct kvm_cpuid_entry2 *e2 = NULL;
516 	int r;
517 
518 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
519 		return -E2BIG;
520 
521 	if (cpuid->nent) {
522 		e2 = vmemdup_array_user(entries, cpuid->nent, sizeof(*e2));
523 		if (IS_ERR(e2))
524 			return PTR_ERR(e2);
525 	}
526 
527 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
528 	if (r)
529 		kvfree(e2);
530 
531 	return r;
532 }
533 
534 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
535 			      struct kvm_cpuid2 *cpuid,
536 			      struct kvm_cpuid_entry2 __user *entries)
537 {
538 	if (cpuid->nent < vcpu->arch.cpuid_nent)
539 		return -E2BIG;
540 
541 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
542 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
543 		return -EFAULT;
544 
545 	cpuid->nent = vcpu->arch.cpuid_nent;
546 	return 0;
547 }
548 
549 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
550 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
551 {
552 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
553 	struct kvm_cpuid_entry2 entry;
554 
555 	reverse_cpuid_check(leaf);
556 
557 	cpuid_count(cpuid.function, cpuid.index,
558 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
559 
560 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
561 }
562 
563 static __always_inline
564 void kvm_cpu_cap_init_kvm_defined(enum kvm_only_cpuid_leafs leaf, u32 mask)
565 {
566 	/* Use kvm_cpu_cap_mask for leafs that aren't KVM-only. */
567 	BUILD_BUG_ON(leaf < NCAPINTS);
568 
569 	kvm_cpu_caps[leaf] = mask;
570 
571 	__kvm_cpu_cap_mask(leaf);
572 }
573 
574 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
575 {
576 	/* Use kvm_cpu_cap_init_kvm_defined for KVM-only leafs. */
577 	BUILD_BUG_ON(leaf >= NCAPINTS);
578 
579 	kvm_cpu_caps[leaf] &= mask;
580 
581 	__kvm_cpu_cap_mask(leaf);
582 }
583 
584 void kvm_set_cpu_caps(void)
585 {
586 #ifdef CONFIG_X86_64
587 	unsigned int f_gbpages = F(GBPAGES);
588 	unsigned int f_lm = F(LM);
589 	unsigned int f_xfd = F(XFD);
590 #else
591 	unsigned int f_gbpages = 0;
592 	unsigned int f_lm = 0;
593 	unsigned int f_xfd = 0;
594 #endif
595 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
596 
597 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
598 		     sizeof(boot_cpu_data.x86_capability));
599 
600 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
601 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
602 
603 	kvm_cpu_cap_mask(CPUID_1_ECX,
604 		/*
605 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
606 		 * advertised to guests via CPUID!
607 		 */
608 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
609 		0 /* DS-CPL, VMX, SMX, EST */ |
610 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
611 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
612 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
613 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
614 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
615 		F(F16C) | F(RDRAND)
616 	);
617 	/* KVM emulates x2apic in software irrespective of host support. */
618 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
619 
620 	kvm_cpu_cap_mask(CPUID_1_EDX,
621 		F(FPU) | F(VME) | F(DE) | F(PSE) |
622 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
623 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
624 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
625 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
626 		0 /* Reserved, DS, ACPI */ | F(MMX) |
627 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
628 		0 /* HTT, TM, Reserved, PBE */
629 	);
630 
631 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
632 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
633 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
634 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
635 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
636 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
637 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
638 		F(AVX512VL));
639 
640 	kvm_cpu_cap_mask(CPUID_7_ECX,
641 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
642 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
643 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
644 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
645 		F(SGX_LC) | F(BUS_LOCK_DETECT)
646 	);
647 	/* Set LA57 based on hardware capability. */
648 	if (cpuid_ecx(7) & F(LA57))
649 		kvm_cpu_cap_set(X86_FEATURE_LA57);
650 
651 	/*
652 	 * PKU not yet implemented for shadow paging and requires OSPKE
653 	 * to be set on the host. Clear it if that is not the case
654 	 */
655 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
656 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
657 
658 	kvm_cpu_cap_mask(CPUID_7_EDX,
659 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
660 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
661 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
662 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
663 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16) | F(FLUSH_L1D)
664 	);
