arch/x86/kvm/cpuid.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  * cpuid support routines
   4  *
   5  * derived from arch/x86/kvm/x86.c
   6  *
   7  * Copyright 2011 Red Hat, Inc. and/or its affiliates.
   8  * Copyright IBM Corporation, 2008
   9  *
  10  * This work is licensed under the terms of the GNU GPL, version 2.  See
  11  * the COPYING file in the top-level directory.
  12  *
  13  */
  14
  15 #include <linux/kvm_host.h>
  16 #include <linux/export.h>
  17 #include <linux/vmalloc.h>
  18 #include <linux/uaccess.h>
  19 #include <linux/sched/stat.h>
  20
  21 #include <asm/processor.h>
  22 #include <asm/user.h>
  23 #include <asm/fpu/xstate.h>
  24 #include "cpuid.h"
  25 #include "lapic.h"
  26 #include "mmu.h"
  27 #include "trace.h"
  28 #include "pmu.h"
  29
  30 static u32 xstate_required_size(u64 xstate_bv, bool compacted)
  31 {
  32         int feature_bit = 0;
  33         u32 ret = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
  34
  35         xstate_bv &= XFEATURE_MASK_EXTEND;
  36         while (xstate_bv) {
  37                 if (xstate_bv & 0x1) {
  38                         u32 eax, ebx, ecx, edx, offset;
  39                         cpuid_count(0xD, feature_bit, &eax, &ebx, &ecx, &edx);
  40                         offset = compacted ? ret : ebx;
  41                         ret = max(ret, offset + eax);
  42                 }
  43
  44                 xstate_bv >>= 1;
  45                 feature_bit++;
  46         }
  47
  48         return ret;
  49 }
  50
  51 bool kvm_mpx_supported(void)
  52 {
  53         return ((host_xcr0 & (XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR))
  54                  && kvm_x86_ops->mpx_supported());
  55 }
  56 EXPORT_SYMBOL_GPL(kvm_mpx_supported);
  57
  58 u64 kvm_supported_xcr0(void)
  59 {
  60         u64 xcr0 = KVM_SUPPORTED_XCR0 & host_xcr0;
  61
  62         if (!kvm_mpx_supported())
  63                 xcr0 &= ~(XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
  64
  65         return xcr0;
  66 }
  67
  68 #define F(x) bit(X86_FEATURE_##x)
  69
  70 /* For scattered features from cpufeatures.h; we currently expose none */
  71 #define KF(x) bit(KVM_CPUID_BIT_##x)
  72
  73 int kvm_update_cpuid(struct kvm_vcpu *vcpu)
  74 {
  75         struct kvm_cpuid_entry2 *best;
  76         struct kvm_lapic *apic = vcpu->arch.apic;
  77
  78         best = kvm_find_cpuid_entry(vcpu, 1, 0);
  79         if (!best)
  80                 return 0;
  81
  82         /* Update OSXSAVE bit */
  83         if (boot_cpu_has(X86_FEATURE_XSAVE) && best->function == 0x1) {
  84                 best->ecx &= ~F(OSXSAVE);
  85                 if (kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE))
  86                         best->ecx |= F(OSXSAVE);
  87         }
  88
  89         best->edx &= ~F(APIC);
  90         if (vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE)
  91                 best->edx |= F(APIC);
  92
  93         if (apic) {
  94                 if (best->ecx & F(TSC_DEADLINE_TIMER))
  95                         apic->lapic_timer.timer_mode_mask = 3 << 17;
  96                 else
  97                         apic->lapic_timer.timer_mode_mask = 1 << 17;
  98         }
  99
 100         best = kvm_find_cpuid_entry(vcpu, 7, 0);
 101         if (best) {
 102                 /* Update OSPKE bit */
 103                 if (boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7) {
 104                         best->ecx &= ~F(OSPKE);
 105                         if (kvm_read_cr4_bits(vcpu, X86_CR4_PKE))
 106                                 best->ecx |= F(OSPKE);
 107                 }
 108         }
 109
 110         best = kvm_find_cpuid_entry(vcpu, 0xD, 0);
 111         if (!best) {
 112                 vcpu->arch.guest_supported_xcr0 = 0;
 113                 vcpu->arch.guest_xstate_size = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
 114         } else {
 115                 vcpu->arch.guest_supported_xcr0 =
 116                         (best->eax | ((u64)best->edx << 32)) &
 117                         kvm_supported_xcr0();
 118                 vcpu->arch.guest_xstate_size = best->ebx =
 119                         xstate_required_size(vcpu->arch.xcr0, false);
 120         }
 121
 122         best = kvm_find_cpuid_entry(vcpu, 0xD, 1);
 123         if (best && (best->eax & (F(XSAVES) | F(XSAVEC))))
 124                 best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
 125
 126         /*
 127          * The existing code assumes virtual address is 48-bit or 57-bit in the
 128          * canonical address checks; exit if it is ever changed.
