xref: /openbmc/linux/arch/x86/kvm/cpuid.c (revision a9886aad)
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 responsibilty 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 	struct kvm_cpuid_entry2 *entry;
318 
319 	entry = cpuid_entry2_find(entries, nent, HYPERV_CPUID_INTERFACE,
320 				  KVM_CPUID_INDEX_NOT_SIGNIFICANT);
321 	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
322 }
323 
324 static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
325 {
326 	struct kvm_lapic *apic = vcpu->arch.apic;
327 	struct kvm_cpuid_entry2 *best;
328 	bool allow_gbpages;
329 
330 	BUILD_BUG_ON(KVM_NR_GOVERNED_FEATURES > KVM_MAX_NR_GOVERNED_FEATURES);
331 	bitmap_zero(vcpu->arch.governed_features.enabled,
332 		    KVM_MAX_NR_GOVERNED_FEATURES);
333 
334 	/*
335 	 * If TDP is enabled, let the guest use GBPAGES if they're supported in
336 	 * hardware.  The hardware page walker doesn't let KVM disable GBPAGES,
337 	 * i.e. won't treat them as reserved, and KVM doesn't redo the GVA->GPA
338 	 * walk for performance and complexity reasons.  Not to mention KVM
339 	 * _can't_ solve the problem because GVA->GPA walks aren't visible to
340 	 * KVM once a TDP translation is installed.  Mimic hardware behavior so
341 	 * that KVM's is at least consistent, i.e. doesn't randomly inject #PF.
342 	 * If TDP is disabled, honor *only* guest CPUID as KVM has full control
343 	 * and can install smaller shadow pages if the host lacks 1GiB support.
344 	 */
345 	allow_gbpages = tdp_enabled ? boot_cpu_has(X86_FEATURE_GBPAGES) :
346 				      guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES);
347 	if (allow_gbpages)
348 		kvm_governed_feature_set(vcpu, X86_FEATURE_GBPAGES);
349 
350 	best = kvm_find_cpuid_entry(vcpu, 1);
351 	if (best && apic) {
352 		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
353 			apic->lapic_timer.timer_mode_mask = 3 << 17;
354 		else
355 			apic->lapic_timer.timer_mode_mask = 1 << 17;
356 
357 		kvm_apic_set_version(vcpu);
358 	}
359 
360 	vcpu->arch.guest_supported_xcr0 =
361 		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
362 
363 	kvm_update_pv_runtime(vcpu);
364 
365 	vcpu->arch.is_amd_compatible = guest_cpuid_is_amd_or_hygon(vcpu);
366 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
367 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
368 
369 	kvm_pmu_refresh(vcpu);
370 	vcpu->arch.cr4_guest_rsvd_bits =
371 	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
372 
373 	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu->arch.cpuid_entries,
374 						    vcpu->arch.cpuid_nent));
375 
376 	/* Invoke the vendor callback only after the above state is updated. */
377 	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
378 
379 	/*
380 	 * Except for the MMU, which needs to do its thing any vendor specific
381 	 * adjustments to the reserved GPA bits.
382 	 */
383 	kvm_mmu_after_set_cpuid(vcpu);
384 }
385 
386 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
387 {
388 	struct kvm_cpuid_entry2 *best;
389 
390 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
391 	if (!best || best->eax < 0x80000008)
392 		goto not_found;
393 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
394 	if (best)
395 		return best->eax & 0xff;
396 not_found:
397 	return 36;
398 }
399 
400 /*
401  * This "raw" version returns the reserved GPA bits without any adjustments for
402  * encryption technologies that usurp bits.  The raw mask should be used if and
403  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
404  */
405 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
406 {
407 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
408 }
409 
410 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
411                         int nent)
412 {
413 	int r;
414 
415 	__kvm_update_cpuid_runtime(vcpu, e2, nent);
416 
417 	/*
418 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
419 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
420 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
421 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
422 	 * the core vCPU model on the fly. It would've been better to forbid any
423 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
424 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
425 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
426 	 * whether the supplied CPUID data is equal to what's already set.