665 
666 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
667 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
668 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
669 
670 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
671 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
672 	if (boot_cpu_has(X86_FEATURE_STIBP))
673 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
674 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
675 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
676 
677 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
678 		F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) |
679 		F(FZRM) | F(FSRS) | F(FSRC) |
680 		F(AMX_FP16) | F(AVX_IFMA) | F(LAM)
681 	);
682 
683 	kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
684 		F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(PREFETCHITI) |
685 		F(AMX_COMPLEX)
686 	);
687 
688 	kvm_cpu_cap_init_kvm_defined(CPUID_7_2_EDX,
689 		F(INTEL_PSFD) | F(IPRED_CTRL) | F(RRSBA_CTRL) | F(DDPD_U) |
690 		F(BHI_CTRL) | F(MCDT_NO)
691 	);
692 
693 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
694 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
695 	);
696 
697 	kvm_cpu_cap_init_kvm_defined(CPUID_12_EAX,
698 		SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
699 	);
700 
701 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
702 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
703 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
704 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
705 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
706 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
707 	);
708 
709 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
710 		F(FPU) | F(VME) | F(DE) | F(PSE) |
711 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
712 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
713 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
714 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
715 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
716 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
717 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
718 	);
719 
720 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
721 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
722 
723 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0007_EDX,
724 		SF(CONSTANT_TSC)
725 	);
726 
727 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
728 		F(CLZERO) | F(XSAVEERPTR) |
729 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
730 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
731 		F(AMD_PSFD)
732 	);
733 
734 	/*
735 	 * AMD has separate bits for each SPEC_CTRL bit.
736 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
737 	 * record that in cpufeatures so use them.
738 	 */
739 	if (boot_cpu_has(X86_FEATURE_IBPB))
740 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
741 	if (boot_cpu_has(X86_FEATURE_IBRS))
742 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
743 	if (boot_cpu_has(X86_FEATURE_STIBP))
744 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
745 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
746 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
747 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
748 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
749 	/*
750 	 * The preference is to use SPEC CTRL MSR instead of the
751 	 * VIRT_SPEC MSR.
752 	 */
753 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
754 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
755 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
756 
757 	/*
758 	 * Hide all SVM features by default, SVM will set the cap bits for
759 	 * features it emulates and/or exposes for L1.
760 	 */
761 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
762 
763 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
764 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
765 		F(SME_COHERENT));
766 
767 	kvm_cpu_cap_mask(CPUID_8000_0021_EAX,
768 		F(NO_NESTED_DATA_BP) | F(LFENCE_RDTSC) | 0 /* SmmPgCfgLock */ |
769 		F(NULL_SEL_CLR_BASE) | F(AUTOIBRS) | 0 /* PrefetchCtlMsr */ |
770 		F(WRMSR_XX_BASE_NS)
771 	);
772 
773 	kvm_cpu_cap_check_and_set(X86_FEATURE_SBPB);
774 	kvm_cpu_cap_check_and_set(X86_FEATURE_IBPB_BRTYPE);
775 	kvm_cpu_cap_check_and_set(X86_FEATURE_SRSO_NO);
776 
777 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0022_EAX,
778 		F(PERFMON_V2)
779 	);
780 
781 	/*
782 	 * Synthesize "LFENCE is serializing" into the AMD-defined entry in
783 	 * KVM's supported CPUID if the feature is reported as supported by the
784 	 * kernel.  LFENCE_RDTSC was a Linux-defined synthetic feature long
785 	 * before AMD joined the bandwagon, e.g. LFENCE is serializing on most
786 	 * CPUs that support SSE2.  On CPUs that don't support AMD's leaf,
787 	 * kvm_cpu_cap_mask() will unfortunately drop the flag due to ANDing
788 	 * the mask with the raw host CPUID, and reporting support in AMD's
789 	 * leaf can make it easier for userspace to detect the feature.
