kernel/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.h>
  20 #include <linux/cpu_pm.h>
  21 #include <linux/errno.h>
  22 #include <linux/err.h>
  23 #include <linux/kvm_host.h>
  24 #include <linux/module.h>
  25 #include <linux/vmalloc.h>
  26 #include <linux/fs.h>
  27 #include <linux/mman.h>
  28 #include <linux/sched.h>
  29 #include <linux/kvm.h>
  30 #include <trace/events/kvm.h>
  31
  32 #define CREATE_TRACE_POINTS
  33 #include "trace.h"
  34
  35 #include <asm/uaccess.h>
  36 #include <asm/ptrace.h>
  37 #include <asm/mman.h>
  38 #include <asm/tlbflush.h>
  39 #include <asm/cacheflush.h>
  40 #include <asm/virt.h>
  41 #include <asm/kvm_arm.h>
  42 #include <asm/kvm_asm.h>
  43 #include <asm/kvm_mmu.h>
  44 #include <asm/kvm_emulate.h>
  45 #include <asm/kvm_coproc.h>
  46 #include <asm/kvm_psci.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 u8 kvm_next_vmid;
  62 static DEFINE_SPINLOCK(kvm_vmid_lock);
  63
  64 static void kvm_arm_set_running_vcpu(struct kvm_vcpu *vcpu)
  65 {
  66         BUG_ON(preemptible());
  67         __this_cpu_write(kvm_arm_running_vcpu, vcpu);
  68 }
  69
  70 /**
  71  * kvm_arm_get_running_vcpu - get the vcpu running on the current CPU.
  72  * Must be called from non-preemptible context
  73  */
  74 struct kvm_vcpu *kvm_arm_get_running_vcpu(void)
  75 {
  76         BUG_ON(preemptible());
  77         return __this_cpu_read(kvm_arm_running_vcpu);
  78 }
  79
  80 /**
  81  * kvm_arm_get_running_vcpus - get the per-CPU array of currently running vcpus.
  82  */
  83 struct kvm_vcpu * __percpu *kvm_get_running_vcpus(void)
  84 {
  85         return &kvm_arm_running_vcpu;
  86 }
  87
  88 int kvm_arch_hardware_enable(void)
  89 {
  90         return 0;
  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 = kvm_vgic_get_max_vcpus();
 136
 137         return ret;
 138 out_free_stage2_pgd:
 139         kvm_free_stage2_pgd(kvm);
 140 out_fail_alloc:
 141         return ret;
 142 }
 143
 144 int kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 145 {
 146         return VM_FAULT_SIGBUS;
 147 }
 148
 149
 150 /**
 151  * kvm_arch_destroy_vm - destroy the VM data structure
 152  * @kvm:        pointer to the KVM struct
 153  */
 154 void kvm_arch_destroy_vm(struct kvm *kvm)
 155 {
 156         int i;
 157
 158         for (i = 0; i < KVM_MAX_VCPUS; ++i) {
 159                 if (kvm->vcpus[i]) {
 160                         kvm_arch_vcpu_free(kvm->vcpus[i]);
 161                         kvm->vcpus[i] = NULL;
 162                 }
 163         }
 164
 165         kvm_vgic_destroy(kvm);
 166 }
 167
 168 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 169 {
 170         int r;
 171         switch (ext) {
 172         case KVM_CAP_IRQCHIP:
 173         case KVM_CAP_IOEVENTFD:
 174         case KVM_CAP_DEVICE_CTRL:
 175         case KVM_CAP_USER_MEMORY:
 176         case KVM_CAP_SYNC_MMU:
 177         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 178         case KVM_CAP_ONE_REG:
 179         case KVM_CAP_ARM_PSCI:
 180         case KVM_CAP_ARM_PSCI_0_2:
 181         case KVM_CAP_READONLY_MEM:
 182         case KVM_CAP_MP_STATE:
 183                 r = 1;
 184                 break;
 185         case KVM_CAP_COALESCED_MMIO:
 186                 r = KVM_COALESCED_MMIO_PAGE_OFFSET;
 187                 break;
 188         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 189                 r = 1;
 190                 break;
 191         case KVM_CAP_NR_VCPUS:
 192                 r = num_online_cpus();
 193                 break;
 194         case KVM_CAP_MAX_VCPUS:
 195                 r = KVM_MAX_VCPUS;
 196                 break;
 197         default:
 198                 r = kvm_arch_dev_ioctl_check_extension(ext);
 199                 break;
 200         }
 201         return r;
 202 }
 203
 204 long kvm_arch_dev_ioctl(struct file *filp,
 205                         unsigned int ioctl, unsigned long arg)
 206 {
 207         return -EINVAL;
 208 }
 209
 210
 211 struct kvm_vcpu *kvm_arch_vcpu_create(struct kvm *kvm, unsigned int id)
 212 {
 213         int err;
 214         struct kvm_vcpu *vcpu;
 215
 216         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm)) {
 217                 err = -EBUSY;
 218                 goto out;
 219         }
 220
 221         if (id >= kvm->arch.max_vcpus) {
 222                 err = -EINVAL;
 223                 goto out;
 224         }
 225
 226         vcpu = kmem_cache_zalloc(kvm_vcpu_cache, GFP_KERNEL);
 227         if (!vcpu) {
 228                 err = -ENOMEM;