427 	 */
428 	if (kvm_vcpu_has_run(vcpu)) {
429 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
430 		if (r)
431 			return r;
432 
433 		kvfree(e2);
434 		return 0;
435 	}
436 
437 	if (kvm_cpuid_has_hyperv(e2, nent)) {
438 		r = kvm_hv_vcpu_init(vcpu);
439 		if (r)
440 			return r;
441 	}
442 
443 	r = kvm_check_cpuid(vcpu, e2, nent);
444 	if (r)
445 		return r;
446 
447 	kvfree(vcpu->arch.cpuid_entries);
448 	vcpu->arch.cpuid_entries = e2;
449 	vcpu->arch.cpuid_nent = nent;
450 
451 	vcpu->arch.kvm_cpuid = kvm_get_hypervisor_cpuid(vcpu, KVM_SIGNATURE);
452 	vcpu->arch.xen.cpuid = kvm_get_hypervisor_cpuid(vcpu, XEN_SIGNATURE);
453 	kvm_vcpu_after_set_cpuid(vcpu);
454 
455 	return 0;
456 }
457 
458 /* when an old userspace process fills a new kernel module */
459 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
460 			     struct kvm_cpuid *cpuid,
461 			     struct kvm_cpuid_entry __user *entries)
462 {
463 	int r, i;
464 	struct kvm_cpuid_entry *e = NULL;
465 	struct kvm_cpuid_entry2 *e2 = NULL;
466 
467 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
468 		return -E2BIG;
469 
470 	if (cpuid->nent) {
471 		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
472 		if (IS_ERR(e))
473 			return PTR_ERR(e);
474 
475 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
476 		if (!e2) {
477 			r = -ENOMEM;
478 			goto out_free_cpuid;
479 		}
480 	}
481 	for (i = 0; i < cpuid->nent; i++) {
482 		e2[i].function = e[i].function;
483 		e2[i].eax = e[i].eax;
484 		e2[i].ebx = e[i].ebx;
485 		e2[i].ecx = e[i].ecx;
486 		e2[i].edx = e[i].edx;
487 		e2[i].index = 0;
488 		e2[i].flags = 0;
489 		e2[i].padding[0] = 0;
490 		e2[i].padding[1] = 0;
491 		e2[i].padding[2] = 0;
492 	}
493 
494 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
495 	if (r)
496 		kvfree(e2);
497 
498 out_free_cpuid:
499 	kvfree(e);
500 
501 	return r;
502 }
503 
504 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
505 			      struct kvm_cpuid2 *cpuid,
506 			      struct kvm_cpuid_entry2 __user *entries)
507 {
508 	struct kvm_cpuid_entry2 *e2 = NULL;
509 	int r;
510 
511 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
512 		return -E2BIG;
513 
514 	if (cpuid->nent) {
515 		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
516 		if (IS_ERR(e2))
517 			return PTR_ERR(e2);
518 	}
519 
520 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
521 	if (r)
522 		kvfree(e2);
523 
524 	return r;
525 }
526 
527 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
528 			      struct kvm_cpuid2 *cpuid,
529 			      struct kvm_cpuid_entry2 __user *entries)
530 {
531 	if (cpuid->nent < vcpu->arch.cpuid_nent)
532 		return -E2BIG;
533 
534 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
535 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
536 		return -EFAULT;
537 
538 	cpuid->nent = vcpu->arch.cpuid_nent;
539 	return 0;
540 }
541 
542 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
543 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
544 {
545 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
546 	struct kvm_cpuid_entry2 entry;
547 
548 	reverse_cpuid_check(leaf);
549 
550 	cpuid_count(cpuid.function, cpuid.index,
551 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
552 
553 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
554 }
555 
556 static __always_inline
557 void kvm_cpu_cap_init_kvm_defined(enum kvm_only_cpuid_leafs leaf, u32 mask)
558 {
559 	/* Use kvm_cpu_cap_mask for leafs that aren't KVM-only. */
560 	BUILD_BUG_ON(leaf < NCAPINTS);
561 
562 	kvm_cpu_caps[leaf] = mask;
563 
564 	__kvm_cpu_cap_mask(leaf);
565 }
566 
567 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
568 {
569 	/* Use kvm_cpu_cap_init_kvm_defined for KVM-only leafs. */
570 	BUILD_BUG_ON(leaf >= NCAPINTS);
571 
572 	kvm_cpu_caps[leaf] &= mask;
573 
574 	__kvm_cpu_cap_mask(leaf);
575 }
576 
577 void kvm_set_cpu_caps(void)
578 {
579 #ifdef CONFIG_X86_64
580 	unsigned int f_gbpages = F(GBPAGES);
581 	unsigned int f_lm = F(LM);
582 	unsigned int f_xfd = F(XFD);
583 #else
584 	unsigned int f_gbpages = 0;
585 	unsigned int f_lm = 0;
586 	unsigned int f_xfd = 0;
587 #endif
588 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
589 
590 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
591 		     sizeof(boot_cpu_data.x86_capability));
592 
593 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
594 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
595 
596 	kvm_cpu_cap_mask(CPUID_1_ECX,
597 		/*
598 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
599 		 * advertised to guests via CPUID!
