arch/arm64/kvm/hyp/nvhe/switch.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2015 - ARM Ltd
   4  * Author: Marc Zyngier <marc.zyngier@arm.com>
   5  */
   6
   7 #include <hyp/switch.h>
   8 #include <hyp/sysreg-sr.h>
   9
  10 #include <linux/arm-smccc.h>
  11 #include <linux/kvm_host.h>
  12 #include <linux/types.h>
  13 #include <linux/jump_label.h>
  14 #include <uapi/linux/psci.h>
  15
  16 #include <kvm/arm_psci.h>
  17
  18 #include <asm/barrier.h>
  19 #include <asm/cpufeature.h>
  20 #include <asm/kprobes.h>
  21 #include <asm/kvm_asm.h>
  22 #include <asm/kvm_emulate.h>
  23 #include <asm/kvm_hyp.h>
  24 #include <asm/kvm_mmu.h>
  25 #include <asm/fpsimd.h>
  26 #include <asm/debug-monitors.h>
  27 #include <asm/processor.h>
  28
  29 #include <nvhe/fixed_config.h>
  30 #include <nvhe/mem_protect.h>
  31
  32 /* Non-VHE specific context */
  33 DEFINE_PER_CPU(struct kvm_host_data, kvm_host_data);
  34 DEFINE_PER_CPU(struct kvm_cpu_context, kvm_hyp_ctxt);
  35 DEFINE_PER_CPU(unsigned long, kvm_hyp_vector);
  36
  37 extern void kvm_nvhe_prepare_backtrace(unsigned long fp, unsigned long pc);
  38
  39 static void __activate_traps(struct kvm_vcpu *vcpu)
  40 {
  41         u64 val;
  42
  43         ___activate_traps(vcpu);
  44         __activate_traps_common(vcpu);
  45
  46         val = vcpu->arch.cptr_el2;
  47         val |= CPTR_EL2_TAM;    /* Same bit irrespective of E2H */
  48         val |= has_hvhe() ? CPACR_EL1_TTA : CPTR_EL2_TTA;
  49         if (cpus_have_final_cap(ARM64_SME)) {
  50                 if (has_hvhe())
  51                         val &= ~(CPACR_EL1_SMEN_EL1EN | CPACR_EL1_SMEN_EL0EN);
  52                 else
  53                         val |= CPTR_EL2_TSM;
  54         }
  55
  56         if (!guest_owns_fp_regs(vcpu)) {
  57                 if (has_hvhe())
  58                         val &= ~(CPACR_EL1_FPEN_EL0EN | CPACR_EL1_FPEN_EL1EN |
  59                                  CPACR_EL1_ZEN_EL0EN | CPACR_EL1_ZEN_EL1EN);
  60                 else
  61                         val |= CPTR_EL2_TFP | CPTR_EL2_TZ;
  62
  63                 __activate_traps_fpsimd32(vcpu);
  64         }
  65
  66         kvm_write_cptr_el2(val);
  67         write_sysreg(__this_cpu_read(kvm_hyp_vector), vbar_el2);
  68
  69         if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
  70                 struct kvm_cpu_context *ctxt = &vcpu->arch.ctxt;
  71
  72                 isb();
  73                 /*
  74                  * At this stage, and thanks to the above isb(), S2 is
  75                  * configured and enabled. We can now restore the guest's S1
  76                  * configuration: SCTLR, and only then TCR.
  77                  */
  78                 write_sysreg_el1(ctxt_sys_reg(ctxt, SCTLR_EL1), SYS_SCTLR);
  79                 isb();
  80                 write_sysreg_el1(ctxt_sys_reg(ctxt, TCR_EL1),   SYS_TCR);
  81         }
  82 }
  83
  84 static void __deactivate_traps(struct kvm_vcpu *vcpu)
  85 {
  86         extern char __kvm_hyp_host_vector[];
  87
  88         ___deactivate_traps(vcpu);
  89
  90         if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
  91                 u64 val;
  92
  93                 /*
  94                  * Set the TCR and SCTLR registers in the exact opposite
  95                  * sequence as __activate_traps (first prevent walks,
  96                  * then force the MMU on). A generous sprinkling of isb()
  97                  * ensure that things happen in this exact order.
