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         /* If SMI is not intercepted, ignore guest SMI intercept as well  */
 159         if (!intercept_smi)
 160                 vmcb_clr_intercept(c, INTERCEPT_SMI);
 161
 162         vmcb_set_intercept(c, INTERCEPT_VMLOAD);
 163         vmcb_set_intercept(c, INTERCEPT_VMSAVE);
 164 }
 165
 166 static void copy_vmcb_control_area(struct vmcb_control_area *dst,
 167                                    struct vmcb_control_area *from)
 168 {
 169         unsigned int i;
 170
 171         for (i = 0; i < MAX_INTERCEPT; i++)
 172                 dst->intercepts[i] = from->intercepts[i];
 173
 174         dst->iopm_base_pa         = from->iopm_base_pa;
 175         dst->msrpm_base_pa        = from->msrpm_base_pa;
 176         dst->tsc_offset           = from->tsc_offset;
 177         /* asid not copied, it is handled manually for svm->vmcb.  */
 178         dst->tlb_ctl              = from->tlb_ctl;
 179         dst->int_ctl              = from->int_ctl;
 180         dst->int_vector           = from->int_vector;
 181         dst->int_state            = from->int_state;
 182         dst->exit_code            = from->exit_code;
 183         dst->exit_code_hi         = from->exit_code_hi;
 184         dst->exit_info_1          = from->exit_info_1;
 185         dst->exit_info_2          = from->exit_info_2;
 186         dst->exit_int_info        = from->exit_int_info;
 187         dst->exit_int_info_err    = from->exit_int_info_err;
 188         dst->nested_ctl           = from->nested_ctl;
 189         dst->event_inj            = from->event_inj;
 190         dst->event_inj_err        = from->event_inj_err;
 191         dst->nested_cr3           = from->nested_cr3;
 192         dst->virt_ext              = from->virt_ext;
 193         dst->pause_filter_count   = from->pause_filter_count;
 194         dst->pause_filter_thresh  = from->pause_filter_thresh;
 195 }
 196
 197 static bool nested_svm_vmrun_msrpm(struct vcpu_svm *svm)
 198 {
 199         /*
 200          * This function merges the msr permission bitmaps of kvm and the
 201          * nested vmcb. It is optimized in that it only merges the parts where
 202          * the kvm msr permission bitmap may contain zero bits
 203          */
 204         int i;
 205
 206         if (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_MSR_PROT)))
 207                 return true;
 208
 209         for (i = 0; i < MSRPM_OFFSETS; i++) {
 210                 u32 value, p;
 211                 u64 offset;
 212
 213                 if (msrpm_offsets[i] == 0xffffffff)
 214                         break;
 215
 216                 p      = msrpm_offsets[i];
 217                 offset = svm->nested.ctl.msrpm_base_pa + (p * 4);
 218
 219                 if (kvm_vcpu_read_guest(&svm->vcpu, offset, &value, 4))
 220                         return false;
 221
 222                 svm->nested.msrpm[p] = svm->msrpm[p] | value;
 223         }
 224
 225         svm->vmcb->control.msrpm_base_pa = __sme_set(__pa(svm->nested.msrpm));
 226
 227         return true;
 228 }
 229
 230 /*
 231  * Bits 11:0 of bitmap address are ignored by hardware
 232  */
 233 static bool nested_svm_check_bitmap_pa(struct kvm_vcpu *vcpu, u64 pa, u32 size)
 234 {
 235         u64 addr = PAGE_ALIGN(pa);
 236
 237         return kvm_vcpu_is_legal_gpa(vcpu, addr) &&
 238             kvm_vcpu_is_legal_gpa(vcpu, addr + size - 1);
 239 }
 240
 241 static bool nested_vmcb_check_controls(struct kvm_vcpu *vcpu,
 242                                        struct vmcb_control_area *control)
 243 {
 244         if (CC(!vmcb_is_intercept(control, INTERCEPT_VMRUN)))
 245                 return false;
 246
 247         if (CC(control->asid == 0))
 248                 return false;
 249
 250         if (CC((control->nested_ctl & SVM_NESTED_CTL_NP_ENABLE) && !npt_enabled))
 251                 return false;
 252
 253         if (CC(!nested_svm_check_bitmap_pa(vcpu, control->msrpm_base_pa,
 254                                            MSRPM_SIZE)))
 255                 return false;
 256         if (CC(!nested_svm_check_bitmap_pa(vcpu, control->iopm_base_pa,
 257                                            IOPM_SIZE)))
 258                 return false;
 259
 260         return true;
 261 }
 262
 263 static bool nested_vmcb_check_cr3_cr4(struct kvm_vcpu *vcpu,
 264                                       struct vmcb_save_area *save)
 265 {
 266         /*
 267          * These checks are also performed by KVM_SET_SREGS,
 268          * except that EFER.LMA is not checked by SVM against
 269          * CR0.PG && EFER.LME.
 270          */
 271         if ((save->efer & EFER_LME) && (save->cr0 & X86_CR0_PG)) {
 272                 if (CC(!(save->cr4 & X86_CR4_PAE)) ||
 273                     CC(!(save->cr0 & X86_CR0_PE)) ||
 274                     CC(kvm_vcpu_is_illegal_gpa(vcpu, save->cr3)))
 275                         return false;
 276         }
 277
 278         if (CC(!kvm_is_valid_cr4(vcpu, save->cr4)))
 279                 return false;
 280
 281         return true;
 282 }
 283
 284 /* Common checks that apply to both L1 and L2 state.  */
 285 static bool nested_vmcb_valid_sregs(struct kvm_vcpu *vcpu,
 286                                     struct vmcb_save_area *save)
 287 {
 288         /*
 289          * FIXME: these should be done after copying the fields,
 290          * to avoid TOC/TOU races.  For these save area checks
 291          * the possible damage is limited since kvm_set_cr0 and
 292          * kvm_set_cr4 handle failure; EFER_SVME is an exception
 293          * so it is force-set later in nested_prepare_vmcb_save.
 294          */
 295         if (CC(!(save->efer & EFER_SVME)))
 296                 return false;
 297
 298         if (CC((save->cr0 & X86_CR0_CD) == 0 && (save->cr0 & X86_CR0_NW)) ||
 299             CC(save->cr0 & ~0xffffffffULL))
 300                 return false;
 301
 302         if (CC(!kvm_dr6_valid(save->dr6)) || CC(!kvm_dr7_valid(save->dr7)))
 303                 return false;
 304
 305         if (!nested_vmcb_check_cr3_cr4(vcpu, save))
 306                 return false;
 307
 308         if (CC(!kvm_valid_efer(vcpu, save->efer)))
 309                 return false;
 310
 311         return true;
 312 }
 313
 314 void nested_load_control_from_vmcb12(struct vcpu_svm *svm,
 315                                      struct vmcb_control_area *control)
 316 {
 317         copy_vmcb_control_area(&svm->nested.ctl, control);
 318
 319         /* Copy it here because nested_svm_check_controls will check it.  */
 320         svm->nested.ctl.asid           = control->asid;
 321         svm->nested.ctl.msrpm_base_pa &= ~0x0fffULL;
 322         svm->nested.ctl.iopm_base_pa  &= ~0x0fffULL;
 323 }
 324
 325 /*
 326  * Synchronize fields that are written by the processor, so that
 327  * they can be copied back into the vmcb12.
