virt/kvm/arm/arch_timer.c

   1 /*
   2  * Copyright (C) 2012 ARM Ltd.
   3  * Author: Marc Zyngier <marc.zyngier@arm.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, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  17  */
  18
  19 #include <linux/cpu.h>
  20 #include <linux/kvm.h>
  21 #include <linux/kvm_host.h>
  22 #include <linux/interrupt.h>
  23 #include <linux/irq.h>
  24 #include <linux/uaccess.h>
  25
  26 #include <clocksource/arm_arch_timer.h>
  27 #include <asm/arch_timer.h>
  28 #include <asm/kvm_hyp.h>
  29
  30 #include <kvm/arm_vgic.h>
  31 #include <kvm/arm_arch_timer.h>
  32
  33 #include "trace.h"
  34
  35 static struct timecounter *timecounter;
  36 static unsigned int host_vtimer_irq;
  37 static u32 host_vtimer_irq_flags;
  38
  39 static const struct kvm_irq_level default_ptimer_irq = {
  40         .irq    = 30,
  41         .level  = 1,
  42 };
  43
  44 static const struct kvm_irq_level default_vtimer_irq = {
  45         .irq    = 27,
  46         .level  = 1,
  47 };
  48
  49 void kvm_timer_vcpu_put(struct kvm_vcpu *vcpu)
  50 {
  51         vcpu_vtimer(vcpu)->active_cleared_last = false;
  52 }
  53
  54 u64 kvm_phys_timer_read(void)
  55 {
  56         return timecounter->cc->read(timecounter->cc);
  57 }
  58
  59 static bool timer_is_armed(struct arch_timer_cpu *timer)
  60 {
  61         return timer->armed;
  62 }
  63
  64 /* timer_arm: as in "arm the timer", not as in ARM the company */
  65 static void timer_arm(struct arch_timer_cpu *timer, u64 ns)
  66 {
  67         timer->armed = true;
  68         hrtimer_start(&timer->timer, ktime_add_ns(ktime_get(), ns),
  69                       HRTIMER_MODE_ABS);
  70 }
  71
  72 static void timer_disarm(struct arch_timer_cpu *timer)
  73 {
  74         if (timer_is_armed(timer)) {
  75                 hrtimer_cancel(&timer->timer);
  76                 cancel_work_sync(&timer->expired);
  77                 timer->armed = false;
  78         }
  79 }
  80
  81 static irqreturn_t kvm_arch_timer_handler(int irq, void *dev_id)
  82 {
  83         struct kvm_vcpu *vcpu = *(struct kvm_vcpu **)dev_id;
  84
  85         /*
  86          * We disable the timer in the world switch and let it be
  87          * handled by kvm_timer_sync_hwstate(). Getting a timer
  88          * interrupt at this point is a sure sign of some major
  89          * breakage.
  90          */
  91         pr_warn("Unexpected interrupt %d on vcpu %p\n", irq, vcpu);
  92         return IRQ_HANDLED;
  93 }
  94
  95 /*
  96  * Work function for handling the backup timer that we schedule when a vcpu is
  97  * no longer running, but had a timer programmed to fire in the future.
  98  */
  99 static void kvm_timer_inject_irq_work(struct work_struct *work)
 100 {
 101         struct kvm_vcpu *vcpu;
 102
 103         vcpu = container_of(work, struct kvm_vcpu, arch.timer_cpu.expired);
 104
 105         /*
 106          * If the vcpu is blocked we want to wake it up so that it will see
 107          * the timer has expired when entering the guest.
 108          */
 109         kvm_vcpu_wake_up(vcpu);
 110 }
 111
 112 static u64 kvm_timer_compute_delta(struct arch_timer_context *timer_ctx)
 113 {
 114         u64 cval, now;
 115
 116         cval = timer_ctx->cnt_cval;
 117         now = kvm_phys_timer_read() - timer_ctx->cntvoff;
 118
 119         if (now < cval) {
 120                 u64 ns;
 121
 122                 ns = cyclecounter_cyc2ns(timecounter->cc,
 123                                          cval - now,
 124                                          timecounter->mask,
 125                                          &timecounter->frac);
 126                 return ns;
 127         }
 128
 129         return 0;
 130 }
 131
 132 static bool kvm_timer_irq_can_fire(struct arch_timer_context *timer_ctx)
 133 {
 134         return !(timer_ctx->cnt_ctl & ARCH_TIMER_CTRL_IT_MASK) &&
 135                 (timer_ctx->cnt_ctl & ARCH_TIMER_CTRL_ENABLE);
 136 }
 137
 138 /*
 139  * Returns the earliest expiration time in ns among guest timers.
