arch/x86/kvm/svm/nested.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel-based Virtual Machine driver for Linux
   4  *
   5  * AMD SVM support
   6  *
   7  * Copyright (C) 2006 Qumranet, Inc.
   8  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   9  *
  10  * Authors:
  11  *   Yaniv Kamay  <yaniv@qumranet.com>
  12  *   Avi Kivity   <avi@qumranet.com>
  13  */
  14
  15 #define pr_fmt(fmt) "SVM: " fmt
  16
  17 #include <linux/kvm_types.h>
  18 #include <linux/kvm_host.h>
  19 #include <linux/kernel.h>
  20
  21 #include <asm/msr-index.h>
  22 #include <asm/debugreg.h>
  23
  24 #include "kvm_emulate.h"
  25 #include "trace.h"
  26 #include "mmu.h"
  27 #include "x86.h"
  28 #include "cpuid.h"
  29 #include "lapic.h"
  30 #include "svm.h"
  31
  32 #define CC KVM_NESTED_VMENTER_CONSISTENCY_CHECK
  33
  34 static void nested_svm_inject_npf_exit(struct kvm_vcpu *vcpu,
  35                                        struct x86_exception *fault)
  36 {
  37         struct vcpu_svm *svm = to_svm(vcpu);
  38
  39         if (svm->vmcb->control.exit_code != SVM_EXIT_NPF) {
  40                 /*
  41                  * TODO: track the cause of the nested page fault, and
  42                  * correctly fill in the high bits of exit_info_1.
  43                  */
  44                 svm->vmcb->control.exit_code = SVM_EXIT_NPF;
  45                 svm->vmcb->control.exit_code_hi = 0;
  46                 svm->vmcb->control.exit_info_1 = (1ULL << 32);
  47                 svm->vmcb->control.exit_info_2 = fault->address;
  48         }
  49
  50         svm->vmcb->control.exit_info_1 &= ~0xffffffffULL;
  51         svm->vmcb->control.exit_info_1 |= fault->error_code;
  52
  53         nested_svm_vmexit(svm);
  54 }
  55
  56 static void svm_inject_page_fault_nested(struct kvm_vcpu *vcpu, struct x86_exception *fault)
  57 {
  58        struct vcpu_svm *svm = to_svm(vcpu);
  59        WARN_ON(!is_guest_mode(vcpu));
  60
  61        if (vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_EXCEPTION_OFFSET + PF_VECTOR) &&
  62            !svm->nested.nested_run_pending) {
  63                svm->vmcb->control.exit_code = SVM_EXIT_EXCP_BASE + PF_VECTOR;
  64                svm->vmcb->control.exit_code_hi = 0;
  65                svm->vmcb->control.exit_info_1 = fault->error_code;
  66                svm->vmcb->control.exit_info_2 = fault->address;
  67                nested_svm_vmexit(svm);
  68        } else {
  69                kvm_inject_page_fault(vcpu, fault);
  70        }
  71 }
  72
  73 static u64 nested_svm_get_tdp_pdptr(struct kvm_vcpu *vcpu, int index)
  74 {
  75         struct vcpu_svm *svm = to_svm(vcpu);
  76         u64 cr3 = svm->nested.ctl.nested_cr3;
  77         u64 pdpte;
  78         int ret;
  79
  80         ret = kvm_vcpu_read_guest_page(vcpu, gpa_to_gfn(cr3), &pdpte,
  81                                        offset_in_page(cr3) + index * 8, 8);
  82         if (ret)
  83                 return 0;
  84         return pdpte;
  85 }
  86
  87 static unsigned long nested_svm_get_tdp_cr3(struct kvm_vcpu *vcpu)
  88 {
  89         struct vcpu_svm *svm = to_svm(vcpu);
  90
  91         return svm->nested.ctl.nested_cr3;
  92 }
  93
  94 static void nested_svm_init_mmu_context(struct kvm_vcpu *vcpu)
  95 {
  96         struct vcpu_svm *svm = to_svm(vcpu);
  97
  98         WARN_ON(mmu_is_nested(vcpu));
  99
 100         vcpu->arch.mmu = &vcpu->arch.guest_mmu;
 101
 102         /*
 103          * The NPT format depends on L1's CR4 and EFER, which is in vmcb01.  Note,
 104          * when called via KVM_SET_NESTED_STATE, that state may _not_ match current
 105          * vCPU state.  CR0.WP is explicitly ignored, while CR0.PG is required.
 106          */
 107         kvm_init_shadow_npt_mmu(vcpu, X86_CR0_PG, svm->vmcb01.ptr->save.cr4,
 108                                 svm->vmcb01.ptr->save.efer,
 109                                 svm->nested.ctl.nested_cr3);
 110         vcpu->arch.mmu->get_guest_pgd     = nested_svm_get_tdp_cr3;
 111         vcpu->arch.mmu->get_pdptr         = nested_svm_get_tdp_pdptr;
 112         vcpu->arch.mmu->inject_page_fault = nested_svm_inject_npf_exit;
 113         vcpu->arch.walk_mmu              = &vcpu->arch.nested_mmu;
 114 }
 115
 116 static void nested_svm_uninit_mmu_context(struct kvm_vcpu *vcpu)
 117 {
 118         vcpu->arch.mmu = &vcpu->arch.root_mmu;
 119         vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
 120 }
 121
 122 void recalc_intercepts(struct vcpu_svm *svm)
 123 {
 124         struct vmcb_control_area *c, *h, *g;
 125         unsigned int i;
 126
 127         vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
 128
 129         if (!is_guest_mode(&svm->vcpu))
 130                 return;
 131
 132         c = &svm->vmcb->control;
 133         h = &svm->vmcb01.ptr->control;
 134         g = &svm->nested.ctl;
 135
 136         for (i = 0; i < MAX_INTERCEPT; i++)
 137                 c->intercepts[i] = h->intercepts[i];
 138
 139         if (g->int_ctl & V_INTR_MASKING_MASK) {
 140                 /* We only want the cr8 intercept bits of L1 */
 141                 vmcb_clr_intercept(c, INTERCEPT_CR8_READ);
 142                 vmcb_clr_intercept(c, INTERCEPT_CR8_WRITE);
 143
 144                 /*
 145                  * Once running L2 with HF_VINTR_MASK, EFLAGS.IF does not
 146                  * affect any interrupt we may want to inject; therefore,
 147                  * interrupt window vmexits are irrelevant to L0.