 129          */
 130         best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
 131         if (best) {
 132                 int vaddr_bits = (best->eax & 0xff00) >> 8;
 133
 134                 if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
 135                         return -EINVAL;
 136         }
 137
 138         /* Update physical-address width */
 139         vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
 140         kvm_mmu_reset_context(vcpu);
 141
 142         kvm_pmu_refresh(vcpu);
 143         return 0;
 144 }
 145
 146 static int is_efer_nx(void)
 147 {
 148         unsigned long long efer = 0;
 149
 150         rdmsrl_safe(MSR_EFER, &efer);
 151         return efer & EFER_NX;
 152 }
 153
 154 static void cpuid_fix_nx_cap(struct kvm_vcpu *vcpu)
 155 {
 156         int i;
 157         struct kvm_cpuid_entry2 *e, *entry;
 158
 159         entry = NULL;
 160         for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
 161                 e = &vcpu->arch.cpuid_entries[i];
 162                 if (e->function == 0x80000001) {
 163                         entry = e;
 164                         break;
 165                 }
 166         }
 167         if (entry && (entry->edx & F(NX)) && !is_efer_nx()) {
 168                 entry->edx &= ~F(NX);
 169                 printk(KERN_INFO "kvm: guest NX capability removed\n");
 170         }
 171 }
 172
 173 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
 174 {
 175         struct kvm_cpuid_entry2 *best;
 176
 177         best = kvm_find_cpuid_entry(vcpu, 0x80000000, 0);
 178         if (!best || best->eax < 0x80000008)
 179                 goto not_found;
 180         best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
 181         if (best)
 182                 return best->eax & 0xff;
 183 not_found:
 184         return 36;
 185 }
 186 EXPORT_SYMBOL_GPL(cpuid_query_maxphyaddr);
 187
 188 /* when an old userspace process fills a new kernel module */
 189 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
 190                              struct kvm_cpuid *cpuid,
 191                              struct kvm_cpuid_entry __user *entries)
 192 {
 193         int r, i;
 194         struct kvm_cpuid_entry *cpuid_entries = NULL;
 195
 196         r = -E2BIG;
 197         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 198                 goto out;
 199         r = -ENOMEM;
 200         if (cpuid->nent) {
 201                 cpuid_entries = vmalloc(sizeof(struct kvm_cpuid_entry) *
 202                                         cpuid->nent);
 203                 if (!cpuid_entries)
 204                         goto out;
 205                 r = -EFAULT;
 206                 if (copy_from_user(cpuid_entries, entries,
 207                                    cpuid->nent * sizeof(struct kvm_cpuid_entry)))
 208                         goto out;
 209         }
 210         for (i = 0; i < cpuid->nent; i++) {
 211                 vcpu->arch.cpuid_entries[i].function = cpuid_entries[i].function;
 212                 vcpu->arch.cpuid_entries[i].eax = cpuid_entries[i].eax;
 213                 vcpu->arch.cpuid_entries[i].ebx = cpuid_entries[i].ebx;
 214                 vcpu->arch.cpuid_entries[i].ecx = cpuid_entries[i].ecx;
 215                 vcpu->arch.cpuid_entries[i].edx = cpuid_entries[i].edx;
 216                 vcpu->arch.cpuid_entries[i].index = 0;
 217                 vcpu->arch.cpuid_entries[i].flags = 0;
 218                 vcpu->arch.cpuid_entries[i].padding[0] = 0;
 219                 vcpu->arch.cpuid_entries[i].padding[1] = 0;
 220                 vcpu->arch.cpuid_entries[i].padding[2] = 0;
 221         }
 222         vcpu->arch.cpuid_nent = cpuid->nent;
 223         cpuid_fix_nx_cap(vcpu);
 224         kvm_apic_set_version(vcpu);
 225         kvm_x86_ops->cpuid_update(vcpu);
 226         r = kvm_update_cpuid(vcpu);
 227
 228 out:
 229         vfree(cpuid_entries);
 230         return r;
 231 }
 232
 233 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
 234                               struct kvm_cpuid2 *cpuid,
 235                               struct kvm_cpuid_entry2 __user *entries)
 236 {
 237         int r;
 238
 239         r = -E2BIG;
 240         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 241                 goto out;
 242         r = -EFAULT;
 243         if (copy_from_user(&vcpu->arch.cpuid_entries, entries,
 244                            cpuid->nent * sizeof(struct kvm_cpuid_entry2)))
 245                 goto out;
 246         vcpu->arch.cpuid_nent = cpuid->nent;
 247         kvm_apic_set_version(vcpu);
 248         kvm_x86_ops->cpuid_update(vcpu);
 249         r = kvm_update_cpuid(vcpu);
 250 out:
 251         return r;
 252 }
 253
 254 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
 255                               struct kvm_cpuid2 *cpuid,
 256                               struct kvm_cpuid_entry2 __user *entries)