600 		 */
601 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
602 		0 /* DS-CPL, VMX, SMX, EST */ |
603 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
604 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
605 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
606 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
607 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
608 		F(F16C) | F(RDRAND)
609 	);
610 	/* KVM emulates x2apic in software irrespective of host support. */
611 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
612 
613 	kvm_cpu_cap_mask(CPUID_1_EDX,
614 		F(FPU) | F(VME) | F(DE) | F(PSE) |
615 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
616 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
617 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
618 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
619 		0 /* Reserved, DS, ACPI */ | F(MMX) |
620 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
621 		0 /* HTT, TM, Reserved, PBE */
622 	);
623 
624 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
625 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
626 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
627 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
628 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
629 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
630 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
631 		F(AVX512VL));
632 
633 	kvm_cpu_cap_mask(CPUID_7_ECX,
634 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
635 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
636 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
637 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
638 		F(SGX_LC) | F(BUS_LOCK_DETECT)
639 	);
640 	/* Set LA57 based on hardware capability. */
641 	if (cpuid_ecx(7) & F(LA57))
642 		kvm_cpu_cap_set(X86_FEATURE_LA57);
643 
644 	/*
645 	 * PKU not yet implemented for shadow paging and requires OSPKE
646 	 * to be set on the host. Clear it if that is not the case
647 	 */
648 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
649 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
650 
651 	kvm_cpu_cap_mask(CPUID_7_EDX,
652 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
653 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
654 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
655 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
656 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16) | F(FLUSH_L1D)
657 	);
658 
659 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
660 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
661 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
662 
663 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
664 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
665 	if (boot_cpu_has(X86_FEATURE_STIBP))
666 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
667 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
668 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
669 
670 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
671 		F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) |
672 		F(FZRM) | F(FSRS) | F(FSRC) |
673 		F(AMX_FP16) | F(AVX_IFMA)
674 	);
675 
676 	kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
677 		F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(PREFETCHITI) |
678 		F(AMX_COMPLEX)
679 	);
680 
681 	kvm_cpu_cap_init_kvm_defined(CPUID_7_2_EDX,
682 		F(INTEL_PSFD) | F(IPRED_CTRL) | F(RRSBA_CTRL) | F(DDPD_U) |
683 		F(BHI_CTRL) | F(MCDT_NO)
684 	);
685 
686 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
687 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
688 	);
689 
690 	kvm_cpu_cap_init_kvm_defined(CPUID_12_EAX,
691 		SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
692 	);
693 
694 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
695 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
696 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
697 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
698 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
699 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
700 	);
701 
702 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
703 		F(FPU) | F(VME) | F(DE) | F(PSE) |
704 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
705 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
706 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
707 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
708 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
709 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
710 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
711 	);
712 
713 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
714 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
715 
716 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0007_EDX,
717 		SF(CONSTANT_TSC)
718 	);
719 
720 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
721 		F(CLZERO) | F(XSAVEERPTR) |
722 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
723 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
724 		F(AMD_PSFD)
725 	);
726 
727 	/*
728 	 * AMD has separate bits for each SPEC_CTRL bit.
729 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
730 	 * record that in cpufeatures so use them.
731 	 */
732 	if (boot_cpu_has(X86_FEATURE_IBPB))
733 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
734 	if (boot_cpu_has(X86_FEATURE_IBRS))
735 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
736 	if (boot_cpu_has(X86_FEATURE_STIBP))
737 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
738 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
739 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
740 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
741 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
742 	/*
743 	 * The preference is to use SPEC CTRL MSR instead of the
744 	 * VIRT_SPEC MSR.
745 	 */
746 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
747 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
748 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
749 
750 	/*
751 	 * Hide all SVM features by default, SVM will set the cap bits for
752 	 * features it emulates and/or exposes for L1.
753 	 */
754 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
755 
756 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
757 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
758 		F(SME_COHERENT));
759 
760 	kvm_cpu_cap_mask(CPUID_8000_0021_EAX,
761 		F(NO_NESTED_DATA_BP) | F(LFENCE_RDTSC) | 0 /* SmmPgCfgLock */ |
762 		F(NULL_SEL_CLR_BASE) | F(AUTOIBRS) | 0 /* PrefetchCtlMsr */
763 	);
764 
765 	if (cpu_feature_enabled(X86_FEATURE_SRSO_NO))
766 		kvm_cpu_cap_set(X86_FEATURE_SRSO_NO);
767 
768 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0022_EAX,
769 		F(PERFMON_V2)
770 	);
771 
772 	/*
773 	 * Synthesize "LFENCE is serializing" into the AMD-defined entry in
774 	 * KVM's supported CPUID if the feature is reported as supported by the
775 	 * kernel.  LFENCE_RDTSC was a Linux-defined synthetic feature long
776 	 * before AMD joined the bandwagon, e.g. LFENCE is serializing on most
777 	 * CPUs that support SSE2.  On CPUs that don't support AMD's leaf,
778 	 * kvm_cpu_cap_mask() will unfortunately drop the flag due to ANDing
779 	 * the mask with the raw host CPUID, and reporting support in AMD's
780 	 * leaf can make it easier for userspace to detect the feature.
