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
  25 #include <clocksource/arm_arch_timer.h>
  26 #include <asm/arch_timer.h>
  27
  28 #include <kvm/arm_vgic.h>
  29 #include <kvm/arm_arch_timer.h>
  30
  31 #include "trace.h"
  32
  33 static struct timecounter *timecounter;
  34 static struct workqueue_struct *wqueue;
  35 static unsigned int host_vtimer_irq;
  36
  37 void kvm_timer_vcpu_put(struct kvm_vcpu *vcpu)
  38 {
  39         vcpu->arch.timer_cpu.active_cleared_last = false;
  40 }
  41
  42 static cycle_t kvm_phys_timer_read(void)
  43 {
  44         return timecounter->cc->read(timecounter->cc);
  45 }
  46
  47 static bool timer_is_armed(struct arch_timer_cpu *timer)
  48 {
  49         return timer->armed;
  50 }
  51
  52 /* timer_arm: as in "arm the timer", not as in ARM the company */
  53 static void timer_arm(struct arch_timer_cpu *timer, u64 ns)
  54 {
  55         timer->armed = true;
  56         hrtimer_start(&timer->timer, ktime_add_ns(ktime_get(), ns),
  57                       HRTIMER_MODE_ABS);
  58 }
  59
  60 static void timer_disarm(struct arch_timer_cpu *timer)
  61 {
  62         if (timer_is_armed(timer)) {
  63                 hrtimer_cancel(&timer->timer);
  64                 cancel_work_sync(&timer->expired);
  65                 timer->armed = false;
  66         }
  67 }
  68
  69 static irqreturn_t kvm_arch_timer_handler(int irq, void *dev_id)
  70 {
  71         struct kvm_vcpu *vcpu = *(struct kvm_vcpu **)dev_id;
  72
  73         /*
  74          * We disable the timer in the world switch and let it be
  75          * handled by kvm_timer_sync_hwstate(). Getting a timer
  76          * interrupt at this point is a sure sign of some major
  77          * breakage.
  78          */
  79         pr_warn("Unexpected interrupt %d on vcpu %p\n", irq, vcpu);
  80         return IRQ_HANDLED;
  81 }
  82
  83 /*
  84  * Work function for handling the backup timer that we schedule when a vcpu is
  85  * no longer running, but had a timer programmed to fire in the future.
  86  */
  87 static void kvm_timer_inject_irq_work(struct work_struct *work)
  88 {
  89         struct kvm_vcpu *vcpu;
  90
  91         vcpu = container_of(work, struct kvm_vcpu, arch.timer_cpu.expired);
  92         vcpu->arch.timer_cpu.armed = false;
  93
  94         WARN_ON(!kvm_timer_should_fire(vcpu));
  95
  96         /*
  97          * If the vcpu is blocked we want to wake it up so that it will see
  98          * the timer has expired when entering the guest.
  99          */
 100         kvm_vcpu_kick(vcpu);
 101 }
 102
 103 static u64 kvm_timer_compute_delta(struct kvm_vcpu *vcpu)
 104 {
 105         cycle_t cval, now;
 106
 107         cval = vcpu->arch.timer_cpu.cntv_cval;
 108         now = kvm_phys_timer_read() - vcpu->kvm->arch.timer.cntvoff;
 109
 110         if (now < cval) {
 111                 u64 ns;
 112
 113                 ns = cyclecounter_cyc2ns(timecounter->cc,
 114                                          cval - now,
 115                                          timecounter->mask,
 116                                          &timecounter->frac);
 117                 return ns;
 118         }
 119
 120         return 0;
 121 }
 122
 123 static enum hrtimer_restart kvm_timer_expire(struct hrtimer *hrt)
 124 {
 125         struct arch_timer_cpu *timer;
 126         struct kvm_vcpu *vcpu;
 127         u64 ns;
 128
 129         timer = container_of(hrt, struct arch_timer_cpu, timer);
 130         vcpu = container_of(timer, struct kvm_vcpu, arch.timer_cpu);
 131
 132         /*
 133          * Check that the timer has really expired from the guest's
 134          * PoV (NTP on the host may have forced it to expire
 135          * early). If we should have slept longer, restart it.