 140  * Note that it will return 0 if none of timers can fire.
 141  */
 142 static u64 kvm_timer_earliest_exp(struct kvm_vcpu *vcpu)
 143 {
 144         u64 min_virt = ULLONG_MAX, min_phys = ULLONG_MAX;
 145         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 146         struct arch_timer_context *ptimer = vcpu_ptimer(vcpu);
 147
 148         if (kvm_timer_irq_can_fire(vtimer))
 149                 min_virt = kvm_timer_compute_delta(vtimer);
 150
 151         if (kvm_timer_irq_can_fire(ptimer))
 152                 min_phys = kvm_timer_compute_delta(ptimer);
 153
 154         /* If none of timers can fire, then return 0 */
 155         if ((min_virt == ULLONG_MAX) && (min_phys == ULLONG_MAX))
 156                 return 0;
 157
 158         return min(min_virt, min_phys);
 159 }
 160
 161 static enum hrtimer_restart kvm_timer_expire(struct hrtimer *hrt)
 162 {
 163         struct arch_timer_cpu *timer;
 164         struct kvm_vcpu *vcpu;
 165         u64 ns;
 166
 167         timer = container_of(hrt, struct arch_timer_cpu, timer);
 168         vcpu = container_of(timer, struct kvm_vcpu, arch.timer_cpu);
 169
 170         /*
 171          * Check that the timer has really expired from the guest's
 172          * PoV (NTP on the host may have forced it to expire
 173          * early). If we should have slept longer, restart it.
 174          */
 175         ns = kvm_timer_earliest_exp(vcpu);
 176         if (unlikely(ns)) {
 177                 hrtimer_forward_now(hrt, ns_to_ktime(ns));
 178                 return HRTIMER_RESTART;
 179         }
 180
 181         schedule_work(&timer->expired);
 182         return HRTIMER_NORESTART;
 183 }
 184
 185 bool kvm_timer_should_fire(struct arch_timer_context *timer_ctx)
 186 {
 187         u64 cval, now;
 188
 189         if (!kvm_timer_irq_can_fire(timer_ctx))
 190                 return false;
 191
 192         cval = timer_ctx->cnt_cval;
 193         now = kvm_phys_timer_read() - timer_ctx->cntvoff;
 194
 195         return cval <= now;
 196 }
 197
 198 /*
 199  * Reflect the timer output level into the kvm_run structure
 200  */
 201 void kvm_timer_update_run(struct kvm_vcpu *vcpu)
 202 {
 203         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 204         struct arch_timer_context *ptimer = vcpu_ptimer(vcpu);
 205         struct kvm_sync_regs *regs = &vcpu->run->s.regs;
 206
 207         /* Populate the device bitmap with the timer states */
 208         regs->device_irq_level &= ~(KVM_ARM_DEV_EL1_VTIMER |
 209                                     KVM_ARM_DEV_EL1_PTIMER);
 210         if (vtimer->irq.level)
 211                 regs->device_irq_level |= KVM_ARM_DEV_EL1_VTIMER;
 212         if (ptimer->irq.level)
 213                 regs->device_irq_level |= KVM_ARM_DEV_EL1_PTIMER;
 214 }
 215
 216 static void kvm_timer_update_irq(struct kvm_vcpu *vcpu, bool new_level,
 217                                  struct arch_timer_context *timer_ctx)
 218 {
 219         int ret;
 220
 221         timer_ctx->active_cleared_last = false;
 222         timer_ctx->irq.level = new_level;
 223         trace_kvm_timer_update_irq(vcpu->vcpu_id, timer_ctx->irq.irq,
 224                                    timer_ctx->irq.level);
 225
 226         if (likely(irqchip_in_kernel(vcpu->kvm))) {
 227                 ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id,
 228                                           timer_ctx->irq.irq,
 229                                           timer_ctx->irq.level,
 230                                           timer_ctx);
 231                 WARN_ON(ret);
 232         }
 233 }
 234
 235 /*
 236  * Check if there was a change in the timer state (should we raise or lower
 237  * the line level to the GIC).