781 	 */
782 	if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
783 		kvm_cpu_cap_set(X86_FEATURE_LFENCE_RDTSC);
784 	if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
785 		kvm_cpu_cap_set(X86_FEATURE_NULL_SEL_CLR_BASE);
786 	kvm_cpu_cap_set(X86_FEATURE_NO_SMM_CTL_MSR);
787 
788 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
789 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
790 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
791 		F(PMM) | F(PMM_EN)
792 	);
793 
794 	/*
795 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
796 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
797 	 * should never happen, but the guest will likely crash if RDTSCP or
798 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
799 	 * the past.  For example, the sanity check may fire if this instance of
800 	 * KVM is running as L1 on top of an older, broken KVM.
801 	 */
802 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
803 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
804 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
805 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
806 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
807 	}
808 }
809 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
810 
811 struct kvm_cpuid_array {
812 	struct kvm_cpuid_entry2 *entries;
813 	int maxnent;
814 	int nent;
815 };
816 
817 static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
818 {
819 	if (array->nent >= array->maxnent)
820 		return NULL;
821 
822 	return &array->entries[array->nent++];
823 }
824 
825 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
826 					      u32 function, u32 index)
827 {
828 	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
829 
830 	if (!entry)
831 		return NULL;
832 
833 	memset(entry, 0, sizeof(*entry));
834 	entry->function = function;
835 	entry->index = index;
836 	switch (function & 0xC0000000) {
837 	case 0x40000000:
838 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
839 		return entry;
840 
841 	case 0x80000000:
842 		/*
843 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
844 		 * would result in out-of-bounds calls to do_host_cpuid.
845 		 */
846 		{
847 			static int max_cpuid_80000000;
848 			if (!READ_ONCE(max_cpuid_80000000))
849 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
850 			if (function > READ_ONCE(max_cpuid_80000000))
851 				return entry;
852 		}
853 		break;
854 
855 	default:
856 		break;
857 	}
858 
859 	cpuid_count(entry->function, entry->index,
860 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
861 
862 	if (cpuid_function_is_indexed(function))
863 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
864 
865 	return entry;
866 }
867 
868 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
869 {
870 	struct kvm_cpuid_entry2 *entry;
871 
872 	if (array->nent >= array->maxnent)
873 		return -E2BIG;
874 
875 	entry = &array->entries[array->nent];
876 	entry->function = func;
877 	entry->index = 0;
878 	entry->flags = 0;
879 
880 	switch (func) {
881 	case 0:
882 		entry->eax = 7;
883 		++array->nent;
884 		break;
885 	case 1:
886 		entry->ecx = F(MOVBE);
887 		++array->nent;
888 		break;
889 	case 7:
890 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
891 		entry->eax = 0;
892 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
893 			entry->ecx = F(RDPID);
894 		++array->nent;
895 		break;
896 	default:
897 		break;
898 	}
899 
900 	return 0;
901 }
902 
903 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
904 {
905 	struct kvm_cpuid_entry2 *entry;
906 	int r, i, max_idx;
907 
908 	/* all calls to cpuid_count() should be made on the same cpu */
909 	get_cpu();
910 
911 	r = -E2BIG;
912 
913 	entry = do_host_cpuid(array, function, 0);
914 	if (!entry)
915 		goto out;
916 
917 	switch (function) {
918 	case 0:
919 		/* Limited to the highest leaf implemented in KVM. */
920 		entry->eax = min(entry->eax, 0x1fU);
921 		break;
922 	case 1:
923 		cpuid_entry_override(entry, CPUID_1_EDX);
924 		cpuid_entry_override(entry, CPUID_1_ECX);
925 		break;
926 	case 2:
927 		/*
928 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
929 		 * CPUID(function=2, index=0) may return different results each
930 		 * time, with the least-significant byte in EAX enumerating the
931 		 * number of times software should do CPUID(2, 0).
932 		 *
933 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
934 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
935 		 * return 01H. Software should ignore this value and not
936 		 * interpret it as an informational descriptor", while AMD's
937 		 * APM states that CPUID(2) is reserved.
938 		 *
939 		 * WARN if a frankenstein CPU that supports virtualization and
940 		 * a stateful CPUID.0x2 is encountered.
941 		 */
942 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
943 		break;
944 	/* functions 4 and 0x8000001d have additional index. */
945 	case 4:
946 	case 0x8000001d:
947 		/*
948 		 * Read entries until the cache type in the previous entry is
949 		 * zero, i.e. indicates an invalid entry.
