target/i386/whpx-all.c

   1 /*
   2  * QEMU Windows Hypervisor Platform accelerator (WHPX)
   3  *
   4  * Copyright Microsoft Corp. 2017
   5  *
   6  * This work is licensed under the terms of the GNU GPL, version 2 or later.
   7  * See the COPYING file in the top-level directory.
   8  *
   9  */
  10
  11 #include "qemu/osdep.h"
  12 #include "cpu.h"
  13 #include "exec/address-spaces.h"
  14 #include "exec/ioport.h"
  15 #include "qemu-common.h"
  16 #include "sysemu/accel.h"
  17 #include "sysemu/whpx.h"
  18 #include "sysemu/cpus.h"
  19 #include "sysemu/runstate.h"
  20 #include "qemu/main-loop.h"
  21 #include "qemu/error-report.h"
  22 #include "qapi/error.h"
  23 #include "migration/blocker.h"
  24 #include "whp-dispatch.h"
  25
  26 #include <WinHvPlatform.h>
  27 #include <WinHvEmulation.h>
  28
  29 struct whpx_state {
  30     uint64_t mem_quota;
  31     WHV_PARTITION_HANDLE partition;
  32 };
  33
  34 static const WHV_REGISTER_NAME whpx_register_names[] = {
  35
  36     /* X64 General purpose registers */
  37     WHvX64RegisterRax,
  38     WHvX64RegisterRcx,
  39     WHvX64RegisterRdx,
  40     WHvX64RegisterRbx,
  41     WHvX64RegisterRsp,
  42     WHvX64RegisterRbp,
  43     WHvX64RegisterRsi,
  44     WHvX64RegisterRdi,
  45     WHvX64RegisterR8,
  46     WHvX64RegisterR9,
  47     WHvX64RegisterR10,
  48     WHvX64RegisterR11,
  49     WHvX64RegisterR12,
  50     WHvX64RegisterR13,
  51     WHvX64RegisterR14,
  52     WHvX64RegisterR15,
  53     WHvX64RegisterRip,
  54     WHvX64RegisterRflags,
  55
  56     /* X64 Segment registers */
  57     WHvX64RegisterEs,
  58     WHvX64RegisterCs,
  59     WHvX64RegisterSs,
  60     WHvX64RegisterDs,
  61     WHvX64RegisterFs,
  62     WHvX64RegisterGs,
  63     WHvX64RegisterLdtr,
  64     WHvX64RegisterTr,
  65
  66     /* X64 Table registers */
  67     WHvX64RegisterIdtr,
  68     WHvX64RegisterGdtr,
  69
  70     /* X64 Control Registers */
  71     WHvX64RegisterCr0,
  72     WHvX64RegisterCr2,
  73     WHvX64RegisterCr3,
  74     WHvX64RegisterCr4,
  75     WHvX64RegisterCr8,
  76
  77     /* X64 Debug Registers */
  78     /*
  79      * WHvX64RegisterDr0,
  80      * WHvX64RegisterDr1,
  81      * WHvX64RegisterDr2,
  82      * WHvX64RegisterDr3,
  83      * WHvX64RegisterDr6,
  84      * WHvX64RegisterDr7,
  85      */
  86
  87     /* X64 Floating Point and Vector Registers */
  88     WHvX64RegisterXmm0,
  89     WHvX64RegisterXmm1,
  90     WHvX64RegisterXmm2,
  91     WHvX64RegisterXmm3,
  92     WHvX64RegisterXmm4,
  93     WHvX64RegisterXmm5,
  94     WHvX64RegisterXmm6,
  95     WHvX64RegisterXmm7,
  96     WHvX64RegisterXmm8,
  97     WHvX64RegisterXmm9,
  98     WHvX64RegisterXmm10,
  99     WHvX64RegisterXmm11,
 100     WHvX64RegisterXmm12,
 101     WHvX64RegisterXmm13,
 102     WHvX64RegisterXmm14,
 103     WHvX64RegisterXmm15,
 104     WHvX64RegisterFpMmx0,
 105     WHvX64RegisterFpMmx1,
 106     WHvX64RegisterFpMmx2,
 107     WHvX64RegisterFpMmx3,
 108     WHvX64RegisterFpMmx4,
 109     WHvX64RegisterFpMmx5,
 110     WHvX64RegisterFpMmx6,
 111     WHvX64RegisterFpMmx7,
 112     WHvX64RegisterFpControlStatus,
 113     WHvX64RegisterXmmControlStatus,
 114
 115     /* X64 MSRs */
 116     WHvX64RegisterTsc,
 117     WHvX64RegisterEfer,
 118 #ifdef TARGET_X86_64
 119     WHvX64RegisterKernelGsBase,
 120 #endif
 121     WHvX64RegisterApicBase,
 122     /* WHvX64RegisterPat, */
 123     WHvX64RegisterSysenterCs,
 124     WHvX64RegisterSysenterEip,
 125     WHvX64RegisterSysenterEsp,
 126     WHvX64RegisterStar,
 127 #ifdef TARGET_X86_64
 128     WHvX64RegisterLstar,
 129     WHvX64RegisterCstar,
 130     WHvX64RegisterSfmask,
 131 #endif
 132
 133     /* Interrupt / Event Registers */
 134     /*
 135      * WHvRegisterPendingInterruption,
 136      * WHvRegisterInterruptState,
 137      * WHvRegisterPendingEvent0,
 138      * WHvRegisterPendingEvent1
 139      * WHvX64RegisterDeliverabilityNotifications,
 140      */
 141 };
 142
 143 struct whpx_register_set {
 144     WHV_REGISTER_VALUE values[RTL_NUMBER_OF(whpx_register_names)];
 145 };
 146
 147 struct whpx_vcpu {
 148     WHV_EMULATOR_HANDLE emulator;
 149     bool window_registered;
 150     bool interruptable;
 151     uint64_t tpr;
 152     uint64_t apic_base;
 153     bool interruption_pending;
 154
 155     /* Must be the last field as it may have a tail */
 156     WHV_RUN_VP_EXIT_CONTEXT exit_ctx;
 157 };
 158
 159 static bool whpx_allowed;
 160 static bool whp_dispatch_initialized;
 161 static HMODULE hWinHvPlatform, hWinHvEmulation;
 162
 163 struct whpx_state whpx_global;
 164 struct WHPDispatch whp_dispatch;
 165
 166
 167 /*
 168  * VP support
 169  */
 170
 171 static struct whpx_vcpu *get_whpx_vcpu(CPUState *cpu)
 172 {
 173     return (struct whpx_vcpu *)cpu->hax_vcpu;
 174 }
 175
 176 static WHV_X64_SEGMENT_REGISTER whpx_seg_q2h(const SegmentCache *qs, int v86,
 177                                              int r86)
 178 {
 179     WHV_X64_SEGMENT_REGISTER hs;
 180     unsigned flags = qs->flags;
 181
 182     hs.Base = qs->base;
 183     hs.Limit = qs->limit;
 184     hs.Selector = qs->selector;
 185
 186     if (v86) {
 187         hs.Attributes = 0;
 188         hs.SegmentType = 3;
 189         hs.Present = 1;
 190         hs.DescriptorPrivilegeLevel = 3;
 191         hs.NonSystemSegment = 1;
 192
 193     } else {
 194         hs.Attributes = (flags >> DESC_TYPE_SHIFT);
 195
 196         if (r86) {
 197             /* hs.Base &= 0xfffff; */
 198         }
 199     }
 200
 201     return hs;
 202 }
 203
 204 static SegmentCache whpx_seg_h2q(const WHV_X64_SEGMENT_REGISTER *hs)
 205 {
 206     SegmentCache qs;
 207
 208     qs.base = hs->Base;
 209     qs.limit = hs->Limit;
 210     qs.selector = hs->Selector;
 211
 212     qs.flags = ((uint32_t)hs->Attributes) << DESC_TYPE_SHIFT;
 213
 214     return qs;
 215 }
 216
 217 static void whpx_set_registers(CPUState *cpu)
 218 {
 219     struct whpx_state *whpx = &whpx_global;
 220     struct whpx_vcpu *vcpu = get_whpx_vcpu(cpu);
 221     struct CPUX86State *env = (CPUArchState *)(cpu->env_ptr);
 222     X86CPU *x86_cpu = X86_CPU(cpu);
 223     struct whpx_register_set vcxt;
 224     HRESULT hr;
 225     int idx;
 226     int idx_next;
 227     int i;
 228     int v86, r86;
 229
 230     assert(cpu_is_stopped(cpu) || qemu_cpu_is_self(cpu));
 231
 232     memset(&vcxt, 0, sizeof(struct whpx_register_set));
 233
 234     v86 = (env->eflags & VM_MASK);
 235     r86 = !(env->cr[0] & CR0_PE_MASK);
