Linux Audio

Check our new training course

Buildroot integration, development and maintenance

Need a Buildroot system for your embedded project?
Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Kernel-based Virtual Machine driver for Linux
   4 * cpuid support routines
   5 *
   6 * derived from arch/x86/kvm/x86.c
   7 *
   8 * Copyright 2011 Red Hat, Inc. and/or its affiliates.
   9 * Copyright IBM Corporation, 2008
  10 */
 
  11
  12#include <linux/kvm_host.h>
 
  13#include <linux/export.h>
  14#include <linux/vmalloc.h>
  15#include <linux/uaccess.h>
  16#include <linux/sched/stat.h>
  17
  18#include <asm/processor.h>
  19#include <asm/user.h>
  20#include <asm/fpu/xstate.h>
  21#include <asm/sgx.h>
  22#include <asm/cpuid.h>
  23#include "cpuid.h"
  24#include "lapic.h"
  25#include "mmu.h"
  26#include "trace.h"
  27#include "pmu.h"
 
  28
  29/*
  30 * Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be
  31 * aligned to sizeof(unsigned long) because it's not accessed via bitops.
  32 */
  33u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
  34EXPORT_SYMBOL_GPL(kvm_cpu_caps);
  35
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  36u32 xstate_required_size(u64 xstate_bv, bool compacted)
  37{
  38	int feature_bit = 0;
  39	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
  40
  41	xstate_bv &= XFEATURE_MASK_EXTEND;
  42	while (xstate_bv) {
  43		if (xstate_bv & 0x1) {
  44		        u32 eax, ebx, ecx, edx, offset;
  45		        cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
 
  46			/* ECX[1]: 64B alignment in compacted form */
  47			if (compacted)
  48				offset = (ecx & 0x2) ? ALIGN(ret, 64) : ret;
  49			else
  50				offset = ebx;
  51			ret = max(ret, offset + eax);
  52		}
  53
  54		xstate_bv >>= 1;
  55		feature_bit++;
  56	}
  57
  58	return ret;
  59}
  60
  61/*
  62 * This one is tied to SSB in the user API, and not
  63 * visible in /proc/cpuinfo.
  64 */
  65#define KVM_X86_FEATURE_AMD_PSFD	(13*32+28) /* Predictive Store Forwarding Disable */
  66
  67#define F feature_bit
  68
  69/* Scattered Flag - For features that are scattered by cpufeatures.h. */
  70#define SF(name)						\
  71({								\
  72	BUILD_BUG_ON(X86_FEATURE_##name >= MAX_CPU_FEATURES);	\
  73	(boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0);	\
  74})
  75
  76/*
  77 * Magic value used by KVM when querying userspace-provided CPUID entries and
  78 * doesn't care about the CPIUD index because the index of the function in
  79 * question is not significant.  Note, this magic value must have at least one
  80 * bit set in bits[63:32] and must be consumed as a u64 by cpuid_entry2_find()
  81 * to avoid false positives when processing guest CPUID input.
  82 */
  83#define KVM_CPUID_INDEX_NOT_SIGNIFICANT -1ull
  84
  85static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
  86	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u64 index)
  87{
  88	struct kvm_cpuid_entry2 *e;
  89	int i;
  90
 
 
 
 
 
 
 
 
 
 
 
 
  91	for (i = 0; i < nent; i++) {
  92		e = &entries[i];
  93
  94		if (e->function != function)
  95			continue;
  96
  97		/*
  98		 * If the index isn't significant, use the first entry with a
  99		 * matching function.  It's userspace's responsibilty to not
 100		 * provide "duplicate" entries in all cases.
 101		 */
 102		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
 103			return e;
 104
 105
 106		/*
 107		 * Similarly, use the first matching entry if KVM is doing a
 108		 * lookup (as opposed to emulating CPUID) for a function that's
 109		 * architecturally defined as not having a significant index.
 110		 */
 111		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
 112			/*
 113			 * Direct lookups from KVM should not diverge from what
 114			 * KVM defines internally (the architectural behavior).
 115			 */
 116			WARN_ON_ONCE(cpuid_function_is_indexed(function));
 117			return e;
 118		}
 119	}
 120
 121	return NULL;
 122}
 123
 124static int kvm_check_cpuid(struct kvm_vcpu *vcpu,
 125			   struct kvm_cpuid_entry2 *entries,
 126			   int nent)
 127{
 128	struct kvm_cpuid_entry2 *best;
 129	u64 xfeatures;
 130
 131	/*
 132	 * The existing code assumes virtual address is 48-bit or 57-bit in the
 133	 * canonical address checks; exit if it is ever changed.
 134	 */
 135	best = cpuid_entry2_find(entries, nent, 0x80000008,
 136				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 137	if (best) {
 138		int vaddr_bits = (best->eax & 0xff00) >> 8;
 139
 140		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
 141			return -EINVAL;
 142	}
 143
 144	/*
 145	 * Exposing dynamic xfeatures to the guest requires additional
 146	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
 147	 */
 148	best = cpuid_entry2_find(entries, nent, 0xd, 0);
 149	if (!best)
 150		return 0;
 151
 152	xfeatures = best->eax | ((u64)best->edx << 32);
 153	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
 154	if (!xfeatures)
 155		return 0;
 156
 157	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
 158}
 159
 160/* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
 161static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
 162				 int nent)
 163{
 164	struct kvm_cpuid_entry2 *orig;
 165	int i;
 166
 167	if (nent != vcpu->arch.cpuid_nent)
 168		return -EINVAL;
 169
 170	for (i = 0; i < nent; i++) {
 171		orig = &vcpu->arch.cpuid_entries[i];
 172		if (e2[i].function != orig->function ||
 173		    e2[i].index != orig->index ||
 174		    e2[i].flags != orig->flags ||
 175		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
 176		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
 177			return -EINVAL;
 178	}
 179
 180	return 0;
 181}
 182
 183static void kvm_update_kvm_cpuid_base(struct kvm_vcpu *vcpu)
 
 184{
 185	u32 function;
 186	struct kvm_cpuid_entry2 *entry;
 
 187
 188	vcpu->arch.kvm_cpuid_base = 0;
 189
 190	for_each_possible_hypervisor_cpuid_base(function) {
 191		entry = kvm_find_cpuid_entry(vcpu, function);
 192
 193		if (entry) {
 194			u32 signature[3];
 195
 196			signature[0] = entry->ebx;
 197			signature[1] = entry->ecx;
 198			signature[2] = entry->edx;
 199
 200			BUILD_BUG_ON(sizeof(signature) > sizeof(KVM_SIGNATURE));
 201			if (!memcmp(signature, KVM_SIGNATURE, sizeof(signature))) {
 202				vcpu->arch.kvm_cpuid_base = function;
 203				break;
 204			}
 205		}
 206	}
 
 
 207}
 208
 209static struct kvm_cpuid_entry2 *__kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu,
 210					      struct kvm_cpuid_entry2 *entries, int nent)
 211{
 212	u32 base = vcpu->arch.kvm_cpuid_base;
 213
 214	if (!base)
 215		return NULL;
 216
 217	return cpuid_entry2_find(entries, nent, base | KVM_CPUID_FEATURES,
 
 
 
 218				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 219}
 220
 221static struct kvm_cpuid_entry2 *kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu)
 222{
 223	return __kvm_find_kvm_cpuid_features(vcpu, vcpu->arch.cpuid_entries,
 224					     vcpu->arch.cpuid_nent);
 
 
 
 
 
 225}
 226
 227void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
 228{
 229	struct kvm_cpuid_entry2 *best = kvm_find_kvm_cpuid_features(vcpu);
 230
 231	/*
 232	 * save the feature bitmap to avoid cpuid lookup for every PV
 233	 * operation
 234	 */
 235	if (best)
 236		vcpu->arch.pv_cpuid.features = best->eax;
 237}
 238
 239/*
 240 * Calculate guest's supported XCR0 taking into account guest CPUID data and
 241 * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
 242 */
 243static u64 cpuid_get_supported_xcr0(struct kvm_cpuid_entry2 *entries, int nent)
 244{
 245	struct kvm_cpuid_entry2 *best;
 246
 247	best = cpuid_entry2_find(entries, nent, 0xd, 0);
 248	if (!best)
 249		return 0;
 250
 251	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
 252}
 253
 254static void __kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *entries,
 255				       int nent)
 256{
 257	struct kvm_cpuid_entry2 *best;
 258	u64 guest_supported_xcr0 = cpuid_get_supported_xcr0(entries, nent);
 259
 260	best = cpuid_entry2_find(entries, nent, 1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 261	if (best) {
 262		/* Update OSXSAVE bit */
 263		if (boot_cpu_has(X86_FEATURE_XSAVE))
 264			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
 265				   kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE));
 266
 267		cpuid_entry_change(best, X86_FEATURE_APIC,
 268			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
 269	}
 270
 271	best = cpuid_entry2_find(entries, nent, 7, 0);
 272	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
 273		cpuid_entry_change(best, X86_FEATURE_OSPKE,
 274				   kvm_read_cr4_bits(vcpu, X86_CR4_PKE));
 275
 276	best = cpuid_entry2_find(entries, nent, 0xD, 0);
 277	if (best)
 278		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
 279
 280	best = cpuid_entry2_find(entries, nent, 0xD, 1);
 281	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
 282		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
 283		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
 284
 285	best = __kvm_find_kvm_cpuid_features(vcpu, entries, nent);
 286	if (kvm_hlt_in_guest(vcpu->kvm) && best &&
 287		(best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
 288		best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
 
 
 289
 290	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
 291		best = cpuid_entry2_find(entries, nent, 0x1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 292		if (best)
 293			cpuid_entry_change(best, X86_FEATURE_MWAIT,
 294					   vcpu->arch.ia32_misc_enable_msr &
 295					   MSR_IA32_MISC_ENABLE_MWAIT);
 296	}
 297
 298	/*
 299	 * Bits 127:0 of the allowed SECS.ATTRIBUTES (CPUID.0x12.0x1) enumerate
 300	 * the supported XSAVE Feature Request Mask (XFRM), i.e. the enclave's
 301	 * requested XCR0 value.  The enclave's XFRM must be a subset of XCRO
 302	 * at the time of EENTER, thus adjust the allowed XFRM by the guest's
 303	 * supported XCR0.  Similar to XCR0 handling, FP and SSE are forced to
 304	 * '1' even on CPUs that don't support XSAVE.
 305	 */
 306	best = cpuid_entry2_find(entries, nent, 0x12, 0x1);
 307	if (best) {
 308		best->ecx &= guest_supported_xcr0 & 0xffffffff;
 309		best->edx &= guest_supported_xcr0 >> 32;
 310		best->ecx |= XFEATURE_MASK_FPSSE;
 311	}
 312}
 313
 314void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
 315{
 316	__kvm_update_cpuid_runtime(vcpu, vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
 317}
 318EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
 319
 320static bool kvm_cpuid_has_hyperv(struct kvm_cpuid_entry2 *entries, int nent)
 321{
 
 322	struct kvm_cpuid_entry2 *entry;
 323
 324	entry = cpuid_entry2_find(entries, nent, HYPERV_CPUID_INTERFACE,
 325				  KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 326	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 327}
 328
 329static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
 330{
 331	struct kvm_lapic *apic = vcpu->arch.apic;
 332	struct kvm_cpuid_entry2 *best;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 333
 334	best = kvm_find_cpuid_entry(vcpu, 1);
 335	if (best && apic) {
 336		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
 337			apic->lapic_timer.timer_mode_mask = 3 << 17;
 338		else
 339			apic->lapic_timer.timer_mode_mask = 1 << 17;
 340
 341		kvm_apic_set_version(vcpu);
 342	}
 343
 344	vcpu->arch.guest_supported_xcr0 =
 345		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
 346
 347	/*
 348	 * FP+SSE can always be saved/restored via KVM_{G,S}ET_XSAVE, even if
 349	 * XSAVE/XCRO are not exposed to the guest, and even if XSAVE isn't
 350	 * supported by the host.
 351	 */
 352	vcpu->arch.guest_fpu.fpstate->user_xfeatures = vcpu->arch.guest_supported_xcr0 |
 353						       XFEATURE_MASK_FPSSE;
 354
 355	kvm_update_pv_runtime(vcpu);
 356
 
 357	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
 358	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
 359
 360	kvm_pmu_refresh(vcpu);
 361	vcpu->arch.cr4_guest_rsvd_bits =
 362	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
 363
 364	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu->arch.cpuid_entries,
 365						    vcpu->arch.cpuid_nent));
 366
 367	/* Invoke the vendor callback only after the above state is updated. */
 368	static_call(kvm_x86_vcpu_after_set_cpuid)(vcpu);
 369
 370	/*
 371	 * Except for the MMU, which needs to do its thing any vendor specific
 372	 * adjustments to the reserved GPA bits.
 373	 */
 374	kvm_mmu_after_set_cpuid(vcpu);
 375}
 376
 377int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
 378{
 379	struct kvm_cpuid_entry2 *best;
 380
 381	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
 382	if (!best || best->eax < 0x80000008)
 383		goto not_found;
 384	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
 385	if (best)
 386		return best->eax & 0xff;
 387not_found:
 388	return 36;
 389}
 390
 391/*
 392 * This "raw" version returns the reserved GPA bits without any adjustments for
 393 * encryption technologies that usurp bits.  The raw mask should be used if and
 394 * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
 395 */
 396u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
 397{
 398	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
 399}
 400
 401static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
 402                        int nent)
 403{
 404	int r;
 405
 406	__kvm_update_cpuid_runtime(vcpu, e2, nent);
 407
 408	/*
 409	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
 410	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
 411	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
 412	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
 413	 * the core vCPU model on the fly. It would've been better to forbid any
 414	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
 415	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
 416	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
 417	 * whether the supplied CPUID data is equal to what's already set.
 418	 */
 419	if (vcpu->arch.last_vmentry_cpu != -1) {
 420		r = kvm_cpuid_check_equal(vcpu, e2, nent);
 421		if (r)
 422			return r;
 423
 424		kvfree(e2);
 425		return 0;
 426	}
 427
 
