xref: /linux/arch/x86/kvm/cpuid.c (revision f4738f56d1dc62aaba69b33702a5ab098f1b8c63)
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 	/*
364 	 * FP+SSE can always be saved/restored via KVM_{G,S}ET_XSAVE, even if
365 	 * XSAVE/XCRO are not exposed to the guest, and even if XSAVE isn't
366 	 * supported by the host.
367 	 */
368 	vcpu->arch.guest_fpu.fpstate->user_xfeatures = vcpu->arch.guest_supported_xcr0 |
369 						       XFEATURE_MASK_FPSSE;
370 
371 	kvm_update_pv_runtime(vcpu);
372 
373 	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
374 	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
375 
376 	kvm_pmu_refresh(vcpu);
377 	vcpu->arch.cr4_guest_rsvd_bits =
378 	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
379 
380 	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu->arch.cpuid_entries,
381 						    vcpu->arch.cpuid_nent));
382 
383 	/* Invoke the vendor callback only after the above state is updated. */
384 	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
385 
386 	/*
387 	 * Except for the MMU, which needs to do its thing any vendor specific
388 	 * adjustments to the reserved GPA bits.
389 	 */
390 	kvm_mmu_after_set_cpuid(vcpu);
391 }
392 
393 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
394 {
395 	struct kvm_cpuid_entry2 *best;
396 
397 	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
398 	if (!best || best->eax < 0x80000008)
399 		goto not_found;
400 	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
401 	if (best)
402 		return best->eax & 0xff;
403 not_found:
404 	return 36;
405 }
406 
407 /*
408  * This "raw" version returns the reserved GPA bits without any adjustments for
409  * encryption technologies that usurp bits.  The raw mask should be used if and
410  * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
411  */
412 u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
413 {
414 	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
415 }
416 
417 static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
418                         int nent)
419 {
420 	int r;
421 
422 	__kvm_update_cpuid_runtime(vcpu, e2, nent);
423 
424 	/*
425 	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
426 	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
427 	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
428 	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
429 	 * the core vCPU model on the fly. It would've been better to forbid any
430 	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
431 	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
432 	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
433 	 * whether the supplied CPUID data is equal to what's already set.
434 	 */
435 	if (kvm_vcpu_has_run(vcpu)) {
436 		r = kvm_cpuid_check_equal(vcpu, e2, nent);
437 		if (r)
438 			return r;
439 
440 		kvfree(e2);
441 		return 0;
442 	}
443 
444 	if (kvm_cpuid_has_hyperv(e2, nent)) {
445 		r = kvm_hv_vcpu_init(vcpu);
446 		if (r)
447 			return r;
448 	}
449 
450 	r = kvm_check_cpuid(vcpu, e2, nent);
451 	if (r)
452 		return r;
453 
454 	kvfree(vcpu->arch.cpuid_entries);
455 	vcpu->arch.cpuid_entries = e2;
456 	vcpu->arch.cpuid_nent = nent;
457 
458 	vcpu->arch.kvm_cpuid = kvm_get_hypervisor_cpuid(vcpu, KVM_SIGNATURE);
459 	vcpu->arch.xen.cpuid = kvm_get_hypervisor_cpuid(vcpu, XEN_SIGNATURE);
460 	kvm_vcpu_after_set_cpuid(vcpu);
461 
462 	return 0;
463 }
464 
465 /* when an old userspace process fills a new kernel module */
466 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
467 			     struct kvm_cpuid *cpuid,
468 			     struct kvm_cpuid_entry __user *entries)
469 {
470 	int r, i;
471 	struct kvm_cpuid_entry *e = NULL;
472 	struct kvm_cpuid_entry2 *e2 = NULL;
473 
474 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
475 		return -E2BIG;
476 
477 	if (cpuid->nent) {
478 		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
479 		if (IS_ERR(e))
480 			return PTR_ERR(e);
481 
482 		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
483 		if (!e2) {
484 			r = -ENOMEM;
485 			goto out_free_cpuid;
486 		}
487 	}
488 	for (i = 0; i < cpuid->nent; i++) {
489 		e2[i].function = e[i].function;
490 		e2[i].eax = e[i].eax;
491 		e2[i].ebx = e[i].ebx;
492 		e2[i].ecx = e[i].ecx;
493 		e2[i].edx = e[i].edx;
494 		e2[i].index = 0;
495 		e2[i].flags = 0;
496 		e2[i].padding[0] = 0;
497 		e2[i].padding[1] = 0;
498 		e2[i].padding[2] = 0;
499 	}
500 
501 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
502 	if (r)
503 		kvfree(e2);
504 
505 out_free_cpuid:
506 	kvfree(e);
507 
508 	return r;
509 }
510 
511 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
512 			      struct kvm_cpuid2 *cpuid,
513 			      struct kvm_cpuid_entry2 __user *entries)
514 {
515 	struct kvm_cpuid_entry2 *e2 = NULL;
516 	int r;
517 
518 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
519 		return -E2BIG;
520 
521 	if (cpuid->nent) {
522 		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
523 		if (IS_ERR(e2))
524 			return PTR_ERR(e2);
525 	}
526 
527 	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
528 	if (r)
529 		kvfree(e2);
530 
531 	return r;
532 }
533 
534 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
535 			      struct kvm_cpuid2 *cpuid,
536 			      struct kvm_cpuid_entry2 __user *entries)
537 {
538 	if (cpuid->nent < vcpu->arch.cpuid_nent)
539 		return -E2BIG;
540 
541 	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
542 			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
543 		return -EFAULT;
544 
545 	cpuid->nent = vcpu->arch.cpuid_nent;
546 	return 0;
547 }
548 
549 /* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
550 static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
551 {
552 	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
553 	struct kvm_cpuid_entry2 entry;
554 
555 	reverse_cpuid_check(leaf);
556 
557 	cpuid_count(cpuid.function, cpuid.index,
558 		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
559 
560 	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
561 }
562 
563 static __always_inline
564 void kvm_cpu_cap_init_kvm_defined(enum kvm_only_cpuid_leafs leaf, u32 mask)
565 {
566 	/* Use kvm_cpu_cap_mask for leafs that aren't KVM-only. */
