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>
16 #include <linux/gfp.h>
18 #include <asm/pvclock.h>
19 #include <asm/xen/hypervisor.h>
20 #include <asm/xen/hypercall.h>
22 #include <xen/events.h>
23 #include <xen/interface/xen.h>
24 #include <xen/interface/vcpu.h>
30 /* Xen may fire a timer up to this many ns early */
31 #define TIMER_SLOP 100000
32 #define NS_PER_TICK (1000000000LL / HZ)
34 /* runstate info updated by Xen */
35 static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate);
37 /* snapshots of runstate info */
38 static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate_snapshot);
40 /* unused ns of stolen and blocked time */
41 static DEFINE_PER_CPU(u64, xen_residual_stolen);
42 static DEFINE_PER_CPU(u64, xen_residual_blocked);
44 /* return an consistent snapshot of 64-bit time/counter value */
45 static u64 get64(const u64 *p)
49 if (BITS_PER_LONG < 64) {
54 * Read high then low, and then make sure high is
55 * still the same; this will only loop if low wraps
56 * and carries into high.
57 * XXX some clean way to make this endian-proof?
64 } while (p32[1] != h);
66 ret = (((u64)h) << 32) | l;
76 static void get_runstate_snapshot(struct vcpu_runstate_info *res)
79 struct vcpu_runstate_info *state;
81 BUG_ON(preemptible());
83 state = &__get_cpu_var(xen_runstate);
86 * The runstate info is always updated by the hypervisor on
87 * the current CPU, so there's no need to use anything
88 * stronger than a compiler barrier when fetching it.
91 state_time = get64(&state->state_entry_time);
95 } while (get64(&state->state_entry_time) != state_time);
98 /* return true when a vcpu could run but has no real cpu to run on */
99 bool xen_vcpu_stolen(int vcpu)
101 return per_cpu(xen_runstate, vcpu).state == RUNSTATE_runnable;
104 void xen_setup_runstate_info(int cpu)
106 struct vcpu_register_runstate_memory_area area;
108 area.addr.v = &per_cpu(xen_runstate, cpu);
110 if (HYPERVISOR_vcpu_op(VCPUOP_register_runstate_memory_area,
115 static void do_stolen_accounting(void)
117 struct vcpu_runstate_info state;
118 struct vcpu_runstate_info *snap;
119 s64 blocked, runnable, offline, stolen;
122 get_runstate_snapshot(&state);
124 WARN_ON(state.state != RUNSTATE_running);
126 snap = &__get_cpu_var(xen_runstate_snapshot);
128 /* work out how much time the VCPU has not been runn*ing* */
129 blocked = state.time[RUNSTATE_blocked] - snap->time[RUNSTATE_blocked];
130 runnable = state.time[RUNSTATE_runnable] - snap->time[RUNSTATE_runnable];
131 offline = state.time[RUNSTATE_offline] - snap->time[RUNSTATE_offline];
135 /* Add the appropriate number of ticks of stolen time,
136 including any left-overs from last time. */
137 stolen = runnable + offline + __get_cpu_var(xen_residual_stolen);
142 ticks = iter_div_u64_rem(stolen, NS_PER_TICK, &stolen);
143 __get_cpu_var(xen_residual_stolen) = stolen;
144 account_steal_ticks(ticks);
146 /* Add the appropriate number of ticks of blocked time,
147 including any left-overs from last time. */
148 blocked += __get_cpu_var(xen_residual_blocked);
153 ticks = iter_div_u64_rem(blocked, NS_PER_TICK, &blocked);
154 __get_cpu_var(xen_residual_blocked) = blocked;
155 account_idle_ticks(ticks);
158 /* Get the TSC speed from Xen */
159 unsigned long xen_tsc_khz(void)
161 struct pvclock_vcpu_time_info *info =
162 &HYPERVISOR_shared_info->vcpu_info[0].time;
164 return pvclock_tsc_khz(info);
167 cycle_t xen_clocksource_read(void)
169 struct pvclock_vcpu_time_info *src;
172 src = &get_cpu_var(xen_vcpu)->time;
173 ret = pvclock_clocksource_read(src);
174 put_cpu_var(xen_vcpu);
178 static cycle_t xen_clocksource_get_cycles(struct clocksource *cs)
180 return xen_clocksource_read();
183 static void xen_read_wallclock(struct timespec *ts)
185 struct shared_info *s = HYPERVISOR_shared_info;
186 struct pvclock_wall_clock *wall_clock = &(s->wc);
187 struct pvclock_vcpu_time_info *vcpu_time;
189 vcpu_time = &get_cpu_var(xen_vcpu)->time;
190 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
191 put_cpu_var(xen_vcpu);
194 unsigned long xen_get_wallclock(void)
198 xen_read_wallclock(&ts);
202 int xen_set_wallclock(unsigned long now)
204 /* do nothing for domU */
208 static struct clocksource xen_clocksource __read_mostly = {
211 .read = xen_clocksource_get_cycles,
213 .mult = 1<<XEN_SHIFT, /* time directly in nanoseconds */
215 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
219 Xen clockevent implementation
221 Xen has two clockevent implementations:
223 The old timer_op one works with all released versions of Xen prior
224 to version 3.0.4. This version of the hypervisor provides a
225 single-shot timer with nanosecond resolution. However, sharing the
226 same event channel is a 100Hz tick which is delivered while the
227 vcpu is running. We don't care about or use this tick, but it will
228 cause the core time code to think the timer fired too soon, and
229 will end up resetting it each time. It could be filtered, but
230 doing so has complications when the ktime clocksource is not yet
231 the xen clocksource (ie, at boot time).
