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/of_irq.h>
  21 #include <linux/kvm.h>
  22 #include <linux/kvm_host.h>
  23 #include <linux/interrupt.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 static struct timecounter *timecounter;
  32 static struct workqueue_struct *wqueue;
  33 static unsigned int host_vtimer_irq;
  34
  35 static cycle_t kvm_phys_timer_read(void)
  36 {
  37         return timecounter->cc->read(timecounter->cc);
  38 }
  39
  40 static bool timer_is_armed(struct arch_timer_cpu *timer)
  41 {
  42         return timer->armed;
  43 }
  44
  45 /* timer_arm: as in "arm the timer", not as in ARM the company */
  46 static void timer_arm(struct arch_timer_cpu *timer, u64 ns)
  47 {
  48         timer->armed = true;
  49         hrtimer_start(&timer->timer, ktime_add_ns(ktime_get(), ns),
  50                       HRTIMER_MODE_ABS);
  51 }
  52
  53 static void timer_disarm(struct arch_timer_cpu *timer)
  54 {
  55         if (timer_is_armed(timer)) {
  56                 hrtimer_cancel(&timer->timer);
  57                 cancel_work_sync(&timer->expired);
  58                 timer->armed = false;
  59         }
  60 }
  61
  62 static void kvm_timer_inject_irq(struct kvm_vcpu *vcpu)
  63 {
  64         int ret;
  65         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
  66
  67         kvm_vgic_set_phys_irq_active(timer->map, true);
  68         ret = kvm_vgic_inject_mapped_irq(vcpu->kvm, vcpu->vcpu_id,
  69                                          timer->map,
  70                                          timer->irq->level);
  71         WARN_ON(ret);
  72 }
  73
  74 static irqreturn_t kvm_arch_timer_handler(int irq, void *dev_id)
  75 {
  76         struct kvm_vcpu *vcpu = *(struct kvm_vcpu **)dev_id;
  77
  78         /*
  79          * We disable the timer in the world switch and let it be
  80          * handled by kvm_timer_sync_hwstate(). Getting a timer
  81          * interrupt at this point is a sure sign of some major
  82          * breakage.
  83          */
  84         pr_warn("Unexpected interrupt %d on vcpu %p\n", irq, vcpu);
  85         return IRQ_HANDLED;
  86 }
  87
  88 /*
  89  * Work function for handling the backup timer that we schedule when a vcpu is
  90  * no longer running, but had a timer programmed to fire in the future.
  91  */
  92 static void kvm_timer_inject_irq_work(struct work_struct *work)
  93 {
  94         struct kvm_vcpu *vcpu;
  95
  96         vcpu = container_of(work, struct kvm_vcpu, arch.timer_cpu.expired);
  97         vcpu->arch.timer_cpu.armed = false;
  98
  99         /*
 100          * If the vcpu is blocked we want to wake it up so that it will see
 101          * the timer has expired when entering the guest.
 102          */
 103         kvm_vcpu_kick(vcpu);
 104 }
 105
 106 static enum hrtimer_restart kvm_timer_expire(struct hrtimer *hrt)
 107 {
 108         struct arch_timer_cpu *timer;
 109         timer = container_of(hrt, struct arch_timer_cpu, timer);
 110         queue_work(wqueue, &timer->expired);
 111         return HRTIMER_NORESTART;
 112 }
 113
 114 bool kvm_timer_should_fire(struct kvm_vcpu *vcpu)
 115 {
 116         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 117         cycle_t cval, now;
 118
 119         if ((timer->cntv_ctl & ARCH_TIMER_CTRL_IT_MASK) ||
 120             !(timer->cntv_ctl & ARCH_TIMER_CTRL_ENABLE) ||
 121             kvm_vgic_get_phys_irq_active(timer->map))
 122                 return false;
 123
 124         cval = timer->cntv_cval;
 125         now = kvm_phys_timer_read() - vcpu->kvm->arch.timer.cntvoff;
 126
 127         return cval <= now;
 128 }
 129
 130 /**
 131  * kvm_timer_flush_hwstate - prepare to move the virt timer to the cpu
 132  * @vcpu: The vcpu pointer
 133  *
 134  * Disarm any pending soft timers, since the world-switch code will write the
 135  * virtual timer state back to the physical CPU.
