arch/x86/kvm/cpuid.c

   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 "cpuid.h"
  22 #include "lapic.h"
  23 #include "mmu.h"
  24 #include "trace.h"
  25 #include "pmu.h"
  26
  27 static u32 xstate_required_size(u64 xstate_bv, bool compacted)
  28 {
  29         int feature_bit = 0;
  30         u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
  31
  32         xstate_bv &= XFEATURE_MASK_EXTEND;
  33         while (xstate_bv) {
  34                 if (xstate_bv & 0x1) {
  35                         u32 eax, ebx, ecx, edx, offset;
  36                         cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
  37                         offset = compacted ? ret : ebx;
  38                         ret = max(ret, offset + eax);
  39                 }
  40
  41                 xstate_bv >>= 1;
  42                 feature_bit++;
  43         }
  44
  45         return ret;
  46 }
  47
  48 bool kvm_mpx_supported(void)
  49 {
  50         return ((host_xcr0 & (XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR))
  51                  && kvm_x86_ops->mpx_supported());
  52 }
  53 EXPORT_SYMBOL_GPL(kvm_mpx_supported);
  54
  55 u64 kvm_supported_xcr0(void)
  56 {
  57         u64 xcr0 = KVM_SUPPORTED_XCR0 & host_xcr0;
  58
  59         if (!kvm_mpx_supported())
  60                 xcr0 &= ~(XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
  61
  62         return xcr0;
  63 }
  64
  65 #define F feature_bit
  66
  67 int kvm_update_cpuid(struct kvm_vcpu *vcpu)
  68 {
  69         struct kvm_cpuid_entry2 *best;
  70         struct kvm_lapic *apic = vcpu->arch.apic;
  71
  72         best = kvm_find_cpuid_entry(vcpu, 1, 0);
  73         if (!best)
  74                 return 0;
  75
  76         /* Update OSXSAVE bit */
  77         if (boot_cpu_has(X86_FEATURE_XSAVE) && best->function == 0x1) {
  78                 best->ecx &= ~F(OSXSAVE);
  79                 if (kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE))
  80                         best->ecx |= F(OSXSAVE);
  81         }
  82
  83         best->edx &= ~F(APIC);
  84         if (vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE)
  85                 best->edx |= F(APIC);
  86
  87         if (apic) {
  88                 if (best->ecx & F(TSC_DEADLINE_TIMER))
  89                         apic->lapic_timer.timer_mode_mask = 3 << 17;
  90                 else
  91                         apic->lapic_timer.timer_mode_mask = 1 << 17;
  92         }
  93
  94         best = kvm_find_cpuid_entry(vcpu, 7, 0);
  95         if (best) {
  96                 /* Update OSPKE bit */
  97                 if (boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7) {
  98                         best->ecx &= ~F(OSPKE);
  99                         if (kvm_read_cr4_bits(vcpu, X86_CR4_PKE))
 100                                 best->ecx |= F(OSPKE);
 101                 }
 102         }
 103
 104         best = kvm_find_cpuid_entry(vcpu, 0xD, 0);
 105         if (!best) {
 106                 vcpu->arch.guest_supported_xcr0 = 0;
 107                 vcpu->arch.guest_xstate_size = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
 108         } else {
 109                 vcpu->arch.guest_supported_xcr0 =
 110                         (best->eax | ((u64)best->edx << 32)) &
 111                         kvm_supported_xcr0();
 112                 vcpu->arch.guest_xstate_size = best->ebx =
 113                         xstate_required_size(vcpu->arch.xcr0, false);
 114         }
 115
 116         best = kvm_find_cpuid_entry(vcpu, 0xD, 1);
 117         if (best && (best->eax & (F(XSAVES) | F(XSAVEC))))
 118                 best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
 119
 120         /*
 121          * The existing code assumes virtual address is 48-bit or 57-bit in the
 122          * canonical address checks; exit if it is ever changed.