 229                 goto out;
 230         }
 231
 232         err = kvm_vcpu_init(vcpu, kvm, id);
 233         if (err)
 234                 goto free_vcpu;
 235
 236         err = create_hyp_mappings(vcpu, vcpu + 1);
 237         if (err)
 238                 goto vcpu_uninit;
 239
 240         return vcpu;
 241 vcpu_uninit:
 242         kvm_vcpu_uninit(vcpu);
 243 free_vcpu:
 244         kmem_cache_free(kvm_vcpu_cache, vcpu);
 245 out:
 246         return ERR_PTR(err);
 247 }
 248
 249 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
 250 {
 251         kvm_vgic_vcpu_early_init(vcpu);
 252 }
 253
 254 void kvm_arch_vcpu_free(struct kvm_vcpu *vcpu)
 255 {
 256         kvm_mmu_free_memory_caches(vcpu);
 257         kvm_timer_vcpu_terminate(vcpu);
 258         kvm_vgic_vcpu_destroy(vcpu);
 259         kmem_cache_free(kvm_vcpu_cache, vcpu);
 260 }
 261
 262 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 263 {
 264         kvm_arch_vcpu_free(vcpu);
 265 }
 266
 267 int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu)
 268 {
 269         return kvm_timer_should_fire(vcpu);
 270 }
 271
 272 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 273 {
 274         kvm_timer_schedule(vcpu);
 275 }
 276
 277 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 278 {
 279         kvm_timer_unschedule(vcpu);
 280 }
 281
 282 int kvm_arch_vcpu_init(struct kvm_vcpu *vcpu)
 283 {
 284         /* Force users to call KVM_ARM_VCPU_INIT */
 285         vcpu->arch.target = -1;
 286         bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
 287
 288         /* Set up the timer */
 289         kvm_timer_vcpu_init(vcpu);
 290
 291         kvm_arm_reset_debug_ptr(vcpu);
 292
 293         return 0;
 294 }
 295
 296 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 297 {
 298         vcpu->cpu = cpu;
 299         vcpu->arch.host_cpu_context = this_cpu_ptr(kvm_host_cpu_state);
 300
 301         kvm_arm_set_running_vcpu(vcpu);
 302 }
 303
 304 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 305 {
 306         /*
 307          * The arch-generic KVM code expects the cpu field of a vcpu to be -1
 308          * if the vcpu is no longer assigned to a cpu.  This is used for the
 309          * optimized make_all_cpus_request path.
 310          */
 311         vcpu->cpu = -1;
 312
 313         kvm_arm_set_running_vcpu(NULL);
 314 }
 315
 316 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 317                                     struct kvm_mp_state *mp_state)
 318 {
 319         if (vcpu->arch.power_off)
 320                 mp_state->mp_state = KVM_MP_STATE_STOPPED;
 321         else
 322                 mp_state->mp_state = KVM_MP_STATE_RUNNABLE;
 323
 324         return 0;
 325 }
 326
 327 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
 328                                     struct kvm_mp_state *mp_state)
 329 {
 330         switch (mp_state->mp_state) {
 331         case KVM_MP_STATE_RUNNABLE:
 332                 vcpu->arch.power_off = false;
 333                 break;
 334         case KVM_MP_STATE_STOPPED:
 335                 vcpu->arch.power_off = true;
 336                 break;
 337         default:
 338                 return -EINVAL;
 339         }
 340
 341         return 0;
 342 }
 343
 344 /**
 345  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
 346  * @v:          The VCPU pointer
 347  *
 348  * If the guest CPU is not waiting for interrupts or an interrupt line is
 349  * asserted, the CPU is by definition runnable.
 350  */
 351 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
 352 {
 353         return ((!!v->arch.irq_lines || kvm_vgic_vcpu_pending_irq(v))
 354                 && !v->arch.power_off && !v->arch.pause);
 355 }
 356
 357 /* Just ensure a guest exit from a particular CPU */
 358 static void exit_vm_noop(void *info)
 359 {
 360 }
 361
 362 void force_vm_exit(const cpumask_t *mask)
 363 {
 364         smp_call_function_many(mask, exit_vm_noop, NULL, true);
 365 }
 366
 367 /**
 368  * need_new_vmid_gen - check that the VMID is still valid
 369  * @kvm: The VM's VMID to checkt
 370  *
 371  * return true if there is a new generation of VMIDs being used
 372  *
 373  * The hardware supports only 256 values with the value zero reserved for the
 374  * host, so we check if an assigned value belongs to a previous generation,
 375  * which which requires us to assign a new value. If we're the first to use a
 376  * VMID for the new generation, we must flush necessary caches and TLBs on all
 377  * CPUs.