 238  */
 239 static void kvm_timer_update_state(struct kvm_vcpu *vcpu)
 240 {
 241         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 242         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 243         struct arch_timer_context *ptimer = vcpu_ptimer(vcpu);
 244
 245         /*
 246          * If userspace modified the timer registers via SET_ONE_REG before
 247          * the vgic was initialized, we mustn't set the vtimer->irq.level value
 248          * because the guest would never see the interrupt.  Instead wait
 249          * until we call this function from kvm_timer_flush_hwstate.
 250          */
 251         if (unlikely(!timer->enabled))
 252                 return;
 253
 254         if (kvm_timer_should_fire(vtimer) != vtimer->irq.level)
 255                 kvm_timer_update_irq(vcpu, !vtimer->irq.level, vtimer);
 256
 257         if (kvm_timer_should_fire(ptimer) != ptimer->irq.level)
 258                 kvm_timer_update_irq(vcpu, !ptimer->irq.level, ptimer);
 259 }
 260
 261 /* Schedule the background timer for the emulated timer. */
 262 static void kvm_timer_emulate(struct kvm_vcpu *vcpu,
 263                               struct arch_timer_context *timer_ctx)
 264 {
 265         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 266
 267         if (kvm_timer_should_fire(timer_ctx))
 268                 return;
 269
 270         if (!kvm_timer_irq_can_fire(timer_ctx))
 271                 return;
 272
 273         /*  The timer has not yet expired, schedule a background timer */
 274         timer_arm(timer, kvm_timer_compute_delta(timer_ctx));
 275 }
 276
 277 /*
 278  * Schedule the background timer before calling kvm_vcpu_block, so that this
 279  * thread is removed from its waitqueue and made runnable when there's a timer
 280  * interrupt to handle.
 281  */
 282 void kvm_timer_schedule(struct kvm_vcpu *vcpu)
 283 {
 284         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 285         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 286         struct arch_timer_context *ptimer = vcpu_ptimer(vcpu);
 287
 288         BUG_ON(timer_is_armed(timer));
 289
 290         /*
 291          * No need to schedule a background timer if any guest timer has
 292          * already expired, because kvm_vcpu_block will return before putting
 293          * the thread to sleep.
 294          */
 295         if (kvm_timer_should_fire(vtimer) || kvm_timer_should_fire(ptimer))
 296                 return;
 297
 298         /*
 299          * If both timers are not capable of raising interrupts (disabled or
 300          * masked), then there's no more work for us to do.
 301          */
 302         if (!kvm_timer_irq_can_fire(vtimer) && !kvm_timer_irq_can_fire(ptimer))
 303                 return;
 304
 305         /*
 306          * The guest timers have not yet expired, schedule a background timer.
 307          * Set the earliest expiration time among the guest timers.
 308          */
 309         timer_arm(timer, kvm_timer_earliest_exp(vcpu));
 310 }
 311
 312 void kvm_timer_unschedule(struct kvm_vcpu *vcpu)
 313 {
 314         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 315         timer_disarm(timer);
 316 }
 317
 318 static void kvm_timer_flush_hwstate_vgic(struct kvm_vcpu *vcpu)
 319 {
 320         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 321         bool phys_active;
 322         int ret;
 323
 324         /*
 325         * If we enter the guest with the virtual input level to the VGIC
 326         * asserted, then we have already told the VGIC what we need to, and
 327         * we don't need to exit from the guest until the guest deactivates
 328         * the already injected interrupt, so therefore we should set the
 329         * hardware active state to prevent unnecessary exits from the guest.
 330         *
 331         * Also, if we enter the guest with the virtual timer interrupt active,
 332         * then it must be active on the physical distributor, because we set
 333         * the HW bit and the guest must be able to deactivate the virtual and
 334         * physical interrupt at the same time.
 335         *
 336         * Conversely, if the virtual input level is deasserted and the virtual
 337         * interrupt is not active, then always clear the hardware active state
 338         * to ensure that hardware interrupts from the timer triggers a guest
 339         * exit.
