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