567 	BUILD_BUG_ON(leaf < NCAPINTS);
568 
569 	kvm_cpu_caps[leaf] = mask;
570 
571 	__kvm_cpu_cap_mask(leaf);
572 }
573 
574 static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
575 {
576 	/* Use kvm_cpu_cap_init_kvm_defined for KVM-only leafs. */
577 	BUILD_BUG_ON(leaf >= NCAPINTS);
578 
579 	kvm_cpu_caps[leaf] &= mask;
580 
581 	__kvm_cpu_cap_mask(leaf);
582 }
583 
584 void kvm_set_cpu_caps(void)
585 {
586 #ifdef CONFIG_X86_64
587 	unsigned int f_gbpages = F(GBPAGES);
588 	unsigned int f_lm = F(LM);
589 	unsigned int f_xfd = F(XFD);
590 #else
591 	unsigned int f_gbpages = 0;
592 	unsigned int f_lm = 0;
593 	unsigned int f_xfd = 0;
594 #endif
595 	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
596 
597 	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
598 		     sizeof(boot_cpu_data.x86_capability));
599 
600 	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
601 	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
602 
603 	kvm_cpu_cap_mask(CPUID_1_ECX,
604 		/*
605 		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
606 		 * advertised to guests via CPUID!
607 		 */
608 		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
609 		0 /* DS-CPL, VMX, SMX, EST */ |
610 		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
611 		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
612 		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
613 		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
614 		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
615 		F(F16C) | F(RDRAND)
616 	);
617 	/* KVM emulates x2apic in software irrespective of host support. */
618 	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
619 
620 	kvm_cpu_cap_mask(CPUID_1_EDX,
621 		F(FPU) | F(VME) | F(DE) | F(PSE) |
622 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
623 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
624 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
625 		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
626 		0 /* Reserved, DS, ACPI */ | F(MMX) |
627 		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
628 		0 /* HTT, TM, Reserved, PBE */
629 	);
630 
631 	kvm_cpu_cap_mask(CPUID_7_0_EBX,
632 		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
633 		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
634 		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
635 		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
636 		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
637 		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
638 		F(AVX512VL));
639 
640 	kvm_cpu_cap_mask(CPUID_7_ECX,
641 		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
642 		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
643 		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
644 		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
645 		F(SGX_LC) | F(BUS_LOCK_DETECT)
646 	);
647 	/* Set LA57 based on hardware capability. */
648 	if (cpuid_ecx(7) & F(LA57))
649 		kvm_cpu_cap_set(X86_FEATURE_LA57);
650 
651 	/*
652 	 * PKU not yet implemented for shadow paging and requires OSPKE
653 	 * to be set on the host. Clear it if that is not the case
654 	 */
655 	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
656 		kvm_cpu_cap_clear(X86_FEATURE_PKU);
657 
658 	kvm_cpu_cap_mask(CPUID_7_EDX,
659 		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
660 		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
661 		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
662 		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
663 		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16) | F(FLUSH_L1D)
664 	);
665 
666 	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
667 	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
668 	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
669 
670 	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
671 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
672 	if (boot_cpu_has(X86_FEATURE_STIBP))
673 		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
674 	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
675 		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
676 
677 	kvm_cpu_cap_mask(CPUID_7_1_EAX,
678 		F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) |
679 		F(FZRM) | F(FSRS) | F(FSRC) |
680 		F(AMX_FP16) | F(AVX_IFMA)
681 	);
682 
683 	kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
684 		F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(PREFETCHITI) |
685 		F(AMX_COMPLEX)
686 	);
687 
688 	kvm_cpu_cap_mask(CPUID_D_1_EAX,
689 		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
690 	);
691 
692 	kvm_cpu_cap_init_kvm_defined(CPUID_12_EAX,
693 		SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
694 	);
695 
696 	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
697 		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
698 		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
699 		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
700 		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
701 		F(TOPOEXT) | 0 /* PERFCTR_CORE */
702 	);
703 
704 	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
705 		F(FPU) | F(VME) | F(DE) | F(PSE) |
706 		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
707 		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
708 		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
709 		F(PAT) | F(PSE36) | 0 /* Reserved */ |
710 		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
711 		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
712 		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
713 	);
714 
715 	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
716 		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
717 
718 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0007_EDX,
719 		SF(CONSTANT_TSC)
720 	);
721 
722 	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
723 		F(CLZERO) | F(XSAVEERPTR) |
724 		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
725 		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
726 		F(AMD_PSFD)
727 	);
728 
729 	/*
730 	 * AMD has separate bits for each SPEC_CTRL bit.