 236
 237     vcpu->tpr = cpu_get_apic_tpr(x86_cpu->apic_state);
 238     vcpu->apic_base = cpu_get_apic_base(x86_cpu->apic_state);
 239
 240     idx = 0;
 241
 242     /* Indexes for first 16 registers match between HV and QEMU definitions */
 243     idx_next = 16;
 244     for (idx = 0; idx < CPU_NB_REGS; idx += 1) {
 245         vcxt.values[idx].Reg64 = (uint64_t)env->regs[idx];
 246     }
 247     idx = idx_next;
 248
 249     /* Same goes for RIP and RFLAGS */
 250     assert(whpx_register_names[idx] == WHvX64RegisterRip);
 251     vcxt.values[idx++].Reg64 = env->eip;
 252
 253     assert(whpx_register_names[idx] == WHvX64RegisterRflags);
 254     vcxt.values[idx++].Reg64 = env->eflags;
 255
 256     /* Translate 6+4 segment registers. HV and QEMU order matches  */
 257     assert(idx == WHvX64RegisterEs);
 258     for (i = 0; i < 6; i += 1, idx += 1) {
 259         vcxt.values[idx].Segment = whpx_seg_q2h(&env->segs[i], v86, r86);
 260     }
 261
 262     assert(idx == WHvX64RegisterLdtr);
 263     vcxt.values[idx++].Segment = whpx_seg_q2h(&env->ldt, 0, 0);
 264
 265     assert(idx == WHvX64RegisterTr);
 266     vcxt.values[idx++].Segment = whpx_seg_q2h(&env->tr, 0, 0);
 267
 268     assert(idx == WHvX64RegisterIdtr);
 269     vcxt.values[idx].Table.Base = env->idt.base;
 270     vcxt.values[idx].Table.Limit = env->idt.limit;
 271     idx += 1;
 272
 273     assert(idx == WHvX64RegisterGdtr);
 274     vcxt.values[idx].Table.Base = env->gdt.base;
 275     vcxt.values[idx].Table.Limit = env->gdt.limit;
 276     idx += 1;
 277
 278     /* CR0, 2, 3, 4, 8 */
 279     assert(whpx_register_names[idx] == WHvX64RegisterCr0);
 280     vcxt.values[idx++].Reg64 = env->cr[0];
 281     assert(whpx_register_names[idx] == WHvX64RegisterCr2);
 282     vcxt.values[idx++].Reg64 = env->cr[2];
 283     assert(whpx_register_names[idx] == WHvX64RegisterCr3);
 284     vcxt.values[idx++].Reg64 = env->cr[3];
 285     assert(whpx_register_names[idx] == WHvX64RegisterCr4);
 286     vcxt.values[idx++].Reg64 = env->cr[4];
 287     assert(whpx_register_names[idx] == WHvX64RegisterCr8);
 288     vcxt.values[idx++].Reg64 = vcpu->tpr;
 289
 290     /* 8 Debug Registers - Skipped */
 291
 292     /* 16 XMM registers */
 293     assert(whpx_register_names[idx] == WHvX64RegisterXmm0);
 294     idx_next = idx + 16;
 295     for (i = 0; i < sizeof(env->xmm_regs) / sizeof(ZMMReg); i += 1, idx += 1) {
 296         vcxt.values[idx].Reg128.Low64 = env->xmm_regs[i].ZMM_Q(0);
 297         vcxt.values[idx].Reg128.High64 = env->xmm_regs[i].ZMM_Q(1);
 298     }
 299     idx = idx_next;
 300
 301     /* 8 FP registers */
 302     assert(whpx_register_names[idx] == WHvX64RegisterFpMmx0);
 303     for (i = 0; i < 8; i += 1, idx += 1) {
 304         vcxt.values[idx].Fp.AsUINT128.Low64 = env->fpregs[i].mmx.MMX_Q(0);
 305         /* vcxt.values[idx].Fp.AsUINT128.High64 =
 306                env->fpregs[i].mmx.MMX_Q(1);
 307         */
 308     }
 309
 310     /* FP control status register */
 311     assert(whpx_register_names[idx] == WHvX64RegisterFpControlStatus);
 312     vcxt.values[idx].FpControlStatus.FpControl = env->fpuc;
 313     vcxt.values[idx].FpControlStatus.FpStatus =
 314         (env->fpus & ~0x3800) | (env->fpstt & 0x7) << 11;
 315     vcxt.values[idx].FpControlStatus.FpTag = 0;
 316     for (i = 0; i < 8; ++i) {
 317         vcxt.values[idx].FpControlStatus.FpTag |= (!env->fptags[i]) << i;
 318     }
 319     vcxt.values[idx].FpControlStatus.Reserved = 0;
 320     vcxt.values[idx].FpControlStatus.LastFpOp = env->fpop;
 321     vcxt.values[idx].FpControlStatus.LastFpRip = env->fpip;
 322     idx += 1;
 323
 324     /* XMM control status register */
 325     assert(whpx_register_names[idx] == WHvX64RegisterXmmControlStatus);
 326     vcxt.values[idx].XmmControlStatus.LastFpRdp = 0;
 327     vcxt.values[idx].XmmControlStatus.XmmStatusControl = env->mxcsr;
 328     vcxt.values[idx].XmmControlStatus.XmmStatusControlMask = 0x0000ffff;
 329     idx += 1;
 330
 331     /* MSRs */
 332     assert(whpx_register_names[idx] == WHvX64RegisterTsc);
 333     vcxt.values[idx++].Reg64 = env->tsc;
 334     assert(whpx_register_names[idx] == WHvX64RegisterEfer);
 335     vcxt.values[idx++].Reg64 = env->efer;
 336 #ifdef TARGET_X86_64
 337     assert(whpx_register_names[idx] == WHvX64RegisterKernelGsBase);
 338     vcxt.values[idx++].Reg64 = env->kernelgsbase;
 339 #endif
 340
 341     assert(whpx_register_names[idx] == WHvX64RegisterApicBase);
 342     vcxt.values[idx++].Reg64 = vcpu->apic_base;
 343
 344     /* WHvX64RegisterPat - Skipped */
 345
 346     assert(whpx_register_names[idx] == WHvX64RegisterSysenterCs);
 347     vcxt.values[idx++].Reg64 = env->sysenter_cs;
 348     assert(whpx_register_names[idx] == WHvX64RegisterSysenterEip);
 349     vcxt.values[idx++].Reg64 = env->sysenter_eip;
 350     assert(whpx_register_names[idx] == WHvX64RegisterSysenterEsp);
 351     vcxt.values[idx++].Reg64 = env->sysenter_esp;
 352     assert(whpx_register_names[idx] == WHvX64RegisterStar);
 353     vcxt.values[idx++].Reg64 = env->star;
 354 #ifdef TARGET_X86_64
 355     assert(whpx_register_names[idx] == WHvX64RegisterLstar);
 356     vcxt.values[idx++].Reg64 = env->lstar;
 357     assert(whpx_register_names[idx] == WHvX64RegisterCstar);
 358     vcxt.values[idx++].Reg64 = env->cstar;
 359     assert(whpx_register_names[idx] == WHvX64RegisterSfmask);
 360     vcxt.values[idx++].Reg64 = env->fmask;
 361 #endif
 362
 363     /* Interrupt / Event Registers - Skipped */
 364
 365     assert(idx == RTL_NUMBER_OF(whpx_register_names));
 366
 367     hr = whp_dispatch.WHvSetVirtualProcessorRegisters(
 368         whpx->partition, cpu->cpu_index,
 369         whpx_register_names,
 370         RTL_NUMBER_OF(whpx_register_names),
 371         &vcxt.values[0]);
 372
 373     if (FAILED(hr)) {
 374         error_report("WHPX: Failed to set virtual processor context, hr=%08lx",
 375                      hr);
 376     }
 377
 378     return;
 379 }
 380
 381 static void whpx_get_registers(CPUState *cpu)
 382 {
 383     struct whpx_state *whpx = &whpx_global;
 384     struct whpx_vcpu *vcpu = get_whpx_vcpu(cpu);
 385     struct CPUX86State *env = (CPUArchState *)(cpu->env_ptr);
 386     X86CPU *x86_cpu = X86_CPU(cpu);
 387     struct whpx_register_set vcxt;
 388     uint64_t tpr, apic_base;
 389     HRESULT hr;
 390     int idx;
 391     int idx_next;
 392     int i;
 393
 394     assert(cpu_is_stopped(cpu) || qemu_cpu_is_self(cpu));
 395