 257 {
 258         int r;
 259
 260         r = -E2BIG;
 261         if (cpuid->nent < vcpu->arch.cpuid_nent)
 262                 goto out;
 263         r = -EFAULT;
 264         if (copy_to_user(entries, &vcpu->arch.cpuid_entries,
 265                          vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
 266                 goto out;
 267         return 0;
 268
 269 out:
 270         cpuid->nent = vcpu->arch.cpuid_nent;
 271         return r;
 272 }
 273
 274 static void cpuid_mask(u32 *word, int wordnum)
 275 {
 276         *word &= boot_cpu_data.x86_capability[wordnum];
 277 }
 278
 279 static void do_cpuid_1_ent(struct kvm_cpuid_entry2 *entry, u32 function,
 280                            u32 index)
 281 {
 282         entry->function = function;
 283         entry->index = index;
 284         cpuid_count(entry->function, entry->index,
 285                     &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
 286         entry->flags = 0;
 287 }
 288
 289 static int __do_cpuid_ent_emulated(struct kvm_cpuid_entry2 *entry,
 290                                    u32 func, u32 index, int *nent, int maxnent)
 291 {
 292         switch (func) {
 293         case 0:
 294                 entry->eax = 7;
 295                 ++*nent;
 296                 break;
 297         case 1:
 298                 entry->ecx = F(MOVBE);
 299                 ++*nent;
 300                 break;
 301         case 7:
 302                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 303                 if (index == 0)
 304                         entry->ecx = F(RDPID);
 305                 ++*nent;
 306         default:
 307                 break;
 308         }
 309
 310         entry->function = func;
 311         entry->index = index;
 312
 313         return 0;
 314 }
 315
 316 static inline int __do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function,
 317                                  u32 index, int *nent, int maxnent)
 318 {
 319         int r;
 320         unsigned f_nx = is_efer_nx() ? F(NX) : 0;
 321 #ifdef CONFIG_X86_64
 322         unsigned f_gbpages = (kvm_x86_ops->get_lpage_level() == PT_PDPE_LEVEL)
 323                                 ? F(GBPAGES) : 0;
 324         unsigned f_lm = F(LM);
 325 #else
 326         unsigned f_gbpages = 0;
 327         unsigned f_lm = 0;
 328 #endif
 329         unsigned f_rdtscp = kvm_x86_ops->rdtscp_supported() ? F(RDTSCP) : 0;
 330         unsigned f_invpcid = kvm_x86_ops->invpcid_supported() ? F(INVPCID) : 0;
 331         unsigned f_mpx = kvm_mpx_supported() ? F(MPX) : 0;
 332         unsigned f_xsaves = kvm_x86_ops->xsaves_supported() ? F(XSAVES) : 0;
 333
 334         /* cpuid 1.edx */
 335         const u32 kvm_cpuid_1_edx_x86_features =
 336                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 337                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 338                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
 339                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 340                 F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
 341                 0 /* Reserved, DS, ACPI */ | F(MMX) |
 342                 F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
 343                 0 /* HTT, TM, Reserved, PBE */;
 344         /* cpuid 0x80000001.edx */
 345         const u32 kvm_cpuid_8000_0001_edx_x86_features =
 346                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 347                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 348                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
 349                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 350                 F(PAT) | F(PSE36) | 0 /* Reserved */ |
 351                 f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
 352                 F(FXSR) | F(FXSR_OPT) | f_gbpages | f_rdtscp |
 353                 0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW);
 354         /* cpuid 1.ecx */
 355         const u32 kvm_cpuid_1_ecx_x86_features =
 356                 /* NOTE: MONITOR (and MWAIT) are emulated as NOP,
 357                  * but *not* advertised to guests via CPUID ! */
 358                 F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
 359                 0 /* DS-CPL, VMX, SMX, EST */ |
 360                 0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
 361                 F(FMA) | F(CX16) | 0 /* xTPR Update, PDCM */ |
 362                 F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
 363                 F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
 364                 0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
 365                 F(F16C) | F(RDRAND);
 366         /* cpuid 0x80000001.ecx */
 367         const u32 kvm_cpuid_8000_0001_ecx_x86_features =
 368                 F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