 123          */
 124         best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
 125         if (best) {
 126                 int vaddr_bits = (best->eax & 0xff00) >> 8;
 127
 128                 if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
 129                         return -EINVAL;
 130         }
 131
 132         best = kvm_find_cpuid_entry(vcpu, KVM_CPUID_FEATURES, 0);
 133         if (kvm_hlt_in_guest(vcpu->kvm) && best &&
 134                 (best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
 135                 best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
 136
 137         if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
 138                 best = kvm_find_cpuid_entry(vcpu, 0x1, 0);
 139                 if (best) {
 140                         if (vcpu->arch.ia32_misc_enable_msr & MSR_IA32_MISC_ENABLE_MWAIT)
 141                                 best->ecx |= F(MWAIT);
 142                         else
 143                                 best->ecx &= ~F(MWAIT);
 144                 }
 145         }
 146
 147         /* Update physical-address width */
 148         vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
 149         kvm_mmu_reset_context(vcpu);
 150
 151         kvm_pmu_refresh(vcpu);
 152         return 0;
 153 }
 154
 155 static int is_efer_nx(void)
 156 {
 157         unsigned long long efer = 0;
 158
 159         rdmsrl_safe(MSR_EFER, &efer);
 160         return efer & EFER_NX;
 161 }
 162
 163 static void cpuid_fix_nx_cap(struct kvm_vcpu *vcpu)
 164 {
 165         int i;
 166         struct kvm_cpuid_entry2 *e, *entry;
 167
 168         entry = NULL;
 169         for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
 170                 e = &vcpu->arch.cpuid_entries[i];
 171                 if (e->function == 0x80000001) {
 172                         entry = e;
 173                         break;
 174                 }
 175         }
 176         if (entry && (entry->edx & F(NX)) && !is_efer_nx()) {
 177                 entry->edx &= ~F(NX);
 178                 printk(KERN_INFO "kvm: guest NX capability removed\n");
 179         }
 180 }
 181
 182 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
 183 {
 184         struct kvm_cpuid_entry2 *best;
 185
 186         best = kvm_find_cpuid_entry(vcpu, 0x80000000, 0);
 187         if (!best || best->eax < 0x80000008)
 188                 goto not_found;
 189         best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
 190         if (best)
 191                 return best->eax & 0xff;
 192 not_found:
 193         return 36;
 194 }
 195 EXPORT_SYMBOL_GPL(cpuid_query_maxphyaddr);
 196
 197 /* when an old userspace process fills a new kernel module */
 198 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
 199                              struct kvm_cpuid *cpuid,
 200                              struct kvm_cpuid_entry __user *entries)
 201 {
 202         int r, i;
 203         struct kvm_cpuid_entry *cpuid_entries = NULL;
 204
 205         r = -E2BIG;
 206         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 207                 goto out;
 208         r = -ENOMEM;
 209         if (cpuid->nent) {
 210                 cpuid_entries =
 211                         vmalloc(array_size(sizeof(struct kvm_cpuid_entry),
 212                                            cpuid->nent));
 213                 if (!cpuid_entries)
 214                         goto out;
 215                 r = -EFAULT;
 216                 if (copy_from_user(cpuid_entries, entries,
 217                                    cpuid->nent * sizeof(struct kvm_cpuid_entry)))
 218                         goto out;
 219         }
 220         for (i = 0; i < cpuid->nent; i++) {
 221                 vcpu->arch.cpuid_entries[i].function = cpuid_entries[i].function;
 222                 vcpu->arch.cpuid_entries[i].eax = cpuid_entries[i].eax;
 223                 vcpu->arch.cpuid_entries[i].ebx = cpuid_entries[i].ebx;
 224                 vcpu->arch.cpuid_entries[i].ecx = cpuid_entries[i].ecx;
 225                 vcpu->arch.cpuid_entries[i].edx = cpuid_entries[i].edx;
 226                 vcpu->arch.cpuid_entries[i].index = 0;
 227                 vcpu->arch.cpuid_entries[i].flags = 0;
 228                 vcpu->arch.cpuid_entries[i].padding[0] = 0;
 229                 vcpu->arch.cpuid_entries[i].padding[1] = 0;
 230                 vcpu->arch.cpuid_entries[i].padding[2] = 0;
 231         }
 232         vcpu->arch.cpuid_nent = cpuid->nent;
 233         cpuid_fix_nx_cap(vcpu);
 234         kvm_apic_set_version(vcpu);
 235         kvm_x86_ops->cpuid_update(vcpu);
 236         r = kvm_update_cpuid(vcpu);
 237
 238 out:
 239         vfree(cpuid_entries);
 240         return r;
 241 }
 242
 243 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
 244                               struct kvm_cpuid2 *cpuid,
 245                               struct kvm_cpuid_entry2 __user *entries)
 246 {
 247         int r;
 248
 249         r = -E2BIG;
 250         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 251                 goto out;
 252         r = -EFAULT;
 253         if (copy_from_user(&vcpu->arch.cpuid_entries, entries,
 254                            cpuid->nent * sizeof(struct kvm_cpuid_entry2)))
 255                 goto out;
 256         vcpu->arch.cpuid_nent = cpuid->nent;
 257         kvm_apic_set_version(vcpu);
 258         kvm_x86_ops->cpuid_update(vcpu);
 259         r = kvm_update_cpuid(vcpu);
 260 out:
 261         return r;
 262 }
 263
 264 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
 265                               struct kvm_cpuid2 *cpuid,
 266                               struct kvm_cpuid_entry2 __user *entries)
 267 {
 268         int r;