950 		 */
951 		for (i = 1; entry->eax & 0x1f; ++i) {
952 			entry = do_host_cpuid(array, function, i);
953 			if (!entry)
954 				goto out;
955 		}
956 		break;
957 	case 6: /* Thermal management */
958 		entry->eax = 0x4; /* allow ARAT */
959 		entry->ebx = 0;
960 		entry->ecx = 0;
961 		entry->edx = 0;
962 		break;
963 	/* function 7 has additional index. */
964 	case 7:
965 		max_idx = entry->eax = min(entry->eax, 2u);
966 		cpuid_entry_override(entry, CPUID_7_0_EBX);
967 		cpuid_entry_override(entry, CPUID_7_ECX);
968 		cpuid_entry_override(entry, CPUID_7_EDX);
969 
970 		/* KVM only supports up to 0x7.2, capped above via min(). */
971 		if (max_idx >= 1) {
972 			entry = do_host_cpuid(array, function, 1);
973 			if (!entry)
974 				goto out;
975 
976 			cpuid_entry_override(entry, CPUID_7_1_EAX);
977 			cpuid_entry_override(entry, CPUID_7_1_EDX);
978 			entry->ebx = 0;
979 			entry->ecx = 0;
980 		}
981 		if (max_idx >= 2) {
982 			entry = do_host_cpuid(array, function, 2);
983 			if (!entry)
984 				goto out;
985 
986 			cpuid_entry_override(entry, CPUID_7_2_EDX);
987 			entry->ecx = 0;
988 			entry->ebx = 0;
989 			entry->eax = 0;
990 		}
991 		break;
992 	case 0xa: { /* Architectural Performance Monitoring */
993 		union cpuid10_eax eax;
994 		union cpuid10_edx edx;
995 
996 		if (!enable_pmu || !static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
997 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
998 			break;
999 		}
1000 
1001 		eax.split.version_id = kvm_pmu_cap.version;
1002 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
1003 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
1004 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
1005 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
1006 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
1007 
1008 		if (kvm_pmu_cap.version)
1009 			edx.split.anythread_deprecated = 1;
1010 		edx.split.reserved1 = 0;
1011 		edx.split.reserved2 = 0;
1012 
1013 		entry->eax = eax.full;
1014 		entry->ebx = kvm_pmu_cap.events_mask;
1015 		entry->ecx = 0;
1016 		entry->edx = edx.full;
1017 		break;
1018 	}
1019 	case 0x1f:
1020 	case 0xb:
1021 		/*
1022 		 * No topology; a valid topology is indicated by the presence
1023 		 * of subleaf 1.
1024 		 */
1025 		entry->eax = entry->ebx = entry->ecx = 0;
1026 		break;
1027 	case 0xd: {
1028 		u64 permitted_xcr0 = kvm_get_filtered_xcr0();
1029 		u64 permitted_xss = kvm_caps.supported_xss;
1030 
1031 		entry->eax &= permitted_xcr0;
1032 		entry->ebx = xstate_required_size(permitted_xcr0, false);
1033 		entry->ecx = entry->ebx;
1034 		entry->edx &= permitted_xcr0 >> 32;
1035 		if (!permitted_xcr0)
1036 			break;
1037 
1038 		entry = do_host_cpuid(array, function, 1);
1039 		if (!entry)
1040 			goto out;
1041 
1042 		cpuid_entry_override(entry, CPUID_D_1_EAX);
1043 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
1044 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
1045 							  true);
1046 		else {
1047 			WARN_ON_ONCE(permitted_xss != 0);
1048 			entry->ebx = 0;
1049 		}
1050 		entry->ecx &= permitted_xss;
1051 		entry->edx &= permitted_xss >> 32;
1052 
1053 		for (i = 2; i < 64; ++i) {
1054 			bool s_state;
1055 			if (permitted_xcr0 & BIT_ULL(i))
1056 				s_state = false;
1057 			else if (permitted_xss & BIT_ULL(i))
1058 				s_state = true;
1059 			else
1060 				continue;
1061 
1062 			entry = do_host_cpuid(array, function, i);
1063 			if (!entry)
1064 				goto out;
1065 
1066 			/*
1067 			 * The supported check above should have filtered out
1068 			 * invalid sub-leafs.  Only valid sub-leafs should
1069 			 * reach this point, and they should have a non-zero
1070 			 * save state size.  Furthermore, check whether the
1071 			 * processor agrees with permitted_xcr0/permitted_xss
1072 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1073 			 */
1074 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1075 				--array->nent;
1076 				continue;
1077 			}
1078 
1079 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1080 				entry->ecx &= ~BIT_ULL(2);
1081 			entry->edx = 0;
1082 		}
1083 		break;
1084 	}
1085 	case 0x12:
1086 		/* Intel SGX */
1087 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1088 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1089 			break;
1090 		}
1091 
1092 		/*
1093 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1094 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1095 		 * are restricted by kernel and KVM capabilities (like most
1096 		 * feature flags), while enclave size is unrestricted.
1097 		 */
1098 		cpuid_entry_override(entry, CPUID_12_EAX);
1099 		entry->ebx &= SGX_MISC_EXINFO;
1100 
1101 		entry = do_host_cpuid(array, function, 1);
1102 		if (!entry)
1103 			goto out;
1104 
1105 		/*
1106 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1107 		 * feature flags.  Advertise all supported flags, including
1108 		 * privileged attributes that require explicit opt-in from
1109 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1110 		 * expected to derive it from supported XCR0.
