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(x) bit(X86_FEATURE_##x)
  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 void cpuid_mask(u32 *word, int wordnum)
 285 {
 286         *word &= boot_cpu_data.x86_capability[wordnum];
 287 }
 288
 289 static void do_host_cpuid(struct kvm_cpuid_entry2 *entry, u32 function,
 290                            u32 index)
 291 {
 292         entry->function = function;
 293         entry->index = index;
 294         entry->flags = 0;
 295
 296         cpuid_count(entry->function, entry->index,
 297                     &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
 298
 299         switch (function) {
 300         case 2:
 301                 entry->flags |= KVM_CPUID_FLAG_STATEFUL_FUNC;
 302                 break;
 303         case 4:
 304         case 7:
 305         case 0xb:
 306         case 0xd:
 307         case 0xf:
 308         case 0x10:
 309         case 0x12:
 310         case 0x14:
 311         case 0x17:
 312         case 0x18:
 313         case 0x1f:
 314         case 0x8000001d:
 315                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 316                 break;
 317         }
 318 }
 319
 320 static int __do_cpuid_func_emulated(struct kvm_cpuid_entry2 *entry,
 321                                     u32 func, int *nent, int maxnent)
 322 {
 323         entry->function = func;
 324         entry->index = 0;
 325         entry->flags = 0;
 326
 327         switch (func) {
 328         case 0:
 329                 entry->eax = 7;
 330                 ++*nent;
 331                 break;
 332         case 1:
 333                 entry->ecx = F(MOVBE);
 334                 ++*nent;
 335                 break;
 336         case 7:
 337                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 338                 entry->eax = 0;
 339                 entry->ecx = F(RDPID);
 340                 ++*nent;
 341         default:
 342                 break;
 343         }
 344
 345         return 0;
 346 }
 347
 348 static inline void do_cpuid_7_mask(struct kvm_cpuid_entry2 *entry, int index)
 349 {
 350         unsigned f_invpcid = kvm_x86_ops->invpcid_supported() ? F(INVPCID) : 0;
 351         unsigned f_mpx = kvm_mpx_supported() ? F(MPX) : 0;
 352         unsigned f_umip = kvm_x86_ops->umip_emulated() ? F(UMIP) : 0;
 353         unsigned f_intel_pt = kvm_x86_ops->pt_supported() ? F(INTEL_PT) : 0;
 354         unsigned f_la57;
 355         unsigned f_pku = kvm_x86_ops->pku_supported() ? F(PKU) : 0;
 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) | 0 /*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                 entry->ecx |= f_pku;
 397                 /* PKU is not yet implemented for shadow paging. */
 398                 if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
 399                         entry->ecx &= ~F(PKU);
 400
 401                 entry->edx &= kvm_cpuid_7_0_edx_x86_features;
 402                 cpuid_mask(&entry->edx, CPUID_7_EDX);
 403                 if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
 404                         entry->edx |= F(SPEC_CTRL);
 405                 if (boot_cpu_has(X86_FEATURE_STIBP))
 406                         entry->edx |= F(INTEL_STIBP);
 407                 if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD) ||
 408                     boot_cpu_has(X86_FEATURE_AMD_SSBD))
 409                         entry->edx |= F(SPEC_CTRL_SSBD);
 410                 /*
 411                  * We emulate ARCH_CAPABILITIES in software even
 412                  * if the host doesn't support it.
