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 int kvm_update_cpuid(struct kvm_vcpu *vcpu)
  71 {
  72         struct kvm_cpuid_entry2 *best;
  73         struct kvm_lapic *apic = vcpu->arch.apic;
  74
  75         best = kvm_find_cpuid_entry(vcpu, 1, 0);
  76         if (!best)
  77                 return 0;
  78
  79         /* Update OSXSAVE bit */
  80         if (boot_cpu_has(X86_FEATURE_XSAVE) && best->function == 0x1) {
  81                 best->ecx &= ~F(OSXSAVE);
  82                 if (kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE))
  83                         best->ecx |= F(OSXSAVE);
  84         }
  85
  86         best->edx &= ~F(APIC);
  87         if (vcpu->arch.apic_base & MSR_IA32_APICBASE_ENABLE)
  88                 best->edx |= F(APIC);
  89
  90         if (apic) {
  91                 if (best->ecx & F(TSC_DEADLINE_TIMER))
  92                         apic->lapic_timer.timer_mode_mask = 3 << 17;
  93                 else
  94                         apic->lapic_timer.timer_mode_mask = 1 << 17;
  95         }
  96
  97         best = kvm_find_cpuid_entry(vcpu, 7, 0);
  98         if (best) {
  99                 /* Update OSPKE bit */
 100                 if (boot_cpu_has(X86_FEATURE_PKU) && best->function == 0x7) {
 101                         best->ecx &= ~F(OSPKE);
 102                         if (kvm_read_cr4_bits(vcpu, X86_CR4_PKE))
 103                                 best->ecx |= F(OSPKE);
 104                 }
 105         }
 106
 107         best = kvm_find_cpuid_entry(vcpu, 0xD, 0);
 108         if (!best) {
 109                 vcpu->arch.guest_supported_xcr0 = 0;
 110                 vcpu->arch.guest_xstate_size = XSAVE_HDR_SIZE + XSAVE_HDR_OFFSET;
 111         } else {
 112                 vcpu->arch.guest_supported_xcr0 =
 113                         (best->eax | ((u64)best->edx << 32)) &
 114                         kvm_supported_xcr0();
 115                 vcpu->arch.guest_xstate_size = best->ebx =
 116                         xstate_required_size(vcpu->arch.xcr0, false);
 117         }
 118
 119         best = kvm_find_cpuid_entry(vcpu, 0xD, 1);
 120         if (best && (best->eax & (F(XSAVES) | F(XSAVEC))))
 121                 best->ebx = xstate_required_size(vcpu->arch.xcr0, true);
 122
 123         /*
 124          * The existing code assumes virtual address is 48-bit or 57-bit in the
 125          * canonical address checks; exit if it is ever changed.
 126          */
 127         best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
 128         if (best) {
 129                 int vaddr_bits = (best->eax & 0xff00) >> 8;
 130
 131                 if (vaddr_bits != 48 && vaddr_bits != 57 && vaddr_bits != 0)
 132                         return -EINVAL;
 133         }
 134
 135         best = kvm_find_cpuid_entry(vcpu, KVM_CPUID_FEATURES, 0);
 136         if (kvm_hlt_in_guest(vcpu->kvm) && best &&
 137                 (best->eax & (1 << KVM_FEATURE_PV_UNHALT)))
 138                 best->eax &= ~(1 << KVM_FEATURE_PV_UNHALT);
 139
 140         if (!kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_MISC_ENABLE_NO_MWAIT)) {
 141                 best = kvm_find_cpuid_entry(vcpu, 0x1, 0);
 142                 if (best) {
 143                         if (vcpu->arch.ia32_misc_enable_msr & MSR_IA32_MISC_ENABLE_MWAIT)
 144                                 best->ecx |= F(MWAIT);
 145                         else
 146                                 best->ecx &= ~F(MWAIT);
 147                 }
 148         }
 149
 150         /* Update physical-address width */
 151         vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu);
 152         kvm_mmu_reset_context(vcpu);
 153
 154         kvm_pmu_refresh(vcpu);
 155         return 0;
 156 }
 157
 158 static int is_efer_nx(void)
 159 {
 160         unsigned long long efer = 0;
 161
 162         rdmsrl_safe(MSR_EFER, &efer);
 163         return efer & EFER_NX;
 164 }
 165
 166 static void cpuid_fix_nx_cap(struct kvm_vcpu *vcpu)
 167 {
 168         int i;
 169         struct kvm_cpuid_entry2 *e, *entry;
 170
 171         entry = NULL;
 172         for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
 173                 e = &vcpu->arch.cpuid_entries[i];
 174                 if (e->function == 0x80000001) {
 175                         entry = e;
 176                         break;
 177                 }
