arch/arm/kvm/arm.c

   1 /*
   2  * Copyright (C) 2012 - Virtual Open Systems and Columbia University
   3  * Author: Christoffer Dall <c.dall@virtualopensystems.com>
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License, version 2, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  17  */
  18
  19 #include <linux/cpu_pm.h>
  20 #include <linux/errno.h>
  21 #include <linux/err.h>
  22 #include <linux/kvm_host.h>
  23 #include <linux/module.h>
  24 #include <linux/vmalloc.h>
  25 #include <linux/fs.h>
  26 #include <linux/mman.h>
  27 #include <linux/sched.h>
  28 #include <linux/kvm.h>
  29 #include <trace/events/kvm.h>
  30
  31 #define CREATE_TRACE_POINTS
  32 #include "trace.h"
  33
  34 #include <asm/uaccess.h>
  35 #include <asm/ptrace.h>
  36 #include <asm/mman.h>
  37 #include <asm/tlbflush.h>
  38 #include <asm/cacheflush.h>
  39 #include <asm/virt.h>
  40 #include <asm/kvm_arm.h>
  41 #include <asm/kvm_asm.h>
  42 #include <asm/kvm_mmu.h>
  43 #include <asm/kvm_emulate.h>
  44 #include <asm/kvm_coproc.h>
  45 #include <asm/kvm_psci.h>
  46 #include <asm/sections.h>
  47
  48 #ifdef REQUIRES_VIRT
  49 __asm__(".arch_extension        virt");
  50 #endif
  51
  52 static DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page);
  53 static kvm_cpu_context_t __percpu *kvm_host_cpu_state;
  54 static unsigned long hyp_default_vectors;
  55
  56 /* Per-CPU variable containing the currently running vcpu. */
  57 static DEFINE_PER_CPU(struct kvm_vcpu *, kvm_arm_running_vcpu);
  58
  59 /* The VMID used in the VTTBR */
  60 static atomic64_t kvm_vmid_gen = ATOMIC64_INIT(1);
  61 static u32 kvm_next_vmid;
  62 static unsigned int kvm_vmid_bits __read_mostly;
  63 static DEFINE_SPINLOCK(kvm_vmid_lock);
  64
  65 static bool vgic_present;
  66
  67 static DEFINE_PER_CPU(unsigned char, kvm_arm_hardware_enabled);
  68
  69 static void kvm_arm_set_running_vcpu(struct kvm_vcpu *vcpu)
  70 {
  71         BUG_ON(preemptible());
  72         __this_cpu_write(kvm_arm_running_vcpu, vcpu);
  73 }
  74
  75 /**
  76  * kvm_arm_get_running_vcpu - get the vcpu running on the current CPU.
  77  * Must be called from non-preemptible context
  78  */
  79 struct kvm_vcpu *kvm_arm_get_running_vcpu(void)
  80 {
  81         BUG_ON(preemptible());
  82         return __this_cpu_read(kvm_arm_running_vcpu);
  83 }
  84
  85 /**
  86  * kvm_arm_get_running_vcpus - get the per-CPU array of currently running vcpus.
  87  */
  88 struct kvm_vcpu * __percpu *kvm_get_running_vcpus(void)
  89 {
  90         return &kvm_arm_running_vcpu;
  91 }
  92
  93 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
  94 {
  95         return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
  96 }
  97
  98 int kvm_arch_hardware_setup(void)
  99 {
 100         return 0;
 101 }
 102
 103 void kvm_arch_check_processor_compat(void *rtn)
 104 {
 105         *(int *)rtn = 0;
 106 }
 107
 108
 109 /**
 110  * kvm_arch_init_vm - initializes a VM data structure
 111  * @kvm:        pointer to the KVM struct
 112  */
 113 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 114 {
 115         int ret = 0;
 116
 117         if (type)
 118                 return -EINVAL;
 119
 120         ret = kvm_alloc_stage2_pgd(kvm);
 121         if (ret)
 122                 goto out_fail_alloc;
 123
 124         ret = create_hyp_mappings(kvm, kvm + 1);
 125         if (ret)
 126                 goto out_free_stage2_pgd;
 127
 128         kvm_vgic_early_init(kvm);
 129         kvm_timer_init(kvm);
 130
 131         /* Mark the initial VMID generation invalid */
 132         kvm->arch.vmid_gen = 0;
 133
 134         /* The maximum number of VCPUs is limited by the host's GIC model */
 135         kvm->arch.max_vcpus = vgic_present ?
