sys/amd64/vmm/x86.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2011 NetApp, Inc.
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  *
  28  * $FreeBSD$
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 #include <sys/param.h>
  35 #include <sys/pcpu.h>
  36 #include <sys/systm.h>
  37 #include <sys/sysctl.h>
  38
  39 #include <machine/clock.h>
  40 #include <machine/cpufunc.h>
  41 #include <machine/md_var.h>
  42 #include <machine/segments.h>
  43 #include <machine/specialreg.h>
  44
  45 #include <machine/vmm.h>
  46
  47 #include "vmm_host.h"
  48 #include "vmm_ktr.h"
  49 #include "vmm_util.h"
  50 #include "x86.h"
  51
  52 SYSCTL_DECL(_hw_vmm);
  53 static SYSCTL_NODE(_hw_vmm, OID_AUTO, topology, CTLFLAG_RD, 0, NULL);
  54
  55 #define CPUID_VM_HIGH           0x40000000
  56
  57 static const char bhyve_id[12] = "bhyve bhyve ";
  58
  59 static uint64_t bhyve_xcpuids;
  60 SYSCTL_ULONG(_hw_vmm, OID_AUTO, bhyve_xcpuids, CTLFLAG_RW, &bhyve_xcpuids, 0,
  61     "Number of times an unknown cpuid leaf was accessed");
  62
  63 #if __FreeBSD_version < 1200060 /* Remove after 11 EOL helps MFCing */
  64 extern u_int threads_per_core;
  65 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, threads_per_core, CTLFLAG_RDTUN,
  66     &threads_per_core, 0, NULL);
  67
  68 extern u_int cores_per_package;
  69 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, cores_per_package, CTLFLAG_RDTUN,
  70     &cores_per_package, 0, NULL);
  71 #endif
  72
  73 static int cpuid_leaf_b = 1;
  74 SYSCTL_INT(_hw_vmm_topology, OID_AUTO, cpuid_leaf_b, CTLFLAG_RDTUN,
  75     &cpuid_leaf_b, 0, NULL);
  76
  77 /*
  78  * Round up to the next power of two, if necessary, and then take log2.
  79  * Returns -1 if argument is zero.
  80  */
  81 static __inline int
  82 log2(u_int x)
  83 {
  84
  85         return (fls(x << (1 - powerof2(x))) - 1);
  86 }
  87
  88 int
  89 x86_emulate_cpuid(struct vm *vm, int vcpu_id,
  90                   uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx)
  91 {
  92         const struct xsave_limits *limits;
  93         uint64_t cr4;
  94         int error, enable_invpcid, level, width, x2apic_id;
  95         unsigned int func, regs[4], logical_cpus;
  96         enum x2apic_state x2apic_state;
  97         uint16_t cores, maxcpus, sockets, threads;
  98
  99         VCPU_CTR2(vm, vcpu_id, "cpuid %#x,%#x", *eax, *ecx);
 100
 101         /*
 102          * Requests for invalid CPUID levels should map to the highest
 103          * available level instead.
 104          */
 105         if (cpu_exthigh != 0 && *eax >= 0x80000000) {
 106                 if (*eax > cpu_exthigh)
 107                         *eax = cpu_exthigh;
 108         } else if (*eax >= 0x40000000) {
 109                 if (*eax > CPUID_VM_HIGH)
 110                         *eax = CPUID_VM_HIGH;
 111         } else if (*eax > cpu_high) {
 112                 *eax = cpu_high;
 113         }
 114
 115         func = *eax;
 116
 117         /*
 118          * In general the approach used for CPU topology is to
 119          * advertise a flat topology where all CPUs are packages with
 120          * no multi-core or SMT.
 121          */
 122         switch (func) {
 123                 /*
 124                  * Pass these through to the guest
 125                  */
 126                 case CPUID_0000_0000:
 127                 case CPUID_0000_0002:
 128                 case CPUID_0000_0003:
 129                 case CPUID_8000_0000:
 130                 case CPUID_8000_0002:
 131                 case CPUID_8000_0003:
 132                 case CPUID_8000_0004:
 133                 case CPUID_8000_0006:
 134                         cpuid_count(*eax, *ecx, regs);
 135                         break;
 136                 case CPUID_8000_0008:
 137                         cpuid_count(*eax, *ecx, regs);
 138                         if (vmm_is_amd()) {
 139                                 /*
 140                                  * XXX this might appear silly because AMD
 141                                  * cpus don't have threads.
 142                                  *
 143                                  * However this matches the logical cpus as
 144                                  * advertised by leaf 0x1 and will work even
 145                                  * if threads is set incorrectly on an AMD host.