1111 		 */
1112 		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1113 		entry->ebx &= 0;
1114 		break;
1115 	/* Intel PT */
1116 	case 0x14:
1117 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1118 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1119 			break;
1120 		}
1121 
1122 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1123 			if (!do_host_cpuid(array, function, i))
1124 				goto out;
1125 		}
1126 		break;
1127 	/* Intel AMX TILE */
1128 	case 0x1d:
1129 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1130 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1131 			break;
1132 		}
1133 
1134 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1135 			if (!do_host_cpuid(array, function, i))
1136 				goto out;
1137 		}
1138 		break;
1139 	case 0x1e: /* TMUL information */
1140 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1141 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1142 			break;
1143 		}
1144 		break;
1145 	case KVM_CPUID_SIGNATURE: {
1146 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1147 		entry->eax = KVM_CPUID_FEATURES;
1148 		entry->ebx = sigptr[0];
1149 		entry->ecx = sigptr[1];
1150 		entry->edx = sigptr[2];
1151 		break;
1152 	}
1153 	case KVM_CPUID_FEATURES:
1154 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1155 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1156 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1157 			     (1 << KVM_FEATURE_ASYNC_PF) |
1158 			     (1 << KVM_FEATURE_PV_EOI) |
1159 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1160 			     (1 << KVM_FEATURE_PV_UNHALT) |
1161 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1162 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1163 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1164 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1165 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1166 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1167 
1168 		if (sched_info_on())
1169 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1170 
1171 		entry->ebx = 0;
1172 		entry->ecx = 0;
1173 		entry->edx = 0;
1174 		break;
1175 	case 0x80000000:
1176 		entry->eax = min(entry->eax, 0x80000022);
1177 		/*
1178 		 * Serializing LFENCE is reported in a multitude of ways, and
1179 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1180 		 * userspace by providing the CPUID leaf ourselves.
1181 		 *
1182 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1183 		 * QEMU thinks that it can query the host blindly for that
1184 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1185 		 * above.  The processor merrily returns values from the
1186 		 * highest Intel leaf which QEMU tries to use as the guest's
1187 		 * 0x8000001d.  Even worse, this can result in an infinite
1188 		 * loop if said highest leaf has no subleaves indexed by ECX.
1189 		 */
1190 		if (entry->eax >= 0x8000001d &&
1191 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1192 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1193 			entry->eax = max(entry->eax, 0x80000021);
1194 		break;
1195 	case 0x80000001:
1196 		entry->ebx &= ~GENMASK(27, 16);
1197 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1198 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1199 		break;
1200 	case 0x80000005:
1201 		/*  Pass host L1 cache and TLB info. */
1202 		break;
1203 	case 0x80000006:
1204 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1205 		entry->edx &= ~GENMASK(17, 16);
1206 		break;
1207 	case 0x80000007: /* Advanced power management */
1208 		cpuid_entry_override(entry, CPUID_8000_0007_EDX);
1209 
1210 		/* mask against host */
1211 		entry->edx &= boot_cpu_data.x86_power;
1212 		entry->eax = entry->ebx = entry->ecx = 0;
1213 		break;
1214 	case 0x80000008: {
1215 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1216 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1217 		unsigned phys_as = entry->eax & 0xff;
1218 
1219 		/*
1220 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1221 		 * the guest operates in the same PA space as the host, i.e.
1222 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1223 		 * paging, too.
1224 		 *
1225 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1226 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1227 		 * the HPAs do not affect GPAs.
