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
 356         /* cpuid 7.0.ebx */
 357         const u32 kvm_cpuid_7_0_ebx_x86_features =
 358                 F(FSGSBASE) | F(BMI1) | F(HLE) | F(AVX2) | F(SMEP) |
 359                 F(BMI2) | F(ERMS) | f_invpcid | F(RTM) | f_mpx | F(RDSEED) |
 360                 F(ADX) | F(SMAP) | F(AVX512IFMA) | F(AVX512F) | F(AVX512PF) |
 361                 F(AVX512ER) | F(AVX512CD) | F(CLFLUSHOPT) | F(CLWB) | F(AVX512DQ) |
 362                 F(SHA_NI) | F(AVX512BW) | F(AVX512VL) | f_intel_pt;
 363
 364         /* cpuid 7.0.ecx*/
 365         const u32 kvm_cpuid_7_0_ecx_x86_features =
 366                 F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ |
 367                 F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
 368                 F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
 369                 F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B) | 0 /*WAITPKG*/;
 370
 371         /* cpuid 7.0.edx*/
 372         const u32 kvm_cpuid_7_0_edx_x86_features =
 373                 F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
 374                 F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
 375                 F(MD_CLEAR);
 376
 377         /* cpuid 7.1.eax */
 378         const u32 kvm_cpuid_7_1_eax_x86_features =
 379                 F(AVX512_BF16);
 380
 381         switch (index) {
 382         case 0:
 383                 entry->eax = min(entry->eax, 1u);
 384                 entry->ebx &= kvm_cpuid_7_0_ebx_x86_features;
 385                 cpuid_mask(&entry->ebx, CPUID_7_0_EBX);
 386                 /* TSC_ADJUST is emulated */
 387                 entry->ebx |= F(TSC_ADJUST);
 388
 389                 entry->ecx &= kvm_cpuid_7_0_ecx_x86_features;
 390                 f_la57 = entry->ecx & F(LA57);
 391                 cpuid_mask(&entry->ecx, CPUID_7_ECX);
 392                 /* Set LA57 based on hardware capability. */
 393                 entry->ecx |= f_la57;
 394                 entry->ecx |= f_umip;
 395                 /* PKU is not yet implemented for shadow paging. */
 396                 if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
 397                         entry->ecx &= ~F(PKU);
 398
 399                 entry->edx &= kvm_cpuid_7_0_edx_x86_features;
 400                 cpuid_mask(&entry->edx, CPUID_7_EDX);
 401                 if (boot_cpu_has(X86_FEATURE_IBPB) && boot_cpu_has(X86_FEATURE_IBRS))
 402                         entry->edx |= F(SPEC_CTRL);
 403                 if (boot_cpu_has(X86_FEATURE_STIBP))
 404                         entry->edx |= F(INTEL_STIBP);
 405                 if (boot_cpu_has(X86_FEATURE_SSBD))
 406                         entry->edx |= F(SPEC_CTRL_SSBD);
 407                 /*
 408                  * We emulate ARCH_CAPABILITIES in software even
 409                  * if the host doesn't support it.