 396     hr = whp_dispatch.WHvGetVirtualProcessorRegisters(
 397         whpx->partition, cpu->cpu_index,
 398         whpx_register_names,
 399         RTL_NUMBER_OF(whpx_register_names),
 400         &vcxt.values[0]);
 401     if (FAILED(hr)) {
 402         error_report("WHPX: Failed to get virtual processor context, hr=%08lx",
 403                      hr);
 404     }
 405
 406     idx = 0;
 407
 408     /* Indexes for first 16 registers match between HV and QEMU definitions */
 409     idx_next = 16;
 410     for (idx = 0; idx < CPU_NB_REGS; idx += 1) {
 411         env->regs[idx] = vcxt.values[idx].Reg64;
 412     }
 413     idx = idx_next;
 414
 415     /* Same goes for RIP and RFLAGS */
 416     assert(whpx_register_names[idx] == WHvX64RegisterRip);
 417     env->eip = vcxt.values[idx++].Reg64;
 418     assert(whpx_register_names[idx] == WHvX64RegisterRflags);
 419     env->eflags = vcxt.values[idx++].Reg64;
 420
 421     /* Translate 6+4 segment registers. HV and QEMU order matches  */
 422     assert(idx == WHvX64RegisterEs);
 423     for (i = 0; i < 6; i += 1, idx += 1) {
 424         env->segs[i] = whpx_seg_h2q(&vcxt.values[idx].Segment);
 425     }
 426
 427     assert(idx == WHvX64RegisterLdtr);
 428     env->ldt = whpx_seg_h2q(&vcxt.values[idx++].Segment);
 429     assert(idx == WHvX64RegisterTr);
 430     env->tr = whpx_seg_h2q(&vcxt.values[idx++].Segment);
 431     assert(idx == WHvX64RegisterIdtr);
 432     env->idt.base = vcxt.values[idx].Table.Base;
 433     env->idt.limit = vcxt.values[idx].Table.Limit;
 434     idx += 1;
 435     assert(idx == WHvX64RegisterGdtr);
 436     env->gdt.base = vcxt.values[idx].Table.Base;
 437     env->gdt.limit = vcxt.values[idx].Table.Limit;
 438     idx += 1;
 439
 440     /* CR0, 2, 3, 4, 8 */
 441     assert(whpx_register_names[idx] == WHvX64RegisterCr0);
 442     env->cr[0] = vcxt.values[idx++].Reg64;
 443     assert(whpx_register_names[idx] == WHvX64RegisterCr2);
 444     env->cr[2] = vcxt.values[idx++].Reg64;
 445     assert(whpx_register_names[idx] == WHvX64RegisterCr3);
 446     env->cr[3] = vcxt.values[idx++].Reg64;
 447     assert(whpx_register_names[idx] == WHvX64RegisterCr4);
 448     env->cr[4] = vcxt.values[idx++].Reg64;
 449     assert(whpx_register_names[idx] == WHvX64RegisterCr8);
 450     tpr = vcxt.values[idx++].Reg64;
 451     if (tpr != vcpu->tpr) {
 452         vcpu->tpr = tpr;
 453         cpu_set_apic_tpr(x86_cpu->apic_state, tpr);
 454     }
 455
 456     /* 8 Debug Registers - Skipped */
 457
 458     /* 16 XMM registers */
 459     assert(whpx_register_names[idx] == WHvX64RegisterXmm0);
 460     idx_next = idx + 16;
 461     for (i = 0; i < sizeof(env->xmm_regs) / sizeof(ZMMReg); i += 1, idx += 1) {
 462         env->xmm_regs[i].ZMM_Q(0) = vcxt.values[idx].Reg128.Low64;
 463         env->xmm_regs[i].ZMM_Q(1) = vcxt.values[idx].Reg128.High64;
 464     }
 465     idx = idx_next;
 466
 467     /* 8 FP registers */
 468     assert(whpx_register_names[idx] == WHvX64RegisterFpMmx0);
 469     for (i = 0; i < 8; i += 1, idx += 1) {
 470         env->fpregs[i].mmx.MMX_Q(0) = vcxt.values[idx].Fp.AsUINT128.Low64;
 471         /* env->fpregs[i].mmx.MMX_Q(1) =
 472                vcxt.values[idx].Fp.AsUINT128.High64;
 473         */
 474     }
 475
 476     /* FP control status register */
 477     assert(whpx_register_names[idx] == WHvX64RegisterFpControlStatus);
 478     env->fpuc = vcxt.values[idx].FpControlStatus.FpControl;
 479     env->fpstt = (vcxt.values[idx].FpControlStatus.FpStatus >> 11) & 0x7;
 480     env->fpus = vcxt.values[idx].FpControlStatus.FpStatus & ~0x3800;
 481     for (i = 0; i < 8; ++i) {
 482         env->fptags[i] = !((vcxt.values[idx].FpControlStatus.FpTag >> i) & 1);
 483     }
 484     env->fpop = vcxt.values[idx].FpControlStatus.LastFpOp;
 485     env->fpip = vcxt.values[idx].FpControlStatus.LastFpRip;
 486     idx += 1;
 487
 488     /* XMM control status register */
 489     assert(whpx_register_names[idx] == WHvX64RegisterXmmControlStatus);
 490     env->mxcsr = vcxt.values[idx].XmmControlStatus.XmmStatusControl;
 491     idx += 1;
 492
 493     /* MSRs */
 494     assert(whpx_register_names[idx] == WHvX64RegisterTsc);
 495     env->tsc = vcxt.values[idx++].Reg64;
 496     assert(whpx_register_names[idx] == WHvX64RegisterEfer);
 497     env->efer = vcxt.values[idx++].Reg64;
 498 #ifdef TARGET_X86_64
 499     assert(whpx_register_names[idx] == WHvX64RegisterKernelGsBase);
 500     env->kernelgsbase = vcxt.values[idx++].Reg64;
 501 #endif
 502
 503     assert(whpx_register_names[idx] == WHvX64RegisterApicBase);
 504     apic_base = vcxt.values[idx++].Reg64;
 505     if (apic_base != vcpu->apic_base) {
 506         vcpu->apic_base = apic_base;
 507         cpu_set_apic_base(x86_cpu->apic_state, vcpu->apic_base);
 508     }
 509
 510     /* WHvX64RegisterPat - Skipped */
 511
 512     assert(whpx_register_names[idx] == WHvX64RegisterSysenterCs);
 513     env->sysenter_cs = vcxt.values[idx++].Reg64;;
 514     assert(whpx_register_names[idx] == WHvX64RegisterSysenterEip);
 515     env->sysenter_eip = vcxt.values[idx++].Reg64;
 516     assert(whpx_register_names[idx] == WHvX64RegisterSysenterEsp);
 517     env->sysenter_esp = vcxt.values[idx++].Reg64;
 518     assert(whpx_register_names[idx] == WHvX64RegisterStar);
 519     env->star = vcxt.values[idx++].Reg64;
 520 #ifdef TARGET_X86_64
 521     assert(whpx_register_names[idx] == WHvX64RegisterLstar);
 522     env->lstar = vcxt.values[idx++].Reg64;
 523     assert(whpx_register_names[idx] == WHvX64RegisterCstar);
 524     env->cstar = vcxt.values[idx++].Reg64;
 525     assert(whpx_register_names[idx] == WHvX64RegisterSfmask);
 526     env->fmask = vcxt.values[idx++].Reg64;
 527 #endif
 528
 529     /* Interrupt / Event Registers - Skipped */
 530
 531     assert(idx == RTL_NUMBER_OF(whpx_register_names));
 532
 533     return;
 534 }
 535
 536 static HRESULT CALLBACK whpx_emu_ioport_callback(
 537     void *ctx,
 538     WHV_EMULATOR_IO_ACCESS_INFO *IoAccess)
 539 {
 540     MemTxAttrs attrs = { 0 };
 541     address_space_rw(&address_space_io, IoAccess->Port, attrs,
 542                      (uint8_t *)&IoAccess->Data, IoAccess->AccessSize,
 543                      IoAccess->Direction);
 544     return S_OK;
 545 }
 546
 547 static HRESULT CALLBACK whpx_emu_mmio_callback(
 548     void *ctx,
 549     WHV_EMULATOR_MEMORY_ACCESS_INFO *ma)
 550 {
 551     cpu_physical_memory_rw(ma->GpaAddress, ma->Data, ma->AccessSize,
 552                            ma->Direction);
