2 * Xen time implementation.
4 * This is implemented in terms of a clocksource driver which uses
5 * the hypervisor clock as a nanosecond timebase, and a clockevent
6 * driver which uses the hypervisor's timer mechanism.
8 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
10 #include <linux/kernel.h>
11 #include <linux/interrupt.h>
12 #include <linux/clocksource.h>
13 #include <linux/clockchips.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/math64.h>
17 #include <asm/pvclock.h>
18 #include <asm/xen/hypervisor.h>
19 #include <asm/xen/hypercall.h>
21 #include <xen/events.h>
22 #include <xen/interface/xen.h>
23 #include <xen/interface/vcpu.h>
29 /* Xen may fire a timer up to this many ns early */
30 #define TIMER_SLOP 100000
31 #define NS_PER_TICK (1000000000LL / HZ)
33 /* runstate info updated by Xen */
34 static DEFINE_PER_CPU(struct vcpu_runstate_info, runstate);
36 /* snapshots of runstate info */
37 static DEFINE_PER_CPU(struct vcpu_runstate_info, runstate_snapshot);
39 /* unused ns of stolen and blocked time */
40 static DEFINE_PER_CPU(u64, residual_stolen);
41 static DEFINE_PER_CPU(u64, residual_blocked);
43 /* return an consistent snapshot of 64-bit time/counter value */
44 static u64 get64(const u64 *p)
48 if (BITS_PER_LONG < 64) {
53 * Read high then low, and then make sure high is
54 * still the same; this will only loop if low wraps
55 * and carries into high.
56 * XXX some clean way to make this endian-proof?
63 } while (p32[1] != h);
65 ret = (((u64)h) << 32) | l;
75 static void get_runstate_snapshot(struct vcpu_runstate_info *res)
78 struct vcpu_runstate_info *state;
80 BUG_ON(preemptible());
82 state = &__get_cpu_var(runstate);
85 * The runstate info is always updated by the hypervisor on
86 * the current CPU, so there's no need to use anything
87 * stronger than a compiler barrier when fetching it.
90 state_time = get64(&state->state_entry_time);
94 } while (get64(&state->state_entry_time) != state_time);
97 /* return true when a vcpu could run but has no real cpu to run on */
98 bool xen_vcpu_stolen(int vcpu)
100 return per_cpu(runstate, vcpu).state == RUNSTATE_runnable;
103 static void setup_runstate_info(int cpu)
105 struct vcpu_register_runstate_memory_area area;
107 area.addr.v = &per_cpu(runstate, cpu);
109 if (HYPERVISOR_vcpu_op(VCPUOP_register_runstate_memory_area,
114 static void do_stolen_accounting(void)
116 struct vcpu_runstate_info state;
117 struct vcpu_runstate_info *snap;
118 s64 blocked, runnable, offline, stolen;
121 get_runstate_snapshot(&state);
123 WARN_ON(state.state != RUNSTATE_running);
125 snap = &__get_cpu_var(runstate_snapshot);
127 /* work out how much time the VCPU has not been runn*ing* */
128 blocked = state.time[RUNSTATE_blocked] - snap->time[RUNSTATE_blocked];
129 runnable = state.time[RUNSTATE_runnable] - snap->time[RUNSTATE_runnable];
130 offline = state.time[RUNSTATE_offline] - snap->time[RUNSTATE_offline];
134 /* Add the appropriate number of ticks of stolen time,
135 including any left-overs from last time. Passing NULL to
136 account_steal_time accounts the time as stolen. */
137 stolen = runnable + offline + __get_cpu_var(residual_stolen);
142 ticks = iter_div_u64_rem(stolen, NS_PER_TICK, &stolen);
143 __get_cpu_var(residual_stolen) = stolen;
144 account_steal_time(NULL, ticks);
146 /* Add the appropriate number of ticks of blocked time,
147 including any left-overs from last time. Passing idle to
148 account_steal_time accounts the time as idle/wait. */
149 blocked += __get_cpu_var(residual_blocked);
154 ticks = iter_div_u64_rem(blocked, NS_PER_TICK, &blocked);
155 __get_cpu_var(residual_blocked) = blocked;
156 account_steal_time(idle_task(smp_processor_id()), ticks);
160 * Xen sched_clock implementation. Returns the number of unstolen
161 * nanoseconds, which is nanoseconds the VCPU spent in RUNNING+BLOCKED
164 unsigned long long xen_sched_clock(void)
166 struct vcpu_runstate_info state;
172 * Ideally sched_clock should be called on a per-cpu basis
173 * anyway, so preempt should already be disabled, but that's
174 * not current practice at the moment.