 178         }
 179         if (entry && (entry->edx & F(NX)) && !is_efer_nx()) {
 180                 entry->edx &= ~F(NX);
 181                 printk(KERN_INFO "kvm: guest NX capability removed\n");
 182         }
 183 }
 184
 185 int cpuid_query_maxphyaddr(struct kvm_vcpu *vcpu)
 186 {
 187         struct kvm_cpuid_entry2 *best;
 188
 189         best = kvm_find_cpuid_entry(vcpu, 0x80000000, 0);
 190         if (!best || best->eax < 0x80000008)
 191                 goto not_found;
 192         best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
 193         if (best)
 194                 return best->eax & 0xff;
 195 not_found:
 196         return 36;
 197 }
 198 EXPORT_SYMBOL_GPL(cpuid_query_maxphyaddr);
 199
 200 /* when an old userspace process fills a new kernel module */
 201 int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
 202                              struct kvm_cpuid *cpuid,
 203                              struct kvm_cpuid_entry __user *entries)
 204 {
 205         int r, i;
 206         struct kvm_cpuid_entry *cpuid_entries = NULL;
 207
 208         r = -E2BIG;
 209         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 210                 goto out;
 211         r = -ENOMEM;
 212         if (cpuid->nent) {
 213                 cpuid_entries =
 214                         vmalloc(array_size(sizeof(struct kvm_cpuid_entry),
 215                                            cpuid->nent));
 216                 if (!cpuid_entries)
 217                         goto out;
 218                 r = -EFAULT;
 219                 if (copy_from_user(cpuid_entries, entries,
 220                                    cpuid->nent * sizeof(struct kvm_cpuid_entry)))
 221                         goto out;
 222         }
 223         for (i = 0; i < cpuid->nent; i++) {
 224                 vcpu->arch.cpuid_entries[i].function = cpuid_entries[i].function;
 225                 vcpu->arch.cpuid_entries[i].eax = cpuid_entries[i].eax;
 226                 vcpu->arch.cpuid_entries[i].ebx = cpuid_entries[i].ebx;
 227                 vcpu->arch.cpuid_entries[i].ecx = cpuid_entries[i].ecx;
 228                 vcpu->arch.cpuid_entries[i].edx = cpuid_entries[i].edx;
 229                 vcpu->arch.cpuid_entries[i].index = 0;
 230                 vcpu->arch.cpuid_entries[i].flags = 0;
 231                 vcpu->arch.cpuid_entries[i].padding[0] = 0;
 232                 vcpu->arch.cpuid_entries[i].padding[1] = 0;
 233                 vcpu->arch.cpuid_entries[i].padding[2] = 0;
 234         }
 235         vcpu->arch.cpuid_nent = cpuid->nent;
 236         cpuid_fix_nx_cap(vcpu);
 237         kvm_apic_set_version(vcpu);
 238         kvm_x86_ops->cpuid_update(vcpu);
 239         r = kvm_update_cpuid(vcpu);
 240
 241 out:
 242         vfree(cpuid_entries);
 243         return r;
 244 }
 245
 246 int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
 247                               struct kvm_cpuid2 *cpuid,
 248                               struct kvm_cpuid_entry2 __user *entries)
 249 {
 250         int r;
 251
 252         r = -E2BIG;
 253         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 254                 goto out;
 255         r = -EFAULT;
 256         if (copy_from_user(&vcpu->arch.cpuid_entries, entries,
 257                            cpuid->nent * sizeof(struct kvm_cpuid_entry2)))
 258                 goto out;
 259         vcpu->arch.cpuid_nent = cpuid->nent;
 260         kvm_apic_set_version(vcpu);
 261         kvm_x86_ops->cpuid_update(vcpu);
 262         r = kvm_update_cpuid(vcpu);
 263 out:
 264         return r;
 265 }
 266
 267 int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
 268                               struct kvm_cpuid2 *cpuid,
 269                               struct kvm_cpuid_entry2 __user *entries)
 270 {
 271         int r;
 272
 273         r = -E2BIG;
 274         if (cpuid->nent < vcpu->arch.cpuid_nent)
 275                 goto out;
 276         r = -EFAULT;
 277         if (copy_to_user(entries, &vcpu->arch.cpuid_entries,
 278                          vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
 279                 goto out;
 280         return 0;
 281
 282 out:
 283         cpuid->nent = vcpu->arch.cpuid_nent;
 284         return r;
 285 }
 286
 287 static void cpuid_mask(u32 *word, int wordnum)
 288 {
 289         *word &= boot_cpu_data.x86_capability[wordnum];