 553     return S_OK;
 554 }
 555
 556 static HRESULT CALLBACK whpx_emu_getreg_callback(
 557     void *ctx,
 558     const WHV_REGISTER_NAME *RegisterNames,
 559     UINT32 RegisterCount,
 560     WHV_REGISTER_VALUE *RegisterValues)
 561 {
 562     HRESULT hr;
 563     struct whpx_state *whpx = &whpx_global;
 564     CPUState *cpu = (CPUState *)ctx;
 565
 566     hr = whp_dispatch.WHvGetVirtualProcessorRegisters(
 567         whpx->partition, cpu->cpu_index,
 568         RegisterNames, RegisterCount,
 569         RegisterValues);
 570     if (FAILED(hr)) {
 571         error_report("WHPX: Failed to get virtual processor registers,"
 572                      " hr=%08lx", hr);
 573     }
 574
 575     return hr;
 576 }
 577
 578 static HRESULT CALLBACK whpx_emu_setreg_callback(
 579     void *ctx,
 580     const WHV_REGISTER_NAME *RegisterNames,
 581     UINT32 RegisterCount,
 582     const WHV_REGISTER_VALUE *RegisterValues)
 583 {
 584     HRESULT hr;
 585     struct whpx_state *whpx = &whpx_global;
 586     CPUState *cpu = (CPUState *)ctx;
 587
 588     hr = whp_dispatch.WHvSetVirtualProcessorRegisters(
 589         whpx->partition, cpu->cpu_index,
 590         RegisterNames, RegisterCount,
 591         RegisterValues);
 592     if (FAILED(hr)) {
 593         error_report("WHPX: Failed to set virtual processor registers,"
 594                      " hr=%08lx", hr);
 595     }
 596
 597     /*
 598      * The emulator just successfully wrote the register state. We clear the
 599      * dirty state so we avoid the double write on resume of the VP.
 600      */
 601     cpu->vcpu_dirty = false;
 602
 603     return hr;
 604 }
 605
 606 static HRESULT CALLBACK whpx_emu_translate_callback(
 607     void *ctx,
 608     WHV_GUEST_VIRTUAL_ADDRESS Gva,
 609     WHV_TRANSLATE_GVA_FLAGS TranslateFlags,
 610     WHV_TRANSLATE_GVA_RESULT_CODE *TranslationResult,
 611     WHV_GUEST_PHYSICAL_ADDRESS *Gpa)
 612 {
 613     HRESULT hr;
 614     struct whpx_state *whpx = &whpx_global;
 615     CPUState *cpu = (CPUState *)ctx;
 616     WHV_TRANSLATE_GVA_RESULT res;
 617
 618     hr = whp_dispatch.WHvTranslateGva(whpx->partition, cpu->cpu_index,
 619                                       Gva, TranslateFlags, &res, Gpa);
 620     if (FAILED(hr)) {
 621         error_report("WHPX: Failed to translate GVA, hr=%08lx", hr);
 622     } else {
 623         *TranslationResult = res.ResultCode;
 624     }
 625
 626     return hr;
 627 }
 628
 629 static const WHV_EMULATOR_CALLBACKS whpx_emu_callbacks = {
 630     .Size = sizeof(WHV_EMULATOR_CALLBACKS),
 631     .WHvEmulatorIoPortCallback = whpx_emu_ioport_callback,
 632     .WHvEmulatorMemoryCallback = whpx_emu_mmio_callback,
 633     .WHvEmulatorGetVirtualProcessorRegisters = whpx_emu_getreg_callback,
 634     .WHvEmulatorSetVirtualProcessorRegisters = whpx_emu_setreg_callback,
 635     .WHvEmulatorTranslateGvaPage = whpx_emu_translate_callback,
 636 };
 637
 638 static int whpx_handle_mmio(CPUState *cpu, WHV_MEMORY_ACCESS_CONTEXT *ctx)
 639 {
 640     HRESULT hr;
 641     struct whpx_vcpu *vcpu = get_whpx_vcpu(cpu);
 642     WHV_EMULATOR_STATUS emu_status;
 643
 644     hr = whp_dispatch.WHvEmulatorTryMmioEmulation(
 645         vcpu->emulator, cpu,
 646         &vcpu->exit_ctx.VpContext, ctx,
 647         &emu_status);
 648     if (FAILED(hr)) {
 649         error_report("WHPX: Failed to parse MMIO access, hr=%08lx", hr);
 650         return -1;
 651     }
 652
 653     if (!emu_status.EmulationSuccessful) {
 654         error_report("WHPX: Failed to emulate MMIO access with"
 655                      " EmulatorReturnStatus: %u", emu_status.AsUINT32);
 656         return -1;
 657     }
 658
 659     return 0;
 660 }
 661
 662 static int whpx_handle_portio(CPUState *cpu,
 663                               WHV_X64_IO_PORT_ACCESS_CONTEXT *ctx)
 664 {
 665     HRESULT hr;
 666     struct whpx_vcpu *vcpu = get_whpx_vcpu(cpu);
 667     WHV_EMULATOR_STATUS emu_status;
 668
 669     hr = whp_dispatch.WHvEmulatorTryIoEmulation(
 670         vcpu->emulator, cpu,
 671         &vcpu->exit_ctx.VpContext, ctx,
 672         &emu_status);
 673     if (FAILED(hr)) {
 674         error_report("WHPX: Failed to parse PortIO access, hr=%08lx", hr);
 675         return -1;
 676     }
 677
 678     if (!emu_status.EmulationSuccessful) {
 679         error_report("WHPX: Failed to emulate PortIO access with"
 680                      " EmulatorReturnStatus: %u", emu_status.AsUINT32);
 681         return -1;
 682     }
 683
 684     return 0;
 685 }
 686
 687 static int whpx_handle_halt(CPUState *cpu)
 688 {
 689     struct CPUX86State *env = (CPUArchState *)(cpu->env_ptr);
 690     int ret = 0;
 691
 692     qemu_mutex_lock_iothread();
 693     if (!((cpu->interrupt_request & CPU_INTERRUPT_HARD) &&
 694           (env->eflags & IF_MASK)) &&
 695         !(cpu->interrupt_request & CPU_INTERRUPT_NMI)) {
 696         cpu->exception_index = EXCP_HLT;
 697         cpu->halted = true;
 698         ret = 1;
 699     }
 700     qemu_mutex_unlock_iothread();
 701
 702     return ret;
 703 }
 704
 705 static void whpx_vcpu_pre_run(CPUState *cpu)
 706 {
 707     HRESULT hr;
 708     struct whpx_state *whpx = &whpx_global;
 709     struct whpx_vcpu *vcpu = get_whpx_vcpu(cpu);
 710     struct CPUX86State *env = (CPUArchState *)(cpu->env_ptr);
 711     X86CPU *x86_cpu = X86_CPU(cpu);
 712     int irq;
 713     uint8_t tpr;
 714     WHV_X64_PENDING_INTERRUPTION_REGISTER new_int;
 715     UINT32 reg_count = 0;
 716     WHV_REGISTER_VALUE reg_values[3];
 717     WHV_REGISTER_NAME reg_names[3];
 718
 719     memset(&new_int, 0, sizeof(new_int));
 720     memset(reg_values, 0, sizeof(reg_values));
 721
 722     qemu_mutex_lock_iothread();
 723
 724     /* Inject NMI */
 725     if (!vcpu->interruption_pending &&
 726         cpu->interrupt_request & (CPU_INTERRUPT_NMI | CPU_INTERRUPT_SMI)) {
 727         if (cpu->interrupt_request & CPU_INTERRUPT_NMI) {
 728             cpu->interrupt_request &= ~CPU_INTERRUPT_NMI;
 729             vcpu->interruptable = false;
 730             new_int.InterruptionType = WHvX64PendingNmi;
 731             new_int.InterruptionPending = 1;
 732             new_int.InterruptionVector = 2;
 733         }
 734         if (cpu->interrupt_request & CPU_INTERRUPT_SMI) {
 735             cpu->interrupt_request &= ~CPU_INTERRUPT_SMI;
 736         }
 737     }
 738
 739     /*
 740      * Force the VCPU out of its inner loop to process any INIT requests or
 741      * commit pending TPR access.