 148                  */
 149                 vmcb_clr_intercept(c, INTERCEPT_VINTR);
 150         }
 151
 152         /* We don't want to see VMMCALLs from a nested guest */
 153         vmcb_clr_intercept(c, INTERCEPT_VMMCALL);
 154
 155         for (i = 0; i < MAX_INTERCEPT; i++)
 156                 c->intercepts[i] |= g->intercepts[i];
 157 }
 158
 159 static void copy_vmcb_control_area(struct vmcb_control_area *dst,
 160                                    struct vmcb_control_area *from)
 161 {
 162         unsigned int i;
 163
 164         for (i = 0; i < MAX_INTERCEPT; i++)
 165                 dst->intercepts[i] = from->intercepts[i];
 166
 167         dst->iopm_base_pa         = from->iopm_base_pa;
 168         dst->msrpm_base_pa        = from->msrpm_base_pa;
 169         dst->tsc_offset           = from->tsc_offset;
 170         /* asid not copied, it is handled manually for svm->vmcb.  */
 171         dst->tlb_ctl              = from->tlb_ctl;
 172         dst->int_ctl              = from->int_ctl;
 173         dst->int_vector           = from->int_vector;
 174         dst->int_state            = from->int_state;
 175         dst->exit_code            = from->exit_code;
 176         dst->exit_code_hi         = from->exit_code_hi;
 177         dst->exit_info_1          = from->exit_info_1;
 178         dst->exit_info_2          = from->exit_info_2;
 179         dst->exit_int_info        = from->exit_int_info;
 180         dst->exit_int_info_err    = from->exit_int_info_err;
 181         dst->nested_ctl           = from->nested_ctl;
 182         dst->event_inj            = from->event_inj;
 183         dst->event_inj_err        = from->event_inj_err;
 184         dst->nested_cr3           = from->nested_cr3;
 185         dst->virt_ext              = from->virt_ext;
 186         dst->pause_filter_count   = from->pause_filter_count;
 187         dst->pause_filter_thresh  = from->pause_filter_thresh;
 188 }
 189
 190 static bool nested_svm_vmrun_msrpm(struct vcpu_svm *svm)
 191 {
 192         /*
 193          * This function merges the msr permission bitmaps of kvm and the
 194          * nested vmcb. It is optimized in that it only merges the parts where
 195          * the kvm msr permission bitmap may contain zero bits
 196          */
 197         int i;
 198
 199         if (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_MSR_PROT)))
 200                 return true;
 201
 202         for (i = 0; i < MSRPM_OFFSETS; i++) {
 203                 u32 value, p;
 204                 u64 offset;
 205
 206                 if (msrpm_offsets[i] == 0xffffffff)
 207                         break;
 208
 209                 p      = msrpm_offsets[i];
 210                 offset = svm->nested.ctl.msrpm_base_pa + (p * 4);
 211
 212                 if (kvm_vcpu_read_guest(&svm->vcpu, offset, &value, 4))
 213                         return false;
 214
 215                 svm->nested.msrpm[p] = svm->msrpm[p] | value;
 216         }
 217
 218         svm->vmcb->control.msrpm_base_pa = __sme_set(__pa(svm->nested.msrpm));
 219
 220         return true;
 221 }
 222
 223 /*
 224  * Bits 11:0 of bitmap address are ignored by hardware
 225  */
 226 static bool nested_svm_check_bitmap_pa(struct kvm_vcpu *vcpu, u64 pa, u32 size)
 227 {
 228         u64 addr = PAGE_ALIGN(pa);
 229
 230         return kvm_vcpu_is_legal_gpa(vcpu, addr) &&
 231             kvm_vcpu_is_legal_gpa(vcpu, addr + size - 1);
 232 }
 233
 234 static bool nested_vmcb_check_controls(struct kvm_vcpu *vcpu,
 235                                        struct vmcb_control_area *control)
 236 {
 237         if (CC(!vmcb_is_intercept(control, INTERCEPT_VMRUN)))
 238                 return false;
 239
 240         if (CC(control->asid == 0))
 241                 return false;
 242
 243         if (CC((control->nested_ctl & SVM_NESTED_CTL_NP_ENABLE) && !npt_enabled))
 244                 return false;
 245
 246         if (CC(!nested_svm_check_bitmap_pa(vcpu, control->msrpm_base_pa,
 247                                            MSRPM_SIZE)))
 248                 return false;
 249         if (CC(!nested_svm_check_bitmap_pa(vcpu, control->iopm_base_pa,
 250                                            IOPM_SIZE)))
 251                 return false;
 252
 253         return true;
 254 }
 255
 256 static bool nested_vmcb_check_cr3_cr4(struct kvm_vcpu *vcpu,
 257                                       struct vmcb_save_area *save)
 258 {
 259         /*
 260          * These checks are also performed by KVM_SET_SREGS,
 261          * except that EFER.LMA is not checked by SVM against
 262          * CR0.PG && EFER.LME.
 263          */
 264         if ((save->efer & EFER_LME) && (save->cr0 & X86_CR0_PG)) {
 265                 if (CC(!(save->cr4 & X86_CR4_PAE)) ||
 266                     CC(!(save->cr0 & X86_CR0_PE)) ||
 267                     CC(kvm_vcpu_is_illegal_gpa(vcpu, save->cr3)))
 268                         return false;
 269         }
 270
 271         if (CC(!kvm_is_valid_cr4(vcpu, save->cr4)))
 272                 return false;
 273
 274         return true;
 275 }
 276
 277 /* Common checks that apply to both L1 and L2 state.  */
 278 static bool nested_vmcb_valid_sregs(struct kvm_vcpu *vcpu,
 279                                     struct vmcb_save_area *save)
 280 {
 281         /*
 282          * FIXME: these should be done after copying the fields,
 283          * to avoid TOC/TOU races.  For these save area checks
 284          * the possible damage is limited since kvm_set_cr0 and
 285          * kvm_set_cr4 handle failure; EFER_SVME is an exception
 286          * so it is force-set later in nested_prepare_vmcb_save.
 287          */
 288         if (CC(!(save->efer & EFER_SVME)))
 289                 return false;
 290
 291         if (CC((save->cr0 & X86_CR0_CD) == 0 && (save->cr0 & X86_CR0_NW)) ||
 292             CC(save->cr0 & ~0xffffffffULL))
 293                 return false;
 294
 295         if (CC(!kvm_dr6_valid(save->dr6)) || CC(!kvm_dr7_valid(save->dr7)))
 296                 return false;
 297
 298         if (!nested_vmcb_check_cr3_cr4(vcpu, save))
 299                 return false;
 300
 301         if (CC(!kvm_valid_efer(vcpu, save->efer)))
 302                 return false;
 303
 304         return true;
 305 }
 306
 307 static void nested_load_control_from_vmcb12(struct vcpu_svm *svm,
 308                                             struct vmcb_control_area *control)
 309 {
 310         copy_vmcb_control_area(&svm->nested.ctl, control);
 311
 312         /* Copy it here because nested_svm_check_controls will check it.  */
 313         svm->nested.ctl.asid           = control->asid;
 314         svm->nested.ctl.msrpm_base_pa &= ~0x0fffULL;
 315         svm->nested.ctl.iopm_base_pa  &= ~0x0fffULL;
 316 }
 317
 318 /*
 319  * Synchronize fields that are written by the processor, so that
 320  * they can be copied back into the vmcb12.
 321  */
 322 void nested_sync_control_from_vmcb02(struct vcpu_svm *svm)
 323 {
 324         u32 mask;
 325         svm->nested.ctl.event_inj      = svm->vmcb->control.event_inj;
 326         svm->nested.ctl.event_inj_err  = svm->vmcb->control.event_inj_err;
 327
 328         /* Only a few fields of int_ctl are written by the processor.  */
 329         mask = V_IRQ_MASK | V_TPR_MASK;
 330         if (!(svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK) &&
 331             svm_is_intercept(svm, INTERCEPT_VINTR)) {
 332                 /*
 333                  * In order to request an interrupt window, L0 is usurping
 334                  * svm->vmcb->control.int_ctl and possibly setting V_IRQ
 335                  * even if it was clear in L1's VMCB.  Restoring it would be
 336                  * wrong.  However, in this case V_IRQ will remain true until
 337                  * interrupt_window_interception calls svm_clear_vintr and
 338                  * restores int_ctl.  We can just leave it aside.
 339                  */
 340                 mask &= ~V_IRQ_MASK;
 341         }
 342         svm->nested.ctl.int_ctl        &= ~mask;
 343         svm->nested.ctl.int_ctl        |= svm->vmcb->control.int_ctl & mask;
 344 }
 345
 346 /*
 347  * Transfer any event that L0 or L1 wanted to inject into L2 to
 348  * EXIT_INT_INFO.