 136                                 kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 137
 138         return ret;
 139 out_free_stage2_pgd:
 140         kvm_free_stage2_pgd(kvm);
 141 out_fail_alloc:
 142         return ret;
 143 }
 144
 145 int kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 146 {
 147         return VM_FAULT_SIGBUS;
 148 }
 149
 150
 151 /**
 152  * kvm_arch_destroy_vm - destroy the VM data structure
 153  * @kvm:        pointer to the KVM struct
 154  */
 155 void kvm_arch_destroy_vm(struct kvm *kvm)
 156 {
 157         int i;
 158
 159         for (i = 0; i < KVM_MAX_VCPUS; ++i) {
 160                 if (kvm->vcpus[i]) {
 161                         kvm_arch_vcpu_free(kvm->vcpus[i]);
 162                         kvm->vcpus[i] = NULL;
 163                 }
 164         }
 165
 166         kvm_vgic_destroy(kvm);
 167 }
 168
 169 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 170 {
 171         int r;
 172         switch (ext) {
 173         case KVM_CAP_IRQCHIP:
 174                 r = vgic_present;
 175                 break;
 176         case KVM_CAP_IOEVENTFD:
 177         case KVM_CAP_DEVICE_CTRL:
 178         case KVM_CAP_USER_MEMORY:
 179         case KVM_CAP_SYNC_MMU:
 180         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 181         case KVM_CAP_ONE_REG:
 182         case KVM_CAP_ARM_PSCI:
 183         case KVM_CAP_ARM_PSCI_0_2:
 184         case KVM_CAP_READONLY_MEM:
 185         case KVM_CAP_MP_STATE:
 186                 r = 1;
 187                 break;
 188         case KVM_CAP_COALESCED_MMIO:
 189                 r = KVM_COALESCED_MMIO_PAGE_OFFSET;
 190                 break;
 191         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 192                 r = 1;
 193                 break;
 194         case KVM_CAP_NR_VCPUS:
 195                 r = num_online_cpus();
 196                 break;
 197         case KVM_CAP_MAX_VCPUS:
 198                 r = KVM_MAX_VCPUS;
 199                 break;
 200         default:
 201                 r = kvm_arch_dev_ioctl_check_extension(ext);
 202                 break;
 203         }
 204         return r;
 205 }
 206
 207 long kvm_arch_dev_ioctl(struct file *filp,
 208                         unsigned int ioctl, unsigned long arg)
 209 {
 210         return -EINVAL;
 211 }
 212
 213
 214 struct kvm_vcpu *kvm_arch_vcpu_create(struct kvm *kvm, unsigned int id)
 215 {
 216         int err;
 217         struct kvm_vcpu *vcpu;
 218
 219         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm)) {
 220                 err = -EBUSY;
 221                 goto out;
 222         }
 223
 224         if (id >= kvm->arch.max_vcpus) {
 225                 err = -EINVAL;
 226                 goto out;
 227         }
 228
 229         vcpu = kmem_cache_zalloc(kvm_vcpu_cache, GFP_KERNEL);
 230         if (!vcpu) {
 231                 err = -ENOMEM;
 232                 goto out;
 233         }
 234
 235         err = kvm_vcpu_init(vcpu, kvm, id);
 236         if (err)
 237                 goto free_vcpu;
 238
 239         err = create_hyp_mappings(vcpu, vcpu + 1);
 240         if (err)
 241                 goto vcpu_uninit;
 242
 243         return vcpu;
 244 vcpu_uninit:
 245         kvm_vcpu_uninit(vcpu);
 246 free_vcpu:
 247         kmem_cache_free(kvm_vcpu_cache, vcpu);
 248 out:
 249         return ERR_PTR(err);
 250 }
 251
 252 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
 253 {
 254         kvm_vgic_vcpu_early_init(vcpu);
 255 }
 256
 257 void kvm_arch_vcpu_free(struct kvm_vcpu *vcpu)
 258 {
 259         kvm_mmu_free_memory_caches(vcpu);
 260         kvm_timer_vcpu_terminate(vcpu);
 261         kvm_vgic_vcpu_destroy(vcpu);
 262         kmem_cache_free(kvm_vcpu_cache, vcpu);
 263 }
 264
 265 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 266 {
 267         kvm_arch_vcpu_free(vcpu);
 268 }
 269
 270 int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu)
 271 {
 272         return kvm_timer_should_fire(vcpu);
 273 }
 274
 275 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 276 {
 277         kvm_timer_schedule(vcpu);
 278 }
 279
 280 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 281 {
 282         kvm_timer_unschedule(vcpu);
 283 }
 284
 285 int kvm_arch_vcpu_init(struct kvm_vcpu *vcpu)
 286 {
 287         /* Force users to call KVM_ARM_VCPU_INIT */
 288         vcpu->arch.target = -1;
 289         bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
 290
 291         /* Set up the timer */
 292         kvm_timer_vcpu_init(vcpu);
 293
 294         kvm_arm_reset_debug_ptr(vcpu);
 295
 296         return 0;
 297 }
 298
 299 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 300 {
 301         vcpu->cpu = cpu;
 302         vcpu->arch.host_cpu_context = this_cpu_ptr(kvm_host_cpu_state);
 303
 304         kvm_arm_set_running_vcpu(vcpu);
 305 }
 306
 307 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 308 {
 309         /*
 310          * The arch-generic KVM code expects the cpu field of a vcpu to be -1
 311          * if the vcpu is no longer assigned to a cpu.  This is used for the
 312          * optimized make_all_cpus_request path.
 313          */
 314         vcpu->cpu = -1;
 315
 316         kvm_arm_set_running_vcpu(NULL);
 317 }
 318
 319 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 320                                     struct kvm_mp_state *mp_state)
 321 {
 322         if (vcpu->arch.power_off)
 323                 mp_state->mp_state = KVM_MP_STATE_STOPPED;
 324         else
 325                 mp_state->mp_state = KVM_MP_STATE_RUNNABLE;
 326
 327         return 0;
 328 }
 329
 330 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
 331                                     struct kvm_mp_state *mp_state)
 332 {
 333         switch (mp_state->mp_state) {
 334         case KVM_MP_STATE_RUNNABLE:
 335                 vcpu->arch.power_off = false;
 336                 break;
 337         case KVM_MP_STATE_STOPPED:
 338                 vcpu->arch.power_off = true;
 339                 break;
 340         default:
 341                 return -EINVAL;
 342         }
 343
 344         return 0;
 345 }
 346
 347 /**
 348  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
 349  * @v:          The VCPU pointer
 350  *
 351  * If the guest CPU is not waiting for interrupts or an interrupt line is
 352  * asserted, the CPU is by definition runnable.