 269
 270         r = -E2BIG;
 271         if (cpuid->nent < vcpu->arch.cpuid_nent)
 272                 goto out;
 273         r = -EFAULT;
 274         if (copy_to_user(entries, &vcpu->arch.cpuid_entries,
 275                          vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
 276                 goto out;
 277         return 0;
 278
 279 out:
 280         cpuid->nent = vcpu->arch.cpuid_nent;
 281         return r;
 282 }
 283
 284 static __always_inline void cpuid_mask(u32 *word, int wordnum)
 285 {
 286         reverse_cpuid_check(wordnum);
 287         *word &= boot_cpu_data.x86_capability[wordnum];
 288 }
 289
 290 static void do_host_cpuid(struct kvm_cpuid_entry2 *entry, u32 function,
 291                            u32 index)
 292 {
 293         entry->function = function;
 294         entry->index = index;
 295         entry->flags = 0;
 296
 297         cpuid_count(entry->function, entry->index,
 298                     &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
 299
 300         switch (function) {
 301         case 2:
 302                 entry->flags |= KVM_CPUID_FLAG_STATEFUL_FUNC;
 303                 break;
 304         case 4:
 305         case 7:
 306         case 0xb:
 307         case 0xd:
 308         case 0xf:
 309         case 0x10:
 310         case 0x12:
 311         case 0x14:
 312         case 0x17:
 313         case 0x18:
 314         case 0x1f:
 315         case 0x8000001d:
 316                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 317                 break;
 318         }
 319 }
 320
 321 static int __do_cpuid_func_emulated(struct kvm_cpuid_entry2 *entry,
 322                                     u32 func, int *nent, int maxnent)
 323 {
 324         entry->function = func;
 325         entry->index = 0;
 326         entry->flags = 0;
 327
 328         switch (func) {
 329         case 0:
 330                 entry->eax = 7;
 331                 ++*nent;
 332                 break;
 333         case 1:
 334                 entry->ecx = F(MOVBE);
 335                 ++*nent;
 336                 break;
 337         case 7:
 338                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 339                 entry->eax = 0;
 340                 entry->ecx = F(RDPID);
 341                 ++*nent;
 342         default:
 343                 break;
 344         }
 345
 346         return 0;
 347 }
 348
 349 static inline void do_cpuid_7_mask(struct kvm_cpuid_entry2 *entry, int index)
 350 {
 351         unsigned f_invpcid = kvm_x86_ops->invpcid_supported() ? F(INVPCID) : 0;
 352         unsigned f_mpx = kvm_mpx_supported() ? F(MPX) : 0;
 353         unsigned f_umip = kvm_x86_ops->umip_emulated() ? F(UMIP) : 0;
 354         unsigned f_intel_pt = kvm_x86_ops->pt_supported() ? F(INTEL_PT) : 0;
 355         unsigned f_la57;
 356
 357         /* cpuid 7.0.ebx */
 358         const u32 kvm_cpuid_7_0_ebx_x86_features =
 359                 F(FSGSBASE) | F(BMI1) | F(HLE) | F(AVX2) | F(SMEP) |
 360                 F(BMI2) | F(ERMS) | f_invpcid | F(RTM) | f_mpx | F(RDSEED) |
 361                 F(ADX) | F(SMAP) | F(AVX512IFMA) | F(AVX512F) | F(AVX512PF) |
 362                 F(AVX512ER) | F(AVX512CD) | F(CLFLUSHOPT) | F(CLWB) | F(AVX512DQ) |
 363                 F(SHA_NI) | F(AVX512BW) | F(AVX512VL) | f_intel_pt;
 364
 365         /* cpuid 7.0.ecx*/
 366         const u32 kvm_cpuid_7_0_ecx_x86_features =
 367                 F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ | F(RDPID) |
 368                 F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
 369                 F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
 370                 F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/;
 371
 372         /* cpuid 7.0.edx*/
 373         const u32 kvm_cpuid_7_0_edx_x86_features =
 374                 F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
 375                 F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
 376                 F(MD_CLEAR);
 377
 378         /* cpuid 7.1.eax */
 379         const u32 kvm_cpuid_7_1_eax_x86_features =
 380                 F(AVX512_BF16);
 381
 382         switch (index) {
 383         case 0:
 384                 entry->eax = min(entry->eax, 1u);
 385                 entry->ebx &= kvm_cpuid_7_0_ebx_x86_features;
 386                 cpuid_mask(&entry->ebx, CPUID_7_0_EBX);
 387                 /* TSC_ADJUST is emulated */
 388                 entry->ebx |= F(TSC_ADJUST);
 389
 390                 entry->ecx &= kvm_cpuid_7_0_ecx_x86_features;
 391                 f_la57 = entry->ecx & F(LA57);
 392                 cpuid_mask(&entry->ecx, CPUID_7_ECX);
 393                 /* Set LA57 based on hardware capability. */
 394                 entry->ecx |= f_la57;
 395                 entry->ecx |= f_umip;
 396                 /* PKU is not yet implemented for shadow paging. */
 397                 if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
 398                         entry->ecx &= ~F(PKU);
 399
 400                 entry->edx &= kvm_cpuid_7_0_edx_x86_features;
 401                 cpuid_mask(&entry->edx, CPUID_7_EDX);
 402                 if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
 403                         entry->edx |= F(SPEC_CTRL);
 404                 if (boot_cpu_has(X86_FEATURE_STIBP))
 405                         entry->edx |= F(INTEL_STIBP);
 406                 if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD) ||
 407                     boot_cpu_has(X86_FEATURE_AMD_SSBD))
 408                         entry->edx |= F(SPEC_CTRL_SSBD);
 409                 /*
 410                  * We emulate ARCH_CAPABILITIES in software even
 411                  * if the host doesn't support it.
