xref: /linux/arch/x86/kvm/cpuid.c (revision 81c29435073355b8194986a2193d3e7b9d449225)
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/export.h>
15 #include <linux/vmalloc.h>
16 #include <linux/uaccess.h>
17 #include <linux/sched/stat.h>
18 
19 #include <asm/processor.h>
20 #include <asm/user.h>
21 #include <asm/fpu/xstate.h>
22 #include <asm/sgx.h>
23 #include <asm/cpuid.h>
24 #include "cpuid.h"
25 #include "lapic.h"
26 #include "mmu.h"
27 #include "trace.h"
28 #include "pmu.h"
29 #include "xen.h"
30 
31 /*
32  * Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be
33  * aligned to sizeof(unsigned long) because it's not accessed via bitops.
34  */
35 u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
36 EXPORT_SYMBOL_GPL(kvm_cpu_caps);
37 
38 u32 xstate_required_size(u64 xstate_bv, bool compacted)
39 {
40 	int feature_bit = 0;
41 	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
42 
43 	xstate_bv &= XFEATURE_MASK_EXTEND;
44 	while (xstate_bv) {
45 		if (xstate_bv & 0x1) {
46 		        u32 eax, ebx, ecx, edx, offset;
47 		        cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
48 			/* ECX[1]: 64B alignment in compacted form */
49 			if (compacted)
50 				offset = (ecx & 0x2) ? ALIGN(ret, 64) : ret;
51 			else
52 				offset = ebx;
53 			ret = max(ret, offset + eax);
54 		}
55 
56 		xstate_bv >>= 1;
57 		feature_bit++;
58 	}
59 
60 	return ret;
61 }
62 
63 #define F feature_bit
64 
65 /* Scattered Flag - For features that are scattered by cpufeatures.h. */
66 #define SF(name)						\
67 ({								\
68 	BUILD_BUG_ON(X86_FEATURE_##name >= MAX_CPU_FEATURES);	\
69 	(boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0);	\
70 })
71 
72 /*
73  * Magic value used by KVM when querying userspace-provided CPUID entries and
74  * doesn't care about the CPIUD index because the index of the function in
75  * question is not significant.  Note, this magic value must have at least one
76  * bit set in bits[63:32] and must be consumed as a u64 by cpuid_entry2_find()
77  * to avoid false positives when processing guest CPUID input.
78  */
79 #define KVM_CPUID_INDEX_NOT_SIGNIFICANT -1ull
80 
81 static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
82 	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u64 index)
83 {
84 	struct kvm_cpuid_entry2 *e;
85 	int i;
86 
87 	for (i = 0; i < nent; i++) {
88 		e = &entries[i];
89 
90 		if (e->function != function)
91 			continue;
92 
93 		/*
94 		 * If the index isn't significant, use the first entry with a
95 		 * matching function.  It's userspace's responsibilty to not
96 		 * provide "duplicate" entries in all cases.
97 		 */
98 		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
99 			return e;
100 
101 
102 		/*
103 		 * Similarly, use the first matching entry if KVM is doing a
104 		 * lookup (as opposed to emulating CPUID) for a function that's
105 		 * architecturally defined as not having a significant index.
106 		 */
107 		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
108 			/*
109 			 * Direct lookups from KVM should not diverge from what
110 			 * KVM defines internally (the architectural behavior).
111 			 */
112 			WARN_ON_ONCE(cpuid_function_is_indexed(function));
113 			return e;
114 		}
115 	}
116 
117 	return NULL;
118 }
119 
120 static int kvm_check_cpuid(struct kvm_vcpu *vcpu,
121 			   struct kvm_cpuid_entry2 *entries,
122 			   int nent)
123 {
124 	struct kvm_cpuid_entry2 *best;
125 	u64 xfeatures;
126 
127 	/*
128 	 * The existing code assumes virtual address is 48-bit or 57-bit in the
129 	 * canonical address checks; exit if it is ever changed.
130 	 */
131 	best = cpuid_entry2_find(entries, nent, 0x80000008,
132 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
133 	if (best) {
134 		int vaddr_bits = (best->eax & 0xff00) >> 8;
135 
136 		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
137 			return -EINVAL;
138 	}
139 
140 	/*
141 	 * Exposing dynamic xfeatures to the guest requires additional
142 	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
143 	 */
144 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
145 	if (!best)
146 		return 0;
147 
148 	xfeatures = best->eax | ((u64)best->edx << 32);
149 	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
150 	if (!xfeatures)
151 		return 0;
152 
153 	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
154 }
155 
156 /* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
157 static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
158 				 int nent)
159 {
160 	struct kvm_cpuid_entry2 *orig;
161 	int i;
162 
163 	if (nent != vcpu->arch.cpuid_nent)
164 		return -EINVAL;
165 
166 	for (i = 0; i < nent; i++) {
167 		orig = &vcpu->arch.cpuid_entries[i];
168 		if (e2[i].function != orig->function ||
169 		    e2[i].index != orig->index ||
170 		    e2[i].flags != orig->flags ||
171 		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
172 		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
173 			return -EINVAL;
174 	}
175 
176 	return 0;
177 }
178 
179 static struct kvm_hypervisor_cpuid kvm_get_hypervisor_cpuid(struct kvm_vcpu *vcpu,
180 							    const char *sig)
181 {
182 	struct kvm_hypervisor_cpuid cpuid = {};
183 	struct kvm_cpuid_entry2 *entry;
184 	u32 base;
185 
186 	for_each_possible_hypervisor_cpuid_base(base) {
187 		entry = kvm_find_cpuid_entry(vcpu, base);
188 
189 		if (entry) {
190 			u32 signature[3];
191 
192 			signature[0] = entry->ebx;