 410                  */
 411                 entry->edx |= F(ARCH_CAPABILITIES);
 412                 break;
 413         case 1:
 414                 entry->eax &= kvm_cpuid_7_1_eax_x86_features;
 415                 entry->ebx = 0;
 416                 entry->ecx = 0;
 417                 entry->edx = 0;
 418                 break;
 419         default:
 420                 WARN_ON_ONCE(1);
 421                 entry->eax = 0;
 422                 entry->ebx = 0;
 423                 entry->ecx = 0;
 424                 entry->edx = 0;
 425                 break;
 426         }
 427 }
 428
 429 static inline int __do_cpuid_func(struct kvm_cpuid_entry2 *entry, u32 function,
 430                                   int *nent, int maxnent)
 431 {
 432         int r;
 433         unsigned f_nx = is_efer_nx() ? F(NX) : 0;
 434 #ifdef CONFIG_X86_64
 435         unsigned f_gbpages = (kvm_x86_ops->get_lpage_level() == PT_PDPE_LEVEL)
 436                                 ? F(GBPAGES) : 0;
 437         unsigned f_lm = F(LM);
 438 #else
 439         unsigned f_gbpages = 0;
 440         unsigned f_lm = 0;
 441 #endif
 442         unsigned f_rdtscp = kvm_x86_ops->rdtscp_supported() ? F(RDTSCP) : 0;
 443         unsigned f_xsaves = kvm_x86_ops->xsaves_supported() ? F(XSAVES) : 0;
 444         unsigned f_intel_pt = kvm_x86_ops->pt_supported() ? F(INTEL_PT) : 0;
 445
 446         /* cpuid 1.edx */
 447         const u32 kvm_cpuid_1_edx_x86_features =
 448                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 449                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 450                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
 451                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 452                 F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
 453                 0 /* Reserved, DS, ACPI */ | F(MMX) |
 454                 F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
 455                 0 /* HTT, TM, Reserved, PBE */;
 456         /* cpuid 0x80000001.edx */
 457         const u32 kvm_cpuid_8000_0001_edx_x86_features =
 458                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 459                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 460                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
 461                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 462                 F(PAT) | F(PSE36) | 0 /* Reserved */ |
 463                 f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
 464                 F(FXSR) | F(FXSR_OPT) | f_gbpages | f_rdtscp |
 465                 0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW);
 466         /* cpuid 1.ecx */
 467         const u32 kvm_cpuid_1_ecx_x86_features =
 468                 /* NOTE: MONITOR (and MWAIT) are emulated as NOP,
 469                  * but *not* advertised to guests via CPUID ! */
 470                 F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
 471                 0 /* DS-CPL, VMX, SMX, EST */ |
 472                 0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
 473                 F(FMA) | F(CX16) | 0 /* xTPR Update, PDCM */ |
 474                 F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
 475                 F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
 476                 0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
 477                 F(F16C) | F(RDRAND);
 478         /* cpuid 0x80000001.ecx */
 479         const u32 kvm_cpuid_8000_0001_ecx_x86_features =
 480                 F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
 481                 F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
 482                 F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
 483                 0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
 484                 F(TOPOEXT) | F(PERFCTR_CORE);
 485
 486         /* cpuid 0x80000008.ebx */
 487         const u32 kvm_cpuid_8000_0008_ebx_x86_features =
 488                 F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
 489                 F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON);
 490
 491         /* cpuid 0xC0000001.edx */
 492         const u32 kvm_cpuid_C000_0001_edx_x86_features =
 493                 F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
 494                 F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
 495                 F(PMM) | F(PMM_EN);
 496
 497         /* cpuid 0xD.1.eax */
 498         const u32 kvm_cpuid_D_1_eax_x86_features =
 499                 F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | f_xsaves;
 500
 501         /* all calls to cpuid_count() should be made on the same cpu */