 428	if (kvm_cpuid_has_hyperv(e2, nent)) {
 429		r = kvm_hv_vcpu_init(vcpu);
 430		if (r)
 431			return r;
 432	}
 
 433
 434	r = kvm_check_cpuid(vcpu, e2, nent);
 435	if (r)
 436		return r;
 437
 438	kvfree(vcpu->arch.cpuid_entries);
 439	vcpu->arch.cpuid_entries = e2;
 440	vcpu->arch.cpuid_nent = nent;
 441
 442	kvm_update_kvm_cpuid_base(vcpu);
 
 
 
 443	kvm_vcpu_after_set_cpuid(vcpu);
 444
 445	return 0;
 446}
 447
 448/* when an old userspace process fills a new kernel module */
 449int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
 450			     struct kvm_cpuid *cpuid,
 451			     struct kvm_cpuid_entry __user *entries)
 452{
 453	int r, i;
 454	struct kvm_cpuid_entry *e = NULL;
 455	struct kvm_cpuid_entry2 *e2 = NULL;
 456
 457	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 458		return -E2BIG;
 459
 460	if (cpuid->nent) {
 461		e = vmemdup_user(entries, array_size(sizeof(*e), cpuid->nent));
 462		if (IS_ERR(e))
 463			return PTR_ERR(e);
 464
 465		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
 466		if (!e2) {
 467			r = -ENOMEM;
 468			goto out_free_cpuid;
 469		}
 470	}
 471	for (i = 0; i < cpuid->nent; i++) {
 472		e2[i].function = e[i].function;
 473		e2[i].eax = e[i].eax;
 474		e2[i].ebx = e[i].ebx;
 475		e2[i].ecx = e[i].ecx;
 476		e2[i].edx = e[i].edx;
 477		e2[i].index = 0;
 478		e2[i].flags = 0;
 479		e2[i].padding[0] = 0;
 480		e2[i].padding[1] = 0;
 481		e2[i].padding[2] = 0;
 482	}
 483
 484	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
 485	if (r)
 486		kvfree(e2);
 487
 488out_free_cpuid:
 489	kvfree(e);
 490
 491	return r;
 492}
 493
 494int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
 495			      struct kvm_cpuid2 *cpuid,
 496			      struct kvm_cpuid_entry2 __user *entries)
 497{
 498	struct kvm_cpuid_entry2 *e2 = NULL;
 499	int r;
 500
 501	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 502		return -E2BIG;
 503
 504	if (cpuid->nent) {
 505		e2 = vmemdup_user(entries, array_size(sizeof(*e2), cpuid->nent));
 506		if (IS_ERR(e2))
 507			return PTR_ERR(e2);
 508	}
 509
 510	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
 511	if (r)
 512		kvfree(e2);
 513
 514	return r;
 515}
 516
 517int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
 518			      struct kvm_cpuid2 *cpuid,
 519			      struct kvm_cpuid_entry2 __user *entries)
 520{
 521	int r;
 522
 523	r = -E2BIG;
 524	if (cpuid->nent < vcpu->arch.cpuid_nent)
 525		goto out;
 526	r = -EFAULT;
 527	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
 528			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
 529		goto out;
 530	return 0;
 531
 532out:
 533	cpuid->nent = vcpu->arch.cpuid_nent;
 534	return r;
 535}
 536
 537/* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
 538static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
 539{
 540	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
 541	struct kvm_cpuid_entry2 entry;
 542
 543	reverse_cpuid_check(leaf);
 544
 545	cpuid_count(cpuid.function, cpuid.index,
 546		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
 547
 548	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
 549}
 550
 551static __always_inline
 552void kvm_cpu_cap_init_kvm_defined(enum kvm_only_cpuid_leafs leaf, u32 mask)
 553{
 554	/* Use kvm_cpu_cap_mask for leafs that aren't KVM-only. */
 555	BUILD_BUG_ON(leaf < NCAPINTS);
 556
 557	kvm_cpu_caps[leaf] = mask;
 558
 559	__kvm_cpu_cap_mask(leaf);
 560}
 561
 562static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
 563{
 564	/* Use kvm_cpu_cap_init_kvm_defined for KVM-only leafs. */
 565	BUILD_BUG_ON(leaf >= NCAPINTS);
 566
 567	kvm_cpu_caps[leaf] &= mask;
 568
 569	__kvm_cpu_cap_mask(leaf);
 570}
 571
 572void kvm_set_cpu_caps(void)
 573{
 574#ifdef CONFIG_X86_64
 575	unsigned int f_gbpages = F(GBPAGES);
 576	unsigned int f_lm = F(LM);
 577	unsigned int f_xfd = F(XFD);
 578#else
 579	unsigned int f_gbpages = 0;
 580	unsigned int f_lm = 0;
 581	unsigned int f_xfd = 0;
 582#endif
 583	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
 584
 585	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
 586		     sizeof(boot_cpu_data.x86_capability));
 587
 588	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
 589	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
 590
 591	kvm_cpu_cap_mask(CPUID_1_ECX,
 592		/*
 593		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
 594		 * advertised to guests via CPUID!
 595		 */
 596		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
 597		0 /* DS-CPL, VMX, SMX, EST */ |
 598		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
 599		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
 600		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
 601		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
 602		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
 603		F(F16C) | F(RDRAND)
 604	);
 605	/* KVM emulates x2apic in software irrespective of host support. */
 606	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
 607
 608	kvm_cpu_cap_mask(CPUID_1_EDX,
 609		F(FPU) | F(VME) | F(DE) | F(PSE) |
 610		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 611		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
 612		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 613		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
 614		0 /* Reserved, DS, ACPI */ | F(MMX) |
 615		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
 616		0 /* HTT, TM, Reserved, PBE */
 617	);
 618
 619	kvm_cpu_cap_mask(CPUID_7_0_EBX,
 620		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
 621		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
 622		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
 623		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
 624		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
 625		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
 626		F(AVX512VL));
 627
 628	kvm_cpu_cap_mask(CPUID_7_ECX,
 629		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
 630		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
 631		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
 632		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
 633		F(SGX_LC) | F(BUS_LOCK_DETECT)
 634	);
 635	/* Set LA57 based on hardware capability. */
 636	if (cpuid_ecx(7) & F(LA57))
 637		kvm_cpu_cap_set(X86_FEATURE_LA57);
 638
 639	/*
 640	 * PKU not yet implemented for shadow paging and requires OSPKE
 641	 * to be set on the host. Clear it if that is not the case
 642	 */
 643	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
 644		kvm_cpu_cap_clear(X86_FEATURE_PKU);
 645
 646	kvm_cpu_cap_mask(CPUID_7_EDX,
 647		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
 648		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
 649		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
 650		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
 651		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16)
 652	);
 653
 654	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
 655	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
 656	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
 657
 658	if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
 
 
 659		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
 660	if (boot_cpu_has(X86_FEATURE_STIBP))
 661		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
 662	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
 663		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
 664
 665	kvm_cpu_cap_mask(CPUID_7_1_EAX,
 666		F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) | F(AMX_FP16) |
 667		F(AVX_IFMA)
 
 668	);
 669
 670	kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
 671		F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(PREFETCHITI)
 
 
 
 
 
 
 672	);
 673
 674	kvm_cpu_cap_mask(CPUID_D_1_EAX,
 675		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
 676	);
 677
 678	kvm_cpu_cap_init_kvm_defined(CPUID_12_EAX,
 679		SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
 680	);
 681
 
 
 
 
 682	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
 683		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
 684		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
 685		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
 686		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
 687		F(TOPOEXT) | 0 /* PERFCTR_CORE */
 688	);
 689
 690	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
 691		F(FPU) | F(VME) | F(DE) | F(PSE) |
 692		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 693		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
 694		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 695		F(PAT) | F(PSE36) | 0 /* Reserved */ |
 696		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
 697		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
 698		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
 699	);
 700
 701	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
 702		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
 703
 
 
 
 
 704	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
 705		F(CLZERO) | F(XSAVEERPTR) |
 706		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
 707		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
 708		__feature_bit(KVM_X86_FEATURE_AMD_PSFD)
 709	);
 710
 711	/*
 712	 * AMD has separate bits for each SPEC_CTRL bit.
 713	 * arch/x86/kernel/cpu/bugs.c is kind enough to
 714	 * record that in cpufeatures so use them.
 715	 */
 716	if (boot_cpu_has(X86_FEATURE_IBPB))
 717		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
 
 
 
 
 718	if (boot_cpu_has(X86_FEATURE_IBRS))
 719		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
 720	if (boot_cpu_has(X86_FEATURE_STIBP))
 721		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
 722	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
 723		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
 724	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
 725		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
 726	/*
 727	 * The preference is to use SPEC CTRL MSR instead of the
 728	 * VIRT_SPEC MSR.
 729	 */
 730	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
 731	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
 732		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
 733
 734	/*
 735	 * Hide all SVM features by default, SVM will set the cap bits for
 736	 * features it emulates and/or exposes for L1.
 737	 */
 738	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
 739
 740	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
 741		0 /* SME */ | F(SEV) | 0 /* VM_PAGE_FLUSH */ | F(SEV_ES) |
 742		F(SME_COHERENT));
 743
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 744	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
 745		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
 746		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
 747		F(PMM) | F(PMM_EN)
 748	);
 749
 750	/*
 751	 * Hide RDTSCP and RDPID if either feature is reported as supported but
 752	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
 753	 * should never happen, but the guest will likely crash if RDTSCP or
 754	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
 755	 * the past.  For example, the sanity check may fire if this instance of
 756	 * KVM is running as L1 on top of an older, broken KVM.
 757	 */
 758	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
 759		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
 760		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
 761		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
 762		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
 763	}
 764}
 765EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
 766
 767struct kvm_cpuid_array {
 768	struct kvm_cpuid_entry2 *entries;
 769	int maxnent;
 770	int nent;
 771};
 772
 773static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
 774{
 775	if (array->nent >= array->maxnent)
 776		return NULL;
 777
 778	return &array->entries[array->nent++];
 779}
 780
 781static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
 782					      u32 function, u32 index)
 783{
 784	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
 785
 786	if (!entry)
 787		return NULL;
 788
 789	memset(entry, 0, sizeof(*entry));
 790	entry->function = function;
 791	entry->index = index;
 792	switch (function & 0xC0000000) {
 793	case 0x40000000:
 794		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
 795		return entry;
 796
 797	case 0x80000000:
 798		/*
 799		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
 800		 * would result in out-of-bounds calls to do_host_cpuid.
 801		 */
 802		{
 803			static int max_cpuid_80000000;
 804			if (!READ_ONCE(max_cpuid_80000000))
 805				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
 806			if (function > READ_ONCE(max_cpuid_80000000))
 807				return entry;
 808		}
 809		break;
 810
 811	default:
 812		break;
 813	}
 814
 815	cpuid_count(entry->function, entry->index,
 816		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
 817
 818	if (cpuid_function_is_indexed(function))
 819		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 820
 821	return entry;
 822}
 823
 824static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
 825{
 826	struct kvm_cpuid_entry2 *entry;
 827
 828	if (array->nent >= array->maxnent)
 829		return -E2BIG;
 830
 831	entry = &array->entries[array->nent];
 832	entry->function = func;
 833	entry->index = 0;
 834	entry->flags = 0;
 835
 836	switch (func) {
 837	case 0:
 838		entry->eax = 7;
 839		++array->nent;
 840		break;
 841	case 1:
 842		entry->ecx = F(MOVBE);
 843		++array->nent;
 844		break;
 845	case 7:
 846		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 847		entry->eax = 0;
 848		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
 849			entry->ecx = F(RDPID);
 850		++array->nent;
 851		break;
 852	default:
 853		break;
 854	}
 855
 856	return 0;
 857}
 858
 859static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
 860{
 861	struct kvm_cpuid_entry2 *entry;
 862	int r, i, max_idx;
 863
 864	/* all calls to cpuid_count() should be made on the same cpu */
 865	get_cpu();
 866
 867	r = -E2BIG;
 868
 869	entry = do_host_cpuid(array, function, 0);
 870	if (!entry)
 871		goto out;
 872
 873	switch (function) {
 874	case 0:
 875		/* Limited to the highest leaf implemented in KVM. */
 876		entry->eax = min(entry->eax, 0x1fU);
 877		break;
 878	case 1:
 879		cpuid_entry_override(entry, CPUID_1_EDX);
 880		cpuid_entry_override(entry, CPUID_1_ECX);
 881		break;
 882	case 2:
 883		/*
 884		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
 885		 * CPUID(function=2, index=0) may return different results each
 886		 * time, with the least-significant byte in EAX enumerating the
 887		 * number of times software should do CPUID(2, 0).
 888		 *
 889		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
 890		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
 891		 * return 01H. Software should ignore this value and not
 892		 * interpret it as an informational descriptor", while AMD's
 893		 * APM states that CPUID(2) is reserved.
 894		 *
 895		 * WARN if a frankenstein CPU that supports virtualization and
 896		 * a stateful CPUID.0x2 is encountered.
 897		 */
 898		WARN_ON_ONCE((entry->eax & 0xff) > 1);
 899		break;
 900	/* functions 4 and 0x8000001d have additional index. */
 901	case 4:
 902	case 0x8000001d:
 903		/*
 904		 * Read entries until the cache type in the previous entry is
 905		 * zero, i.e. indicates an invalid entry.
 906		 */
 907		for (i = 1; entry->eax & 0x1f; ++i) {
 908			entry = do_host_cpuid(array, function, i);
 909			if (!entry)
 910				goto out;
 911		}
 912		break;
 913	case 6: /* Thermal management */
 914		entry->eax = 0x4; /* allow ARAT */
 915		entry->ebx = 0;
 916		entry->ecx = 0;
 917		entry->edx = 0;
 918		break;
 919	/* function 7 has additional index. */
 920	case 7:
 921		entry->eax = min(entry->eax, 1u);
 922		cpuid_entry_override(entry, CPUID_7_0_EBX);
 923		cpuid_entry_override(entry, CPUID_7_ECX);
 924		cpuid_entry_override(entry, CPUID_7_EDX);
 925
 926		/* KVM only supports 0x7.0 and 0x7.1, capped above via min(). */
 927		if (entry->eax == 1) {
 928			entry = do_host_cpuid(array, function, 1);
 929			if (!entry)
 930				goto out;
 931
 932			cpuid_entry_override(entry, CPUID_7_1_EAX);
 933			cpuid_entry_override(entry, CPUID_7_1_EDX);
 934			entry->ebx = 0;
 935			entry->ecx = 0;
 936		}
 