 413                  */
 414                 entry->edx |= F(ARCH_CAPABILITIES);
 415                 break;
 416         case 1:
 417                 entry->eax &= kvm_cpuid_7_1_eax_x86_features;
 418                 entry->ebx = 0;
 419                 entry->ecx = 0;
 420                 entry->edx = 0;
 421                 break;
 422         default:
 423                 WARN_ON_ONCE(1);
 424                 entry->eax = 0;
 425                 entry->ebx = 0;
 426                 entry->ecx = 0;
 427                 entry->edx = 0;
 428                 break;
 429         }
 430 }
 431
 432 static inline int __do_cpuid_func(struct kvm_cpuid_entry2 *entry, u32 function,
 433                                   int *nent, int maxnent)
 434 {
 435         int r;
 436         unsigned f_nx = is_efer_nx() ? F(NX) : 0;
 437 #ifdef CONFIG_X86_64
 438         unsigned f_gbpages = (kvm_x86_ops->get_lpage_level() == PT_PDPE_LEVEL)
 439                                 ? F(GBPAGES) : 0;
 440         unsigned f_lm = F(LM);
 441 #else
 442         unsigned f_gbpages = 0;
 443         unsigned f_lm = 0;
 444 #endif
 445         unsigned f_rdtscp = kvm_x86_ops->rdtscp_supported() ? F(RDTSCP) : 0;
 446         unsigned f_xsaves = kvm_x86_ops->xsaves_supported() ? F(XSAVES) : 0;
 447         unsigned f_intel_pt = kvm_x86_ops->pt_supported() ? F(INTEL_PT) : 0;
 448
 449         /* cpuid 1.edx */
 450         const u32 kvm_cpuid_1_edx_x86_features =
 451                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 452                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 453                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
 454                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 455                 F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
 456                 0 /* Reserved, DS, ACPI */ | F(MMX) |
 457                 F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
 458                 0 /* HTT, TM, Reserved, PBE */;
 459         /* cpuid 0x80000001.edx */
 460         const u32 kvm_cpuid_8000_0001_edx_x86_features =
 461                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 462                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 463                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
 464                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 465                 F(PAT) | F(PSE36) | 0 /* Reserved */ |
 466                 f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
 467                 F(FXSR) | F(FXSR_OPT) | f_gbpages | f_rdtscp |
 468                 0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW);
 469         /* cpuid 1.ecx */
 470         const u32 kvm_cpuid_1_ecx_x86_features =
 471                 /* NOTE: MONITOR (and MWAIT) are emulated as NOP,
 472                  * but *not* advertised to guests via CPUID ! */
 473                 F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
 474                 0 /* DS-CPL, VMX, SMX, EST */ |
 475                 0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
 476                 F(FMA) | F(CX16) | 0 /* xTPR Update, PDCM */ |
 477                 F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
 478                 F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
 479                 0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
 480                 F(F16C) | F(RDRAND);
 481         /* cpuid 0x80000001.ecx */
 482         const u32 kvm_cpuid_8000_0001_ecx_x86_features =
 483                 F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
 484                 F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
 485                 F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
 486                 0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
 487                 F(TOPOEXT) | F(PERFCTR_CORE);
 488
 489         /* cpuid 0x80000008.ebx */
 490         const u32 kvm_cpuid_8000_0008_ebx_x86_features =
 491                 F(CLZERO) | F(XSAVEERPTR) |
 492                 F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
 493                 F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON);
 494
 495         /* cpuid 0xC0000001.edx */
 496         const u32 kvm_cpuid_C000_0001_edx_x86_features =
 497                 F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
 498                 F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
 499                 F(PMM) | F(PMM_EN);
 500
 501         /* cpuid 0xD.1.eax */
 502         const u32 kvm_cpuid_D_1_eax_x86_features =
 503                 F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | f_xsaves;
 504
 505         /* all calls to cpuid_count() should be made on the same cpu */
 506         get_cpu();
 507
 508         r = -E2BIG;
 509
 510         if (WARN_ON(*nent >= maxnent))
 511                 goto out;
 512
 513         do_host_cpuid(entry, function, 0);
 514         ++*nent;
 515
 516         switch (function) {
 517         case 0:
 518                 /* Limited to the highest leaf implemented in KVM. */