 742      */
 743     if (cpu->interrupt_request & (CPU_INTERRUPT_INIT | CPU_INTERRUPT_TPR)) {
 744         if ((cpu->interrupt_request & CPU_INTERRUPT_INIT) &&
 745             !(env->hflags & HF_SMM_MASK)) {
 746             cpu->exit_request = 1;
 747         }
 748         if (cpu->interrupt_request & CPU_INTERRUPT_TPR) {
 749             cpu->exit_request = 1;
 750         }
 751     }
 752
 753     /* Get pending hard interruption or replay one that was overwritten */
 754     if (!vcpu->interruption_pending &&
 755         vcpu->interruptable && (env->eflags & IF_MASK)) {
 756         assert(!new_int.InterruptionPending);
 757         if (cpu->interrupt_request & CPU_INTERRUPT_HARD) {
 758             cpu->interrupt_request &= ~CPU_INTERRUPT_HARD;
 759             irq = cpu_get_pic_interrupt(env);
 760             if (irq >= 0) {
 761                 new_int.InterruptionType = WHvX64PendingInterrupt;
 762                 new_int.InterruptionPending = 1;
 763                 new_int.InterruptionVector = irq;
 764             }
 765         }
 766     }
 767
 768     /* Setup interrupt state if new one was prepared */
 769     if (new_int.InterruptionPending) {
 770         reg_values[reg_count].PendingInterruption = new_int;
 771         reg_names[reg_count] = WHvRegisterPendingInterruption;
 772         reg_count += 1;
 773     }
 774
 775     /* Sync the TPR to the CR8 if was modified during the intercept */
 776     tpr = cpu_get_apic_tpr(x86_cpu->apic_state);
 777     if (tpr != vcpu->tpr) {
 778         vcpu->tpr = tpr;
 779         reg_values[reg_count].Reg64 = tpr;
 780         cpu->exit_request = 1;
 781         reg_names[reg_count] = WHvX64RegisterCr8;
 782         reg_count += 1;
 783     }
 784
 785     /* Update the state of the interrupt delivery notification */
 786     if (!vcpu->window_registered &&
 787         cpu->interrupt_request & CPU_INTERRUPT_HARD) {
 788         reg_values[reg_count].DeliverabilityNotifications.InterruptNotification
 789             = 1;
 790         vcpu->window_registered = 1;
 791         reg_names[reg_count] = WHvX64RegisterDeliverabilityNotifications;
 792         reg_count += 1;
 793     }
 794
 795     qemu_mutex_unlock_iothread();
 796
 797     if (reg_count) {
 798         hr = whp_dispatch.WHvSetVirtualProcessorRegisters(
 799             whpx->partition, cpu->cpu_index,
 800             reg_names, reg_count, reg_values);
 801         if (FAILED(hr)) {
 802             error_report("WHPX: Failed to set interrupt state registers,"
 803                          " hr=%08lx", hr);
 804         }
 805     }
 806
 807     return;
 808 }
 809
 810 static void whpx_vcpu_post_run(CPUState *cpu)
 811 {
 812     struct whpx_vcpu *vcpu = get_whpx_vcpu(cpu);
 813     struct CPUX86State *env = (CPUArchState *)(cpu->env_ptr);
 814     X86CPU *x86_cpu = X86_CPU(cpu);
 815
 816     env->eflags = vcpu->exit_ctx.VpContext.Rflags;
 817
 818     uint64_t tpr = vcpu->exit_ctx.VpContext.Cr8;
 819     if (vcpu->tpr != tpr) {
 820         vcpu->tpr = tpr;
 821         qemu_mutex_lock_iothread();
 822         cpu_set_apic_tpr(x86_cpu->apic_state, vcpu->tpr);
 823         qemu_mutex_unlock_iothread();
 824     }
 825
 826     vcpu->interruption_pending =
 827         vcpu->exit_ctx.VpContext.ExecutionState.InterruptionPending;
 828
 829     vcpu->interruptable =
 830         !vcpu->exit_ctx.VpContext.ExecutionState.InterruptShadow;
 831
 832     return;
 833 }
 834
 835 static void whpx_vcpu_process_async_events(CPUState *cpu)
 836 {
 837     struct CPUX86State *env = (CPUArchState *)(cpu->env_ptr);
 838     X86CPU *x86_cpu = X86_CPU(cpu);
 839     struct whpx_vcpu *vcpu = get_whpx_vcpu(cpu);
 840
 841     if ((cpu->interrupt_request & CPU_INTERRUPT_INIT) &&
 842         !(env->hflags & HF_SMM_MASK)) {
 843
 844         do_cpu_init(x86_cpu);
 845         cpu->vcpu_dirty = true;
 846         vcpu->interruptable = true;
 847     }
 848
 849     if (cpu->interrupt_request & CPU_INTERRUPT_POLL) {
 850         cpu->interrupt_request &= ~CPU_INTERRUPT_POLL;
 851         apic_poll_irq(x86_cpu->apic_state);
 852     }
 853
 854     if (((cpu->interrupt_request & CPU_INTERRUPT_HARD) &&
 855          (env->eflags & IF_MASK)) ||
 856         (cpu->interrupt_request & CPU_INTERRUPT_NMI)) {
 857         cpu->halted = false;
 858     }
 859
 860     if (cpu->interrupt_request & CPU_INTERRUPT_SIPI) {
 861         if (!cpu->vcpu_dirty) {
 862             whpx_get_registers(cpu);
 863         }
 864         do_cpu_sipi(x86_cpu);
 865     }
 866
 867     if (cpu->interrupt_request & CPU_INTERRUPT_TPR) {
 868         cpu->interrupt_request &= ~CPU_INTERRUPT_TPR;
 869         if (!cpu->vcpu_dirty) {
 870             whpx_get_registers(cpu);
 871         }
 872         apic_handle_tpr_access_report(x86_cpu->apic_state, env->eip,
 873                                       env->tpr_access_type);
 874     }
 875
 876     return;
 877 }
 878
 879 static int whpx_vcpu_run(CPUState *cpu)
 880 {
 881     HRESULT hr;
 882     struct whpx_state *whpx = &whpx_global;
 883     struct whpx_vcpu *vcpu = get_whpx_vcpu(cpu);
 884     int ret;
 885
 886     whpx_vcpu_process_async_events(cpu);
 887     if (cpu->halted) {
 888         cpu->exception_index = EXCP_HLT;
 889         atomic_set(&cpu->exit_request, false);
 890         return 0;
 891     }
 892
 893     qemu_mutex_unlock_iothread();
 894     cpu_exec_start(cpu);
 895
 896     do {
 897         if (cpu->vcpu_dirty) {
 898             whpx_set_registers(cpu);
 899             cpu->vcpu_dirty = false;
 900         }
 901
 902         whpx_vcpu_pre_run(cpu);
 903
 904         if (atomic_read(&cpu->exit_request)) {
 905             whpx_vcpu_kick(cpu);
 906         }
 907
 908         hr = whp_dispatch.WHvRunVirtualProcessor(
 909             whpx->partition, cpu->cpu_index,
 910             &vcpu->exit_ctx, sizeof(vcpu->exit_ctx));
 911
 912         if (FAILED(hr)) {
 913             error_report("WHPX: Failed to exec a virtual processor,"
 914                          " hr=%08lx", hr);
 915             ret = -1;
 916             break;
 917         }
 918
 919         whpx_vcpu_post_run(cpu);
 920
 921         switch (vcpu->exit_ctx.ExitReason) {
 922         case WHvRunVpExitReasonMemoryAccess:
 923             ret = whpx_handle_mmio(cpu, &vcpu->exit_ctx.MemoryAccess);
 924             break;