 349  */
 350 static void nested_save_pending_event_to_vmcb12(struct vcpu_svm *svm,
 351                                                 struct vmcb *vmcb12)
 352 {
 353         struct kvm_vcpu *vcpu = &svm->vcpu;
 354         u32 exit_int_info = 0;
 355         unsigned int nr;
 356
 357         if (vcpu->arch.exception.injected) {
 358                 nr = vcpu->arch.exception.nr;
 359                 exit_int_info = nr | SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_EXEPT;
 360
 361                 if (vcpu->arch.exception.has_error_code) {
 362                         exit_int_info |= SVM_EVTINJ_VALID_ERR;
 363                         vmcb12->control.exit_int_info_err =
 364                                 vcpu->arch.exception.error_code;
 365                 }
 366
 367         } else if (vcpu->arch.nmi_injected) {
 368                 exit_int_info = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
 369
 370         } else if (vcpu->arch.interrupt.injected) {
 371                 nr = vcpu->arch.interrupt.nr;
 372                 exit_int_info = nr | SVM_EVTINJ_VALID;
 373
 374                 if (vcpu->arch.interrupt.soft)
 375                         exit_int_info |= SVM_EVTINJ_TYPE_SOFT;
 376                 else
 377                         exit_int_info |= SVM_EVTINJ_TYPE_INTR;
 378         }
 379
 380         vmcb12->control.exit_int_info = exit_int_info;
 381 }
 382
 383 static inline bool nested_npt_enabled(struct vcpu_svm *svm)
 384 {
 385         return svm->nested.ctl.nested_ctl & SVM_NESTED_CTL_NP_ENABLE;
 386 }
 387
 388 static void nested_svm_transition_tlb_flush(struct kvm_vcpu *vcpu)
 389 {
 390         /*
 391          * TODO: optimize unconditional TLB flush/MMU sync.  A partial list of
 392          * things to fix before this can be conditional:
 393          *
 394          *  - Flush TLBs for both L1 and L2 remote TLB flush
 395          *  - Honor L1's request to flush an ASID on nested VMRUN
 396          *  - Sync nested NPT MMU on VMRUN that flushes L2's ASID[*]
 397          *  - Don't crush a pending TLB flush in vmcb02 on nested VMRUN
 398          *  - Flush L1's ASID on KVM_REQ_TLB_FLUSH_GUEST
 399          *
 400          * [*] Unlike nested EPT, SVM's ASID management can invalidate nested
 401          *     NPT guest-physical mappings on VMRUN.
 402          */
 403         kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
 404         kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
 405 }
 406
 407 /*
 408  * Load guest's/host's cr3 on nested vmentry or vmexit. @nested_npt is true
 409  * if we are emulating VM-Entry into a guest with NPT enabled.
 410  */
 411 static int nested_svm_load_cr3(struct kvm_vcpu *vcpu, unsigned long cr3,
 412                                bool nested_npt, bool reload_pdptrs)
 413 {
 414         if (CC(kvm_vcpu_is_illegal_gpa(vcpu, cr3)))
 415                 return -EINVAL;
 416
 417         if (reload_pdptrs && !nested_npt && is_pae_paging(vcpu) &&
 418             CC(!load_pdptrs(vcpu, vcpu->arch.walk_mmu, cr3)))
 419                 return -EINVAL;
 420
 421         if (!nested_npt)
 422                 kvm_mmu_new_pgd(vcpu, cr3);
 423
 424         vcpu->arch.cr3 = cr3;
 425         kvm_register_mark_available(vcpu, VCPU_EXREG_CR3);
 426
 427         /* Re-initialize the MMU, e.g. to pick up CR4 MMU role changes. */
 428         kvm_init_mmu(vcpu);
 429
 430         return 0;
 431 }
 432
 433 void nested_vmcb02_compute_g_pat(struct vcpu_svm *svm)
 434 {
 435         if (!svm->nested.vmcb02.ptr)
 436                 return;
 437
 438         /* FIXME: merge g_pat from vmcb01 and vmcb12.  */
 439         svm->nested.vmcb02.ptr->save.g_pat = svm->vmcb01.ptr->save.g_pat;
 440 }
 441
 442 static void nested_vmcb02_prepare_save(struct vcpu_svm *svm, struct vmcb *vmcb12)
 443 {
 444         bool new_vmcb12 = false;
 445
 446         nested_vmcb02_compute_g_pat(svm);
 447
 448         /* Load the nested guest state */
 449         if (svm->nested.vmcb12_gpa != svm->nested.last_vmcb12_gpa) {
 450                 new_vmcb12 = true;
 451                 svm->nested.last_vmcb12_gpa = svm->nested.vmcb12_gpa;
 452         }
 453
 454         if (unlikely(new_vmcb12 || vmcb_is_dirty(vmcb12, VMCB_SEG))) {
 455                 svm->vmcb->save.es = vmcb12->save.es;
 456                 svm->vmcb->save.cs = vmcb12->save.cs;
 457                 svm->vmcb->save.ss = vmcb12->save.ss;
 458                 svm->vmcb->save.ds = vmcb12->save.ds;
 459                 svm->vmcb->save.cpl = vmcb12->save.cpl;
 460                 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
 461         }
 462
 463         if (unlikely(new_vmcb12 || vmcb_is_dirty(vmcb12, VMCB_DT))) {
 464                 svm->vmcb->save.gdtr = vmcb12->save.gdtr;
 465                 svm->vmcb->save.idtr = vmcb12->save.idtr;
 466                 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
 467         }
 468
 469         kvm_set_rflags(&svm->vcpu, vmcb12->save.rflags | X86_EFLAGS_FIXED);
 470
 471         /*
 472          * Force-set EFER_SVME even though it is checked earlier on the
 473          * VMCB12, because the guest can flip the bit between the check
 474          * and now.  Clearing EFER_SVME would call svm_free_nested.
 475          */
 476         svm_set_efer(&svm->vcpu, vmcb12->save.efer | EFER_SVME);
 477
 478         svm_set_cr0(&svm->vcpu, vmcb12->save.cr0);
 479         svm_set_cr4(&svm->vcpu, vmcb12->save.cr4);
 480
 481         svm->vcpu.arch.cr2 = vmcb12->save.cr2;
 482
 483         kvm_rax_write(&svm->vcpu, vmcb12->save.rax);
 484         kvm_rsp_write(&svm->vcpu, vmcb12->save.rsp);
 485         kvm_rip_write(&svm->vcpu, vmcb12->save.rip);
 486
 487         /* In case we don't even reach vcpu_run, the fields are not updated */
 488         svm->vmcb->save.rax = vmcb12->save.rax;
 489         svm->vmcb->save.rsp = vmcb12->save.rsp;
 490         svm->vmcb->save.rip = vmcb12->save.rip;
 491
 492         /* These bits will be set properly on the first execution when new_vmc12 is true */
 493         if (unlikely(new_vmcb12 || vmcb_is_dirty(vmcb12, VMCB_DR))) {
 494                 svm->vmcb->save.dr7 = vmcb12->save.dr7 | DR7_FIXED_1;
 495                 svm->vcpu.arch.dr6  = vmcb12->save.dr6 | DR6_ACTIVE_LOW;
 496                 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
 497         }
 498 }
 499
 500 static void nested_vmcb02_prepare_control(struct vcpu_svm *svm)
 501 {
 502         const u32 mask = V_INTR_MASKING_MASK | V_GIF_ENABLE_MASK | V_GIF_MASK;
 503         struct kvm_vcpu *vcpu = &svm->vcpu;
 504
 505         /*
 506          * Filled at exit: exit_code, exit_code_hi, exit_info_1, exit_info_2,
 507          * exit_int_info, exit_int_info_err, next_rip, insn_len, insn_bytes.
 508          */
 509
 510         /*
 511          * Also covers avic_vapic_bar, avic_backing_page, avic_logical_id,
 512          * avic_physical_id.