 340         */
 341         phys_active = vtimer->irq.level ||
 342                         kvm_vgic_map_is_active(vcpu, vtimer->irq.irq);
 343
 344         /*
 345          * We want to avoid hitting the (re)distributor as much as
 346          * possible, as this is a potentially expensive MMIO access
 347          * (not to mention locks in the irq layer), and a solution for
 348          * this is to cache the "active" state in memory.
 349          *
 350          * Things to consider: we cannot cache an "active set" state,
 351          * because the HW can change this behind our back (it becomes
 352          * "clear" in the HW). We must then restrict the caching to
 353          * the "clear" state.
 354          *
 355          * The cache is invalidated on:
 356          * - vcpu put, indicating that the HW cannot be trusted to be
 357          *   in a sane state on the next vcpu load,
 358          * - any change in the interrupt state
 359          *
 360          * Usage conditions:
 361          * - cached value is "active clear"
 362          * - value to be programmed is "active clear"
 363          */
 364         if (vtimer->active_cleared_last && !phys_active)
 365                 return;
 366
 367         ret = irq_set_irqchip_state(host_vtimer_irq,
 368                                     IRQCHIP_STATE_ACTIVE,
 369                                     phys_active);
 370         WARN_ON(ret);
 371
 372         vtimer->active_cleared_last = !phys_active;
 373 }
 374
 375 bool kvm_timer_should_notify_user(struct kvm_vcpu *vcpu)
 376 {
 377         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 378         struct arch_timer_context *ptimer = vcpu_ptimer(vcpu);
 379         struct kvm_sync_regs *sregs = &vcpu->run->s.regs;
 380         bool vlevel, plevel;
 381
 382         if (likely(irqchip_in_kernel(vcpu->kvm)))
 383                 return false;
 384
 385         vlevel = sregs->device_irq_level & KVM_ARM_DEV_EL1_VTIMER;
 386         plevel = sregs->device_irq_level & KVM_ARM_DEV_EL1_PTIMER;
 387
 388         return vtimer->irq.level != vlevel ||
 389                ptimer->irq.level != plevel;
 390 }
 391
 392 static void kvm_timer_flush_hwstate_user(struct kvm_vcpu *vcpu)
 393 {
 394         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 395
 396         /*
 397          * To prevent continuously exiting from the guest, we mask the
 398          * physical interrupt such that the guest can make forward progress.
 399          * Once we detect the output level being deasserted, we unmask the
 400          * interrupt again so that we exit from the guest when the timer
 401          * fires.
 402         */
 403         if (vtimer->irq.level)
 404                 disable_percpu_irq(host_vtimer_irq);
 405         else
 406                 enable_percpu_irq(host_vtimer_irq, 0);
 407 }
 408
 409 /**
 410  * kvm_timer_flush_hwstate - prepare timers before running the vcpu
 411  * @vcpu: The vcpu pointer
 412  *
 413  * Check if the virtual timer has expired while we were running in the host,
 414  * and inject an interrupt if that was the case, making sure the timer is
 415  * masked or disabled on the host so that we keep executing.  Also schedule a
 416  * software timer for the physical timer if it is enabled.
 417  */
 418 void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu)
 419 {
 420         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 421
 422         if (unlikely(!timer->enabled))
 423                 return;
 424
 425         kvm_timer_update_state(vcpu);
 426
 427         /* Set the background timer for the physical timer emulation. */
 428         kvm_timer_emulate(vcpu, vcpu_ptimer(vcpu));
 429
 430         if (unlikely(!irqchip_in_kernel(vcpu->kvm)))
 431                 kvm_timer_flush_hwstate_user(vcpu);
 432         else
 433                 kvm_timer_flush_hwstate_vgic(vcpu);
 434 }
 435
 436 /**
 437  * kvm_timer_sync_hwstate - sync timer state from cpu
 438  * @vcpu: The vcpu pointer
 439  *
 440  * Check if any of the timers have expired while we were running in the guest,
 441  * and inject an interrupt if that was the case.
 442  */
 443 void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu)
 444 {
 445         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 446
 447         /*
 448          * This is to cancel the background timer for the physical timer
 449          * emulation if it is set.
 450          */
 451         timer_disarm(timer);
 452
 453         /*
 454          * The guest could have modified the timer registers or the timer
 455          * could have expired, update the timer state.