178 now = xen_clocksource_read();
180 get_runstate_snapshot(&state);
182 WARN_ON(state.state != RUNSTATE_running);
184 offset = now - state.state_entry_time;
188 ret = state.time[RUNSTATE_blocked] +
189 state.time[RUNSTATE_running] +
198 /* Get the TSC speed from Xen */
199 unsigned long xen_tsc_khz(void)
201 struct pvclock_vcpu_time_info *info =
202 &HYPERVISOR_shared_info->vcpu_info[0].time;
204 return pvclock_tsc_khz(info);
207 cycle_t xen_clocksource_read(void)
209 struct pvclock_vcpu_time_info *src;
212 src = &get_cpu_var(xen_vcpu)->time;
213 ret = pvclock_clocksource_read(src);
214 put_cpu_var(xen_vcpu);
218 static void xen_read_wallclock(struct timespec *ts)
220 struct shared_info *s = HYPERVISOR_shared_info;
221 struct pvclock_wall_clock *wall_clock = &(s->wc);
222 struct pvclock_vcpu_time_info *vcpu_time;
224 vcpu_time = &get_cpu_var(xen_vcpu)->time;
225 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
226 put_cpu_var(xen_vcpu);
229 unsigned long xen_get_wallclock(void)
233 xen_read_wallclock(&ts);
237 int xen_set_wallclock(unsigned long now)
239 /* do nothing for domU */
243 static struct clocksource xen_clocksource __read_mostly = {
246 .read = xen_clocksource_read,
248 .mult = 1<<XEN_SHIFT, /* time directly in nanoseconds */
250 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
254 Xen clockevent implementation
256 Xen has two clockevent implementations:
258 The old timer_op one works with all released versions of Xen prior
259 to version 3.0.4. This version of the hypervisor provides a
260 single-shot timer with nanosecond resolution. However, sharing the
261 same event channel is a 100Hz tick which is delivered while the
262 vcpu is running. We don't care about or use this tick, but it will
263 cause the core time code to think the timer fired too soon, and
264 will end up resetting it each time. It could be filtered, but
265 doing so has complications when the ktime clocksource is not yet
266 the xen clocksource (ie, at boot time).
268 The new vcpu_op-based timer interface allows the tick timer period
269 to be changed or turned off. The tick timer is not useful as a
270 periodic timer because events are only delivered to running vcpus.
271 The one-shot timer can report when a timeout is in the past, so
272 set_next_event is capable of returning -ETIME when appropriate.
273 This interface is used when available.
278 Get a hypervisor absolute time. In theory we could maintain an
279 offset between the kernel's time and the hypervisor's time, and
280 apply that to a kernel's absolute timeout. Unfortunately the
281 hypervisor and kernel times can drift even if the kernel is using
282 the Xen clocksource, because ntp can warp the kernel's clocksource.