 513          */
 514         WARN_ON(svm->vmcb01.ptr->control.int_ctl & AVIC_ENABLE_MASK);
 515
 516         /* Copied from vmcb01.  msrpm_base can be overwritten later.  */
 517         svm->vmcb->control.nested_ctl = svm->vmcb01.ptr->control.nested_ctl;
 518         svm->vmcb->control.iopm_base_pa = svm->vmcb01.ptr->control.iopm_base_pa;
 519         svm->vmcb->control.msrpm_base_pa = svm->vmcb01.ptr->control.msrpm_base_pa;
 520
 521         /* Done at vmrun: asid.  */
 522
 523         /* Also overwritten later if necessary.  */
 524         svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
 525
 526         /* nested_cr3.  */
 527         if (nested_npt_enabled(svm))
 528                 nested_svm_init_mmu_context(vcpu);
 529
 530         svm->vmcb->control.tsc_offset = vcpu->arch.tsc_offset =
 531                 vcpu->arch.l1_tsc_offset + svm->nested.ctl.tsc_offset;
 532
 533         svm->vmcb->control.int_ctl             =
 534                 (svm->nested.ctl.int_ctl & ~mask) |
 535                 (svm->vmcb01.ptr->control.int_ctl & mask);
 536
 537         svm->vmcb->control.virt_ext            = svm->nested.ctl.virt_ext;
 538         svm->vmcb->control.int_vector          = svm->nested.ctl.int_vector;
 539         svm->vmcb->control.int_state           = svm->nested.ctl.int_state;
 540         svm->vmcb->control.event_inj           = svm->nested.ctl.event_inj;
 541         svm->vmcb->control.event_inj_err       = svm->nested.ctl.event_inj_err;
 542
 543         svm->vmcb->control.pause_filter_count  = svm->nested.ctl.pause_filter_count;
 544         svm->vmcb->control.pause_filter_thresh = svm->nested.ctl.pause_filter_thresh;
 545
 546         nested_svm_transition_tlb_flush(vcpu);
 547
 548         /* Enter Guest-Mode */
 549         enter_guest_mode(vcpu);
 550
 551         /*
 552          * Merge guest and host intercepts - must be called with vcpu in
 553          * guest-mode to take effect.
 554          */
 555         recalc_intercepts(svm);
 556 }
 557
 558 static void nested_svm_copy_common_state(struct vmcb *from_vmcb, struct vmcb *to_vmcb)
 559 {
 560         /*
 561          * Some VMCB state is shared between L1 and L2 and thus has to be
 562          * moved at the time of nested vmrun and vmexit.
 563          *
 564          * VMLOAD/VMSAVE state would also belong in this category, but KVM
 565          * always performs VMLOAD and VMSAVE from the VMCB01.
 566          */
 567         to_vmcb->save.spec_ctrl = from_vmcb->save.spec_ctrl;
 568 }
 569
 570 int enter_svm_guest_mode(struct kvm_vcpu *vcpu, u64 vmcb12_gpa,
 571                          struct vmcb *vmcb12)
 572 {
 573         struct vcpu_svm *svm = to_svm(vcpu);
 574         int ret;
 575
 576         trace_kvm_nested_vmrun(svm->vmcb->save.rip, vmcb12_gpa,
 577                                vmcb12->save.rip,
 578                                vmcb12->control.int_ctl,
 579                                vmcb12->control.event_inj,
 580                                vmcb12->control.nested_ctl);
 581
 582         trace_kvm_nested_intercepts(vmcb12->control.intercepts[INTERCEPT_CR] & 0xffff,
 583                                     vmcb12->control.intercepts[INTERCEPT_CR] >> 16,
 584                                     vmcb12->control.intercepts[INTERCEPT_EXCEPTION],
 585                                     vmcb12->control.intercepts[INTERCEPT_WORD3],
 586                                     vmcb12->control.intercepts[INTERCEPT_WORD4],
 587                                     vmcb12->control.intercepts[INTERCEPT_WORD5]);
 588
 589
 590         svm->nested.vmcb12_gpa = vmcb12_gpa;
 591
 592         WARN_ON(svm->vmcb == svm->nested.vmcb02.ptr);
 593
 594         nested_svm_copy_common_state(svm->vmcb01.ptr, svm->nested.vmcb02.ptr);
 595
 596         svm_switch_vmcb(svm, &svm->nested.vmcb02);
 597         nested_vmcb02_prepare_control(svm);
 598         nested_vmcb02_prepare_save(svm, vmcb12);
 599
 600         ret = nested_svm_load_cr3(&svm->vcpu, vmcb12->save.cr3,
 601                                   nested_npt_enabled(svm), true);
 602         if (ret)
 603                 return ret;
 604
 605         if (!npt_enabled)
 606                 vcpu->arch.mmu->inject_page_fault = svm_inject_page_fault_nested;
 607
 608         svm_set_gif(svm, true);
 609
 610         return 0;
 611 }
 612
 613 int nested_svm_vmrun(struct kvm_vcpu *vcpu)
 614 {
 615         struct vcpu_svm *svm = to_svm(vcpu);
 616         int ret;
 617         struct vmcb *vmcb12;
 618         struct kvm_host_map map;
 619         u64 vmcb12_gpa;
 620
 621         if (is_smm(vcpu)) {
 622                 kvm_queue_exception(vcpu, UD_VECTOR);
 623                 return 1;
 624         }
 625
 626         vmcb12_gpa = svm->vmcb->save.rax;
 627         ret = kvm_vcpu_map(vcpu, gpa_to_gfn(vmcb12_gpa), &map);
 628         if (ret == -EINVAL) {
 629                 kvm_inject_gp(vcpu, 0);
 630                 return 1;
 631         } else if (ret) {
 632                 return kvm_skip_emulated_instruction(vcpu);
 633         }
 634
 635         ret = kvm_skip_emulated_instruction(vcpu);
 636
 637         vmcb12 = map.hva;
 638
 639         if (WARN_ON_ONCE(!svm->nested.initialized))
 640                 return -EINVAL;
 641
 642         nested_load_control_from_vmcb12(svm, &vmcb12->control);
 643
 644         if (!nested_vmcb_valid_sregs(vcpu, &vmcb12->save) ||
 645             !nested_vmcb_check_controls(vcpu, &svm->nested.ctl)) {
 646                 vmcb12->control.exit_code    = SVM_EXIT_ERR;
 647                 vmcb12->control.exit_code_hi = 0;
 648                 vmcb12->control.exit_info_1  = 0;
 649                 vmcb12->control.exit_info_2  = 0;
 650                 goto out;
 651         }
 652
 653
 654         /* Clear internal status */
 655         kvm_clear_exception_queue(vcpu);
 656         kvm_clear_interrupt_queue(vcpu);
 657
 658         /*
 659          * Since vmcb01 is not in use, we can use it to store some of the L1
 660          * state.