 502         get_cpu();
 503
 504         r = -E2BIG;
 505
 506         if (*nent >= maxnent)
 507                 goto out;
 508
 509         do_host_cpuid(entry, function, 0);
 510         ++*nent;
 511
 512         switch (function) {
 513         case 0:
 514                 /* Limited to the highest leaf implemented in KVM. */
 515                 entry->eax = min(entry->eax, 0x1fU);
 516                 break;
 517         case 1:
 518                 entry->edx &= kvm_cpuid_1_edx_x86_features;
 519                 cpuid_mask(&entry->edx, CPUID_1_EDX);
 520                 entry->ecx &= kvm_cpuid_1_ecx_x86_features;
 521                 cpuid_mask(&entry->ecx, CPUID_1_ECX);
 522                 /* we support x2apic emulation even if host does not support
 523                  * it since we emulate x2apic in software */
 524                 entry->ecx |= F(X2APIC);
 525                 break;
 526         /* function 2 entries are STATEFUL. That is, repeated cpuid commands
 527          * may return different values. This forces us to get_cpu() before
 528          * issuing the first command, and also to emulate this annoying behavior
 529          * in kvm_emulate_cpuid() using KVM_CPUID_FLAG_STATE_READ_NEXT */
 530         case 2: {
 531                 int t, times = entry->eax & 0xff;
 532
 533                 entry->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 534                 for (t = 1; t < times; ++t) {
 535                         if (*nent >= maxnent)
 536                                 goto out;
 537
 538                         do_host_cpuid(&entry[t], function, 0);
 539                         ++*nent;
 540                 }
 541                 break;
 542         }
 543         /* functions 4 and 0x8000001d have additional index. */
 544         case 4:
 545         case 0x8000001d: {
 546                 int i, cache_type;
 547
 548                 /* read more entries until cache_type is zero */
 549                 for (i = 1; ; ++i) {
 550                         if (*nent >= maxnent)
 551                                 goto out;
 552
 553                         cache_type = entry[i - 1].eax & 0x1f;
 554                         if (!cache_type)
 555                                 break;
 556                         do_host_cpuid(&entry[i], function, i);
 557                         ++*nent;
 558                 }
 559                 break;
 560         }
 561         case 6: /* Thermal management */
 562                 entry->eax = 0x4; /* allow ARAT */
 563                 entry->ebx = 0;
 564                 entry->ecx = 0;
 565                 entry->edx = 0;
 566                 break;
 567         /* function 7 has additional index. */
 568         case 7: {
 569                 int i;
 570
 571                 for (i = 0; ; ) {
 572                         do_cpuid_7_mask(&entry[i], i);
 573                         if (i == entry->eax)
 574                                 break;
 575                         if (*nent >= maxnent)
 576                                 goto out;
 577
 578                         ++i;
 579                         do_host_cpuid(&entry[i], function, i);
 580                         ++*nent;
 581                 }
 582                 break;
 583         }
 584         case 9:
 585                 break;
 586         case 0xa: { /* Architectural Performance Monitoring */
 587                 struct x86_pmu_capability cap;
 588                 union cpuid10_eax eax;
 589                 union cpuid10_edx edx;
 590
 591                 perf_get_x86_pmu_capability(&cap);
 592
 593                 /*
 594                  * Only support guest architectural pmu on a host
 595                  * with architectural pmu.
 596                  */
 597                 if (!cap.version)
 598                         memset(&cap, 0, sizeof(cap));
 599
 600                 eax.split.version_id = min(cap.version, 2);
 601                 eax.split.num_counters = cap.num_counters_gp;
 602                 eax.split.bit_width = cap.bit_width_gp;
 603                 eax.split.mask_length = cap.events_mask_len;
 604
 605                 edx.split.num_counters_fixed = cap.num_counters_fixed;
 606                 edx.split.bit_width_fixed = cap.bit_width_fixed;
 607                 edx.split.reserved = 0;
 608
 609                 entry->eax = eax.full;
 610                 entry->ebx = cap.events_mask;
 611                 entry->ecx = 0;
 612                 entry->edx = edx.full;
 613                 break;
 614         }
 615         /*
 616          * Per Intel's SDM, the 0x1f is a superset of 0xb,
 617          * thus they can be handled by common code.