 369                 F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
 370                 F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
 371                 0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM);
 372
 373         /* cpuid 0x80000008.ebx */
 374         const u32 kvm_cpuid_8000_0008_ebx_x86_features =
 375                 F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
 376                 F(AMD_SSB_NO) | F(AMD_STIBP);
 377
 378         /* cpuid 0xC0000001.edx */
 379         const u32 kvm_cpuid_C000_0001_edx_x86_features =
 380                 F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
 381                 F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
 382                 F(PMM) | F(PMM_EN);
 383
 384         /* cpuid 7.0.ebx */
 385         const u32 kvm_cpuid_7_0_ebx_x86_features =
 386                 F(FSGSBASE) | F(BMI1) | F(HLE) | F(AVX2) | F(SMEP) |
 387                 F(BMI2) | F(ERMS) | f_invpcid | F(RTM) | f_mpx | F(RDSEED) |
 388                 F(ADX) | F(SMAP) | F(AVX512IFMA) | F(AVX512F) | F(AVX512PF) |
 389                 F(AVX512ER) | F(AVX512CD) | F(CLFLUSHOPT) | F(CLWB) | F(AVX512DQ) |
 390                 F(SHA_NI) | F(AVX512BW) | F(AVX512VL);
 391
 392         /* cpuid 0xD.1.eax */
 393         const u32 kvm_cpuid_D_1_eax_x86_features =
 394                 F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | f_xsaves;
 395
 396         /* cpuid 7.0.ecx*/
 397         const u32 kvm_cpuid_7_0_ecx_x86_features =
 398                 F(AVX512VBMI) | F(LA57) | F(PKU) |
 399                 0 /*OSPKE*/ | F(AVX512_VPOPCNTDQ);
 400
 401         /* cpuid 7.0.edx*/
 402         const u32 kvm_cpuid_7_0_edx_x86_features =
 403                 F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
 404                 F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
 405                 F(MD_CLEAR);
 406
 407         /* all calls to cpuid_count() should be made on the same cpu */
 408         get_cpu();
 409
 410         r = -E2BIG;
 411
 412         if (WARN_ON(*nent >= maxnent))
 413                 goto out;
 414
 415         do_cpuid_1_ent(entry, function, index);
 416         ++*nent;
 417
 418         switch (function) {
 419         case 0:
 420                 entry->eax = min(entry->eax, (u32)0xd);
 421                 break;
 422         case 1:
 423                 entry->edx &= kvm_cpuid_1_edx_x86_features;
 424                 cpuid_mask(&entry->edx, CPUID_1_EDX);
 425                 entry->ecx &= kvm_cpuid_1_ecx_x86_features;
 426                 cpuid_mask(&entry->ecx, CPUID_1_ECX);
 427                 /* we support x2apic emulation even if host does not support
 428                  * it since we emulate x2apic in software */
 429                 entry->ecx |= F(X2APIC);
 430                 break;
 431         /* function 2 entries are STATEFUL. That is, repeated cpuid commands
 432          * may return different values. This forces us to get_cpu() before
 433          * issuing the first command, and also to emulate this annoying behavior
 434          * in kvm_emulate_cpuid() using KVM_CPUID_FLAG_STATE_READ_NEXT */
 435         case 2: {
 436                 int t, times = entry->eax & 0xff;
 437
 438                 entry->flags |= KVM_CPUID_FLAG_STATEFUL_FUNC;
 439                 entry->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 440                 for (t = 1; t < times; ++t) {
 441                         if (*nent >= maxnent)
 442                                 goto out;
 443
 444                         do_cpuid_1_ent(&entry[t], function, 0);
 445                         entry[t].flags |= KVM_CPUID_FLAG_STATEFUL_FUNC;
 446                         ++*nent;
 447                 }
 448                 break;
 449         }
 450         /* function 4 has additional index. */
 451         case 4: {
 452                 int i, cache_type;
 453
 454                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 455                 /* read more entries until cache_type is zero */
 456                 for (i = 1; ; ++i) {
 457                         if (*nent >= maxnent)
 458                                 goto out;
 459
 460                         cache_type = entry[i - 1].eax & 0x1f;
 461                         if (!cache_type)
 462                                 break;
 463                         do_cpuid_1_ent(&entry[i], function, i);
 464                         entry[i].flags |=
 465                                KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 466                         ++*nent;
 467                 }
 468                 break;
 469         }
 470         case 6: /* Thermal management */
 471                 entry->eax = 0x4; /* allow ARAT */
 472                 entry->ebx = 0;
 473                 entry->ecx = 0;
 474                 entry->edx = 0;
 475                 break;
 476         case 7: {