 146                                  */
 147                                 vm_get_topology(vm, &sockets, &cores, &threads,
 148                                     &maxcpus);
 149                                 logical_cpus = threads * cores;
 150                                 regs[2] = logical_cpus - 1;
 151                         }
 152                         break;
 153
 154                 case CPUID_8000_0001:
 155                         cpuid_count(*eax, *ecx, regs);
 156
 157                         /*
 158                          * Hide SVM and Topology Extension features from guest.
 159                          */
 160                         regs[2] &= ~(AMDID2_SVM | AMDID2_TOPOLOGY);
 161
 162                         /*
 163                          * Don't advertise extended performance counter MSRs
 164                          * to the guest.
 165                          */
 166                         regs[2] &= ~AMDID2_PCXC;
 167                         regs[2] &= ~AMDID2_PNXC;
 168                         regs[2] &= ~AMDID2_PTSCEL2I;
 169
 170                         /*
 171                          * Don't advertise Instruction Based Sampling feature.
 172                          */
 173                         regs[2] &= ~AMDID2_IBS;
 174
 175                         /* NodeID MSR not available */
 176                         regs[2] &= ~AMDID2_NODE_ID;
 177
 178                         /* Don't advertise the OS visible workaround feature */
 179                         regs[2] &= ~AMDID2_OSVW;
 180
 181                         /* Hide mwaitx/monitorx capability from the guest */
 182                         regs[2] &= ~AMDID2_MWAITX;
 183
 184                         /*
 185                          * Hide rdtscp/ia32_tsc_aux until we know how
 186                          * to deal with them.
 187                          */
 188                         regs[3] &= ~AMDID_RDTSCP;
 189                         break;
 190
 191                 case CPUID_8000_0007:
 192                         /*
 193                          * AMD uses this leaf to advertise the processor's
 194                          * power monitoring and RAS capabilities. These
 195                          * features are hardware-specific and exposing
 196                          * them to a guest doesn't make a lot of sense.
 197                          *
 198                          * Intel uses this leaf only to advertise the
 199                          * "Invariant TSC" feature with all other bits
 200                          * being reserved (set to zero).
 201                          */
 202                         regs[0] = 0;
 203                         regs[1] = 0;
 204                         regs[2] = 0;
 205                         regs[3] = 0;
 206
 207                         /*
 208                          * "Invariant TSC" can be advertised to the guest if:
 209                          * - host TSC frequency is invariant
 210                          * - host TSCs are synchronized across physical cpus
 211                          *
 212                          * XXX This still falls short because the vcpu
 213                          * can observe the TSC moving backwards as it
 214                          * migrates across physical cpus. But at least
 215                          * it should discourage the guest from using the
 216                          * TSC to keep track of time.
 217                          */
 218                         if (tsc_is_invariant && smp_tsc)
 219                                 regs[3] |= AMDPM_TSC_INVARIANT;
 220                         break;
 221
 222                 case CPUID_0000_0001:
 223                         do_cpuid(1, regs);
 224
 225                         error = vm_get_x2apic_state(vm, vcpu_id, &x2apic_state);
 226                         if (error) {
 227                                 panic("x86_emulate_cpuid: error %d "
 228                                       "fetching x2apic state", error);
 229                         }
 230
 231                         /*
 232                          * Override the APIC ID only in ebx
 233                          */
 234                         regs[1] &= ~(CPUID_LOCAL_APIC_ID);
 235                         regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT);
 236
 237                         /*
 238                          * Don't expose VMX, SpeedStep, TME or SMX capability.
 239                          * Advertise x2APIC capability and Hypervisor guest.
 240                          */
 241                         regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2);
 242                         regs[2] &= ~(CPUID2_SMX);
 243
 244                         regs[2] |= CPUID2_HV;
 245
 246                         if (x2apic_state != X2APIC_DISABLED)
 247                                 regs[2] |= CPUID2_X2APIC;
 248                         else
 249                                 regs[2] &= ~CPUID2_X2APIC;
 250
 251                         /*
 252                          * Only advertise CPUID2_XSAVE in the guest if
 253                          * the host is using XSAVE.
 254                          */
 255                         if (!(regs[2] & CPUID2_OSXSAVE))
 256                                 regs[2] &= ~CPUID2_XSAVE;
 257
 258                         /*
 259                          * If CPUID2_XSAVE is being advertised and the
 260                          * guest has set CR4_XSAVE, set
 261                          * CPUID2_OSXSAVE.
 262                          */
 263                         regs[2] &= ~CPUID2_OSXSAVE;
 264                         if (regs[2] & CPUID2_XSAVE) {
 265                                 error = vm_get_register(vm, vcpu_id,
 266                                     VM_REG_GUEST_CR4, &cr4);
 267                                 if (error)
 268                                         panic("x86_emulate_cpuid: error %d "
 269                                               "fetching %%cr4", error);
 270                                 if (cr4 & CR4_XSAVE)
 271                                         regs[2] |= CPUID2_OSXSAVE;
 272                         }
 273
 274                         /*
 275                          * Hide monitor/mwait until we know how to deal with
 276                          * these instructions.