 136  */
 137 void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu)
 138 {
 139         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 140
 141         /*
 142          * We're about to run this vcpu again, so there is no need to
 143          * keep the background timer running, as we're about to
 144          * populate the CPU timer again.
 145          */
 146         timer_disarm(timer);
 147
 148         /*
 149          * If the timer expired while we were not scheduled, now is the time
 150          * to inject it.
 151          */
 152         if (kvm_timer_should_fire(vcpu))
 153                 kvm_timer_inject_irq(vcpu);
 154 }
 155
 156 /**
 157  * kvm_timer_sync_hwstate - sync timer state from cpu
 158  * @vcpu: The vcpu pointer
 159  *
 160  * Check if the virtual timer was armed and either schedule a corresponding
 161  * soft timer or inject directly if already expired.
 162  */
 163 void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu)
 164 {
 165         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 166         cycle_t cval, now;
 167         u64 ns;
 168
 169         BUG_ON(timer_is_armed(timer));
 170
 171         if (kvm_timer_should_fire(vcpu)) {
 172                 /*
 173                  * Timer has already expired while we were not
 174                  * looking. Inject the interrupt and carry on.
 175                  */
 176                 kvm_timer_inject_irq(vcpu);
 177                 return;
 178         }
 179
 180         cval = timer->cntv_cval;
 181         now = kvm_phys_timer_read() - vcpu->kvm->arch.timer.cntvoff;
 182
 183         ns = cyclecounter_cyc2ns(timecounter->cc, cval - now, timecounter->mask,
 184                                  &timecounter->frac);
 185         timer_arm(timer, ns);
 186 }
 187
 188 int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu,
 189                          const struct kvm_irq_level *irq)
 190 {
 191         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 192         struct irq_phys_map *map;
 193
 194         /*
 195          * The vcpu timer irq number cannot be determined in
 196          * kvm_timer_vcpu_init() because it is called much before
 197          * kvm_vcpu_set_target(). To handle this, we determine
 198          * vcpu timer irq number when the vcpu is reset.
 199          */
 200         timer->irq = irq;
 201
 202         /*
 203          * Tell the VGIC that the virtual interrupt is tied to a
 204          * physical interrupt. We do that once per VCPU.
 205          */
 206         map = kvm_vgic_map_phys_irq(vcpu, irq->irq, host_vtimer_irq);
 207         if (WARN_ON(IS_ERR(map)))
 208                 return PTR_ERR(map);
 209
 210         timer->map = map;
 211         return 0;
 212 }
 213
 214 void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu)
 215 {
 216         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 217
 218         INIT_WORK(&timer->expired, kvm_timer_inject_irq_work);
 219         hrtimer_init(&timer->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 220         timer->timer.function = kvm_timer_expire;
 221 }
 222
 223 static void kvm_timer_init_interrupt(void *info)
 224 {
 225         enable_percpu_irq(host_vtimer_irq, 0);
 226 }
 227
 228 int kvm_arm_timer_set_reg(struct kvm_vcpu *vcpu, u64 regid, u64 value)
 229 {
 230         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 231
 232         switch (regid) {
 233         case KVM_REG_ARM_TIMER_CTL:
 234                 timer->cntv_ctl = value;
 235                 break;
 236         case KVM_REG_ARM_TIMER_CNT:
 237                 vcpu->kvm->arch.timer.cntvoff = kvm_phys_timer_read() - value;
 238                 break;
 239         case KVM_REG_ARM_TIMER_CVAL:
 240                 timer->cntv_cval = value;
 241                 break;
 242         default:
 243                 return -1;
 244         }
 245         return 0;
 246 }
 247
 248 u64 kvm_arm_timer_get_reg(struct kvm_vcpu *vcpu, u64 regid)
 249 {
 250         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 251
 252         switch (regid) {