 618          */
 619         case 0x1f:
 620         case 0xb: {
 621                 int i, level_type;
 622
 623                 /* read more entries until level_type is zero */
 624                 for (i = 1; ; ++i) {
 625                         if (*nent >= maxnent)
 626                                 goto out;
 627
 628                         level_type = entry[i - 1].ecx & 0xff00;
 629                         if (!level_type)
 630                                 break;
 631                         do_host_cpuid(&entry[i], function, i);
 632                         ++*nent;
 633                 }
 634                 break;
 635         }
 636         case 0xd: {
 637                 int idx, i;
 638                 u64 supported = kvm_supported_xcr0();
 639
 640                 entry->eax &= supported;
 641                 entry->ebx = xstate_required_size(supported, false);
 642                 entry->ecx = entry->ebx;
 643                 entry->edx &= supported >> 32;
 644                 if (!supported)
 645                         break;
 646
 647                 for (idx = 1, i = 1; idx < 64; ++idx) {
 648                         u64 mask = ((u64)1 << idx);
 649                         if (*nent >= maxnent)
 650                                 goto out;
 651
 652                         do_host_cpuid(&entry[i], function, idx);
 653                         if (idx == 1) {
 654                                 entry[i].eax &= kvm_cpuid_D_1_eax_x86_features;
 655                                 cpuid_mask(&entry[i].eax, CPUID_D_1_EAX);
 656                                 entry[i].ebx = 0;
 657                                 if (entry[i].eax & (F(XSAVES)|F(XSAVEC)))
 658                                         entry[i].ebx =
 659                                                 xstate_required_size(supported,
 660                                                                      true);
 661                         } else {
 662                                 if (entry[i].eax == 0 || !(supported & mask))
 663                                         continue;
 664                                 if (WARN_ON_ONCE(entry[i].ecx & 1))
 665                                         continue;
 666                         }
 667                         entry[i].ecx = 0;
 668                         entry[i].edx = 0;
 669                         ++*nent;
 670                         ++i;
 671                 }
 672                 break;
 673         }
 674         /* Intel PT */
 675         case 0x14: {
 676                 int t, times = entry->eax;
 677
 678                 if (!f_intel_pt)
 679                         break;
 680
 681                 for (t = 1; t <= times; ++t) {
 682                         if (*nent >= maxnent)
 683                                 goto out;
 684                         do_host_cpuid(&entry[t], function, t);
 685                         ++*nent;
 686                 }
 687                 break;
 688         }
 689         case KVM_CPUID_SIGNATURE: {
 690                 static const char signature[12] = "KVMKVMKVM\0\0";
 691                 const u32 *sigptr = (const u32 *)signature;
 692                 entry->eax = KVM_CPUID_FEATURES;
 693                 entry->ebx = sigptr[0];
 694                 entry->ecx = sigptr[1];
 695                 entry->edx = sigptr[2];
 696                 break;
 697         }
 698         case KVM_CPUID_FEATURES:
 699                 entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
 700                              (1 << KVM_FEATURE_NOP_IO_DELAY) |
 701                              (1 << KVM_FEATURE_CLOCKSOURCE2) |
 702                              (1 << KVM_FEATURE_ASYNC_PF) |
 703                              (1 << KVM_FEATURE_PV_EOI) |
 704                              (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
 705                              (1 << KVM_FEATURE_PV_UNHALT) |
 706                              (1 << KVM_FEATURE_PV_TLB_FLUSH) |
 707                              (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
 708                              (1 << KVM_FEATURE_PV_SEND_IPI) |
 709                              (1 << KVM_FEATURE_POLL_CONTROL) |
 710                              (1 << KVM_FEATURE_PV_SCHED_YIELD);
 711
 712                 if (sched_info_on())
 713                         entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
 714
 715                 entry->ebx = 0;
 716                 entry->ecx = 0;
 717                 entry->edx = 0;
 718                 break;
 719         case 0x80000000:
 720                 entry->eax = min(entry->eax, 0x8000001f);
 721                 break;
 722         case 0x80000001:
 723                 entry->edx &= kvm_cpuid_8000_0001_edx_x86_features;
 724                 cpuid_mask(&entry->edx, CPUID_8000_0001_EDX);
 725                 entry->ecx &= kvm_cpuid_8000_0001_ecx_x86_features;
 726                 cpuid_mask(&entry->ecx, CPUID_8000_0001_ECX);
 727                 break;
 728         case 0x80000007: /* Advanced power management */
 729                 /* invariant TSC is CPUID.80000007H:EDX[8] */
 730                 entry->edx &= (1 << 8);
 731                 /* mask against host */
 732                 entry->edx &= boot_cpu_data.x86_power;
 733                 entry->eax = entry->ebx = entry->ecx = 0;
 734                 break;
 735         case 0x80000008: {
 736                 unsigned g_phys_as = (entry->eax >> 16) & 0xff;
 737                 unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
 738                 unsigned phys_as = entry->eax & 0xff;
 739
 740                 if (!g_phys_as)
 741                         g_phys_as = phys_as;
 742                 entry->eax = g_phys_as | (virt_as << 8);
 743                 entry->edx = 0;
 744                 entry->ebx &= kvm_cpuid_8000_0008_ebx_x86_features;
 745                 cpuid_mask(&entry->ebx, CPUID_8000_0008_EBX);
 746                 /*
 747                  * AMD has separate bits for each SPEC_CTRL bit.