 
 
 
 
 
 
 
 
 
 937		break;
 938	case 0xa: { /* Architectural Performance Monitoring */
 939		union cpuid10_eax eax;
 940		union cpuid10_edx edx;
 941
 942		if (!static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
 943			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
 944			break;
 945		}
 946
 947		eax.split.version_id = kvm_pmu_cap.version;
 948		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
 949		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
 950		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
 951		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
 952		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
 953
 954		if (kvm_pmu_cap.version)
 955			edx.split.anythread_deprecated = 1;
 956		edx.split.reserved1 = 0;
 957		edx.split.reserved2 = 0;
 958
 959		entry->eax = eax.full;
 960		entry->ebx = kvm_pmu_cap.events_mask;
 961		entry->ecx = 0;
 962		entry->edx = edx.full;
 963		break;
 964	}
 965	case 0x1f:
 966	case 0xb:
 967		/*
 968		 * No topology; a valid topology is indicated by the presence
 969		 * of subleaf 1.
 970		 */
 971		entry->eax = entry->ebx = entry->ecx = 0;
 972		break;
 973	case 0xd: {
 974		u64 permitted_xcr0 = kvm_caps.supported_xcr0 & xstate_get_guest_group_perm();
 975		u64 permitted_xss = kvm_caps.supported_xss;
 976
 977		entry->eax &= permitted_xcr0;
 978		entry->ebx = xstate_required_size(permitted_xcr0, false);
 979		entry->ecx = entry->ebx;
 980		entry->edx &= permitted_xcr0 >> 32;
 981		if (!permitted_xcr0)
 982			break;
 983
 984		entry = do_host_cpuid(array, function, 1);
 985		if (!entry)
 986			goto out;
 987
 988		cpuid_entry_override(entry, CPUID_D_1_EAX);
 989		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
 990			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
 991							  true);
 992		else {
 993			WARN_ON_ONCE(permitted_xss != 0);
 994			entry->ebx = 0;
 995		}
 996		entry->ecx &= permitted_xss;
 997		entry->edx &= permitted_xss >> 32;
 998
 999		for (i = 2; i < 64; ++i) {
1000			bool s_state;
1001			if (permitted_xcr0 & BIT_ULL(i))
1002				s_state = false;
1003			else if (permitted_xss & BIT_ULL(i))
1004				s_state = true;
1005			else
1006				continue;
1007
1008			entry = do_host_cpuid(array, function, i);
1009			if (!entry)
1010				goto out;
1011
1012			/*
1013			 * The supported check above should have filtered out
1014			 * invalid sub-leafs.  Only valid sub-leafs should
1015			 * reach this point, and they should have a non-zero
1016			 * save state size.  Furthermore, check whether the
1017			 * processor agrees with permitted_xcr0/permitted_xss
1018			 * on whether this is an XCR0- or IA32_XSS-managed area.
1019			 */
1020			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1021				--array->nent;
1022				continue;
1023			}
1024
1025			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1026				entry->ecx &= ~BIT_ULL(2);
1027			entry->edx = 0;
1028		}
1029		break;
1030	}
1031	case 0x12:
1032		/* Intel SGX */
1033		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1034			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1035			break;
1036		}
1037
1038		/*
1039		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1040		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1041		 * are restricted by kernel and KVM capabilities (like most
1042		 * feature flags), while enclave size is unrestricted.
1043		 */
1044		cpuid_entry_override(entry, CPUID_12_EAX);
1045		entry->ebx &= SGX_MISC_EXINFO;
1046
1047		entry = do_host_cpuid(array, function, 1);
1048		if (!entry)
1049			goto out;
1050
1051		/*
1052		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1053		 * feature flags.  Advertise all supported flags, including
1054		 * privileged attributes that require explicit opt-in from
1055		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1056		 * expected to derive it from supported XCR0.
1057		 */
1058		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1059		entry->ebx &= 0;
1060		break;
1061	/* Intel PT */
1062	case 0x14:
1063		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1064			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1065			break;
1066		}
1067
1068		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1069			if (!do_host_cpuid(array, function, i))
1070				goto out;
1071		}
1072		break;
1073	/* Intel AMX TILE */
1074	case 0x1d:
1075		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1076			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1077			break;
1078		}
1079
1080		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1081			if (!do_host_cpuid(array, function, i))
1082				goto out;
1083		}
1084		break;
1085	case 0x1e: /* TMUL information */
1086		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1087			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1088			break;
1089		}
1090		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1091	case KVM_CPUID_SIGNATURE: {
1092		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1093		entry->eax = KVM_CPUID_FEATURES;
1094		entry->ebx = sigptr[0];
1095		entry->ecx = sigptr[1];
1096		entry->edx = sigptr[2];
1097		break;
1098	}
1099	case KVM_CPUID_FEATURES:
1100		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1101			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1102			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1103			     (1 << KVM_FEATURE_ASYNC_PF) |
1104			     (1 << KVM_FEATURE_PV_EOI) |
1105			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1106			     (1 << KVM_FEATURE_PV_UNHALT) |
1107			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1108			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1109			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1110			     (1 << KVM_FEATURE_POLL_CONTROL) |
1111			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1112			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1113
1114		if (sched_info_on())
1115			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1116
1117		entry->ebx = 0;
1118		entry->ecx = 0;
1119		entry->edx = 0;
1120		break;
1121	case 0x80000000:
1122		entry->eax = min(entry->eax, 0x80000021);
1123		/*
1124		 * Serializing LFENCE is reported in a multitude of ways, and
1125		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1126		 * userspace by providing the CPUID leaf ourselves.
1127		 *
1128		 * However, only do it if the host has CPUID leaf 0x8000001d.
1129		 * QEMU thinks that it can query the host blindly for that
1130		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1131		 * above.  The processor merrily returns values from the
1132		 * highest Intel leaf which QEMU tries to use as the guest's
1133		 * 0x8000001d.  Even worse, this can result in an infinite
1134		 * loop if said highest leaf has no subleaves indexed by ECX.
1135		 */
1136		if (entry->eax >= 0x8000001d &&
1137		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1138		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1139			entry->eax = max(entry->eax, 0x80000021);
1140		break;
1141	case 0x80000001:
1142		entry->ebx &= ~GENMASK(27, 16);
1143		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1144		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1145		break;
 
 
 
1146	case 0x80000006:
1147		/* Drop reserved bits, pass host L2 cache and TLB info. */
1148		entry->edx &= ~GENMASK(17, 16);
1149		break;
1150	case 0x80000007: /* Advanced power management */
1151		/* invariant TSC is CPUID.80000007H:EDX[8] */
1152		entry->edx &= (1 << 8);
1153		/* mask against host */
1154		entry->edx &= boot_cpu_data.x86_power;
1155		entry->eax = entry->ebx = entry->ecx = 0;
1156		break;
1157	case 0x80000008: {
1158		unsigned g_phys_as = (entry->eax >> 16) & 0xff;
1159		unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
1160		unsigned phys_as = entry->eax & 0xff;
 
 
 
 
 
 
 
 
 
 
 
 
 
1161
1162		/*
1163		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1164		 * the guest operates in the same PA space as the host, i.e.
1165		 * reductions in MAXPHYADDR for memory encryption affect shadow
1166		 * paging, too.
1167		 *
1168		 * If TDP is enabled but an explicit guest MAXPHYADDR is not
1169		 * provided, use the raw bare metal MAXPHYADDR as reductions to
1170		 * the HPAs do not affect GPAs.
 
 
 
1171		 */
1172		if (!tdp_enabled)
1173			g_phys_as = boot_cpu_data.x86_phys_bits;
1174		else if (!g_phys_as)
 
 
1175			g_phys_as = phys_as;
 
 
 
1176
1177		entry->eax = g_phys_as | (virt_as << 8);
1178		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1179		entry->edx = 0;
1180		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1181		break;
1182	}
1183	case 0x8000000A:
1184		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1185			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1186			break;
1187		}
1188		entry->eax = 1; /* SVM revision 1 */
1189		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1190				   ASID emulation to nested SVM */
1191		entry->ecx = 0; /* Reserved */
1192		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1193		break;
1194	case 0x80000019:
1195		entry->ecx = entry->edx = 0;
1196		break;
1197	case 0x8000001a:
1198		entry->eax &= GENMASK(2, 0);
1199		entry->ebx = entry->ecx = entry->edx = 0;
1200		break;
1201	case 0x8000001e:
1202		/* Do not return host topology information.  */
1203		entry->eax = entry->ebx = entry->ecx = 0;
1204		entry->edx = 0; /* reserved */
1205		break;
1206	case 0x8000001F:
1207		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1208			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1209		} else {
1210			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1211			/* Clear NumVMPL since KVM does not support VMPL.  */
1212			entry->ebx &= ~GENMASK(31, 12);
1213			/*
1214			 * Enumerate '0' for "PA bits reduction", the adjusted
1215			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1216			 */
1217			entry->ebx &= ~GENMASK(11, 6);
1218		}
1219		break;
1220	case 0x80000020:
1221		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1222		break;
1223	case 0x80000021:
1224		entry->ebx = entry->ecx = entry->edx = 0;
1225		/*
1226		 * Pass down these bits:
1227		 *    EAX      0      NNDBP, Processor ignores nested data breakpoints
1228		 *    EAX      2      LAS, LFENCE always serializing
1229		 *    EAX      6      NSCB, Null selector clear base
1230		 *
1231		 * Other defined bits are for MSRs that KVM does not expose:
1232		 *   EAX      3      SPCL, SMM page configuration lock
1233		 *   EAX      13     PCMSR, Prefetch control MSR
1234		 *
1235		 * KVM doesn't support SMM_CTL.
1236		 *   EAX       9     SMM_CTL MSR is not supported
1237		 */
1238		entry->eax &= BIT(0) | BIT(2) | BIT(6);
1239		entry->eax |= BIT(9);
1240		if (static_cpu_has(X86_FEATURE_LFENCE_RDTSC))
1241			entry->eax |= BIT(2);
1242		if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
1243			entry->eax |= BIT(6);
1244		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1245	/*Add support for Centaur's CPUID instruction*/
1246	case 0xC0000000:
1247		/*Just support up to 0xC0000004 now*/
1248		entry->eax = min(entry->eax, 0xC0000004);
1249		break;
1250	case 0xC0000001:
1251		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1252		break;
1253	case 3: /* Processor serial number */
1254	case 5: /* MONITOR/MWAIT */
1255	case 0xC0000002:
1256	case 0xC0000003:
1257	case 0xC0000004:
1258	default:
1259		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1260		break;
1261	}
1262
1263	r = 0;
1264
1265out:
1266	put_cpu();
1267
1268	return r;
1269}
1270
1271static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1272			 unsigned int type)
1273{
1274	if (type == KVM_GET_EMULATED_CPUID)
1275		return __do_cpuid_func_emulated(array, func);
1276
1277	return __do_cpuid_func(array, func);
1278}
1279
1280#define CENTAUR_CPUID_SIGNATURE 0xC0000000
1281
1282static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1283			  unsigned int type)
1284{
1285	u32 limit;
1286	int r;
1287
1288	if (func == CENTAUR_CPUID_SIGNATURE &&
1289	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1290		return 0;
1291
1292	r = do_cpuid_func(array, func, type);
1293	if (r)
1294		return r;
1295
1296	limit = array->entries[array->nent - 1].eax;
1297	for (func = func + 1; func <= limit; ++func) {
1298		r = do_cpuid_func(array, func, type);
1299		if (r)
1300			break;
1301	}
1302
1303	return r;
1304}
1305
1306static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1307				 __u32 num_entries, unsigned int ioctl_type)
1308{
1309	int i;
1310	__u32 pad[3];
1311
1312	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1313		return false;
1314
1315	/*
1316	 * We want to make sure that ->padding is being passed clean from
1317	 * userspace in case we want to use it for something in the future.
1318	 *
1319	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1320	 * have to give ourselves satisfied only with the emulated side. /me
1321	 * sheds a tear.
1322	 */
1323	for (i = 0; i < num_entries; i++) {
1324		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1325			return true;
1326
1327		if (pad[0] || pad[1] || pad[2])
1328			return true;
1329	}
1330	return false;
1331}
1332
1333int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1334			    struct kvm_cpuid_entry2 __user *entries,
1335			    unsigned int type)
1336{
1337	static const u32 funcs[] = {
1338		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1339	};
1340
1341	struct kvm_cpuid_array array = {
1342		.nent = 0,
1343	};
1344	int r, i;
1345
1346	if (cpuid->nent < 1)
1347		return -E2BIG;
1348	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1349		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1350
1351	if (sanity_check_entries(entries, cpuid->nent, type))
1352		return -EINVAL;
1353
1354	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1355	if (!array.entries)
1356		return -ENOMEM;
1357
1358	array.maxnent = cpuid->nent;
1359
1360	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1361		r = get_cpuid_func(&array, funcs[i], type);
1362		if (r)
1363			goto out_free;
1364	}
1365	cpuid->nent = array.nent;
1366
1367	if (copy_to_user(entries, array.entries,
1368			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1369		r = -EFAULT;
1370
1371out_free:
1372	kvfree(array.entries);
1373	return r;
1374}
1375
1376struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1377						    u32 function, u32 index)
1378{
1379	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1380				 function, index);
1381}
1382EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1383
1384struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1385					      u32 function)
1386{
1387	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1388				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1389}
1390EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1391
1392/*
1393 * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1394 * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1395 * returns all zeroes for any undefined leaf, whether or not the leaf is in
1396 * range.  Centaur/VIA follows Intel semantics.
1397 *
1398 * A leaf is considered out-of-range if its function is higher than the maximum
1399 * supported leaf of its associated class or if its associated class does not
1400 * exist.
1401 *
1402 * There are three primary classes to be considered, with their respective
1403 * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1404 * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1405 * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1406 *
1407 *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1408 *  - Hypervisor: 0x40000000 - 0x4fffffff
1409 *  - Extended:   0x80000000 - 0xbfffffff
1410 *  - Centaur:    0xc0000000 - 0xcfffffff
1411 *
1412 * The Hypervisor class is further subdivided into sub-classes that each act as
1413 * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1414 * is advertising support for both HyperV and KVM, the resulting Hypervisor
1415 * CPUID sub-classes are:
1416 *
1417 *  - HyperV:     0x40000000 - 0x400000ff
1418 *  - KVM:        0x40000100 - 0x400001ff
1419 */
1420static struct kvm_cpuid_entry2 *
1421get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1422{
1423	struct kvm_cpuid_entry2 *basic, *class;
1424	u32 function = *fn_ptr;
1425
1426	basic = kvm_find_cpuid_entry(vcpu, 0);
1427	if (!basic)
1428		return NULL;
1429
1430	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1431	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1432		return NULL;
1433
1434	if (function >= 0x40000000 && function <= 0x4fffffff)
1435		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1436	else if (function >= 0xc0000000)
1437		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1438	else
1439		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1440
1441	if (class && function <= class->eax)
1442		return NULL;
1443
1444	/*
1445	 * Leaf specific adjustments are also applied when redirecting to the
1446	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1447	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1448	 * needs to be pulled from CPUID.0xb.1.
1449	 */
1450	*fn_ptr = basic->eax;
1451
1452	/*
1453	 * The class does not exist or the requested function is out of range;
1454	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1455	 * the original requested leaf is observed!
1456	 */
1457	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1458}
1459
1460bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1461	       u32 *ecx, u32 *edx, bool exact_only)
1462{
1463	u32 orig_function = *eax, function = *eax, index = *ecx;
1464	struct kvm_cpuid_entry2 *entry;
1465	bool exact, used_max_basic = false;
1466
1467	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1468	exact = !!entry;
1469
1470	if (!entry && !exact_only) {
1471		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1472		used_max_basic = !!entry;
1473	}
1474
1475	if (entry) {
1476		*eax = entry->eax;
1477		*ebx = entry->ebx;
1478		*ecx = entry->ecx;
1479		*edx = entry->edx;
1480		if (function == 7 && index == 0) {
1481			u64 data;
1482		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1483			    (data & TSX_CTRL_CPUID_CLEAR))
1484				*ebx &= ~(F(RTM) | F(HLE));
 