 925
 926         case WHvRunVpExitReasonX64IoPortAccess:
 927             ret = whpx_handle_portio(cpu, &vcpu->exit_ctx.IoPortAccess);
 928             break;
 929
 930         case WHvRunVpExitReasonX64InterruptWindow:
 931             vcpu->window_registered = 0;
 932             ret = 0;
 933             break;
 934
 935         case WHvRunVpExitReasonX64Halt:
 936             ret = whpx_handle_halt(cpu);
 937             break;
 938
 939         case WHvRunVpExitReasonCanceled:
 940             cpu->exception_index = EXCP_INTERRUPT;
 941             ret = 1;
 942             break;
 943
 944         case WHvRunVpExitReasonX64MsrAccess: {
 945             WHV_REGISTER_VALUE reg_values[3] = {0};
 946             WHV_REGISTER_NAME reg_names[3];
 947             UINT32 reg_count;
 948
 949             reg_names[0] = WHvX64RegisterRip;
 950             reg_names[1] = WHvX64RegisterRax;
 951             reg_names[2] = WHvX64RegisterRdx;
 952
 953             reg_values[0].Reg64 =
 954                 vcpu->exit_ctx.VpContext.Rip +
 955                 vcpu->exit_ctx.VpContext.InstructionLength;
 956
 957             /*
 958              * For all unsupported MSR access we:
 959              *     ignore writes
 960              *     return 0 on read.
 961              */
 962             reg_count = vcpu->exit_ctx.MsrAccess.AccessInfo.IsWrite ?
 963                         1 : 3;
 964
 965             hr = whp_dispatch.WHvSetVirtualProcessorRegisters(
 966                 whpx->partition,
 967                 cpu->cpu_index,
 968                 reg_names, reg_count,
 969                 reg_values);
 970
 971             if (FAILED(hr)) {
 972                 error_report("WHPX: Failed to set MsrAccess state "
 973                              " registers, hr=%08lx", hr);
 974             }
 975             ret = 0;
 976             break;
 977         }
 978         case WHvRunVpExitReasonX64Cpuid: {
 979             WHV_REGISTER_VALUE reg_values[5];
 980             WHV_REGISTER_NAME reg_names[5];
 981             UINT32 reg_count = 5;
 982             UINT64 rip, rax, rcx, rdx, rbx;
 983
 984             memset(reg_values, 0, sizeof(reg_values));
 985
 986             rip = vcpu->exit_ctx.VpContext.Rip +
 987                   vcpu->exit_ctx.VpContext.InstructionLength;
 988             switch (vcpu->exit_ctx.CpuidAccess.Rax) {
 989             case 1:
 990                 rax = vcpu->exit_ctx.CpuidAccess.DefaultResultRax;
 991                 /* Advertise that we are running on a hypervisor */
 992                 rcx =
 993                     vcpu->exit_ctx.CpuidAccess.DefaultResultRcx |
 994                     CPUID_EXT_HYPERVISOR;
 995
 996                 rdx = vcpu->exit_ctx.CpuidAccess.DefaultResultRdx;
 997                 rbx = vcpu->exit_ctx.CpuidAccess.DefaultResultRbx;
 998                 break;
 999             case 0x80000001:
1000                 rax = vcpu->exit_ctx.CpuidAccess.DefaultResultRax;
1001                 /* Remove any support of OSVW */
1002                 rcx =
1003                     vcpu->exit_ctx.CpuidAccess.DefaultResultRcx &
1004                     ~CPUID_EXT3_OSVW;
1005
1006                 rdx = vcpu->exit_ctx.CpuidAccess.DefaultResultRdx;
1007                 rbx = vcpu->exit_ctx.CpuidAccess.DefaultResultRbx;
1008                 break;
1009             default:
1010                 rax = vcpu->exit_ctx.CpuidAccess.DefaultResultRax;
1011                 rcx = vcpu->exit_ctx.CpuidAccess.DefaultResultRcx;
1012                 rdx = vcpu->exit_ctx.CpuidAccess.DefaultResultRdx;
1013                 rbx = vcpu->exit_ctx.CpuidAccess.DefaultResultRbx;
1014             }
1015
1016             reg_names[0] = WHvX64RegisterRip;
1017             reg_names[1] = WHvX64RegisterRax;
1018             reg_names[2] = WHvX64RegisterRcx;
1019             reg_names[3] = WHvX64RegisterRdx;
1020             reg_names[4] = WHvX64RegisterRbx;
1021
1022             reg_values[0].Reg64 = rip;
1023             reg_values[1].Reg64 = rax;
1024             reg_values[2].Reg64 = rcx;
1025             reg_values[3].Reg64 = rdx;
1026             reg_values[4].Reg64 = rbx;
1027
1028             hr = whp_dispatch.WHvSetVirtualProcessorRegisters(
1029                 whpx->partition, cpu->cpu_index,
1030                 reg_names,
1031                 reg_count,
1032                 reg_values);
1033
1034             if (FAILED(hr)) {
1035                 error_report("WHPX: Failed to set CpuidAccess state registers,"
1036                              " hr=%08lx", hr);
1037             }
1038             ret = 0;
1039             break;
1040         }
1041         case WHvRunVpExitReasonNone:
1042         case WHvRunVpExitReasonUnrecoverableException:
1043         case WHvRunVpExitReasonInvalidVpRegisterValue:
1044         case WHvRunVpExitReasonUnsupportedFeature:
1045         case WHvRunVpExitReasonException:
1046         default:
1047             error_report("WHPX: Unexpected VP exit code %d",
1048                          vcpu->exit_ctx.ExitReason);
1049             whpx_get_registers(cpu);
1050             qemu_mutex_lock_iothread();
1051             qemu_system_guest_panicked(cpu_get_crash_info(cpu));
1052             qemu_mutex_unlock_iothread();
1053             break;
1054         }
1055
1056     } while (!ret);
1057
1058     cpu_exec_end(cpu);
1059     qemu_mutex_lock_iothread();
1060     current_cpu = cpu;
1061
1062     atomic_set(&cpu->exit_request, false);
1063
1064     return ret < 0;
1065 }
1066
1067 static void do_whpx_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
1068 {
1069     whpx_get_registers(cpu);
1070     cpu->vcpu_dirty = true;
1071 }
1072
1073 static void do_whpx_cpu_synchronize_post_reset(CPUState *cpu,
1074                                                run_on_cpu_data arg)
1075 {
1076     whpx_set_registers(cpu);
1077     cpu->vcpu_dirty = false;
1078 }
1079
1080 static void do_whpx_cpu_synchronize_post_init(CPUState *cpu,
1081                                               run_on_cpu_data arg)
1082 {
1083     whpx_set_registers(cpu);
1084     cpu->vcpu_dirty = false;
1085 }
1086
1087 static void do_whpx_cpu_synchronize_pre_loadvm(CPUState *cpu,
1088                                                run_on_cpu_data arg)
1089 {
1090     cpu->vcpu_dirty = true;
1091 }
1092
1093 /*
1094  * CPU support.