 456          */
 457         kvm_timer_update_state(vcpu);
 458 }
 459
 460 int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu)
 461 {
 462         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 463         struct arch_timer_context *ptimer = vcpu_ptimer(vcpu);
 464
 465         /*
 466          * The bits in CNTV_CTL are architecturally reset to UNKNOWN for ARMv8
 467          * and to 0 for ARMv7.  We provide an implementation that always
 468          * resets the timer to be disabled and unmasked and is compliant with
 469          * the ARMv7 architecture.
 470          */
 471         vtimer->cnt_ctl = 0;
 472         ptimer->cnt_ctl = 0;
 473         kvm_timer_update_state(vcpu);
 474
 475         return 0;
 476 }
 477
 478 /* Make the updates of cntvoff for all vtimer contexts atomic */
 479 static void update_vtimer_cntvoff(struct kvm_vcpu *vcpu, u64 cntvoff)
 480 {
 481         int i;
 482         struct kvm *kvm = vcpu->kvm;
 483         struct kvm_vcpu *tmp;
 484
 485         mutex_lock(&kvm->lock);
 486         kvm_for_each_vcpu(i, tmp, kvm)
 487                 vcpu_vtimer(tmp)->cntvoff = cntvoff;
 488
 489         /*
 490          * When called from the vcpu create path, the CPU being created is not
 491          * included in the loop above, so we just set it here as well.
 492          */
 493         vcpu_vtimer(vcpu)->cntvoff = cntvoff;
 494         mutex_unlock(&kvm->lock);
 495 }
 496
 497 void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu)
 498 {
 499         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 500         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 501         struct arch_timer_context *ptimer = vcpu_ptimer(vcpu);
 502
 503         /* Synchronize cntvoff across all vtimers of a VM. */
 504         update_vtimer_cntvoff(vcpu, kvm_phys_timer_read());
 505         vcpu_ptimer(vcpu)->cntvoff = 0;
 506
 507         INIT_WORK(&timer->expired, kvm_timer_inject_irq_work);
 508         hrtimer_init(&timer->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 509         timer->timer.function = kvm_timer_expire;
 510
 511         vtimer->irq.irq = default_vtimer_irq.irq;
 512         ptimer->irq.irq = default_ptimer_irq.irq;
 513 }
 514
 515 static void kvm_timer_init_interrupt(void *info)
 516 {
 517         enable_percpu_irq(host_vtimer_irq, host_vtimer_irq_flags);
 518 }
 519
 520 int kvm_arm_timer_set_reg(struct kvm_vcpu *vcpu, u64 regid, u64 value)
 521 {
 522         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 523
 524         switch (regid) {
 525         case KVM_REG_ARM_TIMER_CTL:
 526                 vtimer->cnt_ctl = value;
 527                 break;
 528         case KVM_REG_ARM_TIMER_CNT:
 529                 update_vtimer_cntvoff(vcpu, kvm_phys_timer_read() - value);
 530                 break;
 531         case KVM_REG_ARM_TIMER_CVAL:
 532                 vtimer->cnt_cval = value;
 533                 break;
 534         default:
 535                 return -1;
 536         }
 537
 538         kvm_timer_update_state(vcpu);
 539         return 0;
 540 }
 541
 542 u64 kvm_arm_timer_get_reg(struct kvm_vcpu *vcpu, u64 regid)
 543 {
 544         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 545
 546         switch (regid) {
 547         case KVM_REG_ARM_TIMER_CTL:
 548                 return vtimer->cnt_ctl;
 549         case KVM_REG_ARM_TIMER_CNT:
 550                 return kvm_phys_timer_read() - vtimer->cntvoff;
 551         case KVM_REG_ARM_TIMER_CVAL:
 552                 return vtimer->cnt_cval;
 553         }
 554         return (u64)-1;
 555 }
 556
 557 static int kvm_timer_starting_cpu(unsigned int cpu)
 558 {
 559         kvm_timer_init_interrupt(NULL);
 560         return 0;
 561 }
 562
 563 static int kvm_timer_dying_cpu(unsigned int cpu)
 564 {
 565         disable_percpu_irq(host_vtimer_irq);
 566         return 0;
 567 }
 568
 569 int kvm_timer_hyp_init(void)
 570 {
 571         struct arch_timer_kvm_info *info;
 572         int err;
 573
 574         info = arch_timer_get_kvm_info();
 575         timecounter = &info->timecounter;
 576