731 	 * arch/x86/kernel/cpu/bugs.c is kind enough to
732 	 * record that in cpufeatures so use them.
733 	 */
734 	if (boot_cpu_has(X86_FEATURE_IBPB))
735 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
736 	if (boot_cpu_has(X86_FEATURE_IBRS))
737 		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
738 	if (boot_cpu_has(X86_FEATURE_STIBP))
739 		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
740 	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
741 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
742 	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
743 		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
744 	/*
745 	 * The preference is to use SPEC CTRL MSR instead of the
746 	 * VIRT_SPEC MSR.
747 	 */
748 	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
749 	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
750 		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
751 
752 	/*
753 	 * Hide all SVM features by default, SVM will set the cap bits for
754 	 * features it emulates and/or exposes for L1.
755 	 */
756 	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
757 
758 	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
759 		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
760 		F(SME_COHERENT));
761 
762 	kvm_cpu_cap_mask(CPUID_8000_0021_EAX,
763 		F(NO_NESTED_DATA_BP) | F(LFENCE_RDTSC) | 0 /* SmmPgCfgLock */ |
764 		F(NULL_SEL_CLR_BASE) | F(AUTOIBRS) | 0 /* PrefetchCtlMsr */
765 	);
766 
767 	if (cpu_feature_enabled(X86_FEATURE_SRSO_NO))
768 		kvm_cpu_cap_set(X86_FEATURE_SRSO_NO);
769 
770 	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0022_EAX,
771 		F(PERFMON_V2)
772 	);
773 
774 	/*
775 	 * Synthesize "LFENCE is serializing" into the AMD-defined entry in
776 	 * KVM's supported CPUID if the feature is reported as supported by the
777 	 * kernel.  LFENCE_RDTSC was a Linux-defined synthetic feature long
778 	 * before AMD joined the bandwagon, e.g. LFENCE is serializing on most
779 	 * CPUs that support SSE2.  On CPUs that don't support AMD's leaf,
780 	 * kvm_cpu_cap_mask() will unfortunately drop the flag due to ANDing
781 	 * the mask with the raw host CPUID, and reporting support in AMD's
782 	 * leaf can make it easier for userspace to detect the feature.
783 	 */
784 	if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
785 		kvm_cpu_cap_set(X86_FEATURE_LFENCE_RDTSC);
786 	if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
787 		kvm_cpu_cap_set(X86_FEATURE_NULL_SEL_CLR_BASE);
788 	kvm_cpu_cap_set(X86_FEATURE_NO_SMM_CTL_MSR);
789 
790 	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
791 		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
792 		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
793 		F(PMM) | F(PMM_EN)
794 	);
795 
796 	/*
797 	 * Hide RDTSCP and RDPID if either feature is reported as supported but
798 	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
799 	 * should never happen, but the guest will likely crash if RDTSCP or
800 	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
801 	 * the past.  For example, the sanity check may fire if this instance of
802 	 * KVM is running as L1 on top of an older, broken KVM.