1095  */
1096
1097 void whpx_cpu_synchronize_state(CPUState *cpu)
1098 {
1099     if (!cpu->vcpu_dirty) {
1100         run_on_cpu(cpu, do_whpx_cpu_synchronize_state, RUN_ON_CPU_NULL);
1101     }
1102 }
1103
1104 void whpx_cpu_synchronize_post_reset(CPUState *cpu)
1105 {
1106     run_on_cpu(cpu, do_whpx_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
1107 }
1108
1109 void whpx_cpu_synchronize_post_init(CPUState *cpu)
1110 {
1111     run_on_cpu(cpu, do_whpx_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
1112 }
1113
1114 void whpx_cpu_synchronize_pre_loadvm(CPUState *cpu)
1115 {
1116     run_on_cpu(cpu, do_whpx_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
1117 }
1118
1119 /*
1120  * Vcpu support.
1121  */
1122
1123 static Error *whpx_migration_blocker;
1124
1125 int whpx_init_vcpu(CPUState *cpu)
1126 {
1127     HRESULT hr;
1128     struct whpx_state *whpx = &whpx_global;
1129     struct whpx_vcpu *vcpu;
1130     Error *local_error = NULL;
1131
1132     /* Add migration blockers for all unsupported features of the
1133      * Windows Hypervisor Platform
1134      */
1135     if (whpx_migration_blocker == NULL) {
1136         error_setg(&whpx_migration_blocker,
1137                "State blocked due to non-migratable CPUID feature support,"
1138                "dirty memory tracking support, and XSAVE/XRSTOR support");
1139
1140         (void)migrate_add_blocker(whpx_migration_blocker, &local_error);
1141         if (local_error) {
1142             error_report_err(local_error);
1143             migrate_del_blocker(whpx_migration_blocker);
1144             error_free(whpx_migration_blocker);
1145             return -EINVAL;
1146         }
1147     }
1148
1149     vcpu = g_malloc0(sizeof(struct whpx_vcpu));
1150
1151     if (!vcpu) {
1152         error_report("WHPX: Failed to allocte VCPU context.");
1153         return -ENOMEM;
1154     }
1155
1156     hr = whp_dispatch.WHvEmulatorCreateEmulator(
1157         &whpx_emu_callbacks,
1158         &vcpu->emulator);
1159     if (FAILED(hr)) {
1160         error_report("WHPX: Failed to setup instruction completion support,"
1161                      " hr=%08lx", hr);
1162         g_free(vcpu);
1163         return -EINVAL;
1164     }
1165
1166     hr = whp_dispatch.WHvCreateVirtualProcessor(
1167         whpx->partition, cpu->cpu_index, 0);
1168     if (FAILED(hr)) {
1169         error_report("WHPX: Failed to create a virtual processor,"
1170                      " hr=%08lx", hr);
1171         whp_dispatch.WHvEmulatorDestroyEmulator(vcpu->emulator);
1172         g_free(vcpu);
1173         return -EINVAL;
1174     }
1175
1176     vcpu->interruptable = true;
1177
1178     cpu->vcpu_dirty = true;
1179     cpu->hax_vcpu = (struct hax_vcpu_state *)vcpu;
1180
1181     return 0;
1182 }
1183
1184 int whpx_vcpu_exec(CPUState *cpu)
1185 {
1186     int ret;
1187     int fatal;
1188
1189     for (;;) {
1190         if (cpu->exception_index >= EXCP_INTERRUPT) {
1191             ret = cpu->exception_index;
1192             cpu->exception_index = -1;
1193             break;
1194         }
1195
1196         fatal = whpx_vcpu_run(cpu);
1197
1198         if (fatal) {
1199             error_report("WHPX: Failed to exec a virtual processor");
1200             abort();
1201         }
1202     }
1203
1204     return ret;
1205 }
1206
1207 void whpx_destroy_vcpu(CPUState *cpu)
1208 {
1209     struct whpx_state *whpx = &whpx_global;
1210     struct whpx_vcpu *vcpu = get_whpx_vcpu(cpu);
1211
1212     whp_dispatch.WHvDeleteVirtualProcessor(whpx->partition, cpu->cpu_index);
1213     whp_dispatch.WHvEmulatorDestroyEmulator(vcpu->emulator);
1214     g_free(cpu->hax_vcpu);
1215     return;
1216 }
1217
1218 void whpx_vcpu_kick(CPUState *cpu)
1219 {
1220     struct whpx_state *whpx = &whpx_global;
1221     whp_dispatch.WHvCancelRunVirtualProcessor(
1222         whpx->partition, cpu->cpu_index, 0);
1223 }
1224
1225 /*
1226  * Memory support.
1227  */
1228
1229 static void whpx_update_mapping(hwaddr start_pa, ram_addr_t size,
1230                                 void *host_va, int add, int rom,
1231                                 const char *name)
1232 {
1233     struct whpx_state *whpx = &whpx_global;
1234     HRESULT hr;
1235
1236     /*
1237     if (add) {
1238         printf("WHPX: ADD PA:%p Size:%p, Host:%p, %s, '%s'\n",
1239                (void*)start_pa, (void*)size, host_va,
1240                (rom ? "ROM" : "RAM"), name);
1241     } else {
1242         printf("WHPX: DEL PA:%p Size:%p, Host:%p,      '%s'\n",
1243                (void*)start_pa, (void*)size, host_va, name);
1244     }
1245     */
1246
1247     if (add) {
1248         hr = whp_dispatch.WHvMapGpaRange(whpx->partition,
1249                                          host_va,
1250                                          start_pa,
1251                                          size,
1252                                          (WHvMapGpaRangeFlagRead |
1253                                           WHvMapGpaRangeFlagExecute |
1254                                           (rom ? 0 : WHvMapGpaRangeFlagWrite)));
1255     } else {
1256         hr = whp_dispatch.WHvUnmapGpaRange(whpx->partition,
1257                                            start_pa,
1258                                            size);
1259     }
1260
1261     if (FAILED(hr)) {
1262         error_report("WHPX: Failed to %s GPA range '%s' PA:%p, Size:%p bytes,"
1263                      " Host:%p, hr=%08lx",
1264                      (add ? "MAP" : "UNMAP"), name,
1265                      (void *)(uintptr_t)start_pa, (void *)size, host_va, hr);
1266     }
1267 }
1268
1269 static void whpx_process_section(MemoryRegionSection *section, int add)
1270 {
1271     MemoryRegion *mr = section->mr;
1272     hwaddr start_pa = section->offset_within_address_space;
1273     ram_addr_t size = int128_get64(section->size);
1274     unsigned int delta;
1275     uint64_t host_va;
1276
1277     if (!memory_region_is_ram(mr)) {
1278         return;
1279     }
1280
1281     delta = qemu_real_host_page_size - (start_pa & ~qemu_real_host_page_mask);
1282     delta &= ~qemu_real_host_page_mask;
1283     if (delta > size) {
1284         return;
1285     }
1286     start_pa += delta;
1287     size -= delta;
1288     size &= qemu_real_host_page_mask;
1289     if (!size || (start_pa & ~qemu_real_host_page_mask)) {
1290         return;
1291     }
1292
1293     host_va = (uintptr_t)memory_region_get_ram_ptr(mr)
1294             + section->offset_within_region + delta;
1295
1296     whpx_update_mapping(start_pa, size, (void *)(uintptr_t)host_va, add,
1297                         memory_region_is_rom(mr), mr->name);
1298 }
1299
1300 static void whpx_region_add(MemoryListener *listener,
1301                            MemoryRegionSection *section)
1302 {
1303     memory_region_ref(section->mr);
1304     whpx_process_section(section, 1);
1305 }
1306