 519                 entry->eax = min(entry->eax, 0x1fU);
 520                 break;
 521         case 1:
 522                 entry->edx &= kvm_cpuid_1_edx_x86_features;
 523                 cpuid_mask(&entry->edx, CPUID_1_EDX);
 524                 entry->ecx &= kvm_cpuid_1_ecx_x86_features;
 525                 cpuid_mask(&entry->ecx, CPUID_1_ECX);
 526                 /* we support x2apic emulation even if host does not support
 527                  * it since we emulate x2apic in software */
 528                 entry->ecx |= F(X2APIC);
 529                 break;
 530         /* function 2 entries are STATEFUL. That is, repeated cpuid commands
 531          * may return different values. This forces us to get_cpu() before
 532          * issuing the first command, and also to emulate this annoying behavior
 533          * in kvm_emulate_cpuid() using KVM_CPUID_FLAG_STATE_READ_NEXT */
 534         case 2: {
 535                 int t, times = entry->eax & 0xff;
 536
 537                 entry->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 538                 for (t = 1; t < times; ++t) {
 539                         if (*nent >= maxnent)
 540                                 goto out;
 541
 542                         do_host_cpuid(&entry[t], function, 0);
 543                         ++*nent;
 544                 }
 545                 break;
 546         }
 547         /* functions 4 and 0x8000001d have additional index. */
 548         case 4:
 549         case 0x8000001d: {
 550                 int i, cache_type;
 551
 552                 /* read more entries until cache_type is zero */
 553                 for (i = 1; ; ++i) {
 554                         if (*nent >= maxnent)
 555                                 goto out;
 556
 557                         cache_type = entry[i - 1].eax & 0x1f;
 558                         if (!cache_type)
 559                                 break;
 560                         do_host_cpuid(&entry[i], function, i);
 561                         ++*nent;
 562                 }
 563                 break;
 564         }
 565         case 6: /* Thermal management */
 566                 entry->eax = 0x4; /* allow ARAT */
 567                 entry->ebx = 0;
 568                 entry->ecx = 0;
 569                 entry->edx = 0;
 570                 break;
 571         /* function 7 has additional index. */
 572         case 7: {
 573                 int i;
 574
 575                 for (i = 0; ; ) {
 576                         do_cpuid_7_mask(&entry[i], i);
 577                         if (i == entry->eax)
 578                                 break;
 579                         if (*nent >= maxnent)
 580                                 goto out;
 581
 582                         ++i;
 583                         do_host_cpuid(&entry[i], function, i);
 584                         ++*nent;
 585                 }
 586                 break;
 587         }
 588         case 9:
 589                 break;
 590         case 0xa: { /* Architectural Performance Monitoring */
 591                 struct x86_pmu_capability cap;
 592                 union cpuid10_eax eax;
 593                 union cpuid10_edx edx;
 594
 595                 if (!static_cpu_has(X86_FEATURE_ARCH_PERFMON)) {
 596                         entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
 597                         break;
 598                 }
 599
 600                 perf_get_x86_pmu_capability(&cap);
 601
 602                 /*
 603                  * Only support guest architectural pmu on a host
 604                  * with architectural pmu.
 605                  */
 606                 if (!cap.version)
 607                         memset(&cap, 0, sizeof(cap));
 608
 609                 eax.split.version_id = min(cap.version, 2);
 610                 eax.split.num_counters = cap.num_counters_gp;
 611                 eax.split.bit_width = cap.bit_width_gp;
 612                 eax.split.mask_length = cap.events_mask_len;
 613
 614                 edx.split.num_counters_fixed = cap.num_counters_fixed;
 615                 edx.split.bit_width_fixed = cap.bit_width_fixed;
 616                 edx.split.reserved = 0;
 617
 618                 entry->eax = eax.full;
 619                 entry->ebx = cap.events_mask;
 620                 entry->ecx = 0;
 621                 entry->edx = edx.full;
 622                 break;
 623         }
 624         /*
 625          * Per Intel's SDM, the 0x1f is a superset of 0xb,
 626          * thus they can be handled by common code.
 627          */
 628         case 0x1f:
 629         case 0xb: {
 630                 int i;
 631
 632                 /*
 633                  * We filled in entry[0] for CPUID(EAX=<function>,
 634                  * ECX=00H) above.  If its level type (ECX[15:8]) is
 635                  * zero, then the leaf is unimplemented, and we're
 636                  * done.  Otherwise, continue to populate entries
 637                  * until the level type (ECX[15:8]) of the previously
 638                  * added entry is zero.
