Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / 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         timer->cntv_ctl |= ARCH_TIMER_CTRL_IT_MASK;
68         ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id,
69                                   timer->irq->irq,
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                 return false;
122
123         cval = timer->cntv_cval;
124         now = kvm_phys_timer_read() - vcpu->kvm->arch.timer.cntvoff;
125
126         return cval <= now;
127 }
128
129 /**
130  * kvm_timer_flush_hwstate - prepare to move the virt timer to the cpu
131  * @vcpu: The vcpu pointer
132  *
133  * Disarm any pending soft timers, since the world-switch code will write the
134  * virtual timer state back to the physical CPU.
135  */
136 void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu)
137 {
138         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
139
140         /*
141          * We're about to run this vcpu again, so there is no need to
142          * keep the background timer running, as we're about to
143          * populate the CPU timer again.
144          */
145         timer_disarm(timer);
146
147         /*
148          * If the timer expired while we were not scheduled, now is the time
149          * to inject it.
150          */
151         if (kvm_timer_should_fire(vcpu))
152                 kvm_timer_inject_irq(vcpu);
153 }
154
155 /**
156  * kvm_timer_sync_hwstate - sync timer state from cpu
157  * @vcpu: The vcpu pointer
158  *
159  * Check if the virtual timer was armed and either schedule a corresponding
160  * soft timer or inject directly if already expired.
161  */
162 void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu)
163 {
164         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
165         cycle_t cval, now;
166         u64 ns;
167
168         BUG_ON(timer_is_armed(timer));
169
170         if (kvm_timer_should_fire(vcpu)) {
171                 /*
172                  * Timer has already expired while we were not
173                  * looking. Inject the interrupt and carry on.
174                  */
175                 kvm_timer_inject_irq(vcpu);
176                 return;
177         }
178
179         cval = timer->cntv_cval;
180         now = kvm_phys_timer_read() - vcpu->kvm->arch.timer.cntvoff;
181
182         ns = cyclecounter_cyc2ns(timecounter->cc, cval - now, timecounter->mask,
183                                  &timecounter->frac);
184         timer_arm(timer, ns);
185 }
186
187 void kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu,
188                           const struct kvm_irq_level *irq)
189 {
190         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
191
192         /*
193          * The vcpu timer irq number cannot be determined in
194          * kvm_timer_vcpu_init() because it is called much before
195          * kvm_vcpu_set_target(). To handle this, we determine
196          * vcpu timer irq number when the vcpu is reset.
197          */
198         timer->irq = irq;
199 }
200
201 void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu)
202 {
203         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
204
205         INIT_WORK(&timer->expired, kvm_timer_inject_irq_work);
206         hrtimer_init(&timer->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
207         timer->timer.function = kvm_timer_expire;
208 }
209
210 static void kvm_timer_init_interrupt(void *info)
211 {
212         enable_percpu_irq(host_vtimer_irq, 0);
213 }
214
215 int kvm_arm_timer_set_reg(struct kvm_vcpu *vcpu, u64 regid, u64 value)
216 {
217         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
218
219         switch (regid) {
220         case KVM_REG_ARM_TIMER_CTL:
221                 timer->cntv_ctl = value;
222                 break;
223         case KVM_REG_ARM_TIMER_CNT:
224                 vcpu->kvm->arch.timer.cntvoff = kvm_phys_timer_read() - value;
225                 break;
226         case KVM_REG_ARM_TIMER_CVAL:
227                 timer->cntv_cval = value;
228                 break;
229         default:
230                 return -1;
231         }
232         return 0;
233 }
234
235 u64 kvm_arm_timer_get_reg(struct kvm_vcpu *vcpu, u64 regid)
236 {
237         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
238
239         switch (regid) {
240         case KVM_REG_ARM_TIMER_CTL:
241                 return timer->cntv_ctl;
242         case KVM_REG_ARM_TIMER_CNT:
243                 return kvm_phys_timer_read() - vcpu->kvm->arch.timer.cntvoff;
244         case KVM_REG_ARM_TIMER_CVAL:
245                 return timer->cntv_cval;
246         }
247         return (u64)-1;
248 }
249
250 static int kvm_timer_cpu_notify(struct notifier_block *self,
251                                 unsigned long action, void *cpu)
252 {
253         switch (action) {
254         case CPU_STARTING:
255         case CPU_STARTING_FROZEN:
256                 kvm_timer_init_interrupt(NULL);
257                 break;
258         case CPU_DYING:
259         case CPU_DYING_FROZEN:
260                 disable_percpu_irq(host_vtimer_irq);
261                 break;
262         }
263
264         return NOTIFY_OK;
265 }
266
267 static struct notifier_block kvm_timer_cpu_nb = {
268         .notifier_call = kvm_timer_cpu_notify,
269 };
270
271 static const struct of_device_id arch_timer_of_match[] = {
272         { .compatible   = "arm,armv7-timer",    },
273         { .compatible   = "arm,armv8-timer",    },
274         {},
275 };
276
277 int kvm_timer_hyp_init(void)
278 {
279         struct device_node *np;
280         unsigned int ppi;
281         int err;
282
283         timecounter = arch_timer_get_timecounter();
284         if (!timecounter)
285                 return -ENODEV;
286
287         np = of_find_matching_node(NULL, arch_timer_of_match);
288         if (!np) {
289                 kvm_err("kvm_arch_timer: can't find DT node\n");
290                 return -ENODEV;
291         }
292
293         ppi = irq_of_parse_and_map(np, 2);
294         if (!ppi) {
295                 kvm_err("kvm_arch_timer: no virtual timer interrupt\n");
296                 err = -EINVAL;
297                 goto out;
298         }
299
300         err = request_percpu_irq(ppi, kvm_arch_timer_handler,
301                                  "kvm guest timer", kvm_get_running_vcpus());
302         if (err) {
303                 kvm_err("kvm_arch_timer: can't request interrupt %d (%d)\n",
304                         ppi, err);
305                 goto out;
306         }
307
308         host_vtimer_irq = ppi;
309
310         err = __register_cpu_notifier(&kvm_timer_cpu_nb);
311         if (err) {
312                 kvm_err("Cannot register timer CPU notifier\n");
313                 goto out_free;
314         }
315
316         wqueue = create_singlethread_workqueue("kvm_arch_timer");
317         if (!wqueue) {
318                 err = -ENOMEM;
319                 goto out_free;
320         }
321
322         kvm_info("%s IRQ%d\n", np->name, ppi);
323         on_each_cpu(kvm_timer_init_interrupt, NULL, 1);
324
325         goto out;
326 out_free:
327         free_percpu_irq(ppi, kvm_get_running_vcpus());
328 out:
329         of_node_put(np);
330         return err;
331 }
332
333 void kvm_timer_vcpu_terminate(struct kvm_vcpu *vcpu)
334 {
335         struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
336
337         timer_disarm(timer);
338 }
339
340 void kvm_timer_enable(struct kvm *kvm)
341 {
342         if (kvm->arch.timer.enabled)
343                 return;
344
345         /*
346          * There is a potential race here between VCPUs starting for the first
347          * time, which may be enabling the timer multiple times.  That doesn't
348          * hurt though, because we're just setting a variable to the same
349          * variable that it already was.  The important thing is that all
350          * VCPUs have the enabled variable set, before entering the guest, if
351          * the arch timers are enabled.
352          */
353         if (timecounter && wqueue)
354                 kvm->arch.timer.enabled = 1;
355 }
356
357 void kvm_timer_init(struct kvm *kvm)
358 {
359         kvm->arch.timer.cntvoff = kvm_phys_timer_read();
360 }