 
 
1485		}
1486	} else {
1487		*eax = *ebx = *ecx = *edx = 0;
1488		/*
1489		 * When leaf 0BH or 1FH is defined, CL is pass-through
1490		 * and EDX is always the x2APIC ID, even for undefined
1491		 * subleaves. Index 1 will exist iff the leaf is
1492		 * implemented, so we pass through CL iff leaf 1
1493		 * exists. EDX can be copied from any existing index.
1494		 */
1495		if (function == 0xb || function == 0x1f) {
1496			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1497			if (entry) {
1498				*ecx = index & 0xff;
1499				*edx = entry->edx;
1500			}
1501		}
1502	}
1503	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1504			used_max_basic);
1505	return exact;
1506}
1507EXPORT_SYMBOL_GPL(kvm_cpuid);
1508
1509int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1510{
1511	u32 eax, ebx, ecx, edx;
1512
1513	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1514		return 1;
1515
1516	eax = kvm_rax_read(vcpu);
1517	ecx = kvm_rcx_read(vcpu);
1518	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1519	kvm_rax_write(vcpu, eax);
1520	kvm_rbx_write(vcpu, ebx);
1521	kvm_rcx_write(vcpu, ecx);
1522	kvm_rdx_write(vcpu, edx);
1523	return kvm_skip_emulated_instruction(vcpu);
1524}
1525EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
v6.13.7
   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 */
  36u32 kvm_cpu_caps[NR_KVM_CPU_CAPS] __read_mostly;
  37EXPORT_SYMBOL_GPL(kvm_cpu_caps);
  38
  39struct cpuid_xstate_sizes {
  40	u32 eax;
  41	u32 ebx;
  42	u32 ecx;
  43};
  44
  45static struct cpuid_xstate_sizes xstate_sizes[XFEATURE_MAX] __ro_after_init;
  46
  47void __init kvm_init_xstate_sizes(void)
  48{
  49	u32 ign;
  50	int i;
  51
  52	for (i = XFEATURE_YMM; i < ARRAY_SIZE(xstate_sizes); i++) {
  53		struct cpuid_xstate_sizes *xs = &xstate_sizes[i];
  54
  55		cpuid_count(0xD, i, &xs->eax, &xs->ebx, &xs->ecx, &ign);
  56	}
  57}
  58
  59u32 xstate_required_size(u64 xstate_bv, bool compacted)
  60{
  61	int feature_bit = 0;
  62	u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
  63
  64	xstate_bv &= XFEATURE_MASK_EXTEND;
  65	while (xstate_bv) {
  66		if (xstate_bv & 0x1) {
  67			struct cpuid_xstate_sizes *xs = &xstate_sizes[feature_bit];
  68			u32 offset;
  69
  70			/* ECX[1]: 64B alignment in compacted form */
  71			if (compacted)
  72				offset = (xs->ecx & 0x2) ? ALIGN(ret, 64) : ret;
  73			else
  74				offset = xs->ebx;
  75			ret = max(ret, offset + xs->eax);
  76		}
  77
  78		xstate_bv >>= 1;
  79		feature_bit++;
  80	}
  81
  82	return ret;
  83}
  84
 
 
 
 
 
 
  85#define F feature_bit
  86
  87/* Scattered Flag - For features that are scattered by cpufeatures.h. */
  88#define SF(name)						\
  89({								\
  90	BUILD_BUG_ON(X86_FEATURE_##name >= MAX_CPU_FEATURES);	\
  91	(boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0);	\
  92})
  93
  94/*
  95 * Magic value used by KVM when querying userspace-provided CPUID entries and
  96 * doesn't care about the CPIUD index because the index of the function in
  97 * question is not significant.  Note, this magic value must have at least one
  98 * bit set in bits[63:32] and must be consumed as a u64 by cpuid_entry2_find()
  99 * to avoid false positives when processing guest CPUID input.
 100 */
 101#define KVM_CPUID_INDEX_NOT_SIGNIFICANT -1ull
 102
 103static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
 104	struct kvm_cpuid_entry2 *entries, int nent, u32 function, u64 index)
 105{
 106	struct kvm_cpuid_entry2 *e;
 107	int i;
 108
 109	/*
 110	 * KVM has a semi-arbitrary rule that querying the guest's CPUID model
 111	 * with IRQs disabled is disallowed.  The CPUID model can legitimately
 112	 * have over one hundred entries, i.e. the lookup is slow, and IRQs are
 113	 * typically disabled in KVM only when KVM is in a performance critical
 114	 * path, e.g. the core VM-Enter/VM-Exit run loop.  Nothing will break
 115	 * if this rule is violated, this assertion is purely to flag potential
 116	 * performance issues.  If this fires, consider moving the lookup out
 117	 * of the hotpath, e.g. by caching information during CPUID updates.
 118	 */
 119	lockdep_assert_irqs_enabled();
 120
 121	for (i = 0; i < nent; i++) {
 122		e = &entries[i];
 123
 124		if (e->function != function)
 125			continue;
 126
 127		/*
 128		 * If the index isn't significant, use the first entry with a
 129		 * matching function.  It's userspace's responsibility to not
 130		 * provide "duplicate" entries in all cases.
 131		 */
 132		if (!(e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) || e->index == index)
 133			return e;
 134
 135
 136		/*
 137		 * Similarly, use the first matching entry if KVM is doing a
 138		 * lookup (as opposed to emulating CPUID) for a function that's
 139		 * architecturally defined as not having a significant index.
 140		 */
 141		if (index == KVM_CPUID_INDEX_NOT_SIGNIFICANT) {
 142			/*
 143			 * Direct lookups from KVM should not diverge from what
 144			 * KVM defines internally (the architectural behavior).
 145			 */
 146			WARN_ON_ONCE(cpuid_function_is_indexed(function));
 147			return e;
 148		}
 149	}
 150
 151	return NULL;
 152}
 153
 154static int kvm_check_cpuid(struct kvm_vcpu *vcpu,
 155			   struct kvm_cpuid_entry2 *entries,
 156			   int nent)
 157{
 158	struct kvm_cpuid_entry2 *best;
 159	u64 xfeatures;
 160
 161	/*
 162	 * The existing code assumes virtual address is 48-bit or 57-bit in the
 163	 * canonical address checks; exit if it is ever changed.
 164	 */
 165	best = cpuid_entry2_find(entries, nent, 0x80000008,
 166				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 167	if (best) {
 168		int vaddr_bits = (best->eax & 0xff00) >> 8;
 169
 170		if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
 171			return -EINVAL;
 172	}
 173
 174	/*
 175	 * Exposing dynamic xfeatures to the guest requires additional
 176	 * enabling in the FPU, e.g. to expand the guest XSAVE state size.
 177	 */
 178	best = cpuid_entry2_find(entries, nent, 0xd, 0);
 179	if (!best)
 180		return 0;
 181
 182	xfeatures = best->eax | ((u64)best->edx << 32);
 183	xfeatures &= XFEATURE_MASK_USER_DYNAMIC;
 184	if (!xfeatures)
 185		return 0;
 186
 187	return fpu_enable_guest_xfd_features(&vcpu->arch.guest_fpu, xfeatures);
 188}
 189
 190/* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
 191static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
 192				 int nent)
 193{
 194	struct kvm_cpuid_entry2 *orig;
 195	int i;
 196
 197	if (nent != vcpu->arch.cpuid_nent)
 198		return -EINVAL;
 199
 200	for (i = 0; i < nent; i++) {
 201		orig = &vcpu->arch.cpuid_entries[i];
 202		if (e2[i].function != orig->function ||
 203		    e2[i].index != orig->index ||
 204		    e2[i].flags != orig->flags ||
 205		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
 206		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
 207			return -EINVAL;
 208	}
 209
 210	return 0;
 211}
 212
 213static struct kvm_hypervisor_cpuid __kvm_get_hypervisor_cpuid(struct kvm_cpuid_entry2 *entries,
 214							      int nent, const char *sig)
 215{
 216	struct kvm_hypervisor_cpuid cpuid = {};
 217	struct kvm_cpuid_entry2 *entry;
 218	u32 base;
 219
 220	for_each_possible_hypervisor_cpuid_base(base) {
 221		entry = cpuid_entry2_find(entries, nent, base, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 
 
 222
 223		if (entry) {
 224			u32 signature[3];
 225
 226			signature[0] = entry->ebx;
 227			signature[1] = entry->ecx;
 228			signature[2] = entry->edx;
 229
 230			if (!memcmp(signature, sig, sizeof(signature))) {
 231				cpuid.base = base;
 232				cpuid.limit = entry->eax;
 233				break;
 234			}
 235		}
 236	}
 237
 238	return cpuid;
 239}
 240
 241static struct kvm_hypervisor_cpuid kvm_get_hypervisor_cpuid(struct kvm_vcpu *vcpu,
 242							    const char *sig)
 243{
 244	return __kvm_get_hypervisor_cpuid(vcpu->arch.cpuid_entries,
 245					  vcpu->arch.cpuid_nent, sig);
 246}
 