 412                  */
 413                 entry->edx |= F(ARCH_CAPABILITIES);
 414                 break;
 415         case 1:
 416                 entry->eax &= kvm_cpuid_7_1_eax_x86_features;
 417                 entry->ebx = 0;
 418                 entry->ecx = 0;
 419                 entry->edx = 0;
 420                 break;
 421         default:
 422                 WARN_ON_ONCE(1);
 423                 entry->eax = 0;
 424                 entry->ebx = 0;
 425                 entry->ecx = 0;
 426                 entry->edx = 0;
 427                 break;
 428         }
 429 }
 430
 431 static inline int __do_cpuid_func(struct kvm_cpuid_entry2 *entry, u32 function,
 432                                   int *nent, int maxnent)
 433 {
 434         int r;
 435         unsigned f_nx = is_efer_nx() ? F(NX) : 0;
 436 #ifdef CONFIG_X86_64
 437         unsigned f_gbpages = (kvm_x86_ops->get_lpage_level() == PT_PDPE_LEVEL)
 438                                 ? F(GBPAGES) : 0;
 439         unsigned f_lm = F(LM);
 440 #else
 441         unsigned f_gbpages = 0;
 442         unsigned f_lm = 0;
 443 #endif
 444         unsigned f_rdtscp = kvm_x86_ops->rdtscp_supported() ? F(RDTSCP) : 0;
 445         unsigned f_xsaves = kvm_x86_ops->xsaves_supported() ? F(XSAVES) : 0;
 446         unsigned f_intel_pt = kvm_x86_ops->pt_supported() ? F(INTEL_PT) : 0;
 447
 448         /* cpuid 1.edx */
 449         const u32 kvm_cpuid_1_edx_x86_features =
 450                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 451                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 452                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
 453                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 454                 F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
 455                 0 /* Reserved, DS, ACPI */ | F(MMX) |
 456                 F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
 457                 0 /* HTT, TM, Reserved, PBE */;
 458         /* cpuid 0x80000001.edx */
 459         const u32 kvm_cpuid_8000_0001_edx_x86_features =
 460                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 461                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 462                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
 463                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 464                 F(PAT) | F(PSE36) | 0 /* Reserved */ |
 465                 f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
 466                 F(FXSR) | F(FXSR_OPT) | f_gbpages | f_rdtscp |
 467                 0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW);
 468         /* cpuid 1.ecx */
 469         const u32 kvm_cpuid_1_ecx_x86_features =
 470                 /* NOTE: MONITOR (and MWAIT) are emulated as NOP,
 471                  * but *not* advertised to guests via CPUID ! */
 472                 F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
 473                 0 /* DS-CPL, VMX, SMX, EST */ |
 474                 0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
 475                 F(FMA) | F(CX16) | 0 /* xTPR Update, PDCM */ |
 476                 F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
 477                 F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
 478                 0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
 479                 F(F16C) | F(RDRAND);
 480         /* cpuid 0x80000001.ecx */
 481         const u32 kvm_cpuid_8000_0001_ecx_x86_features =
 482                 F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
 483                 F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
 484                 F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
 485                 0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
 486                 F(TOPOEXT) | F(PERFCTR_CORE);
 487
 488         /* cpuid 0x80000008.ebx */
 489         const u32 kvm_cpuid_8000_0008_ebx_x86_features =
 490                 F(CLZERO) | F(XSAVEERPTR) |
 491                 F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
 492                 F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON);
 493
 494         /* cpuid 0xC0000001.edx */
 495         const u32 kvm_cpuid_C000_0001_edx_x86_features =
 496                 F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
 497                 F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
 498                 F(PMM) | F(PMM_EN);
 499
 500         /* cpuid 0xD.1.eax */
 501         const u32 kvm_cpuid_D_1_eax_x86_features =
 502                 F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | f_xsaves;
 503
 504         /* all calls to cpuid_count() should be made on the same cpu */
 505         get_cpu();
 506
 507         r = -E2BIG;
 508
 509         if (WARN_ON(*nent >= maxnent))
 510                 goto out;
 511
 512         do_host_cpuid(entry, function, 0);
 513         ++*nent;
 514
 515         switch (function) {
 516         case 0:
 517                 /* Limited to the highest leaf implemented in KVM. */
 518                 entry->eax = min(entry->eax, 0x1fU);
 519                 break;
 520         case 1:
 521                 entry->edx &= kvm_cpuid_1_edx_x86_features;
 522                 cpuid_mask(&entry->edx, CPUID_1_EDX);
 523                 entry->ecx &= kvm_cpuid_1_ecx_x86_features;
 524                 cpuid_mask(&entry->ecx, CPUID_1_ECX);
 525                 /* we support x2apic emulation even if host does not support