 639                  */
 640                 for (i = 1; entry[i - 1].ecx & 0xff00; ++i) {
 641                         if (*nent >= maxnent)
 642                                 goto out;
 643
 644                         do_host_cpuid(&entry[i], function, i);
 645                         ++*nent;
 646                 }
 647                 break;
 648         }
 649         case 0xd: {
 650                 int idx, i;
 651                 u64 supported = kvm_supported_xcr0();
 652
 653                 entry->eax &= supported;
 654                 entry->ebx = xstate_required_size(supported, false);
 655                 entry->ecx = entry->ebx;
 656                 entry->edx &= supported >> 32;
 657                 if (!supported)
 658                         break;
 659
 660                 for (idx = 1, i = 1; idx < 64; ++idx) {
 661                         u64 mask = ((u64)1 << idx);
 662                         if (*nent >= maxnent)
 663                                 goto out;
 664
 665                         do_host_cpuid(&entry[i], function, idx);
 666                         if (idx == 1) {
 667                                 entry[i].eax &= kvm_cpuid_D_1_eax_x86_features;
 668                                 cpuid_mask(&entry[i].eax, CPUID_D_1_EAX);
 669                                 entry[i].ebx = 0;
 670                                 if (entry[i].eax & (F(XSAVES)|F(XSAVEC)))
 671                                         entry[i].ebx =
 672                                                 xstate_required_size(supported,
 673                                                                      true);
 674                         } else {
 675                                 if (entry[i].eax == 0 || !(supported & mask))
 676                                         continue;
 677                                 if (WARN_ON_ONCE(entry[i].ecx & 1))
 678                                         continue;
 679                         }
 680                         entry[i].ecx = 0;
 681                         entry[i].edx = 0;
 682                         ++*nent;
 683                         ++i;
 684                 }
 685                 break;
 686         }
 687         /* Intel PT */
 688         case 0x14: {
 689                 int t, times = entry->eax;
 690
 691                 if (!f_intel_pt)
 692                         break;
 693
 694                 for (t = 1; t <= times; ++t) {
 695                         if (*nent >= maxnent)
 696                                 goto out;
 697                         do_host_cpuid(&entry[t], function, t);
 698                         ++*nent;
 699                 }
 700                 break;
 701         }
 702         case KVM_CPUID_SIGNATURE: {
 703                 static const char signature[12] = "KVMKVMKVM\0\0";
 704                 const u32 *sigptr = (const u32 *)signature;
 705                 entry->eax = KVM_CPUID_FEATURES;
 706                 entry->ebx = sigptr[0];
 707                 entry->ecx = sigptr[1];
 708                 entry->edx = sigptr[2];
 709                 break;
 710         }
 711         case KVM_CPUID_FEATURES:
 712                 entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
 713                              (1 << KVM_FEATURE_NOP_IO_DELAY) |
 714                              (1 << KVM_FEATURE_CLOCKSOURCE2) |
 715                              (1 << KVM_FEATURE_ASYNC_PF) |
 716                              (1 << KVM_FEATURE_PV_EOI) |
 717                              (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
 718                              (1 << KVM_FEATURE_PV_UNHALT) |
 719                              (1 << KVM_FEATURE_PV_TLB_FLUSH) |
 720                              (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
 721                              (1 << KVM_FEATURE_PV_SEND_IPI) |
 722                              (1 << KVM_FEATURE_POLL_CONTROL) |
 723                              (1 << KVM_FEATURE_PV_SCHED_YIELD);
 724
 725                 if (sched_info_on())
 726                         entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
 727
 728                 entry->ebx = 0;
 729                 entry->ecx = 0;
 730                 entry->edx = 0;
 731                 break;
 732         case 0x80000000:
 733                 entry->eax = min(entry->eax, 0x8000001f);
 734                 break;
 735         case 0x80000001:
 736                 entry->edx &= kvm_cpuid_8000_0001_edx_x86_features;
 737                 cpuid_mask(&entry->edx, CPUID_8000_0001_EDX);
 738                 entry->ecx &= kvm_cpuid_8000_0001_ecx_x86_features;
 739                 cpuid_mask(&entry->ecx, CPUID_8000_0001_ECX);
 740                 break;
 741         case 0x80000007: /* Advanced power management */
 742                 /* invariant TSC is CPUID.80000007H:EDX[8] */
 743                 entry->edx &= (1 << 8);
 744                 /* mask against host */
 745                 entry->edx &= boot_cpu_data.x86_power;
 746                 entry->eax = entry->ebx = entry->ecx = 0;
 747                 break;
 748         case 0x80000008: {
 749                 unsigned g_phys_as = (entry->eax >> 16) & 0xff;
 750                 unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
 751                 unsigned phys_as = entry->eax & 0xff;
 752
 753                 /*
 754                  * Use bare metal's MAXPHADDR if the CPU doesn't report guest
 755                  * MAXPHYADDR separately, or if TDP (NPT) is disabled, as the
 756                  * guest version "applies only to guests using nested paging".