 247
 248static struct kvm_cpuid_entry2 *__kvm_find_kvm_cpuid_features(struct kvm_cpuid_entry2 *entries,
 249							      int nent, u32 kvm_cpuid_base)
 250{
 251	return cpuid_entry2_find(entries, nent, kvm_cpuid_base | KVM_CPUID_FEATURES,
 252				 KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 253}
 254
 255static struct kvm_cpuid_entry2 *kvm_find_kvm_cpuid_features(struct kvm_vcpu *vcpu)
 256{
 257	u32 base = vcpu->arch.kvm_cpuid.base;
 258
 259	if (!base)
 260		return NULL;
 261
 262	return __kvm_find_kvm_cpuid_features(vcpu->arch.cpuid_entries,
 263					     vcpu->arch.cpuid_nent, base);
 264}
 265
 266void kvm_update_pv_runtime(struct kvm_vcpu *vcpu)
 267{
 268	struct kvm_cpuid_entry2 *best = kvm_find_kvm_cpuid_features(vcpu);
 269
 270	/*
 271	 * save the feature bitmap to avoid cpuid lookup for every PV
 272	 * operation
 273	 */
 274	if (best)
 275		vcpu->arch.pv_cpuid.features = best->eax;
 276}
 277
 278/*
 279 * Calculate guest's supported XCR0 taking into account guest CPUID data and
 280 * KVM's supported XCR0 (comprised of host's XCR0 and KVM_SUPPORTED_XCR0).
 281 */
 282static u64 cpuid_get_supported_xcr0(struct kvm_cpuid_entry2 *entries, int nent)
 283{
 284	struct kvm_cpuid_entry2 *best;
 285
 286	best = cpuid_entry2_find(entries, nent, 0xd, 0);
 287	if (!best)
 288		return 0;
 289
 290	return (best->eax | ((u64)best->edx << 32)) & kvm_caps.supported_xcr0;
 291}
 292
 293static void __kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *entries,
 294				       int nent)
 295{
 296	struct kvm_cpuid_entry2 *best;
 297	struct kvm_hypervisor_cpuid kvm_cpuid;
 298
 299	best = cpuid_entry2_find(entries, nent, 1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 300	if (best) {
 301		/* Update OSXSAVE bit */
 302		if (boot_cpu_has(X86_FEATURE_XSAVE))
 303			cpuid_entry_change(best, X86_FEATURE_OSXSAVE,
 304					   kvm_is_cr4_bit_set(vcpu, X86_CR4_OSXSAVE));
 305
 306		cpuid_entry_change(best, X86_FEATURE_APIC,
 307			   vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE);
 308	}
 309
 310	best = cpuid_entry2_find(entries, nent, 7, 0);
 311	if (best && boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7)
 312		cpuid_entry_change(best, X86_FEATURE_OSPKE,
 313				   kvm_is_cr4_bit_set(vcpu, X86_CR4_PKE));
 314
 315	best = cpuid_entry2_find(entries, nent, 0xD, 0);
 316	if (best)
 317		best->ebx = xstate_required_size(vcpu->arch.xcr0, false);
 318
 319	best = cpuid_entry2_find(entries, nent, 0xD, 1);
 320	if (best && (cpuid_entry_has(best, X86_FEATURE_XSAVES) ||
 321		     cpuid_entry_has(best, X86_FEATURE_XSAVEC)))
 322		best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
 323
 324	kvm_cpuid = __kvm_get_hypervisor_cpuid(entries, nent, KVM_SIGNATURE);
 325	if (kvm_cpuid.base) {
 326		best = __kvm_find_kvm_cpuid_features(entries, nent, kvm_cpuid.base);
 327		if (kvm_hlt_in_guest(vcpu->kvm) && best)
 328			best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
 329	}
 330
 331	if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
 332		best = cpuid_entry2_find(entries, nent, 0x1, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 333		if (best)
 334			cpuid_entry_change(best, X86_FEATURE_MWAIT,
 335					   vcpu->arch.ia32_misc_enable_msr &
 336					   MSR_IA32_MISC_ENABLE_MWAIT);
 337	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 338}
 339
 340void kvm_update_cpuid_runtime(struct kvm_vcpu *vcpu)
 341{
 342	__kvm_update_cpuid_runtime(vcpu, vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
 343}
 344EXPORT_SYMBOL_GPL(kvm_update_cpuid_runtime);
 345
 346static bool kvm_cpuid_has_hyperv(struct kvm_cpuid_entry2 *entries, int nent)
 347{
 348#ifdef CONFIG_KVM_HYPERV
 349	struct kvm_cpuid_entry2 *entry;
 350
 351	entry = cpuid_entry2_find(entries, nent, HYPERV_CPUID_INTERFACE,
 352				  KVM_CPUID_INDEX_NOT_SIGNIFICANT);
 353	return entry && entry->eax == HYPERV_CPUID_SIGNATURE_EAX;
 354#else
 355	return false;
 356#endif
 357}
 358
 359static bool guest_cpuid_is_amd_or_hygon(struct kvm_vcpu *vcpu)
 360{
 361	struct kvm_cpuid_entry2 *entry;
 362
 363	entry = kvm_find_cpuid_entry(vcpu, 0);
 364	if (!entry)
 365		return false;
 366
 367	return is_guest_vendor_amd(entry->ebx, entry->ecx, entry->edx) ||
 368	       is_guest_vendor_hygon(entry->ebx, entry->ecx, entry->edx);
 369}
 370
 371static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)
 372{
 373	struct kvm_lapic *apic = vcpu->arch.apic;
 374	struct kvm_cpuid_entry2 *best;
 375	bool allow_gbpages;
 376
 377	BUILD_BUG_ON(KVM_NR_GOVERNED_FEATURES > KVM_MAX_NR_GOVERNED_FEATURES);
 378	bitmap_zero(vcpu->arch.governed_features.enabled,
 379		    KVM_MAX_NR_GOVERNED_FEATURES);
 380
 381	/*
 382	 * If TDP is enabled, let the guest use GBPAGES if they're supported in
 383	 * hardware.  The hardware page walker doesn't let KVM disable GBPAGES,
 384	 * i.e. won't treat them as reserved, and KVM doesn't redo the GVA->GPA
 385	 * walk for performance and complexity reasons.  Not to mention KVM
 386	 * _can't_ solve the problem because GVA->GPA walks aren't visible to
 387	 * KVM once a TDP translation is installed.  Mimic hardware behavior so
 388	 * that KVM's is at least consistent, i.e. doesn't randomly inject #PF.
 389	 * If TDP is disabled, honor *only* guest CPUID as KVM has full control
 390	 * and can install smaller shadow pages if the host lacks 1GiB support.
 391	 */
 392	allow_gbpages = tdp_enabled ? boot_cpu_has(X86_FEATURE_GBPAGES) :
 393				      guest_cpuid_has(vcpu, X86_FEATURE_GBPAGES);
 394	if (allow_gbpages)
 395		kvm_governed_feature_set(vcpu, X86_FEATURE_GBPAGES);
 396
 397	best = kvm_find_cpuid_entry(vcpu, 1);
 398	if (best && apic) {
 399		if (cpuid_entry_has(best, X86_FEATURE_TSC_DEADLINE_TIMER))
 400			apic->lapic_timer.timer_mode_mask = 3 << 17;
 401		else
 402			apic->lapic_timer.timer_mode_mask = 1 << 17;
 403
 404		kvm_apic_set_version(vcpu);
 405	}
 406
 407	vcpu->arch.guest_supported_xcr0 =
 408		cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent);
 409
 
 
 
 
 
 
 
 
 410	kvm_update_pv_runtime(vcpu);
 411
 412	vcpu->arch.is_amd_compatible = guest_cpuid_is_amd_or_hygon(vcpu);
 413	vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
 414	vcpu->arch.reserved_gpa_bits = kvm_vcpu_reserved_gpa_bits_raw(vcpu);
 415
 416	kvm_pmu_refresh(vcpu);
 417	vcpu->arch.cr4_guest_rsvd_bits =
 418	    __cr4_reserved_bits(guest_cpuid_has, vcpu);
 419
 420	kvm_hv_set_cpuid(vcpu, kvm_cpuid_has_hyperv(vcpu->arch.cpuid_entries,
 421						    vcpu->arch.cpuid_nent));
 422
 423	/* Invoke the vendor callback only after the above state is updated. */
 424	kvm_x86_call(vcpu_after_set_cpuid)(vcpu);
 425
 426	/*
 427	 * Except for the MMU, which needs to do its thing any vendor specific
 428	 * adjustments to the reserved GPA bits.
 429	 */
 430	kvm_mmu_after_set_cpuid(vcpu);
 431}
 432
 433int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
 434{
 435	struct kvm_cpuid_entry2 *best;
 436
 437	best = kvm_find_cpuid_entry(vcpu, 0x80000000);
 438	if (!best || best->eax < 0x80000008)
 439		goto not_found;
 440	best = kvm_find_cpuid_entry(vcpu, 0x80000008);
 441	if (best)
 442		return best->eax & 0xff;
 443not_found:
 444	return 36;
 445}
 446
 447/*
 448 * This "raw" version returns the reserved GPA bits without any adjustments for
 449 * encryption technologies that usurp bits.  The raw mask should be used if and
 450 * only if hardware does _not_ strip the usurped bits, e.g. in virtual MTRRs.
 451 */
 452u64 kvm_vcpu_reserved_gpa_bits_raw(struct kvm_vcpu *vcpu)
 453{
 454	return rsvd_bits(cpuid_maxphyaddr(vcpu), 63);
 455}
 456
 457static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
 458                        int nent)
 459{
 460	int r;
 461
 462	__kvm_update_cpuid_runtime(vcpu, e2, nent);
 463
 464	/*
 465	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
 466	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
 467	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
 468	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
 469	 * the core vCPU model on the fly. It would've been better to forbid any
 470	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
 471	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
 472	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
 473	 * whether the supplied CPUID data is equal to what's already set.
 474	 */
 475	if (kvm_vcpu_has_run(vcpu)) {
 476		r = kvm_cpuid_check_equal(vcpu, e2, nent);
 477		if (r)
 478			return r;
 479
 480		kvfree(e2);
 481		return 0;
 482	}
 483
 484#ifdef CONFIG_KVM_HYPERV
 485	if (kvm_cpuid_has_hyperv(e2, nent)) {
 486		r = kvm_hv_vcpu_init(vcpu);
 487		if (r)
 488			return r;
 489	}
 490#endif
 491
 492	r = kvm_check_cpuid(vcpu, e2, nent);
 493	if (r)
 494		return r;
 495
 496	kvfree(vcpu->arch.cpuid_entries);
 497	vcpu->arch.cpuid_entries = e2;
 498	vcpu->arch.cpuid_nent = nent;
 499
 500	vcpu->arch.kvm_cpuid = kvm_get_hypervisor_cpuid(vcpu, KVM_SIGNATURE);
 501#ifdef CONFIG_KVM_XEN
 502	vcpu->arch.xen.cpuid = kvm_get_hypervisor_cpuid(vcpu, XEN_SIGNATURE);
 503#endif
 504	kvm_vcpu_after_set_cpuid(vcpu);
 505
 506	return 0;
 507}
 508
 509/* when an old userspace process fills a new kernel module */
 510int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
 511			     struct kvm_cpuid *cpuid,
 512			     struct kvm_cpuid_entry __user *entries)
 513{
 514	int r, i;
 515	struct kvm_cpuid_entry *e = NULL;
 516	struct kvm_cpuid_entry2 *e2 = NULL;
 517
 518	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 519		return -E2BIG;
 520
 521	if (cpuid->nent) {
 522		e = vmemdup_array_user(entries, cpuid->nent, sizeof(*e));
 523		if (IS_ERR(e))
 524			return PTR_ERR(e);
 525
 526		e2 = kvmalloc_array(cpuid->nent, sizeof(*e2), GFP_KERNEL_ACCOUNT);
 527		if (!e2) {
 528			r = -ENOMEM;
 529			goto out_free_cpuid;
 530		}
 531	}
 532	for (i = 0; i < cpuid->nent; i++) {
 533		e2[i].function = e[i].function;
 534		e2[i].eax = e[i].eax;
 535		e2[i].ebx = e[i].ebx;
 536		e2[i].ecx = e[i].ecx;
 537		e2[i].edx = e[i].edx;
 538		e2[i].index = 0;
 539		e2[i].flags = 0;
 540		e2[i].padding[0] = 0;
 541		e2[i].padding[1] = 0;
 542		e2[i].padding[2] = 0;
 543	}
 544
 545	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
 546	if (r)
 547		kvfree(e2);
 548
 549out_free_cpuid:
 550	kvfree(e);
 551
 552	return r;
 553}
 554
 555int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
 556			      struct kvm_cpuid2 *cpuid,
 557			      struct kvm_cpuid_entry2 __user *entries)
 558{
 559	struct kvm_cpuid_entry2 *e2 = NULL;
 560	int r;
 561
 562	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 563		return -E2BIG;
 564
 565	if (cpuid->nent) {
 566		e2 = vmemdup_array_user(entries, cpuid->nent, sizeof(*e2));
 567		if (IS_ERR(e2))
 568			return PTR_ERR(e2);
 569	}
 570
 571	r = kvm_set_cpuid(vcpu, e2, cpuid->nent);
 572	if (r)
 573		kvfree(e2);
 574
 575	return r;
 576}
 577
 578int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
 579			      struct kvm_cpuid2 *cpuid,
 580			      struct kvm_cpuid_entry2 __user *entries)
 581{
 
 
 
 582	if (cpuid->nent < vcpu->arch.cpuid_nent)
 583		return -E2BIG;
 584
 585	if (copy_to_user(entries, vcpu->arch.cpuid_entries,
 586			 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
 587		return -EFAULT;
 