 661          */
 662         svm->vmcb01.ptr->save.efer   = vcpu->arch.efer;
 663         svm->vmcb01.ptr->save.cr0    = kvm_read_cr0(vcpu);
 664         svm->vmcb01.ptr->save.cr4    = vcpu->arch.cr4;
 665         svm->vmcb01.ptr->save.rflags = kvm_get_rflags(vcpu);
 666         svm->vmcb01.ptr->save.rip    = kvm_rip_read(vcpu);
 667
 668         if (!npt_enabled)
 669                 svm->vmcb01.ptr->save.cr3 = kvm_read_cr3(vcpu);
 670
 671         svm->nested.nested_run_pending = 1;
 672
 673         if (enter_svm_guest_mode(vcpu, vmcb12_gpa, vmcb12))
 674                 goto out_exit_err;
 675
 676         if (nested_svm_vmrun_msrpm(svm))
 677                 goto out;
 678
 679 out_exit_err:
 680         svm->nested.nested_run_pending = 0;
 681
 682         svm->vmcb->control.exit_code    = SVM_EXIT_ERR;
 683         svm->vmcb->control.exit_code_hi = 0;
 684         svm->vmcb->control.exit_info_1  = 0;
 685         svm->vmcb->control.exit_info_2  = 0;
 686
 687         nested_svm_vmexit(svm);
 688
 689 out:
 690         kvm_vcpu_unmap(vcpu, &map, true);
 691
 692         return ret;
 693 }
 694
 695 void nested_svm_vmloadsave(struct vmcb *from_vmcb, struct vmcb *to_vmcb)
 696 {
 697         to_vmcb->save.fs = from_vmcb->save.fs;
 698         to_vmcb->save.gs = from_vmcb->save.gs;
 699         to_vmcb->save.tr = from_vmcb->save.tr;
 700         to_vmcb->save.ldtr = from_vmcb->save.ldtr;
 701         to_vmcb->save.kernel_gs_base = from_vmcb->save.kernel_gs_base;
 702         to_vmcb->save.star = from_vmcb->save.star;
 703         to_vmcb->save.lstar = from_vmcb->save.lstar;
 704         to_vmcb->save.cstar = from_vmcb->save.cstar;
 705         to_vmcb->save.sfmask = from_vmcb->save.sfmask;
 706         to_vmcb->save.sysenter_cs = from_vmcb->save.sysenter_cs;
 707         to_vmcb->save.sysenter_esp = from_vmcb->save.sysenter_esp;
 708         to_vmcb->save.sysenter_eip = from_vmcb->save.sysenter_eip;
 709 }
 710
 711 int nested_svm_vmexit(struct vcpu_svm *svm)
 712 {
 713         struct kvm_vcpu *vcpu = &svm->vcpu;
 714         struct vmcb *vmcb12;
 715         struct vmcb *vmcb = svm->vmcb;
 716         struct kvm_host_map map;
 717         int rc;
 718
 719         /* Triple faults in L2 should never escape. */
 720         WARN_ON_ONCE(kvm_check_request(KVM_REQ_TRIPLE_FAULT, vcpu));
 721
 722         rc = kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.vmcb12_gpa), &map);
 723         if (rc) {
 724                 if (rc == -EINVAL)
 725                         kvm_inject_gp(vcpu, 0);
 726                 return 1;
 727         }
 728
 729         vmcb12 = map.hva;
 730
 731         /* Exit Guest-Mode */
 732         leave_guest_mode(vcpu);
 733         svm->nested.vmcb12_gpa = 0;
 734         WARN_ON_ONCE(svm->nested.nested_run_pending);
 735
 736         kvm_clear_request(KVM_REQ_GET_NESTED_STATE_PAGES, vcpu);
 737
 738         /* in case we halted in L2 */
 739         svm->vcpu.arch.mp_state = KVM_MP_STATE_RUNNABLE;
 740
 741         /* Give the current vmcb to the guest */
 742
 743         vmcb12->save.es     = vmcb->save.es;
 744         vmcb12->save.cs     = vmcb->save.cs;
 745         vmcb12->save.ss     = vmcb->save.ss;
 746         vmcb12->save.ds     = vmcb->save.ds;
 747         vmcb12->save.gdtr   = vmcb->save.gdtr;
 748         vmcb12->save.idtr   = vmcb->save.idtr;
 749         vmcb12->save.efer   = svm->vcpu.arch.efer;
 750         vmcb12->save.cr0    = kvm_read_cr0(vcpu);
 751         vmcb12->save.cr3    = kvm_read_cr3(vcpu);
 752         vmcb12->save.cr2    = vmcb->save.cr2;
 753         vmcb12->save.cr4    = svm->vcpu.arch.cr4;
 754         vmcb12->save.rflags = kvm_get_rflags(vcpu);
 755         vmcb12->save.rip    = kvm_rip_read(vcpu);
 756         vmcb12->save.rsp    = kvm_rsp_read(vcpu);
 757         vmcb12->save.rax    = kvm_rax_read(vcpu);
 758         vmcb12->save.dr7    = vmcb->save.dr7;
 759         vmcb12->save.dr6    = svm->vcpu.arch.dr6;
 760         vmcb12->save.cpl    = vmcb->save.cpl;
 761
 762         vmcb12->control.int_state         = vmcb->control.int_state;
 763         vmcb12->control.exit_code         = vmcb->control.exit_code;
 764         vmcb12->control.exit_code_hi      = vmcb->control.exit_code_hi;
 765         vmcb12->control.exit_info_1       = vmcb->control.exit_info_1;
 766         vmcb12->control.exit_info_2       = vmcb->control.exit_info_2;
 767
 768         if (vmcb12->control.exit_code != SVM_EXIT_ERR)
 769                 nested_save_pending_event_to_vmcb12(svm, vmcb12);
 770
 771         if (svm->nrips_enabled)
 772                 vmcb12->control.next_rip  = vmcb->control.next_rip;
 773
 774         vmcb12->control.int_ctl           = svm->nested.ctl.int_ctl;
 775         vmcb12->control.tlb_ctl           = svm->nested.ctl.tlb_ctl;
 776         vmcb12->control.event_inj         = svm->nested.ctl.event_inj;
 777         vmcb12->control.event_inj_err     = svm->nested.ctl.event_inj_err;
 778
 779         vmcb12->control.pause_filter_count =
 780                 svm->vmcb->control.pause_filter_count;
 781         vmcb12->control.pause_filter_thresh =
 782                 svm->vmcb->control.pause_filter_thresh;
 783
 784         nested_svm_copy_common_state(svm->nested.vmcb02.ptr, svm->vmcb01.ptr);
 785
 786         svm_switch_vmcb(svm, &svm->vmcb01);
 787
 788         /*
 789          * On vmexit the  GIF is set to false and
 790          * no event can be injected in L1.
 791          */
 792         svm_set_gif(svm, false);
 793         svm->vmcb->control.exit_int_info = 0;
 794
 795         svm->vcpu.arch.tsc_offset = svm->vcpu.arch.l1_tsc_offset;
 796         if (svm->vmcb->control.tsc_offset != svm->vcpu.arch.tsc_offset) {
 797                 svm->vmcb->control.tsc_offset = svm->vcpu.arch.tsc_offset;
 798                 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
 799         }
 800
 801         svm->nested.ctl.nested_cr3 = 0;
 802
 803         /*
 804          * Restore processor state that had been saved in vmcb01
 805          */
 806         kvm_set_rflags(vcpu, svm->vmcb->save.rflags);
 807         svm_set_efer(vcpu, svm->vmcb->save.efer);
 808         svm_set_cr0(vcpu, svm->vmcb->save.cr0 | X86_CR0_PE);
 809         svm_set_cr4(vcpu, svm->vmcb->save.cr4);
 810         kvm_rax_write(vcpu, svm->vmcb->save.rax);
 811         kvm_rsp_write(vcpu, svm->vmcb->save.rsp);
 812         kvm_rip_write(vcpu, svm->vmcb->save.rip);
 813
 814         svm->vcpu.arch.dr7 = DR7_FIXED_1;
 815         kvm_update_dr7(&svm->vcpu);
 816
 817         trace_kvm_nested_vmexit_inject(vmcb12->control.exit_code,
 818                                        vmcb12->control.exit_info_1,
 819                                        vmcb12->control.exit_info_2,
 820                                        vmcb12->control.exit_int_info,
 821                                        vmcb12->control.exit_int_info_err,
 822                                        KVM_ISA_SVM);
 823
 824         kvm_vcpu_unmap(vcpu, &map, true);
 825
 826         nested_svm_transition_tlb_flush(vcpu);
 827
 828         nested_svm_uninit_mmu_context(vcpu);
 829
 830         rc = nested_svm_load_cr3(vcpu, svm->vmcb->save.cr3, false, true);
 831         if (rc)
 832                 return 1;
 833
 834         /*
 835          * Drop what we picked up for L2 via svm_complete_interrupts() so it
 836          * doesn't end up in L1.
 837          */
 838         svm->vcpu.arch.nmi_injected = false;
 839         kvm_clear_exception_queue(vcpu);
 840         kvm_clear_interrupt_queue(vcpu);
 841
 842         /*
 843          * If we are here following the completion of a VMRUN that
 844          * is being single-stepped, queue the pending #DB intercept
 845          * right now so that it an be accounted for before we execute
 846          * L1's next instruction.