 477                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 478                 /* Mask ebx against host capability word 9 */
 479                 if (index == 0) {
 480                         entry->ebx &= kvm_cpuid_7_0_ebx_x86_features;
 481                         cpuid_mask(&entry->ebx, CPUID_7_0_EBX);
 482                         // TSC_ADJUST is emulated
 483                         entry->ebx |= F(TSC_ADJUST);
 484                         entry->ecx &= kvm_cpuid_7_0_ecx_x86_features;
 485                         cpuid_mask(&entry->ecx, CPUID_7_ECX);
 486                         /* PKU is not yet implemented for shadow paging. */
 487                         if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
 488                                 entry->ecx &= ~F(PKU);
 489
 490                         entry->edx &= kvm_cpuid_7_0_edx_x86_features;
 491                         cpuid_mask(&entry->edx, CPUID_7_EDX);
 492                         if (boot_cpu_has(X86_FEATURE_IBPB) &&
 493                             boot_cpu_has(X86_FEATURE_IBRS))
 494                                 entry->edx |= F(SPEC_CTRL);
 495                         if (boot_cpu_has(X86_FEATURE_STIBP))
 496                                 entry->edx |= F(INTEL_STIBP);
 497                         if (boot_cpu_has(X86_FEATURE_SPEC_CTRL_SSBD) ||
 498                             boot_cpu_has(X86_FEATURE_AMD_SSBD))
 499                                 entry->edx |= F(SPEC_CTRL_SSBD);
 500                         /*
 501                          * We emulate ARCH_CAPABILITIES in software even
 502                          * if the host doesn't support it.
 503                          */
 504                         entry->edx |= F(ARCH_CAPABILITIES);
 505                 } else {
 506                         entry->ebx = 0;
 507                         entry->ecx = 0;
 508                         entry->edx = 0;
 509                 }
 510                 entry->eax = 0;
 511                 break;
 512         }
 513         case 9:
 514                 break;
 515         case 0xa: { /* Architectural Performance Monitoring */
 516                 struct x86_pmu_capability cap;
 517                 union cpuid10_eax eax;
 518                 union cpuid10_edx edx;
 519
 520                 perf_get_x86_pmu_capability(&cap);
 521
 522                 /*
 523                  * Only support guest architectural pmu on a host
 524                  * with architectural pmu.
 525                  */
 526                 if (!cap.version)
 527                         memset(&cap, 0, sizeof(cap));
 528
 529                 eax.split.version_id = min(cap.version, 2);
 530                 eax.split.num_counters = cap.num_counters_gp;
 531                 eax.split.bit_width = cap.bit_width_gp;
 532                 eax.split.mask_length = cap.events_mask_len;
 533
 534                 edx.split.num_counters_fixed = cap.num_counters_fixed;
 535                 edx.split.bit_width_fixed = cap.bit_width_fixed;
 536                 edx.split.reserved = 0;
 537
 538                 entry->eax = eax.full;
 539                 entry->ebx = cap.events_mask;
 540                 entry->ecx = 0;
 541                 entry->edx = edx.full;
 542                 break;
 543         }
 544         /* function 0xb has additional index. */
 545         case 0xb: {
 546                 int i, level_type;
 547
 548                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 549                 /* read more entries until level_type is zero */
 550                 for (i = 1; ; ++i) {
 551                         if (*nent >= maxnent)
 552                                 goto out;
 553
 554                         level_type = entry[i - 1].ecx & 0xff00;
 555                         if (!level_type)
 556                                 break;
 557                         do_cpuid_1_ent(&entry[i], function, i);
 558                         entry[i].flags |=
 559                                KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 560                         ++*nent;
 561                 }
 562                 break;
 563         }
 564         case 0xd: {
 565                 int idx, i;
 566                 u64 supported = kvm_supported_xcr0();
 567
 568                 entry->eax &= supported;
 569                 entry->ebx = xstate_required_size(supported, false);
 570                 entry->ecx = entry->ebx;
 571                 entry->edx &= supported >> 32;
 572                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 573                 if (!supported)
 574                         break;
 575
 576                 for (idx = 1, i = 1; idx < 64; ++idx) {
 577                         u64 mask = ((u64)1 << idx);
 578                         if (*nent >= maxnent)
 579                                 goto out;
 580
 581                         do_cpuid_1_ent(&entry[i], function, idx);