233 The new vcpu_op-based timer interface allows the tick timer period
234 to be changed or turned off. The tick timer is not useful as a
235 periodic timer because events are only delivered to running vcpus.
236 The one-shot timer can report when a timeout is in the past, so
237 set_next_event is capable of returning -ETIME when appropriate.
238 This interface is used when available.
243 Get a hypervisor absolute time. In theory we could maintain an
244 offset between the kernel's time and the hypervisor's time, and
245 apply that to a kernel's absolute timeout. Unfortunately the
246 hypervisor and kernel times can drift even if the kernel is using
247 the Xen clocksource, because ntp can warp the kernel's clocksource.
249 static s64 get_abs_timeout(unsigned long delta)
251 return xen_clocksource_read() + delta;
254 static void xen_timerop_set_mode(enum clock_event_mode mode,
255 struct clock_event_device *evt)
258 case CLOCK_EVT_MODE_PERIODIC:
263 case CLOCK_EVT_MODE_ONESHOT:
264 case CLOCK_EVT_MODE_RESUME:
267 case CLOCK_EVT_MODE_UNUSED:
268 case CLOCK_EVT_MODE_SHUTDOWN:
269 HYPERVISOR_set_timer_op(0); /* cancel timeout */
274 static int xen_timerop_set_next_event(unsigned long delta,
275 struct clock_event_device *evt)
277 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
279 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
282 /* We may have missed the deadline, but there's no real way of
283 knowing for sure. If the event was in the past, then we'll
284 get an immediate interrupt. */
289 static const struct clock_event_device xen_timerop_clockevent = {
291 .features = CLOCK_EVT_FEAT_ONESHOT,
293 .max_delta_ns = 0xffffffff,
294 .min_delta_ns = TIMER_SLOP,
300 .set_mode = xen_timerop_set_mode,
301 .set_next_event = xen_timerop_set_next_event,
306 static void xen_vcpuop_set_mode(enum clock_event_mode mode,
307 struct clock_event_device *evt)
309 int cpu = smp_processor_id();
312 case CLOCK_EVT_MODE_PERIODIC:
313 WARN_ON(1); /* unsupported */
316 case CLOCK_EVT_MODE_ONESHOT:
317 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
321 case CLOCK_EVT_MODE_UNUSED:
322 case CLOCK_EVT_MODE_SHUTDOWN:
323 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, cpu, NULL) ||
324 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
327 case CLOCK_EVT_MODE_RESUME:
332 static int xen_vcpuop_set_next_event(unsigned long delta,
333 struct clock_event_device *evt)
335 int cpu = smp_processor_id();
336 struct vcpu_set_singleshot_timer single;
339 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
341 single.timeout_abs_ns = get_abs_timeout(delta);
342 single.flags = VCPU_SSHOTTMR_future;
344 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single);
346 BUG_ON(ret != 0 && ret != -ETIME);
351 static const struct clock_event_device xen_vcpuop_clockevent = {
353 .features = CLOCK_EVT_FEAT_ONESHOT,
355 .max_delta_ns = 0xffffffff,
356 .min_delta_ns = TIMER_SLOP,
362 .set_mode = xen_vcpuop_set_mode,
363 .set_next_event = xen_vcpuop_set_next_event,
366 static const struct clock_event_device *xen_clockevent =
367 &xen_timerop_clockevent;
368 static DEFINE_PER_CPU(struct clock_event_device, xen_clock_events);
370 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
372 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events);
376 if (evt->event_handler) {
377 evt->event_handler(evt);
381 do_stolen_accounting();
386 void xen_setup_timer(int cpu)
389 struct clock_event_device *evt;
392 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
394 name = kasprintf(GFP_KERNEL, "timer%d", cpu);
396 name = "<timer kasprintf failed>";
398 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
399 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING|IRQF_TIMER,
402 evt = &per_cpu(xen_clock_events, cpu);
403 memcpy(evt, xen_clockevent, sizeof(*evt));
405 evt->cpumask = cpumask_of(cpu);
409 void xen_teardown_timer(int cpu)
411 struct clock_event_device *evt;
413 evt = &per_cpu(xen_clock_events, cpu);
414 unbind_from_irqhandler(evt->irq, NULL);
417 void xen_setup_cpu_clockevents(void)
419 BUG_ON(preemptible());
421 clockevents_register_device(&__get_cpu_var(xen_clock_events));
424 void xen_timer_resume(void)
428 if (xen_clockevent != &xen_vcpuop_clockevent)
431 for_each_online_cpu(cpu) {
432 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
437 __init void xen_time_init(void)
439 int cpu = smp_processor_id();
442 clocksource_register(&xen_clocksource);
444 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) {
445 /* Successfully turned off 100Hz tick, so we have the
446 vcpuop-based timer interface */
447 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
448 xen_clockevent = &xen_vcpuop_clockevent;
451 /* Set initial system time with full resolution */
452 xen_read_wallclock(&tp);
453 do_settimeofday(&tp);
455 setup_force_cpu_cap(X86_FEATURE_TSC);
457 xen_setup_runstate_info(cpu);
458 xen_setup_timer(cpu);
459 xen_setup_cpu_clockevents();