790 	 */
791 	if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
792 		kvm_cpu_cap_set(X86_FEATURE_LFENCE_RDTSC);
793 	if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
794 		kvm_cpu_cap_set(X86_FEATURE_NULL_SEL_CLR_BASE);
795 	kvm_cpu_cap_set(X86_FEATURE_NO_SMM_CTL_MSR);
796 
797 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
798 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
799 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
800 		F(PMM) | F(PMM_EN)
801 	);
802 
803 	/*
804 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
805 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
806 	 * should never happen, but the guest will likely crash if RDTSCP or
807 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
808 	 * the past.  For example, the sanity check may fire if this instance of
809 	 * KVM is running as L1 on top of an older, broken KVM.
810 	 */
811 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
812 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
813 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
814 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
815 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
816 	}
817 }
818 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
819 
820 struct kvm_cpuid_array {
821 	struct kvm_cpuid_entry2 *entries;
822 	int maxnent;
823 	int nent;
824 };
825 
826 static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
827 {
828 	if (array->nent >= array->maxnent)
829 		return NULL;
830 
831 	return &array->entries[array->nent++];
832 }
833 
834 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
835 					      u32 function, u32 index)
836 {
837 	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
838 
839 	if (!entry)
840 		return NULL;
841 
842 	memset(entry, 0, sizeof(*entry));
843 	entry->function = function;
844 	entry->index = index;
845 	switch (function & 0xC0000000) {
846 	case 0x40000000:
847 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
848 		return entry;
849 
850 	case 0x80000000:
851 		/*
852 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
853 		 * would result in out-of-bounds calls to do_host_cpuid.
854 		 */
855 		{
856 			static int max_cpuid_80000000;
857 			if (!READ_ONCE(max_cpuid_80000000))
858 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
859 			if (function > READ_ONCE(max_cpuid_80000000))
860 				return entry;
861 		}
862 		break;
863 
864 	default:
865 		break;
866 	}
867 
868 	cpuid_count(entry->function, entry->index,
869 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
870 
871 	if (cpuid_function_is_indexed(function))
872 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
873 
874 	return entry;
875 }
876 
877 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
878 {
879 	struct kvm_cpuid_entry2 *entry;
880 
881 	if (array->nent >= array->maxnent)
882 		return -E2BIG;
883 
884 	entry = &array->entries[array->nent];
885 	entry->function = func;
886 	entry->index = 0;
887 	entry->flags = 0;
888 
889 	switch (func) {
890 	case 0:
891 		entry->eax = 7;
892 		++array->nent;
893 		break;
894 	case 1:
895 		entry->ecx = F(MOVBE);
896 		++array->nent;
897 		break;
898 	case 7:
899 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
900 		entry->eax = 0;
901 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
902 			entry->ecx = F(RDPID);
903 		++array->nent;
904 		break;
905 	default:
906 		break;
907 	}
908 
909 	return 0;
910 }
911 
912 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
913 {
914 	struct kvm_cpuid_entry2 *entry;
915 	int r, i, max_idx;
916 
917 	/* all calls to cpuid_count() should be made on the same cpu */
918 	get_cpu();
919 
920 	r = -E2BIG;
921 
922 	entry = do_host_cpuid(array, function, 0);
923 	if (!entry)
924 		goto out;
925 
926 	switch (function) {
927 	case 0:
928 		/* Limited to the highest leaf implemented in KVM. */
929 		entry->eax = min(entry->eax, 0x1fU);
930 		break;
931 	case 1:
932 		cpuid_entry_override(entry, CPUID_1_EDX);
933 		cpuid_entry_override(entry, CPUID_1_ECX);
934 		break;
935 	case 2:
936 		/*
937 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
938 		 * CPUID(function=2, index=0) may return different results each
939 		 * time, with the least-significant byte in EAX enumerating the
940 		 * number of times software should do CPUID(2, 0).
941 		 *
942 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
943 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
944 		 * return 01H. Software should ignore this value and not
945 		 * interpret it as an informational descriptor", while AMD's
946 		 * APM states that CPUID(2) is reserved.