 136          */
 137         ns = kvm_timer_compute_delta(vcpu);
 138         if (unlikely(ns)) {
 139                 hrtimer_forward_now(hrt, ns_to_ktime(ns));
 140                 return HRTIMER_RESTART;
 141         }
 142
 143         queue_work(wqueue, &timer->expired);
 144         return HRTIMER_NORESTART;
 145 }
 146
 147 static bool kvm_timer_irq_can_fire(struct kvm_vcpu *vcpu)
 148 {
 149         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 150
 151         return !(timer->cntv_ctl & ARCH_TIMER_CTRL_IT_MASK) &&
 152                 (timer->cntv_ctl & ARCH_TIMER_CTRL_ENABLE);
 153 }
 154
 155 bool kvm_timer_should_fire(struct kvm_vcpu *vcpu)
 156 {
 157         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 158         cycle_t cval, now;
 159
 160         if (!kvm_timer_irq_can_fire(vcpu))
 161                 return false;
 162
 163         cval = timer->cntv_cval;
 164         now = kvm_phys_timer_read() - vcpu->kvm->arch.timer.cntvoff;
 165
 166         return cval <= now;
 167 }
 168
 169 static void kvm_timer_update_irq(struct kvm_vcpu *vcpu, bool new_level)
 170 {
 171         int ret;
 172         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 173
 174         BUG_ON(!vgic_initialized(vcpu->kvm));
 175
 176         timer->active_cleared_last = false;
 177         timer->irq.level = new_level;
 178         trace_kvm_timer_update_irq(vcpu->vcpu_id, timer->irq.irq,
 179                                    timer->irq.level);
 180         ret = kvm_vgic_inject_mapped_irq(vcpu->kvm, vcpu->vcpu_id,
 181                                          timer->irq.irq,
 182                                          timer->irq.level);
 183         WARN_ON(ret);
 184 }
 185
 186 /*
 187  * Check if there was a change in the timer state (should we raise or lower
 188  * the line level to the GIC).
 189  */
 190 static int kvm_timer_update_state(struct kvm_vcpu *vcpu)
 191 {
 192         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 193
 194         /*
 195          * If userspace modified the timer registers via SET_ONE_REG before
 196          * the vgic was initialized, we mustn't set the timer->irq.level value
 197          * because the guest would never see the interrupt.  Instead wait
 198          * until we call this function from kvm_timer_flush_hwstate.
 199          */
 200         if (!vgic_initialized(vcpu->kvm) || !timer->enabled)
 201                 return -ENODEV;
 202
 203         if (kvm_timer_should_fire(vcpu) != timer->irq.level)
 204                 kvm_timer_update_irq(vcpu, !timer->irq.level);
 205
 206         return 0;
 207 }
 208
 209 /*
 210  * Schedule the background timer before calling kvm_vcpu_block, so that this
 211  * thread is removed from its waitqueue and made runnable when there's a timer
 212  * interrupt to handle.
 213  */
 214 void kvm_timer_schedule(struct kvm_vcpu *vcpu)
 215 {
 216         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 217
 218         BUG_ON(timer_is_armed(timer));
 219
 220         /*
 221          * No need to schedule a background timer if the guest timer has
 222          * already expired, because kvm_vcpu_block will return before putting
 223          * the thread to sleep.
 224          */
 225         if (kvm_timer_should_fire(vcpu))
 226                 return;
 227
 228         /*
 229          * If the timer is not capable of raising interrupts (disabled or
 230          * masked), then there's no more work for us to do.
 231          */
 232         if (!kvm_timer_irq_can_fire(vcpu))
 233                 return;
 234
 235         /*  The timer has not yet expired, schedule a background timer */
 236         timer_arm(timer, kvm_timer_compute_delta(vcpu));
 237 }
 238
 239 void kvm_timer_unschedule(struct kvm_vcpu *vcpu)
 240 {
 241         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 242         timer_disarm(timer);
 243 }
 244
 245 /**
 246  * kvm_timer_flush_hwstate - prepare to move the virt timer to the cpu
 247  * @vcpu: The vcpu pointer
 248  *
 249  * Check if the virtual timer has expired while we were running in the host,
 250  * and inject an interrupt if that was the case.
 251  */
 252 void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu)
 253 {
 254         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 255         bool phys_active;
 256         int ret;
 257
 258         if (kvm_timer_update_state(vcpu))
 259                 return;
 260
 261         /*
 262         * If we enter the guest with the virtual input level to the VGIC
 263         * asserted, then we have already told the VGIC what we need to, and
 264         * we don't need to exit from the guest until the guest deactivates
 265         * the already injected interrupt, so therefore we should set the
 266         * hardware active state to prevent unnecessary exits from the guest.