 353  */
 354 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
 355 {
 356         return ((!!v->arch.irq_lines || kvm_vgic_vcpu_pending_irq(v))
 357                 && !v->arch.power_off && !v->arch.pause);
 358 }
 359
 360 /* Just ensure a guest exit from a particular CPU */
 361 static void exit_vm_noop(void *info)
 362 {
 363 }
 364
 365 void force_vm_exit(const cpumask_t *mask)
 366 {
 367         smp_call_function_many(mask, exit_vm_noop, NULL, true);
 368 }
 369
 370 /**
 371  * need_new_vmid_gen - check that the VMID is still valid
 372  * @kvm: The VM's VMID to checkt
 373  *
 374  * return true if there is a new generation of VMIDs being used
 375  *
 376  * The hardware supports only 256 values with the value zero reserved for the
 377  * host, so we check if an assigned value belongs to a previous generation,
 378  * which which requires us to assign a new value. If we're the first to use a
 379  * VMID for the new generation, we must flush necessary caches and TLBs on all
 380  * CPUs.
 381  */
 382 static bool need_new_vmid_gen(struct kvm *kvm)
 383 {
 384         return unlikely(kvm->arch.vmid_gen != atomic64_read(&kvm_vmid_gen));
 385 }
 386
 387 /**
 388  * update_vttbr - Update the VTTBR with a valid VMID before the guest runs
 389  * @kvm The guest that we are about to run
 390  *
 391  * Called from kvm_arch_vcpu_ioctl_run before entering the guest to ensure the
 392  * VM has a valid VMID, otherwise assigns a new one and flushes corresponding
 393  * caches and TLBs.
 394  */
 395 static void update_vttbr(struct kvm *kvm)
 396 {
 397         phys_addr_t pgd_phys;
 398         u64 vmid;
 399
 400         if (!need_new_vmid_gen(kvm))
 401                 return;
 402
 403         spin_lock(&kvm_vmid_lock);
 404
 405         /*
 406          * We need to re-check the vmid_gen here to ensure that if another vcpu
 407          * already allocated a valid vmid for this vm, then this vcpu should
 408          * use the same vmid.
 409          */
 410         if (!need_new_vmid_gen(kvm)) {
 411                 spin_unlock(&kvm_vmid_lock);
 412                 return;
 413         }
 414
 415         /* First user of a new VMID generation? */
 416         if (unlikely(kvm_next_vmid == 0)) {
 417                 atomic64_inc(&kvm_vmid_gen);
 418                 kvm_next_vmid = 1;
 419
 420                 /*
 421                  * On SMP we know no other CPUs can use this CPU's or each
 422                  * other's VMID after force_vm_exit returns since the
 423                  * kvm_vmid_lock blocks them from reentry to the guest.
 424                  */
 425                 force_vm_exit(cpu_all_mask);
 426                 /*
 427                  * Now broadcast TLB + ICACHE invalidation over the inner
 428                  * shareable domain to make sure all data structures are
 429                  * clean.
 430                  */
 431                 kvm_call_hyp(__kvm_flush_vm_context);
 432         }
 433
 434         kvm->arch.vmid_gen = atomic64_read(&kvm_vmid_gen);
 435         kvm->arch.vmid = kvm_next_vmid;
 436         kvm_next_vmid++;
 437         kvm_next_vmid &= (1 << kvm_vmid_bits) - 1;
 438
 439         /* update vttbr to be used with the new vmid */
 440         pgd_phys = virt_to_phys(kvm_get_hwpgd(kvm));
 441         BUG_ON(pgd_phys & ~VTTBR_BADDR_MASK);
 442         vmid = ((u64)(kvm->arch.vmid) << VTTBR_VMID_SHIFT) & VTTBR_VMID_MASK(kvm_vmid_bits);
 443         kvm->arch.vttbr = pgd_phys | vmid;
 444
 445         spin_unlock(&kvm_vmid_lock);
 446 }
 447
 448 static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
 449 {
 450         struct kvm *kvm = vcpu->kvm;
 451         int ret;
 452
 453         if (likely(vcpu->arch.has_run_once))
 454                 return 0;
 455
 456         vcpu->arch.has_run_once = true;
 457
 458         /*
 459          * Map the VGIC hardware resources before running a vcpu the first
 460          * time on this VM.
 461          */
 462         if (unlikely(irqchip_in_kernel(kvm) && !vgic_ready(kvm))) {
 463                 ret = kvm_vgic_map_resources(kvm);
 464                 if (ret)
 465                         return ret;
 466         }
 467
 468         /*
 469          * Enable the arch timers only if we have an in-kernel VGIC
 470          * and it has been properly initialized, since we cannot handle
 471          * interrupts from the virtual timer with a userspace gic.