947 		 *
948 		 * WARN if a frankenstein CPU that supports virtualization and
949 		 * a stateful CPUID.0x2 is encountered.
950 		 */
951 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
952 		break;
953 	/* functions 4 and 0x8000001d have additional index. */
954 	case 4:
955 	case 0x8000001d:
956 		/*
957 		 * Read entries until the cache type in the previous entry is
958 		 * zero, i.e. indicates an invalid entry.
959 		 */
960 		for (i = 1; entry->eax & 0x1f; ++i) {
961 			entry = do_host_cpuid(array, function, i);
962 			if (!entry)
963 				goto out;
964 		}
965 		break;
966 	case 6: /* Thermal management */
967 		entry->eax = 0x4; /* allow ARAT */
968 		entry->ebx = 0;
969 		entry->ecx = 0;
970 		entry->edx = 0;
971 		break;
972 	/* function 7 has additional index. */
973 	case 7:
974 		max_idx = entry->eax = min(entry->eax, 2u);
975 		cpuid_entry_override(entry, CPUID_7_0_EBX);
976 		cpuid_entry_override(entry, CPUID_7_ECX);
977 		cpuid_entry_override(entry, CPUID_7_EDX);
978 
979 		/* KVM only supports up to 0x7.2, capped above via min(). */
980 		if (max_idx >= 1) {
981 			entry = do_host_cpuid(array, function, 1);
982 			if (!entry)
983 				goto out;
984 
985 			cpuid_entry_override(entry, CPUID_7_1_EAX);
986 			cpuid_entry_override(entry, CPUID_7_1_EDX);
987 			entry->ebx = 0;
988 			entry->ecx = 0;
989 		}
990 		if (max_idx >= 2) {
991 			entry = do_host_cpuid(array, function, 2);
992 			if (!entry)
993 				goto out;
994 
995 			cpuid_entry_override(entry, CPUID_7_2_EDX);
996 			entry->ecx = 0;
997 			entry->ebx = 0;
998 			entry->eax = 0;
999 		}
1000 		break;
1001 	case 0xa: { /* Architectural Performance Monitoring */
1002 		union cpuid10_eax eax;
1003 		union cpuid10_edx edx;
1004 
1005 		if (!enable_pmu || !static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
1006 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1007 			break;
1008 		}
1009 
1010 		eax.split.version_id = kvm_pmu_cap.version;
1011 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
1012 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
1013 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
1014 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
1015 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
1016 
1017 		if (kvm_pmu_cap.version)
1018 			edx.split.anythread_deprecated = 1;
1019 		edx.split.reserved1 = 0;
1020 		edx.split.reserved2 = 0;
1021 
1022 		entry->eax = eax.full;
1023 		entry->ebx = kvm_pmu_cap.events_mask;
1024 		entry->ecx = 0;
1025 		entry->edx = edx.full;
1026 		break;
1027 	}
1028 	case 0x1f:
1029 	case 0xb:
1030 		/*
1031 		 * No topology; a valid topology is indicated by the presence
1032 		 * of subleaf 1.