 267         *
 268         * Also, if we enter the guest with the virtual timer interrupt active,
 269         * then it must be active on the physical distributor, because we set
 270         * the HW bit and the guest must be able to deactivate the virtual and
 271         * physical interrupt at the same time.
 272         *
 273         * Conversely, if the virtual input level is deasserted and the virtual
 274         * interrupt is not active, then always clear the hardware active state
 275         * to ensure that hardware interrupts from the timer triggers a guest
 276         * exit.
 277         */
 278         phys_active = timer->irq.level ||
 279                         kvm_vgic_map_is_active(vcpu, timer->irq.irq);
 280
 281         /*
 282          * We want to avoid hitting the (re)distributor as much as
 283          * possible, as this is a potentially expensive MMIO access
 284          * (not to mention locks in the irq layer), and a solution for
 285          * this is to cache the "active" state in memory.
 286          *
 287          * Things to consider: we cannot cache an "active set" state,
 288          * because the HW can change this behind our back (it becomes
 289          * "clear" in the HW). We must then restrict the caching to
 290          * the "clear" state.
 291          *
 292          * The cache is invalidated on:
 293          * - vcpu put, indicating that the HW cannot be trusted to be
 294          *   in a sane state on the next vcpu load,
 295          * - any change in the interrupt state
 296          *
 297          * Usage conditions:
 298          * - cached value is "active clear"
 299          * - value to be programmed is "active clear"
 300          */
 301         if (timer->active_cleared_last && !phys_active)
 302                 return;
 303
 304         ret = irq_set_irqchip_state(host_vtimer_irq,
 305                                     IRQCHIP_STATE_ACTIVE,
 306                                     phys_active);
 307         WARN_ON(ret);
 308
 309         timer->active_cleared_last = !phys_active;
 310 }
 311
 312 /**
 313  * kvm_timer_sync_hwstate - sync timer state from cpu
 314  * @vcpu: The vcpu pointer
 315  *
 316  * Check if the virtual timer has expired while we were running in the guest,
 317  * and inject an interrupt if that was the case.
 318  */
 319 void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu)
 320 {
 321         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 322
 323         BUG_ON(timer_is_armed(timer));
 324
 325         /*
 326          * The guest could have modified the timer registers or the timer
 327          * could have expired, update the timer state.
 328          */
 329         kvm_timer_update_state(vcpu);
 330 }
 331
 332 int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu,
 333                          const struct kvm_irq_level *irq)
 334 {
 335         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 336
 337         /*
 338          * The vcpu timer irq number cannot be determined in
 339          * kvm_timer_vcpu_init() because it is called much before
 340          * kvm_vcpu_set_target(). To handle this, we determine
 341          * vcpu timer irq number when the vcpu is reset.
 342          */
 343         timer->irq.irq = irq->irq;
 344
 345         /*
 346          * The bits in CNTV_CTL are architecturally reset to UNKNOWN for ARMv8
 347          * and to 0 for ARMv7.  We provide an implementation that always
 348          * resets the timer to be disabled and unmasked and is compliant with
 349          * the ARMv7 architecture.
 350          */
 351         timer->cntv_ctl = 0;
 352         kvm_timer_update_state(vcpu);
 353
 354         return 0;
 355 }
 356
 357 void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu)
 358 {
 359         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 360
 361         INIT_WORK(&timer->expired, kvm_timer_inject_irq_work);
 362         hrtimer_init(&timer->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 363         timer->timer.function = kvm_timer_expire;
 364 }
 365
 366 static void kvm_timer_init_interrupt(void *info)
 367 {
 368         enable_percpu_irq(host_vtimer_irq, 0);
 369 }
 370
 371 int kvm_arm_timer_set_reg(struct kvm_vcpu *vcpu, u64 regid, u64 value)
 372 {
 373         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 374
 375         switch (regid) {
 376         case KVM_REG_ARM_TIMER_CTL:
 377                 timer->cntv_ctl = value;
 378                 break;
 379         case KVM_REG_ARM_TIMER_CNT:
 380                 vcpu->kvm->arch.timer.cntvoff = kvm_phys_timer_read() - value;