 328  */
 329 void nested_sync_control_from_vmcb02(struct vcpu_svm *svm)
 330 {
 331         u32 mask;
 332         svm->nested.ctl.event_inj      = svm->vmcb->control.event_inj;
 333         svm->nested.ctl.event_inj_err  = svm->vmcb->control.event_inj_err;
 334
 335         /* Only a few fields of int_ctl are written by the processor.  */
 336         mask = V_IRQ_MASK | V_TPR_MASK;
 337         if (!(svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK) &&
 338             svm_is_intercept(svm, INTERCEPT_VINTR)) {
 339                 /*
 340                  * In order to request an interrupt window, L0 is usurping
 341                  * svm->vmcb->control.int_ctl and possibly setting V_IRQ
 342                  * even if it was clear in L1's VMCB.  Restoring it would be
 343                  * wrong.  However, in this case V_IRQ will remain true until
 344                  * interrupt_window_interception calls svm_clear_vintr and
 345                  * restores int_ctl.  We can just leave it aside.
 346                  */
 347                 mask &= ~V_IRQ_MASK;
 348         }
 349         svm->nested.ctl.int_ctl        &= ~mask;
 350         svm->nested.ctl.int_ctl        |= svm->vmcb->control.int_ctl & mask;
 351 }
 352
 353 /*
 354  * Transfer any event that L0 or L1 wanted to inject into L2 to
 355  * EXIT_INT_INFO.
 356  */
 357 static void nested_save_pending_event_to_vmcb12(struct vcpu_svm *svm,
 358                                                 struct vmcb *vmcb12)
 359 {
 360         struct kvm_vcpu *vcpu = &svm->vcpu;
 361         u32 exit_int_info = 0;
 362         unsigned int nr;
 363
 364         if (vcpu->arch.exception.injected) {
 365                 nr = vcpu->arch.exception.nr;
 366                 exit_int_info = nr | SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_EXEPT;
 367
 368                 if (vcpu->arch.exception.has_error_code) {
 369                         exit_int_info |= SVM_EVTINJ_VALID_ERR;
 370                         vmcb12->control.exit_int_info_err =
 371                                 vcpu->arch.exception.error_code;
 372                 }
 373
 374         } else if (vcpu->arch.nmi_injected) {
 375                 exit_int_info = SVM_EVTINJ_VALID | SVM_EVTINJ_TYPE_NMI;
 376
 377         } else if (vcpu->arch.interrupt.injected) {
 378                 nr = vcpu->arch.interrupt.nr;
 379                 exit_int_info = nr | SVM_EVTINJ_VALID;
 380
 381                 if (vcpu->arch.interrupt.soft)
 382                         exit_int_info |= SVM_EVTINJ_TYPE_SOFT;
 383                 else
 384                         exit_int_info |= SVM_EVTINJ_TYPE_INTR;
 385         }
 386
 387         vmcb12->control.exit_int_info = exit_int_info;
 388 }
 389
 390 static inline bool nested_npt_enabled(struct vcpu_svm *svm)
 391 {
 392         return svm->nested.ctl.nested_ctl & SVM_NESTED_CTL_NP_ENABLE;
 393 }
 394
 395 static void nested_svm_transition_tlb_flush(struct kvm_vcpu *vcpu)
 396 {
 397         /*
 398          * TODO: optimize unconditional TLB flush/MMU sync.  A partial list of
 399          * things to fix before this can be conditional:
 400          *
 401          *  - Flush TLBs for both L1 and L2 remote TLB flush
 402          *  - Honor L1's request to flush an ASID on nested VMRUN
 403          *  - Sync nested NPT MMU on VMRUN that flushes L2's ASID[*]
 404          *  - Don't crush a pending TLB flush in vmcb02 on nested VMRUN
 405          *  - Flush L1's ASID on KVM_REQ_TLB_FLUSH_GUEST
 406          *
 407          * [*] Unlike nested EPT, SVM's ASID management can invalidate nested
 408          *     NPT guest-physical mappings on VMRUN.
 409          */
 410         kvm_make_request(KVM_REQ_MMU_SYNC, vcpu);
 411         kvm_make_request(KVM_REQ_TLB_FLUSH_CURRENT, vcpu);
 412 }
 413
 414 /*
 415  * Load guest's/host's cr3 on nested vmentry or vmexit. @nested_npt is true
 416  * if we are emulating VM-Entry into a guest with NPT enabled.
 417  */
 418 static int nested_svm_load_cr3(struct kvm_vcpu *vcpu, unsigned long cr3,
 419                                bool nested_npt, bool reload_pdptrs)
 420 {
 421         if (CC(kvm_vcpu_is_illegal_gpa(vcpu, cr3)))
 422                 return -EINVAL;
 423
 424         if (reload_pdptrs && !nested_npt && is_pae_paging(vcpu) &&
 425             CC(!load_pdptrs(vcpu, vcpu->arch.walk_mmu, cr3)))
 426                 return -EINVAL;
 427
 428         if (!nested_npt)
 429                 kvm_mmu_new_pgd(vcpu, cr3);
 430
 431         vcpu->arch.cr3 = cr3;
 432         kvm_register_mark_available(vcpu, VCPU_EXREG_CR3);
 433
 434         /* Re-initialize the MMU, e.g. to pick up CR4 MMU role changes. */
 435         kvm_init_mmu(vcpu);
 436
 437         return 0;
 438 }
 439
 440 void nested_vmcb02_compute_g_pat(struct vcpu_svm *svm)
 441 {
 442         if (!svm->nested.vmcb02.ptr)
 443                 return;
 444
 445         /* FIXME: merge g_pat from vmcb01 and vmcb12.  */
 446         svm->nested.vmcb02.ptr->save.g_pat = svm->vmcb01.ptr->save.g_pat;
 447 }
 448
 449 static void nested_vmcb02_prepare_save(struct vcpu_svm *svm, struct vmcb *vmcb12)
 450 {
 451         bool new_vmcb12 = false;
 452
 453         nested_vmcb02_compute_g_pat(svm);
 454
 455         /* Load the nested guest state */
 456         if (svm->nested.vmcb12_gpa != svm->nested.last_vmcb12_gpa) {
 457                 new_vmcb12 = true;
 458                 svm->nested.last_vmcb12_gpa = svm->nested.vmcb12_gpa;
 459         }
 460
 461         if (unlikely(new_vmcb12 || vmcb_is_dirty(vmcb12, VMCB_SEG))) {
 462                 svm->vmcb->save.es = vmcb12->save.es;
 463                 svm->vmcb->save.cs = vmcb12->save.cs;
 464                 svm->vmcb->save.ss = vmcb12->save.ss;
 465                 svm->vmcb->save.ds = vmcb12->save.ds;
 466                 svm->vmcb->save.cpl = vmcb12->save.cpl;
 467                 vmcb_mark_dirty(svm->vmcb, VMCB_SEG);
 468         }
 469
 470         if (unlikely(new_vmcb12 || vmcb_is_dirty(vmcb12, VMCB_DT))) {
 471                 svm->vmcb->save.gdtr = vmcb12->save.gdtr;
 472                 svm->vmcb->save.idtr = vmcb12->save.idtr;
 473                 vmcb_mark_dirty(svm->vmcb, VMCB_DT);
 474         }
 475
 476         kvm_set_rflags(&svm->vcpu, vmcb12->save.rflags | X86_EFLAGS_FIXED);
 477
 478         /*
 479          * Force-set EFER_SVME even though it is checked earlier on the
 480          * VMCB12, because the guest can flip the bit between the check
 481          * and now.  Clearing EFER_SVME would call svm_free_nested.