1033 		 */
1034 		entry->eax = entry->ebx = entry->ecx = 0;
1035 		break;
1036 	case 0xd: {
1037 		u64 permitted_xcr0 = kvm_get_filtered_xcr0();
1038 		u64 permitted_xss = kvm_caps.supported_xss;
1039 
1040 		entry->eax &= permitted_xcr0;
1041 		entry->ebx = xstate_required_size(permitted_xcr0, false);
1042 		entry->ecx = entry->ebx;
1043 		entry->edx &= permitted_xcr0 >> 32;
1044 		if (!permitted_xcr0)
1045 			break;
1046 
1047 		entry = do_host_cpuid(array, function, 1);
1048 		if (!entry)
1049 			goto out;
1050 
1051 		cpuid_entry_override(entry, CPUID_D_1_EAX);
1052 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
1053 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
1054 							  true);
1055 		else {
1056 			WARN_ON_ONCE(permitted_xss != 0);
1057 			entry->ebx = 0;
1058 		}
1059 		entry->ecx &= permitted_xss;
1060 		entry->edx &= permitted_xss >> 32;
1061 
1062 		for (i = 2; i < 64; ++i) {
1063 			bool s_state;
1064 			if (permitted_xcr0 & BIT_ULL(i))
1065 				s_state = false;
1066 			else if (permitted_xss & BIT_ULL(i))
1067 				s_state = true;
1068 			else
1069 				continue;
1070 
1071 			entry = do_host_cpuid(array, function, i);
1072 			if (!entry)
1073 				goto out;
1074 
1075 			/*
1076 			 * The supported check above should have filtered out
1077 			 * invalid sub-leafs.  Only valid sub-leafs should
1078 			 * reach this point, and they should have a non-zero
1079 			 * save state size.  Furthermore, check whether the
1080 			 * processor agrees with permitted_xcr0/permitted_xss
1081 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1082 			 */
1083 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1084 				--array->nent;
1085 				continue;
1086 			}
1087 
1088 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1089 				entry->ecx &= ~BIT_ULL(2);
1090 			entry->edx = 0;
1091 		}
1092 		break;
1093 	}
1094 	case 0x12:
1095 		/* Intel SGX */
1096 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1097 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1098 			break;
1099 		}
1100 
1101 		/*
1102 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1103 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1104 		 * are restricted by kernel and KVM capabilities (like most
1105 		 * feature flags), while enclave size is unrestricted.
1106 		 */
1107 		cpuid_entry_override(entry, CPUID_12_EAX);
1108 		entry->ebx &= SGX_MISC_EXINFO;
1109 
1110 		entry = do_host_cpuid(array, function, 1);
1111 		if (!entry)
1112 			goto out;
1113 
1114 		/*
1115 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1116 		 * feature flags.  Advertise all supported flags, including
1117 		 * privileged attributes that require explicit opt-in from
1118 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1119 		 * expected to derive it from supported XCR0.
1120 		 */
1121 		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1122 		entry->ebx &= 0;
1123 		break;
1124 	/* Intel PT */
1125 	case 0x14:
1126 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1127 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1128 			break;
1129 		}
1130 
1131 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1132 			if (!do_host_cpuid(array, function, i))
1133 				goto out;
1134 		}
1135 		break;
1136 	/* Intel AMX TILE */
1137 	case 0x1d:
1138 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1139 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1140 			break;
1141 		}
1142 
1143 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1144 			if (!do_host_cpuid(array, function, i))
1145 				goto out;
1146 		}
1147 		break;
1148 	case 0x1e: /* TMUL information */
1149 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1150 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1151 			break;
1152 		}
1153 		break;
1154 	case KVM_CPUID_SIGNATURE: {
1155 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1156 		entry->eax = KVM_CPUID_FEATURES;
1157 		entry->ebx = sigptr[0];
1158 		entry->ecx = sigptr[1];
1159 		entry->edx = sigptr[2];
1160 		break;
1161 	}
1162 	case KVM_CPUID_FEATURES:
1163 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1164 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1165 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1166 			     (1 << KVM_FEATURE_ASYNC_PF) |
1167 			     (1 << KVM_FEATURE_PV_EOI) |
1168 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1169 			     (1 << KVM_FEATURE_PV_UNHALT) |
1170 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1171 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1172 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1173 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1174 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1175 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1176 
1177 		if (sched_info_on())
1178 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1179 
1180 		entry->ebx = 0;
1181 		entry->ecx = 0;
1182 		entry->edx = 0;
1183 		break;
1184 	case 0x80000000:
1185 		entry->eax = min(entry->eax, 0x80000022);
1186 		/*
1187 		 * Serializing LFENCE is reported in a multitude of ways, and
1188 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1189 		 * userspace by providing the CPUID leaf ourselves.
1190 		 *
1191 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1192 		 * QEMU thinks that it can query the host blindly for that
1193 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1194 		 * above.  The processor merrily returns values from the
1195 		 * highest Intel leaf which QEMU tries to use as the guest's
1196 		 * 0x8000001d.  Even worse, this can result in an infinite
1197 		 * loop if said highest leaf has no subleaves indexed by ECX.