 757                  */
 758                 if (!g_phys_as || !tdp_enabled)
 759                         g_phys_as = phys_as;
 760
 761                 entry->eax = g_phys_as | (virt_as << 8);
 762                 entry->ecx &= ~(GENMASK(31, 16) | GENMASK(11, 8));
 763                 entry->edx = 0;
 764                 entry->ebx &= kvm_cpuid_8000_0008_ebx_x86_features;
 765                 cpuid_mask(&entry->ebx, CPUID_8000_0008_EBX);
 766                 /*
 767                  * AMD has separate bits for each SPEC_CTRL bit.
 768                  * arch/x86/kernel/cpu/bugs.c is kind enough to
 769                  * record that in cpufeatures so use them.
 770                  */
 771                 if (boot_cpu_has(X86_FEATURE_IBPB))
 772                         entry->ebx |= F(AMD_IBPB);
 773                 if (boot_cpu_has(X86_FEATURE_IBRS))
 774                         entry->ebx |= F(AMD_IBRS);
 775                 if (boot_cpu_has(X86_FEATURE_STIBP))
 776                         entry->ebx |= F(AMD_STIBP);
 777                 if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD) ||
 778                     boot_cpu_has(X86_FEATURE_AMD_SSBD))
 779                         entry->ebx |= F(AMD_SSBD);
 780                 if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
 781                         entry->ebx |= F(AMD_SSB_NO);
 782                 /*
 783                  * The preference is to use SPEC CTRL MSR instead of the
 784                  * VIRT_SPEC MSR.
 785                  */
 786                 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
 787                     !boot_cpu_has(X86_FEATURE_AMD_SSBD))
 788                         entry->ebx |= F(VIRT_SSBD);
 789                 break;
 790         }
 791         case 0x80000019:
 792                 entry->ecx = entry->edx = 0;
 793                 break;
 794         case 0x8000001a:
 795                 entry->eax &= GENMASK(2, 0);
 796                 entry->ebx = entry->ecx = entry->edx = 0;
 797                 break;
 798         case 0x8000001e:
 799                 break;
 800         /*Add support for Centaur's CPUID instruction*/
 801         case 0xC0000000:
 802                 /*Just support up to 0xC0000004 now*/
 803                 entry->eax = min(entry->eax, 0xC0000004);
 804                 break;
 805         case 0xC0000001:
 806                 entry->edx &= kvm_cpuid_C000_0001_edx_x86_features;
 807                 cpuid_mask(&entry->edx, CPUID_C000_0001_EDX);
 808                 break;
 809         case 3: /* Processor serial number */
 810         case 5: /* MONITOR/MWAIT */
 811         case 0xC0000002:
 812         case 0xC0000003:
 813         case 0xC0000004:
 814         default:
 815                 entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
 816                 break;
 817         }
 818
 819         kvm_x86_ops->set_supported_cpuid(function, entry);
 820
 821         r = 0;
 822
 823 out:
 824         put_cpu();
 825
 826         return r;
 827 }
 828
 829 static int do_cpuid_func(struct kvm_cpuid_entry2 *entry, u32 func,
 830                          int *nent, int maxnent, unsigned int type)
 831 {
 832         if (*nent >= maxnent)
 833                 return -E2BIG;
 834
 835         if (type == KVM_GET_EMULATED_CPUID)
 836                 return __do_cpuid_func_emulated(entry, func, nent, maxnent);
 837
 838         return __do_cpuid_func(entry, func, nent, maxnent);
 839 }
 840
 841 #undef F
 842
 843 struct kvm_cpuid_param {
 844         u32 func;
 845         bool (*qualifier)(const struct kvm_cpuid_param *param);
 846 };
 847
 848 static bool is_centaur_cpu(const struct kvm_cpuid_param *param)
 849 {
 850         return boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR;
 851 }
 852
 853 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
 854                                  __u32 num_entries, unsigned int ioctl_type)
 855 {
 856         int i;
 857         __u32 pad[3];
 858
 859         if (ioctl_type != KVM_GET_EMULATED_CPUID)
 860                 return false;
 861
 862         /*
 863          * We want to make sure that ->padding is being passed clean from
 864          * userspace in case we want to use it for something in the future.