 847          */
 848         if (unlikely(svm->vmcb->save.rflags & X86_EFLAGS_TF))
 849                 kvm_queue_exception(&(svm->vcpu), DB_VECTOR);
 850
 851         return 0;
 852 }
 853
 854 static void nested_svm_triple_fault(struct kvm_vcpu *vcpu)
 855 {
 856         nested_svm_simple_vmexit(to_svm(vcpu), SVM_EXIT_SHUTDOWN);
 857 }
 858
 859 int svm_allocate_nested(struct vcpu_svm *svm)
 860 {
 861         struct page *vmcb02_page;
 862
 863         if (svm->nested.initialized)
 864                 return 0;
 865
 866         vmcb02_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 867         if (!vmcb02_page)
 868                 return -ENOMEM;
 869         svm->nested.vmcb02.ptr = page_address(vmcb02_page);
 870         svm->nested.vmcb02.pa = __sme_set(page_to_pfn(vmcb02_page) << PAGE_SHIFT);
 871
 872         svm->nested.msrpm = svm_vcpu_alloc_msrpm();
 873         if (!svm->nested.msrpm)
 874                 goto err_free_vmcb02;
 875         svm_vcpu_init_msrpm(&svm->vcpu, svm->nested.msrpm);
 876
 877         svm->nested.initialized = true;
 878         return 0;
 879
 880 err_free_vmcb02:
 881         __free_page(vmcb02_page);
 882         return -ENOMEM;
 883 }
 884
 885 void svm_free_nested(struct vcpu_svm *svm)
 886 {
 887         if (!svm->nested.initialized)
 888                 return;
 889
 890         svm_vcpu_free_msrpm(svm->nested.msrpm);
 891         svm->nested.msrpm = NULL;
 892
 893         __free_page(virt_to_page(svm->nested.vmcb02.ptr));
 894         svm->nested.vmcb02.ptr = NULL;
 895
 896         /*
 897          * When last_vmcb12_gpa matches the current vmcb12 gpa,
 898          * some vmcb12 fields are not loaded if they are marked clean
 899          * in the vmcb12, since in this case they are up to date already.
 900          *
 901          * When the vmcb02 is freed, this optimization becomes invalid.
 902          */
 903         svm->nested.last_vmcb12_gpa = INVALID_GPA;
 904
 905         svm->nested.initialized = false;
 906 }
 907
 908 /*
 909  * Forcibly leave nested mode in order to be able to reset the VCPU later on.
 910  */
 911 void svm_leave_nested(struct vcpu_svm *svm)
 912 {
 913         struct kvm_vcpu *vcpu = &svm->vcpu;
 914
 915         if (is_guest_mode(vcpu)) {
 916                 svm->nested.nested_run_pending = 0;
 917                 svm->nested.vmcb12_gpa = INVALID_GPA;
 918
 919                 leave_guest_mode(vcpu);
 920
 921                 svm_switch_vmcb(svm, &svm->vmcb01);
 922
 923                 nested_svm_uninit_mmu_context(vcpu);
 924                 vmcb_mark_all_dirty(svm->vmcb);
 925         }
 926
 927         kvm_clear_request(KVM_REQ_GET_NESTED_STATE_PAGES, vcpu);
 928 }
 929
 930 static int nested_svm_exit_handled_msr(struct vcpu_svm *svm)
 931 {
 932         u32 offset, msr, value;
 933         int write, mask;
 934
 935         if (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_MSR_PROT)))
 936                 return NESTED_EXIT_HOST;
 937
 938         msr    = svm->vcpu.arch.regs[VCPU_REGS_RCX];
 939         offset = svm_msrpm_offset(msr);
 940         write  = svm->vmcb->control.exit_info_1 & 1;
 941         mask   = 1 << ((2 * (msr & 0xf)) + write);
 942
 943         if (offset == MSR_INVALID)
 944                 return NESTED_EXIT_DONE;
 945
 946         /* Offset is in 32 bit units but need in 8 bit units */
 947         offset *= 4;
 948
 949         if (kvm_vcpu_read_guest(&svm->vcpu, svm->nested.ctl.msrpm_base_pa + offset, &value, 4))
 950                 return NESTED_EXIT_DONE;
 951
 952         return (value & mask) ? NESTED_EXIT_DONE : NESTED_EXIT_HOST;
 953 }
 954
 955 static int nested_svm_intercept_ioio(struct vcpu_svm *svm)
 956 {
 957         unsigned port, size, iopm_len;
 958         u16 val, mask;
 959         u8 start_bit;
 960         u64 gpa;
 961
 962         if (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_IOIO_PROT)))
 963                 return NESTED_EXIT_HOST;
 964
 965         port = svm->vmcb->control.exit_info_1 >> 16;
 966         size = (svm->vmcb->control.exit_info_1 & SVM_IOIO_SIZE_MASK) >>
 967                 SVM_IOIO_SIZE_SHIFT;
 968         gpa  = svm->nested.ctl.iopm_base_pa + (port / 8);
 969         start_bit = port % 8;
 970         iopm_len = (start_bit + size > 8) ? 2 : 1;
 971         mask = (0xf >> (4 - size)) << start_bit;
 972         val = 0;
 973
 974         if (kvm_vcpu_read_guest(&svm->vcpu, gpa, &val, iopm_len))
 975                 return NESTED_EXIT_DONE;
 976
 977         return (val & mask) ? NESTED_EXIT_DONE : NESTED_EXIT_HOST;
 978 }
 979
 980 static int nested_svm_intercept(struct vcpu_svm *svm)
 981 {
 982         u32 exit_code = svm->vmcb->control.exit_code;
 983         int vmexit = NESTED_EXIT_HOST;
 984
 985         switch (exit_code) {
 986         case SVM_EXIT_MSR:
 987                 vmexit = nested_svm_exit_handled_msr(svm);
 988                 break;
 989         case SVM_EXIT_IOIO:
 990                 vmexit = nested_svm_intercept_ioio(svm);
 991                 break;
 992         case SVM_EXIT_READ_CR0 ... SVM_EXIT_WRITE_CR8: {
 993                 if (vmcb_is_intercept(&svm->nested.ctl, exit_code))
 994                         vmexit = NESTED_EXIT_DONE;
 995                 break;
 996         }
 997         case SVM_EXIT_READ_DR0 ... SVM_EXIT_WRITE_DR7: {
 998                 if (vmcb_is_intercept(&svm->nested.ctl, exit_code))
 999                         vmexit = NESTED_EXIT_DONE;
1000                 break;
1001         }
1002         case SVM_EXIT_EXCP_BASE ... SVM_EXIT_EXCP_BASE + 0x1f: {
1003                 /*
1004                  * Host-intercepted exceptions have been checked already in
1005                  * nested_svm_exit_special.  There is nothing to do here,
1006                  * the vmexit is injected by svm_check_nested_events.
1007                  */
1008                 vmexit = NESTED_EXIT_DONE;
1009                 break;
1010         }
1011         case SVM_EXIT_ERR: {
1012                 vmexit = NESTED_EXIT_DONE;
1013                 break;
1014         }
1015         default: {
1016                 if (vmcb_is_intercept(&svm->nested.ctl, exit_code))
1017                         vmexit = NESTED_EXIT_DONE;
1018         }
1019         }
1020
1021         return vmexit;
1022 }
1023
1024 int nested_svm_exit_handled(struct vcpu_svm *svm)
1025 {
1026         int vmexit;
1027
1028         vmexit = nested_svm_intercept(svm);
1029
1030         if (vmexit == NESTED_EXIT_DONE)
1031                 nested_svm_vmexit(svm);
1032
1033         return vmexit;
1034 }
1035
1036 int nested_svm_check_permissions(struct kvm_vcpu *vcpu)
1037 {
1038         if (!(vcpu->arch.efer & EFER_SVME) || !is_paging(vcpu)) {
1039                 kvm_queue_exception(vcpu, UD_VECTOR);
1040                 return 1;
1041         }
1042
1043         if (to_svm(vcpu)->vmcb->save.cpl) {
1044                 kvm_inject_gp(vcpu, 0);
1045                 return 1;
1046         }
1047
1048         return 0;
1049 }
1050
1051 static bool nested_exit_on_exception(struct vcpu_svm *svm)
1052 {
1053         unsigned int nr = svm->vcpu.arch.exception.nr;
1054
1055         return (svm->nested.ctl.intercepts[INTERCEPT_EXCEPTION] & BIT(nr));
1056 }
1057
1058 static void nested_svm_inject_exception_vmexit(struct vcpu_svm *svm)
1059 {
1060         unsigned int nr = svm->vcpu.arch.exception.nr;
1061
1062         svm->vmcb->control.exit_code = SVM_EXIT_EXCP_BASE + nr;
1063         svm->vmcb->control.exit_code_hi = 0;
1064
1065         if (svm->vcpu.arch.exception.has_error_code)
1066                 svm->vmcb->control.exit_info_1 = svm->vcpu.arch.exception.error_code;
1067
1068         /*
1069          * EXITINFO2 is undefined for all exception intercepts other
1070          * than #PF.