 526                  * it since we emulate x2apic in software */
 527                 entry->ecx |= F(X2APIC);
 528                 break;
 529         /* function 2 entries are STATEFUL. That is, repeated cpuid commands
 530          * may return different values. This forces us to get_cpu() before
 531          * issuing the first command, and also to emulate this annoying behavior
 532          * in kvm_emulate_cpuid() using KVM_CPUID_FLAG_STATE_READ_NEXT */
 533         case 2: {
 534                 int t, times = entry->eax & 0xff;
 535
 536                 entry->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 537                 for (t = 1; t < times; ++t) {
 538                         if (*nent >= maxnent)
 539                                 goto out;
 540
 541                         do_host_cpuid(&entry[t], function, 0);
 542                         ++*nent;
 543                 }
 544                 break;
 545         }
 546         /* functions 4 and 0x8000001d have additional index. */
 547         case 4:
 548         case 0x8000001d: {
 549                 int i, cache_type;
 550
 551                 /* read more entries until cache_type is zero */
 552                 for (i = 1; ; ++i) {
 553                         if (*nent >= maxnent)
 554                                 goto out;
 555
 556                         cache_type = entry[i - 1].eax & 0x1f;
 557                         if (!cache_type)
 558                                 break;
 559                         do_host_cpuid(&entry[i], function, i);
 560                         ++*nent;
 561                 }
 562                 break;
 563         }
 564         case 6: /* Thermal management */
 565                 entry->eax = 0x4; /* allow ARAT */
 566                 entry->ebx = 0;
 567                 entry->ecx = 0;
 568                 entry->edx = 0;
 569                 break;
 570         /* function 7 has additional index. */
 571         case 7: {
 572                 int i;
 573
 574                 for (i = 0; ; ) {
 575                         do_cpuid_7_mask(&entry[i], i);
 576                         if (i == entry->eax)
 577                                 break;
 578                         if (*nent >= maxnent)
 579                                 goto out;
 580
 581                         ++i;
 582                         do_host_cpuid(&entry[i], function, i);
 583                         ++*nent;
 584                 }
 585                 break;
 586         }
 587         case 9:
 588                 break;
 589         case 0xa: { /* Architectural Performance Monitoring */
 590                 struct x86_pmu_capability cap;
 591                 union cpuid10_eax eax;
 592                 union cpuid10_edx edx;
 593
 594                 perf_get_x86_pmu_capability(&cap);
 595
 596                 /*
 597                  * Only support guest architectural pmu on a host
 598                  * with architectural pmu.
 599                  */
 600                 if (!cap.version)
 601                         memset(&cap, 0, sizeof(cap));
 602
 603                 eax.split.version_id = min(cap.version, 2);
 604                 eax.split.num_counters = cap.num_counters_gp;
 605                 eax.split.bit_width = cap.bit_width_gp;
 606                 eax.split.mask_length = cap.events_mask_len;
 607
 608                 edx.split.num_counters_fixed = cap.num_counters_fixed;
 609                 edx.split.bit_width_fixed = cap.bit_width_fixed;
 610                 edx.split.reserved = 0;
 611
 612                 entry->eax = eax.full;
 613                 entry->ebx = cap.events_mask;
 614                 entry->ecx = 0;
 615                 entry->edx = edx.full;
 616                 break;
 617         }
 618         /*
 619          * Per Intel's SDM, the 0x1f is a superset of 0xb,
 620          * thus they can be handled by common code.
 621          */
 622         case 0x1f:
 623         case 0xb: {
 624                 int i;
 625
 626                 /*
 627                  * We filled in entry[0] for CPUID(EAX=<function>,
 628                  * ECX=00H) above.  If its level type (ECX[15:8]) is
 629                  * zero, then the leaf is unimplemented, and we're
 630                  * done.  Otherwise, continue to populate entries
 631                  * until the level type (ECX[15:8]) of the previously
 632                  * added entry is zero.
 633                  */
 634                 for (i = 1; entry[i - 1].ecx & 0xff00; ++i) {
 635                         if (*nent >= maxnent)
 636                                 goto out;
 637
 638                         do_host_cpuid(&entry[i], function, i);
 639                         ++*nent;
 640                 }
 641                 break;
 642         }
 643         case 0xd: {
 644                 int idx, i;
 645                 u64 supported = kvm_supported_xcr0();
 646
 647                 entry->eax &= supported;
 648                 entry->ebx = xstate_required_size(supported, false);
 649                 entry->ecx = entry->ebx;
 650                 entry->edx &= supported >> 32;
 651                 if (!supported)
 652                         break;
 653
 654                 for (idx = 1, i = 1; idx < 64; ++idx) {
 655                         u64 mask = ((u64)1 << idx);
 656                         if (*nent >= maxnent)
 657                                 goto out;
 658
 659                         do_host_cpuid(&entry[i], function, idx);
 660                         if (idx == 1) {
 661                                 entry[i].eax &= kvm_cpuid_D_1_eax_x86_features;