 577         if (!timecounter->cc) {
 578                 kvm_err("kvm_arch_timer: uninitialized timecounter\n");
 579                 return -ENODEV;
 580         }
 581
 582         if (info->virtual_irq <= 0) {
 583                 kvm_err("kvm_arch_timer: invalid virtual timer IRQ: %d\n",
 584                         info->virtual_irq);
 585                 return -ENODEV;
 586         }
 587         host_vtimer_irq = info->virtual_irq;
 588
 589         host_vtimer_irq_flags = irq_get_trigger_type(host_vtimer_irq);
 590         if (host_vtimer_irq_flags != IRQF_TRIGGER_HIGH &&
 591             host_vtimer_irq_flags != IRQF_TRIGGER_LOW) {
 592                 kvm_err("Invalid trigger for IRQ%d, assuming level low\n",
 593                         host_vtimer_irq);
 594                 host_vtimer_irq_flags = IRQF_TRIGGER_LOW;
 595         }
 596
 597         err = request_percpu_irq(host_vtimer_irq, kvm_arch_timer_handler,
 598                                  "kvm guest timer", kvm_get_running_vcpus());
 599         if (err) {
 600                 kvm_err("kvm_arch_timer: can't request interrupt %d (%d)\n",
 601                         host_vtimer_irq, err);
 602                 return err;
 603         }
 604
 605         kvm_debug("virtual timer IRQ%d\n", host_vtimer_irq);
 606
 607         cpuhp_setup_state(CPUHP_AP_KVM_ARM_TIMER_STARTING,
 608                           "kvm/arm/timer:starting", kvm_timer_starting_cpu,
 609                           kvm_timer_dying_cpu);
 610         return err;
 611 }
 612
 613 void kvm_timer_vcpu_terminate(struct kvm_vcpu *vcpu)
 614 {
 615         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 616         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 617
 618         timer_disarm(timer);
 619         kvm_vgic_unmap_phys_irq(vcpu, vtimer->irq.irq);
 620 }
 621
 622 static bool timer_irqs_are_valid(struct kvm_vcpu *vcpu)
 623 {
 624         int vtimer_irq, ptimer_irq;
 625         int i, ret;
 626
 627         vtimer_irq = vcpu_vtimer(vcpu)->irq.irq;
 628         ret = kvm_vgic_set_owner(vcpu, vtimer_irq, vcpu_vtimer(vcpu));
 629         if (ret)
 630                 return false;
 631
 632         ptimer_irq = vcpu_ptimer(vcpu)->irq.irq;
 633         ret = kvm_vgic_set_owner(vcpu, ptimer_irq, vcpu_ptimer(vcpu));
 634         if (ret)
 635                 return false;
 636
 637         kvm_for_each_vcpu(i, vcpu, vcpu->kvm) {
 638                 if (vcpu_vtimer(vcpu)->irq.irq != vtimer_irq ||
 639                     vcpu_ptimer(vcpu)->irq.irq != ptimer_irq)
 640                         return false;
 641         }
 642
 643         return true;
 644 }
 645
 646 int kvm_timer_enable(struct kvm_vcpu *vcpu)
 647 {
 648         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 649         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 650         struct irq_desc *desc;
 651         struct irq_data *data;
 652         int phys_irq;
 653         int ret;
 654
 655         if (timer->enabled)
 656                 return 0;
 657
 658         /* Without a VGIC we do not map virtual IRQs to physical IRQs */
 659         if (!irqchip_in_kernel(vcpu->kvm))
 660                 goto no_vgic;
 661
 662         if (!vgic_initialized(vcpu->kvm))
 663                 return -ENODEV;
 664
 665         if (!timer_irqs_are_valid(vcpu)) {
 666                 kvm_debug("incorrectly configured timer irqs\n");
 667                 return -EINVAL;
 668         }
 669
 670         /*
 671          * Find the physical IRQ number corresponding to the host_vtimer_irq
 672          */
 673         desc = irq_to_desc(host_vtimer_irq);
 674         if (!desc) {
 675                 kvm_err("%s: no interrupt descriptor\n", __func__);
 676                 return -EINVAL;
 677         }
 678
 679         data = irq_desc_get_irq_data(desc);
 680         while (data->parent_data)
 681                 data = data->parent_data;
 682
 683         phys_irq = data->hwirq;
 684
 685         /*
 686          * Tell the VGIC that the virtual interrupt is tied to a
 687          * physical interrupt. We do that once per VCPU.