1071          */
1072         if (nr == PF_VECTOR) {
1073                 if (svm->vcpu.arch.exception.nested_apf)
1074                         svm->vmcb->control.exit_info_2 = svm->vcpu.arch.apf.nested_apf_token;
1075                 else if (svm->vcpu.arch.exception.has_payload)
1076                         svm->vmcb->control.exit_info_2 = svm->vcpu.arch.exception.payload;
1077                 else
1078                         svm->vmcb->control.exit_info_2 = svm->vcpu.arch.cr2;
1079         } else if (nr == DB_VECTOR) {
1080                 /* See inject_pending_event.  */
1081                 kvm_deliver_exception_payload(&svm->vcpu);
1082                 if (svm->vcpu.arch.dr7 & DR7_GD) {
1083                         svm->vcpu.arch.dr7 &= ~DR7_GD;
1084                         kvm_update_dr7(&svm->vcpu);
1085                 }
1086         } else
1087                 WARN_ON(svm->vcpu.arch.exception.has_payload);
1088
1089         nested_svm_vmexit(svm);
1090 }
1091
1092 static inline bool nested_exit_on_init(struct vcpu_svm *svm)
1093 {
1094         return vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_INIT);
1095 }
1096
1097 static int svm_check_nested_events(struct kvm_vcpu *vcpu)
1098 {
1099         struct vcpu_svm *svm = to_svm(vcpu);
1100         bool block_nested_events =
1101                 kvm_event_needs_reinjection(vcpu) || svm->nested.nested_run_pending;
1102         struct kvm_lapic *apic = vcpu->arch.apic;
1103
1104         if (lapic_in_kernel(vcpu) &&
1105             test_bit(KVM_APIC_INIT, &apic->pending_events)) {
1106                 if (block_nested_events)
1107                         return -EBUSY;
1108                 if (!nested_exit_on_init(svm))
1109                         return 0;
1110                 nested_svm_simple_vmexit(svm, SVM_EXIT_INIT);
1111                 return 0;
1112         }
1113
1114         if (vcpu->arch.exception.pending) {
1115                 /*
1116                  * Only a pending nested run can block a pending exception.
1117                  * Otherwise an injected NMI/interrupt should either be
1118                  * lost or delivered to the nested hypervisor in the EXITINTINFO
1119                  * vmcb field, while delivering the pending exception.
1120                  */
1121                 if (svm->nested.nested_run_pending)
1122                         return -EBUSY;
1123                 if (!nested_exit_on_exception(svm))
1124                         return 0;
1125                 nested_svm_inject_exception_vmexit(svm);
1126                 return 0;
1127         }
1128
1129         if (vcpu->arch.smi_pending && !svm_smi_blocked(vcpu)) {
1130                 if (block_nested_events)
1131                         return -EBUSY;
1132                 if (!nested_exit_on_smi(svm))
1133                         return 0;
1134                 nested_svm_simple_vmexit(svm, SVM_EXIT_SMI);
1135                 return 0;
1136         }
1137
1138         if (vcpu->arch.nmi_pending && !svm_nmi_blocked(vcpu)) {
1139                 if (block_nested_events)
1140                         return -EBUSY;
1141                 if (!nested_exit_on_nmi(svm))
1142                         return 0;
1143                 nested_svm_simple_vmexit(svm, SVM_EXIT_NMI);
1144                 return 0;
1145         }
1146
1147         if (kvm_cpu_has_interrupt(vcpu) && !svm_interrupt_blocked(vcpu)) {
1148                 if (block_nested_events)
1149                         return -EBUSY;
1150                 if (!nested_exit_on_intr(svm))
1151                         return 0;
1152                 trace_kvm_nested_intr_vmexit(svm->vmcb->save.rip);
1153                 nested_svm_simple_vmexit(svm, SVM_EXIT_INTR);
1154                 return 0;
1155         }
1156
1157         return 0;
1158 }
1159
1160 int nested_svm_exit_special(struct vcpu_svm *svm)
1161 {
1162         u32 exit_code = svm->vmcb->control.exit_code;
1163
1164         switch (exit_code) {
1165         case SVM_EXIT_INTR:
1166         case SVM_EXIT_NMI:
1167         case SVM_EXIT_NPF:
1168                 return NESTED_EXIT_HOST;
1169         case SVM_EXIT_EXCP_BASE ... SVM_EXIT_EXCP_BASE + 0x1f: {
1170                 u32 excp_bits = 1 << (exit_code - SVM_EXIT_EXCP_BASE);
1171
1172                 if (svm->vmcb01.ptr->control.intercepts[INTERCEPT_EXCEPTION] &
1173                     excp_bits)
1174                         return NESTED_EXIT_HOST;
1175                 else if (exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR &&
1176                          svm->vcpu.arch.apf.host_apf_flags)
1177                         /* Trap async PF even if not shadowing */
1178                         return NESTED_EXIT_HOST;
1179                 break;
1180         }
1181         default:
1182                 break;
1183         }
1184
1185         return NESTED_EXIT_CONTINUE;
1186 }
1187
1188 static int svm_get_nested_state(struct kvm_vcpu *vcpu,
1189                                 struct kvm_nested_state __user *user_kvm_nested_state,
1190                                 u32 user_data_size)
1191 {
1192         struct vcpu_svm *svm;
1193         struct kvm_nested_state kvm_state = {
1194                 .flags = 0,
1195                 .format = KVM_STATE_NESTED_FORMAT_SVM,
1196                 .size = sizeof(kvm_state),
1197         };
1198         struct vmcb __user *user_vmcb = (struct vmcb __user *)
1199                 &user_kvm_nested_state->data.svm[0];
1200
1201         if (!vcpu)
1202                 return kvm_state.size + KVM_STATE_NESTED_SVM_VMCB_SIZE;
1203
1204         svm = to_svm(vcpu);
1205
1206         if (user_data_size < kvm_state.size)
1207                 goto out;
1208
1209         /* First fill in the header and copy it out.  */
1210         if (is_guest_mode(vcpu)) {
1211                 kvm_state.hdr.svm.vmcb_pa = svm->nested.vmcb12_gpa;
1212                 kvm_state.size += KVM_STATE_NESTED_SVM_VMCB_SIZE;
1213                 kvm_state.flags |= KVM_STATE_NESTED_GUEST_MODE;
1214
1215                 if (svm->nested.nested_run_pending)
1216                         kvm_state.flags |= KVM_STATE_NESTED_RUN_PENDING;
1217         }
1218
1219         if (gif_set(svm))
1220                 kvm_state.flags |= KVM_STATE_NESTED_GIF_SET;
1221
1222         if (copy_to_user(user_kvm_nested_state, &kvm_state, sizeof(kvm_state)))
1223                 return -EFAULT;
1224
1225         if (!is_guest_mode(vcpu))
1226                 goto out;
1227
1228         /*
1229          * Copy over the full size of the VMCB rather than just the size
1230          * of the structs.
1231          */
1232         if (clear_user(user_vmcb, KVM_STATE_NESTED_SVM_VMCB_SIZE))
1233                 return -EFAULT;
1234         if (copy_to_user(&user_vmcb->control, &svm->nested.ctl,
1235                          sizeof(user_vmcb->control)))
1236                 return -EFAULT;
1237         if (copy_to_user(&user_vmcb->save, &svm->vmcb01.ptr->save,
1238                          sizeof(user_vmcb->save)))
1239                 return -EFAULT;
1240 out:
1241         return kvm_state.size;
1242 }
1243
1244 static int svm_set_nested_state(struct kvm_vcpu *vcpu,
1245                                 struct kvm_nested_state __user *user_kvm_nested_state,
1246                                 struct kvm_nested_state *kvm_state)
1247 {
1248         struct vcpu_svm *svm = to_svm(vcpu);
1249         struct vmcb __user *user_vmcb = (struct vmcb __user *)
1250                 &user_kvm_nested_state->data.svm[0];
1251         struct vmcb_control_area *ctl;
1252         struct vmcb_save_area *save;
1253         unsigned long cr0;
1254         int ret;
1255
1256         BUILD_BUG_ON(sizeof(struct vmcb_control_area) + sizeof(struct vmcb_save_area) >
1257                      KVM_STATE_NESTED_SVM_VMCB_SIZE);
1258
1259         if (kvm_state->format != KVM_STATE_NESTED_FORMAT_SVM)
1260                 return -EINVAL;
1261
1262         if (kvm_state->flags & ~(KVM_STATE_NESTED_GUEST_MODE |
1263                                  KVM_STATE_NESTED_RUN_PENDING |
1264                                  KVM_STATE_NESTED_GIF_SET))
1265                 return -EINVAL;
1266
1267         /*
1268          * If in guest mode, vcpu->arch.efer actually refers to the L2 guest's
1269          * EFER.SVME, but EFER.SVME still has to be 1 for VMRUN to succeed.