1307 static void whpx_region_del(MemoryListener *listener,
1308                            MemoryRegionSection *section)
1309 {
1310     whpx_process_section(section, 0);
1311     memory_region_unref(section->mr);
1312 }
1313
1314 static void whpx_transaction_begin(MemoryListener *listener)
1315 {
1316 }
1317
1318 static void whpx_transaction_commit(MemoryListener *listener)
1319 {
1320 }
1321
1322 static void whpx_log_sync(MemoryListener *listener,
1323                          MemoryRegionSection *section)
1324 {
1325     MemoryRegion *mr = section->mr;
1326
1327     if (!memory_region_is_ram(mr)) {
1328         return;
1329     }
1330
1331     memory_region_set_dirty(mr, 0, int128_get64(section->size));
1332 }
1333
1334 static MemoryListener whpx_memory_listener = {
1335     .begin = whpx_transaction_begin,
1336     .commit = whpx_transaction_commit,
1337     .region_add = whpx_region_add,
1338     .region_del = whpx_region_del,
1339     .log_sync = whpx_log_sync,
1340     .priority = 10,
1341 };
1342
1343 static void whpx_memory_init(void)
1344 {
1345     memory_listener_register(&whpx_memory_listener, &address_space_memory);
1346 }
1347
1348 static void whpx_handle_interrupt(CPUState *cpu, int mask)
1349 {
1350     cpu->interrupt_request |= mask;
1351
1352     if (!qemu_cpu_is_self(cpu)) {
1353         qemu_cpu_kick(cpu);
1354     }
1355 }
1356
1357 /*
1358  * Partition support
1359  */
1360
1361 static int whpx_accel_init(MachineState *ms)
1362 {
1363     struct whpx_state *whpx;
1364     int ret;
1365     HRESULT hr;
1366     WHV_CAPABILITY whpx_cap;
1367     UINT32 whpx_cap_size;
1368     WHV_PARTITION_PROPERTY prop;
1369
1370     whpx = &whpx_global;
1371
1372     if (!init_whp_dispatch()) {
1373         ret = -ENOSYS;
1374         goto error;
1375     }
1376
1377     memset(whpx, 0, sizeof(struct whpx_state));
1378     whpx->mem_quota = ms->ram_size;
1379
1380     hr = whp_dispatch.WHvGetCapability(
1381         WHvCapabilityCodeHypervisorPresent, &whpx_cap,
1382         sizeof(whpx_cap), &whpx_cap_size);
1383     if (FAILED(hr) || !whpx_cap.HypervisorPresent) {
1384         error_report("WHPX: No accelerator found, hr=%08lx", hr);
1385         ret = -ENOSPC;
1386         goto error;
1387     }
1388
1389     hr = whp_dispatch.WHvCreatePartition(&whpx->partition);
1390     if (FAILED(hr)) {
1391         error_report("WHPX: Failed to create partition, hr=%08lx", hr);
1392         ret = -EINVAL;
1393         goto error;
1394     }
1395
1396     memset(&prop, 0, sizeof(WHV_PARTITION_PROPERTY));
1397     prop.ProcessorCount = ms->smp.cpus;
1398     hr = whp_dispatch.WHvSetPartitionProperty(
1399         whpx->partition,
1400         WHvPartitionPropertyCodeProcessorCount,
1401         &prop,
1402         sizeof(WHV_PARTITION_PROPERTY));
1403
1404     if (FAILED(hr)) {
1405         error_report("WHPX: Failed to set partition core count to %d,"
1406                      " hr=%08lx", ms->smp.cores, hr);
1407         ret = -EINVAL;
1408         goto error;
1409     }
1410
1411     memset(&prop, 0, sizeof(WHV_PARTITION_PROPERTY));
1412     prop.ExtendedVmExits.X64MsrExit = 1;
1413     prop.ExtendedVmExits.X64CpuidExit = 1;
1414     hr = whp_dispatch.WHvSetPartitionProperty(
1415         whpx->partition,
1416         WHvPartitionPropertyCodeExtendedVmExits,
1417         &prop,
1418         sizeof(WHV_PARTITION_PROPERTY));
1419
1420     if (FAILED(hr)) {
1421         error_report("WHPX: Failed to enable partition extended X64MsrExit and"
1422                      " X64CpuidExit hr=%08lx", hr);
1423         ret = -EINVAL;
1424         goto error;
1425     }
1426
1427     UINT32 cpuidExitList[] = {1, 0x80000001};
1428     hr = whp_dispatch.WHvSetPartitionProperty(
1429         whpx->partition,
1430         WHvPartitionPropertyCodeCpuidExitList,
1431         cpuidExitList,
1432         RTL_NUMBER_OF(cpuidExitList) * sizeof(UINT32));
1433
1434     if (FAILED(hr)) {
1435         error_report("WHPX: Failed to set partition CpuidExitList hr=%08lx",
1436                      hr);
1437         ret = -EINVAL;
1438         goto error;
1439     }
1440
1441     hr = whp_dispatch.WHvSetupPartition(whpx->partition);
1442     if (FAILED(hr)) {
1443         error_report("WHPX: Failed to setup partition, hr=%08lx", hr);
1444         ret = -EINVAL;
1445         goto error;
1446     }
1447
1448     whpx_memory_init();
1449
1450     cpu_interrupt_handler = whpx_handle_interrupt;
1451
1452     printf("Windows Hypervisor Platform accelerator is operational\n");
1453     return 0;
1454
1455   error:
1456
1457     if (NULL != whpx->partition) {
1458         whp_dispatch.WHvDeletePartition(whpx->partition);
1459         whpx->partition = NULL;
1460     }
1461
1462
1463     return ret;
1464 }
1465
1466 int whpx_enabled(void)
1467 {
1468     return whpx_allowed;
1469 }
1470
1471 static void whpx_accel_class_init(ObjectClass *oc, void *data)
1472 {
1473     AccelClass *ac = ACCEL_CLASS(oc);
1474     ac->name = "WHPX";
1475     ac->init_machine = whpx_accel_init;
1476     ac->allowed = &whpx_allowed;
1477 }
1478
1479 static const TypeInfo whpx_accel_type = {
1480     .name = ACCEL_CLASS_NAME("whpx"),
1481     .parent = TYPE_ACCEL,
1482     .class_init = whpx_accel_class_init,
1483 };
1484
1485 static void whpx_type_init(void)
1486 {
1487     type_register_static(&whpx_accel_type);
1488 }
1489
1490 bool init_whp_dispatch(void)
1491 {
1492     const char *lib_name;
1493     HMODULE hLib;
1494
1495     if (whp_dispatch_initialized) {
1496         return true;
1497     }
1498
1499     #define WHP_LOAD_FIELD(return_type, function_name, signature) \
1500         whp_dispatch.function_name = \
1501             (function_name ## _t)GetProcAddress(hLib, #function_name); \
1502         if (!whp_dispatch.function_name) { \
1503             error_report("Could not load function %s from library %s.", \
1504                          #function_name, lib_name); \
1505             goto error; \
1506         } \
1507
1508     lib_name = "WinHvPlatform.dll";
1509     hWinHvPlatform = LoadLibrary(lib_name);
1510     if (!hWinHvPlatform) {
1511         error_report("Could not load library %s.", lib_name);
1512         goto error;
1513     }
1514     hLib = hWinHvPlatform;
1515     LIST_WINHVPLATFORM_FUNCTIONS(WHP_LOAD_FIELD)
1516
1517     lib_name = "WinHvEmulation.dll";
1518     hWinHvEmulation = LoadLibrary(lib_name);
1519     if (!hWinHvEmulation) {
1520         error_report("Could not load library %s.", lib_name);
1521         goto error;
1522     }
1523     hLib = hWinHvEmulation;
1524     LIST_WINHVEMULATION_FUNCTIONS(WHP_LOAD_FIELD)
1525
1526     whp_dispatch_initialized = true;
1527     return true;
1528
1529     error:
1530
1531     if (hWinHvPlatform) {
1532         FreeLibrary(hWinHvPlatform);
1533     }
1534     if (hWinHvEmulation) {
1535         FreeLibrary(hWinHvEmulation);
1536     }
1537     return false;
1538 }
1539
1540 type_init(whpx_type_init);