 482          */
 483         svm_set_efer(&svm->vcpu, vmcb12->save.efer | EFER_SVME);
 484
 485         svm_set_cr0(&svm->vcpu, vmcb12->save.cr0);
 486         svm_set_cr4(&svm->vcpu, vmcb12->save.cr4);
 487
 488         svm->vcpu.arch.cr2 = vmcb12->save.cr2;
 489
 490         kvm_rax_write(&svm->vcpu, vmcb12->save.rax);
 491         kvm_rsp_write(&svm->vcpu, vmcb12->save.rsp);
 492         kvm_rip_write(&svm->vcpu, vmcb12->save.rip);
 493
 494         /* In case we don't even reach vcpu_run, the fields are not updated */
 495         svm->vmcb->save.rax = vmcb12->save.rax;
 496         svm->vmcb->save.rsp = vmcb12->save.rsp;
 497         svm->vmcb->save.rip = vmcb12->save.rip;
 498
 499         /* These bits will be set properly on the first execution when new_vmc12 is true */
 500         if (unlikely(new_vmcb12 || vmcb_is_dirty(vmcb12, VMCB_DR))) {
 501                 svm->vmcb->save.dr7 = vmcb12->save.dr7 | DR7_FIXED_1;
 502                 svm->vcpu.arch.dr6  = vmcb12->save.dr6 | DR6_ACTIVE_LOW;
 503                 vmcb_mark_dirty(svm->vmcb, VMCB_DR);
 504         }
 505 }
 506
 507 static void nested_vmcb02_prepare_control(struct vcpu_svm *svm)
 508 {
 509         const u32 int_ctl_vmcb01_bits =
 510                 V_INTR_MASKING_MASK | V_GIF_MASK | V_GIF_ENABLE_MASK;
 511
 512         const u32 int_ctl_vmcb12_bits = V_TPR_MASK | V_IRQ_INJECTION_BITS_MASK;
 513
 514         struct kvm_vcpu *vcpu = &svm->vcpu;
 515
 516         /*
 517          * Filled at exit: exit_code, exit_code_hi, exit_info_1, exit_info_2,
 518          * exit_int_info, exit_int_info_err, next_rip, insn_len, insn_bytes.
 519          */
 520
 521         /*
 522          * Also covers avic_vapic_bar, avic_backing_page, avic_logical_id,
 523          * avic_physical_id.
 524          */
 525         WARN_ON(kvm_apicv_activated(svm->vcpu.kvm));
 526
 527         /* Copied from vmcb01.  msrpm_base can be overwritten later.  */
 528         svm->vmcb->control.nested_ctl = svm->vmcb01.ptr->control.nested_ctl;
 529         svm->vmcb->control.iopm_base_pa = svm->vmcb01.ptr->control.iopm_base_pa;
 530         svm->vmcb->control.msrpm_base_pa = svm->vmcb01.ptr->control.msrpm_base_pa;
 531
 532         /* Done at vmrun: asid.  */
 533
 534         /* Also overwritten later if necessary.  */
 535         svm->vmcb->control.tlb_ctl = TLB_CONTROL_DO_NOTHING;
 536
 537         /* nested_cr3.  */
 538         if (nested_npt_enabled(svm))
 539                 nested_svm_init_mmu_context(vcpu);
 540
 541         svm->vmcb->control.tsc_offset = vcpu->arch.tsc_offset =
 542                 vcpu->arch.l1_tsc_offset + svm->nested.ctl.tsc_offset;
 543
 544         svm->vmcb->control.int_ctl             =
 545                 (svm->nested.ctl.int_ctl & int_ctl_vmcb12_bits) |
 546                 (svm->vmcb01.ptr->control.int_ctl & int_ctl_vmcb01_bits);
 547
 548         svm->vmcb->control.virt_ext            = svm->nested.ctl.virt_ext;
 549         svm->vmcb->control.int_vector          = svm->nested.ctl.int_vector;
 550         svm->vmcb->control.int_state           = svm->nested.ctl.int_state;
 551         svm->vmcb->control.event_inj           = svm->nested.ctl.event_inj;
 552         svm->vmcb->control.event_inj_err       = svm->nested.ctl.event_inj_err;
 553
 554         svm->vmcb->control.pause_filter_count  = svm->nested.ctl.pause_filter_count;
 555         svm->vmcb->control.pause_filter_thresh = svm->nested.ctl.pause_filter_thresh;
 556
 557         nested_svm_transition_tlb_flush(vcpu);
 558
 559         /* Enter Guest-Mode */
 560         enter_guest_mode(vcpu);
 561
 562         /*
 563          * Merge guest and host intercepts - must be called with vcpu in
 564          * guest-mode to take effect.
 565          */
 566         recalc_intercepts(svm);
 567 }
 568
 569 static void nested_svm_copy_common_state(struct vmcb *from_vmcb, struct vmcb *to_vmcb)
 570 {
 571         /*
 572          * Some VMCB state is shared between L1 and L2 and thus has to be
 573          * moved at the time of nested vmrun and vmexit.
 574          *
 575          * VMLOAD/VMSAVE state would also belong in this category, but KVM
 576          * always performs VMLOAD and VMSAVE from the VMCB01.