1228 		 */
1229 		if (!tdp_enabled)
1230 			g_phys_as = boot_cpu_data.x86_phys_bits;
1231 		else if (!g_phys_as)
1232 			g_phys_as = phys_as;
1233 
1234 		entry->eax = g_phys_as | (virt_as << 8);
1235 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1236 		entry->edx = 0;
1237 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1238 		break;
1239 	}
1240 	case 0x8000000A:
1241 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1242 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1243 			break;
1244 		}
1245 		entry->eax = 1; /* SVM revision 1 */
1246 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1247 				   ASID emulation to nested SVM */
1248 		entry->ecx = 0; /* Reserved */
1249 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1250 		break;
1251 	case 0x80000019:
1252 		entry->ecx = entry->edx = 0;
1253 		break;
1254 	case 0x8000001a:
1255 		entry->eax &= GENMASK(2, 0);
1256 		entry->ebx = entry->ecx = entry->edx = 0;
1257 		break;
1258 	case 0x8000001e:
1259 		/* Do not return host topology information.  */
1260 		entry->eax = entry->ebx = entry->ecx = 0;
1261 		entry->edx = 0; /* reserved */
1262 		break;
1263 	case 0x8000001F:
1264 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1265 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1266 		} else {
1267 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1268 			/* Clear NumVMPL since KVM does not support VMPL.  */
1269 			entry->ebx &= ~GENMASK(31, 12);
1270 			/*
1271 			 * Enumerate '0' for "PA bits reduction", the adjusted
1272 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1273 			 */
1274 			entry->ebx &= ~GENMASK(11, 6);
1275 		}
1276 		break;
1277 	case 0x80000020:
1278 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1279 		break;
1280 	case 0x80000021:
1281 		entry->ebx = entry->ecx = entry->edx = 0;
1282 		cpuid_entry_override(entry, CPUID_8000_0021_EAX);
1283 		break;
1284 	/* AMD Extended Performance Monitoring and Debug */
1285 	case 0x80000022: {
1286 		union cpuid_0x80000022_ebx ebx;
1287 
1288 		entry->ecx = entry->edx = 0;
1289 		if (!enable_pmu || !kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2)) {
1290 			entry->eax = entry->ebx;
1291 			break;
1292 		}
1293 
1294 		cpuid_entry_override(entry, CPUID_8000_0022_EAX);
1295 
1296 		if (kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2))
1297 			ebx.split.num_core_pmc = kvm_pmu_cap.num_counters_gp;
1298 		else if (kvm_cpu_cap_has(X86_FEATURE_PERFCTR_CORE))
1299 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS_CORE;
1300 		else
1301 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS;
1302 
1303 		entry->ebx = ebx.full;
1304 		break;
1305 	}
1306 	/*Add support for Centaur's CPUID instruction*/
1307 	case 0xC0000000:
1308 		/*Just support up to 0xC0000004 now*/
1309 		entry->eax = min(entry->eax, 0xC0000004);
1310 		break;
1311 	case 0xC0000001:
1312 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1313 		break;
1314 	case 3: /* Processor serial number */
1315 	case 5: /* MONITOR/MWAIT */
1316 	case 0xC0000002:
1317 	case 0xC0000003:
1318 	case 0xC0000004:
1319 	default:
1320 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1321 		break;
1322 	}
1323 
1324 	r = 0;
1325 
1326 out:
1327 	put_cpu();
1328 
1329 	return r;
1330 }
1331 
1332 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1333 			 unsigned int type)
1334 {
1335 	if (type == KVM_GET_EMULATED_CPUID)
1336 		return __do_cpuid_func_emulated(array, func);
1337 
1338 	return __do_cpuid_func(array, func);
1339 }
1340 
1341 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1342 
1343 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1344 			  unsigned int type)
1345 {
1346 	u32 limit;
1347 	int r;
1348 
1349 	if (func == CENTAUR_CPUID_SIGNATURE &&
1350 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1351 		return 0;
1352 
1353 	r = do_cpuid_func(array, func, type);
1354 	if (r)
1355 		return r;
1356 
1357 	limit = array->entries[array->nent - 1].eax;
1358 	for (func = func + 1; func <= limit; ++func) {
1359 		r = do_cpuid_func(array, func, type);
1360 		if (r)
1361 			break;
1362 	}
1363 
1364 	return r;
1365 }
1366 
1367 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1368 				 __u32 num_entries, unsigned int ioctl_type)
1369 {
1370 	int i;
1371 	__u32 pad[3];
1372 
1373 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1374 		return false;
1375 
1376 	/*
1377 	 * We want to make sure that ->padding is being passed clean from
1378 	 * userspace in case we want to use it for something in the future.
1379 	 *
1380 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1381 	 * have to give ourselves satisfied only with the emulated side. /me
1382 	 * sheds a tear.
1383 	 */
1384 	for (i = 0; i < num_entries; i++) {
1385 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1386 			return true;
1387 
1388 		if (pad[0] || pad[1] || pad[2])
1389 			return true;
1390 	}
1391 	return false;
1392 }
1393 
1394 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1395 			    struct kvm_cpuid_entry2 __user *entries,
1396 			    unsigned int type)
1397 {
1398 	static const u32 funcs[] = {
1399 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1400 	};
1401 
1402 	struct kvm_cpuid_array array = {
1403 		.nent = 0,
1404 	};
1405 	int r, i;
1406 
1407 	if (cpuid->nent < 1)
1408 		return -E2BIG;
1409 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1410 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1411 
1412 	if (sanity_check_entries(entries, cpuid->nent, type))
1413 		return -EINVAL;
1414 
1415 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1416 	if (!array.entries)
1417 		return -ENOMEM;
1418 
1419 	array.maxnent = cpuid->nent;
1420 
1421 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1422 		r = get_cpuid_func(&array, funcs[i], type);
1423 		if (r)
1424 			goto out_free;
1425 	}
1426 	cpuid->nent = array.nent;
1427 
1428 	if (copy_to_user(entries, array.entries,
1429 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1430 		r = -EFAULT;
1431 
1432 out_free:
1433 	kvfree(array.entries);
1434 	return r;
1435 }
1436 
1437 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1438 						    u32 function, u32 index)
1439 {
1440 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1441 				 function, index);
1442 }
1443 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1444 
1445 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1446 					      u32 function)
1447 {
1448 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1449 				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1450 }
1451 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1452 
1453 /*
1454  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1455  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1456  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1457  * range.  Centaur/VIA follows Intel semantics.