193 			signature[1] = entry->ecx;
194 			signature[2] = entry->edx;
195 
196 			if (!memcmp(signature, sig, sizeof(signature))) {
197 				cpuid.base = base;
198 				cpuid.limit = entry->eax;
199 				break;
200 			}
201 		}
202 	}
203 
204 	return cpuid;
205 }
206 
207 static struct kvm_cpuid_entry2 *__kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu,
208 					      struct kvm_cpuid_entry2 *entries, int nent)
209 {
210 	u32 base = vcpu->arch.kvm_cpuid.base;
211 
212 	if (!base)
213 		return NULL;
214 
215 	return cpuid_entry2_find(entries, nent, base | KVM_CPUID_FEATURES,
216 				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
217 }
218 
219 static struct kvm_cpuid_entry2 *kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu)
220 {
221 	return __kvm_find_kvm_cpuid_features(vcpu, vcpu->arch.cpuid_entries,
222 					     vcpu->arch.cpuid_nent);
223 }
224 
225 void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
226 {
227 	struct kvm_cpuid_entry2 *best = kvm_find_kvm_cpuid_features(vcpu);
228 
229 	/*
230 	 * save the feature bitmap to avoid cpuid lookup for every PV
231 	 * operation
232 	 */
233 	if (best)
234 		vcpu->arch.pv_cpuid.features = best->eax;
235 }
236 
237 /*
238  * Calculate guest's supported XCR0 taking into account guest CPUID data and
239  * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
240  */
241 static u64 cpuid_get_supported_xcr0(struct kvm_cpuid_entry2 *entries, int nent)
242 {
243 	struct kvm_cpuid_entry2 *best;
244 
245 	best = cpuid_entry2_find(entries, nent, 0xd, 0);
246 	if (!best)
247 		return 0;
248 
249 	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
250 }
251 
252 static void __kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *entries,
253 				       int nent)
254 {
255 	struct kvm_cpuid_entry2 *best;
256 	u64 guest_supported_xcr0 = cpuid_get_supported_xcr0(entries, nent);
257 
258 	best = cpuid_entry2_find(entries, nent, 1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
259 	if (best) {
260 		/* Update OSXSAVE bit */
261 		if (boot_cpu_has(X86_FEATURE_XSAVE))
262 			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
263 					   kvm_is_cr4_bit_set(vcpu, X86_CR4_OSXSAVE));
264 
265 		cpuid_entry_change(best, X86_FEATURE_APIC,
266 			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
267 	}
268 
269 	best = cpuid_entry2_find(entries, nent, 7, 0);
270 	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
271 		cpuid_entry_change(best, X86_FEATURE_OSPKE,
272 				   kvm_is_cr4_bit_set(vcpu, X86_CR4_PKE));
273 
274 	best = cpuid_entry2_find(entries, nent, 0xD, 0);
275 	if (best)
276 		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
277 
278 	best = cpuid_entry2_find(entries, nent, 0xD, 1);
279 	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
280 		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
281 		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
282 
283 	best = __kvm_find_kvm_cpuid_features(vcpu, entries, nent);
284 	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
285 		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
286 		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
287 
288 	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
289 		best = cpuid_entry2_find(entries, nent, 0x1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
290 		if (best)
291 			cpuid_entry_change(best, X86_FEATURE_MWAIT,
292 					   vcpu->arch.ia32_misc_enable_msr &
293 					   MSR_IA32_MISC_ENABLE_MWAIT);
294 	}
295 
296 	/*
297 	 * Bits 127:0 of the allowed SECS.ATTRIBUTES (CPUID.0x12.0x1) enumerate
298 	 * the supported XSAVE Feature Request Mask (XFRM), i.e. the enclave's
299 	 * requested XCR0 value.  The enclave's XFRM must be a subset of XCRO
300 	 * at the time of EENTER, thus adjust the allowed XFRM by the guest's
301 	 * supported XCR0.  Similar to XCR0 handling, FP and SSE are forced to
302 	 * '1' even on CPUs that don't support XSAVE.
303 	 */
304 	best = cpuid_entry2_find(entries, nent, 0x12, 0x1);
305 	if (best) {
306 		best->ecx &= guest_supported_xcr0 & 0xffffffff;
307 		best->edx &= guest_supported_xcr0 >> 32;
308 		best->ecx |= XFEATURE_MASK_FPSSE;
309 	}
310 }
311 
312 void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
313 {
314 	__kvm_update_cpuid_runtime(vcpu, vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
315 }
316 EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
317 
318 static bool kvm_cpuid_has_hyperv(struct kvm_cpuid_entry2 *entries, int nent)
319 {
320 	struct kvm_cpuid_entry2 *entry;
321 
322 	entry = cpuid_entry2_find(entries, nent, HYPERV_CPUID_INTERFACE,
323 				  KVM_CPUID_INDEX_NOT_SIGNIFICANT);
324 	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
325 }
326 
327 static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
328 {
329 	struct kvm_lapic *apic = vcpu->arch.apic;
330 	struct kvm_cpuid_entry2 *best;
331 
332 	best = kvm_find_cpuid_entry(vcpu, 1);
333 	if (best && apic) {
334 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
335 			apic->lapic_timer.timer_mode_mask = 3 << 17;
336 		else
337 			apic->lapic_timer.timer_mode_mask = 1 << 17;
338 
339 		kvm_apic_set_version(vcpu);
340 	}
341 
342 	vcpu->arch.guest_supported_xcr0 =
343 		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
344 
345 	/*
346 	 * FP+SSE can always be saved/restored via KVM_{G,S}ET_XSAVE, even if
347 	 * XSAVE/XCRO are not exposed to the guest, and even if XSAVE isn't
348 	 * supported by the host.