 865          *
 866          * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
 867          * have to give ourselves satisfied only with the emulated side. /me
 868          * sheds a tear.
 869          */
 870         for (i = 0; i < num_entries; i++) {
 871                 if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
 872                         return true;
 873
 874                 if (pad[0] || pad[1] || pad[2])
 875                         return true;
 876         }
 877         return false;
 878 }
 879
 880 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
 881                             struct kvm_cpuid_entry2 __user *entries,
 882                             unsigned int type)
 883 {
 884         struct kvm_cpuid_entry2 *cpuid_entries;
 885         int limit, nent = 0, r = -E2BIG, i;
 886         u32 func;
 887         static const struct kvm_cpuid_param param[] = {
 888                 { .func = 0 },
 889                 { .func = 0x80000000 },
 890                 { .func = 0xC0000000, .qualifier = is_centaur_cpu },
 891                 { .func = KVM_CPUID_SIGNATURE },
 892         };
 893
 894         if (cpuid->nent < 1)
 895                 goto out;
 896         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 897                 cpuid->nent = KVM_MAX_CPUID_ENTRIES;
 898
 899         if (sanity_check_entries(entries, cpuid->nent, type))
 900                 return -EINVAL;
 901
 902         r = -ENOMEM;
 903         cpuid_entries = vzalloc(array_size(sizeof(struct kvm_cpuid_entry2),
 904                                            cpuid->nent));
 905         if (!cpuid_entries)
 906                 goto out;
 907
 908         r = 0;
 909         for (i = 0; i < ARRAY_SIZE(param); i++) {
 910                 const struct kvm_cpuid_param *ent = &param[i];
 911
 912                 if (ent->qualifier && !ent->qualifier(ent))
 913                         continue;
 914
 915                 r = do_cpuid_func(&cpuid_entries[nent], ent->func,
 916                                   &nent, cpuid->nent, type);
 917
 918                 if (r)
 919                         goto out_free;
 920
 921                 limit = cpuid_entries[nent - 1].eax;
 922                 for (func = ent->func + 1; func <= limit && nent < cpuid->nent && r == 0; ++func)
 923                         r = do_cpuid_func(&cpuid_entries[nent], func,
 924                                           &nent, cpuid->nent, type);
 925
 926                 if (r)
 927                         goto out_free;
 928         }
 929
 930         r = -EFAULT;
 931         if (copy_to_user(entries, cpuid_entries,
 932                          nent * sizeof(struct kvm_cpuid_entry2)))
 933                 goto out_free;
 934         cpuid->nent = nent;
 935         r = 0;
 936
 937 out_free:
 938         vfree(cpuid_entries);
 939 out:
 940         return r;
 941 }
 942
 943 static int move_to_next_stateful_cpuid_entry(struct kvm_vcpu *vcpu, int i)
 944 {
 945         struct kvm_cpuid_entry2 *e = &vcpu->arch.cpuid_entries[i];
 946         struct kvm_cpuid_entry2 *ej;
 947         int j = i;
 948         int nent = vcpu->arch.cpuid_nent;
 949
 950         e->flags &= ~KVM_CPUID_FLAG_STATE_READ_NEXT;
 951         /* when no next entry is found, the current entry[i] is reselected */
 952         do {
 953                 j = (j + 1) % nent;
 954                 ej = &vcpu->arch.cpuid_entries[j];
 955         } while (ej->function != e->function);
 956
 957         ej->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 958
 959         return j;
 960 }
 961
 962 /* find an entry with matching function, matching index (if needed), and that
 963  * should be read next (if it's stateful) */
 964 static int is_matching_cpuid_entry(struct kvm_cpuid_entry2 *e,
 965         u32 function, u32 index)
 966 {
 967         if (e->function != function)
 968                 return 0;
 969         if ((e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) && e->index != index)
 970                 return 0;
 971         if ((e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC) &&
 972             !(e->flags & KVM_CPUID_FLAG_STATE_READ_NEXT))
 973                 return 0;
 974         return 1;
 975 }
 976
 977 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
 978                                               u32 function, u32 index)
 979 {
 980         int i;
 981         struct kvm_cpuid_entry2 *best = NULL;
 982
 983         for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
 984                 struct kvm_cpuid_entry2 *e;
 985
 986                 e = &vcpu->arch.cpuid_entries[i];
 987                 if (is_matching_cpuid_entry(e, function, index)) {
 988                         if (e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC)
 989                                 move_to_next_stateful_cpuid_entry(vcpu, i);
 990                         best = e;
 991                         break;
 992                 }
 993         }
 994         return best;
 995 }
 996 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
 997
 998 /*
 999  * If the basic or extended CPUID leaf requested is higher than the
1000  * maximum supported basic or extended leaf, respectively, then it is
1001  * out of range.