 290 }
 291
 292 static void do_host_cpuid(struct kvm_cpuid_entry2 *entry, u32 function,
 293                            u32 index)
 294 {
 295         entry->function = function;
 296         entry->index = index;
 297         entry->flags = 0;
 298
 299         cpuid_count(entry->function, entry->index,
 300                     &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
 301
 302         switch (function) {
 303         case 2:
 304                 entry->flags |= KVM_CPUID_FLAG_STATEFUL_FUNC;
 305                 break;
 306         case 4:
 307         case 7:
 308         case 0xb:
 309         case 0xd:
 310         case 0x14:
 311         case 0x8000001d:
 312                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 313                 break;
 314         }
 315 }
 316
 317 static int __do_cpuid_func_emulated(struct kvm_cpuid_entry2 *entry,
 318                                     u32 func, int *nent, int maxnent)
 319 {
 320         entry->function = func;
 321         entry->index = 0;
 322         entry->flags = 0;
 323
 324         switch (func) {
 325         case 0:
 326                 entry->eax = 7;
 327                 ++*nent;
 328                 break;
 329         case 1:
 330                 entry->ecx = F(MOVBE);
 331                 ++*nent;
 332                 break;
 333         case 7:
 334                 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
 335                 entry->eax = 0;
 336                 entry->ecx = F(RDPID);
 337                 ++*nent;
 338         default:
 339                 break;
 340         }
 341
 342         return 0;
 343 }
 344
 345 static inline void do_cpuid_7_mask(struct kvm_cpuid_entry2 *entry, int index)
 346 {
 347         unsigned f_invpcid = kvm_x86_ops->invpcid_supported() ? F(INVPCID) : 0;
 348         unsigned f_mpx = kvm_mpx_supported() ? F(MPX) : 0;
 349         unsigned f_umip = kvm_x86_ops->umip_emulated() ? F(UMIP) : 0;
 350         unsigned f_intel_pt = kvm_x86_ops->pt_supported() ? F(INTEL_PT) : 0;
 351         unsigned f_la57;
 352
 353         /* cpuid 7.0.ebx */
 354         const u32 kvm_cpuid_7_0_ebx_x86_features =
 355                 F(FSGSBASE) | F(BMI1) | F(HLE) | F(AVX2) | F(SMEP) |
 356                 F(BMI2) | F(ERMS) | f_invpcid | F(RTM) | f_mpx | F(RDSEED) |
 357                 F(ADX) | F(SMAP) | F(AVX512IFMA) | F(AVX512F) | F(AVX512PF) |
 358                 F(AVX512ER) | F(AVX512CD) | F(CLFLUSHOPT) | F(CLWB) | F(AVX512DQ) |
 359                 F(SHA_NI) | F(AVX512BW) | F(AVX512VL) | f_intel_pt;
 360
 361         /* cpuid 7.0.ecx*/
 362         const u32 kvm_cpuid_7_0_ecx_x86_features =
 363                 F(AVX512VBMI) | F(LA57) | F(PKU) | 0 /*OSPKE*/ |
 364                 F(AVX512_VPOPCNTDQ) | F(UMIP) | F(AVX512_VBMI2) | F(GFNI) |
 365                 F(VAES) | F(VPCLMULQDQ) | F(AVX512_VNNI) | F(AVX512_BITALG) |
 366                 F(CLDEMOTE) | F(MOVDIRI) | F(MOVDIR64B);
 367
 368         /* cpuid 7.0.edx*/
 369         const u32 kvm_cpuid_7_0_edx_x86_features =
 370                 F(AVX512_4VNNIW) | F(AVX512_4FMAPS) | F(SPEC_CTRL) |
 371                 F(SPEC_CTRL_SSBD) | F(ARCH_CAPABILITIES) | F(INTEL_STIBP) |
 372                 F(MD_CLEAR);
 373
 374         switch (index) {
 375         case 0:
 376                 entry->eax = 0;
 377                 entry->ebx &= kvm_cpuid_7_0_ebx_x86_features;
 378                 cpuid_mask(&entry->ebx, CPUID_7_0_EBX);
 379                 /* TSC_ADJUST is emulated */
 380                 entry->ebx |= F(TSC_ADJUST);
 381
 382                 entry->ecx &= kvm_cpuid_7_0_ecx_x86_features;
 383                 f_la57 = entry->ecx & F(LA57);
 384                 cpuid_mask(&entry->ecx, CPUID_7_ECX);
 385                 /* Set LA57 based on hardware capability. */
 386                 entry->ecx |= f_la57;
 387                 entry->ecx |= f_umip;
 388                 /* PKU is not yet implemented for shadow paging. */
 389                 if (!tdp_enabled || !boot_cpu_has(X86_FEATURE_OSPKE))
 390                         entry->ecx &= ~F(PKU);
 391
 392                 entry->edx &= kvm_cpuid_7_0_edx_x86_features;
 393                 cpuid_mask(&entry->edx, CPUID_7_EDX);
 394                 /*
 395                  * We emulate ARCH_CAPABILITIES in software even
 396                  * if the host doesn't support it.