349 	 */
350 	vcpu->arch.guest_fpu.fpstate->user_xfeatures = vcpu->arch.guest_supported_xcr0 |
351 						       XFEATURE_MASK_FPSSE;
352 
353 	kvm_update_pv_runtime(vcpu);
354 
355 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
356 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
357 
358 	kvm_pmu_refresh(vcpu);
359 	vcpu->arch.cr4_guest_rsvd_bits =
360 	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
361 
362 	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu->arch.cpuid_entries,
363 						    vcpu->arch.cpuid_nent));
364 
365 	/* Invoke the vendor callback only after the above state is updated. */
366 	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
367 
368 	/*
369 	 * Except for the MMU, which needs to do its thing any vendor specific
370 	 * adjustments to the reserved GPA bits.
371 	 */
372 	kvm_mmu_after_set_cpuid(vcpu);
373 }
374 
375 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
376 {
377 	struct kvm_cpuid_entry2 *best;
378 
379 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
380 	if (!best || best->eax < 0x80000008)
381 		goto not_found;
382 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
383 	if (best)
384 		return best->eax & 0xff;
385 not_found:
386 	return 36;
387 }
388 
389 /*
390  * This "raw" version returns the reserved GPA bits without any adjustments for
391  * encryption technologies that usurp bits.  The raw mask should be used if and
392  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
393  */
394 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
395 {
396 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
397 }
398 
399 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
400                         int nent)
401 {
402 	int r;
403 
404 	__kvm_update_cpuid_runtime(vcpu, e2, nent);
405 
406 	/*
407 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
408 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
409 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
410 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
411 	 * the core vCPU model on the fly. It would've been better to forbid any
412 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
413 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
414 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
415 	 * whether the supplied CPUID data is equal to what's already set.
416 	 */
417 	if (kvm_vcpu_has_run(vcpu)) {
418 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
419 		if (r)
420 			return r;
421 
422 		kvfree(e2);
423 		return 0;
424 	}
425 
426 	if (kvm_cpuid_has_hyperv(e2, nent)) {
427 		r = kvm_hv_vcpu_init(vcpu);
428 		if (r)
429 			return r;
430 	}
431 
432 	r = kvm_check_cpuid(vcpu, e2, nent);
433 	if (r)
434 		return r;
435 
436 	kvfree(vcpu->arch.cpuid_entries);
437 	vcpu->arch.cpuid_entries = e2;
438 	vcpu->arch.cpuid_nent = nent;
439 
440 	vcpu->arch.kvm_cpuid = kvm_get_hypervisor_cpuid(vcpu, KVM_SIGNATURE);
441 	vcpu->arch.xen.cpuid = kvm_get_hypervisor_cpuid(vcpu, XEN_SIGNATURE);
442 	kvm_vcpu_after_set_cpuid(vcpu);
443 
444 	return 0;
445 }
446 
447 /* when an old userspace process fills a new kernel module */
448 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
449 			     struct kvm_cpuid *cpuid,
450 			     struct kvm_cpuid_entry __user *entries)
451 {
452 	int r, i;
453 	struct kvm_cpuid_entry *e = NULL;
454 	struct kvm_cpuid_entry2 *e2 = NULL;
455 
456 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
457 		return -E2BIG;
458 
459 	if (cpuid->nent) {
460 		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
461 		if (IS_ERR(e))
462 			return PTR_ERR(e);
463 
464 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
465 		if (!e2) {
466 			r = -ENOMEM;
467 			goto out_free_cpuid;
468 		}
469 	}
470 	for (i = 0; i < cpuid->nent; i++) {
471 		e2[i].function = e[i].function;
472 		e2[i].eax = e[i].eax;
473 		e2[i].ebx = e[i].ebx;
474 		e2[i].ecx = e[i].ecx;
475 		e2[i].edx = e[i].edx;
476 		e2[i].index = 0;
477 		e2[i].flags = 0;
478 		e2[i].padding[0] = 0;
479 		e2[i].padding[1] = 0;
480 		e2[i].padding[2] = 0;
481 	}
482 
483 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
484 	if (r)
485 		kvfree(e2);
486 
487 out_free_cpuid:
488 	kvfree(e);
489 
490 	return r;
491 }
492 
493 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
494 			      struct kvm_cpuid2 *cpuid,
495 			      struct kvm_cpuid_entry2 __user *entries)
496 {
497 	struct kvm_cpuid_entry2 *e2 = NULL;
498 	int r;
499 
500 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
501 		return -E2BIG;
502 
503 	if (cpuid->nent) {
504 		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
505 		if (IS_ERR(e2))
506 			return PTR_ERR(e2);
507 	}
508 
509 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
510 	if (r)
511 		kvfree(e2);
512 
513 	return r;
514 }
515 
516 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
517 			      struct kvm_cpuid2 *cpuid,
518 			      struct kvm_cpuid_entry2 __user *entries)
519 {
520 	int r;
521 
522 	r = -E2BIG;
523 	if (cpuid->nent < vcpu->arch.cpuid_nent)
524 		goto out;
525 	r = -EFAULT;
526 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
527 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
528 		goto out;
529 	return 0;
530 
531 out:
532 	cpuid->nent = vcpu->arch.cpuid_nent;
533 	return r;
534 }
535 
536 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
537 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
538 {
539 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
540 	struct kvm_cpuid_entry2 entry;
541 
542 	reverse_cpuid_check(leaf);
543 
544 	cpuid_count(cpuid.function, cpuid.index,
545 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
546 
547 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
548 }
549 
550 static __always_inline
551 void kvm_cpu_cap_init_kvm_defined(enum kvm_only_cpuid_leafs leaf, u32 mask)
552 {
553 	/* Use kvm_cpu_cap_mask for leafs that aren't KVM-only. */
554 	BUILD_BUG_ON(leaf < NCAPINTS);
555 
556 	kvm_cpu_caps[leaf] = mask;
557 
558 	__kvm_cpu_cap_mask(leaf);
559 }
560 
561 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
562 {
563 	/* Use kvm_cpu_cap_init_kvm_defined for KVM-only leafs. */
564 	BUILD_BUG_ON(leaf >= NCAPINTS);
565 
566 	kvm_cpu_caps[leaf] &= mask;
567 
568 	__kvm_cpu_cap_mask(leaf);
569 }
570 
571 void kvm_set_cpu_caps(void)
572 {
573 #ifdef CONFIG_X86_64
574 	unsigned int f_gbpages = F(GBPAGES);
575 	unsigned int f_lm = F(LM);
576 	unsigned int f_xfd = F(XFD);
577 #else
578 	unsigned int f_gbpages = 0;
579 	unsigned int f_lm = 0;
580 	unsigned int f_xfd = 0;
581 #endif
582 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
583 
584 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
585 		     sizeof(boot_cpu_data.x86_capability));
586 
587 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
588 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
589 
590 	kvm_cpu_cap_mask(CPUID_1_ECX,
591 		/*
592 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
593 		 * advertised to guests via CPUID!
