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