 397                  */
 398                 entry->edx |= F(ARCH_CAPABILITIES);
 399                 break;
 400         default:
 401                 WARN_ON_ONCE(1);
 402                 entry->eax = 0;
 403                 entry->ebx = 0;
 404                 entry->ecx = 0;
 405                 entry->edx = 0;
 406                 break;
 407         }
 408 }
 409
 410 static inline int __do_cpuid_func(struct kvm_cpuid_entry2 *entry, u32 function,
 411                                   int *nent, int maxnent)
 412 {
 413         int r;
 414         unsigned f_nx = is_efer_nx() ? F(NX) : 0;
 415 #ifdef CONFIG_X86_64
 416         unsigned f_gbpages = (kvm_x86_ops->get_lpage_level() == PT_PDPE_LEVEL)
 417                                 ? F(GBPAGES) : 0;
 418         unsigned f_lm = F(LM);
 419 #else
 420         unsigned f_gbpages = 0;
 421         unsigned f_lm = 0;
 422 #endif
 423         unsigned f_rdtscp = kvm_x86_ops->rdtscp_supported() ? F(RDTSCP) : 0;
 424         unsigned f_xsaves = kvm_x86_ops->xsaves_supported() ? F(XSAVES) : 0;
 425         unsigned f_intel_pt = kvm_x86_ops->pt_supported() ? F(INTEL_PT) : 0;
 426
 427         /* cpuid 1.edx */
 428         const u32 kvm_cpuid_1_edx_x86_features =
 429                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 430                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 431                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
 432                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 433                 F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLUSH) |
 434                 0 /* Reserved, DS, ACPI */ | F(MMX) |
 435                 F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
 436                 0 /* HTT, TM, Reserved, PBE */;
 437         /* cpuid 0x80000001.edx */
 438         const u32 kvm_cpuid_8000_0001_edx_x86_features =
 439                 F(FPU) | F(VME) | F(DE) | F(PSE) |
 440                 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
 441                 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
 442                 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
 443                 F(PAT) | F(PSE36) | 0 /* Reserved */ |
 444                 f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
 445                 F(FXSR) | F(FXSR_OPT) | f_gbpages | f_rdtscp |
 446                 0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW);
 447         /* cpuid 1.ecx */
 448         const u32 kvm_cpuid_1_ecx_x86_features =
 449                 /* NOTE: MONITOR (and MWAIT) are emulated as NOP,
 450                  * but *not* advertised to guests via CPUID ! */
 451                 F(XMM3) | F(PCLMULQDQ) | 0 /* DTES64, MONITOR */ |
 452                 0 /* DS-CPL, VMX, SMX, EST */ |
 453                 0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
 454                 F(FMA) | F(CX16) | 0 /* xTPR Update, PDCM */ |
 455                 F(PCID) | 0 /* Reserved, DCA */ | F(XMM4_1) |
 456                 F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
 457                 0 /* Reserved*/ | F(AES) | F(XSAVE) | 0 /* OSXSAVE */ | F(AVX) |
 458                 F(F16C) | F(RDRAND);
 459         /* cpuid 0x80000001.ecx */
 460         const u32 kvm_cpuid_8000_0001_ecx_x86_features =
 461                 F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ |
 462                 F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
 463                 F(3DNOWPREFETCH) | F(OSVW) | 0 /* IBS */ | F(XOP) |
 464                 0 /* SKINIT, WDT, LWP */ | F(FMA4) | F(TBM) |
 465                 F(TOPOEXT) | F(PERFCTR_CORE);
 466
 467         /* cpuid 0x80000008.ebx */
 468         const u32 kvm_cpuid_8000_0008_ebx_x86_features =
 469                 F(WBNOINVD) | F(AMD_IBPB) | F(AMD_IBRS) | F(AMD_SSBD) | F(VIRT_SSBD) |
 470                 F(AMD_SSB_NO) | F(AMD_STIBP) | F(AMD_STIBP_ALWAYS_ON);
 471
 472         /* cpuid 0xC0000001.edx */
 473         const u32 kvm_cpuid_C000_0001_edx_x86_features =
 474                 F(XSTORE) | F(XSTORE_EN) | F(XCRYPT) | F(XCRYPT_EN) |
 475                 F(ACE2) | F(ACE2_EN) | F(PHE) | F(PHE_EN) |
 476                 F(PMM) | F(PMM_EN);
 477
 478         /* cpuid 0xD.1.eax */
 479         const u32 kvm_cpuid_D_1_eax_x86_features =
 480                 F(XSAVEOPT) | F(XSAVEC) | F(XGETBV1) | f_xsaves;
 481
 482         /* all calls to cpuid_count() should be made on the same cpu */
 483         get_cpu();
 484
 485         r = -E2BIG;
 486
 487         if (*nent >= maxnent)
 488                 goto out;
 489
 490         do_host_cpuid(entry, function, 0);
 491         ++*nent;
 492
 493         switch (function) {
 494         case 0:
 495                 /* Limited to the highest leaf implemented in KVM. */
 496                 entry->eax = min(entry->eax, 0x1fU);
 497                 break;
 498         case 1:
 499                 entry->edx &= kvm_cpuid_1_edx_x86_features;
 500                 cpuid_mask(&entry->edx, CPUID_1_EDX);
 501                 entry->ecx &= kvm_cpuid_1_ecx_x86_features;
 502                 cpuid_mask(&entry->ecx, CPUID_1_ECX);
 503                 /* we support x2apic emulation even if host does not support