  98                  */
  99                 val = read_sysreg_el1(SYS_TCR);
 100                 write_sysreg_el1(val | TCR_EPD1_MASK | TCR_EPD0_MASK, SYS_TCR);
 101                 isb();
 102                 val = read_sysreg_el1(SYS_SCTLR);
 103                 write_sysreg_el1(val | SCTLR_ELx_M, SYS_SCTLR);
 104                 isb();
 105         }
 106
 107         __deactivate_traps_common(vcpu);
 108
 109         write_sysreg(this_cpu_ptr(&kvm_init_params)->hcr_el2, hcr_el2);
 110
 111         kvm_reset_cptr_el2(vcpu);
 112         write_sysreg(__kvm_hyp_host_vector, vbar_el2);
 113 }
 114
 115 /* Save VGICv3 state on non-VHE systems */
 116 static void __hyp_vgic_save_state(struct kvm_vcpu *vcpu)
 117 {
 118         if (static_branch_unlikely(&kvm_vgic_global_state.gicv3_cpuif)) {
 119                 __vgic_v3_save_state(&vcpu->arch.vgic_cpu.vgic_v3);
 120                 __vgic_v3_deactivate_traps(&vcpu->arch.vgic_cpu.vgic_v3);
 121         }
 122 }
 123
 124 /* Restore VGICv3 state on non-VHE systems */
 125 static void __hyp_vgic_restore_state(struct kvm_vcpu *vcpu)
 126 {
 127         if (static_branch_unlikely(&kvm_vgic_global_state.gicv3_cpuif)) {
 128                 __vgic_v3_activate_traps(&vcpu->arch.vgic_cpu.vgic_v3);
 129                 __vgic_v3_restore_state(&vcpu->arch.vgic_cpu.vgic_v3);
 130         }
 131 }
 132
 133 /*
 134  * Disable host events, enable guest events
 135  */
 136 #ifdef CONFIG_HW_PERF_EVENTS
 137 static bool __pmu_switch_to_guest(struct kvm_vcpu *vcpu)
 138 {
 139         struct kvm_pmu_events *pmu = &vcpu->arch.pmu.events;
 140
 141         if (pmu->events_host)
 142                 write_sysreg(pmu->events_host, pmcntenclr_el0);
 143
 144         if (pmu->events_guest)
 145                 write_sysreg(pmu->events_guest, pmcntenset_el0);
 146
 147         return (pmu->events_host || pmu->events_guest);
 148 }
 149
 150 /*
 151  * Disable guest events, enable host events
 152  */
 153 static void __pmu_switch_to_host(struct kvm_vcpu *vcpu)
 154 {
 155         struct kvm_pmu_events *pmu = &vcpu->arch.pmu.events;
 156
 157         if (pmu->events_guest)
 158                 write_sysreg(pmu->events_guest, pmcntenclr_el0);
 159
 160         if (pmu->events_host)
 161                 write_sysreg(pmu->events_host, pmcntenset_el0);
 162 }
 163 #else
 164 #define __pmu_switch_to_guest(v)        ({ false; })
 165 #define __pmu_switch_to_host(v)         do {} while (0)
 166 #endif
 167
 168 /*
 169  * Handler for protected VM MSR, MRS or System instruction execution in AArch64.
 170  *
 171  * Returns true if the hypervisor has handled the exit, and control should go
 172  * back to the guest, or false if it hasn't.
 173  */
 174 static bool kvm_handle_pvm_sys64(struct kvm_vcpu *vcpu, u64 *exit_code)
 175 {
 176         /*
 177          * Make sure we handle the exit for workarounds and ptrauth
 178          * before the pKVM handling, as the latter could decide to
 179          * UNDEF.