 588
 
 589	cpuid->nent = vcpu->arch.cpuid_nent;
 590	return 0;
 591}
 592
 593/* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */
 594static __always_inline void __kvm_cpu_cap_mask(unsigned int leaf)
 595{
 596	const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
 597	struct kvm_cpuid_entry2 entry;
 598
 599	reverse_cpuid_check(leaf);
 600
 601	cpuid_count(cpuid.function, cpuid.index,
 602		    &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
 603
 604	kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
 605}
 606
 607static __always_inline
 608void kvm_cpu_cap_init_kvm_defined(enum kvm_only_cpuid_leafs leaf, u32 mask)
 609{
 610	/* Use kvm_cpu_cap_mask for leafs that aren't KVM-only. */
 611	BUILD_BUG_ON(leaf < NCAPINTS);
 612
 613	kvm_cpu_caps[leaf] = mask;
 614
 615	__kvm_cpu_cap_mask(leaf);
 616}
 617
 618static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask)
 619{
 620	/* Use kvm_cpu_cap_init_kvm_defined for KVM-only leafs. */
 621	BUILD_BUG_ON(leaf >= NCAPINTS);
 622
 623	kvm_cpu_caps[leaf] &= mask;
 624
 625	__kvm_cpu_cap_mask(leaf);
 626}
 627
 628void kvm_set_cpu_caps(void)
 629{
 630#ifdef CONFIG_X86_64
 631	unsigned int f_gbpages = F(GBPAGES);
 632	unsigned int f_lm = F(LM);
 633	unsigned int f_xfd = F(XFD);
 634#else
 635	unsigned int f_gbpages = 0;
 636	unsigned int f_lm = 0;
 637	unsigned int f_xfd = 0;
 638#endif
 639	memset(kvm_cpu_caps, 0, sizeof(kvm_cpu_caps));
 640
 641	BUILD_BUG_ON(sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)) >
 642		     sizeof(boot_cpu_data.x86_capability));
 643
 644	memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
 645	       sizeof(kvm_cpu_caps) - (NKVMCAPINTS * sizeof(*kvm_cpu_caps)));
 646
 647	kvm_cpu_cap_mask(CPUID_1_ECX,
 648		/*
 649		 * NOTE: MONITOR (and MWAIT) are emulated as NOP, but *not*
 650		 * advertised to guests via CPUID!
 651		 */
 652		F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
 653		0 /* DS-CPL, VMX, SMX, EST */ |
 654		0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
 655		F(FMA) | F(CX16) | 0 /* xTPR Update */ | F(PDCM) |
 656		F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
 657		F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
 658		0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
 659		F(F16C) | F(RDRAND)
 660	);
 661	/* KVM emulates x2apic in software irrespective of host support. */
 662	kvm_cpu_cap_set(X86_FEATURE_X2APIC);
 663
 664	kvm_cpu_cap_mask(CPUID_1_EDX,
 665		F(FPU) | F(VME) | F(DE) | F(PSE) |
 666		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 667		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
 668		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 669		F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
 670		0 /* Reserved, DS, ACPI */ | F(MMX) |
 671		F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
 672		0 /* HTT, TM, Reserved, PBE */
 673	);
 674
 675	kvm_cpu_cap_mask(CPUID_7_0_EBX,
 676		F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) |
 677		F(FDP_EXCPTN_ONLY) | F(SMEP) | F(BMI2) | F(ERMS) | F(INVPCID) |
 678		F(RTM) | F(ZERO_FCS_FDS) | 0 /*MPX*/ | F(AVX512F) |
 679		F(AVX512DQ) | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) |
 680		F(CLFLUSHOPT) | F(CLWB) | 0 /*INTEL_PT*/ | F(AVX512PF) |
 681		F(AVX512ER) | F(AVX512CD) | F(SHA_NI) | F(AVX512BW) |
 682		F(AVX512VL));
 683
 684	kvm_cpu_cap_mask(CPUID_7_ECX,
 685		F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
 686		F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
 687		F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
 688		F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/ |
 689		F(SGX_LC) | F(BUS_LOCK_DETECT)
 690	);
 691	/* Set LA57 based on hardware capability. */
 692	if (cpuid_ecx(7) & F(LA57))
 693		kvm_cpu_cap_set(X86_FEATURE_LA57);
 694
 695	/*
 696	 * PKU not yet implemented for shadow paging and requires OSPKE
 697	 * to be set on the host. Clear it if that is not the case
 698	 */
 699	if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
 700		kvm_cpu_cap_clear(X86_FEATURE_PKU);
 701
 702	kvm_cpu_cap_mask(CPUID_7_EDX,
 703		F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
 704		F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
 705		F(MD_CLEAR) | F(AVX512_VP2INTERSECT) | F(FSRM) |
 706		F(SERIALIZE) | F(TSXLDTRK) | F(AVX512_FP16) |
 707		F(AMX_TILE) | F(AMX_INT8) | F(AMX_BF16) | F(FLUSH_L1D)
 708	);
 709
 710	/* TSC_ADJUST and ARCH_CAPABILITIES are emulated in software. */
 711	kvm_cpu_cap_set(X86_FEATURE_TSC_ADJUST);
 712	kvm_cpu_cap_set(X86_FEATURE_ARCH_CAPABILITIES);
 713
 714	if (boot_cpu_has(X86_FEATURE_AMD_IBPB_RET) &&
 715	    boot_cpu_has(X86_FEATURE_AMD_IBPB) &&
 716	    boot_cpu_has(X86_FEATURE_AMD_IBRS))
 717		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL);
 718	if (boot_cpu_has(X86_FEATURE_STIBP))
 719		kvm_cpu_cap_set(X86_FEATURE_INTEL_STIBP);
 720	if (boot_cpu_has(X86_FEATURE_AMD_SSBD))
 721		kvm_cpu_cap_set(X86_FEATURE_SPEC_CTRL_SSBD);
 722
 723	kvm_cpu_cap_mask(CPUID_7_1_EAX,
 724		F(SHA512) | F(SM3) | F(SM4) | F(AVX_VNNI) | F(AVX512_BF16) |
 725		F(CMPCCXADD) | F(FZRM) | F(FSRS) | F(FSRC) | F(AMX_FP16) |
 726		F(AVX_IFMA) | F(LAM)
 727	);
 728
 729	kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX,
 730		F(AVX_VNNI_INT8) | F(AVX_NE_CONVERT) | F(AMX_COMPLEX) |
 731		F(AVX_VNNI_INT16) | F(PREFETCHITI) | F(AVX10)
 732	);
 733
 734	kvm_cpu_cap_init_kvm_defined(CPUID_7_2_EDX,
 735		F(INTEL_PSFD) | F(IPRED_CTRL) | F(RRSBA_CTRL) | F(DDPD_U) |
 736		F(BHI_CTRL) | F(MCDT_NO)
 737	);
 738
 739	kvm_cpu_cap_mask(CPUID_D_1_EAX,
 740		F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | F(XSAVES) | f_xfd
 741	);
 742
 743	kvm_cpu_cap_init_kvm_defined(CPUID_12_EAX,
 744		SF(SGX1) | SF(SGX2) | SF(SGX_EDECCSSA)
 745	);
 746
 747	kvm_cpu_cap_init_kvm_defined(CPUID_24_0_EBX,
 748		F(AVX10_128) | F(AVX10_256) | F(AVX10_512)
 749	);
 750
 751	kvm_cpu_cap_mask(CPUID_8000_0001_ECX,
 752		F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
 753		F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
 754		F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
 755		0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
 756		F(TOPOEXT) | 0 /* PERFCTR_CORE */
 757	);
 758
 759	kvm_cpu_cap_mask(CPUID_8000_0001_EDX,
 760		F(FPU) | F(VME) | F(DE) | F(PSE) |
 761		F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 762		F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
 763		F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 764		F(PAT) | F(PSE36) | 0 /* Reserved */ |
 765		F(NX) | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
 766		F(FXSR) | F(FXSR_OPT) | f_gbpages | F(RDTSCP) |
 767		0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW)
 768	);
 769
 770	if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
 771		kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
 772
 773	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0007_EDX,
 774		SF(CONSTANT_TSC)
 775	);
 776
 777	kvm_cpu_cap_mask(CPUID_8000_0008_EBX,
 778		F(CLZERO) | F(XSAVEERPTR) |
 779		F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
 780		F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON) |
 781		F(AMD_PSFD) | F(AMD_IBPB_RET)
 782	);
 783
 784	/*
 785	 * AMD has separate bits for each SPEC_CTRL bit.
 786	 * arch/x86/kernel/cpu/bugs.c is kind enough to
 787	 * record that in cpufeatures so use them.
 788	 */
 789	if (boot_cpu_has(X86_FEATURE_IBPB)) {
 790		kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB);
 791		if (boot_cpu_has(X86_FEATURE_SPEC_CTRL) &&
 792		    !boot_cpu_has_bug(X86_BUG_EIBRS_PBRSB))
 793			kvm_cpu_cap_set(X86_FEATURE_AMD_IBPB_RET);
 794	}
 795	if (boot_cpu_has(X86_FEATURE_IBRS))
 796		kvm_cpu_cap_set(X86_FEATURE_AMD_IBRS);
 797	if (boot_cpu_has(X86_FEATURE_STIBP))
 798		kvm_cpu_cap_set(X86_FEATURE_AMD_STIBP);
 799	if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD))
 800		kvm_cpu_cap_set(X86_FEATURE_AMD_SSBD);
 801	if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
 802		kvm_cpu_cap_set(X86_FEATURE_AMD_SSB_NO);
 803	/*
 804	 * The preference is to use SPEC CTRL MSR instead of the
 805	 * VIRT_SPEC MSR.
 806	 */
 807	if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
 808	    !boot_cpu_has(X86_FEATURE_AMD_SSBD))
 809		kvm_cpu_cap_set(X86_FEATURE_VIRT_SSBD);
 810
 811	/*
 812	 * Hide all SVM features by default, SVM will set the cap bits for
 813	 * features it emulates and/or exposes for L1.
 814	 */
 815	kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0);
 816
 817	kvm_cpu_cap_mask(CPUID_8000_001F_EAX,
 818		0 /* SME */ | 0 /* SEV */ | 0 /* VM_PAGE_FLUSH */ | 0 /* SEV_ES */ |
 819		F(SME_COHERENT));
 820
 821	kvm_cpu_cap_mask(CPUID_8000_0021_EAX,
 822		F(NO_NESTED_DATA_BP) | F(LFENCE_RDTSC) | 0 /* SmmPgCfgLock */ |
 823		F(NULL_SEL_CLR_BASE) | F(AUTOIBRS) | 0 /* PrefetchCtlMsr */ |
 824		F(WRMSR_XX_BASE_NS)
 825	);
 826
 827	kvm_cpu_cap_check_and_set(X86_FEATURE_SBPB);
 828	kvm_cpu_cap_check_and_set(X86_FEATURE_IBPB_BRTYPE);
 829	kvm_cpu_cap_check_and_set(X86_FEATURE_SRSO_NO);
 830
 831	kvm_cpu_cap_init_kvm_defined(CPUID_8000_0022_EAX,
 832		F(PERFMON_V2)
 833	);
 834
 835	/*
 836	 * Synthesize "LFENCE is serializing" into the AMD-defined entry in
 837	 * KVM's supported CPUID if the feature is reported as supported by the
 838	 * kernel.  LFENCE_RDTSC was a Linux-defined synthetic feature long
 839	 * before AMD joined the bandwagon, e.g. LFENCE is serializing on most
 840	 * CPUs that support SSE2.  On CPUs that don't support AMD's leaf,
 841	 * kvm_cpu_cap_mask() will unfortunately drop the flag due to ANDing
 842	 * the mask with the raw host CPUID, and reporting support in AMD's
 843	 * leaf can make it easier for userspace to detect the feature.
 844	 */
 845	if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
 846		kvm_cpu_cap_set(X86_FEATURE_LFENCE_RDTSC);
 847	if (!static_cpu_has_bug(X86_BUG_NULL_SEG))
 848		kvm_cpu_cap_set(X86_FEATURE_NULL_SEL_CLR_BASE);
 849	kvm_cpu_cap_set(X86_FEATURE_NO_SMM_CTL_MSR);
 850
 851	kvm_cpu_cap_mask(CPUID_C000_0001_EDX,
 852		F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
 853		F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
 854		F(PMM) | F(PMM_EN)
 855	);
 856
 857	/*
 858	 * Hide RDTSCP and RDPID if either feature is reported as supported but
 859	 * probing MSR_TSC_AUX failed.  This is purely a sanity check and
 860	 * should never happen, but the guest will likely crash if RDTSCP or
 861	 * RDPID is misreported, and KVM has botched MSR_TSC_AUX emulation in
 862	 * the past.  For example, the sanity check may fire if this instance of
 863	 * KVM is running as L1 on top of an older, broken KVM.
 864	 */
 865	if (WARN_ON((kvm_cpu_cap_has(X86_FEATURE_RDTSCP) ||
 866		     kvm_cpu_cap_has(X86_FEATURE_RDPID)) &&
 867		     !kvm_is_supported_user_return_msr(MSR_TSC_AUX))) {
 868		kvm_cpu_cap_clear(X86_FEATURE_RDTSCP);
 869		kvm_cpu_cap_clear(X86_FEATURE_RDPID);
 870	}
 871}
 872EXPORT_SYMBOL_GPL(kvm_set_cpu_caps);
 873
 874struct kvm_cpuid_array {
 875	struct kvm_cpuid_entry2 *entries;
 876	int maxnent;
 877	int nent;
 878};
 879
 880static struct kvm_cpuid_entry2 *get_next_cpuid(struct kvm_cpuid_array *array)
 881{
 882	if (array->nent >= array->maxnent)
 883		return NULL;
 884
 885	return &array->entries[array->nent++];
 886}
 887
 888static struct kvm_cpuid_entry2 *do_host_cpuid(struct kvm_cpuid_array *array,
 889					      u32 function, u32 index)
 890{
 891	struct kvm_cpuid_entry2 *entry = get_next_cpuid(array);
 892
 893	if (!entry)
 894		return NULL;
 895
 896	memset(entry, 0, sizeof(*entry));
 897	entry->function = function;
 898	entry->index = index;
 899	switch (function & 0xC0000000) {
 900	case 0x40000000:
 901		/* Hypervisor leaves are always synthesized by __do_cpuid_func.  */
 902		return entry;
 903
 904	case 0x80000000:
 905		/*
 906		 * 0x80000021 is sometimes synthesized by __do_cpuid_func, which
 907		 * would result in out-of-bounds calls to do_host_cpuid.
 908		 */
 909		{
 910			static int max_cpuid_80000000;
 911			if (!READ_ONCE(max_cpuid_80000000))
 912				WRITE_ONCE(max_cpuid_80000000, cpuid_eax(0x80000000));
 913			if (function > READ_ONCE(max_cpuid_80000000))
 914				return entry;
 915		}
 916		break;
 917
 918	default:
 919		break;
 920	}
 921
 922	cpuid_count(entry->function, entry->index,
 923		    &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
 924
 925	if (cpuid_function_is_indexed(function))
 926		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 927