803 	 */
804 	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
805 		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
806 		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
807 		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
808 		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
809 	}
810 }
811 EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
812 
813 struct kvm_cpuid_array {
814 	struct kvm_cpuid_entry2 *entries;
815 	int maxnent;
816 	int nent;
817 };
818 
819 static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
820 {
821 	if (array->nent >= array->maxnent)
822 		return NULL;
823 
824 	return &array->entries[array->nent++];
825 }
826 
827 static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
828 					      u32 function, u32 index)
829 {
830 	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
831 
832 	if (!entry)
833 		return NULL;
834 
835 	memset(entry, 0, sizeof(*entry));
836 	entry->function = function;
837 	entry->index = index;
838 	switch (function & 0xC0000000) {
839 	case 0x40000000:
840 		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
841 		return entry;
842 
843 	case 0x80000000:
844 		/*
845 		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
846 		 * would result in out-of-bounds calls to do_host_cpuid.
847 		 */
848 		{
849 			static int max_cpuid_80000000;
850 			if (!READ_ONCE(max_cpuid_80000000))
851 				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
852 			if (function > READ_ONCE(max_cpuid_80000000))
853 				return entry;
854 		}
855 		break;
856 
857 	default:
858 		break;
859 	}
860 
861 	cpuid_count(entry->function, entry->index,
862 		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
863 
864 	if (cpuid_function_is_indexed(function))
865 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
866 
867 	return entry;
868 }
869 
870 static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
871 {
872 	struct kvm_cpuid_entry2 *entry;
873 
874 	if (array->nent >= array->maxnent)
875 		return -E2BIG;
876 
877 	entry = &array->entries[array->nent];
878 	entry->function = func;
879 	entry->index = 0;
880 	entry->flags = 0;
881 
882 	switch (func) {
883 	case 0:
884 		entry->eax = 7;
885 		++array->nent;
886 		break;
887 	case 1:
888 		entry->ecx = F(MOVBE);
889 		++array->nent;
890 		break;
891 	case 7:
892 		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
893 		entry->eax = 0;
894 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
895 			entry->ecx = F(RDPID);
896 		++array->nent;
897 		break;
898 	default:
899 		break;
900 	}
901 
902 	return 0;
903 }
904 
905 static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
906 {
907 	struct kvm_cpuid_entry2 *entry;
908 	int r, i, max_idx;
909 
910 	/* all calls to cpuid_count() should be made on the same cpu */
911 	get_cpu();
912 
913 	r = -E2BIG;
914 
915 	entry = do_host_cpuid(array, function, 0);
916 	if (!entry)
917 		goto out;
918 
919 	switch (function) {
920 	case 0:
921 		/* Limited to the highest leaf implemented in KVM. */
922 		entry->eax = min(entry->eax, 0x1fU);
923 		break;
924 	case 1:
925 		cpuid_entry_override(entry, CPUID_1_EDX);
926 		cpuid_entry_override(entry, CPUID_1_ECX);
927 		break;
928 	case 2:
929 		/*
930 		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
931 		 * CPUID(function=2, index=0) may return different results each
932 		 * time, with the least-significant byte in EAX enumerating the
933 		 * number of times software should do CPUID(2, 0).
934 		 *
935 		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
936 		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
937 		 * return 01H. Software should ignore this value and not
938 		 * interpret it as an informational descriptor", while AMD's
939 		 * APM states that CPUID(2) is reserved.
940 		 *
941 		 * WARN if a frankenstein CPU that supports virtualization and
942 		 * a stateful CPUID.0x2 is encountered.
943 		 */
944 		WARN_ON_ONCE((entry->eax & 0xff) > 1);
945 		break;
946 	/* functions 4 and 0x8000001d have additional index. */
947 	case 4:
948 	case 0x8000001d:
949 		/*
950 		 * Read entries until the cache type in the previous entry is
951 		 * zero, i.e. indicates an invalid entry.
952 		 */
953 		for (i = 1; entry->eax & 0x1f; ++i) {
954 			entry = do_host_cpuid(array, function, i);
955 			if (!entry)
956 				goto out;
957 		}
958 		break;
959 	case 6: /* Thermal management */
960 		entry->eax = 0x4; /* allow ARAT */
961 		entry->ebx = 0;
962 		entry->ecx = 0;
963 		entry->edx = 0;
964 		break;
965 	/* function 7 has additional index. */
966 	case 7:
967 		entry->eax = min(entry->eax, 1u);
968 		cpuid_entry_override(entry, CPUID_7_0_EBX);
969 		cpuid_entry_override(entry, CPUID_7_ECX);
970 		cpuid_entry_override(entry, CPUID_7_EDX);
971 
972 		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
973 		if (entry->eax == 1) {
974 			entry = do_host_cpuid(array, function, 1);
975 			if (!entry)
976 				goto out;
977 
978 			cpuid_entry_override(entry, CPUID_7_1_EAX);
979 			cpuid_entry_override(entry, CPUID_7_1_EDX);
980 			entry->ebx = 0;
981 			entry->ecx = 0;
982 		}
983 		break;
984 	case 0xa: { /* Architectural Performance Monitoring */
985 		union cpuid10_eax eax;
986 		union cpuid10_edx edx;
987 
988 		if (!enable_pmu || !static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
989 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
990 			break;
991 		}
992 
993 		eax.split.version_id = kvm_pmu_cap.version;
994 		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
995 		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
996 		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
997 		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
998 		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
999 
1000 		if (kvm_pmu_cap.version)
1001 			edx.split.anythread_deprecated = 1;
1002 		edx.split.reserved1 = 0;
1003 		edx.split.reserved2 = 0;
1004 
1005 		entry->eax = eax.full;
1006 		entry->ebx = kvm_pmu_cap.events_mask;
1007 		entry->ecx = 0;
1008 		entry->edx = edx.full;
1009 		break;
1010 	}
1011 	case 0x1f:
1012 	case 0xb:
1013 		/*
1014 		 * No topology; a valid topology is indicated by the presence
1015 		 * of subleaf 1.