 748                  * arch/x86/kernel/cpu/bugs.c is kind enough to
 749                  * record that in cpufeatures so use them.
 750                  */
 751                 if (boot_cpu_has(X86_FEATURE_IBPB))
 752                         entry->ebx |= F(AMD_IBPB);
 753                 if (boot_cpu_has(X86_FEATURE_IBRS))
 754                         entry->ebx |= F(AMD_IBRS);
 755                 if (boot_cpu_has(X86_FEATURE_STIBP))
 756                         entry->ebx |= F(AMD_STIBP);
 757                 if (boot_cpu_has(X86_FEATURE_SSBD))
 758                         entry->ebx |= F(AMD_SSBD);
 759                 if (!boot_cpu_has_bug(X86_BUG_SPEC_STORE_BYPASS))
 760                         entry->ebx |= F(AMD_SSB_NO);
 761                 /*
 762                  * The preference is to use SPEC CTRL MSR instead of the
 763                  * VIRT_SPEC MSR.
 764                  */
 765                 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
 766                     !boot_cpu_has(X86_FEATURE_AMD_SSBD))
 767                         entry->ebx |= F(VIRT_SSBD);
 768                 break;
 769         }
 770         case 0x80000019:
 771                 entry->ecx = entry->edx = 0;
 772                 break;
 773         case 0x8000001a:
 774         case 0x8000001e:
 775                 break;
 776         /*Add support for Centaur's CPUID instruction*/
 777         case 0xC0000000:
 778                 /*Just support up to 0xC0000004 now*/
 779                 entry->eax = min(entry->eax, 0xC0000004);
 780                 break;
 781         case 0xC0000001:
 782                 entry->edx &= kvm_cpuid_C000_0001_edx_x86_features;
 783                 cpuid_mask(&entry->edx, CPUID_C000_0001_EDX);
 784                 break;
 785         case 3: /* Processor serial number */
 786         case 5: /* MONITOR/MWAIT */
 787         case 0xC0000002:
 788         case 0xC0000003:
 789         case 0xC0000004:
 790         default:
 791                 entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
 792                 break;
 793         }
 794
 795         kvm_x86_ops->set_supported_cpuid(function, entry);
 796
 797         r = 0;
 798
 799 out:
 800         put_cpu();
 801
 802         return r;
 803 }
 804
 805 static int do_cpuid_func(struct kvm_cpuid_entry2 *entry, u32 func,
 806                          int *nent, int maxnent, unsigned int type)
 807 {
 808         if (type == KVM_GET_EMULATED_CPUID)
 809                 return __do_cpuid_func_emulated(entry, func, nent, maxnent);
 810
 811         return __do_cpuid_func(entry, func, nent, maxnent);
 812 }
 813
 814 #undef F
 815
 816 struct kvm_cpuid_param {
 817         u32 func;
 818         bool (*qualifier)(const struct kvm_cpuid_param *param);
 819 };
 820
 821 static bool is_centaur_cpu(const struct kvm_cpuid_param *param)
 822 {
 823         return boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR;
 824 }
 825
 826 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
 827                                  __u32 num_entries, unsigned int ioctl_type)
 828 {
 829         int i;
 830         __u32 pad[3];
 831
 832         if (ioctl_type != KVM_GET_EMULATED_CPUID)
 833                 return false;
 834
 835         /*
 836          * We want to make sure that ->padding is being passed clean from
 837          * userspace in case we want to use it for something in the future.