 277                          */
 278                         regs[2] &= ~CPUID2_MON;
 279
 280                         /*
 281                          * Hide the performance and debug features.
 282                          */
 283                         regs[2] &= ~CPUID2_PDCM;
 284
 285                         /*
 286                          * No TSC deadline support in the APIC yet
 287                          */
 288                         regs[2] &= ~CPUID2_TSCDLT;
 289
 290                         /*
 291                          * Hide thermal monitoring
 292                          */
 293                         regs[3] &= ~(CPUID_ACPI | CPUID_TM);
 294
 295                         /*
 296                          * Hide the debug store capability.
 297                          */
 298                         regs[3] &= ~CPUID_DS;
 299
 300                         /*
 301                          * Advertise the Machine Check and MTRR capability.
 302                          *
 303                          * Some guest OSes (e.g. Windows) will not boot if
 304                          * these features are absent.
 305                          */
 306                         regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR);
 307
 308                         vm_get_topology(vm, &sockets, &cores, &threads,
 309                             &maxcpus);
 310                         logical_cpus = threads * cores;
 311                         regs[1] &= ~CPUID_HTT_CORES;
 312                         regs[1] |= (logical_cpus & 0xff) << 16;
 313                         regs[3] |= CPUID_HTT;
 314                         break;
 315
 316                 case CPUID_0000_0004:
 317                         cpuid_count(*eax, *ecx, regs);
 318
 319                         if (regs[0] || regs[1] || regs[2] || regs[3]) {
 320                                 vm_get_topology(vm, &sockets, &cores, &threads,
 321                                     &maxcpus);
 322                                 regs[0] &= 0x3ff;
 323                                 regs[0] |= (cores - 1) << 26;
 324                                 /*
 325                                  * Cache topology:
 326                                  * - L1 and L2 are shared only by the logical
 327                                  *   processors in a single core.
 328                                  * - L3 and above are shared by all logical
 329                                  *   processors in the package.
 330                                  */
 331                                 logical_cpus = threads;
 332                                 level = (regs[0] >> 5) & 0x7;
 333                                 if (level >= 3)
 334                                         logical_cpus *= cores;
 335                                 regs[0] |= (logical_cpus - 1) << 14;
 336                         }
 337                         break;
 338
 339                 case CPUID_0000_0007:
 340                         regs[0] = 0;
 341                         regs[1] = 0;
 342                         regs[2] = 0;
 343                         regs[3] = 0;
 344
 345                         /* leaf 0 */
 346                         if (*ecx == 0) {
 347                                 cpuid_count(*eax, *ecx, regs);
 348
 349                                 /* Only leaf 0 is supported */
 350                                 regs[0] = 0;
 351
 352                                 /*
 353                                  * Expose known-safe features.
 354                                  */
 355                                 regs[1] &= (CPUID_STDEXT_FSGSBASE |
 356                                     CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE |
 357                                     CPUID_STDEXT_AVX2 | CPUID_STDEXT_BMI2 |
 358                                     CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM |
 359                                     CPUID_STDEXT_AVX512F |
 360                                     CPUID_STDEXT_AVX512PF |
 361                                     CPUID_STDEXT_AVX512ER |
 362                                     CPUID_STDEXT_AVX512CD);
 363                                 regs[2] = 0;
 364                                 regs[3] = 0;
 365
 366                                 /* Advertise INVPCID if it is enabled. */
 367                                 error = vm_get_capability(vm, vcpu_id,
 368                                     VM_CAP_ENABLE_INVPCID, &enable_invpcid);
 369                                 if (error == 0 && enable_invpcid)
 370                                         regs[1] |= CPUID_STDEXT_INVPCID;
 371                         }
 372                         break;
 373
 374                 case CPUID_0000_0006:
 375                         regs[0] = CPUTPM1_ARAT;
 376                         regs[1] = 0;
 377                         regs[2] = 0;
 378                         regs[3] = 0;
 379                         break;
 380
 381                 case CPUID_0000_000A:
 382                         /*
 383                          * Handle the access, but report 0 for
 384                          * all options
 385                          */
 386                         regs[0] = 0;
 387                         regs[1] = 0;
 388                         regs[2] = 0;
 389                         regs[3] = 0;
 390                         break;
 391
 392                 case CPUID_0000_000B:
 393                         /*
 394                          * Processor topology enumeration
 395                          */
 396                         vm_get_topology(vm, &sockets, &cores, &threads,
 397                             &maxcpus);
 398                         if (*ecx == 0) {
 399                                 logical_cpus = threads;