 472          */
 473         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 474                 kvm_timer_enable(kvm);
 475
 476         return 0;
 477 }
 478
 479 bool kvm_arch_intc_initialized(struct kvm *kvm)
 480 {
 481         return vgic_initialized(kvm);
 482 }
 483
 484 static void kvm_arm_halt_guest(struct kvm *kvm) __maybe_unused;
 485 static void kvm_arm_resume_guest(struct kvm *kvm) __maybe_unused;
 486
 487 static void kvm_arm_halt_guest(struct kvm *kvm)
 488 {
 489         int i;
 490         struct kvm_vcpu *vcpu;
 491
 492         kvm_for_each_vcpu(i, vcpu, kvm)
 493                 vcpu->arch.pause = true;
 494         force_vm_exit(cpu_all_mask);
 495 }
 496
 497 static void kvm_arm_resume_guest(struct kvm *kvm)
 498 {
 499         int i;
 500         struct kvm_vcpu *vcpu;
 501
 502         kvm_for_each_vcpu(i, vcpu, kvm) {
 503                 wait_queue_head_t *wq = kvm_arch_vcpu_wq(vcpu);
 504
 505                 vcpu->arch.pause = false;
 506                 wake_up_interruptible(wq);
 507         }
 508 }
 509
 510 static void vcpu_sleep(struct kvm_vcpu *vcpu)
 511 {
 512         wait_queue_head_t *wq = kvm_arch_vcpu_wq(vcpu);
 513
 514         wait_event_interruptible(*wq, ((!vcpu->arch.power_off) &&
 515                                        (!vcpu->arch.pause)));
 516 }
 517
 518 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu)
 519 {
 520         return vcpu->arch.target >= 0;
 521 }
 522
 523 /**
 524  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 525  * @vcpu:       The VCPU pointer
 526  * @run:        The kvm_run structure pointer used for userspace state exchange
 527  *
 528  * This function is called through the VCPU_RUN ioctl called from user space. It
 529  * will execute VM code in a loop until the time slice for the process is used
 530  * or some emulation is needed from user space in which case the function will
 531  * return with return value 0 and with the kvm_run structure filled in with the
 532  * required data for the requested emulation.
 533  */
 534 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run)
 535 {
 536         int ret;
 537         sigset_t sigsaved;
 538
 539         if (unlikely(!kvm_vcpu_initialized(vcpu)))
 540                 return -ENOEXEC;
 541
 542         ret = kvm_vcpu_first_run_init(vcpu);
 543         if (ret)
 544                 return ret;
 545
 546         if (run->exit_reason == KVM_EXIT_MMIO) {
 547                 ret = kvm_handle_mmio_return(vcpu, vcpu->run);
 548                 if (ret)
 549                         return ret;
 550         }
 551
 552         if (vcpu->sigset_active)
 553                 sigprocmask(SIG_SETMASK, &vcpu->sigset, &sigsaved);
 554
 555         ret = 1;
 556         run->exit_reason = KVM_EXIT_UNKNOWN;
 557         while (ret > 0) {
 558                 /*
 559                  * Check conditions before entering the guest
 560                  */
 561                 cond_resched();
 562
 563                 update_vttbr(vcpu->kvm);
 564
 565                 if (vcpu->arch.power_off || vcpu->arch.pause)
 566                         vcpu_sleep(vcpu);
 567
 568                 /*
 569                  * Preparing the interrupts to be injected also
 570                  * involves poking the GIC, which must be done in a
 571                  * non-preemptible context.
 572                  */
 573                 preempt_disable();
 574                 kvm_timer_flush_hwstate(vcpu);
 575                 kvm_vgic_flush_hwstate(vcpu);
 576
 577                 local_irq_disable();
 578
 579                 /*
 580                  * Re-check atomic conditions
 581                  */
 582                 if (signal_pending(current)) {
 583                         ret = -EINTR;
 584                         run->exit_reason = KVM_EXIT_INTR;
 585                 }
 586
 587                 if (ret <= 0 || need_new_vmid_gen(vcpu->kvm) ||
 588                         vcpu->arch.power_off || vcpu->arch.pause) {
 589                         local_irq_enable();
 590                         kvm_timer_sync_hwstate(vcpu);
 591                         kvm_vgic_sync_hwstate(vcpu);
 592                         preempt_enable();
 593                         continue;
 594                 }
 595
 596                 kvm_arm_setup_debug(vcpu);
 597
 598                 /**************************************************************
 599                  * Enter the guest
 600                  */
 601                 trace_kvm_entry(*vcpu_pc(vcpu));
 602                 __kvm_guest_enter();
 603                 vcpu->mode = IN_GUEST_MODE;
 604
 605                 ret = kvm_call_hyp(__kvm_vcpu_run, vcpu);
 606
 607                 vcpu->mode = OUTSIDE_GUEST_MODE;
 608                 /*
 609                  * Back from guest
 610                  *************************************************************/
 611
 612                 kvm_arm_clear_debug(vcpu);
 613
 614                 /*
 615                  * We may have taken a host interrupt in HYP mode (ie
 616                  * while executing the guest). This interrupt is still
 617                  * pending, as we haven't serviced it yet!
 618                  *
 619                  * We're now back in SVC mode, with interrupts
 620                  * disabled.  Enabling the interrupts now will have
 621                  * the effect of taking the interrupt again, in SVC
 622                  * mode this time.
 623                  */
 624                 local_irq_enable();
 625
 626                 /*
 627                  * We do local_irq_enable() before calling kvm_guest_exit() so
 628                  * that if a timer interrupt hits while running the guest we
 629                  * account that tick as being spent in the guest.  We enable
 630                  * preemption after calling kvm_guest_exit() so that if we get
 631                  * preempted we make sure ticks after that is not counted as
 632                  * guest time.
 633                  */
 634                 kvm_guest_exit();
 635                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
 636
 637                 /*
 638                  * We must sync the timer state before the vgic state so that
 639                  * the vgic can properly sample the updated state of the
 640                  * interrupt line.