 662                                 cpuid_mask(&entry[i].eax, CPUID_D_1_EAX);
 663                                 entry[i].ebx = 0;
 664                                 if (entry[i].eax & (F(XSAVES)|F(XSAVEC)))
 665                                         entry[i].ebx =
 666                                                 xstate_required_size(supported,
 667                                                                      true);
 668                         } else {
 669                                 if (entry[i].eax == 0 || !(supported & mask))
 670                                         continue;
 671                                 if (WARN_ON_ONCE(entry[i].ecx & 1))
 672                                         continue;
 673                         }
 674                         entry[i].ecx = 0;
 675                         entry[i].edx = 0;
 676                         ++*nent;
 677                         ++i;
 678                 }
 679                 break;
 680         }
 681         /* Intel PT */
 682         case 0x14: {
 683                 int t, times = entry->eax;
 684
 685                 if (!f_intel_pt)
 686                         break;
 687
 688                 for (t = 1; t <= times; ++t) {
 689                         if (*nent >= maxnent)
 690                                 goto out;
 691                         do_host_cpuid(&entry[t], function, t);
 692                         ++*nent;
 693                 }
 694                 break;
 695         }
 696         case KVM_CPUID_SIGNATURE: {
 697                 static const char signature[12] = "KVMKVMKVM\0\0";
 698                 const u32 *sigptr = (const u32 *)signature;
 699                 entry->eax = KVM_CPUID_FEATURES;
 700                 entry->ebx = sigptr[0];
 701                 entry->ecx = sigptr[1];
 702                 entry->edx = sigptr[2];
 703                 break;
 704         }
 705         case KVM_CPUID_FEATURES:
 706                 entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
 707                              (1 << KVM_FEATURE_NOP_IO_DELAY) |
 708                              (1 << KVM_FEATURE_CLOCKSOURCE2) |
 709                              (1 << KVM_FEATURE_ASYNC_PF) |
 710                              (1 << KVM_FEATURE_PV_EOI) |
 711                              (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
 712                              (1 << KVM_FEATURE_PV_UNHALT) |
 713                              (1 << KVM_FEATURE_PV_TLB_FLUSH) |
 714                              (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
 715                              (1 << KVM_FEATURE_PV_SEND_IPI) |
 716                              (1 << KVM_FEATURE_POLL_CONTROL) |
 717                              (1 << KVM_FEATURE_PV_SCHED_YIELD);
 718
 719                 if (sched_info_on())
 720                         entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
 721
 722                 entry->ebx = 0;
 723                 entry->ecx = 0;
 724                 entry->edx = 0;
 725                 break;
 726         case 0x80000000:
 727                 entry->eax = min(entry->eax, 0x8000001f);
 728                 break;
 729         case 0x80000001:
 730                 entry->edx &= kvm_cpuid_8000_0001_edx_x86_features;
 731                 cpuid_mask(&entry->edx, CPUID_8000_0001_EDX);
 732                 entry->ecx &= kvm_cpuid_8000_0001_ecx_x86_features;
 733                 cpuid_mask(&entry->ecx, CPUID_8000_0001_ECX);
 734                 break;
 735         case 0x80000007: /* Advanced power management */
 736                 /* invariant TSC is CPUID.80000007H:EDX[8] */
 737                 entry->edx &= (1 << 8);
 738                 /* mask against host */
 739                 entry->edx &= boot_cpu_data.x86_power;
 740                 entry->eax = entry->ebx = entry->ecx = 0;
 741                 break;
 742         case 0x80000008: {
 743                 unsigned g_phys_as = (entry->eax >> 16) & 0xff;
 744                 unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
 745                 unsigned phys_as = entry->eax & 0xff;
 746
 747                 if (!g_phys_as)
 748                         g_phys_as = phys_as;
 749                 entry->eax = g_phys_as | (virt_as << 8);
 750                 entry->edx = 0;
 751                 entry->ebx &= kvm_cpuid_8000_0008_ebx_x86_features;
 752                 cpuid_mask(&entry->ebx, CPUID_8000_0008_EBX);
 753                 /*
 754                  * AMD has separate bits for each SPEC_CTRL bit.
 755                  * arch/x86/kernel/cpu/bugs.c is kind enough to
 756                  * record that in cpufeatures so use them.
 757                  */
 758                 if (boot_cpu_has(X86_FEATURE_IBPB))
 759                         entry->ebx |= F(AMD_IBPB);
 760                 if (boot_cpu_has(X86_FEATURE_IBRS))
 761                         entry->ebx |= F(AMD_IBRS);
 762                 if (boot_cpu_has(X86_FEATURE_STIBP))
 763                         entry->ebx |= F(AMD_STIBP);
 764                 if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD) ||
 765                     boot_cpu_has(X86_FEATURE_AMD_SSBD))
 766                         entry->ebx |= F(AMD_SSBD);
 767                 if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
 768                         entry->ebx |= F(AMD_SSB_NO);
 769                 /*
 770                  * The preference is to use SPEC CTRL MSR instead of the
 771                  * VIRT_SPEC MSR.