1270          */
1271         if (!(vcpu->arch.efer & EFER_SVME)) {
1272                 /* GIF=1 and no guest mode are required if SVME=0.  */
1273                 if (kvm_state->flags != KVM_STATE_NESTED_GIF_SET)
1274                         return -EINVAL;
1275         }
1276
1277         /* SMM temporarily disables SVM, so we cannot be in guest mode.  */
1278         if (is_smm(vcpu) && (kvm_state->flags & KVM_STATE_NESTED_GUEST_MODE))
1279                 return -EINVAL;
1280
1281         if (!(kvm_state->flags & KVM_STATE_NESTED_GUEST_MODE)) {
1282                 svm_leave_nested(svm);
1283                 svm_set_gif(svm, !!(kvm_state->flags & KVM_STATE_NESTED_GIF_SET));
1284                 return 0;
1285         }
1286
1287         if (!page_address_valid(vcpu, kvm_state->hdr.svm.vmcb_pa))
1288                 return -EINVAL;
1289         if (kvm_state->size < sizeof(*kvm_state) + KVM_STATE_NESTED_SVM_VMCB_SIZE)
1290                 return -EINVAL;
1291
1292         ret  = -ENOMEM;
1293         ctl  = kzalloc(sizeof(*ctl),  GFP_KERNEL_ACCOUNT);
1294         save = kzalloc(sizeof(*save), GFP_KERNEL_ACCOUNT);
1295         if (!ctl || !save)
1296                 goto out_free;
1297
1298         ret = -EFAULT;
1299         if (copy_from_user(ctl, &user_vmcb->control, sizeof(*ctl)))
1300                 goto out_free;
1301         if (copy_from_user(save, &user_vmcb->save, sizeof(*save)))
1302                 goto out_free;
1303
1304         ret = -EINVAL;
1305         if (!nested_vmcb_check_controls(vcpu, ctl))
1306                 goto out_free;
1307
1308         /*
1309          * Processor state contains L2 state.  Check that it is
1310          * valid for guest mode (see nested_vmcb_check_save).
1311          */
1312         cr0 = kvm_read_cr0(vcpu);
1313         if (((cr0 & X86_CR0_CD) == 0) && (cr0 & X86_CR0_NW))
1314                 goto out_free;
1315
1316         /*
1317          * Validate host state saved from before VMRUN (see
1318          * nested_svm_check_permissions).
1319          */
1320         if (!(save->cr0 & X86_CR0_PG) ||
1321             !(save->cr0 & X86_CR0_PE) ||
1322             (save->rflags & X86_EFLAGS_VM) ||
1323             !nested_vmcb_valid_sregs(vcpu, save))
1324                 goto out_free;
1325
1326         /*
1327          * While the nested guest CR3 is already checked and set by
1328          * KVM_SET_SREGS, it was set when nested state was yet loaded,
1329          * thus MMU might not be initialized correctly.
1330          * Set it again to fix this.
1331          */
1332
1333         ret = nested_svm_load_cr3(&svm->vcpu, vcpu->arch.cr3,
1334                                   nested_npt_enabled(svm), false);
1335         if (WARN_ON_ONCE(ret))
1336                 goto out_free;
1337
1338
1339         /*
1340          * All checks done, we can enter guest mode. Userspace provides
1341          * vmcb12.control, which will be combined with L1 and stored into
1342          * vmcb02, and the L1 save state which we store in vmcb01.
1343          * L2 registers if needed are moved from the current VMCB to VMCB02.
1344          */
1345
1346         if (is_guest_mode(vcpu))
1347                 svm_leave_nested(svm);
1348         else
1349                 svm->nested.vmcb02.ptr->save = svm->vmcb01.ptr->save;
1350
1351         svm_set_gif(svm, !!(kvm_state->flags & KVM_STATE_NESTED_GIF_SET));
1352
1353         svm->nested.nested_run_pending =
1354                 !!(kvm_state->flags & KVM_STATE_NESTED_RUN_PENDING);
1355
1356         svm->nested.vmcb12_gpa = kvm_state->hdr.svm.vmcb_pa;
1357
1358         svm->vmcb01.ptr->save.es = save->es;
1359         svm->vmcb01.ptr->save.cs = save->cs;
1360         svm->vmcb01.ptr->save.ss = save->ss;
1361         svm->vmcb01.ptr->save.ds = save->ds;
1362         svm->vmcb01.ptr->save.gdtr = save->gdtr;
1363         svm->vmcb01.ptr->save.idtr = save->idtr;
1364         svm->vmcb01.ptr->save.rflags = save->rflags | X86_EFLAGS_FIXED;
1365         svm->vmcb01.ptr->save.efer = save->efer;
1366         svm->vmcb01.ptr->save.cr0 = save->cr0;
1367         svm->vmcb01.ptr->save.cr3 = save->cr3;
1368         svm->vmcb01.ptr->save.cr4 = save->cr4;
1369         svm->vmcb01.ptr->save.rax = save->rax;
1370         svm->vmcb01.ptr->save.rsp = save->rsp;
1371         svm->vmcb01.ptr->save.rip = save->rip;
1372         svm->vmcb01.ptr->save.cpl = 0;
1373
1374         nested_load_control_from_vmcb12(svm, ctl);
1375
1376         svm_switch_vmcb(svm, &svm->nested.vmcb02);
1377
1378         nested_vmcb02_prepare_control(svm);
1379
1380         kvm_make_request(KVM_REQ_GET_NESTED_STATE_PAGES, vcpu);
1381         ret = 0;
1382 out_free:
1383         kfree(save);
1384         kfree(ctl);
1385
1386         return ret;
1387 }
1388
1389 static bool svm_get_nested_state_pages(struct kvm_vcpu *vcpu)
1390 {
1391         struct vcpu_svm *svm = to_svm(vcpu);
1392
1393         if (WARN_ON(!is_guest_mode(vcpu)))
1394                 return true;
1395
1396         if (!vcpu->arch.pdptrs_from_userspace &&
1397             !nested_npt_enabled(svm) && is_pae_paging(vcpu))
1398                 /*
1399                  * Reload the guest's PDPTRs since after a migration
1400                  * the guest CR3 might be restored prior to setting the nested
1401                  * state which can lead to a load of wrong PDPTRs.
1402                  */
1403                 if (CC(!load_pdptrs(vcpu, vcpu->arch.walk_mmu, vcpu->arch.cr3)))
1404                         return false;
1405
1406         if (!nested_svm_vmrun_msrpm(svm)) {
1407                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
1408                 vcpu->run->internal.suberror =
1409                         KVM_INTERNAL_ERROR_EMULATION;
1410                 vcpu->run->internal.ndata = 0;
1411                 return false;
1412         }
1413
1414         return true;
1415 }
1416
1417 struct kvm_x86_nested_ops svm_nested_ops = {
1418         .check_events = svm_check_nested_events,
1419         .triple_fault = nested_svm_triple_fault,
1420         .get_nested_state_pages = svm_get_nested_state_pages,
1421         .get_state = svm_get_nested_state,
1422         .set_state = svm_set_nested_state,
1423 };