1198 		 */
1199 		if (entry->eax >= 0x8000001d &&
1200 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1201 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1202 			entry->eax = max(entry->eax, 0x80000021);
1203 		break;
1204 	case 0x80000001:
1205 		entry->ebx &= ~GENMASK(27, 16);
1206 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1207 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1208 		break;
1209 	case 0x80000005:
1210 		/*  Pass host L1 cache and TLB info. */
1211 		break;
1212 	case 0x80000006:
1213 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1214 		entry->edx &= ~GENMASK(17, 16);
1215 		break;
1216 	case 0x80000007: /* Advanced power management */
1217 		cpuid_entry_override(entry, CPUID_8000_0007_EDX);
1218 
1219 		/* mask against host */
1220 		entry->edx &= boot_cpu_data.x86_power;
1221 		entry->eax = entry->ebx = entry->ecx = 0;
1222 		break;
1223 	case 0x80000008: {
1224 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1225 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1226 		unsigned phys_as = entry->eax & 0xff;
1227 
1228 		/*
1229 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1230 		 * the guest operates in the same PA space as the host, i.e.
1231 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1232 		 * paging, too.
1233 		 *
1234 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1235 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1236 		 * the HPAs do not affect GPAs.
1237 		 */
1238 		if (!tdp_enabled)
1239 			g_phys_as = boot_cpu_data.x86_phys_bits;
1240 		else if (!g_phys_as)
1241 			g_phys_as = phys_as;
1242 
1243 		entry->eax = g_phys_as | (virt_as << 8);
1244 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1245 		entry->edx = 0;
1246 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1247 		break;
1248 	}
1249 	case 0x8000000A:
1250 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1251 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1252 			break;
1253 		}
1254 		entry->eax = 1; /* SVM revision 1 */
1255 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1256 				   ASID emulation to nested SVM */
1257 		entry->ecx = 0; /* Reserved */
1258 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1259 		break;
1260 	case 0x80000019:
1261 		entry->ecx = entry->edx = 0;
1262 		break;
1263 	case 0x8000001a:
1264 		entry->eax &= GENMASK(2, 0);
1265 		entry->ebx = entry->ecx = entry->edx = 0;
1266 		break;
1267 	case 0x8000001e:
1268 		/* Do not return host topology information.  */
1269 		entry->eax = entry->ebx = entry->ecx = 0;
1270 		entry->edx = 0; /* reserved */
1271 		break;
1272 	case 0x8000001F:
1273 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1274 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1275 		} else {
1276 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1277 			/* Clear NumVMPL since KVM does not support VMPL.  */
1278 			entry->ebx &= ~GENMASK(31, 12);
1279 			/*
1280 			 * Enumerate '0' for "PA bits reduction", the adjusted
1281 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1282 			 */
1283 			entry->ebx &= ~GENMASK(11, 6);
1284 		}
1285 		break;
1286 	case 0x80000020:
1287 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1288 		break;
1289 	case 0x80000021:
1290 		entry->ebx = entry->ecx = entry->edx = 0;
1291 		cpuid_entry_override(entry, CPUID_8000_0021_EAX);
1292 		break;
1293 	/* AMD Extended Performance Monitoring and Debug */
1294 	case 0x80000022: {
1295 		union cpuid_0x80000022_ebx ebx;
1296 
1297 		entry->ecx = entry->edx = 0;
1298 		if (!enable_pmu || !kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2)) {
1299 			entry->eax = entry->ebx;
1300 			break;
1301 		}
1302 
1303 		cpuid_entry_override(entry, CPUID_8000_0022_EAX);
1304 
1305 		if (kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2))
1306 			ebx.split.num_core_pmc = kvm_pmu_cap.num_counters_gp;
1307 		else if (kvm_cpu_cap_has(X86_FEATURE_PERFCTR_CORE))
1308 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS_CORE;
1309 		else
1310 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS;
1311 
1312 		entry->ebx = ebx.full;
1313 		break;
1314 	}
1315 	/*Add support for Centaur's CPUID instruction*/
1316 	case 0xC0000000:
1317 		/*Just support up to 0xC0000004 now*/