 772                  */
 773                 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
 774                     !boot_cpu_has(X86_FEATURE_AMD_SSBD))
 775                         entry->ebx |= F(VIRT_SSBD);
 776                 break;
 777         }
 778         case 0x80000019:
 779                 entry->ecx = entry->edx = 0;
 780                 break;
 781         case 0x8000001a:
 782         case 0x8000001e:
 783                 break;
 784         /* Support memory encryption cpuid if host supports it */
 785         case 0x8000001F:
 786                 if (!boot_cpu_has(X86_FEATURE_SEV))
 787                         entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
 788                 break;
 789         /*Add support for Centaur's CPUID instruction*/
 790         case 0xC0000000:
 791                 /*Just support up to 0xC0000004 now*/
 792                 entry->eax = min(entry->eax, 0xC0000004);
 793                 break;
 794         case 0xC0000001:
 795                 entry->edx &= kvm_cpuid_C000_0001_edx_x86_features;
 796                 cpuid_mask(&entry->edx, CPUID_C000_0001_EDX);
 797                 break;
 798         case 3: /* Processor serial number */
 799         case 5: /* MONITOR/MWAIT */
 800         case 0xC0000002:
 801         case 0xC0000003:
 802         case 0xC0000004:
 803         default:
 804                 entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
 805                 break;
 806         }
 807
 808         kvm_x86_ops->set_supported_cpuid(function, entry);
 809
 810         r = 0;
 811
 812 out:
 813         put_cpu();
 814
 815         return r;
 816 }
 817
 818 static int do_cpuid_func(struct kvm_cpuid_entry2 *entry, u32 func,
 819                          int *nent, int maxnent, unsigned int type)
 820 {
 821         if (*nent >= maxnent)
 822                 return -E2BIG;
 823
 824         if (type == KVM_GET_EMULATED_CPUID)
 825                 return __do_cpuid_func_emulated(entry, func, nent, maxnent);
 826
 827         return __do_cpuid_func(entry, func, nent, maxnent);
 828 }
 829
 830 struct kvm_cpuid_param {
 831         u32 func;
 832         bool (*qualifier)(const struct kvm_cpuid_param *param);
 833 };
 834
 835 static bool is_centaur_cpu(const struct kvm_cpuid_param *param)
 836 {
 837         return boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR;
 838 }
 839
 840 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
 841                                  __u32 num_entries, unsigned int ioctl_type)
 842 {
 843         int i;
 844         __u32 pad[3];
 845
 846         if (ioctl_type != KVM_GET_EMULATED_CPUID)
 847                 return false;
 848
 849         /*
 850          * We want to make sure that ->padding is being passed clean from
 851          * userspace in case we want to use it for something in the future.
 852          *
 853          * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
 854          * have to give ourselves satisfied only with the emulated side. /me
 855          * sheds a tear.
 856          */
 857         for (i = 0; i < num_entries; i++) {
 858                 if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
 859                         return true;
 860
 861                 if (pad[0] || pad[1] || pad[2])
 862                         return true;
 863         }
 864         return false;
 865 }
 866
 867 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
 868                             struct kvm_cpuid_entry2 __user *entries,
 869                             unsigned int type)
 870 {
 871         struct kvm_cpuid_entry2 *cpuid_entries;
 872         int limit, nent = 0, r = -E2BIG, i;
 873         u32 func;
 874         static const struct kvm_cpuid_param param[] = {
 875                 { .func = 0 },
 876                 { .func = 0x80000000 },
 877                 { .func = 0xC0000000, .qualifier = is_centaur_cpu },
 878                 { .func = KVM_CPUID_SIGNATURE },
 879         };
 880
 881         if (cpuid->nent < 1)
 882                 goto out;
 883         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 884                 cpuid->nent = KVM_MAX_CPUID_ENTRIES;
 885
 886         if (sanity_check_entries(entries, cpuid->nent, type))
 887                 return -EINVAL;
 888
 889         r = -ENOMEM;
 890         cpuid_entries = vzalloc(array_size(sizeof(struct kvm_cpuid_entry2),
 891                                            cpuid->nent));
 892         if (!cpuid_entries)
 893                 goto out;
 894
 895         r = 0;
 896         for (i = 0; i < ARRAY_SIZE(param); i++) {
 897                 const struct kvm_cpuid_param *ent = &param[i];
 898
 899                 if (ent->qualifier && !ent->qualifier(ent))
 900                         continue;
 901
 902                 r = do_cpuid_func(&cpuid_entries[nent], ent->func,
 903                                   &nent, cpuid->nent, type);
 904
 905                 if (r)
 906                         goto out_free;
 907
 908                 limit = cpuid_entries[nent - 1].eax;
 909                 for (func = ent->func + 1; func <= limit && nent < cpuid->nent && r == 0; ++func)
 910                         r = do_cpuid_func(&cpuid_entries[nent], func,
 911                                           &nent, cpuid->nent, type);
 912
 913                 if (r)
 914                         goto out_free;
 915         }
 916
 917         r = -EFAULT;
 918         if (copy_to_user(entries, cpuid_entries,
 919                          nent * sizeof(struct kvm_cpuid_entry2)))
 920                 goto out_free;
 921         cpuid->nent = nent;
 922         r = 0;
 923
 924 out_free:
 925         vfree(cpuid_entries);
 926 out:
 927         return r;
 928 }
 929
 930 static int move_to_next_stateful_cpuid_entry(struct kvm_vcpu *vcpu, int i)