 838          *
 839          * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
 840          * have to give ourselves satisfied only with the emulated side. /me
 841          * sheds a tear.
 842          */
 843         for (i = 0; i < num_entries; i++) {
 844                 if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
 845                         return true;
 846
 847                 if (pad[0] || pad[1] || pad[2])
 848                         return true;
 849         }
 850         return false;
 851 }
 852
 853 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
 854                             struct kvm_cpuid_entry2 __user *entries,
 855                             unsigned int type)
 856 {
 857         struct kvm_cpuid_entry2 *cpuid_entries;
 858         int limit, nent = 0, r = -E2BIG, i;
 859         u32 func;
 860         static const struct kvm_cpuid_param param[] = {
 861                 { .func = 0 },
 862                 { .func = 0x80000000 },
 863                 { .func = 0xC0000000, .qualifier = is_centaur_cpu },
 864                 { .func = KVM_CPUID_SIGNATURE },
 865         };
 866
 867         if (cpuid->nent < 1)
 868                 goto out;
 869         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 870                 cpuid->nent = KVM_MAX_CPUID_ENTRIES;
 871
 872         if (sanity_check_entries(entries, cpuid->nent, type))
 873                 return -EINVAL;
 874
 875         r = -ENOMEM;
 876         cpuid_entries = vzalloc(array_size(sizeof(struct kvm_cpuid_entry2),
 877                                            cpuid->nent));
 878         if (!cpuid_entries)
 879                 goto out;
 880
 881         r = 0;
 882         for (i = 0; i < ARRAY_SIZE(param); i++) {
 883                 const struct kvm_cpuid_param *ent = &param[i];
 884
 885                 if (ent->qualifier && !ent->qualifier(ent))
 886                         continue;
 887
 888                 r = do_cpuid_func(&cpuid_entries[nent], ent->func,
 889                                   &nent, cpuid->nent, type);
 890
 891                 if (r)
 892                         goto out_free;
 893
 894                 limit = cpuid_entries[nent - 1].eax;
 895                 for (func = ent->func + 1; func <= limit && nent < cpuid->nent && r == 0; ++func)
 896                         r = do_cpuid_func(&cpuid_entries[nent], func,
 897                                           &nent, cpuid->nent, type);
 898
 899                 if (r)
 900                         goto out_free;
 901         }
 902
 903         r = -EFAULT;
 904         if (copy_to_user(entries, cpuid_entries,
 905                          nent * sizeof(struct kvm_cpuid_entry2)))
 906                 goto out_free;
 907         cpuid->nent = nent;
 908         r = 0;
 909
 910 out_free:
 911         vfree(cpuid_entries);
 912 out:
 913         return r;
 914 }
 915
 916 static int move_to_next_stateful_cpuid_entry(struct kvm_vcpu *vcpu, int i)
 917 {
 918         struct kvm_cpuid_entry2 *e = &vcpu->arch.cpuid_entries[i];
 919         struct kvm_cpuid_entry2 *ej;
 920         int j = i;
 921         int nent = vcpu->arch.cpuid_nent;
 922
 923         e->flags &= ~KVM_CPUID_FLAG_STATE_READ_NEXT;
 924         /* when no next entry is found, the current entry[i] is reselected */
 925         do {
 926                 j = (j + 1) % nent;
 927                 ej = &vcpu->arch.cpuid_entries[j];
 928         } while (ej->function != e->function);
 929
 930         ej->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 931
 932         return j;
 933 }
 934
 935 /* find an entry with matching function, matching index (if needed), and that
 936  * should be read next (if it's stateful) */
 937 static int is_matching_cpuid_entry(struct kvm_cpuid_entry2 *e,