 641                  */
 642                 kvm_timer_sync_hwstate(vcpu);
 643
 644                 kvm_vgic_sync_hwstate(vcpu);
 645
 646                 preempt_enable();
 647
 648                 ret = handle_exit(vcpu, run, ret);
 649         }
 650
 651         if (vcpu->sigset_active)
 652                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 653         return ret;
 654 }
 655
 656 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
 657 {
 658         int bit_index;
 659         bool set;
 660         unsigned long *ptr;
 661
 662         if (number == KVM_ARM_IRQ_CPU_IRQ)
 663                 bit_index = __ffs(HCR_VI);
 664         else /* KVM_ARM_IRQ_CPU_FIQ */
 665                 bit_index = __ffs(HCR_VF);
 666
 667         ptr = (unsigned long *)&vcpu->arch.irq_lines;
 668         if (level)
 669                 set = test_and_set_bit(bit_index, ptr);
 670         else
 671                 set = test_and_clear_bit(bit_index, ptr);
 672
 673         /*
 674          * If we didn't change anything, no need to wake up or kick other CPUs
 675          */
 676         if (set == level)
 677                 return 0;
 678
 679         /*
 680          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
 681          * trigger a world-switch round on the running physical CPU to set the
 682          * virtual IRQ/FIQ fields in the HCR appropriately.
 683          */
 684         kvm_vcpu_kick(vcpu);
 685
 686         return 0;
 687 }
 688
 689 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
 690                           bool line_status)
 691 {
 692         u32 irq = irq_level->irq;
 693         unsigned int irq_type, vcpu_idx, irq_num;
 694         int nrcpus = atomic_read(&kvm->online_vcpus);
 695         struct kvm_vcpu *vcpu = NULL;
 696         bool level = irq_level->level;
 697
 698         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
 699         vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
 700         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
 701
 702         trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level);
 703
 704         switch (irq_type) {
 705         case KVM_ARM_IRQ_TYPE_CPU:
 706                 if (irqchip_in_kernel(kvm))
 707                         return -ENXIO;
 708
 709                 if (vcpu_idx >= nrcpus)
 710                         return -EINVAL;
 711
 712                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
 713                 if (!vcpu)
 714                         return -EINVAL;
 715
 716                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
 717                         return -EINVAL;
 718
 719                 return vcpu_interrupt_line(vcpu, irq_num, level);
 720         case KVM_ARM_IRQ_TYPE_PPI:
 721                 if (!irqchip_in_kernel(kvm))
 722                         return -ENXIO;
 723
 724                 if (vcpu_idx >= nrcpus)
 725                         return -EINVAL;
 726
 727                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
 728                 if (!vcpu)
 729                         return -EINVAL;
 730
 731                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
 732                         return -EINVAL;
 733
 734                 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level);
 735         case KVM_ARM_IRQ_TYPE_SPI:
 736                 if (!irqchip_in_kernel(kvm))
 737                         return -ENXIO;
 738
 739                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
 740                         return -EINVAL;
 741
 742                 return kvm_vgic_inject_irq(kvm, 0, irq_num, level);
 743         }
 744
 745         return -EINVAL;
 746 }
 747
 748 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
 749                                const struct kvm_vcpu_init *init)
 750 {
 751         unsigned int i;
 752         int phys_target = kvm_target_cpu();
 753
 754         if (init->target != phys_target)
 755                 return -EINVAL;
 756
 757         /*
 758          * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
 759          * use the same target.
 760          */
 761         if (vcpu->arch.target != -1 && vcpu->arch.target != init->target)
 762                 return -EINVAL;
 763
 764         /* -ENOENT for unknown features, -EINVAL for invalid combinations. */
 765         for (i = 0; i < sizeof(init->features) * 8; i++) {
 766                 bool set = (init->features[i / 32] & (1 << (i % 32)));
 767
 768                 if (set && i >= KVM_VCPU_MAX_FEATURES)
 769                         return -ENOENT;
 770
 771                 /*
 772                  * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
 773                  * use the same feature set.
 774                  */
 775                 if (vcpu->arch.target != -1 && i < KVM_VCPU_MAX_FEATURES &&
 776                     test_bit(i, vcpu->arch.features) != set)
 777                         return -EINVAL;
 778
 779                 if (set)
 780                         set_bit(i, vcpu->arch.features);
 781         }
 782
 783         vcpu->arch.target = phys_target;
 784
 785         /* Now we know what it is, we can reset it. */
 786         return kvm_reset_vcpu(vcpu);
 787 }
 788
 789
 790 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
 791                                          struct kvm_vcpu_init *init)
 792 {
 793         int ret;
 794
 795         ret = kvm_vcpu_set_target(vcpu, init);
 796         if (ret)
 797                 return ret;
 798
 799         /*
 800          * Ensure a rebooted VM will fault in RAM pages and detect if the
 801          * guest MMU is turned off and flush the caches as needed.
 802          */
 803         if (vcpu->arch.has_run_once)
 804                 stage2_unmap_vm(vcpu->kvm);
 805
 806         vcpu_reset_hcr(vcpu);
 807
 808         /*
 809          * Handle the "start in power-off" case.