 378  */
 379 static bool need_new_vmid_gen(struct kvm *kvm)
 380 {
 381         return unlikely(kvm->arch.vmid_gen != atomic64_read(&kvm_vmid_gen));
 382 }
 383
 384 /**
 385  * update_vttbr - Update the VTTBR with a valid VMID before the guest runs
 386  * @kvm The guest that we are about to run
 387  *
 388  * Called from kvm_arch_vcpu_ioctl_run before entering the guest to ensure the
 389  * VM has a valid VMID, otherwise assigns a new one and flushes corresponding
 390  * caches and TLBs.
 391  */
 392 static void update_vttbr(struct kvm *kvm)
 393 {
 394         phys_addr_t pgd_phys;
 395         u64 vmid;
 396
 397         if (!need_new_vmid_gen(kvm))
 398                 return;
 399
 400         spin_lock(&kvm_vmid_lock);
 401
 402         /*
 403          * We need to re-check the vmid_gen here to ensure that if another vcpu
 404          * already allocated a valid vmid for this vm, then this vcpu should
 405          * use the same vmid.
 406          */
 407         if (!need_new_vmid_gen(kvm)) {
 408                 spin_unlock(&kvm_vmid_lock);
 409                 return;
 410         }
 411
 412         /* First user of a new VMID generation? */
 413         if (unlikely(kvm_next_vmid == 0)) {
 414                 atomic64_inc(&kvm_vmid_gen);
 415                 kvm_next_vmid = 1;
 416
 417                 /*
 418                  * On SMP we know no other CPUs can use this CPU's or each
 419                  * other's VMID after force_vm_exit returns since the
 420                  * kvm_vmid_lock blocks them from reentry to the guest.
 421                  */
 422                 force_vm_exit(cpu_all_mask);
 423                 /*
 424                  * Now broadcast TLB + ICACHE invalidation over the inner
 425                  * shareable domain to make sure all data structures are
 426                  * clean.
 427                  */
 428                 kvm_call_hyp(__kvm_flush_vm_context);
 429         }
 430
 431         kvm->arch.vmid_gen = atomic64_read(&kvm_vmid_gen);
 432         kvm->arch.vmid = kvm_next_vmid;
 433         kvm_next_vmid++;
 434
 435         /* update vttbr to be used with the new vmid */
 436         pgd_phys = virt_to_phys(kvm_get_hwpgd(kvm));
 437         BUG_ON(pgd_phys & ~VTTBR_BADDR_MASK);
 438         vmid = ((u64)(kvm->arch.vmid) << VTTBR_VMID_SHIFT) & VTTBR_VMID_MASK;
 439         kvm->arch.vttbr = pgd_phys | vmid;
 440
 441         spin_unlock(&kvm_vmid_lock);
 442 }
 443
 444 static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
 445 {
 446         struct kvm *kvm = vcpu->kvm;
 447         int ret;
 448
 449         if (likely(vcpu->arch.has_run_once))
 450                 return 0;
 451
 452         vcpu->arch.has_run_once = true;
 453
 454         /*
 455          * Map the VGIC hardware resources before running a vcpu the first
 456          * time on this VM.
 457          */
 458         if (unlikely(irqchip_in_kernel(kvm) && !vgic_ready(kvm))) {
 459                 ret = kvm_vgic_map_resources(kvm);
 460                 if (ret)
 461                         return ret;
 462         }
 463
 464         /*
 465          * Enable the arch timers only if we have an in-kernel VGIC
 466          * and it has been properly initialized, since we cannot handle
 467          * interrupts from the virtual timer with a userspace gic.