284 static s64 get_abs_timeout(unsigned long delta)
286 return xen_clocksource_read() + delta;
289 static void xen_timerop_set_mode(enum clock_event_mode mode,
290 struct clock_event_device *evt)
293 case CLOCK_EVT_MODE_PERIODIC:
298 case CLOCK_EVT_MODE_ONESHOT:
299 case CLOCK_EVT_MODE_RESUME:
302 case CLOCK_EVT_MODE_UNUSED:
303 case CLOCK_EVT_MODE_SHUTDOWN:
304 HYPERVISOR_set_timer_op(0); /* cancel timeout */
309 static int xen_timerop_set_next_event(unsigned long delta,
310 struct clock_event_device *evt)
312 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
314 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
317 /* We may have missed the deadline, but there's no real way of
318 knowing for sure. If the event was in the past, then we'll
319 get an immediate interrupt. */
324 static const struct clock_event_device xen_timerop_clockevent = {
326 .features = CLOCK_EVT_FEAT_ONESHOT,
328 .max_delta_ns = 0xffffffff,
329 .min_delta_ns = TIMER_SLOP,
335 .set_mode = xen_timerop_set_mode,
336 .set_next_event = xen_timerop_set_next_event,
341 static void xen_vcpuop_set_mode(enum clock_event_mode mode,
342 struct clock_event_device *evt)
344 int cpu = smp_processor_id();
347 case CLOCK_EVT_MODE_PERIODIC:
348 WARN_ON(1); /* unsupported */
351 case CLOCK_EVT_MODE_ONESHOT:
352 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
356 case CLOCK_EVT_MODE_UNUSED:
357 case CLOCK_EVT_MODE_SHUTDOWN:
358 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, cpu, NULL) ||
359 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
362 case CLOCK_EVT_MODE_RESUME:
367 static int xen_vcpuop_set_next_event(unsigned long delta,
368 struct clock_event_device *evt)
370 int cpu = smp_processor_id();
371 struct vcpu_set_singleshot_timer single;
374 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
376 single.timeout_abs_ns = get_abs_timeout(delta);
377 single.flags = VCPU_SSHOTTMR_future;
379 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single);
381 BUG_ON(ret != 0 && ret != -ETIME);
386 static const struct clock_event_device xen_vcpuop_clockevent = {
388 .features = CLOCK_EVT_FEAT_ONESHOT,
390 .max_delta_ns = 0xffffffff,
391 .min_delta_ns = TIMER_SLOP,
397 .set_mode = xen_vcpuop_set_mode,
398 .set_next_event = xen_vcpuop_set_next_event,
401 static const struct clock_event_device *xen_clockevent =
402 &xen_timerop_clockevent;
403 static DEFINE_PER_CPU(struct clock_event_device, xen_clock_events);
405 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
407 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events);
411 if (evt->event_handler) {
412 evt->event_handler(evt);
416 do_stolen_accounting();
421 void xen_setup_timer(int cpu)
424 struct clock_event_device *evt;
427 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
429 name = kasprintf(GFP_KERNEL, "timer%d", cpu);
431 name = "<timer kasprintf failed>";
433 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
434 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
437 evt = &per_cpu(xen_clock_events, cpu);
438 memcpy(evt, xen_clockevent, sizeof(*evt));
440 evt->cpumask = cpumask_of(cpu);
443 setup_runstate_info(cpu);
446 void xen_teardown_timer(int cpu)
448 struct clock_event_device *evt;
450 evt = &per_cpu(xen_clock_events, cpu);
451 unbind_from_irqhandler(evt->irq, NULL);
454 void xen_setup_cpu_clockevents(void)
456 BUG_ON(preemptible());
458 clockevents_register_device(&__get_cpu_var(xen_clock_events));
461 void xen_timer_resume(void)
465 if (xen_clockevent != &xen_vcpuop_clockevent)
468 for_each_online_cpu(cpu) {
469 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
474 __init void xen_time_init(void)
476 int cpu = smp_processor_id();
478 clocksource_register(&xen_clocksource);
480 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) {
481 /* Successfully turned off 100Hz tick, so we have the
482 vcpuop-based timer interface */
483 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
484 xen_clockevent = &xen_vcpuop_clockevent;
487 /* Set initial system time with full resolution */
488 xen_read_wallclock(&xtime);
489 set_normalized_timespec(&wall_to_monotonic,
490 -xtime.tv_sec, -xtime.tv_nsec);
492 setup_force_cpu_cap(X86_FEATURE_TSC);
494 xen_setup_timer(cpu);
495 xen_setup_cpu_clockevents();