1016 		 */
1017 		entry->eax = entry->ebx = entry->ecx = 0;
1018 		break;
1019 	case 0xd: {
1020 		u64 permitted_xcr0 = kvm_get_filtered_xcr0();
1021 		u64 permitted_xss = kvm_caps.supported_xss;
1022 
1023 		entry->eax &= permitted_xcr0;
1024 		entry->ebx = xstate_required_size(permitted_xcr0, false);
1025 		entry->ecx = entry->ebx;
1026 		entry->edx &= permitted_xcr0 >> 32;
1027 		if (!permitted_xcr0)
1028 			break;
1029 
1030 		entry = do_host_cpuid(array, function, 1);
1031 		if (!entry)
1032 			goto out;
1033 
1034 		cpuid_entry_override(entry, CPUID_D_1_EAX);
1035 		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
1036 			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
1037 							  true);
1038 		else {
1039 			WARN_ON_ONCE(permitted_xss != 0);
1040 			entry->ebx = 0;
1041 		}
1042 		entry->ecx &= permitted_xss;
1043 		entry->edx &= permitted_xss >> 32;
1044 
1045 		for (i = 2; i < 64; ++i) {
1046 			bool s_state;
1047 			if (permitted_xcr0 & BIT_ULL(i))
1048 				s_state = false;
1049 			else if (permitted_xss & BIT_ULL(i))
1050 				s_state = true;
1051 			else
1052 				continue;
1053 
1054 			entry = do_host_cpuid(array, function, i);
1055 			if (!entry)
1056 				goto out;
1057 
1058 			/*
1059 			 * The supported check above should have filtered out
1060 			 * invalid sub-leafs.  Only valid sub-leafs should
1061 			 * reach this point, and they should have a non-zero
1062 			 * save state size.  Furthermore, check whether the
1063 			 * processor agrees with permitted_xcr0/permitted_xss
1064 			 * on whether this is an XCR0- or IA32_XSS-managed area.
1065 			 */
1066 			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1067 				--array->nent;
1068 				continue;
1069 			}
1070 
1071 			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1072 				entry->ecx &= ~BIT_ULL(2);
1073 			entry->edx = 0;
1074 		}
1075 		break;
1076 	}
1077 	case 0x12:
1078 		/* Intel SGX */
1079 		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1080 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1081 			break;
1082 		}
1083 
1084 		/*
1085 		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1086 		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1087 		 * are restricted by kernel and KVM capabilities (like most
1088 		 * feature flags), while enclave size is unrestricted.
1089 		 */
1090 		cpuid_entry_override(entry, CPUID_12_EAX);
1091 		entry->ebx &= SGX_MISC_EXINFO;
1092 
1093 		entry = do_host_cpuid(array, function, 1);
1094 		if (!entry)
1095 			goto out;
1096 
1097 		/*
1098 		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1099 		 * feature flags.  Advertise all supported flags, including
1100 		 * privileged attributes that require explicit opt-in from
1101 		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1102 		 * expected to derive it from supported XCR0.