 468          */
 469         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 470                 kvm_timer_enable(kvm);
 471
 472         return 0;
 473 }
 474
 475 bool kvm_arch_intc_initialized(struct kvm *kvm)
 476 {
 477         return vgic_initialized(kvm);
 478 }
 479
 480 static void kvm_arm_halt_guest(struct kvm *kvm) __maybe_unused;
 481 static void kvm_arm_resume_guest(struct kvm *kvm) __maybe_unused;
 482
 483 static void kvm_arm_halt_guest(struct kvm *kvm)
 484 {
 485         int i;
 486         struct kvm_vcpu *vcpu;
 487
 488         kvm_for_each_vcpu(i, vcpu, kvm)
 489                 vcpu->arch.pause = true;
 490         force_vm_exit(cpu_all_mask);
 491 }
 492
 493 static void kvm_arm_resume_guest(struct kvm *kvm)
 494 {
 495         int i;
 496         struct kvm_vcpu *vcpu;
 497
 498         kvm_for_each_vcpu(i, vcpu, kvm) {
 499                 struct swait_queue_head *wq = kvm_arch_vcpu_wq(vcpu);
 500
 501                 vcpu->arch.pause = false;
 502                 swake_up(wq);
 503         }
 504 }
 505
 506 static void vcpu_sleep(struct kvm_vcpu *vcpu)
 507 {
 508         struct swait_queue_head *wq = kvm_arch_vcpu_wq(vcpu);
 509
 510         swait_event_interruptible(*wq, ((!vcpu->arch.power_off) &&
 511                                        (!vcpu->arch.pause)));
 512 }
 513
 514 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu)
 515 {
 516         return vcpu->arch.target >= 0;
 517 }
 518
 519 /**
 520  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 521  * @vcpu:       The VCPU pointer
 522  * @run:        The kvm_run structure pointer used for userspace state exchange
 523  *
 524  * This function is called through the VCPU_RUN ioctl called from user space. It
 525  * will execute VM code in a loop until the time slice for the process is used
 526  * or some emulation is needed from user space in which case the function will
 527  * return with return value 0 and with the kvm_run structure filled in with the
 528  * required data for the requested emulation.
 529  */
 530 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run)
 531 {
 532         int ret;
 533         sigset_t sigsaved;
 534
 535         if (unlikely(!kvm_vcpu_initialized(vcpu)))
 536                 return -ENOEXEC;
 537
 538         ret = kvm_vcpu_first_run_init(vcpu);
 539         if (ret)
 540                 return ret;
 541
 542         if (run->exit_reason == KVM_EXIT_MMIO) {
 543                 ret = kvm_handle_mmio_return(vcpu, vcpu->run);
 544                 if (ret)
 545                         return ret;
 546         }
 547
 548         if (vcpu->sigset_active)
 549                 sigprocmask(SIG_SETMASK, &vcpu->sigset, &sigsaved);
 550
 551         ret = 1;
 552         run->exit_reason = KVM_EXIT_UNKNOWN;
 553         while (ret > 0) {
 554                 /*
 555                  * Check conditions before entering the guest
 556                  */
 557                 cond_resched();
 558
 559                 update_vttbr(vcpu->kvm);
 560
 561                 if (vcpu->arch.power_off || vcpu->arch.pause)
 562                         vcpu_sleep(vcpu);
 563
 564                 /*
 565                  * Preparing the interrupts to be injected also
 566                  * involves poking the GIC, which must be done in a
 567                  * non-preemptible context.
 568                  */
 569                 migrate_disable();
 570                 kvm_timer_flush_hwstate(vcpu);
 571                 kvm_vgic_flush_hwstate(vcpu);
 572
 573                 local_irq_disable();
 574
 575                 /*
 576                  * Re-check atomic conditions
 577                  */
 578                 if (signal_pending(current)) {
 579                         ret = -EINTR;
 580                         run->exit_reason = KVM_EXIT_INTR;
 581                 }
 582
 583                 if (ret <= 0 || need_new_vmid_gen(vcpu->kvm) ||
 584                         vcpu->arch.power_off || vcpu->arch.pause) {
 585                         local_irq_enable();
 586                         kvm_timer_sync_hwstate(vcpu);
 587                         kvm_vgic_sync_hwstate(vcpu);
 588                         migrate_enable();
 589                         continue;
 590                 }
 591
 592                 kvm_arm_setup_debug(vcpu);
 593
 594                 /**************************************************************
 595                  * Enter the guest
 596                  */
 597                 trace_kvm_entry(*vcpu_pc(vcpu));
 598                 __kvm_guest_enter();
 599                 vcpu->mode = IN_GUEST_MODE;
 600
 601                 ret = kvm_call_hyp(__kvm_vcpu_run, vcpu);
 602
 603                 vcpu->mode = OUTSIDE_GUEST_MODE;
 604                 /*
 605                  * Back from guest
 606                  *************************************************************/
 607
 608                 kvm_arm_clear_debug(vcpu);
 609
 610                 /*
 611                  * We may have taken a host interrupt in HYP mode (ie
 612                  * while executing the guest). This interrupt is still
 613                  * pending, as we haven't serviced it yet!
 614                  *
 615                  * We're now back in SVC mode, with interrupts
 616                  * disabled.  Enabling the interrupts now will have
 617                  * the effect of taking the interrupt again, in SVC
 618                  * mode this time.
 619                  */
 620                 local_irq_enable();
 621
 622                 /*
 623                  * We do local_irq_enable() before calling kvm_guest_exit() so
 624                  * that if a timer interrupt hits while running the guest we
 625                  * account that tick as being spent in the guest.  We enable
 626                  * preemption after calling kvm_guest_exit() so that if we get
 627                  * preempted we make sure ticks after that is not counted as
 628                  * guest time.