 582                         if (idx == 1) {
 583                                 entry[i].eax &= kvm_cpuid_D_1_eax_x86_features;
 584                                 cpuid_mask(&entry[i].eax, CPUID_D_1_EAX);
 585                                 entry[i].ebx = 0;
 586                                 if (entry[i].eax & (F(XSAVES)|F(XSAVEC)))
 587                                         entry[i].ebx =
 588                                                 xstate_required_size(supported,
 589                                                                      true);
 590                         } else {
 591                                 if (entry[i].eax == 0 || !(supported & mask))
 592                                         continue;
 593                                 if (WARN_ON_ONCE(entry[i].ecx & 1))
 594                                         continue;
 595                         }
 596                         entry[i].ecx = 0;
 597                         entry[i].edx = 0;
 598                         entry[i].flags |=
 599                                KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 600                         ++*nent;
 601                         ++i;
 602                 }
 603                 break;
 604         }
 605         case KVM_CPUID_SIGNATURE: {
 606                 static const char signature[12] = "KVMKVMKVM\0\0";
 607                 const u32 *sigptr = (const u32 *)signature;
 608                 entry->eax = KVM_CPUID_FEATURES;
 609                 entry->ebx = sigptr[0];
 610                 entry->ecx = sigptr[1];
 611                 entry->edx = sigptr[2];
 612                 break;
 613         }
 614         case KVM_CPUID_FEATURES:
 615                 entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
 616                              (1 << KVM_FEATURE_NOP_IO_DELAY) |
 617                              (1 << KVM_FEATURE_CLOCKSOURCE2) |
 618                              (1 << KVM_FEATURE_ASYNC_PF) |
 619                              (1 << KVM_FEATURE_PV_EOI) |
 620                              (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
 621                              (1 << KVM_FEATURE_PV_UNHALT) |
 622                              (1 << KVM_FEATURE_ASYNC_PF_VMEXIT);
 623
 624                 if (sched_info_on())
 625                         entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
 626
 627                 entry->ebx = 0;
 628                 entry->ecx = 0;
 629                 entry->edx = 0;
 630                 break;
 631         case 0x80000000:
 632                 entry->eax = min(entry->eax, 0x8000001a);
 633                 break;
 634         case 0x80000001:
 635                 entry->edx &= kvm_cpuid_8000_0001_edx_x86_features;
 636                 cpuid_mask(&entry->edx, CPUID_8000_0001_EDX);
 637                 entry->ecx &= kvm_cpuid_8000_0001_ecx_x86_features;
 638                 cpuid_mask(&entry->ecx, CPUID_8000_0001_ECX);
 639                 break;
 640         case 0x80000007: /* Advanced power management */
 641                 /* invariant TSC is CPUID.80000007H:EDX[8] */
 642                 entry->edx &= (1 << 8);
 643                 /* mask against host */
 644                 entry->edx &= boot_cpu_data.x86_power;
 645                 entry->eax = entry->ebx = entry->ecx = 0;
 646                 break;
 647         case 0x80000008: {
 648                 unsigned g_phys_as = (entry->eax >> 16) & 0xff;
 649                 unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
 650                 unsigned phys_as = entry->eax & 0xff;
 651
 652                 /*
 653                  * Use bare metal's MAXPHADDR if the CPU doesn't report guest
 654                  * MAXPHYADDR separately, or if TDP (NPT) is disabled, as the
 655                  * guest version "applies only to guests using nested paging".
 656                  */
 657                 if (!g_phys_as || !tdp_enabled)
 658                         g_phys_as = phys_as;
 659
 660                 entry->eax = g_phys_as | (virt_as << 8);
 661                 entry->edx = 0;
 662                 /*
 663                  * IBRS, IBPB and VIRT_SSBD aren't necessarily present in
 664                  * hardware cpuid
 665                  */
 666                 if (boot_cpu_has(X86_FEATURE_AMD_IBPB))
 667                         entry->ebx |= F(AMD_IBPB);
 668                 if (boot_cpu_has(X86_FEATURE_AMD_IBRS))
 669                         entry->ebx |= F(AMD_IBRS);
 670                 if (boot_cpu_has(X86_FEATURE_VIRT_SSBD))
 671                         entry->ebx |= F(VIRT_SSBD);
 672                 entry->ebx &= kvm_cpuid_8000_0008_ebx_x86_features;
 673                 cpuid_mask(&entry->ebx, CPUID_8000_0008_EBX);
 674                 /*
 675                  * The preference is to use SPEC CTRL MSR instead of the
 676                  * VIRT_SPEC MSR.