594 		 */
595 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
596 		0 /* DS-CPL, VMX, SMX, EST */ |
597 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
598 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
599 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
600 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
601 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
602 		F(F16C) | F(RDRAND)
603 	);
604 	/* KVM emulates x2apic in software irrespective of host support. */
605 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
606 
607 	kvm_cpu_cap_mask(CPUID_1_EDX,
608 		F(FPU) | F(VME) | F(DE) | F(PSE) |
609 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
610 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
611 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
612 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
613 		0 /* Reserved, DS, ACPI */ | F(MMX) |
614 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
615 		0 /* HTT, TM, Reserved, PBE */
616 	);
617 
618 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
619 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
620 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
621 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
622 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
623 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
624 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
625 		F(AVX512VL));
626 
627 	kvm_cpu_cap_mask(CPUID_7_ECX,
628 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
629 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
630 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
631 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
632 		F(SGX_LC) | F(BUS_LOCK_DETECT)
633 	);
634 	/* Set LA57 based on hardware capability. */
635 	if (cpuid_ecx(7) & F(LA57))
636 		kvm_cpu_cap_set(X86_FEATURE_LA57);
637 
638 	/*
639 	 * PKU not yet implemented for shadow paging and requires OSPKE
640 	 * to be set on the host. Clear it if that is not the case
641 	 */
642 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
643 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
644 
645 	kvm_cpu_cap_mask(CPUID_7_EDX,
646 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
647 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
648 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
649 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
650 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16) | F(FLUSH_L1D)
651 	);
652 
653 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
654 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
655 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
656 
657 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
658 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
659 	if (boot_cpu_has(X86_FEATURE_STIBP))
660 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
661 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
662 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
663 
664 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
665 		F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) |
666 		F(FZRM) | F(FSRS) | F(FSRC) |
667 		F(AMX_FP16) | F(AVX_IFMA)
668 	);
669 
670 	kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
671 		F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(PREFETCHITI)
672 	);
673 
674 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
675 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
676 	);
677 
678 	kvm_cpu_cap_init_kvm_defined(CPUID_12_EAX,
679 		SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
680 	);
681 
682 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
683 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
684 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
685 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
686 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
687 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
688 	);
689 
690 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
691 		F(FPU) | F(VME) | F(DE) | F(PSE) |
692 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
693 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
694 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
695 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
696 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
697 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
698 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
699 	);
700 
701 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
702 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
703 
704 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0007_EDX,
705 		SF(CONSTANT_TSC)
706 	);
707 
708 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
709 		F(CLZERO) | F(XSAVEERPTR) |
710 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
711 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
712 		F(AMD_PSFD)
713 	);
714 
715 	/*
716 	 * AMD has separate bits for each SPEC_CTRL bit.
717 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
718 	 * record that in cpufeatures so use them.
719 	 */
720 	if (boot_cpu_has(X86_FEATURE_IBPB))
721 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
722 	if (boot_cpu_has(X86_FEATURE_IBRS))
723 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
724 	if (boot_cpu_has(X86_FEATURE_STIBP))
725 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
726 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
727 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
728 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
729 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
730 	/*
731 	 * The preference is to use SPEC CTRL MSR instead of the
732 	 * VIRT_SPEC MSR.
733 	 */
734 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
735 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
736 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
737 
738 	/*
739 	 * Hide all SVM features by default, SVM will set the cap bits for
740 	 * features it emulates and/or exposes for L1.
741 	 */
742 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
743 
744 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
745 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
746 		F(SME_COHERENT));
747 
748 	kvm_cpu_cap_mask(CPUID_8000_0021_EAX,
749 		F(NO_NESTED_DATA_BP) | F(LFENCE_RDTSC) | 0 /* SmmPgCfgLock */ |
750 		F(NULL_SEL_CLR_BASE) | F(AUTOIBRS) | 0 /* PrefetchCtlMsr */
751 	);
752 
753 	/*
754 	 * Synthesize "LFENCE is serializing" into the AMD-defined entry in
755 	 * KVM's supported CPUID if the feature is reported as supported by the
756 	 * kernel.  LFENCE_RDTSC was a Linux-defined synthetic feature long
757 	 * before AMD joined the bandwagon, e.g. LFENCE is serializing on most
758 	 * CPUs that support SSE2.  On CPUs that don't support AMD's leaf,
759 	 * kvm_cpu_cap_mask() will unfortunately drop the flag due to ANDing
760 	 * the mask with the raw host CPUID, and reporting support in AMD's
761 	 * leaf can make it easier for userspace to detect the feature.