 938         u32 function, u32 index)
 939 {
 940         if (e->function != function)
 941                 return 0;
 942         if ((e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) && e->index != index)
 943                 return 0;
 944         if ((e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC) &&
 945             !(e->flags & KVM_CPUID_FLAG_STATE_READ_NEXT))
 946                 return 0;
 947         return 1;
 948 }
 949
 950 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
 951                                               u32 function, u32 index)
 952 {
 953         int i;
 954         struct kvm_cpuid_entry2 *best = NULL;
 955
 956         for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
 957                 struct kvm_cpuid_entry2 *e;
 958
 959                 e = &vcpu->arch.cpuid_entries[i];
 960                 if (is_matching_cpuid_entry(e, function, index)) {
 961                         if (e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC)
 962                                 move_to_next_stateful_cpuid_entry(vcpu, i);
 963                         best = e;
 964                         break;
 965                 }
 966         }
 967         return best;
 968 }
 969 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
 970
 971 /*
 972  * If no match is found, check whether we exceed the vCPU's limit
 973  * and return the content of the highest valid _standard_ leaf instead.
 974  * This is to satisfy the CPUID specification.
 975  */
 976 static struct kvm_cpuid_entry2* check_cpuid_limit(struct kvm_vcpu *vcpu,
 977                                                   u32 function, u32 index)
 978 {
 979         struct kvm_cpuid_entry2 *maxlevel;
 980
 981         maxlevel = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
 982         if (!maxlevel || maxlevel->eax >= function)
 983                 return NULL;
 984         if (function & 0x80000000) {
 985                 maxlevel = kvm_find_cpuid_entry(vcpu, 0, 0);
 986                 if (!maxlevel)
 987                         return NULL;
 988         }
 989         return kvm_find_cpuid_entry(vcpu, maxlevel->eax, index);
 990 }
 991
 992 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
 993                u32 *ecx, u32 *edx, bool check_limit)
 994 {
 995         u32 function = *eax, index = *ecx;
 996         struct kvm_cpuid_entry2 *best;
 997         bool entry_found = true;
 998
 999         best = kvm_find_cpuid_entry(vcpu, function, index);
1000
1001         if (!best) {
1002                 entry_found = false;
1003                 if (!check_limit)
1004                         goto out;
1005
1006                 best = check_cpuid_limit(vcpu, function, index);
1007         }
1008
1009 out:
1010         if (best) {
1011                 *eax = best->eax;
1012                 *ebx = best->ebx;
1013                 *ecx = best->ecx;
1014                 *edx = best->edx;
1015         } else
1016                 *eax = *ebx = *ecx = *edx = 0;
1017         trace_kvm_cpuid(function, *eax, *ebx, *ecx, *edx, entry_found);
1018         return entry_found;
1019 }
1020 EXPORT_SYMBOL_GPL(kvm_cpuid);
1021
1022 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1023 {
1024         u32 eax, ebx, ecx, edx;
1025
1026         if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1027                 return 1;
1028
1029         eax = kvm_rax_read(vcpu);
1030         ecx = kvm_rcx_read(vcpu);
1031         kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, true);
1032         kvm_rax_write(vcpu, eax);
1033         kvm_rbx_write(vcpu, ebx);
1034         kvm_rcx_write(vcpu, ecx);
1035         kvm_rdx_write(vcpu, edx);
1036         return kvm_skip_emulated_instruction(vcpu);
1037 }
1038 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);