 688          */
 689         ret = kvm_vgic_map_phys_irq(vcpu, vtimer->irq.irq, phys_irq);
 690         if (ret)
 691                 return ret;
 692
 693 no_vgic:
 694         timer->enabled = 1;
 695         return 0;
 696 }
 697
 698 /*
 699  * On VHE system, we only need to configure trap on physical timer and counter
 700  * accesses in EL0 and EL1 once, not for every world switch.
 701  * The host kernel runs at EL2 with HCR_EL2.TGE == 1,
 702  * and this makes those bits have no effect for the host kernel execution.
 703  */
 704 void kvm_timer_init_vhe(void)
 705 {
 706         /* When HCR_EL2.E2H ==1, EL1PCEN and EL1PCTEN are shifted by 10 */
 707         u32 cnthctl_shift = 10;
 708         u64 val;
 709
 710         /*
 711          * Disallow physical timer access for the guest.
 712          * Physical counter access is allowed.
 713          */
 714         val = read_sysreg(cnthctl_el2);
 715         val &= ~(CNTHCTL_EL1PCEN << cnthctl_shift);
 716         val |= (CNTHCTL_EL1PCTEN << cnthctl_shift);
 717         write_sysreg(val, cnthctl_el2);
 718 }
 719
 720 static void set_timer_irqs(struct kvm *kvm, int vtimer_irq, int ptimer_irq)
 721 {
 722         struct kvm_vcpu *vcpu;
 723         int i;
 724
 725         kvm_for_each_vcpu(i, vcpu, kvm) {
 726                 vcpu_vtimer(vcpu)->irq.irq = vtimer_irq;
 727                 vcpu_ptimer(vcpu)->irq.irq = ptimer_irq;
 728         }
 729 }
 730
 731 int kvm_arm_timer_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
 732 {
 733         int __user *uaddr = (int __user *)(long)attr->addr;
 734         struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
 735         struct arch_timer_context *ptimer = vcpu_ptimer(vcpu);
 736         int irq;
 737
 738         if (!irqchip_in_kernel(vcpu->kvm))
 739                 return -EINVAL;
 740
 741         if (get_user(irq, uaddr))
 742                 return -EFAULT;
 743
 744         if (!(irq_is_ppi(irq)))
 745                 return -EINVAL;
 746
 747         if (vcpu->arch.timer_cpu.enabled)
 748                 return -EBUSY;
 749
 750         switch (attr->attr) {
 751         case KVM_ARM_VCPU_TIMER_IRQ_VTIMER:
 752                 set_timer_irqs(vcpu->kvm, irq, ptimer->irq.irq);
 753                 break;
 754         case KVM_ARM_VCPU_TIMER_IRQ_PTIMER:
 755                 set_timer_irqs(vcpu->kvm, vtimer->irq.irq, irq);
 756                 break;
 757         default:
 758                 return -ENXIO;
 759         }
 760
 761         return 0;
 762 }
 763
 764 int kvm_arm_timer_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
 765 {
 766         int __user *uaddr = (int __user *)(long)attr->addr;
 767         struct arch_timer_context *timer;
 768         int irq;
 769
 770         switch (attr->attr) {
 771         case KVM_ARM_VCPU_TIMER_IRQ_VTIMER:
 772                 timer = vcpu_vtimer(vcpu);
 773                 break;
 774         case KVM_ARM_VCPU_TIMER_IRQ_PTIMER:
 775                 timer = vcpu_ptimer(vcpu);
 776                 break;
 777         default:
 778                 return -ENXIO;
 779         }
 780
 781         irq = timer->irq.irq;
 782         return put_user(irq, uaddr);
 783 }
 784
 785 int kvm_arm_timer_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
 786 {
 787         switch (attr->attr) {
 788         case KVM_ARM_VCPU_TIMER_IRQ_VTIMER:
 789         case KVM_ARM_VCPU_TIMER_IRQ_PTIMER:
 790                 return 0;
 791         }
 792
 793         return -ENXIO;
 794 }