762 	 */
763 	if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
764 		kvm_cpu_cap_set(X86_FEATURE_LFENCE_RDTSC);
765 	if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
766 		kvm_cpu_cap_set(X86_FEATURE_NULL_SEL_CLR_BASE);
767 	kvm_cpu_cap_set(X86_FEATURE_NO_SMM_CTL_MSR);
768 
769 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
770 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
771 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
772 		F(PMM) | F(PMM_EN)
773 	);
774 
775 	/*
776 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
777 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
778 	 * should never happen, but the guest will likely crash if RDTSCP or
779 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
780 	 * the past.  For example, the sanity check may fire if this instance of
781 	 * KVM is running as L1 on top of an older, broken KVM.
782 	 */
783 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
784 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
785 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
786 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
787 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
788 	}
789 }
790 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
791 
792 struct kvm_cpuid_array {
793 	struct kvm_cpuid_entry2 *entries;
794 	int maxnent;
795 	int nent;
796 };
797 
798 static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
799 {
800 	if (array->nent >= array->maxnent)
801 		return NULL;
802 
803 	return &array->entries[array->nent++];
804 }
805 
806 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
807 					      u32 function, u32 index)
808 {
809 	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
810 
811 	if (!entry)
812 		return NULL;
813 
814 	memset(entry, 0, sizeof(*entry));
815 	entry->function = function;
816 	entry->index = index;
817 	switch (function & 0xC0000000) {
818 	case 0x40000000:
819 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
820 		return entry;
821 
822 	case 0x80000000:
823 		/*
824 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
825 		 * would result in out-of-bounds calls to do_host_cpuid.
826 		 */
827 		{
828 			static int max_cpuid_80000000;
829 			if (!READ_ONCE(max_cpuid_80000000))
830 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
831 			if (function > READ_ONCE(max_cpuid_80000000))
832 				return entry;
833 		}
834 		break;
835 
836 	default:
837 		break;
838 	}
839 
840 	cpuid_count(entry->function, entry->index,
841 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
842 
843 	if (cpuid_function_is_indexed(function))
844 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
845 
846 	return entry;
847 }
848 
849 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
850 {
851 	struct kvm_cpuid_entry2 *entry;
852 
853 	if (array->nent >= array->maxnent)
854 		return -E2BIG;
855 
856 	entry = &array->entries[array->nent];
857 	entry->function = func;
858 	entry->index = 0;
859 	entry->flags = 0;
860 
861 	switch (func) {
862 	case 0:
863 		entry->eax = 7;
864 		++array->nent;
865 		break;
866 	case 1:
867 		entry->ecx = F(MOVBE);
868 		++array->nent;
869 		break;
870 	case 7:
871 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
872 		entry->eax = 0;
873 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
874 			entry->ecx = F(RDPID);
875 		++array->nent;
876 		break;
877 	default:
878 		break;
879 	}
880 
881 	return 0;
882 }
883 
884 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
885 {
886 	struct kvm_cpuid_entry2 *entry;
887 	int r, i, max_idx;
888 
889 	/* all calls to cpuid_count() should be made on the same cpu */
890 	get_cpu();
891 
892 	r = -E2BIG;
893 
894 	entry = do_host_cpuid(array, function, 0);
895 	if (!entry)
896 		goto out;
897 
898 	switch (function) {
899 	case 0:
900 		/* Limited to the highest leaf implemented in KVM. */
901 		entry->eax = min(entry->eax, 0x1fU);
902 		break;
903 	case 1:
904 		cpuid_entry_override(entry, CPUID_1_EDX);
905 		cpuid_entry_override(entry, CPUID_1_ECX);
906 		break;
907 	case 2:
908 		/*
909 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
910 		 * CPUID(function=2, index=0) may return different results each
911 		 * time, with the least-significant byte in EAX enumerating the
912 		 * number of times software should do CPUID(2, 0).
913 		 *
914 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
915 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
916 		 * return 01H. Software should ignore this value and not
917 		 * interpret it as an informational descriptor", while AMD's
918 		 * APM states that CPUID(2) is reserved.
919 		 *
920 		 * WARN if a frankenstein CPU that supports virtualization and
921 		 * a stateful CPUID.0x2 is encountered.
922 		 */
923 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
924 		break;
925 	/* functions 4 and 0x8000001d have additional index. */
926 	case 4:
927 	case 0x8000001d:
928 		/*
929 		 * Read entries until the cache type in the previous entry is
930 		 * zero, i.e. indicates an invalid entry.
931 		 */
932 		for (i = 1; entry->eax & 0x1f; ++i) {
933 			entry = do_host_cpuid(array, function, i);
934 			if (!entry)
935 				goto out;
936 		}
937 		break;
938 	case 6: /* Thermal management */
939 		entry->eax = 0x4; /* allow ARAT */
940 		entry->ebx = 0;
941 		entry->ecx = 0;
942 		entry->edx = 0;
943 		break;
944 	/* function 7 has additional index. */
945 	case 7:
946 		entry->eax = min(entry->eax, 1u);
947 		cpuid_entry_override(entry, CPUID_7_0_EBX);
948 		cpuid_entry_override(entry, CPUID_7_ECX);
949 		cpuid_entry_override(entry, CPUID_7_EDX);
950 
951 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
952 		if (entry->eax == 1) {
953 			entry = do_host_cpuid(array, function, 1);
954 			if (!entry)
955 				goto out;
956 
957 			cpuid_entry_override(entry, CPUID_7_1_EAX);
958 			cpuid_entry_override(entry, CPUID_7_1_EDX);
959 			entry->ebx = 0;
960 			entry->ecx = 0;
961 		}
962 		break;
963 	case 0xa: { /* Architectural Performance Monitoring */
964 		union cpuid10_eax eax;
965 		union cpuid10_edx edx;
966 
967 		if (!static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
968 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
969 			break;
970 		}
971 
972 		eax.split.version_id = kvm_pmu_cap.version;
973 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
974 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
975 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
976 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
977 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
978 
979 		if (kvm_pmu_cap.version)
980 			edx.split.anythread_deprecated = 1;
981 		edx.split.reserved1 = 0;
982 		edx.split.reserved2 = 0;
983 
984 		entry->eax = eax.full;
985 		entry->ebx = kvm_pmu_cap.events_mask;
986 		entry->ecx = 0;
987 		entry->edx = edx.full;
988 		break;
989 	}
990 	case 0x1f:
991 	case 0xb:
992 		/*
993 		 * No topology; a valid topology is indicated by the presence
994 		 * of subleaf 1.