1458  *
1459  * A leaf is considered out-of-range if its function is higher than the maximum
1460  * supported leaf of its associated class or if its associated class does not
1461  * exist.
1462  *
1463  * There are three primary classes to be considered, with their respective
1464  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1465  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1466  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1467  *
1468  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1469  *  - Hypervisor: 0x40000000 - 0x4fffffff
1470  *  - Extended:   0x80000000 - 0xbfffffff
1471  *  - Centaur:    0xc0000000 - 0xcfffffff
1472  *
1473  * The Hypervisor class is further subdivided into sub-classes that each act as
1474  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1475  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1476  * CPUID sub-classes are:
1477  *
1478  *  - HyperV:     0x40000000 - 0x400000ff
1479  *  - KVM:        0x40000100 - 0x400001ff
1480  */
1481 static struct kvm_cpuid_entry2 *
1482 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1483 {
1484 	struct kvm_cpuid_entry2 *basic, *class;
1485 	u32 function = *fn_ptr;
1486 
1487 	basic = kvm_find_cpuid_entry(vcpu, 0);
1488 	if (!basic)
1489 		return NULL;
1490 
1491 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1492 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1493 		return NULL;
1494 
1495 	if (function >= 0x40000000 && function <= 0x4fffffff)
1496 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1497 	else if (function >= 0xc0000000)
1498 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1499 	else
1500 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1501 
1502 	if (class && function <= class->eax)
1503 		return NULL;
1504 
1505 	/*
1506 	 * Leaf specific adjustments are also applied when redirecting to the
1507 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1508 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1509 	 * needs to be pulled from CPUID.0xb.1.
1510 	 */
1511 	*fn_ptr = basic->eax;
1512 
1513 	/*
1514 	 * The class does not exist or the requested function is out of range;
1515 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1516 	 * the original requested leaf is observed!
1517 	 */
1518 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1519 }
1520 
1521 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1522 	       u32 *ecx, u32 *edx, bool exact_only)
1523 {
1524 	u32 orig_function = *eax, function = *eax, index = *ecx;
1525 	struct kvm_cpuid_entry2 *entry;
1526 	bool exact, used_max_basic = false;
1527 
1528 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1529 	exact = !!entry;
1530 
1531 	if (!entry && !exact_only) {
1532 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1533 		used_max_basic = !!entry;
1534 	}
1535 
1536 	if (entry) {
1537 		*eax = entry->eax;
1538 		*ebx = entry->ebx;
1539 		*ecx = entry->ecx;
1540 		*edx = entry->edx;
1541 		if (function == 7 && index == 0) {
1542 			u64 data;
1543 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1544 			    (data & TSX_CTRL_CPUID_CLEAR))
1545 				*ebx &= ~(F(RTM) | F(HLE));
1546 		} else if (function == 0x80000007) {
1547 			if (kvm_hv_invtsc_suppressed(vcpu))
1548 				*edx &= ~SF(CONSTANT_TSC);
1549 		}
1550 	} else {
1551 		*eax = *ebx = *ecx = *edx = 0;
1552 		/*
1553 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1554 		 * and EDX is always the x2APIC ID, even for undefined
1555 		 * subleaves. Index 1 will exist iff the leaf is
1556 		 * implemented, so we pass through CL iff leaf 1
1557 		 * exists. EDX can be copied from any existing index.
1558 		 */
1559 		if (function == 0xb || function == 0x1f) {
1560 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1561 			if (entry) {
1562 				*ecx = index & 0xff;
1563 				*edx = entry->edx;
1564 			}
1565 		}
1566 	}
1567 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1568 			used_max_basic);
1569 	return exact;
1570 }
1571 EXPORT_SYMBOL_GPL(kvm_cpuid);
1572 
1573 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1574 {
1575 	u32 eax, ebx, ecx, edx;
1576 
1577 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1578 		return 1;
1579 
1580 	eax = kvm_rax_read(vcpu);
1581 	ecx = kvm_rcx_read(vcpu);
1582 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1583 	kvm_rax_write(vcpu, eax);
1584 	kvm_rbx_write(vcpu, ebx);
1585 	kvm_rcx_write(vcpu, ecx);
1586 	kvm_rdx_write(vcpu, edx);
1587 	return kvm_skip_emulated_instruction(vcpu);
1588 }
1589 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1590