 677                  */
 678                 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
 679                     !boot_cpu_has(X86_FEATURE_AMD_SSBD))
 680                         entry->ebx |= F(VIRT_SSBD);
 681                 break;
 682         }
 683         case 0x80000019:
 684                 entry->ecx = entry->edx = 0;
 685                 break;
 686         case 0x8000001a:
 687                 break;
 688         case 0x8000001d:
 689                 break;
 690         /*Add support for Centaur's CPUID instruction*/
 691         case 0xC0000000:
 692                 /*Just support up to 0xC0000004 now*/
 693                 entry->eax = min(entry->eax, 0xC0000004);
 694                 break;
 695         case 0xC0000001:
 696                 entry->edx &= kvm_cpuid_C000_0001_edx_x86_features;
 697                 cpuid_mask(&entry->edx, CPUID_C000_0001_EDX);
 698                 break;
 699         case 3: /* Processor serial number */
 700         case 5: /* MONITOR/MWAIT */
 701         case 0xC0000002:
 702         case 0xC0000003:
 703         case 0xC0000004:
 704         default:
 705                 entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
 706                 break;
 707         }
 708
 709         kvm_x86_ops->set_supported_cpuid(function, entry);
 710
 711         r = 0;
 712
 713 out:
 714         put_cpu();
 715
 716         return r;
 717 }
 718
 719 static int do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 func,
 720                         u32 idx, int *nent, int maxnent, unsigned int type)
 721 {
 722         if (*nent >= maxnent)
 723                 return -E2BIG;
 724
 725         if (type == KVM_GET_EMULATED_CPUID)
 726                 return __do_cpuid_ent_emulated(entry, func, idx, nent, maxnent);
 727
 728         return __do_cpuid_ent(entry, func, idx, nent, maxnent);
 729 }
 730
 731 #undef F
 732
 733 struct kvm_cpuid_param {
 734         u32 func;
 735         u32 idx;
 736         bool has_leaf_count;
 737         bool (*qualifier)(const struct kvm_cpuid_param *param);
 738 };
 739
 740 static bool is_centaur_cpu(const struct kvm_cpuid_param *param)
 741 {
 742         return boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR;
 743 }
 744
 745 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
 746                                  __u32 num_entries, unsigned int ioctl_type)
 747 {
 748         int i;
 749         __u32 pad[3];
 750
 751         if (ioctl_type != KVM_GET_EMULATED_CPUID)
 752                 return false;
 753
 754         /*
 755          * We want to make sure that ->padding is being passed clean from
 756          * userspace in case we want to use it for something in the future.
 757          *
 758          * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
 759          * have to give ourselves satisfied only with the emulated side. /me
 760          * sheds a tear.
 761          */
 762         for (i = 0; i < num_entries; i++) {
 763                 if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
 764                         return true;
 765
 766                 if (pad[0] || pad[1] || pad[2])
 767                         return true;
 768         }
 769         return false;
 770 }
 771
 772 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
 773                             struct kvm_cpuid_entry2 __user *entries,
 774                             unsigned int type)
 775 {
 776         struct kvm_cpuid_entry2 *cpuid_entries;
 777         int limit, nent = 0, r = -E2BIG, i;
 778         u32 func;
 779         static const struct kvm_cpuid_param param[] = {
 780                 { .func = 0, .has_leaf_count = true },
 781                 { .func = 0x80000000, .has_leaf_count = true },
 782                 { .func = 0xC0000000, .qualifier = is_centaur_cpu, .has_leaf_count = true },
 783                 { .func = KVM_CPUID_SIGNATURE },
 784                 { .func = KVM_CPUID_FEATURES },
 785         };
 786
 787         if (cpuid->nent < 1)
 788                 goto out;
 789         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 790                 cpuid->nent = KVM_MAX_CPUID_ENTRIES;
 791
 792         if (sanity_check_entries(entries, cpuid->nent, type))
 793                 return -EINVAL;
 794
 795         r = -ENOMEM;
 796         cpuid_entries = vzalloc(sizeof(struct kvm_cpuid_entry2) * cpuid->nent);
 797         if (!cpuid_entries)
 798                 goto out;
 799
 800         r = 0;
 801         for (i = 0; i < ARRAY_SIZE(param); i++) {
 802                 const struct kvm_cpuid_param *ent = &param[i];
 803
 804                 if (ent->qualifier && !ent->qualifier(ent))
 805                         continue;
 806
 807                 r = do_cpuid_ent(&cpuid_entries[nent], ent->func, ent->idx,
 808                                 &nent, cpuid->nent, type);
 809
 810                 if (r)
 811                         goto out_free;
 812
 813                 if (!ent->has_leaf_count)
 814                         continue;
 815
 816                 limit = cpuid_entries[nent - 1].eax;
 817                 for (func = ent->func + 1; func <= limit && nent < cpuid->nent && r == 0; ++func)
 818                         r = do_cpuid_ent(&cpuid_entries[nent], func, ent->idx,