995 		 */
996 		entry->eax = entry->ebx = entry->ecx = 0;
997 		break;
998 	case 0xd: {
999 		u64 permitted_xcr0 = kvm_get_filtered_xcr0();
1000 		u64 permitted_xss = kvm_caps.supported_xss;
1001 
1002 		entry->eax &= permitted_xcr0;
1003 		entry->ebx = xstate_required_size(permitted_xcr0, false);
1004 		entry->ecx = entry->ebx;
1005 		entry->edx &= permitted_xcr0 >> 32;
1006 		if (!permitted_xcr0)
1007 			break;
1008 
1009 		entry = do_host_cpuid(array, function, 1);
1010 		if (!entry)
1011 			goto out;
1012 
1013 		cpuid_entry_override(entry, CPUID_D_1_EAX);
1014 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
1015 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
1016 							  true);
1017 		else {
1018 			WARN_ON_ONCE(permitted_xss != 0);
1019 			entry->ebx = 0;
1020 		}
1021 		entry->ecx &= permitted_xss;
1022 		entry->edx &= permitted_xss >> 32;
1023 
1024 		for (i = 2; i < 64; ++i) {
1025 			bool s_state;
1026 			if (permitted_xcr0 & BIT_ULL(i))
1027 				s_state = false;
1028 			else if (permitted_xss & BIT_ULL(i))
1029 				s_state = true;
1030 			else
1031 				continue;
1032 
1033 			entry = do_host_cpuid(array, function, i);
1034 			if (!entry)
1035 				goto out;
1036 
1037 			/*
1038 			 * The supported check above should have filtered out
1039 			 * invalid sub-leafs.  Only valid sub-leafs should
1040 			 * reach this point, and they should have a non-zero
1041 			 * save state size.  Furthermore, check whether the
1042 			 * processor agrees with permitted_xcr0/permitted_xss
1043 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1044 			 */
1045 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1046 				--array->nent;
1047 				continue;
1048 			}
1049 
1050 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1051 				entry->ecx &= ~BIT_ULL(2);
1052 			entry->edx = 0;
1053 		}
1054 		break;
1055 	}
1056 	case 0x12:
1057 		/* Intel SGX */
1058 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1059 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1060 			break;
1061 		}
1062 
1063 		/*
1064 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1065 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1066 		 * are restricted by kernel and KVM capabilities (like most
1067 		 * feature flags), while enclave size is unrestricted.
1068 		 */
1069 		cpuid_entry_override(entry, CPUID_12_EAX);
1070 		entry->ebx &= SGX_MISC_EXINFO;
1071 
1072 		entry = do_host_cpuid(array, function, 1);
1073 		if (!entry)
1074 			goto out;
1075 
1076 		/*
1077 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1078 		 * feature flags.  Advertise all supported flags, including
1079 		 * privileged attributes that require explicit opt-in from
1080 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1081 		 * expected to derive it from supported XCR0.
1082 		 */
1083 		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1084 		entry->ebx &= 0;
1085 		break;
1086 	/* Intel PT */
1087 	case 0x14:
1088 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1089 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1090 			break;
1091 		}
1092 
1093 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1094 			if (!do_host_cpuid(array, function, i))
1095 				goto out;
1096 		}
1097 		break;
1098 	/* Intel AMX TILE */
1099 	case 0x1d:
1100 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1101 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1102 			break;
1103 		}
1104 
1105 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1106 			if (!do_host_cpuid(array, function, i))
1107 				goto out;
1108 		}
1109 		break;
1110 	case 0x1e: /* TMUL information */
1111 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1112 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1113 			break;
1114 		}
1115 		break;
1116 	case KVM_CPUID_SIGNATURE: {
1117 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1118 		entry->eax = KVM_CPUID_FEATURES;
1119 		entry->ebx = sigptr[0];
1120 		entry->ecx = sigptr[1];
1121 		entry->edx = sigptr[2];
1122 		break;
1123 	}
1124 	case KVM_CPUID_FEATURES:
1125 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1126 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1127 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1128 			     (1 << KVM_FEATURE_ASYNC_PF) |
1129 			     (1 << KVM_FEATURE_PV_EOI) |
1130 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1131 			     (1 << KVM_FEATURE_PV_UNHALT) |
1132 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1133 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1134 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1135 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1136 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1137 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1138 
1139 		if (sched_info_on())
1140 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1141 
1142 		entry->ebx = 0;
1143 		entry->ecx = 0;
1144 		entry->edx = 0;
1145 		break;
1146 	case 0x80000000:
1147 		entry->eax = min(entry->eax, 0x80000021);
1148 		/*
1149 		 * Serializing LFENCE is reported in a multitude of ways, and
1150 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1151 		 * userspace by providing the CPUID leaf ourselves.