 629                  */
 630                 kvm_guest_exit();
 631                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
 632
 633                 /*
 634                  * We must sync the timer state before the vgic state so that
 635                  * the vgic can properly sample the updated state of the
 636                  * interrupt line.
 637                  */
 638                 kvm_timer_sync_hwstate(vcpu);
 639
 640                 kvm_vgic_sync_hwstate(vcpu);
 641
 642                 migrate_enable();
 643
 644                 ret = handle_exit(vcpu, run, ret);
 645         }
 646
 647         if (vcpu->sigset_active)
 648                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
 649         return ret;
 650 }
 651
 652 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
 653 {
 654         int bit_index;
 655         bool set;
 656         unsigned long *ptr;
 657
 658         if (number == KVM_ARM_IRQ_CPU_IRQ)
 659                 bit_index = __ffs(HCR_VI);
 660         else /* KVM_ARM_IRQ_CPU_FIQ */
 661                 bit_index = __ffs(HCR_VF);
 662
 663         ptr = (unsigned long *)&vcpu->arch.irq_lines;
 664         if (level)
 665                 set = test_and_set_bit(bit_index, ptr);
 666         else
 667                 set = test_and_clear_bit(bit_index, ptr);
 668
 669         /*
 670          * If we didn't change anything, no need to wake up or kick other CPUs
 671          */
 672         if (set == level)
 673                 return 0;
 674
 675         /*
 676          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
 677          * trigger a world-switch round on the running physical CPU to set the
 678          * virtual IRQ/FIQ fields in the HCR appropriately.
 679          */
 680         kvm_vcpu_kick(vcpu);
 681
 682         return 0;
 683 }
 684
 685 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
 686                           bool line_status)
 687 {
 688         u32 irq = irq_level->irq;
 689         unsigned int irq_type, vcpu_idx, irq_num;
 690         int nrcpus = atomic_read(&kvm->online_vcpus);
 691         struct kvm_vcpu *vcpu = NULL;
 692         bool level = irq_level->level;
 693
 694         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
 695         vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
 696         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
 697
 698         trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level);
 699
 700         switch (irq_type) {
 701         case KVM_ARM_IRQ_TYPE_CPU:
 702                 if (irqchip_in_kernel(kvm))
 703                         return -ENXIO;
 704
 705                 if (vcpu_idx >= nrcpus)
 706                         return -EINVAL;
 707
 708                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
 709                 if (!vcpu)
 710                         return -EINVAL;
 711
 712                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
 713                         return -EINVAL;
 714
 715                 return vcpu_interrupt_line(vcpu, irq_num, level);
 716         case KVM_ARM_IRQ_TYPE_PPI:
 717                 if (!irqchip_in_kernel(kvm))
 718                         return -ENXIO;
 719
 720                 if (vcpu_idx >= nrcpus)
 721                         return -EINVAL;
 722
 723                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
 724                 if (!vcpu)
 725                         return -EINVAL;
 726
 727                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
 728                         return -EINVAL;
 729
 730                 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level);
 731         case KVM_ARM_IRQ_TYPE_SPI:
 732                 if (!irqchip_in_kernel(kvm))
 733                         return -ENXIO;
 734
 735                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
 736                         return -EINVAL;
 737
 738                 return kvm_vgic_inject_irq(kvm, 0, irq_num, level);
 739         }
 740
 741         return -EINVAL;
 742 }
 743
 744 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
 745                                const struct kvm_vcpu_init *init)
 746 {
 747         unsigned int i;
 748         int phys_target = kvm_target_cpu();
 749
 750         if (init->target != phys_target)
 751                 return -EINVAL;
 752
 753         /*
 754          * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
 755          * use the same target.
 756          */
 757         if (vcpu->arch.target != -1 && vcpu->arch.target != init->target)
 758                 return -EINVAL;
 759
 760         /* -ENOENT for unknown features, -EINVAL for invalid combinations. */
 761         for (i = 0; i < sizeof(init->features) * 8; i++) {
 762                 bool set = (init->features[i / 32] & (1 << (i % 32)));
 763
 764                 if (set && i >= KVM_VCPU_MAX_FEATURES)
 765                         return -ENOENT;
 766
 767                 /*
 768                  * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
 769                  * use the same feature set.
 770                  */
 771                 if (vcpu->arch.target != -1 && i < KVM_VCPU_MAX_FEATURES &&
 772                     test_bit(i, vcpu->arch.features) != set)
 773                         return -EINVAL;
 774
 775                 if (set)
 776                         set_bit(i, vcpu->arch.features);
 777         }
 778
 779         vcpu->arch.target = phys_target;
 780
 781         /* Now we know what it is, we can reset it. */
 782         return kvm_reset_vcpu(vcpu);
 783 }
 784
 785
 786 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
 787                                          struct kvm_vcpu_init *init)
 788 {
 789         int ret;
 790
 791         ret = kvm_vcpu_set_target(vcpu, init);
 792         if (ret)
 793                 return ret;
 794
 795         /*
 796          * Ensure a rebooted VM will fault in RAM pages and detect if the
 797          * guest MMU is turned off and flush the caches as needed.