 400                                 width = log2(logical_cpus);
 401                                 level = CPUID_TYPE_SMT;
 402                                 x2apic_id = vcpu_id;
 403                         }
 404
 405                         if (*ecx == 1) {
 406                                 logical_cpus = threads * cores;
 407                                 width = log2(logical_cpus);
 408                                 level = CPUID_TYPE_CORE;
 409                                 x2apic_id = vcpu_id;
 410                         }
 411
 412                         if (!cpuid_leaf_b || *ecx >= 2) {
 413                                 width = 0;
 414                                 logical_cpus = 0;
 415                                 level = 0;
 416                                 x2apic_id = 0;
 417                         }
 418
 419                         regs[0] = width & 0x1f;
 420                         regs[1] = logical_cpus & 0xffff;
 421                         regs[2] = (level << 8) | (*ecx & 0xff);
 422                         regs[3] = x2apic_id;
 423                         break;
 424
 425                 case CPUID_0000_000D:
 426                         limits = vmm_get_xsave_limits();
 427                         if (!limits->xsave_enabled) {
 428                                 regs[0] = 0;
 429                                 regs[1] = 0;
 430                                 regs[2] = 0;
 431                                 regs[3] = 0;
 432                                 break;
 433                         }
 434
 435                         cpuid_count(*eax, *ecx, regs);
 436                         switch (*ecx) {
 437                         case 0:
 438                                 /*
 439                                  * Only permit the guest to use bits
 440                                  * that are active in the host in
 441                                  * %xcr0.  Also, claim that the
 442                                  * maximum save area size is
 443                                  * equivalent to the host's current
 444                                  * save area size.  Since this runs
 445                                  * "inside" of vmrun(), it runs with
 446                                  * the guest's xcr0, so the current
 447                                  * save area size is correct as-is.
 448                                  */
 449                                 regs[0] &= limits->xcr0_allowed;
 450                                 regs[2] = limits->xsave_max_size;
 451                                 regs[3] &= (limits->xcr0_allowed >> 32);
 452                                 break;
 453                         case 1:
 454                                 /* Only permit XSAVEOPT. */
 455                                 regs[0] &= CPUID_EXTSTATE_XSAVEOPT;
 456                                 regs[1] = 0;
 457                                 regs[2] = 0;
 458                                 regs[3] = 0;
 459                                 break;
 460                         default:
 461                                 /*
 462                                  * If the leaf is for a permitted feature,
 463                                  * pass through as-is, otherwise return
 464                                  * all zeroes.
 465                                  */
 466                                 if (!(limits->xcr0_allowed & (1ul << *ecx))) {
 467                                         regs[0] = 0;
 468                                         regs[1] = 0;
 469                                         regs[2] = 0;
 470                                         regs[3] = 0;
 471                                 }
 472                                 break;
 473                         }
 474                         break;
 475
 476                 case 0x40000000:
 477                         regs[0] = CPUID_VM_HIGH;
 478                         bcopy(bhyve_id, &regs[1], 4);
 479                         bcopy(bhyve_id + 4, &regs[2], 4);
 480                         bcopy(bhyve_id + 8, &regs[3], 4);
 481                         break;
 482
 483                 default:
 484                         /*
 485                          * The leaf value has already been clamped so
 486                          * simply pass this through, keeping count of
 487                          * how many unhandled leaf values have been seen.
 488                          */
 489                         atomic_add_long(&bhyve_xcpuids, 1);
 490                         cpuid_count(*eax, *ecx, regs);
 491                         break;
 492         }
 493
 494         *eax = regs[0];
 495         *ebx = regs[1];
 496         *ecx = regs[2];
 497         *edx = regs[3];
 498
 499         return (1);
 500 }
 501
 502 bool
 503 vm_cpuid_capability(struct vm *vm, int vcpuid, enum vm_cpuid_capability cap)
 504 {
 505         bool rv;
 506
 507         KASSERT(cap > 0 && cap < VCC_LAST, ("%s: invalid vm_cpu_capability %d",
 508             __func__, cap));
 509
 510         /*
 511          * Simply passthrough the capabilities of the host cpu for now.
 512          */
 513         rv = false;
 514         switch (cap) {
 515         case VCC_NO_EXECUTE:
 516                 if (amd_feature & AMDID_NX)
 517                         rv = true;
 518                 break;
 519         case VCC_FFXSR:
 520                 if (amd_feature & AMDID_FFXSR)
 521                         rv = true;
 522                 break;
 523         case VCC_TCE:
 524                 if (amd_feature2 & AMDID2_TCE)
 525                         rv = true;
 526                 break;
 527         default:
 528                 panic("%s: unknown vm_cpu_capability %d", __func__, cap);
 529         }
 530         return (rv);
 531 }