1002  */
1003 static bool cpuid_function_in_range(struct kvm_vcpu *vcpu, u32 function)
1004 {
1005         struct kvm_cpuid_entry2 *max;
1006
1007         max = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
1008         return max && function <= max->eax;
1009 }
1010
1011 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
1012                u32 *ecx, u32 *edx, bool check_limit)
1013 {
1014         u32 function = *eax, index = *ecx;
1015         struct kvm_cpuid_entry2 *entry;
1016         struct kvm_cpuid_entry2 *max;
1017         bool found;
1018
1019         entry = kvm_find_cpuid_entry(vcpu, function, index);
1020         found = entry;
1021         /*
1022          * Intel CPUID semantics treats any query for an out-of-range
1023          * leaf as if the highest basic leaf (i.e. CPUID.0H:EAX) were
1024          * requested. AMD CPUID semantics returns all zeroes for any
1025          * undefined leaf, whether or not the leaf is in range.
1026          */
1027         if (!entry && check_limit && !guest_cpuid_is_amd(vcpu) &&
1028             !cpuid_function_in_range(vcpu, function)) {
1029                 max = kvm_find_cpuid_entry(vcpu, 0, 0);
1030                 if (max) {
1031                         function = max->eax;
1032                         entry = kvm_find_cpuid_entry(vcpu, function, index);
1033                 }
1034         }
1035         if (entry) {
1036                 *eax = entry->eax;
1037                 *ebx = entry->ebx;
1038                 *ecx = entry->ecx;
1039                 *edx = entry->edx;
1040         } else {
1041                 *eax = *ebx = *ecx = *edx = 0;
1042                 /*
1043                  * When leaf 0BH or 1FH is defined, CL is pass-through
1044                  * and EDX is always the x2APIC ID, even for undefined
1045                  * subleaves. Index 1 will exist iff the leaf is
1046                  * implemented, so we pass through CL iff leaf 1
1047                  * exists. EDX can be copied from any existing index.
1048                  */
1049                 if (function == 0xb || function == 0x1f) {
1050                         entry = kvm_find_cpuid_entry(vcpu, function, 1);
1051                         if (entry) {
1052                                 *ecx = index & 0xff;
1053                                 *edx = entry->edx;
1054                         }
1055                 }
1056         }
1057         trace_kvm_cpuid(function, *eax, *ebx, *ecx, *edx, found);
1058         return found;
1059 }
1060 EXPORT_SYMBOL_GPL(kvm_cpuid);
1061
1062 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1063 {
1064         u32 eax, ebx, ecx, edx;
1065
1066         if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1067                 return 1;
1068
1069         eax = kvm_rax_read(vcpu);
1070         ecx = kvm_rcx_read(vcpu);
1071         kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, true);
1072         kvm_rax_write(vcpu, eax);
1073         kvm_rbx_write(vcpu, ebx);
1074         kvm_rcx_write(vcpu, ecx);
1075         kvm_rdx_write(vcpu, edx);
1076         return kvm_skip_emulated_instruction(vcpu);
1077 }
1078 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);