 798          */
 799         if (vcpu->arch.has_run_once)
 800                 stage2_unmap_vm(vcpu->kvm);
 801
 802         vcpu_reset_hcr(vcpu);
 803
 804         /*
 805          * Handle the "start in power-off" case.
 806          */
 807         if (test_bit(KVM_ARM_VCPU_POWER_OFF, vcpu->arch.features))
 808                 vcpu->arch.power_off = true;
 809         else
 810                 vcpu->arch.power_off = false;
 811
 812         return 0;
 813 }
 814
 815 long kvm_arch_vcpu_ioctl(struct file *filp,
 816                          unsigned int ioctl, unsigned long arg)
 817 {
 818         struct kvm_vcpu *vcpu = filp->private_data;
 819         void __user *argp = (void __user *)arg;
 820
 821         switch (ioctl) {
 822         case KVM_ARM_VCPU_INIT: {
 823                 struct kvm_vcpu_init init;
 824
 825                 if (copy_from_user(&init, argp, sizeof(init)))
 826                         return -EFAULT;
 827
 828                 return kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
 829         }
 830         case KVM_SET_ONE_REG:
 831         case KVM_GET_ONE_REG: {
 832                 struct kvm_one_reg reg;
 833
 834                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
 835                         return -ENOEXEC;
 836
 837                 if (copy_from_user(&reg, argp, sizeof(reg)))
 838                         return -EFAULT;
 839                 if (ioctl == KVM_SET_ONE_REG)
 840                         return kvm_arm_set_reg(vcpu, &reg);
 841                 else
 842                         return kvm_arm_get_reg(vcpu, &reg);
 843         }
 844         case KVM_GET_REG_LIST: {
 845                 struct kvm_reg_list __user *user_list = argp;
 846                 struct kvm_reg_list reg_list;
 847                 unsigned n;
 848
 849                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
 850                         return -ENOEXEC;
 851
 852                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
 853                         return -EFAULT;
 854                 n = reg_list.n;
 855                 reg_list.n = kvm_arm_num_regs(vcpu);
 856                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
 857                         return -EFAULT;
 858                 if (n < reg_list.n)
 859                         return -E2BIG;
 860                 return kvm_arm_copy_reg_indices(vcpu, user_list->reg);
 861         }
 862         default:
 863                 return -EINVAL;
 864         }
 865 }
 866
 867 /**
 868  * kvm_vm_ioctl_get_dirty_log - get and clear the log of dirty pages in a slot
 869  * @kvm: kvm instance
 870  * @log: slot id and address to which we copy the log
 871  *
 872  * Steps 1-4 below provide general overview of dirty page logging. See
 873  * kvm_get_dirty_log_protect() function description for additional details.
 874  *
 875  * We call kvm_get_dirty_log_protect() to handle steps 1-3, upon return we
 876  * always flush the TLB (step 4) even if previous step failed  and the dirty
 877  * bitmap may be corrupt. Regardless of previous outcome the KVM logging API
 878  * does not preclude user space subsequent dirty log read. Flushing TLB ensures
 879  * writes will be marked dirty for next log read.
 880  *
 881  *   1. Take a snapshot of the bit and clear it if needed.
 882  *   2. Write protect the corresponding page.
 883  *   3. Copy the snapshot to the userspace.
 884  *   4. Flush TLB's if needed.