 810          */
 811         if (test_bit(KVM_ARM_VCPU_POWER_OFF, vcpu->arch.features))
 812                 vcpu->arch.power_off = true;
 813         else
 814                 vcpu->arch.power_off = false;
 815
 816         return 0;
 817 }
 818
 819 long kvm_arch_vcpu_ioctl(struct file *filp,
 820                          unsigned int ioctl, unsigned long arg)
 821 {
 822         struct kvm_vcpu *vcpu = filp->private_data;
 823         void __user *argp = (void __user *)arg;
 824
 825         switch (ioctl) {
 826         case KVM_ARM_VCPU_INIT: {
 827                 struct kvm_vcpu_init init;
 828
 829                 if (copy_from_user(&init, argp, sizeof(init)))
 830                         return -EFAULT;
 831
 832                 return kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
 833         }
 834         case KVM_SET_ONE_REG:
 835         case KVM_GET_ONE_REG: {
 836                 struct kvm_one_reg reg;
 837
 838                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
 839                         return -ENOEXEC;
 840
 841                 if (copy_from_user(&reg, argp, sizeof(reg)))
 842                         return -EFAULT;
 843                 if (ioctl == KVM_SET_ONE_REG)
 844                         return kvm_arm_set_reg(vcpu, &reg);
 845                 else
 846                         return kvm_arm_get_reg(vcpu, &reg);
 847         }
 848         case KVM_GET_REG_LIST: {
 849                 struct kvm_reg_list __user *user_list = argp;
 850                 struct kvm_reg_list reg_list;
 851                 unsigned n;
 852
 853                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
 854                         return -ENOEXEC;
 855
 856                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
 857                         return -EFAULT;
 858                 n = reg_list.n;
 859                 reg_list.n = kvm_arm_num_regs(vcpu);
 860                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
 861                         return -EFAULT;
 862                 if (n < reg_list.n)
 863                         return -E2BIG;
 864                 return kvm_arm_copy_reg_indices(vcpu, user_list->reg);
 865         }
 866         default:
 867                 return -EINVAL;
 868         }
 869 }
 870
 871 /**
 872  * kvm_vm_ioctl_get_dirty_log - get and clear the log of dirty pages in a slot
 873  * @kvm: kvm instance
 874  * @log: slot id and address to which we copy the log
 875  *
 876  * Steps 1-4 below provide general overview of dirty page logging. See
 877  * kvm_get_dirty_log_protect() function description for additional details.
 878  *
 879  * We call kvm_get_dirty_log_protect() to handle steps 1-3, upon return we
 880  * always flush the TLB (step 4) even if previous step failed  and the dirty
 881  * bitmap may be corrupt. Regardless of previous outcome the KVM logging API
 882  * does not preclude user space subsequent dirty log read. Flushing TLB ensures
 883  * writes will be marked dirty for next log read.
 884  *
 885  *   1. Take a snapshot of the bit and clear it if needed.
 886  *   2. Write protect the corresponding page.
 887  *   3. Copy the snapshot to the userspace.
 888  *   4. Flush TLB's if needed.
 889  */
 890 int kvm_vm_ioctl_get_dirty_log(struct kvm *kvm, struct kvm_dirty_log *log)
 891 {
 892         bool is_dirty = false;
 893         int r;
 894
 895         mutex_lock(&kvm->slots_lock);
 896
 897         r = kvm_get_dirty_log_protect(kvm, log, &is_dirty);
 898
 899         if (is_dirty)
 900                 kvm_flush_remote_tlbs(kvm);
 901
 902         mutex_unlock(&kvm->slots_lock);
 903         return r;
 904 }
 905
 906 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
 907                                         struct kvm_arm_device_addr *dev_addr)
 908 {
 909         unsigned long dev_id, type;
 910
 911         dev_id = (dev_addr->id & KVM_ARM_DEVICE_ID_MASK) >>
 912                 KVM_ARM_DEVICE_ID_SHIFT;
 913         type = (dev_addr->id & KVM_ARM_DEVICE_TYPE_MASK) >>
 914                 KVM_ARM_DEVICE_TYPE_SHIFT;
 915
 916         switch (dev_id) {
 917         case KVM_ARM_DEVICE_VGIC_V2:
 918                 if (!vgic_present)
 919                         return -ENXIO;
 920                 return kvm_vgic_addr(kvm, type, &dev_addr->addr, true);
 921         default:
 922                 return -ENODEV;
 923         }
 924 }
 925
 926 long kvm_arch_vm_ioctl(struct file *filp,
 927                        unsigned int ioctl, unsigned long arg)
 928 {
 929         struct kvm *kvm = filp->private_data;
 930         void __user *argp = (void __user *)arg;
 931
 932         switch (ioctl) {
 933         case KVM_CREATE_IRQCHIP: {
 934                 if (!vgic_present)
 935                         return -ENXIO;
 936                 return kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
 937         }
 938         case KVM_ARM_SET_DEVICE_ADDR: {
 939                 struct kvm_arm_device_addr dev_addr;
 940
 941                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
 942                         return -EFAULT;
 943                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
 944         }
 945         case KVM_ARM_PREFERRED_TARGET: {
 946                 int err;
 947                 struct kvm_vcpu_init init;
 948
 949                 err = kvm_vcpu_preferred_target(&init);
 950                 if (err)
 951                         return err;
 952
 953                 if (copy_to_user(argp, &init, sizeof(init)))
 954                         return -EFAULT;
 955
 956                 return 0;
 957         }
 958         default:
 959                 return -EINVAL;
 960         }
 961 }
 962
 963 static void cpu_init_hyp_mode(void *dummy)
 964 {
 965         phys_addr_t boot_pgd_ptr;
 966         phys_addr_t pgd_ptr;
 967         unsigned long hyp_stack_ptr;
 968         unsigned long stack_page;
 969         unsigned long vector_ptr;
 970
 971         /* Switch from the HYP stub to our own HYP init vector */
 972         __hyp_set_vectors(kvm_get_idmap_vector());
 973
 974         boot_pgd_ptr = kvm_mmu_get_boot_httbr();
 975         pgd_ptr = kvm_mmu_get_httbr();
 976         stack_page = __this_cpu_read(kvm_arm_hyp_stack_page);
 977         hyp_stack_ptr = stack_page + PAGE_SIZE;
 978         vector_ptr = (unsigned long)kvm_ksym_ref(__kvm_hyp_vector);
 979
 980         __cpu_init_hyp_mode(boot_pgd_ptr, pgd_ptr, hyp_stack_ptr, vector_ptr);
 981         __cpu_init_stage2();
 982
 983         kvm_arm_init_debug();
 984 }
 985
 986 static void cpu_hyp_reinit(void)
 987 {
 988         if (is_kernel_in_hyp_mode()) {
 989                 /*
 990                  * __cpu_init_stage2() is safe to call even if the PM
 991                  * event was cancelled before the CPU was reset.