 931 {
 932         struct kvm_cpuid_entry2 *e = &vcpu->arch.cpuid_entries[i];
 933         struct kvm_cpuid_entry2 *ej;
 934         int j = i;
 935         int nent = vcpu->arch.cpuid_nent;
 936
 937         e->flags &= ~KVM_CPUID_FLAG_STATE_READ_NEXT;
 938         /* when no next entry is found, the current entry[i] is reselected */
 939         do {
 940                 j = (j + 1) % nent;
 941                 ej = &vcpu->arch.cpuid_entries[j];
 942         } while (ej->function != e->function);
 943
 944         ej->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 945
 946         return j;
 947 }
 948
 949 /* find an entry with matching function, matching index (if needed), and that
 950  * should be read next (if it's stateful) */
 951 static int is_matching_cpuid_entry(struct kvm_cpuid_entry2 *e,
 952         u32 function, u32 index)
 953 {
 954         if (e->function != function)
 955                 return 0;
 956         if ((e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) && e->index != index)
 957                 return 0;
 958         if ((e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC) &&
 959             !(e->flags & KVM_CPUID_FLAG_STATE_READ_NEXT))
 960                 return 0;
 961         return 1;
 962 }
 963
 964 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
 965                                               u32 function, u32 index)
 966 {
 967         int i;
 968         struct kvm_cpuid_entry2 *best = NULL;
 969
 970         for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
 971                 struct kvm_cpuid_entry2 *e;
 972
 973                 e = &vcpu->arch.cpuid_entries[i];
 974                 if (is_matching_cpuid_entry(e, function, index)) {
 975                         if (e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC)
 976                                 move_to_next_stateful_cpuid_entry(vcpu, i);
 977                         best = e;
 978                         break;
 979                 }
 980         }
 981         return best;
 982 }
 983 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
 984
 985 /*
 986  * If the basic or extended CPUID leaf requested is higher than the
 987  * maximum supported basic or extended leaf, respectively, then it is
 988  * out of range.
 989  */
 990 static bool cpuid_function_in_range(struct kvm_vcpu *vcpu, u32 function)
 991 {
 992         struct kvm_cpuid_entry2 *max;
 993
 994         max = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
 995         return max && function <= max->eax;
 996 }
 997
 998 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
 999                u32 *ecx, u32 *edx, bool check_limit)
1000 {
1001         u32 function = *eax, index = *ecx;
1002         struct kvm_cpuid_entry2 *entry;
1003         struct kvm_cpuid_entry2 *max;
1004         bool found;
1005
1006         entry = kvm_find_cpuid_entry(vcpu, function, index);
1007         found = entry;
1008         /*
1009          * Intel CPUID semantics treats any query for an out-of-range
1010          * leaf as if the highest basic leaf (i.e. CPUID.0H:EAX) were
1011          * requested. AMD CPUID semantics returns all zeroes for any
1012          * undefined leaf, whether or not the leaf is in range.
1013          */
1014         if (!entry && check_limit && !guest_cpuid_is_amd(vcpu) &&
1015             !cpuid_function_in_range(vcpu, function)) {
1016                 max = kvm_find_cpuid_entry(vcpu, 0, 0);
1017                 if (max) {
1018                         function = max->eax;
1019                         entry = kvm_find_cpuid_entry(vcpu, function, index);
1020                 }
1021         }
1022         if (entry) {
1023                 *eax = entry->eax;
1024                 *ebx = entry->ebx;
1025                 *ecx = entry->ecx;
1026                 *edx = entry->edx;
1027                 if (function == 7 && index == 0) {
1028                         u64 data;
1029                         if (!__kvm_get_msr(vcpu, MSR_IA32_TSX_CTRL, &data, true) &&
1030                             (data & TSX_CTRL_CPUID_CLEAR))
1031                                 *ebx &= ~(F(RTM) | F(HLE));
1032                 }
1033         } else {
1034                 *eax = *ebx = *ecx = *edx = 0;
1035                 /*
1036                  * When leaf 0BH or 1FH is defined, CL is pass-through
1037                  * and EDX is always the x2APIC ID, even for undefined
1038                  * subleaves. Index 1 will exist iff the leaf is
1039                  * implemented, so we pass through CL iff leaf 1
1040                  * exists. EDX can be copied from any existing index.
1041                  */
1042                 if (function == 0xb || function == 0x1f) {
1043                         entry = kvm_find_cpuid_entry(vcpu, function, 1);
1044                         if (entry) {
1045                                 *ecx = index & 0xff;
1046                                 *edx = entry->edx;
1047                         }
1048                 }
1049         }
1050         trace_kvm_cpuid(function, *eax, *ebx, *ecx, *edx, found);
1051         return found;
1052 }
1053 EXPORT_SYMBOL_GPL(kvm_cpuid);
1054
1055 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1056 {
1057         u32 eax, ebx, ecx, edx;
1058
1059         if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1060                 return 1;
1061
1062         eax = kvm_rax_read(vcpu);
1063         ecx = kvm_rcx_read(vcpu);
1064         kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, true);
1065         kvm_rax_write(vcpu, eax);
1066         kvm_rbx_write(vcpu, ebx);
1067         kvm_rcx_write(vcpu, ecx);
1068         kvm_rdx_write(vcpu, edx);
1069         return kvm_skip_emulated_instruction(vcpu);
1070 }
1071 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);