 577          */
 578         to_vmcb->save.spec_ctrl = from_vmcb->save.spec_ctrl;
 579 }
 580
 581 int enter_svm_guest_mode(struct kvm_vcpu *vcpu, u64 vmcb12_gpa,
 582                          struct vmcb *vmcb12)
 583 {
 584         struct vcpu_svm *svm = to_svm(vcpu);
 585         int ret;
 586
 587         trace_kvm_nested_vmrun(svm->vmcb->save.rip, vmcb12_gpa,
 588                                vmcb12->save.rip,
 589                                vmcb12->control.int_ctl,
 590                                vmcb12->control.event_inj,
 591                                vmcb12->control.nested_ctl);
 592
 593         trace_kvm_nested_intercepts(vmcb12->control.intercepts[INTERCEPT_CR] & 0xffff,
 594                                     vmcb12->control.intercepts[INTERCEPT_CR] >> 16,
 595                                     vmcb12->control.intercepts[INTERCEPT_EXCEPTION],
 596                                     vmcb12->control.intercepts[INTERCEPT_WORD3],
 597                                     vmcb12->control.intercepts[INTERCEPT_WORD4],
 598                                     vmcb12->control.intercepts[INTERCEPT_WORD5]);
 599
 600
 601         svm->nested.vmcb12_gpa = vmcb12_gpa;
 602
 603         WARN_ON(svm->vmcb == svm->nested.vmcb02.ptr);
 604
 605         nested_svm_copy_common_state(svm->vmcb01.ptr, svm->nested.vmcb02.ptr);
 606
 607         svm_switch_vmcb(svm, &svm->nested.vmcb02);
 608         nested_vmcb02_prepare_control(svm);
 609         nested_vmcb02_prepare_save(svm, vmcb12);
 610
 611         ret = nested_svm_load_cr3(&svm->vcpu, vmcb12->save.cr3,
 612                                   nested_npt_enabled(svm), true);
 613         if (ret)
 614                 return ret;
 615
 616         if (!npt_enabled)
 617                 vcpu->arch.mmu->inject_page_fault = svm_inject_page_fault_nested;
 618
 619         svm_set_gif(svm, true);
 620
 621         return 0;
 622 }
 623
 624 int nested_svm_vmrun(struct kvm_vcpu *vcpu)
 625 {
 626         struct vcpu_svm *svm = to_svm(vcpu);
 627         int ret;
 628         struct vmcb *vmcb12;
 629         struct kvm_host_map map;
 630         u64 vmcb12_gpa;
 631
 632         if (!svm->nested.hsave_msr) {
 633                 kvm_inject_gp(vcpu, 0);
 634                 return 1;
 635         }
 636
 637         if (is_smm(vcpu)) {
 638                 kvm_queue_exception(vcpu, UD_VECTOR);
 639                 return 1;
 640         }
 641
 642         vmcb12_gpa = svm->vmcb->save.rax;
 643         ret = kvm_vcpu_map(vcpu, gpa_to_gfn(vmcb12_gpa), &map);
 644         if (ret == -EINVAL) {
 645                 kvm_inject_gp(vcpu, 0);
 646                 return 1;
 647         } else if (ret) {
 648                 return kvm_skip_emulated_instruction(vcpu);
 649         }
 650
 651         ret = kvm_skip_emulated_instruction(vcpu);
 652
 653         vmcb12 = map.hva;
 654
 655         if (WARN_ON_ONCE(!svm->nested.initialized))
 656                 return -EINVAL;
 657
 658         nested_load_control_from_vmcb12(svm, &vmcb12->control);
 659
 660         if (!nested_vmcb_valid_sregs(vcpu, &vmcb12->save) ||
 661             !nested_vmcb_check_controls(vcpu, &svm->nested.ctl)) {
 662                 vmcb12->control.exit_code    = SVM_EXIT_ERR;
 663                 vmcb12->control.exit_code_hi = 0;
 664                 vmcb12->control.exit_info_1  = 0;
 665                 vmcb12->control.exit_info_2  = 0;
 666                 goto out;
 667         }
 668
 669
 670         /* Clear internal status */
 671         kvm_clear_exception_queue(vcpu);
 672         kvm_clear_interrupt_queue(vcpu);
 673
 674         /*
 675          * Since vmcb01 is not in use, we can use it to store some of the L1
 676          * state.
 677          */
 678         svm->vmcb01.ptr->save.efer   = vcpu->arch.efer;
 679         svm->vmcb01.ptr->save.cr0    = kvm_read_cr0(vcpu);
 680         svm->vmcb01.ptr->save.cr4    = vcpu->arch.cr4;
 681         svm->vmcb01.ptr->save.rflags = kvm_get_rflags(vcpu);
 682         svm->vmcb01.ptr->save.rip    = kvm_rip_read(vcpu);
 683
 684         if (!npt_enabled)
 685                 svm->vmcb01.ptr->save.cr3 = kvm_read_cr3(vcpu);
 686
 687         svm->nested.nested_run_pending = 1;
 688
 689         if (enter_svm_guest_mode(vcpu, vmcb12_gpa, vmcb12))
 690                 goto out_exit_err;
 691
 692         if (nested_svm_vmrun_msrpm(svm))
 693                 goto out;
 694
 695 out_exit_err:
 696         svm->nested.nested_run_pending = 0;
 697
 698         svm->vmcb->control.exit_code    = SVM_EXIT_ERR;
 699         svm->vmcb->control.exit_code_hi = 0;
 700         svm->vmcb->control.exit_info_1  = 0;
 701         svm->vmcb->control.exit_info_2  = 0;
 702
 703         nested_svm_vmexit(svm);
 704
 705 out:
 706         kvm_vcpu_unmap(vcpu, &map, true);
 707
 708         return ret;
 709 }
 710
 711 /* Copy state save area fields which are handled by VMRUN */
 712 void svm_copy_vmrun_state(struct vmcb_save_area *to_save,
 713                           struct vmcb_save_area *from_save)
 714 {
 715         to_save->es = from_save->es;
 716         to_save->cs = from_save->cs;
 717         to_save->ss = from_save->ss;
 718         to_save->ds = from_save->ds;
 719         to_save->gdtr = from_save->gdtr;
 720         to_save->idtr = from_save->idtr;
 721         to_save->rflags = from_save->rflags | X86_EFLAGS_FIXED;
 722         to_save->efer = from_save->efer;
 723         to_save->cr0 = from_save->cr0;
 724         to_save->cr3 = from_save->cr3;
 725         to_save->cr4 = from_save->cr4;
 726         to_save->rax = from_save->rax;
 727         to_save->rsp = from_save->rsp;
 728         to_save->rip = from_save->rip;
 729         to_save->cpl = 0;
 730 }
 731
 732 void svm_copy_vmloadsave_state(struct vmcb *to_vmcb, struct vmcb *from_vmcb)
 733 {
 734         to_vmcb->save.fs = from_vmcb->save.fs;
 735         to_vmcb->save.gs = from_vmcb->save.gs;
 736         to_vmcb->save.tr = from_vmcb->save.tr;
 737         to_vmcb->save.ldtr = from_vmcb->save.ldtr;
 738         to_vmcb->save.kernel_gs_base = from_vmcb->save.kernel_gs_base;
 739         to_vmcb->save.star = from_vmcb->save.star;
 740         to_vmcb->save.lstar = from_vmcb->save.lstar;
 741         to_vmcb->save.cstar = from_vmcb->save.cstar;
 742         to_vmcb->save.sfmask = from_vmcb->save.sfmask;
 743         to_vmcb->save.sysenter_cs = from_vmcb->save.sysenter_cs;
 744         to_vmcb->save.sysenter_esp = from_vmcb->save.sysenter_esp;
 745         to_vmcb->save.sysenter_eip = from_vmcb->save.sysenter_eip;
 746 }
 747
 748 int nested_svm_vmexit(struct vcpu_svm *svm)
 749 {
 750         struct kvm_vcpu *vcpu = &svm->vcpu;
 751         struct vmcb *vmcb12;
 752         struct vmcb *vmcb = svm->vmcb;
 753         struct kvm_host_map map;
 754         int rc;
 755
 756         /* Triple faults in L2 should never escape. */
 757         WARN_ON_ONCE(kvm_check_request(KVM_REQ_TRIPLE_FAULT, vcpu));
 758
 759         rc = kvm_vcpu_map(vcpu, gpa_to_gfn(svm->nested.vmcb12_gpa), &map);
 760         if (rc) {
 761                 if (rc == -EINVAL)