 928	return entry;
 929}
 930
 931static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
 932{
 933	struct kvm_cpuid_entry2 *entry;
 934
 935	if (array->nent >= array->maxnent)
 936		return -E2BIG;
 937
 938	entry = &array->entries[array->nent];
 939	entry->function = func;
 940	entry->index = 0;
 941	entry->flags = 0;
 942
 943	switch (func) {
 944	case 0:
 945		entry->eax = 7;
 946		++array->nent;
 947		break;
 948	case 1:
 949		entry->ecx = F(MOVBE);
 950		++array->nent;
 951		break;
 952	case 7:
 953		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 954		entry->eax = 0;
 955		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
 956			entry->ecx = F(RDPID);
 957		++array->nent;
 958		break;
 959	default:
 960		break;
 961	}
 962
 963	return 0;
 964}
 965
 966static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
 967{
 968	struct kvm_cpuid_entry2 *entry;
 969	int r, i, max_idx;
 970
 971	/* all calls to cpuid_count() should be made on the same cpu */
 972	get_cpu();
 973
 974	r = -E2BIG;
 975
 976	entry = do_host_cpuid(array, function, 0);
 977	if (!entry)
 978		goto out;
 979
 980	switch (function) {
 981	case 0:
 982		/* Limited to the highest leaf implemented in KVM. */
 983		entry->eax = min(entry->eax, 0x24U);
 984		break;
 985	case 1:
 986		cpuid_entry_override(entry, CPUID_1_EDX);
 987		cpuid_entry_override(entry, CPUID_1_ECX);
 988		break;
 989	case 2:
 990		/*
 991		 * On ancient CPUs, function 2 entries are STATEFUL.  That is,
 992		 * CPUID(function=2, index=0) may return different results each
 993		 * time, with the least-significant byte in EAX enumerating the
 994		 * number of times software should do CPUID(2, 0).
 995		 *
 996		 * Modern CPUs, i.e. every CPU KVM has *ever* run on are less
 997		 * idiotic.  Intel's SDM states that EAX & 0xff "will always
 998		 * return 01H. Software should ignore this value and not
 999		 * interpret it as an informational descriptor", while AMD's
1000		 * APM states that CPUID(2) is reserved.
1001		 *
1002		 * WARN if a frankenstein CPU that supports virtualization and
1003		 * a stateful CPUID.0x2 is encountered.
1004		 */
1005		WARN_ON_ONCE((entry->eax & 0xff) > 1);
1006		break;
1007	/* functions 4 and 0x8000001d have additional index. */
1008	case 4:
1009	case 0x8000001d:
1010		/*
1011		 * Read entries until the cache type in the previous entry is
1012		 * zero, i.e. indicates an invalid entry.
1013		 */
1014		for (i = 1; entry->eax & 0x1f; ++i) {
1015			entry = do_host_cpuid(array, function, i);
1016			if (!entry)
1017				goto out;
1018		}
1019		break;
1020	case 6: /* Thermal management */
1021		entry->eax = 0x4; /* allow ARAT */
1022		entry->ebx = 0;
1023		entry->ecx = 0;
1024		entry->edx = 0;
1025		break;
1026	/* function 7 has additional index. */
1027	case 7:
1028		max_idx = entry->eax = min(entry->eax, 2u);
1029		cpuid_entry_override(entry, CPUID_7_0_EBX);
1030		cpuid_entry_override(entry, CPUID_7_ECX);
1031		cpuid_entry_override(entry, CPUID_7_EDX);
1032
1033		/* KVM only supports up to 0x7.2, capped above via min(). */
1034		if (max_idx >= 1) {
1035			entry = do_host_cpuid(array, function, 1);
1036			if (!entry)
1037				goto out;
1038
1039			cpuid_entry_override(entry, CPUID_7_1_EAX);
1040			cpuid_entry_override(entry, CPUID_7_1_EDX);
1041			entry->ebx = 0;
1042			entry->ecx = 0;
1043		}
1044		if (max_idx >= 2) {
1045			entry = do_host_cpuid(array, function, 2);
1046			if (!entry)
1047				goto out;
1048
1049			cpuid_entry_override(entry, CPUID_7_2_EDX);
1050			entry->ecx = 0;
1051			entry->ebx = 0;
1052			entry->eax = 0;
1053		}
1054		break;
1055	case 0xa: { /* Architectural Performance Monitoring */
1056		union cpuid10_eax eax;
1057		union cpuid10_edx edx;
1058
1059		if (!enable_pmu || !static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
1060			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1061			break;
1062		}
1063
1064		eax.split.version_id = kvm_pmu_cap.version;
1065		eax.split.num_counters = kvm_pmu_cap.num_counters_gp;
1066		eax.split.bit_width = kvm_pmu_cap.bit_width_gp;
1067		eax.split.mask_length = kvm_pmu_cap.events_mask_len;
1068		edx.split.num_counters_fixed = kvm_pmu_cap.num_counters_fixed;
1069		edx.split.bit_width_fixed = kvm_pmu_cap.bit_width_fixed;
1070
1071		if (kvm_pmu_cap.version)
1072			edx.split.anythread_deprecated = 1;
1073		edx.split.reserved1 = 0;
1074		edx.split.reserved2 = 0;
1075
1076		entry->eax = eax.full;
1077		entry->ebx = kvm_pmu_cap.events_mask;
1078		entry->ecx = 0;
1079		entry->edx = edx.full;
1080		break;
1081	}
1082	case 0x1f:
1083	case 0xb:
1084		/*
1085		 * No topology; a valid topology is indicated by the presence
1086		 * of subleaf 1.
1087		 */
1088		entry->eax = entry->ebx = entry->ecx = 0;
1089		break;
1090	case 0xd: {
1091		u64 permitted_xcr0 = kvm_get_filtered_xcr0();
1092		u64 permitted_xss = kvm_caps.supported_xss;
1093
1094		entry->eax &= permitted_xcr0;
1095		entry->ebx = xstate_required_size(permitted_xcr0, false);
1096		entry->ecx = entry->ebx;
1097		entry->edx &= permitted_xcr0 >> 32;
1098		if (!permitted_xcr0)
1099			break;
1100
1101		entry = do_host_cpuid(array, function, 1);
1102		if (!entry)
1103			goto out;
1104
1105		cpuid_entry_override(entry, CPUID_D_1_EAX);
1106		if (entry->eax & (F(XSAVES)|F(XSAVEC)))
1107			entry->ebx = xstate_required_size(permitted_xcr0 | permitted_xss,
1108							  true);
1109		else {
1110			WARN_ON_ONCE(permitted_xss != 0);
1111			entry->ebx = 0;
1112		}
1113		entry->ecx &= permitted_xss;
1114		entry->edx &= permitted_xss >> 32;
1115
1116		for (i = 2; i < 64; ++i) {
1117			bool s_state;
1118			if (permitted_xcr0 & BIT_ULL(i))
1119				s_state = false;
1120			else if (permitted_xss & BIT_ULL(i))
1121				s_state = true;
1122			else
1123				continue;
1124
1125			entry = do_host_cpuid(array, function, i);
1126			if (!entry)
1127				goto out;
1128
1129			/*
1130			 * The supported check above should have filtered out
1131			 * invalid sub-leafs.  Only valid sub-leafs should
1132			 * reach this point, and they should have a non-zero
1133			 * save state size.  Furthermore, check whether the
1134			 * processor agrees with permitted_xcr0/permitted_xss
1135			 * on whether this is an XCR0- or IA32_XSS-managed area.
1136			 */
1137			if (WARN_ON_ONCE(!entry->eax || (entry->ecx & 0x1) != s_state)) {
1138				--array->nent;
1139				continue;
1140			}
1141
1142			if (!kvm_cpu_cap_has(X86_FEATURE_XFD))
1143				entry->ecx &= ~BIT_ULL(2);
1144			entry->edx = 0;
1145		}
1146		break;
1147	}
1148	case 0x12:
1149		/* Intel SGX */
1150		if (!kvm_cpu_cap_has(X86_FEATURE_SGX)) {
1151			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1152			break;
1153		}
1154
1155		/*
1156		 * Index 0: Sub-features, MISCSELECT (a.k.a extended features)
1157		 * and max enclave sizes.   The SGX sub-features and MISCSELECT
1158		 * are restricted by kernel and KVM capabilities (like most
1159		 * feature flags), while enclave size is unrestricted.
1160		 */
1161		cpuid_entry_override(entry, CPUID_12_EAX);
1162		entry->ebx &= SGX_MISC_EXINFO;
1163
1164		entry = do_host_cpuid(array, function, 1);
1165		if (!entry)
1166			goto out;
1167
1168		/*
1169		 * Index 1: SECS.ATTRIBUTES.  ATTRIBUTES are restricted a la
1170		 * feature flags.  Advertise all supported flags, including
1171		 * privileged attributes that require explicit opt-in from
1172		 * userspace.  ATTRIBUTES.XFRM is not adjusted as userspace is
1173		 * expected to derive it from supported XCR0.
1174		 */
1175		entry->eax &= SGX_ATTR_PRIV_MASK | SGX_ATTR_UNPRIV_MASK;
1176		entry->ebx &= 0;
1177		break;
1178	/* Intel PT */
1179	case 0x14:
1180		if (!kvm_cpu_cap_has(X86_FEATURE_INTEL_PT)) {
1181			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1182			break;
1183		}
1184
1185		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1186			if (!do_host_cpuid(array, function, i))
1187				goto out;
1188		}
1189		break;
1190	/* Intel AMX TILE */
1191	case 0x1d:
1192		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1193			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1194			break;
1195		}
1196
1197		for (i = 1, max_idx = entry->eax; i <= max_idx; ++i) {
1198			if (!do_host_cpuid(array, function, i))
1199				goto out;
1200		}
1201		break;
1202	case 0x1e: /* TMUL information */
1203		if (!kvm_cpu_cap_has(X86_FEATURE_AMX_TILE)) {
1204			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1205			break;
1206		}
1207		break;
1208	case 0x24: {
1209		u8 avx10_version;
1210
1211		if (!kvm_cpu_cap_has(X86_FEATURE_AVX10)) {
1212			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1213			break;
1214		}
1215
1216		/*
1217		 * The AVX10 version is encoded in EBX[7:0].  Note, the version
1218		 * is guaranteed to be >=1 if AVX10 is supported.  Note #2, the
1219		 * version needs to be captured before overriding EBX features!
1220		 */
1221		avx10_version = min_t(u8, entry->ebx & 0xff, 1);
1222		cpuid_entry_override(entry, CPUID_24_0_EBX);
1223		entry->ebx |= avx10_version;
1224
1225		entry->eax = 0;
1226		entry->ecx = 0;
1227		entry->edx = 0;
1228		break;
1229	}
1230	case KVM_CPUID_SIGNATURE: {
1231		const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
1232		entry->eax = KVM_CPUID_FEATURES;
1233		entry->ebx = sigptr[0];
1234		entry->ecx = sigptr[1];
1235		entry->edx = sigptr[2];
1236		break;
1237	}
1238	case KVM_CPUID_FEATURES:
1239		entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
1240			     (1 << KVM_FEATURE_NOP_IO_DELAY) |
1241			     (1 << KVM_FEATURE_CLOCKSOURCE2) |
1242			     (1 << KVM_FEATURE_ASYNC_PF) |
1243			     (1 << KVM_FEATURE_PV_EOI) |
1244			     (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
1245			     (1 << KVM_FEATURE_PV_UNHALT) |
1246			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
1247			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
1248			     (1 << KVM_FEATURE_PV_SEND_IPI) |
1249			     (1 << KVM_FEATURE_POLL_CONTROL) |
1250			     (1 << KVM_FEATURE_PV_SCHED_YIELD) |
1251			     (1 << KVM_FEATURE_ASYNC_PF_INT);
1252
1253		if (sched_info_on())
1254			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
1255
1256		entry->ebx = 0;
1257		entry->ecx = 0;
1258		entry->edx = 0;
1259		break;
1260	case 0x80000000:
1261		entry->eax = min(entry->eax, 0x80000022);
1262		/*
1263		 * Serializing LFENCE is reported in a multitude of ways, and
1264		 * NullSegClearsBase is not reported in CPUID on Zen2; help
1265		 * userspace by providing the CPUID leaf ourselves.
1266		 *
1267		 * However, only do it if the host has CPUID leaf 0x8000001d.
1268		 * QEMU thinks that it can query the host blindly for that
1269		 * CPUID leaf if KVM reports that it supports 0x8000001d or
1270		 * above.  The processor merrily returns values from the
1271		 * highest Intel leaf which QEMU tries to use as the guest's
1272		 * 0x8000001d.  Even worse, this can result in an infinite
1273		 * loop if said highest leaf has no subleaves indexed by ECX.
1274		 */
1275		if (entry->eax >= 0x8000001d &&
1276		    (static_cpu_has(X86_FEATURE_LFENCE_RDTSC)
1277		     || !static_cpu_has_bug(X86_BUG_NULL_SEG)))
1278			entry->eax = max(entry->eax, 0x80000021);
1279		break;
1280	case 0x80000001:
1281		entry->ebx &= ~GENMASK(27, 16);
1282		cpuid_entry_override(entry, CPUID_8000_0001_EDX);
1283		cpuid_entry_override(entry, CPUID_8000_0001_ECX);
1284		break;
1285	case 0x80000005:
1286		/*  Pass host L1 cache and TLB info. */
1287		break;
1288	case 0x80000006:
1289		/* Drop reserved bits, pass host L2 cache and TLB info. */
1290		entry->edx &= ~GENMASK(17, 16);
1291		break;
1292	case 0x80000007: /* Advanced power management */
1293		cpuid_entry_override(entry, CPUID_8000_0007_EDX);
1294
1295		/* mask against host */
1296		entry->edx &= boot_cpu_data.x86_power;
1297		entry->eax = entry->ebx = entry->ecx = 0;
1298		break;
1299	case 0x80000008: {
1300		/*
1301		 * GuestPhysAddrSize (EAX[23:16]) is intended for software
1302		 * use.
1303		 *
1304		 * KVM's ABI is to report the effective MAXPHYADDR for the
1305		 * guest in PhysAddrSize (phys_as), and the maximum
1306		 * *addressable* GPA in GuestPhysAddrSize (g_phys_as).
1307		 *
1308		 * GuestPhysAddrSize is valid if and only if TDP is enabled,
1309		 * in which case the max GPA that can be addressed by KVM may
1310		 * be less than the max GPA that can be legally generated by
1311		 * the guest, e.g. if MAXPHYADDR>48 but the CPU doesn't
1312		 * support 5-level TDP.
1313		 */
1314		unsigned int virt_as = max((entry->eax >> 8) & 0xff, 48U);
1315		unsigned int phys_as, g_phys_as;