 381                 break;
 382         case KVM_REG_ARM_TIMER_CVAL:
 383                 timer->cntv_cval = value;
 384                 break;
 385         default:
 386                 return -1;
 387         }
 388
 389         kvm_timer_update_state(vcpu);
 390         return 0;
 391 }
 392
 393 u64 kvm_arm_timer_get_reg(struct kvm_vcpu *vcpu, u64 regid)
 394 {
 395         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 396
 397         switch (regid) {
 398         case KVM_REG_ARM_TIMER_CTL:
 399                 return timer->cntv_ctl;
 400         case KVM_REG_ARM_TIMER_CNT:
 401                 return kvm_phys_timer_read() - vcpu->kvm->arch.timer.cntvoff;
 402         case KVM_REG_ARM_TIMER_CVAL:
 403                 return timer->cntv_cval;
 404         }
 405         return (u64)-1;
 406 }
 407
 408 static int kvm_timer_starting_cpu(unsigned int cpu)
 409 {
 410         kvm_timer_init_interrupt(NULL);
 411         return 0;
 412 }
 413
 414 static int kvm_timer_dying_cpu(unsigned int cpu)
 415 {
 416         disable_percpu_irq(host_vtimer_irq);
 417         return 0;
 418 }
 419
 420 int kvm_timer_hyp_init(void)
 421 {
 422         struct arch_timer_kvm_info *info;
 423         int err;
 424
 425         info = arch_timer_get_kvm_info();
 426         timecounter = &info->timecounter;
 427
 428         if (info->virtual_irq <= 0) {
 429                 kvm_err("kvm_arch_timer: invalid virtual timer IRQ: %d\n",
 430                         info->virtual_irq);
 431                 return -ENODEV;
 432         }
 433         host_vtimer_irq = info->virtual_irq;
 434
 435         err = request_percpu_irq(host_vtimer_irq, kvm_arch_timer_handler,
 436                                  "kvm guest timer", kvm_get_running_vcpus());
 437         if (err) {
 438                 kvm_err("kvm_arch_timer: can't request interrupt %d (%d)\n",
 439                         host_vtimer_irq, err);
 440                 goto out;
 441         }
 442
 443         wqueue = create_singlethread_workqueue("kvm_arch_timer");
 444         if (!wqueue) {
 445                 err = -ENOMEM;
 446                 goto out_free;
 447         }
 448
 449         kvm_info("virtual timer IRQ%d\n", host_vtimer_irq);
 450
 451         cpuhp_setup_state(CPUHP_AP_KVM_ARM_TIMER_STARTING,
 452                           "AP_KVM_ARM_TIMER_STARTING", kvm_timer_starting_cpu,
 453                           kvm_timer_dying_cpu);
 454         goto out;
 455 out_free:
 456         free_percpu_irq(host_vtimer_irq, kvm_get_running_vcpus());
 457 out:
 458         return err;
 459 }
 460
 461 void kvm_timer_vcpu_terminate(struct kvm_vcpu *vcpu)
 462 {
 463         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 464
 465         timer_disarm(timer);
 466         kvm_vgic_unmap_phys_irq(vcpu, timer->irq.irq);
 467 }
 468
 469 int kvm_timer_enable(struct kvm_vcpu *vcpu)
 470 {
 471         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 472         struct irq_desc *desc;
 473         struct irq_data *data;
 474         int phys_irq;
 475         int ret;
 476
 477         if (timer->enabled)
 478                 return 0;
 479
 480         /*
 481          * Find the physical IRQ number corresponding to the host_vtimer_irq
 482          */
 483         desc = irq_to_desc(host_vtimer_irq);
 484         if (!desc) {
 485                 kvm_err("%s: no interrupt descriptor\n", __func__);
 486                 return -EINVAL;
 487         }
 488
 489         data = irq_desc_get_irq_data(desc);
 490         while (data->parent_data)
 491                 data = data->parent_data;
 492
 493         phys_irq = data->hwirq;
 494
 495         /*
 496          * Tell the VGIC that the virtual interrupt is tied to a
 497          * physical interrupt. We do that once per VCPU.
 498          */
 499         ret = kvm_vgic_map_phys_irq(vcpu, timer->irq.irq, phys_irq);
 500         if (ret)
 501                 return ret;
 502
 503
 504         /*
 505          * There is a potential race here between VCPUs starting for the first
 506          * time, which may be enabling the timer multiple times.  That doesn't
 507          * hurt though, because we're just setting a variable to the same
 508          * variable that it already was.  The important thing is that all
 509          * VCPUs have the enabled variable set, before entering the guest, if
 510          * the arch timers are enabled.
 511          */
 512         if (timecounter && wqueue)
 513                 timer->enabled = 1;
 514
 515         return 0;
 516 }
 517
 518 void kvm_timer_init(struct kvm *kvm)
 519 {
 520         kvm->arch.timer.cntvoff = kvm_phys_timer_read();
 521 }