 762                         kvm_inject_gp(vcpu, 0);
 763                 return 1;
 764         }
 765
 766         vmcb12 = map.hva;
 767
 768         /* Exit Guest-Mode */
 769         leave_guest_mode(vcpu);
 770         svm->nested.vmcb12_gpa = 0;
 771         WARN_ON_ONCE(svm->nested.nested_run_pending);
 772
 773         kvm_clear_request(KVM_REQ_GET_NESTED_STATE_PAGES, vcpu);
 774
 775         /* in case we halted in L2 */
 776         svm->vcpu.arch.mp_state = KVM_MP_STATE_RUNNABLE;
 777
 778         /* Give the current vmcb to the guest */
 779
 780         vmcb12->save.es     = vmcb->save.es;
 781         vmcb12->save.cs     = vmcb->save.cs;
 782         vmcb12->save.ss     = vmcb->save.ss;
 783         vmcb12->save.ds     = vmcb->save.ds;
 784         vmcb12->save.gdtr   = vmcb->save.gdtr;
 785         vmcb12->save.idtr   = vmcb->save.idtr;
 786         vmcb12->save.efer   = svm->vcpu.arch.efer;
 787         vmcb12->save.cr0    = kvm_read_cr0(vcpu);
 788         vmcb12->save.cr3    = kvm_read_cr3(vcpu);
 789         vmcb12->save.cr2    = vmcb->save.cr2;
 790         vmcb12->save.cr4    = svm->vcpu.arch.cr4;
 791         vmcb12->save.rflags = kvm_get_rflags(vcpu);
 792         vmcb12->save.rip    = kvm_rip_read(vcpu);
 793         vmcb12->save.rsp    = kvm_rsp_read(vcpu);
 794         vmcb12->save.rax    = kvm_rax_read(vcpu);
 795         vmcb12->save.dr7    = vmcb->save.dr7;
 796         vmcb12->save.dr6    = svm->vcpu.arch.dr6;
 797         vmcb12->save.cpl    = vmcb->save.cpl;
 798
 799         vmcb12->control.int_state         = vmcb->control.int_state;
 800         vmcb12->control.exit_code         = vmcb->control.exit_code;
 801         vmcb12->control.exit_code_hi      = vmcb->control.exit_code_hi;
 802         vmcb12->control.exit_info_1       = vmcb->control.exit_info_1;
 803         vmcb12->control.exit_info_2       = vmcb->control.exit_info_2;
 804
 805         if (vmcb12->control.exit_code != SVM_EXIT_ERR)
 806                 nested_save_pending_event_to_vmcb12(svm, vmcb12);
 807
 808         if (svm->nrips_enabled)
 809                 vmcb12->control.next_rip  = vmcb->control.next_rip;
 810
 811         vmcb12->control.int_ctl           = svm->nested.ctl.int_ctl;
 812         vmcb12->control.tlb_ctl           = svm->nested.ctl.tlb_ctl;
 813         vmcb12->control.event_inj         = svm->nested.ctl.event_inj;
 814         vmcb12->control.event_inj_err     = svm->nested.ctl.event_inj_err;
 815
 816         vmcb12->control.pause_filter_count =
 817                 svm->vmcb->control.pause_filter_count;
 818         vmcb12->control.pause_filter_thresh =
 819                 svm->vmcb->control.pause_filter_thresh;
 820
 821         nested_svm_copy_common_state(svm->nested.vmcb02.ptr, svm->vmcb01.ptr);
 822
 823         svm_switch_vmcb(svm, &svm->vmcb01);
 824
 825         /*
 826          * On vmexit the  GIF is set to false and
 827          * no event can be injected in L1.
 828          */
 829         svm_set_gif(svm, false);
 830         svm->vmcb->control.exit_int_info = 0;
 831
 832         svm->vcpu.arch.tsc_offset = svm->vcpu.arch.l1_tsc_offset;
 833         if (svm->vmcb->control.tsc_offset != svm->vcpu.arch.tsc_offset) {
 834                 svm->vmcb->control.tsc_offset = svm->vcpu.arch.tsc_offset;
 835                 vmcb_mark_dirty(svm->vmcb, VMCB_INTERCEPTS);
 836         }
 837
 838         svm->nested.ctl.nested_cr3 = 0;
 839
 840         /*
 841          * Restore processor state that had been saved in vmcb01
 842          */
 843         kvm_set_rflags(vcpu, svm->vmcb->save.rflags);
 844         svm_set_efer(vcpu, svm->vmcb->save.efer);
 845         svm_set_cr0(vcpu, svm->vmcb->save.cr0 | X86_CR0_PE);
 846         svm_set_cr4(vcpu, svm->vmcb->save.cr4);
 847         kvm_rax_write(vcpu, svm->vmcb->save.rax);
 848         kvm_rsp_write(vcpu, svm->vmcb->save.rsp);
 849         kvm_rip_write(vcpu, svm->vmcb->save.rip);
 850
 851         svm->vcpu.arch.dr7 = DR7_FIXED_1;
 852         kvm_update_dr7(&svm->vcpu);
 853
 854         trace_kvm_nested_vmexit_inject(vmcb12->control.exit_code,
 855                                        vmcb12->control.exit_info_1,
 856                                        vmcb12->control.exit_info_2,
 857                                        vmcb12->control.exit_int_info,
 858                                        vmcb12->control.exit_int_info_err,
 859                                        KVM_ISA_SVM);
 860
 861         kvm_vcpu_unmap(vcpu, &map, true);
 862
 863         nested_svm_transition_tlb_flush(vcpu);
 864
 865         nested_svm_uninit_mmu_context(vcpu);
 866
 867         rc = nested_svm_load_cr3(vcpu, svm->vmcb->save.cr3, false, true);
 868         if (rc)
 869                 return 1;
 870
 871         /*
 872          * Drop what we picked up for L2 via svm_complete_interrupts() so it
 873          * doesn't end up in L1.
 874          */
 875         svm->vcpu.arch.nmi_injected = false;
 876         kvm_clear_exception_queue(vcpu);
 877         kvm_clear_interrupt_queue(vcpu);
 878
 879         /*
 880          * If we are here following the completion of a VMRUN that
 881          * is being single-stepped, queue the pending #DB intercept
 882          * right now so that it an be accounted for before we execute
 883          * L1's next instruction.
 884          */
 885         if (unlikely(svm->vmcb->save.rflags & X86_EFLAGS_TF))
 886                 kvm_queue_exception(&(svm->vcpu), DB_VECTOR);
 887
 888         return 0;
 889 }
 890
 891 static void nested_svm_triple_fault(struct kvm_vcpu *vcpu)
 892 {
 893         nested_svm_simple_vmexit(to_svm(vcpu), SVM_EXIT_SHUTDOWN);
 894 }
 895
 896 int svm_allocate_nested(struct vcpu_svm *svm)
 897 {
 898         struct page *vmcb02_page;
 899
 900         if (svm->nested.initialized)
 901                 return 0;
 902
 903         vmcb02_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 904         if (!vmcb02_page)
 905                 return -ENOMEM;
 906         svm->nested.vmcb02.ptr = page_address(vmcb02_page);
 907         svm->nested.vmcb02.pa = __sme_set(page_to_pfn(vmcb02_page) << PAGE_SHIFT);
 908
 909         svm->nested.msrpm = svm_vcpu_alloc_msrpm();
 910         if (!svm->nested.msrpm)
 911                 goto err_free_vmcb02;
 912         svm_vcpu_init_msrpm(&svm->vcpu, svm->nested.msrpm);
 913
 914         svm->nested.initialized = true;
 915         return 0;
 916
 917 err_free_vmcb02:
 918         __free_page(vmcb02_page);
 919         return -ENOMEM;
 920 }
 921
 922 void svm_free_nested(struct vcpu_svm *svm)
 923 {
 924         if (!svm->nested.initialized)
 925                 return;
 926
 927         svm_vcpu_free_msrpm(svm->nested.msrpm);
 928         svm->nested.msrpm = NULL;
 929
 930         __free_page(virt_to_page(svm->nested.vmcb02.ptr));
 931         svm->nested.vmcb02.ptr = NULL;
 932
 933         /*
 934          * When last_vmcb12_gpa matches the current vmcb12 gpa,
 935          * some vmcb12 fields are not loaded if they are marked clean
 936          * in the vmcb12, since in this case they are up to date already.