 504                  * it since we emulate x2apic in software */
 505                 entry->ecx |= F(X2APIC);
 506                 break;
 507         /* function 2 entries are STATEFUL. That is, repeated cpuid commands
 508          * may return different values. This forces us to get_cpu() before
 509          * issuing the first command, and also to emulate this annoying behavior
 510          * in kvm_emulate_cpuid() using KVM_CPUID_FLAG_STATE_READ_NEXT */
 511         case 2: {
 512                 int t, times = entry->eax & 0xff;
 513
 514                 entry->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 515                 for (t = 1; t < times; ++t) {
 516                         if (*nent >= maxnent)
 517                                 goto out;
 518
 519                         do_host_cpuid(&entry[t], function, 0);
 520                         ++*nent;
 521                 }
 522                 break;
 523         }
 524         /* functions 4 and 0x8000001d have additional index. */
 525         case 4:
 526         case 0x8000001d: {
 527                 int i, cache_type;
 528
 529                 /* read more entries until cache_type is zero */
 530                 for (i = 1; ; ++i) {
 531                         if (*nent >= maxnent)
 532                                 goto out;
 533
 534                         cache_type = entry[i - 1].eax & 0x1f;
 535                         if (!cache_type)
 536                                 break;
 537                         do_host_cpuid(&entry[i], function, i);
 538                         ++*nent;
 539                 }
 540                 break;
 541         }
 542         case 6: /* Thermal management */
 543                 entry->eax = 0x4; /* allow ARAT */
 544                 entry->ebx = 0;
 545                 entry->ecx = 0;
 546                 entry->edx = 0;
 547                 break;
 548         /* function 7 has additional index. */
 549         case 7: {
 550                 int i;
 551
 552                 for (i = 0; ; ) {
 553                         do_cpuid_7_mask(&entry[i], i);
 554                         if (i == entry->eax)
 555                                 break;
 556                         if (*nent >= maxnent)
 557                                 goto out;
 558
 559                         ++i;
 560                         do_host_cpuid(&entry[i], function, i);
 561                         ++*nent;
 562                 }
 563                 break;
 564         }
 565         case 9:
 566                 break;
 567         case 0xa: { /* Architectural Performance Monitoring */
 568                 struct x86_pmu_capability cap;
 569                 union cpuid10_eax eax;
 570                 union cpuid10_edx edx;
 571
 572                 perf_get_x86_pmu_capability(&cap);
 573
 574                 /*
 575                  * Only support guest architectural pmu on a host
 576                  * with architectural pmu.
 577                  */
 578                 if (!cap.version)
 579                         memset(&cap, 0, sizeof(cap));
 580
 581                 eax.split.version_id = min(cap.version, 2);
 582                 eax.split.num_counters = cap.num_counters_gp;
 583                 eax.split.bit_width = cap.bit_width_gp;
 584                 eax.split.mask_length = cap.events_mask_len;
 585
 586                 edx.split.num_counters_fixed = cap.num_counters_fixed;
 587                 edx.split.bit_width_fixed = cap.bit_width_fixed;
 588                 edx.split.reserved = 0;
 589
 590                 entry->eax = eax.full;
 591                 entry->ebx = cap.events_mask;
 592                 entry->ecx = 0;
 593                 entry->edx = edx.full;
 594                 break;
 595         }
 596         /*
 597          * Per Intel's SDM, the 0x1f is a superset of 0xb,
 598          * thus they can be handled by common code.
 599          */
 600         case 0x1f:
 601         case 0xb: {
 602                 int i, level_type;
 603
 604                 /* read more entries until level_type is zero */
 605                 for (i = 1; ; ++i) {
 606                         if (*nent >= maxnent)
 607                                 goto out;
 608
 609                         level_type = entry[i - 1].ecx & 0xff00;
 610                         if (!level_type)
 611                                 break;
 612                         do_host_cpuid(&entry[i], function, i);
 613                         ++*nent;
 614                 }
 615                 break;
 616         }
 617         case 0xd: {
 618                 int idx, i;
 619                 u64 supported = kvm_supported_xcr0();
 620
 621                 entry->eax &= supported;
 622                 entry->ebx = xstate_required_size(supported, false);
 623                 entry->ecx = entry->ebx;
 624                 entry->edx &= supported >> 32;
 625                 if (!supported)
 626                         break;
 627
 628                 for (idx = 1, i = 1; idx < 64; ++idx) {
 629                         u64 mask = ((u64)1 << idx);
 630                         if (*nent >= maxnent)
 631                                 goto out;
 632
 633                         do_host_cpuid(&entry[i], function, idx);