 180          */
 181         return (kvm_hyp_handle_sysreg(vcpu, exit_code) ||
 182                 kvm_handle_pvm_sysreg(vcpu, exit_code));
 183 }
 184
 185 static const exit_handler_fn hyp_exit_handlers[] = {
 186         [0 ... ESR_ELx_EC_MAX]          = NULL,
 187         [ESR_ELx_EC_CP15_32]            = kvm_hyp_handle_cp15_32,
 188         [ESR_ELx_EC_SYS64]              = kvm_hyp_handle_sysreg,
 189         [ESR_ELx_EC_SVE]                = kvm_hyp_handle_fpsimd,
 190         [ESR_ELx_EC_FP_ASIMD]           = kvm_hyp_handle_fpsimd,
 191         [ESR_ELx_EC_IABT_LOW]           = kvm_hyp_handle_iabt_low,
 192         [ESR_ELx_EC_DABT_LOW]           = kvm_hyp_handle_dabt_low,
 193         [ESR_ELx_EC_WATCHPT_LOW]        = kvm_hyp_handle_watchpt_low,
 194         [ESR_ELx_EC_PAC]                = kvm_hyp_handle_ptrauth,
 195         [ESR_ELx_EC_MOPS]               = kvm_hyp_handle_mops,
 196 };
 197
 198 static const exit_handler_fn pvm_exit_handlers[] = {
 199         [0 ... ESR_ELx_EC_MAX]          = NULL,
 200         [ESR_ELx_EC_SYS64]              = kvm_handle_pvm_sys64,
 201         [ESR_ELx_EC_SVE]                = kvm_handle_pvm_restricted,
 202         [ESR_ELx_EC_FP_ASIMD]           = kvm_hyp_handle_fpsimd,
 203         [ESR_ELx_EC_IABT_LOW]           = kvm_hyp_handle_iabt_low,
 204         [ESR_ELx_EC_DABT_LOW]           = kvm_hyp_handle_dabt_low,
 205         [ESR_ELx_EC_WATCHPT_LOW]        = kvm_hyp_handle_watchpt_low,
 206         [ESR_ELx_EC_PAC]                = kvm_hyp_handle_ptrauth,
 207         [ESR_ELx_EC_MOPS]               = kvm_hyp_handle_mops,
 208 };
 209
 210 static const exit_handler_fn *kvm_get_exit_handler_array(struct kvm_vcpu *vcpu)
 211 {
 212         if (unlikely(kvm_vm_is_protected(kern_hyp_va(vcpu->kvm))))
 213                 return pvm_exit_handlers;
 214
 215         return hyp_exit_handlers;
 216 }
 217
 218 /*
 219  * Some guests (e.g., protected VMs) are not be allowed to run in AArch32.
 220  * The ARMv8 architecture does not give the hypervisor a mechanism to prevent a
 221  * guest from dropping to AArch32 EL0 if implemented by the CPU. If the
 222  * hypervisor spots a guest in such a state ensure it is handled, and don't
 223  * trust the host to spot or fix it.  The check below is based on the one in
 224  * kvm_arch_vcpu_ioctl_run().
 225  *
 226  * Returns false if the guest ran in AArch32 when it shouldn't have, and
 227  * thus should exit to the host, or true if a the guest run loop can continue.
 228  */
 229 static void early_exit_filter(struct kvm_vcpu *vcpu, u64 *exit_code)
 230 {
 231         struct kvm *kvm = kern_hyp_va(vcpu->kvm);
 232
 233         if (kvm_vm_is_protected(kvm) && vcpu_mode_is_32bit(vcpu)) {
 234                 /*
 235                  * As we have caught the guest red-handed, decide that it isn't
 236                  * fit for purpose anymore by making the vcpu invalid. The VMM
 237                  * can try and fix it by re-initializing the vcpu with
 238                  * KVM_ARM_VCPU_INIT, however, this is likely not possible for
 239                  * protected VMs.
 240                  */
 241                 vcpu_clear_flag(vcpu, VCPU_INITIALIZED);
 242                 *exit_code &= BIT(ARM_EXIT_WITH_SERROR_BIT);
 243                 *exit_code |= ARM_EXCEPTION_IL;
 244         }
 245 }
 246
 247 /* Switch to the guest for legacy non-VHE systems */
 248 int __kvm_vcpu_run(struct kvm_vcpu *vcpu)
 249 {
 250         struct kvm_cpu_context *host_ctxt;
 251         struct kvm_cpu_context *guest_ctxt;
 252         struct kvm_s2_mmu *mmu;
 253         bool pmu_switch_needed;
 254         u64 exit_code;
 255
 256         /*
 257          * Having IRQs masked via PMR when entering the guest means the GIC
 258          * will not signal the CPU of interrupts of lower priority, and the
 259          * only way to get out will be via guest exceptions.
 260          * Naturally, we want to avoid this.
 261          */
 262         if (system_uses_irq_prio_masking()) {
 263                 gic_write_pmr(GIC_PRIO_IRQON | GIC_PRIO_PSR_I_SET);
 264                 pmr_sync();
 265         }
 266
 267         host_ctxt = &this_cpu_ptr(&kvm_host_data)->host_ctxt;
 268         host_ctxt->__hyp_running_vcpu = vcpu;
 269         guest_ctxt = &vcpu->arch.ctxt;
 270
 271         pmu_switch_needed = __pmu_switch_to_guest(vcpu);
 272
 273         __sysreg_save_state_nvhe(host_ctxt);
 274         /*
 275          * We must flush and disable the SPE buffer for nVHE, as
 276          * the translation regime(EL1&0) is going to be loaded with
 277          * that of the guest. And we must do this before we change the
 278          * translation regime to EL2 (via MDCR_EL2_E2PB == 0) and
 279          * before we load guest Stage1.