 937          *
 938          * When the vmcb02 is freed, this optimization becomes invalid.
 939          */
 940         svm->nested.last_vmcb12_gpa = INVALID_GPA;
 941
 942         svm->nested.initialized = false;
 943 }
 944
 945 /*
 946  * Forcibly leave nested mode in order to be able to reset the VCPU later on.
 947  */
 948 void svm_leave_nested(struct vcpu_svm *svm)
 949 {
 950         struct kvm_vcpu *vcpu = &svm->vcpu;
 951
 952         if (is_guest_mode(vcpu)) {
 953                 svm->nested.nested_run_pending = 0;
 954                 svm->nested.vmcb12_gpa = INVALID_GPA;
 955
 956                 leave_guest_mode(vcpu);
 957
 958                 svm_switch_vmcb(svm, &svm->vmcb01);
 959
 960                 nested_svm_uninit_mmu_context(vcpu);
 961                 vmcb_mark_all_dirty(svm->vmcb);
 962         }
 963
 964         kvm_clear_request(KVM_REQ_GET_NESTED_STATE_PAGES, vcpu);
 965 }
 966
 967 static int nested_svm_exit_handled_msr(struct vcpu_svm *svm)
 968 {
 969         u32 offset, msr, value;
 970         int write, mask;
 971
 972         if (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_MSR_PROT)))
 973                 return NESTED_EXIT_HOST;
 974
 975         msr    = svm->vcpu.arch.regs[VCPU_REGS_RCX];
 976         offset = svm_msrpm_offset(msr);
 977         write  = svm->vmcb->control.exit_info_1 & 1;
 978         mask   = 1 << ((2 * (msr & 0xf)) + write);
 979
 980         if (offset == MSR_INVALID)
 981                 return NESTED_EXIT_DONE;
 982
 983         /* Offset is in 32 bit units but need in 8 bit units */
 984         offset *= 4;
 985
 986         if (kvm_vcpu_read_guest(&svm->vcpu, svm->nested.ctl.msrpm_base_pa + offset, &value, 4))
 987                 return NESTED_EXIT_DONE;
 988
 989         return (value & mask) ? NESTED_EXIT_DONE : NESTED_EXIT_HOST;
 990 }
 991
 992 static int nested_svm_intercept_ioio(struct vcpu_svm *svm)
 993 {
 994         unsigned port, size, iopm_len;
 995         u16 val, mask;
 996         u8 start_bit;
 997         u64 gpa;
 998
 999         if (!(vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_IOIO_PROT)))
1000                 return NESTED_EXIT_HOST;
1001
1002         port = svm->vmcb->control.exit_info_1 >> 16;
1003         size = (svm->vmcb->control.exit_info_1 & SVM_IOIO_SIZE_MASK) >>
1004                 SVM_IOIO_SIZE_SHIFT;
1005         gpa  = svm->nested.ctl.iopm_base_pa + (port / 8);
1006         start_bit = port % 8;
1007         iopm_len = (start_bit + size > 8) ? 2 : 1;
1008         mask = (0xf >> (4 - size)) << start_bit;
1009         val = 0;
1010
1011         if (kvm_vcpu_read_guest(&svm->vcpu, gpa, &val, iopm_len))
1012                 return NESTED_EXIT_DONE;
1013
1014         return (val & mask) ? NESTED_EXIT_DONE : NESTED_EXIT_HOST;
1015 }
1016
1017 static int nested_svm_intercept(struct vcpu_svm *svm)
1018 {
1019         u32 exit_code = svm->vmcb->control.exit_code;
1020         int vmexit = NESTED_EXIT_HOST;
1021
1022         switch (exit_code) {
1023         case SVM_EXIT_MSR:
1024                 vmexit = nested_svm_exit_handled_msr(svm);
1025                 break;
1026         case SVM_EXIT_IOIO:
1027                 vmexit = nested_svm_intercept_ioio(svm);
1028                 break;
1029         case SVM_EXIT_READ_CR0 ... SVM_EXIT_WRITE_CR8: {
1030                 if (vmcb_is_intercept(&svm->nested.ctl, exit_code))
1031                         vmexit = NESTED_EXIT_DONE;
1032                 break;
1033         }
1034         case SVM_EXIT_READ_DR0 ... SVM_EXIT_WRITE_DR7: {
1035                 if (vmcb_is_intercept(&svm->nested.ctl, exit_code))
1036                         vmexit = NESTED_EXIT_DONE;
1037                 break;
1038         }
1039         case SVM_EXIT_EXCP_BASE ... SVM_EXIT_EXCP_BASE + 0x1f: {
1040                 /*
1041                  * Host-intercepted exceptions have been checked already in
1042                  * nested_svm_exit_special.  There is nothing to do here,
1043                  * the vmexit is injected by svm_check_nested_events.
1044                  */
1045                 vmexit = NESTED_EXIT_DONE;
1046                 break;
1047         }
1048         case SVM_EXIT_ERR: {
1049                 vmexit = NESTED_EXIT_DONE;
1050                 break;
1051         }
1052         default: {
1053                 if (vmcb_is_intercept(&svm->nested.ctl, exit_code))
1054                         vmexit = NESTED_EXIT_DONE;
1055         }
1056         }
1057
1058         return vmexit;
1059 }
1060
1061 int nested_svm_exit_handled(struct vcpu_svm *svm)
1062 {
1063         int vmexit;
1064
1065         vmexit = nested_svm_intercept(svm);
1066
1067         if (vmexit == NESTED_EXIT_DONE)
1068                 nested_svm_vmexit(svm);
1069
1070         return vmexit;
1071 }
1072
1073 int nested_svm_check_permissions(struct kvm_vcpu *vcpu)
1074 {
1075         if (!(vcpu->arch.efer & EFER_SVME) || !is_paging(vcpu)) {
1076                 kvm_queue_exception(vcpu, UD_VECTOR);
1077                 return 1;
1078         }
1079
1080         if (to_svm(vcpu)->vmcb->save.cpl) {
1081                 kvm_inject_gp(vcpu, 0);
1082                 return 1;
1083         }
1084
1085         return 0;
1086 }
1087
1088 static bool nested_exit_on_exception(struct vcpu_svm *svm)
1089 {
1090         unsigned int nr = svm->vcpu.arch.exception.nr;
1091
1092         return (svm->nested.ctl.intercepts[INTERCEPT_EXCEPTION] & BIT(nr));
1093 }
1094
1095 static void nested_svm_inject_exception_vmexit(struct vcpu_svm *svm)
1096 {
1097         unsigned int nr = svm->vcpu.arch.exception.nr;
1098
1099         svm->vmcb->control.exit_code = SVM_EXIT_EXCP_BASE + nr;
1100         svm->vmcb->control.exit_code_hi = 0;
1101
1102         if (svm->vcpu.arch.exception.has_error_code)
1103                 svm->vmcb->control.exit_info_1 = svm->vcpu.arch.exception.error_code;
1104
1105         /*
1106          * EXITINFO2 is undefined for all exception intercepts other
1107          * than #PF.