1103 		 */
1104 		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1105 		entry->ebx &= 0;
1106 		break;
1107 	/* Intel PT */
1108 	case 0x14:
1109 		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1110 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1111 			break;
1112 		}
1113 
1114 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1115 			if (!do_host_cpuid(array, function, i))
1116 				goto out;
1117 		}
1118 		break;
1119 	/* Intel AMX TILE */
1120 	case 0x1d:
1121 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1122 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1123 			break;
1124 		}
1125 
1126 		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1127 			if (!do_host_cpuid(array, function, i))
1128 				goto out;
1129 		}
1130 		break;
1131 	case 0x1e: /* TMUL information */
1132 		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1133 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1134 			break;
1135 		}
1136 		break;
1137 	case KVM_CPUID_SIGNATURE: {
1138 		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1139 		entry->eax = KVM_CPUID_FEATURES;
1140 		entry->ebx = sigptr[0];
1141 		entry->ecx = sigptr[1];
1142 		entry->edx = sigptr[2];
1143 		break;
1144 	}
1145 	case KVM_CPUID_FEATURES:
1146 		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1147 			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1148 			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1149 			     (1 << KVM_FEATURE_ASYNC_PF) |
1150 			     (1 << KVM_FEATURE_PV_EOI) |
1151 			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1152 			     (1 << KVM_FEATURE_PV_UNHALT) |
1153 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1154 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1155 			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1156 			     (1 << KVM_FEATURE_POLL_CONTROL) |
1157 			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1158 			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1159 
1160 		if (sched_info_on())
1161 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1162 
1163 		entry->ebx = 0;
1164 		entry->ecx = 0;
1165 		entry->edx = 0;
1166 		break;
1167 	case 0x80000000:
1168 		entry->eax = min(entry->eax, 0x80000022);
1169 		/*
1170 		 * Serializing LFENCE is reported in a multitude of ways, and
1171 		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1172 		 * userspace by providing the CPUID leaf ourselves.
1173 		 *
1174 		 * However, only do it if the host has CPUID leaf 0x8000001d.
1175 		 * QEMU thinks that it can query the host blindly for that
1176 		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1177 		 * above.  The processor merrily returns values from the
1178 		 * highest Intel leaf which QEMU tries to use as the guest's
1179 		 * 0x8000001d.  Even worse, this can result in an infinite
1180 		 * loop if said highest leaf has no subleaves indexed by ECX.
1181 		 */
1182 		if (entry->eax >= 0x8000001d &&
1183 		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1184 		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1185 			entry->eax = max(entry->eax, 0x80000021);
1186 		break;
1187 	case 0x80000001:
1188 		entry->ebx &= ~GENMASK(27, 16);
1189 		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1190 		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1191 		break;
1192 	case 0x80000005:
1193 		/*  Pass host L1 cache and TLB info. */
1194 		break;
1195 	case 0x80000006:
1196 		/* Drop reserved bits, pass host L2 cache and TLB info. */
1197 		entry->edx &= ~GENMASK(17, 16);
1198 		break;
1199 	case 0x80000007: /* Advanced power management */
1200 		cpuid_entry_override(entry, CPUID_8000_0007_EDX);
1201 
1202 		/* mask against host */
1203 		entry->edx &= boot_cpu_data.x86_power;
1204 		entry->eax = entry->ebx = entry->ecx = 0;
1205 		break;
1206 	case 0x80000008: {
1207 		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1208 		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1209 		unsigned phys_as = entry->eax & 0xff;
1210 
1211 		/*
1212 		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1213 		 * the guest operates in the same PA space as the host, i.e.
1214 		 * reductions in MAXPHYADDR for memory encryption affect shadow
1215 		 * paging, too.
1216 		 *
1217 		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1218 		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1219 		 * the HPAs do not affect GPAs.
1220 		 */
1221 		if (!tdp_enabled)
1222 			g_phys_as = boot_cpu_data.x86_phys_bits;
1223 		else if (!g_phys_as)
1224 			g_phys_as = phys_as;
1225 
1226 		entry->eax = g_phys_as | (virt_as << 8);
1227 		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1228 		entry->edx = 0;
1229 		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1230 		break;
1231 	}
1232 	case 0x8000000A:
1233 		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1234 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1235 			break;
1236 		}
1237 		entry->eax = 1; /* SVM revision 1 */
1238 		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1239 				   ASID emulation to nested SVM */
1240 		entry->ecx = 0; /* Reserved */
1241 		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1242 		break;
1243 	case 0x80000019:
1244 		entry->ecx = entry->edx = 0;
1245 		break;
1246 	case 0x8000001a:
1247 		entry->eax &= GENMASK(2, 0);
1248 		entry->ebx = entry->ecx = entry->edx = 0;
1249 		break;
1250 	case 0x8000001e:
1251 		/* Do not return host topology information.  */
1252 		entry->eax = entry->ebx = entry->ecx = 0;
1253 		entry->edx = 0; /* reserved */
1254 		break;
1255 	case 0x8000001F:
1256 		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1257 			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1258 		} else {
