sys/amd64/vmm/x86.c

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