 885  */
 886 int kvm_vm_ioctl_get_dirty_log(struct kvm *kvm, struct kvm_dirty_log *log)
 887 {
 888         bool is_dirty = false;
 889         int r;
 890
 891         mutex_lock(&kvm->slots_lock);
 892
 893         r = kvm_get_dirty_log_protect(kvm, log, &is_dirty);
 894
 895         if (is_dirty)
 896                 kvm_flush_remote_tlbs(kvm);
 897
 898         mutex_unlock(&kvm->slots_lock);
 899         return r;
 900 }
 901
 902 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
 903                                         struct kvm_arm_device_addr *dev_addr)
 904 {
 905         unsigned long dev_id, type;
 906
 907         dev_id = (dev_addr->id & KVM_ARM_DEVICE_ID_MASK) >>
 908                 KVM_ARM_DEVICE_ID_SHIFT;
 909         type = (dev_addr->id & KVM_ARM_DEVICE_TYPE_MASK) >>
 910                 KVM_ARM_DEVICE_TYPE_SHIFT;
 911
 912         switch (dev_id) {
 913         case KVM_ARM_DEVICE_VGIC_V2:
 914                 return kvm_vgic_addr(kvm, type, &dev_addr->addr, true);
 915         default:
 916                 return -ENODEV;
 917         }
 918 }
 919
 920 long kvm_arch_vm_ioctl(struct file *filp,
 921                        unsigned int ioctl, unsigned long arg)
 922 {
 923         struct kvm *kvm = filp->private_data;
 924         void __user *argp = (void __user *)arg;
 925
 926         switch (ioctl) {
 927         case KVM_CREATE_IRQCHIP: {
 928                 return kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
 929         }
 930         case KVM_ARM_SET_DEVICE_ADDR: {
 931                 struct kvm_arm_device_addr dev_addr;
 932
 933                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
 934                         return -EFAULT;
 935                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
 936         }
 937         case KVM_ARM_PREFERRED_TARGET: {
 938                 int err;
 939                 struct kvm_vcpu_init init;
 940
 941                 err = kvm_vcpu_preferred_target(&init);
 942                 if (err)
 943                         return err;
 944
 945                 if (copy_to_user(argp, &init, sizeof(init)))
 946                         return -EFAULT;
 947
 948                 return 0;
 949         }
 950         default:
 951                 return -EINVAL;
 952         }
 953 }
 954
 955 static void cpu_init_hyp_mode(void *dummy)
 956 {
 957         phys_addr_t boot_pgd_ptr;
 958         phys_addr_t pgd_ptr;
 959         unsigned long hyp_stack_ptr;
 960         unsigned long stack_page;
 961         unsigned long vector_ptr;
 962
 963         /* Switch from the HYP stub to our own HYP init vector */
 964         __hyp_set_vectors(kvm_get_idmap_vector());
 965
 966         boot_pgd_ptr = kvm_mmu_get_boot_httbr();
 967         pgd_ptr = kvm_mmu_get_httbr();
 968         stack_page = __this_cpu_read(kvm_arm_hyp_stack_page);
 969         hyp_stack_ptr = stack_page + PAGE_SIZE;
 970         vector_ptr = (unsigned long)__kvm_hyp_vector;
 971
 972         __cpu_init_hyp_mode(boot_pgd_ptr, pgd_ptr, hyp_stack_ptr, vector_ptr);
 973
 974         kvm_arm_init_debug();
 975 }
 976
 977 static int hyp_init_cpu_notify(struct notifier_block *self,
 978                                unsigned long action, void *cpu)
 979 {
 980         switch (action) {
 981         case CPU_STARTING:
 982         case CPU_STARTING_FROZEN:
 983                 if (__hyp_get_vectors() == hyp_default_vectors)
 984                         cpu_init_hyp_mode(NULL);
 985                 break;
 986         }
 987
 988         return NOTIFY_OK;
 989 }
 990
 991 static struct notifier_block hyp_init_cpu_nb = {
 992         .notifier_call = hyp_init_cpu_notify,
 993 };
 994
 995 #ifdef CONFIG_CPU_PM
 996 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
 997                                     unsigned long cmd,
 998                                     void *v)
 999 {
1000         if (cmd == CPU_PM_EXIT &&
1001             __hyp_get_vectors() == hyp_default_vectors) {
1002                 cpu_init_hyp_mode(NULL);
1003                 return NOTIFY_OK;
1004         }
1005
1006         return NOTIFY_DONE;
1007 }
1008
1009 static struct notifier_block hyp_init_cpu_pm_nb = {
1010         .notifier_call = hyp_init_cpu_pm_notifier,
1011 };
1012
1013 static void __init hyp_cpu_pm_init(void)
1014 {
1015         cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1016 }
1017 #else
1018 static inline void hyp_cpu_pm_init(void)
1019 {
1020 }
1021 #endif
1022
1023 /**
1024  * Inits Hyp-mode on all online CPUs
1025  */
1026 static int init_hyp_mode(void)
1027 {
1028         int cpu;
1029         int err = 0;
1030
1031         /*
1032          * Allocate Hyp PGD and setup Hyp identity mapping
1033          */
1034         err = kvm_mmu_init();
1035         if (err)
1036                 goto out_err;
1037
1038         /*
1039          * It is probably enough to obtain the default on one
1040          * CPU. It's unlikely to be different on the others.