 992                  */
 993                 __cpu_init_stage2();
 994         } else {
 995                 if (__hyp_get_vectors() == hyp_default_vectors)
 996                         cpu_init_hyp_mode(NULL);
 997         }
 998 }
 999
1000 static void cpu_hyp_reset(void)
1001 {
1002         phys_addr_t boot_pgd_ptr;
1003         phys_addr_t phys_idmap_start;
1004
1005         if (!is_kernel_in_hyp_mode()) {
1006                 boot_pgd_ptr = kvm_mmu_get_boot_httbr();
1007                 phys_idmap_start = kvm_get_idmap_start();
1008
1009                 __cpu_reset_hyp_mode(boot_pgd_ptr, phys_idmap_start);
1010         }
1011 }
1012
1013 static void _kvm_arch_hardware_enable(void *discard)
1014 {
1015         if (!__this_cpu_read(kvm_arm_hardware_enabled)) {
1016                 cpu_hyp_reinit();
1017                 __this_cpu_write(kvm_arm_hardware_enabled, 1);
1018         }
1019 }
1020
1021 int kvm_arch_hardware_enable(void)
1022 {
1023         _kvm_arch_hardware_enable(NULL);
1024         return 0;
1025 }
1026
1027 static void _kvm_arch_hardware_disable(void *discard)
1028 {
1029         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1030                 cpu_hyp_reset();
1031                 __this_cpu_write(kvm_arm_hardware_enabled, 0);
1032         }
1033 }
1034
1035 void kvm_arch_hardware_disable(void)
1036 {
1037         _kvm_arch_hardware_disable(NULL);
1038 }
1039
1040 #ifdef CONFIG_CPU_PM
1041 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
1042                                     unsigned long cmd,
1043                                     void *v)
1044 {
1045         /*
1046          * kvm_arm_hardware_enabled is left with its old value over
1047          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
1048          * re-enable hyp.
1049          */
1050         switch (cmd) {
1051         case CPU_PM_ENTER:
1052                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1053                         /*
1054                          * don't update kvm_arm_hardware_enabled here
1055                          * so that the hardware will be re-enabled
1056                          * when we resume. See below.
1057                          */
1058                         cpu_hyp_reset();
1059
1060                 return NOTIFY_OK;
1061         case CPU_PM_EXIT:
1062                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1063                         /* The hardware was enabled before suspend. */
1064                         cpu_hyp_reinit();
1065
1066                 return NOTIFY_OK;
1067
1068         default:
1069                 return NOTIFY_DONE;
1070         }
1071 }
1072
1073 static struct notifier_block hyp_init_cpu_pm_nb = {
1074         .notifier_call = hyp_init_cpu_pm_notifier,
1075 };
1076
1077 static void __init hyp_cpu_pm_init(void)
1078 {
1079         cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1080 }
1081 #else
1082 static inline void hyp_cpu_pm_init(void)
1083 {
1084 }
1085 #endif
1086
1087 static void teardown_common_resources(void)
1088 {
1089         free_percpu(kvm_host_cpu_state);
1090 }
1091
1092 static int init_common_resources(void)
1093 {
1094         kvm_host_cpu_state = alloc_percpu(kvm_cpu_context_t);
1095         if (!kvm_host_cpu_state) {
1096                 kvm_err("Cannot allocate host CPU state\n");
1097                 return -ENOMEM;
1098         }
1099
1100         return 0;
1101 }
1102
1103 static int init_subsystems(void)
1104 {
1105         int err = 0;
1106
1107         /*
1108          * Enable hardware so that subsystem initialisation can access EL2.
1109          */
1110         on_each_cpu(_kvm_arch_hardware_enable, NULL, 1);
1111
1112         /*
1113          * Register CPU lower-power notifier
1114          */
1115         hyp_cpu_pm_init();
1116
1117         /*
1118          * Init HYP view of VGIC
1119          */
1120         err = kvm_vgic_hyp_init();
1121         switch (err) {
1122         case 0:
1123                 vgic_present = true;
1124                 break;
1125         case -ENODEV:
1126         case -ENXIO:
1127                 vgic_present = false;
1128                 err = 0;
1129                 break;
1130         default:
1131                 goto out;
1132         }
1133
1134         /*
1135          * Init HYP architected timer support
1136          */
1137         err = kvm_timer_hyp_init();
1138         if (err)
1139                 goto out;
1140
1141         kvm_perf_init();
1142         kvm_coproc_table_init();
1143
1144 out:
1145         on_each_cpu(_kvm_arch_hardware_disable, NULL, 1);
1146
1147         return err;
1148 }
1149
1150 static void teardown_hyp_mode(void)
1151 {
1152         int cpu;
1153
1154         if (is_kernel_in_hyp_mode())
1155                 return;
1156
1157         free_hyp_pgds();
1158         for_each_possible_cpu(cpu)
1159                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
1160 }
1161
1162 static int init_vhe_mode(void)
1163 {
1164         /* set size of VMID supported by CPU */
1165         kvm_vmid_bits = kvm_get_vmid_bits();
1166         kvm_info("%d-bit VMID\n", kvm_vmid_bits);
1167
1168         kvm_info("VHE mode initialized successfully\n");
1169         return 0;
1170 }
1171
1172 /**
1173  * Inits Hyp-mode on all online CPUs
1174  */
1175 static int init_hyp_mode(void)
1176 {
1177         int cpu;
1178         int err = 0;
1179
1180         /*
1181          * Allocate Hyp PGD and setup Hyp identity mapping
1182          */
1183         err = kvm_mmu_init();
1184         if (err)
1185                 goto out_err;
1186
1187         /*
1188          * It is probably enough to obtain the default on one
1189          * CPU. It's unlikely to be different on the others.