1108          */
1109         if (nr == PF_VECTOR) {
1110                 if (svm->vcpu.arch.exception.nested_apf)
1111                         svm->vmcb->control.exit_info_2 = svm->vcpu.arch.apf.nested_apf_token;
1112                 else if (svm->vcpu.arch.exception.has_payload)
1113                         svm->vmcb->control.exit_info_2 = svm->vcpu.arch.exception.payload;
1114                 else
1115                         svm->vmcb->control.exit_info_2 = svm->vcpu.arch.cr2;
1116         } else if (nr == DB_VECTOR) {
1117                 /* See inject_pending_event.  */
1118                 kvm_deliver_exception_payload(&svm->vcpu);
1119                 if (svm->vcpu.arch.dr7 & DR7_GD) {
1120                         svm->vcpu.arch.dr7 &= ~DR7_GD;
1121                         kvm_update_dr7(&svm->vcpu);
1122                 }
1123         } else
1124                 WARN_ON(svm->vcpu.arch.exception.has_payload);
1125
1126         nested_svm_vmexit(svm);
1127 }
1128
1129 static inline bool nested_exit_on_init(struct vcpu_svm *svm)
1130 {
1131         return vmcb_is_intercept(&svm->nested.ctl, INTERCEPT_INIT);
1132 }
1133
1134 static int svm_check_nested_events(struct kvm_vcpu *vcpu)
1135 {
1136         struct vcpu_svm *svm = to_svm(vcpu);
1137         bool block_nested_events =
1138                 kvm_event_needs_reinjection(vcpu) || svm->nested.nested_run_pending;
1139         struct kvm_lapic *apic = vcpu->arch.apic;
1140
1141         if (lapic_in_kernel(vcpu) &&
1142             test_bit(KVM_APIC_INIT, &apic->pending_events)) {
1143                 if (block_nested_events)
1144                         return -EBUSY;
1145                 if (!nested_exit_on_init(svm))
1146                         return 0;
1147                 nested_svm_simple_vmexit(svm, SVM_EXIT_INIT);
1148                 return 0;
1149         }
1150
1151         if (vcpu->arch.exception.pending) {
1152                 /*
1153                  * Only a pending nested run can block a pending exception.
1154                  * Otherwise an injected NMI/interrupt should either be
1155                  * lost or delivered to the nested hypervisor in the EXITINTINFO
1156                  * vmcb field, while delivering the pending exception.
1157                  */
1158                 if (svm->nested.nested_run_pending)
1159                         return -EBUSY;
1160                 if (!nested_exit_on_exception(svm))
1161                         return 0;
1162                 nested_svm_inject_exception_vmexit(svm);
1163                 return 0;
1164         }
1165
1166         if (vcpu->arch.smi_pending && !svm_smi_blocked(vcpu)) {
1167                 if (block_nested_events)
1168                         return -EBUSY;
1169                 if (!nested_exit_on_smi(svm))
1170                         return 0;
1171                 nested_svm_simple_vmexit(svm, SVM_EXIT_SMI);
1172                 return 0;
1173         }
1174
1175         if (vcpu->arch.nmi_pending && !svm_nmi_blocked(vcpu)) {
1176                 if (block_nested_events)
1177                         return -EBUSY;
1178                 if (!nested_exit_on_nmi(svm))
1179                         return 0;
1180                 nested_svm_simple_vmexit(svm, SVM_EXIT_NMI);
1181                 return 0;
1182         }
1183
1184         if (kvm_cpu_has_interrupt(vcpu) && !svm_interrupt_blocked(vcpu)) {
1185                 if (block_nested_events)
1186                         return -EBUSY;
1187                 if (!nested_exit_on_intr(svm))
1188                         return 0;
1189                 trace_kvm_nested_intr_vmexit(svm->vmcb->save.rip);
1190                 nested_svm_simple_vmexit(svm, SVM_EXIT_INTR);
1191                 return 0;
1192         }
1193
1194         return 0;
1195 }
1196
1197 int nested_svm_exit_special(struct vcpu_svm *svm)
1198 {
1199         u32 exit_code = svm->vmcb->control.exit_code;
1200
1201         switch (exit_code) {
1202         case SVM_EXIT_INTR:
1203         case SVM_EXIT_NMI:
1204         case SVM_EXIT_NPF:
1205                 return NESTED_EXIT_HOST;
1206         case SVM_EXIT_EXCP_BASE ... SVM_EXIT_EXCP_BASE + 0x1f: {
1207                 u32 excp_bits = 1 << (exit_code - SVM_EXIT_EXCP_BASE);
1208
1209                 if (svm->vmcb01.ptr->control.intercepts[INTERCEPT_EXCEPTION] &
1210                     excp_bits)
1211                         return NESTED_EXIT_HOST;
1212                 else if (exit_code == SVM_EXIT_EXCP_BASE + PF_VECTOR &&
1213                          svm->vcpu.arch.apf.host_apf_flags)
1214                         /* Trap async PF even if not shadowing */
1215                         return NESTED_EXIT_HOST;
1216                 break;
1217         }
1218         default:
1219                 break;
1220         }
1221
1222         return NESTED_EXIT_CONTINUE;
1223 }
1224
1225 static int svm_get_nested_state(struct kvm_vcpu *vcpu,
1226                                 struct kvm_nested_state __user *user_kvm_nested_state,
1227                                 u32 user_data_size)
1228 {
1229         struct vcpu_svm *svm;
1230         struct kvm_nested_state kvm_state = {
1231                 .flags = 0,
1232                 .format = KVM_STATE_NESTED_FORMAT_SVM,
1233                 .size = sizeof(kvm_state),
1234         };
1235         struct vmcb __user *user_vmcb = (struct vmcb __user *)
1236                 &user_kvm_nested_state->data.svm[0];
1237
1238         if (!vcpu)
1239                 return kvm_state.size + KVM_STATE_NESTED_SVM_VMCB_SIZE;
1240
1241         svm = to_svm(vcpu);
1242
1243         if (user_data_size < kvm_state.size)
1244                 goto out;
1245
1246         /* First fill in the header and copy it out.  */
1247         if (is_guest_mode(vcpu)) {
1248                 kvm_state.hdr.svm.vmcb_pa = svm->nested.vmcb12_gpa;
1249                 kvm_state.size += KVM_STATE_NESTED_SVM_VMCB_SIZE;
1250                 kvm_state.flags |= KVM_STATE_NESTED_GUEST_MODE;
1251
1252                 if (svm->nested.nested_run_pending)
1253                         kvm_state.flags |= KVM_STATE_NESTED_RUN_PENDING;
1254         }
1255
1256         if (gif_set(svm))
1257                 kvm_state.flags |= KVM_STATE_NESTED_GIF_SET;
1258
1259         if (copy_to_user(user_kvm_nested_state, &kvm_state, sizeof(kvm_state)))
1260                 return -EFAULT;
1261
1262         if (!is_guest_mode(vcpu))
1263                 goto out;
1264
1265         /*
1266          * Copy over the full size of the VMCB rather than just the size
1267          * of the structs.