 253         case KVM_REG_ARM_TIMER_CTL:
 254                 return timer->cntv_ctl;
 255         case KVM_REG_ARM_TIMER_CNT:
 256                 return kvm_phys_timer_read() - vcpu->kvm->arch.timer.cntvoff;
 257         case KVM_REG_ARM_TIMER_CVAL:
 258                 return timer->cntv_cval;
 259         }
 260         return (u64)-1;
 261 }
 262
 263 static int kvm_timer_cpu_notify(struct notifier_block *self,
 264                                 unsigned long action, void *cpu)
 265 {
 266         switch (action) {
 267         case CPU_STARTING:
 268         case CPU_STARTING_FROZEN:
 269                 kvm_timer_init_interrupt(NULL);
 270                 break;
 271         case CPU_DYING:
 272         case CPU_DYING_FROZEN:
 273                 disable_percpu_irq(host_vtimer_irq);
 274                 break;
 275         }
 276
 277         return NOTIFY_OK;
 278 }
 279
 280 static struct notifier_block kvm_timer_cpu_nb = {
 281         .notifier_call = kvm_timer_cpu_notify,
 282 };
 283
 284 static const struct of_device_id arch_timer_of_match[] = {
 285         { .compatible   = "arm,armv7-timer",    },
 286         { .compatible   = "arm,armv8-timer",    },
 287         {},
 288 };
 289
 290 int kvm_timer_hyp_init(void)
 291 {
 292         struct device_node *np;
 293         unsigned int ppi;
 294         int err;
 295
 296         timecounter = arch_timer_get_timecounter();
 297         if (!timecounter)
 298                 return -ENODEV;
 299
 300         np = of_find_matching_node(NULL, arch_timer_of_match);
 301         if (!np) {
 302                 kvm_err("kvm_arch_timer: can't find DT node\n");
 303                 return -ENODEV;
 304         }
 305
 306         ppi = irq_of_parse_and_map(np, 2);
 307         if (!ppi) {
 308                 kvm_err("kvm_arch_timer: no virtual timer interrupt\n");
 309                 err = -EINVAL;
 310                 goto out;
 311         }
 312
 313         err = request_percpu_irq(ppi, kvm_arch_timer_handler,
 314                                  "kvm guest timer", kvm_get_running_vcpus());
 315         if (err) {
 316                 kvm_err("kvm_arch_timer: can't request interrupt %d (%d)\n",
 317                         ppi, err);
 318                 goto out;
 319         }
 320
 321         host_vtimer_irq = ppi;
 322
 323         err = __register_cpu_notifier(&kvm_timer_cpu_nb);
 324         if (err) {
 325                 kvm_err("Cannot register timer CPU notifier\n");
 326                 goto out_free;
 327         }
 328
 329         wqueue = create_singlethread_workqueue("kvm_arch_timer");
 330         if (!wqueue) {
 331                 err = -ENOMEM;
 332                 goto out_free;
 333         }
 334
 335         kvm_info("%s IRQ%d\n", np->name, ppi);
 336         on_each_cpu(kvm_timer_init_interrupt, NULL, 1);
 337
 338         goto out;
 339 out_free:
 340         free_percpu_irq(ppi, kvm_get_running_vcpus());
 341 out:
 342         of_node_put(np);
 343         return err;
 344 }
 345
 346 void kvm_timer_vcpu_terminate(struct kvm_vcpu *vcpu)
 347 {
 348         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 349
 350         timer_disarm(timer);
 351         if (timer->map)
 352                 kvm_vgic_unmap_phys_irq(vcpu, timer->map);
 353 }
 354
 355 void kvm_timer_enable(struct kvm *kvm)
 356 {
 357         if (kvm->arch.timer.enabled)
 358                 return;
 359
 360         /*
 361          * There is a potential race here between VCPUs starting for the first
 362          * time, which may be enabling the timer multiple times.  That doesn't
 363          * hurt though, because we're just setting a variable to the same
 364          * variable that it already was.  The important thing is that all
 365          * VCPUs have the enabled variable set, before entering the guest, if
 366          * the arch timers are enabled.
 367          */
 368         if (timecounter && wqueue)
 369                 kvm->arch.timer.enabled = 1;
 370 }
 371
 372 void kvm_timer_init(struct kvm *kvm)
 373 {
 374         kvm->arch.timer.cntvoff = kvm_phys_timer_read();
 375 }