1190          */
1191         hyp_default_vectors = __hyp_get_vectors();
1192
1193         /*
1194          * Allocate stack pages for Hypervisor-mode
1195          */
1196         for_each_possible_cpu(cpu) {
1197                 unsigned long stack_page;
1198
1199                 stack_page = __get_free_page(GFP_KERNEL);
1200                 if (!stack_page) {
1201                         err = -ENOMEM;
1202                         goto out_err;
1203                 }
1204
1205                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
1206         }
1207
1208         /*
1209          * Map the Hyp-code called directly from the host
1210          */
1211         err = create_hyp_mappings(kvm_ksym_ref(__kvm_hyp_code_start),
1212                                   kvm_ksym_ref(__kvm_hyp_code_end));
1213         if (err) {
1214                 kvm_err("Cannot map world-switch code\n");
1215                 goto out_err;
1216         }
1217
1218         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
1219                                   kvm_ksym_ref(__end_rodata));
1220         if (err) {
1221                 kvm_err("Cannot map rodata section\n");
1222                 goto out_err;
1223         }
1224
1225         /*
1226          * Map the Hyp stack pages
1227          */
1228         for_each_possible_cpu(cpu) {
1229                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
1230                 err = create_hyp_mappings(stack_page, stack_page + PAGE_SIZE);
1231
1232                 if (err) {
1233                         kvm_err("Cannot map hyp stack\n");
1234                         goto out_err;
1235                 }
1236         }
1237
1238         for_each_possible_cpu(cpu) {
1239                 kvm_cpu_context_t *cpu_ctxt;
1240
1241                 cpu_ctxt = per_cpu_ptr(kvm_host_cpu_state, cpu);
1242                 err = create_hyp_mappings(cpu_ctxt, cpu_ctxt + 1);
1243
1244                 if (err) {
1245                         kvm_err("Cannot map host CPU state: %d\n", err);
1246                         goto out_err;
1247                 }
1248         }
1249
1250 #ifndef CONFIG_HOTPLUG_CPU
1251         free_boot_hyp_pgd();
1252 #endif
1253
1254         /* set size of VMID supported by CPU */
1255         kvm_vmid_bits = kvm_get_vmid_bits();
1256         kvm_info("%d-bit VMID\n", kvm_vmid_bits);
1257
1258         kvm_info("Hyp mode initialized successfully\n");
1259
1260         return 0;
1261
1262 out_err:
1263         teardown_hyp_mode();
1264         kvm_err("error initializing Hyp mode: %d\n", err);
1265         return err;
1266 }
1267
1268 static void check_kvm_target_cpu(void *ret)
1269 {
1270         *(int *)ret = kvm_target_cpu();
1271 }
1272
1273 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
1274 {
1275         struct kvm_vcpu *vcpu;
1276         int i;
1277
1278         mpidr &= MPIDR_HWID_BITMASK;
1279         kvm_for_each_vcpu(i, vcpu, kvm) {
1280                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
1281                         return vcpu;
1282         }
1283         return NULL;
1284 }
1285
1286 /**
1287  * Initialize Hyp-mode and memory mappings on all CPUs.
1288  */
1289 int kvm_arch_init(void *opaque)
1290 {
1291         int err;
1292         int ret, cpu;
1293
1294         if (!is_hyp_mode_available()) {
1295                 kvm_err("HYP mode not available\n");
1296                 return -ENODEV;
1297         }
1298
1299         for_each_online_cpu(cpu) {
1300                 smp_call_function_single(cpu, check_kvm_target_cpu, &ret, 1);
1301                 if (ret < 0) {
1302                         kvm_err("Error, CPU %d not supported!\n", cpu);
1303                         return -ENODEV;
1304                 }
1305         }
1306
1307         err = init_common_resources();
1308         if (err)
1309                 return err;
1310
1311         if (is_kernel_in_hyp_mode())
1312                 err = init_vhe_mode();
1313         else
1314                 err = init_hyp_mode();
1315         if (err)
1316                 goto out_err;
1317
1318         err = init_subsystems();
1319         if (err)
1320                 goto out_hyp;
1321
1322         return 0;
1323
1324 out_hyp:
1325         teardown_hyp_mode();
1326 out_err:
1327         teardown_common_resources();
1328         return err;
1329 }
1330
1331 /* NOP: Compiling as a module not supported */
1332 void kvm_arch_exit(void)
1333 {
1334         kvm_perf_teardown();
1335 }
1336
1337 static int arm_init(void)
1338 {
1339         int rc = kvm_init(NULL, sizeof(struct kvm_vcpu), 0, THIS_MODULE);
1340         return rc;
1341 }
1342
1343 module_init(arm_init);