1268          */
1269         if (clear_user(user_vmcb, KVM_STATE_NESTED_SVM_VMCB_SIZE))
1270                 return -EFAULT;
1271         if (copy_to_user(&user_vmcb->control, &svm->nested.ctl,
1272                          sizeof(user_vmcb->control)))
1273                 return -EFAULT;
1274         if (copy_to_user(&user_vmcb->save, &svm->vmcb01.ptr->save,
1275                          sizeof(user_vmcb->save)))
1276                 return -EFAULT;
1277 out:
1278         return kvm_state.size;
1279 }
1280
1281 static int svm_set_nested_state(struct kvm_vcpu *vcpu,
1282                                 struct kvm_nested_state __user *user_kvm_nested_state,
1283                                 struct kvm_nested_state *kvm_state)
1284 {
1285         struct vcpu_svm *svm = to_svm(vcpu);
1286         struct vmcb __user *user_vmcb = (struct vmcb __user *)
1287                 &user_kvm_nested_state->data.svm[0];
1288         struct vmcb_control_area *ctl;
1289         struct vmcb_save_area *save;
1290         unsigned long cr0;
1291         int ret;
1292
1293         BUILD_BUG_ON(sizeof(struct vmcb_control_area) + sizeof(struct vmcb_save_area) >
1294                      KVM_STATE_NESTED_SVM_VMCB_SIZE);
1295
1296         if (kvm_state->format != KVM_STATE_NESTED_FORMAT_SVM)
1297                 return -EINVAL;
1298
1299         if (kvm_state->flags & ~(KVM_STATE_NESTED_GUEST_MODE |
1300                                  KVM_STATE_NESTED_RUN_PENDING |
1301                                  KVM_STATE_NESTED_GIF_SET))
1302                 return -EINVAL;
1303
1304         /*
1305          * If in guest mode, vcpu->arch.efer actually refers to the L2 guest's
1306          * EFER.SVME, but EFER.SVME still has to be 1 for VMRUN to succeed.
1307          */
1308         if (!(vcpu->arch.efer & EFER_SVME)) {
1309                 /* GIF=1 and no guest mode are required if SVME=0.  */
1310                 if (kvm_state->flags != KVM_STATE_NESTED_GIF_SET)
1311                         return -EINVAL;
1312         }
1313
1314         /* SMM temporarily disables SVM, so we cannot be in guest mode.  */
1315         if (is_smm(vcpu) && (kvm_state->flags & KVM_STATE_NESTED_GUEST_MODE))
1316                 return -EINVAL;
1317
1318         if (!(kvm_state->flags & KVM_STATE_NESTED_GUEST_MODE)) {
1319                 svm_leave_nested(svm);
1320                 svm_set_gif(svm, !!(kvm_state->flags & KVM_STATE_NESTED_GIF_SET));
1321                 return 0;
1322         }
1323
1324         if (!page_address_valid(vcpu, kvm_state->hdr.svm.vmcb_pa))
1325                 return -EINVAL;
1326         if (kvm_state->size < sizeof(*kvm_state) + KVM_STATE_NESTED_SVM_VMCB_SIZE)
1327                 return -EINVAL;
1328
1329         ret  = -ENOMEM;
1330         ctl  = kzalloc(sizeof(*ctl),  GFP_KERNEL_ACCOUNT);
1331         save = kzalloc(sizeof(*save), GFP_KERNEL_ACCOUNT);
1332         if (!ctl || !save)
1333                 goto out_free;
1334
1335         ret = -EFAULT;
1336         if (copy_from_user(ctl, &user_vmcb->control, sizeof(*ctl)))
1337                 goto out_free;
1338         if (copy_from_user(save, &user_vmcb->save, sizeof(*save)))
1339                 goto out_free;
1340
1341         ret = -EINVAL;
1342         if (!nested_vmcb_check_controls(vcpu, ctl))
1343                 goto out_free;
1344
1345         /*
1346          * Processor state contains L2 state.  Check that it is
1347          * valid for guest mode (see nested_vmcb_check_save).
1348          */
1349         cr0 = kvm_read_cr0(vcpu);
1350         if (((cr0 & X86_CR0_CD) == 0) && (cr0 & X86_CR0_NW))
1351                 goto out_free;
1352
1353         /*
1354          * Validate host state saved from before VMRUN (see
1355          * nested_svm_check_permissions).
1356          */
1357         if (!(save->cr0 & X86_CR0_PG) ||
1358             !(save->cr0 & X86_CR0_PE) ||
1359             (save->rflags & X86_EFLAGS_VM) ||
1360             !nested_vmcb_valid_sregs(vcpu, save))
1361                 goto out_free;
1362
1363         /*
1364          * While the nested guest CR3 is already checked and set by
1365          * KVM_SET_SREGS, it was set when nested state was yet loaded,
1366          * thus MMU might not be initialized correctly.
1367          * Set it again to fix this.
1368          */
1369
1370         ret = nested_svm_load_cr3(&svm->vcpu, vcpu->arch.cr3,
1371                                   nested_npt_enabled(svm), false);
1372         if (WARN_ON_ONCE(ret))
1373                 goto out_free;
1374
1375
1376         /*
1377          * All checks done, we can enter guest mode. Userspace provides
1378          * vmcb12.control, which will be combined with L1 and stored into
1379          * vmcb02, and the L1 save state which we store in vmcb01.
1380          * L2 registers if needed are moved from the current VMCB to VMCB02.
1381          */
1382
1383         if (is_guest_mode(vcpu))
1384                 svm_leave_nested(svm);
1385         else
1386                 svm->nested.vmcb02.ptr->save = svm->vmcb01.ptr->save;
1387
1388         svm_set_gif(svm, !!(kvm_state->flags & KVM_STATE_NESTED_GIF_SET));
1389
1390         svm->nested.nested_run_pending =
1391                 !!(kvm_state->flags & KVM_STATE_NESTED_RUN_PENDING);
1392
1393         svm->nested.vmcb12_gpa = kvm_state->hdr.svm.vmcb_pa;
1394
1395         svm_copy_vmrun_state(&svm->vmcb01.ptr->save, save);
1396         nested_load_control_from_vmcb12(svm, ctl);
1397
1398         svm_switch_vmcb(svm, &svm->nested.vmcb02);
1399         nested_vmcb02_prepare_control(svm);
1400         kvm_make_request(KVM_REQ_GET_NESTED_STATE_PAGES, vcpu);
1401         ret = 0;
1402 out_free:
1403         kfree(save);
1404         kfree(ctl);
1405
1406         return ret;
1407 }
1408
1409 static bool svm_get_nested_state_pages(struct kvm_vcpu *vcpu)
1410 {
1411         struct vcpu_svm *svm = to_svm(vcpu);
1412
1413         if (WARN_ON(!is_guest_mode(vcpu)))
1414                 return true;
1415
1416         if (!vcpu->arch.pdptrs_from_userspace &&
1417             !nested_npt_enabled(svm) && is_pae_paging(vcpu))
1418                 /*
1419                  * Reload the guest's PDPTRs since after a migration
1420                  * the guest CR3 might be restored prior to setting the nested
1421                  * state which can lead to a load of wrong PDPTRs.
1422                  */
1423                 if (CC(!load_pdptrs(vcpu, vcpu->arch.walk_mmu, vcpu->arch.cr3)))
1424                         return false;
1425
1426         if (!nested_svm_vmrun_msrpm(svm)) {
1427                 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
1428                 vcpu->run->internal.suberror =
1429                         KVM_INTERNAL_ERROR_EMULATION;
1430                 vcpu->run->internal.ndata = 0;
1431                 return false;
1432         }
1433
1434         return true;
1435 }
1436
1437 struct kvm_x86_nested_ops svm_nested_ops = {
1438         .check_events = svm_check_nested_events,
1439         .triple_fault = nested_svm_triple_fault,
1440         .get_nested_state_pages = svm_get_nested_state_pages,
1441         .get_state = svm_get_nested_state,
1442         .set_state = svm_set_nested_state,
1443 };