1152 		 *
1153 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1154 		 * QEMU thinks that it can query the host blindly for that
1155 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1156 		 * above.  The processor merrily returns values from the
1157 		 * highest Intel leaf which QEMU tries to use as the guest's
1158 		 * 0x8000001d.  Even worse, this can result in an infinite
1159 		 * loop if said highest leaf has no subleaves indexed by ECX.
1160 		 */
1161 		if (entry->eax >= 0x8000001d &&
1162 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1163 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1164 			entry->eax = max(entry->eax, 0x80000021);
1165 		break;
1166 	case 0x80000001:
1167 		entry->ebx &= ~GENMASK(27, 16);
1168 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1169 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1170 		break;
1171 	case 0x80000006:
1172 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1173 		entry->edx &= ~GENMASK(17, 16);
1174 		break;
1175 	case 0x80000007: /* Advanced power management */
1176 		cpuid_entry_override(entry, CPUID_8000_0007_EDX);
1177 
1178 		/* mask against host */
1179 		entry->edx &= boot_cpu_data.x86_power;
1180 		entry->eax = entry->ebx = entry->ecx = 0;
1181 		break;
1182 	case 0x80000008: {
1183 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1184 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1185 		unsigned phys_as = entry->eax & 0xff;
1186 
1187 		/*
1188 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1189 		 * the guest operates in the same PA space as the host, i.e.
1190 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1191 		 * paging, too.
1192 		 *
1193 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1194 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1195 		 * the HPAs do not affect GPAs.
1196 		 */
1197 		if (!tdp_enabled)
1198 			g_phys_as = boot_cpu_data.x86_phys_bits;
1199 		else if (!g_phys_as)
1200 			g_phys_as = phys_as;
1201 
1202 		entry->eax = g_phys_as | (virt_as << 8);
1203 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1204 		entry->edx = 0;
1205 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1206 		break;
1207 	}
1208 	case 0x8000000A:
1209 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1210 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1211 			break;
1212 		}
1213 		entry->eax = 1; /* SVM revision 1 */
1214 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1215 				   ASID emulation to nested SVM */
1216 		entry->ecx = 0; /* Reserved */
1217 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1218 		break;
1219 	case 0x80000019:
1220 		entry->ecx = entry->edx = 0;
1221 		break;
1222 	case 0x8000001a:
1223 		entry->eax &= GENMASK(2, 0);
1224 		entry->ebx = entry->ecx = entry->edx = 0;
1225 		break;
1226 	case 0x8000001e:
1227 		/* Do not return host topology information.  */
1228 		entry->eax = entry->ebx = entry->ecx = 0;
1229 		entry->edx = 0; /* reserved */
1230 		break;
1231 	case 0x8000001F:
1232 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1233 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1234 		} else {
1235 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1236 			/* Clear NumVMPL since KVM does not support VMPL.  */
1237 			entry->ebx &= ~GENMASK(31, 12);
1238 			/*
1239 			 * Enumerate '0' for "PA bits reduction", the adjusted
1240 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1241 			 */
1242 			entry->ebx &= ~GENMASK(11, 6);
1243 		}
1244 		break;
1245 	case 0x80000020:
1246 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1247 		break;
1248 	case 0x80000021:
1249 		entry->ebx = entry->ecx = entry->edx = 0;
1250 		cpuid_entry_override(entry, CPUID_8000_0021_EAX);
1251 		break;
1252 	/*Add support for Centaur's CPUID instruction*/
1253 	case 0xC0000000:
1254 		/*Just support up to 0xC0000004 now*/
1255 		entry->eax = min(entry->eax, 0xC0000004);
1256 		break;
1257 	case 0xC0000001:
1258 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1259 		break;
1260 	case 3: /* Processor serial number */
1261 	case 5: /* MONITOR/MWAIT */
1262 	case 0xC0000002:
1263 	case 0xC0000003:
1264 	case 0xC0000004:
1265 	default:
1266 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1267 		break;
1268 	}
1269 
1270 	r = 0;
1271 
1272 out:
1273 	put_cpu();
1274 
1275 	return r;
1276 }
1277 
1278 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1279 			 unsigned int type)
1280 {
1281 	if (type == KVM_GET_EMULATED_CPUID)
1282 		return __do_cpuid_func_emulated(array, func);
1283 
1284 	return __do_cpuid_func(array, func);
1285 }
1286 
1287 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1288 
1289 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1290 			  unsigned int type)
1291 {
1292 	u32 limit;
1293 	int r;
1294 
1295 	if (func == CENTAUR_CPUID_SIGNATURE &&
1296 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1297 		return 0;
1298 
1299 	r = do_cpuid_func(array, func, type);
1300 	if (r)
1301 		return r;
1302 
1303 	limit = array->entries[array->nent - 1].eax;
1304 	for (func = func + 1; func <= limit; ++func) {
1305 		r = do_cpuid_func(array, func, type);
1306 		if (r)
1307 			break;
1308 	}
1309 
1310 	return r;
1311 }
1312 
1313 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1314 				 __u32 num_entries, unsigned int ioctl_type)
1315 {
1316 	int i;
1317 	__u32 pad[3];
1318 
1319 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1320 		return false;
1321 
1322 	/*
1323 	 * We want to make sure that ->padding is being passed clean from
1324 	 * userspace in case we want to use it for something in the future.
1325 	 *
1326 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1327 	 * have to give ourselves satisfied only with the emulated side. /me
1328 	 * sheds a tear.