1259 			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1260 			/* Clear NumVMPL since KVM does not support VMPL.  */
1261 			entry->ebx &= ~GENMASK(31, 12);
1262 			/*
1263 			 * Enumerate '0' for "PA bits reduction", the adjusted
1264 			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1265 			 */
1266 			entry->ebx &= ~GENMASK(11, 6);
1267 		}
1268 		break;
1269 	case 0x80000020:
1270 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1271 		break;
1272 	case 0x80000021:
1273 		entry->ebx = entry->ecx = entry->edx = 0;
1274 		cpuid_entry_override(entry, CPUID_8000_0021_EAX);
1275 		break;
1276 	/* AMD Extended Performance Monitoring and Debug */
1277 	case 0x80000022: {
1278 		union cpuid_0x80000022_ebx ebx;
1279 
1280 		entry->ecx = entry->edx = 0;
1281 		if (!enable_pmu || !kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2)) {
1282 			entry->eax = entry->ebx;
1283 			break;
1284 		}
1285 
1286 		cpuid_entry_override(entry, CPUID_8000_0022_EAX);
1287 
1288 		if (kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2))
1289 			ebx.split.num_core_pmc = kvm_pmu_cap.num_counters_gp;
1290 		else if (kvm_cpu_cap_has(X86_FEATURE_PERFCTR_CORE))
1291 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS_CORE;
1292 		else
1293 			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS;
1294 
1295 		entry->ebx = ebx.full;
1296 		break;
1297 	}
1298 	/*Add support for Centaur's CPUID instruction*/
1299 	case 0xC0000000:
1300 		/*Just support up to 0xC0000004 now*/
1301 		entry->eax = min(entry->eax, 0xC0000004);
1302 		break;
1303 	case 0xC0000001:
1304 		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1305 		break;
1306 	case 3: /* Processor serial number */
1307 	case 5: /* MONITOR/MWAIT */
1308 	case 0xC0000002:
1309 	case 0xC0000003:
1310 	case 0xC0000004:
1311 	default:
1312 		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1313 		break;
1314 	}
1315 
1316 	r = 0;
1317 
1318 out:
1319 	put_cpu();
1320 
1321 	return r;
1322 }
1323 
1324 static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1325 			 unsigned int type)
1326 {
1327 	if (type == KVM_GET_EMULATED_CPUID)
1328 		return __do_cpuid_func_emulated(array, func);
1329 
1330 	return __do_cpuid_func(array, func);
1331 }
1332 
1333 #define CENTAUR_CPUID_SIGNATURE 0xC0000000
1334 
1335 static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1336 			  unsigned int type)
1337 {
1338 	u32 limit;
1339 	int r;
1340 
1341 	if (func == CENTAUR_CPUID_SIGNATURE &&
1342 	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1343 		return 0;
1344 
1345 	r = do_cpuid_func(array, func, type);
1346 	if (r)
1347 		return r;
1348 
1349 	limit = array->entries[array->nent - 1].eax;
1350 	for (func = func + 1; func <= limit; ++func) {
1351 		r = do_cpuid_func(array, func, type);
1352 		if (r)
1353 			break;
1354 	}
1355 
1356 	return r;
1357 }
1358 
1359 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1360 				 __u32 num_entries, unsigned int ioctl_type)
1361 {
1362 	int i;
1363 	__u32 pad[3];
1364 
1365 	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1366 		return false;
1367 
1368 	/*
1369 	 * We want to make sure that ->padding is being passed clean from
1370 	 * userspace in case we want to use it for something in the future.
1371 	 *
1372 	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1373 	 * have to give ourselves satisfied only with the emulated side. /me
1374 	 * sheds a tear.
1375 	 */
1376 	for (i = 0; i < num_entries; i++) {
1377 		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1378 			return true;
1379 
1380 		if (pad[0] || pad[1] || pad[2])
1381 			return true;
1382 	}
1383 	return false;
1384 }
1385 
1386 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1387 			    struct kvm_cpuid_entry2 __user *entries,
1388 			    unsigned int type)
1389 {
1390 	static const u32 funcs[] = {
1391 		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1392 	};
1393 
1394 	struct kvm_cpuid_array array = {
1395 		.nent = 0,
1396 	};
1397 	int r, i;
1398 
1399 	if (cpuid->nent < 1)
1400 		return -E2BIG;
1401 	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1402 		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1403 
1404 	if (sanity_check_entries(entries, cpuid->nent, type))
1405 		return -EINVAL;
1406 
1407 	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1408 	if (!array.entries)
1409 		return -ENOMEM;
1410 
1411 	array.maxnent = cpuid->nent;
1412 
1413 	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1414 		r = get_cpuid_func(&array, funcs[i], type);
1415 		if (r)
1416 			goto out_free;
1417 	}
1418 	cpuid->nent = array.nent;
1419 
1420 	if (copy_to_user(entries, array.entries,
1421 			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1422 		r = -EFAULT;
1423 
1424 out_free:
1425 	kvfree(array.entries);
1426 	return r;
1427 }
1428 
1429 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1430 						    u32 function, u32 index)
1431 {
1432 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1433 				 function, index);
1434 }
1435 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1436 
1437 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1438 					      u32 function)
1439 {
1440 	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1441 				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1442 }
1443 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1444 
1445 /*
1446  * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1447  * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1448  * returns all zeroes for any undefined leaf, whether or not the leaf is in
1449  * range.  Centaur/VIA follows Intel semantics.
