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