1041          */
1042         hyp_default_vectors = __hyp_get_vectors();
1043
1044         /*
1045          * Allocate stack pages for Hypervisor-mode
1046          */
1047         for_each_possible_cpu(cpu) {
1048                 unsigned long stack_page;
1049
1050                 stack_page = __get_free_page(GFP_KERNEL);
1051                 if (!stack_page) {
1052                         err = -ENOMEM;
1053                         goto out_free_stack_pages;
1054                 }
1055
1056                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
1057         }
1058
1059         /*
1060          * Map the Hyp-code called directly from the host
1061          */
1062         err = create_hyp_mappings(__kvm_hyp_code_start, __kvm_hyp_code_end);
1063         if (err) {
1064                 kvm_err("Cannot map world-switch code\n");
1065                 goto out_free_mappings;
1066         }
1067
1068         /*
1069          * Map the Hyp stack pages
1070          */
1071         for_each_possible_cpu(cpu) {
1072                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
1073                 err = create_hyp_mappings(stack_page, stack_page + PAGE_SIZE);
1074
1075                 if (err) {
1076                         kvm_err("Cannot map hyp stack\n");
1077                         goto out_free_mappings;
1078                 }
1079         }
1080
1081         /*
1082          * Map the host CPU structures
1083          */
1084         kvm_host_cpu_state = alloc_percpu(kvm_cpu_context_t);
1085         if (!kvm_host_cpu_state) {
1086                 err = -ENOMEM;
1087                 kvm_err("Cannot allocate host CPU state\n");
1088                 goto out_free_mappings;
1089         }
1090
1091         for_each_possible_cpu(cpu) {
1092                 kvm_cpu_context_t *cpu_ctxt;
1093
1094                 cpu_ctxt = per_cpu_ptr(kvm_host_cpu_state, cpu);
1095                 err = create_hyp_mappings(cpu_ctxt, cpu_ctxt + 1);
1096
1097                 if (err) {
1098                         kvm_err("Cannot map host CPU state: %d\n", err);
1099                         goto out_free_context;
1100                 }
1101         }
1102
1103         /*
1104          * Execute the init code on each CPU.
1105          */
1106         on_each_cpu(cpu_init_hyp_mode, NULL, 1);
1107
1108         /*
1109          * Init HYP view of VGIC
1110          */
1111         err = kvm_vgic_hyp_init();
1112         if (err)
1113                 goto out_free_context;
1114
1115         /*
1116          * Init HYP architected timer support
1117          */
1118         err = kvm_timer_hyp_init();
1119         if (err)
1120                 goto out_free_context;
1121
1122 #ifndef CONFIG_HOTPLUG_CPU
1123         free_boot_hyp_pgd();
1124 #endif
1125
1126         kvm_perf_init();
1127
1128         kvm_info("Hyp mode initialized successfully\n");
1129
1130         return 0;
1131 out_free_context:
1132         free_percpu(kvm_host_cpu_state);
1133 out_free_mappings:
1134         free_hyp_pgds();
1135 out_free_stack_pages:
1136         for_each_possible_cpu(cpu)
1137                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
1138 out_err:
1139         kvm_err("error initializing Hyp mode: %d\n", err);
1140         return err;
1141 }
1142
1143 static void check_kvm_target_cpu(void *ret)
1144 {
1145         *(int *)ret = kvm_target_cpu();
1146 }
1147
1148 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
1149 {
1150         struct kvm_vcpu *vcpu;
1151         int i;
1152
1153         mpidr &= MPIDR_HWID_BITMASK;
1154         kvm_for_each_vcpu(i, vcpu, kvm) {
1155                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
1156                         return vcpu;
1157         }
1158         return NULL;
1159 }
1160
1161 /**
1162  * Initialize Hyp-mode and memory mappings on all CPUs.
1163  */
1164 int kvm_arch_init(void *opaque)
1165 {
1166         int err;
1167         int ret, cpu;
1168
1169         if (!is_hyp_mode_available()) {
1170                 kvm_err("HYP mode not available\n");
1171                 return -ENODEV;
1172         }
1173
1174         for_each_online_cpu(cpu) {
1175                 smp_call_function_single(cpu, check_kvm_target_cpu, &ret, 1);
1176                 if (ret < 0) {
1177                         kvm_err("Error, CPU %d not supported!\n", cpu);
1178                         return -ENODEV;
1179                 }
1180         }
1181
1182         cpu_notifier_register_begin();
1183
1184         err = init_hyp_mode();
1185         if (err)
1186                 goto out_err;
1187
1188         err = __register_cpu_notifier(&hyp_init_cpu_nb);
1189         if (err) {
1190                 kvm_err("Cannot register HYP init CPU notifier (%d)\n", err);
1191                 goto out_err;
1192         }
1193
1194         cpu_notifier_register_done();
1195
1196         hyp_cpu_pm_init();
1197
1198         kvm_coproc_table_init();
1199         return 0;
1200 out_err:
1201         cpu_notifier_register_done();
1202         return err;
1203 }
1204
1205 /* NOP: Compiling as a module not supported */
1206 void kvm_arch_exit(void)
1207 {
1208         kvm_perf_teardown();
1209 }
1210
1211 static int arm_init(void)
1212 {
1213         int rc = kvm_init(NULL, sizeof(struct kvm_vcpu), 0, THIS_MODULE);
1214         return rc;
1215 }
1216
1217 module_init(arm_init);