1318 		entry->eax = min(entry->eax, 0xC0000004);
1319 		break;
1320 	case 0xC0000001:
1321 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1322 		break;
1323 	case 3: /* Processor serial number */
1324 	case 5: /* MONITOR/MWAIT */
1325 	case 0xC0000002:
1326 	case 0xC0000003:
1327 	case 0xC0000004:
1328 	default:
1329 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1330 		break;
1331 	}
1332 
1333 	r = 0;
1334 
1335 out:
1336 	put_cpu();
1337 
1338 	return r;
1339 }
1340 
1341 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1342 			 unsigned int type)
1343 {
1344 	if (type == KVM_GET_EMULATED_CPUID)
1345 		return __do_cpuid_func_emulated(array, func);
1346 
1347 	return __do_cpuid_func(array, func);
1348 }
1349 
1350 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1351 
1352 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1353 			  unsigned int type)
1354 {
1355 	u32 limit;
1356 	int r;
1357 
1358 	if (func == CENTAUR_CPUID_SIGNATURE &&
1359 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1360 		return 0;
1361 
1362 	r = do_cpuid_func(array, func, type);
1363 	if (r)
1364 		return r;
1365 
1366 	limit = array->entries[array->nent - 1].eax;
1367 	for (func = func + 1; func <= limit; ++func) {
1368 		r = do_cpuid_func(array, func, type);
1369 		if (r)
1370 			break;
1371 	}
1372 
1373 	return r;
1374 }
1375 
1376 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1377 				 __u32 num_entries, unsigned int ioctl_type)
1378 {
1379 	int i;
1380 	__u32 pad[3];
1381 
1382 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1383 		return false;
1384 
1385 	/*
1386 	 * We want to make sure that ->padding is being passed clean from
1387 	 * userspace in case we want to use it for something in the future.
1388 	 *
1389 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1390 	 * have to give ourselves satisfied only with the emulated side. /me
1391 	 * sheds a tear.
1392 	 */
1393 	for (i = 0; i < num_entries; i++) {
1394 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1395 			return true;
1396 
1397 		if (pad[0] || pad[1] || pad[2])
1398 			return true;
1399 	}
1400 	return false;
1401 }
1402 
1403 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1404 			    struct kvm_cpuid_entry2 __user *entries,
1405 			    unsigned int type)
1406 {
1407 	static const u32 funcs[] = {
1408 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1409 	};
1410 
1411 	struct kvm_cpuid_array array = {
1412 		.nent = 0,
1413 	};
1414 	int r, i;
1415 
1416 	if (cpuid->nent < 1)
1417 		return -E2BIG;
1418 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1419 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1420 
1421 	if (sanity_check_entries(entries, cpuid->nent, type))
1422 		return -EINVAL;
1423 
1424 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1425 	if (!array.entries)
1426 		return -ENOMEM;
1427 
1428 	array.maxnent = cpuid->nent;
1429 
1430 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1431 		r = get_cpuid_func(&array, funcs[i], type);
1432 		if (r)
1433 			goto out_free;
1434 	}
1435 	cpuid->nent = array.nent;
1436 
1437 	if (copy_to_user(entries, array.entries,
1438 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1439 		r = -EFAULT;
1440 
1441 out_free:
1442 	kvfree(array.entries);
1443 	return r;
1444 }
1445 
1446 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1447 						    u32 function, u32 index)
1448 {
1449 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1450 				 function, index);
1451 }
1452 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1453 
1454 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1455 					      u32 function)
1456 {
1457 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1458 				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1459 }
1460 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1461 
1462 /*
1463  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1464  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1465  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1466  * range.  Centaur/VIA follows Intel semantics.
1467  *
1468  * A leaf is considered out-of-range if its function is higher than the maximum
1469  * supported leaf of its associated class or if its associated class does not
1470  * exist.