 819                                      &nent, cpuid->nent, type);
 820
 821                 if (r)
 822                         goto out_free;
 823         }
 824
 825         r = -EFAULT;
 826         if (copy_to_user(entries, cpuid_entries,
 827                          nent * sizeof(struct kvm_cpuid_entry2)))
 828                 goto out_free;
 829         cpuid->nent = nent;
 830         r = 0;
 831
 832 out_free:
 833         vfree(cpuid_entries);
 834 out:
 835         return r;
 836 }
 837
 838 static int move_to_next_stateful_cpuid_entry(struct kvm_vcpu *vcpu, int i)
 839 {
 840         struct kvm_cpuid_entry2 *e = &vcpu->arch.cpuid_entries[i];
 841         struct kvm_cpuid_entry2 *ej;
 842         int j = i;
 843         int nent = vcpu->arch.cpuid_nent;
 844
 845         e->flags &= ~KVM_CPUID_FLAG_STATE_READ_NEXT;
 846         /* when no next entry is found, the current entry[i] is reselected */
 847         do {
 848                 j = (j + 1) % nent;
 849                 ej = &vcpu->arch.cpuid_entries[j];
 850         } while (ej->function != e->function);
 851
 852         ej->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 853
 854         return j;
 855 }
 856
 857 /* find an entry with matching function, matching index (if needed), and that
 858  * should be read next (if it's stateful) */
 859 static int is_matching_cpuid_entry(struct kvm_cpuid_entry2 *e,
 860         u32 function, u32 index)
 861 {
 862         if (e->function != function)
 863                 return 0;
 864         if ((e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) && e->index != index)
 865                 return 0;
 866         if ((e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC) &&
 867             !(e->flags & KVM_CPUID_FLAG_STATE_READ_NEXT))
 868                 return 0;
 869         return 1;
 870 }
 871
 872 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
 873                                               u32 function, u32 index)
 874 {
 875         int i;
 876         struct kvm_cpuid_entry2 *best = NULL;
 877
 878         for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
 879                 struct kvm_cpuid_entry2 *e;
 880
 881                 e = &vcpu->arch.cpuid_entries[i];
 882                 if (is_matching_cpuid_entry(e, function, index)) {
 883                         if (e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC)
 884                                 move_to_next_stateful_cpuid_entry(vcpu, i);
 885                         best = e;
 886                         break;
 887                 }
 888         }
 889         return best;
 890 }
 891 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
 892
 893 /*
 894  * If no match is found, check whether we exceed the vCPU's limit
 895  * and return the content of the highest valid _standard_ leaf instead.
 896  * This is to satisfy the CPUID specification.
 897  */
 898 static struct kvm_cpuid_entry2* check_cpuid_limit(struct kvm_vcpu *vcpu,
 899                                                   u32 function, u32 index)
 900 {
 901         struct kvm_cpuid_entry2 *maxlevel;
 902
 903         maxlevel = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
 904         if (!maxlevel || maxlevel->eax >= function)
 905                 return NULL;
 906         if (function & 0x80000000) {
 907                 maxlevel = kvm_find_cpuid_entry(vcpu, 0, 0);
 908                 if (!maxlevel)
 909                         return NULL;
 910         }
 911         return kvm_find_cpuid_entry(vcpu, maxlevel->eax, index);
 912 }
 913
 914 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
 915                u32 *ecx, u32 *edx, bool check_limit)
 916 {
 917         u32 function = *eax, index = *ecx;
 918         struct kvm_cpuid_entry2 *best;
 919         bool entry_found = true;
 920
 921         best = kvm_find_cpuid_entry(vcpu, function, index);
 922
 923         if (!best) {
 924                 entry_found = false;
 925                 if (!check_limit)
 926                         goto out;
 927
 928                 best = check_cpuid_limit(vcpu, function, index);
 929         }
 930
 931 out:
 932         if (best) {
 933                 *eax = best->eax;
 934                 *ebx = best->ebx;
 935                 *ecx = best->ecx;
 936                 *edx = best->edx;
 937         } else
 938                 *eax = *ebx = *ecx = *edx = 0;
 939         trace_kvm_cpuid(function, *eax, *ebx, *ecx, *edx, entry_found);
 940         return entry_found;
 941 }
 942 EXPORT_SYMBOL_GPL(kvm_cpuid);
 943
 944 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
 945 {
 946         u32 eax, ebx, ecx, edx;
 947
 948         if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
 949                 return 1;
 950
 951         eax = kvm_register_read(vcpu, VCPU_REGS_RAX);
 952         ecx = kvm_register_read(vcpu, VCPU_REGS_RCX);
 953         kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, true);
 954         kvm_register_write(vcpu, VCPU_REGS_RAX, eax);
 955         kvm_register_write(vcpu, VCPU_REGS_RBX, ebx);
 956         kvm_register_write(vcpu, VCPU_REGS_RCX, ecx);
 957         kvm_register_write(vcpu, VCPU_REGS_RDX, edx);
 958         return kvm_skip_emulated_instruction(vcpu);
 959 }
 960 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);