1316
1317		/*
1318		 * If TDP (NPT) is disabled use the adjusted host MAXPHYADDR as
1319		 * the guest operates in the same PA space as the host, i.e.
1320		 * reductions in MAXPHYADDR for memory encryption affect shadow
1321		 * paging, too.
1322		 *
1323		 * If TDP is enabled, use the raw bare metal MAXPHYADDR as
1324		 * reductions to the HPAs do not affect GPAs.  The max
1325		 * addressable GPA is the same as the max effective GPA, except
1326		 * that it's capped at 48 bits if 5-level TDP isn't supported
1327		 * (hardware processes bits 51:48 only when walking the fifth
1328		 * level page table).
1329		 */
1330		if (!tdp_enabled) {
1331			phys_as = boot_cpu_data.x86_phys_bits;
1332			g_phys_as = 0;
1333		} else {
1334			phys_as = entry->eax & 0xff;
1335			g_phys_as = phys_as;
1336			if (kvm_mmu_get_max_tdp_level() < 5)
1337				g_phys_as = min(g_phys_as, 48);
1338		}
1339
1340		entry->eax = phys_as | (virt_as << 8) | (g_phys_as << 16);
1341		entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
1342		entry->edx = 0;
1343		cpuid_entry_override(entry, CPUID_8000_0008_EBX);
1344		break;
1345	}
1346	case 0x8000000A:
1347		if (!kvm_cpu_cap_has(X86_FEATURE_SVM)) {
1348			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1349			break;
1350		}
1351		entry->eax = 1; /* SVM revision 1 */
1352		entry->ebx = 8; /* Lets support 8 ASIDs in case we add proper
1353				   ASID emulation to nested SVM */
1354		entry->ecx = 0; /* Reserved */
1355		cpuid_entry_override(entry, CPUID_8000_000A_EDX);
1356		break;
1357	case 0x80000019:
1358		entry->ecx = entry->edx = 0;
1359		break;
1360	case 0x8000001a:
1361		entry->eax &= GENMASK(2, 0);
1362		entry->ebx = entry->ecx = entry->edx = 0;
1363		break;
1364	case 0x8000001e:
1365		/* Do not return host topology information.  */
1366		entry->eax = entry->ebx = entry->ecx = 0;
1367		entry->edx = 0; /* reserved */
1368		break;
1369	case 0x8000001F:
1370		if (!kvm_cpu_cap_has(X86_FEATURE_SEV)) {
1371			entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1372		} else {
1373			cpuid_entry_override(entry, CPUID_8000_001F_EAX);
1374			/* Clear NumVMPL since KVM does not support VMPL.  */
1375			entry->ebx &= ~GENMASK(31, 12);
1376			/*
1377			 * Enumerate '0' for "PA bits reduction", the adjusted
1378			 * MAXPHYADDR is enumerated directly (see 0x80000008).
1379			 */
1380			entry->ebx &= ~GENMASK(11, 6);
1381		}
1382		break;
1383	case 0x80000020:
1384		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1385		break;
1386	case 0x80000021:
1387		entry->ebx = entry->ecx = entry->edx = 0;
1388		cpuid_entry_override(entry, CPUID_8000_0021_EAX);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1389		break;
1390	/* AMD Extended Performance Monitoring and Debug */
1391	case 0x80000022: {
1392		union cpuid_0x80000022_ebx ebx;
1393
1394		entry->ecx = entry->edx = 0;
1395		if (!enable_pmu || !kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2)) {
1396			entry->eax = entry->ebx = 0;
1397			break;
1398		}
1399
1400		cpuid_entry_override(entry, CPUID_8000_0022_EAX);
1401
1402		if (kvm_cpu_cap_has(X86_FEATURE_PERFMON_V2))
1403			ebx.split.num_core_pmc = kvm_pmu_cap.num_counters_gp;
1404		else if (kvm_cpu_cap_has(X86_FEATURE_PERFCTR_CORE))
1405			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS_CORE;
1406		else
1407			ebx.split.num_core_pmc = AMD64_NUM_COUNTERS;
1408
1409		entry->ebx = ebx.full;
1410		break;
1411	}
1412	/*Add support for Centaur's CPUID instruction*/
1413	case 0xC0000000:
1414		/*Just support up to 0xC0000004 now*/
1415		entry->eax = min(entry->eax, 0xC0000004);
1416		break;
1417	case 0xC0000001:
1418		cpuid_entry_override(entry, CPUID_C000_0001_EDX);
1419		break;
1420	case 3: /* Processor serial number */
1421	case 5: /* MONITOR/MWAIT */
1422	case 0xC0000002:
1423	case 0xC0000003:
1424	case 0xC0000004:
1425	default:
1426		entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
1427		break;
1428	}
1429
1430	r = 0;
1431
1432out:
1433	put_cpu();
1434
1435	return r;
1436}
1437
1438static int do_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1439			 unsigned int type)
1440{
1441	if (type == KVM_GET_EMULATED_CPUID)
1442		return __do_cpuid_func_emulated(array, func);
1443
1444	return __do_cpuid_func(array, func);
1445}
1446
1447#define CENTAUR_CPUID_SIGNATURE 0xC0000000
1448
1449static int get_cpuid_func(struct kvm_cpuid_array *array, u32 func,
1450			  unsigned int type)
1451{
1452	u32 limit;
1453	int r;
1454
1455	if (func == CENTAUR_CPUID_SIGNATURE &&
1456	    boot_cpu_data.x86_vendor != X86_VENDOR_CENTAUR)
1457		return 0;
1458
1459	r = do_cpuid_func(array, func, type);
1460	if (r)
1461		return r;
1462
1463	limit = array->entries[array->nent - 1].eax;
1464	for (func = func + 1; func <= limit; ++func) {
1465		r = do_cpuid_func(array, func, type);
1466		if (r)
1467			break;
1468	}
1469
1470	return r;
1471}
1472
1473static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
1474				 __u32 num_entries, unsigned int ioctl_type)
1475{
1476	int i;
1477	__u32 pad[3];
1478
1479	if (ioctl_type != KVM_GET_EMULATED_CPUID)
1480		return false;
1481
1482	/*
1483	 * We want to make sure that ->padding is being passed clean from
1484	 * userspace in case we want to use it for something in the future.
1485	 *
1486	 * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
1487	 * have to give ourselves satisfied only with the emulated side. /me
1488	 * sheds a tear.
1489	 */
1490	for (i = 0; i < num_entries; i++) {
1491		if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
1492			return true;
1493
1494		if (pad[0] || pad[1] || pad[2])
1495			return true;
1496	}
1497	return false;
1498}
1499
1500int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
1501			    struct kvm_cpuid_entry2 __user *entries,
1502			    unsigned int type)
1503{
1504	static const u32 funcs[] = {
1505		0, 0x80000000, CENTAUR_CPUID_SIGNATURE, KVM_CPUID_SIGNATURE,
1506	};
1507
1508	struct kvm_cpuid_array array = {
1509		.nent = 0,
1510	};
1511	int r, i;
1512
1513	if (cpuid->nent < 1)
1514		return -E2BIG;
1515	if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1516		cpuid->nent = KVM_MAX_CPUID_ENTRIES;
1517
1518	if (sanity_check_entries(entries, cpuid->nent, type))
1519		return -EINVAL;
1520
1521	array.entries = kvcalloc(cpuid->nent, sizeof(struct kvm_cpuid_entry2), GFP_KERNEL);
1522	if (!array.entries)
1523		return -ENOMEM;
1524
1525	array.maxnent = cpuid->nent;
1526
1527	for (i = 0; i < ARRAY_SIZE(funcs); i++) {
1528		r = get_cpuid_func(&array, funcs[i], type);
1529		if (r)
1530			goto out_free;
1531	}
1532	cpuid->nent = array.nent;
1533
1534	if (copy_to_user(entries, array.entries,
1535			 array.nent * sizeof(struct kvm_cpuid_entry2)))
1536		r = -EFAULT;
1537
1538out_free:
1539	kvfree(array.entries);
1540	return r;
1541}
1542
1543struct kvm_cpuid_entry2 *kvm_find_cpuid_entry_index(struct kvm_vcpu *vcpu,
1544						    u32 function, u32 index)
1545{
1546	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1547				 function, index);
1548}
1549EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry_index);
1550
1551struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
1552					      u32 function)
1553{
1554	return cpuid_entry2_find(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent,
1555				 function, KVM_CPUID_INDEX_NOT_SIGNIFICANT);
1556}
1557EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
1558
1559/*
1560 * Intel CPUID semantics treats any query for an out-of-range leaf as if the
1561 * highest basic leaf (i.e. CPUID.0H:EAX) were requested.  AMD CPUID semantics
1562 * returns all zeroes for any undefined leaf, whether or not the leaf is in
1563 * range.  Centaur/VIA follows Intel semantics.
1564 *
1565 * A leaf is considered out-of-range if its function is higher than the maximum
1566 * supported leaf of its associated class or if its associated class does not
1567 * exist.
1568 *
1569 * There are three primary classes to be considered, with their respective
1570 * ranges described as "<base> - <top>[,<base2> - <top2>] inclusive.  A primary
1571 * class exists if a guest CPUID entry for its <base> leaf exists.  For a given
1572 * class, CPUID.<base>.EAX contains the max supported leaf for the class.
1573 *
1574 *  - Basic:      0x00000000 - 0x3fffffff, 0x50000000 - 0x7fffffff
1575 *  - Hypervisor: 0x40000000 - 0x4fffffff
1576 *  - Extended:   0x80000000 - 0xbfffffff
1577 *  - Centaur:    0xc0000000 - 0xcfffffff
1578 *
1579 * The Hypervisor class is further subdivided into sub-classes that each act as
1580 * their own independent class associated with a 0x100 byte range.  E.g. if Qemu
1581 * is advertising support for both HyperV and KVM, the resulting Hypervisor
1582 * CPUID sub-classes are:
1583 *
1584 *  - HyperV:     0x40000000 - 0x400000ff
1585 *  - KVM:        0x40000100 - 0x400001ff
1586 */
1587static struct kvm_cpuid_entry2 *
1588get_out_of_range_cpuid_entry(struct kvm_vcpu *vcpu, u32 *fn_ptr, u32 index)
1589{
1590	struct kvm_cpuid_entry2 *basic, *class;
1591	u32 function = *fn_ptr;
1592
1593	basic = kvm_find_cpuid_entry(vcpu, 0);
1594	if (!basic)
1595		return NULL;
1596
1597	if (is_guest_vendor_amd(basic->ebx, basic->ecx, basic->edx) ||
1598	    is_guest_vendor_hygon(basic->ebx, basic->ecx, basic->edx))
1599		return NULL;
1600
1601	if (function >= 0x40000000 && function <= 0x4fffffff)
1602		class = kvm_find_cpuid_entry(vcpu, function & 0xffffff00);
1603	else if (function >= 0xc0000000)
1604		class = kvm_find_cpuid_entry(vcpu, 0xc0000000);
1605	else
1606		class = kvm_find_cpuid_entry(vcpu, function & 0x80000000);
1607
1608	if (class && function <= class->eax)
1609		return NULL;
1610
1611	/*
1612	 * Leaf specific adjustments are also applied when redirecting to the
1613	 * max basic entry, e.g. if the max basic leaf is 0xb but there is no
1614	 * entry for CPUID.0xb.index (see below), then the output value for EDX
1615	 * needs to be pulled from CPUID.0xb.1.
1616	 */
1617	*fn_ptr = basic->eax;
1618
1619	/*
1620	 * The class does not exist or the requested function is out of range;
1621	 * the effective CPUID entry is the max basic leaf.  Note, the index of
1622	 * the original requested leaf is observed!
1623	 */
1624	return kvm_find_cpuid_entry_index(vcpu, basic->eax, index);
1625}
1626
1627bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1628	       u32 *ecx, u32 *edx, bool exact_only)
1629{
1630	u32 orig_function = *eax, function = *eax, index = *ecx;
1631	struct kvm_cpuid_entry2 *entry;
1632	bool exact, used_max_basic = false;
1633
1634	entry = kvm_find_cpuid_entry_index(vcpu, function, index);
1635	exact = !!entry;
1636
1637	if (!entry && !exact_only) {
1638		entry = get_out_of_range_cpuid_entry(vcpu, &function, index);
1639		used_max_basic = !!entry;
1640	}
1641
1642	if (entry) {
1643		*eax = entry->eax;
1644		*ebx = entry->ebx;
1645		*ecx = entry->ecx;
1646		*edx = entry->edx;
1647		if (function == 7 && index == 0) {
1648			u64 data;
1649		        if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1650			    (data & TSX_CTRL_CPUID_CLEAR))
1651				*ebx &= ~(F(RTM) | F(HLE));
1652		} else if (function == 0x80000007) {
1653			if (kvm_hv_invtsc_suppressed(vcpu))
1654				*edx &= ~SF(CONSTANT_TSC);
1655		}
1656	} else {
1657		*eax = *ebx = *ecx = *edx = 0;
1658		/*
1659		 * When leaf 0BH or 1FH is defined, CL is pass-through
1660		 * and EDX is always the x2APIC ID, even for undefined
1661		 * subleaves. Index 1 will exist iff the leaf is
1662		 * implemented, so we pass through CL iff leaf 1
1663		 * exists. EDX can be copied from any existing index.
1664		 */
1665		if (function == 0xb || function == 0x1f) {
1666			entry = kvm_find_cpuid_entry_index(vcpu, function, 1);
1667			if (entry) {
1668				*ecx = index & 0xff;
1669				*edx = entry->edx;
1670			}
1671		}
1672	}
1673	trace_kvm_cpuid(orig_function, index, *eax, *ebx, *ecx, *edx, exact,
1674			used_max_basic);
1675	return exact;
1676}
1677EXPORT_SYMBOL_GPL(kvm_cpuid);
1678
1679int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1680{
1681	u32 eax, ebx, ecx, edx;
1682
1683	if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1684		return 1;
1685
1686	eax = kvm_rax_read(vcpu);
1687	ecx = kvm_rcx_read(vcpu);
1688	kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, false);
1689	kvm_rax_write(vcpu, eax);
1690	kvm_rbx_write(vcpu, ebx);
1691	kvm_rcx_write(vcpu, ecx);
1692	kvm_rdx_write(vcpu, edx);
1693	return kvm_skip_emulated_instruction(vcpu);
1694}
1695EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);