 634                         if (idx == 1) {
 635                                 entry[i].eax &= kvm_cpuid_D_1_eax_x86_features;
 636                                 cpuid_mask(&entry[i].eax, CPUID_D_1_EAX);
 637                                 entry[i].ebx = 0;
 638                                 if (entry[i].eax & (F(XSAVES)|F(XSAVEC)))
 639                                         entry[i].ebx =
 640                                                 xstate_required_size(supported,
 641                                                                      true);
 642                         } else {
 643                                 if (entry[i].eax == 0 || !(supported & mask))
 644                                         continue;
 645                                 if (WARN_ON_ONCE(entry[i].ecx & 1))
 646                                         continue;
 647                         }
 648                         entry[i].ecx = 0;
 649                         entry[i].edx = 0;
 650                         ++*nent;
 651                         ++i;
 652                 }
 653                 break;
 654         }
 655         /* Intel PT */
 656         case 0x14: {
 657                 int t, times = entry->eax;
 658
 659                 if (!f_intel_pt)
 660                         break;
 661
 662                 for (t = 1; t <= times; ++t) {
 663                         if (*nent >= maxnent)
 664                                 goto out;
 665                         do_host_cpuid(&entry[t], function, t);
 666                         ++*nent;
 667                 }
 668                 break;
 669         }
 670         case KVM_CPUID_SIGNATURE: {
 671                 static const char signature[12] = "KVMKVMKVM\0\0";
 672                 const u32 *sigptr = (const u32 *)signature;
 673                 entry->eax = KVM_CPUID_FEATURES;
 674                 entry->ebx = sigptr[0];
 675                 entry->ecx = sigptr[1];
 676                 entry->edx = sigptr[2];
 677                 break;
 678         }
 679         case KVM_CPUID_FEATURES:
 680                 entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
 681                              (1 << KVM_FEATURE_NOP_IO_DELAY) |
 682                              (1 << KVM_FEATURE_CLOCKSOURCE2) |
 683                              (1 << KVM_FEATURE_ASYNC_PF) |
 684                              (1 << KVM_FEATURE_PV_EOI) |
 685                              (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) |
 686                              (1 << KVM_FEATURE_PV_UNHALT) |
 687                              (1 << KVM_FEATURE_PV_TLB_FLUSH) |
 688                              (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
 689                              (1 << KVM_FEATURE_PV_SEND_IPI) |
 690                              (1 << KVM_FEATURE_POLL_CONTROL) |
 691                              (1 << KVM_FEATURE_PV_SCHED_YIELD);
 692
 693                 if (sched_info_on())
 694                         entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
 695
 696                 entry->ebx = 0;
 697                 entry->ecx = 0;
 698                 entry->edx = 0;
 699                 break;
 700         case 0x80000000:
 701                 entry->eax = min(entry->eax, 0x8000001f);
 702                 break;
 703         case 0x80000001:
 704                 entry->edx &= kvm_cpuid_8000_0001_edx_x86_features;
 705                 cpuid_mask(&entry->edx, CPUID_8000_0001_EDX);
 706                 entry->ecx &= kvm_cpuid_8000_0001_ecx_x86_features;
 707                 cpuid_mask(&entry->ecx, CPUID_8000_0001_ECX);
 708                 break;
 709         case 0x80000007: /* Advanced power management */
 710                 /* invariant TSC is CPUID.80000007H:EDX[8] */
 711                 entry->edx &= (1 << 8);
 712                 /* mask against host */
 713                 entry->edx &= boot_cpu_data.x86_power;
 714                 entry->eax = entry->ebx = entry->ecx = 0;
 715                 break;
 716         case 0x80000008: {
 717                 unsigned g_phys_as = (entry->eax >> 16) & 0xff;
 718                 unsigned virt_as = max((entry->eax >> 8) & 0xff, 48U);
 719                 unsigned phys_as = entry->eax & 0xff;
 720
 721                 if (!g_phys_as)
 722                         g_phys_as = phys_as;
 723                 entry->eax = g_phys_as | (virt_as << 8);
 724                 entry->edx = 0;
 725                 /*
 726                  * IBRS, IBPB and VIRT_SSBD aren't necessarily present in
 727                  * hardware cpuid
 728                  */
 729                 if (boot_cpu_has(X86_FEATURE_AMD_IBPB))
 730                         entry->ebx |= F(AMD_IBPB);
 731                 if (boot_cpu_has(X86_FEATURE_AMD_IBRS))
 732                         entry->ebx |= F(AMD_IBRS);
 733                 if (boot_cpu_has(X86_FEATURE_VIRT_SSBD))
 734                         entry->ebx |= F(VIRT_SSBD);
 735                 entry->ebx &= kvm_cpuid_8000_0008_ebx_x86_features;
 736                 cpuid_mask(&entry->ebx, CPUID_8000_0008_EBX);
 737                 /*
 738                  * The preference is to use SPEC CTRL MSR instead of the
 739                  * VIRT_SPEC MSR.