1329 	 */
1330 	for (i = 0; i < num_entries; i++) {
1331 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1332 			return true;
1333 
1334 		if (pad[0] || pad[1] || pad[2])
1335 			return true;
1336 	}
1337 	return false;
1338 }
1339 
1340 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1341 			    struct kvm_cpuid_entry2 __user *entries,
1342 			    unsigned int type)
1343 {
1344 	static const u32 funcs[] = {
1345 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1346 	};
1347 
1348 	struct kvm_cpuid_array array = {
1349 		.nent = 0,
1350 	};
1351 	int r, i;
1352 
1353 	if (cpuid->nent < 1)
1354 		return -E2BIG;
1355 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1356 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1357 
1358 	if (sanity_check_entries(entries, cpuid->nent, type))
1359 		return -EINVAL;
1360 
1361 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1362 	if (!array.entries)
1363 		return -ENOMEM;
1364 
1365 	array.maxnent = cpuid->nent;
1366 
1367 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1368 		r = get_cpuid_func(&array, funcs[i], type);
1369 		if (r)
1370 			goto out_free;
1371 	}
1372 	cpuid->nent = array.nent;
1373 
1374 	if (copy_to_user(entries, array.entries,
1375 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1376 		r = -EFAULT;
1377 
1378 out_free:
1379 	kvfree(array.entries);
1380 	return r;
1381 }
1382 
1383 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1384 						    u32 function, u32 index)
1385 {
1386 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1387 				 function, index);
1388 }
1389 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1390 
1391 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1392 					      u32 function)
1393 {
1394 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1395 				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1396 }
1397 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1398 
1399 /*
1400  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1401  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1402  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1403  * range.  Centaur/VIA follows Intel semantics.
1404  *
1405  * A leaf is considered out-of-range if its function is higher than the maximum
1406  * supported leaf of its associated class or if its associated class does not
1407  * exist.
1408  *
1409  * There are three primary classes to be considered, with their respective
1410  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1411  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1412  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1413  *
1414  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1415  *  - Hypervisor: 0x40000000 - 0x4fffffff
1416  *  - Extended:   0x80000000 - 0xbfffffff
1417  *  - Centaur:    0xc0000000 - 0xcfffffff
1418  *
1419  * The Hypervisor class is further subdivided into sub-classes that each act as
1420  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1421  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1422  * CPUID sub-classes are:
1423  *
1424  *  - HyperV:     0x40000000 - 0x400000ff
1425  *  - KVM:        0x40000100 - 0x400001ff
1426  */
1427 static struct kvm_cpuid_entry2 *
1428 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1429 {
1430 	struct kvm_cpuid_entry2 *basic, *class;
1431 	u32 function = *fn_ptr;
1432 
1433 	basic = kvm_find_cpuid_entry(vcpu, 0);
1434 	if (!basic)
1435 		return NULL;
1436 
1437 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1438 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1439 		return NULL;
1440 
1441 	if (function >= 0x40000000 && function <= 0x4fffffff)
1442 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1443 	else if (function >= 0xc0000000)
1444 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1445 	else
1446 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1447 
1448 	if (class && function <= class->eax)
1449 		return NULL;
1450 
1451 	/*
1452 	 * Leaf specific adjustments are also applied when redirecting to the
1453 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1454 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1455 	 * needs to be pulled from CPUID.0xb.1.
1456 	 */
1457 	*fn_ptr = basic->eax;
1458 
1459 	/*
1460 	 * The class does not exist or the requested function is out of range;
1461 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1462 	 * the original requested leaf is observed!
1463 	 */
1464 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1465 }
1466 
1467 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1468 	       u32 *ecx, u32 *edx, bool exact_only)
1469 {
1470 	u32 orig_function = *eax, function = *eax, index = *ecx;
1471 	struct kvm_cpuid_entry2 *entry;
1472 	bool exact, used_max_basic = false;
1473 
1474 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1475 	exact = !!entry;
1476 
1477 	if (!entry && !exact_only) {
1478 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1479 		used_max_basic = !!entry;
1480 	}
1481 
1482 	if (entry) {
1483 		*eax = entry->eax;
1484 		*ebx = entry->ebx;
1485 		*ecx = entry->ecx;
1486 		*edx = entry->edx;
1487 		if (function == 7 && index == 0) {
1488 			u64 data;
1489 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1490 			    (data & TSX_CTRL_CPUID_CLEAR))
1491 				*ebx &= ~(F(RTM) | F(HLE));
1492 		} else if (function == 0x80000007) {
1493 			if (kvm_hv_invtsc_suppressed(vcpu))
1494 				*edx &= ~SF(CONSTANT_TSC);
1495 		}
1496 	} else {
1497 		*eax = *ebx = *ecx = *edx = 0;
1498 		/*
1499 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1500 		 * and EDX is always the x2APIC ID, even for undefined
1501 		 * subleaves. Index 1 will exist iff the leaf is
1502 		 * implemented, so we pass through CL iff leaf 1
1503 		 * exists. EDX can be copied from any existing index.
1504 		 */
1505 		if (function == 0xb || function == 0x1f) {
1506 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1507 			if (entry) {
1508 				*ecx = index & 0xff;
1509 				*edx = entry->edx;
1510 			}
1511 		}
1512 	}
1513 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1514 			used_max_basic);
1515 	return exact;
1516 }
1517 EXPORT_SYMBOL_GPL(kvm_cpuid);
1518 
1519 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1520 {
1521 	u32 eax, ebx, ecx, edx;
1522 
1523 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1524 		return 1;
1525 
1526 	eax = kvm_rax_read(vcpu);
1527 	ecx = kvm_rcx_read(vcpu);
1528 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1529 	kvm_rax_write(vcpu, eax);
1530 	kvm_rbx_write(vcpu, ebx);
1531 	kvm_rcx_write(vcpu, ecx);
1532 	kvm_rdx_write(vcpu, edx);
1533 	return kvm_skip_emulated_instruction(vcpu);
1534 }
1535 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1536