1450  *
1451  * A leaf is considered out-of-range if its function is higher than the maximum
1452  * supported leaf of its associated class or if its associated class does not
1453  * exist.
1454  *
1455  * There are three primary classes to be considered, with their respective
1456  * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1457  * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1458  * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1459  *
1460  *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1461  *  - Hypervisor: 0x40000000 - 0x4fffffff
1462  *  - Extended:   0x80000000 - 0xbfffffff
1463  *  - Centaur:    0xc0000000 - 0xcfffffff
1464  *
1465  * The Hypervisor class is further subdivided into sub-classes that each act as
1466  * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1467  * is advertising support for both HyperV and KVM, the resulting Hypervisor
1468  * CPUID sub-classes are:
1469  *
1470  *  - HyperV:     0x40000000 - 0x400000ff
1471  *  - KVM:        0x40000100 - 0x400001ff
1472  */
1473 static struct kvm_cpuid_entry2 *
1474 get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1475 {
1476 	struct kvm_cpuid_entry2 *basic, *class;
1477 	u32 function = *fn_ptr;
1478 
1479 	basic = kvm_find_cpuid_entry(vcpu, 0);
1480 	if (!basic)
1481 		return NULL;
1482 
1483 	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1484 	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1485 		return NULL;
1486 
1487 	if (function >= 0x40000000 && function <= 0x4fffffff)
1488 		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1489 	else if (function >= 0xc0000000)
1490 		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1491 	else
1492 		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1493 
1494 	if (class && function <= class->eax)
1495 		return NULL;
1496 
1497 	/*
1498 	 * Leaf specific adjustments are also applied when redirecting to the
1499 	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1500 	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1501 	 * needs to be pulled from CPUID.0xb.1.
1502 	 */
1503 	*fn_ptr = basic->eax;
1504 
1505 	/*
1506 	 * The class does not exist or the requested function is out of range;
1507 	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1508 	 * the original requested leaf is observed!
1509 	 */
1510 	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1511 }
1512 
1513 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1514 	       u32 *ecx, u32 *edx, bool exact_only)
1515 {
1516 	u32 orig_function = *eax, function = *eax, index = *ecx;
1517 	struct kvm_cpuid_entry2 *entry;
1518 	bool exact, used_max_basic = false;
1519 
1520 	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1521 	exact = !!entry;
1522 
1523 	if (!entry && !exact_only) {
1524 		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1525 		used_max_basic = !!entry;
1526 	}
1527 
1528 	if (entry) {
1529 		*eax = entry->eax;
1530 		*ebx = entry->ebx;
1531 		*ecx = entry->ecx;
1532 		*edx = entry->edx;
1533 		if (function == 7 && index == 0) {
1534 			u64 data;
1535 		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1536 			    (data & TSX_CTRL_CPUID_CLEAR))
1537 				*ebx &= ~(F(RTM) | F(HLE));
1538 		} else if (function == 0x80000007) {
1539 			if (kvm_hv_invtsc_suppressed(vcpu))
1540 				*edx &= ~SF(CONSTANT_TSC);
1541 		}
1542 	} else {
1543 		*eax = *ebx = *ecx = *edx = 0;
1544 		/*
1545 		 * When leaf 0BH or 1FH is defined, CL is pass-through
1546 		 * and EDX is always the x2APIC ID, even for undefined
1547 		 * subleaves. Index 1 will exist iff the leaf is
1548 		 * implemented, so we pass through CL iff leaf 1
1549 		 * exists. EDX can be copied from any existing index.
1550 		 */
1551 		if (function == 0xb || function == 0x1f) {
1552 			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1553 			if (entry) {
1554 				*ecx = index & 0xff;
1555 				*edx = entry->edx;
1556 			}
1557 		}
1558 	}
1559 	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1560 			used_max_basic);
1561 	return exact;
1562 }
1563 EXPORT_SYMBOL_GPL(kvm_cpuid);
1564 
1565 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1566 {
1567 	u32 eax, ebx, ecx, edx;
1568 
1569 	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1570 		return 1;
1571 
1572 	eax = kvm_rax_read(vcpu);
1573 	ecx = kvm_rcx_read(vcpu);
1574 	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1575 	kvm_rax_write(vcpu, eax);
1576 	kvm_rbx_write(vcpu, ebx);
1577 	kvm_rcx_write(vcpu, ecx);
1578 	kvm_rdx_write(vcpu, edx);
1579 	return kvm_skip_emulated_instruction(vcpu);
1580 }
1581 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
1582