1471  *
1472  * There are three primary classes to be considered, with their respective
1473  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1474  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1475  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1476  *
1477  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1478  *  - Hypervisor: 0x40000000 - 0x4fffffff
1479  *  - Extended:   0x80000000 - 0xbfffffff
1480  *  - Centaur:    0xc0000000 - 0xcfffffff
1481  *
1482  * The Hypervisor class is further subdivided into sub-classes that each act as
1483  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1484  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1485  * CPUID sub-classes are:
1486  *
1487  *  - HyperV:     0x40000000 - 0x400000ff
1488  *  - KVM:        0x40000100 - 0x400001ff
1489  */
1490 static struct kvm_cpuid_entry2 *
1491 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1492 {
1493 	struct kvm_cpuid_entry2 *basic, *class;
1494 	u32 function = *fn_ptr;
1495 
1496 	basic = kvm_find_cpuid_entry(vcpu, 0);
1497 	if (!basic)
1498 		return NULL;
1499 
1500 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1501 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1502 		return NULL;
1503 
1504 	if (function >= 0x40000000 && function <= 0x4fffffff)
1505 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1506 	else if (function >= 0xc0000000)
1507 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1508 	else
1509 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1510 
1511 	if (class && function <= class->eax)
1512 		return NULL;
1513 
1514 	/*
1515 	 * Leaf specific adjustments are also applied when redirecting to the
1516 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1517 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1518 	 * needs to be pulled from CPUID.0xb.1.
1519 	 */
1520 	*fn_ptr = basic->eax;
1521 
1522 	/*
1523 	 * The class does not exist or the requested function is out of range;
1524 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1525 	 * the original requested leaf is observed!
1526 	 */
1527 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1528 }
1529 
1530 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1531 	       u32 *ecx, u32 *edx, bool exact_only)
1532 {
1533 	u32 orig_function = *eax, function = *eax, index = *ecx;
1534 	struct kvm_cpuid_entry2 *entry;
1535 	bool exact, used_max_basic = false;
1536 
1537 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1538 	exact = !!entry;
1539 
1540 	if (!entry && !exact_only) {
1541 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1542 		used_max_basic = !!entry;
1543 	}
1544 
1545 	if (entry) {
1546 		*eax = entry->eax;
1547 		*ebx = entry->ebx;
1548 		*ecx = entry->ecx;
1549 		*edx = entry->edx;
1550 		if (function == 7 && index == 0) {
1551 			u64 data;
1552 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1553 			    (data & TSX_CTRL_CPUID_CLEAR))
1554 				*ebx &= ~(F(RTM) | F(HLE));
1555 		} else if (function == 0x80000007) {
1556 			if (kvm_hv_invtsc_suppressed(vcpu))
1557 				*edx &= ~SF(CONSTANT_TSC);
1558 		}
1559 	} else {
1560 		*eax = *ebx = *ecx = *edx = 0;
1561 		/*
1562 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1563 		 * and EDX is always the x2APIC ID, even for undefined
1564 		 * subleaves. Index 1 will exist iff the leaf is
1565 		 * implemented, so we pass through CL iff leaf 1
1566 		 * exists. EDX can be copied from any existing index.
1567 		 */
1568 		if (function == 0xb || function == 0x1f) {
1569 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1570 			if (entry) {
1571 				*ecx = index & 0xff;
1572 				*edx = entry->edx;
1573 			}
1574 		}
1575 	}
1576 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1577 			used_max_basic);
1578 	return exact;
1579 }
1580 EXPORT_SYMBOL_GPL(kvm_cpuid);
1581 
1582 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1583 {
1584 	u32 eax, ebx, ecx, edx;
1585 
1586 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1587 		return 1;
1588 
1589 	eax = kvm_rax_read(vcpu);
1590 	ecx = kvm_rcx_read(vcpu);
1591 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1592 	kvm_rax_write(vcpu, eax);
1593 	kvm_rbx_write(vcpu, ebx);
1594 	kvm_rcx_write(vcpu, ecx);
1595 	kvm_rdx_write(vcpu, edx);
1596 	return kvm_skip_emulated_instruction(vcpu);
1597 }
1598 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1599