 740                  */
 741                 if (boot_cpu_has(X86_FEATURE_LS_CFG_SSBD) &&
 742                     !boot_cpu_has(X86_FEATURE_AMD_SSBD))
 743                         entry->ebx |= F(VIRT_SSBD);
 744                 break;
 745         }
 746         case 0x80000019:
 747                 entry->ecx = entry->edx = 0;
 748                 break;
 749         case 0x8000001a:
 750         case 0x8000001e:
 751                 break;
 752         /*Add support for Centaur's CPUID instruction*/
 753         case 0xC0000000:
 754                 /*Just support up to 0xC0000004 now*/
 755                 entry->eax = min(entry->eax, 0xC0000004);
 756                 break;
 757         case 0xC0000001:
 758                 entry->edx &= kvm_cpuid_C000_0001_edx_x86_features;
 759                 cpuid_mask(&entry->edx, CPUID_C000_0001_EDX);
 760                 break;
 761         case 3: /* Processor serial number */
 762         case 5: /* MONITOR/MWAIT */
 763         case 0xC0000002:
 764         case 0xC0000003:
 765         case 0xC0000004:
 766         default:
 767                 entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
 768                 break;
 769         }
 770
 771         kvm_x86_ops->set_supported_cpuid(function, entry);
 772
 773         r = 0;
 774
 775 out:
 776         put_cpu();
 777
 778         return r;
 779 }
 780
 781 static int do_cpuid_func(struct kvm_cpuid_entry2 *entry, u32 func,
 782                          int *nent, int maxnent, unsigned int type)
 783 {
 784         if (type == KVM_GET_EMULATED_CPUID)
 785                 return __do_cpuid_func_emulated(entry, func, nent, maxnent);
 786
 787         return __do_cpuid_func(entry, func, nent, maxnent);
 788 }
 789
 790 #undef F
 791
 792 struct kvm_cpuid_param {
 793         u32 func;
 794         bool has_leaf_count;
 795         bool (*qualifier)(const struct kvm_cpuid_param *param);
 796 };
 797
 798 static bool is_centaur_cpu(const struct kvm_cpuid_param *param)
 799 {
 800         return boot_cpu_data.x86_vendor == X86_VENDOR_CENTAUR;
 801 }
 802
 803 static bool sanity_check_entries(struct kvm_cpuid_entry2 __user *entries,
 804                                  __u32 num_entries, unsigned int ioctl_type)
 805 {
 806         int i;
 807         __u32 pad[3];
 808
 809         if (ioctl_type != KVM_GET_EMULATED_CPUID)
 810                 return false;
 811
 812         /*
 813          * We want to make sure that ->padding is being passed clean from
 814          * userspace in case we want to use it for something in the future.
 815          *
 816          * Sadly, this wasn't enforced for KVM_GET_SUPPORTED_CPUID and so we
 817          * have to give ourselves satisfied only with the emulated side. /me
 818          * sheds a tear.
 819          */
 820         for (i = 0; i < num_entries; i++) {
 821                 if (copy_from_user(pad, entries[i].padding, sizeof(pad)))
 822                         return true;
 823
 824                 if (pad[0] || pad[1] || pad[2])
 825                         return true;
 826         }
 827         return false;
 828 }
 829
 830 int kvm_dev_ioctl_get_cpuid(struct kvm_cpuid2 *cpuid,
 831                             struct kvm_cpuid_entry2 __user *entries,
 832                             unsigned int type)
 833 {
 834         struct kvm_cpuid_entry2 *cpuid_entries;
 835         int limit, nent = 0, r = -E2BIG, i;
 836         u32 func;
 837         static const struct kvm_cpuid_param param[] = {
 838                 { .func = 0, .has_leaf_count = true },
 839                 { .func = 0x80000000, .has_leaf_count = true },
 840                 { .func = 0xC0000000, .qualifier = is_centaur_cpu, .has_leaf_count = true },
 841                 { .func = KVM_CPUID_SIGNATURE },
 842                 { .func = KVM_CPUID_FEATURES },
 843         };
 844
 845         if (cpuid->nent < 1)
 846                 goto out;
 847         if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
 848                 cpuid->nent = KVM_MAX_CPUID_ENTRIES;
 849
 850         if (sanity_check_entries(entries, cpuid->nent, type))
 851                 return -EINVAL;
 852
 853         r = -ENOMEM;
 854         cpuid_entries = vzalloc(array_size(sizeof(struct kvm_cpuid_entry2),
 855                                            cpuid->nent));
 856         if (!cpuid_entries)
 857                 goto out;
 858
 859         r = 0;
 860         for (i = 0; i < ARRAY_SIZE(param); i++) {
 861                 const struct kvm_cpuid_param *ent = &param[i];
 862
 863                 if (ent->qualifier && !ent->qualifier(ent))
 864                         continue;
 865
 866                 r = do_cpuid_func(&cpuid_entries[nent], ent->func,
 867                                   &nent, cpuid->nent, type);
 868
 869                 if (r)
 870                         goto out_free;
 871
 872                 if (!ent->has_leaf_count)
 873                         continue;
 874
 875                 limit = cpuid_entries[nent - 1].eax;
 876                 for (func = ent->func + 1; func <= limit && nent < cpuid->nent && r == 0; ++func)
 877                         r = do_cpuid_func(&cpuid_entries[nent], func,