 280          */
 281         __debug_save_host_buffers_nvhe(vcpu);
 282
 283         /*
 284          * We're about to restore some new MMU state. Make sure
 285          * ongoing page-table walks that have started before we
 286          * trapped to EL2 have completed. This also synchronises the
 287          * above disabling of SPE and TRBE.
 288          *
 289          * See DDI0487I.a D8.1.5 "Out-of-context translation regimes",
 290          * rule R_LFHQG and subsequent information statements.
 291          */
 292         dsb(nsh);
 293
 294         __kvm_adjust_pc(vcpu);
 295
 296         /*
 297          * We must restore the 32-bit state before the sysregs, thanks
 298          * to erratum #852523 (Cortex-A57) or #853709 (Cortex-A72).
 299          *
 300          * Also, and in order to be able to deal with erratum #1319537 (A57)
 301          * and #1319367 (A72), we must ensure that all VM-related sysreg are
 302          * restored before we enable S2 translation.
 303          */
 304         __sysreg32_restore_state(vcpu);
 305         __sysreg_restore_state_nvhe(guest_ctxt);
 306
 307         mmu = kern_hyp_va(vcpu->arch.hw_mmu);
 308         __load_stage2(mmu, kern_hyp_va(mmu->arch));
 309         __activate_traps(vcpu);
 310
 311         __hyp_vgic_restore_state(vcpu);
 312         __timer_enable_traps(vcpu);
 313
 314         __debug_switch_to_guest(vcpu);
 315
 316         do {
 317                 /* Jump in the fire! */
 318                 exit_code = __guest_enter(vcpu);
 319
 320                 /* And we're baaack! */
 321         } while (fixup_guest_exit(vcpu, &exit_code));
 322
 323         __sysreg_save_state_nvhe(guest_ctxt);
 324         __sysreg32_save_state(vcpu);
 325         __timer_disable_traps(vcpu);
 326         __hyp_vgic_save_state(vcpu);
 327
 328         /*
 329          * Same thing as before the guest run: we're about to switch
 330          * the MMU context, so let's make sure we don't have any
 331          * ongoing EL1&0 translations.
 332          */
 333         dsb(nsh);
 334
 335         __deactivate_traps(vcpu);
 336         __load_host_stage2();
 337
 338         __sysreg_restore_state_nvhe(host_ctxt);
 339
 340         if (vcpu->arch.fp_state == FP_STATE_GUEST_OWNED)
 341                 __fpsimd_save_fpexc32(vcpu);
 342
 343         __debug_switch_to_host(vcpu);
 344         /*
 345          * This must come after restoring the host sysregs, since a non-VHE
 346          * system may enable SPE here and make use of the TTBRs.
 347          */
 348         __debug_restore_host_buffers_nvhe(vcpu);
 349
 350         if (pmu_switch_needed)
 351                 __pmu_switch_to_host(vcpu);
 352
 353         /* Returning to host will clear PSR.I, remask PMR if needed */
 354         if (system_uses_irq_prio_masking())
 355                 gic_write_pmr(GIC_PRIO_IRQOFF);
 356
 357         host_ctxt->__hyp_running_vcpu = NULL;
 358
 359         return exit_code;
 360 }
 361
 362 asmlinkage void __noreturn hyp_panic(void)
 363 {
 364         u64 spsr = read_sysreg_el2(SYS_SPSR);
 365         u64 elr = read_sysreg_el2(SYS_ELR);
 366         u64 par = read_sysreg_par();
 367         struct kvm_cpu_context *host_ctxt;
 368         struct kvm_vcpu *vcpu;
 369
 370         host_ctxt = &this_cpu_ptr(&kvm_host_data)->host_ctxt;
 371         vcpu = host_ctxt->__hyp_running_vcpu;
 372
 373         if (vcpu) {
 374                 __timer_disable_traps(vcpu);
 375                 __deactivate_traps(vcpu);
 376                 __load_host_stage2();
 377                 __sysreg_restore_state_nvhe(host_ctxt);
 378         }
 379
 380         /* Prepare to dump kvm nvhe hyp stacktrace */
 381         kvm_nvhe_prepare_backtrace((unsigned long)__builtin_frame_address(0),
 382                                    _THIS_IP_);
 383
 384         __hyp_do_panic(host_ctxt, spsr, elr, par);
 385         unreachable();
 386 }
 387
 388 asmlinkage void __noreturn hyp_panic_bad_stack(void)
 389 {
 390         hyp_panic();
 391 }
 392
 393 asmlinkage void kvm_unexpected_el2_exception(void)
 394 {
 395         __kvm_unexpected_el2_exception();
 396 }