 878                                           &nent, cpuid->nent, type);
 879
 880                 if (r)
 881                         goto out_free;
 882         }
 883
 884         r = -EFAULT;
 885         if (copy_to_user(entries, cpuid_entries,
 886                          nent * sizeof(struct kvm_cpuid_entry2)))
 887                 goto out_free;
 888         cpuid->nent = nent;
 889         r = 0;
 890
 891 out_free:
 892         vfree(cpuid_entries);
 893 out:
 894         return r;
 895 }
 896
 897 static int move_to_next_stateful_cpuid_entry(struct kvm_vcpu *vcpu, int i)
 898 {
 899         struct kvm_cpuid_entry2 *e = &vcpu->arch.cpuid_entries[i];
 900         struct kvm_cpuid_entry2 *ej;
 901         int j = i;
 902         int nent = vcpu->arch.cpuid_nent;
 903
 904         e->flags &= ~KVM_CPUID_FLAG_STATE_READ_NEXT;
 905         /* when no next entry is found, the current entry[i] is reselected */
 906         do {
 907                 j = (j + 1) % nent;
 908                 ej = &vcpu->arch.cpuid_entries[j];
 909         } while (ej->function != e->function);
 910
 911         ej->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
 912
 913         return j;
 914 }
 915
 916 /* find an entry with matching function, matching index (if needed), and that
 917  * should be read next (if it's stateful) */
 918 static int is_matching_cpuid_entry(struct kvm_cpuid_entry2 *e,
 919         u32 function, u32 index)
 920 {
 921         if (e->function != function)
 922                 return 0;
 923         if ((e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) && e->index != index)
 924                 return 0;
 925         if ((e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC) &&
 926             !(e->flags & KVM_CPUID_FLAG_STATE_READ_NEXT))
 927                 return 0;
 928         return 1;
 929 }
 930
 931 struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
 932                                               u32 function, u32 index)
 933 {
 934         int i;
 935         struct kvm_cpuid_entry2 *best = NULL;
 936
 937         for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
 938                 struct kvm_cpuid_entry2 *e;
 939
 940                 e = &vcpu->arch.cpuid_entries[i];
 941                 if (is_matching_cpuid_entry(e, function, index)) {
 942                         if (e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC)
 943                                 move_to_next_stateful_cpuid_entry(vcpu, i);
 944                         best = e;
 945                         break;
 946                 }
 947         }
 948         return best;
 949 }
 950 EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
 951
 952 /*
 953  * If no match is found, check whether we exceed the vCPU's limit
 954  * and return the content of the highest valid _standard_ leaf instead.
 955  * This is to satisfy the CPUID specification.
 956  */
 957 static struct kvm_cpuid_entry2* check_cpuid_limit(struct kvm_vcpu *vcpu,
 958                                                   u32 function, u32 index)
 959 {
 960         struct kvm_cpuid_entry2 *maxlevel;
 961
 962         maxlevel = kvm_find_cpuid_entry(vcpu, function & 0x80000000, 0);
 963         if (!maxlevel || maxlevel->eax >= function)
 964                 return NULL;
 965         if (function & 0x80000000) {
 966                 maxlevel = kvm_find_cpuid_entry(vcpu, 0, 0);
 967                 if (!maxlevel)
 968                         return NULL;
 969         }
 970         return kvm_find_cpuid_entry(vcpu, maxlevel->eax, index);
 971 }
 972
 973 bool kvm_cpuid(struct kvm_vcpu *vcpu, u32 *eax, u32 *ebx,
 974                u32 *ecx, u32 *edx, bool check_limit)
 975 {
 976         u32 function = *eax, index = *ecx;
 977         struct kvm_cpuid_entry2 *best;
 978         bool entry_found = true;
 979
 980         best = kvm_find_cpuid_entry(vcpu, function, index);
 981
 982         if (!best) {
 983                 entry_found = false;
 984                 if (!check_limit)
 985                         goto out;
 986
 987                 best = check_cpuid_limit(vcpu, function, index);
 988         }
 989
 990 out:
 991         if (best) {
 992                 *eax = best->eax;
 993                 *ebx = best->ebx;
 994                 *ecx = best->ecx;
 995                 *edx = best->edx;
 996         } else
 997                 *eax = *ebx = *ecx = *edx = 0;
 998         trace_kvm_cpuid(function, *eax, *ebx, *ecx, *edx, entry_found);
 999         return entry_found;
1000 }
1001 EXPORT_SYMBOL_GPL(kvm_cpuid);
1002
1003 int kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
1004 {
1005         u32 eax, ebx, ecx, edx;
1006
1007         if (cpuid_fault_enabled(vcpu) && !kvm_require_cpl(vcpu, 0))
1008                 return 1;
1009
1010         eax = kvm_rax_read(vcpu);
1011         ecx = kvm_rcx_read(vcpu);
1012         kvm_cpuid(vcpu, &eax, &ebx, &ecx, &edx, true);
1013         kvm_rax_write(vcpu, eax);
1014         kvm_rbx_write(vcpu, ebx);
1015         kvm_rcx_write(vcpu, ecx);
1016         kvm_rdx_write(vcpu, edx);
1017         return kvm_skip_emulated_instruction(vcpu);
1018 }
1019 EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);