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 u64 xen_khz = 1000000ULL << 32;
202 const struct pvclock_vcpu_time_info *info =
203 &HYPERVISOR_shared_info->vcpu_info[0].time;
205 do_div(xen_khz, info->tsc_to_system_mul);
206 if (info->tsc_shift < 0)
207 xen_khz <<= -info->tsc_shift;
209 xen_khz >>= info->tsc_shift;
214 cycle_t xen_clocksource_read(void)
216 struct pvclock_vcpu_time_info *src;
219 src = &get_cpu_var(xen_vcpu)->time;
220 ret = pvclock_clocksource_read(src);
221 put_cpu_var(xen_vcpu);
225 static void xen_read_wallclock(struct timespec *ts)
227 struct shared_info *s = HYPERVISOR_shared_info;
228 struct pvclock_wall_clock *wall_clock = &(s->wc);
229 struct pvclock_vcpu_time_info *vcpu_time;
231 vcpu_time = &get_cpu_var(xen_vcpu)->time;
232 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
233 put_cpu_var(xen_vcpu);
236 unsigned long xen_get_wallclock(void)
240 xen_read_wallclock(&ts);
244 int xen_set_wallclock(unsigned long now)
246 /* do nothing for domU */
250 static struct clocksource xen_clocksource __read_mostly = {
253 .read = xen_clocksource_read,
255 .mult = 1<<XEN_SHIFT, /* time directly in nanoseconds */
257 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
261 Xen clockevent implementation
263 Xen has two clockevent implementations:
265 The old timer_op one works with all released versions of Xen prior
266 to version 3.0.4. This version of the hypervisor provides a
267 single-shot timer with nanosecond resolution. However, sharing the
268 same event channel is a 100Hz tick which is delivered while the
269 vcpu is running. We don't care about or use this tick, but it will
270 cause the core time code to think the timer fired too soon, and
271 will end up resetting it each time. It could be filtered, but
272 doing so has complications when the ktime clocksource is not yet
273 the xen clocksource (ie, at boot time).
275 The new vcpu_op-based timer interface allows the tick timer period
276 to be changed or turned off. The tick timer is not useful as a
277 periodic timer because events are only delivered to running vcpus.
278 The one-shot timer can report when a timeout is in the past, so
279 set_next_event is capable of returning -ETIME when appropriate.
280 This interface is used when available.
285 Get a hypervisor absolute time. In theory we could maintain an
286 offset between the kernel's time and the hypervisor's time, and
287 apply that to a kernel's absolute timeout. Unfortunately the
288 hypervisor and kernel times can drift even if the kernel is using
289 the Xen clocksource, because ntp can warp the kernel's clocksource.
291 static s64 get_abs_timeout(unsigned long delta)
293 return xen_clocksource_read() + delta;
296 static void xen_timerop_set_mode(enum clock_event_mode mode,
297 struct clock_event_device *evt)
300 case CLOCK_EVT_MODE_PERIODIC:
305 case CLOCK_EVT_MODE_ONESHOT:
306 case CLOCK_EVT_MODE_RESUME:
309 case CLOCK_EVT_MODE_UNUSED:
310 case CLOCK_EVT_MODE_SHUTDOWN:
311 HYPERVISOR_set_timer_op(0); /* cancel timeout */
316 static int xen_timerop_set_next_event(unsigned long delta,
317 struct clock_event_device *evt)
319 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
321 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
324 /* We may have missed the deadline, but there's no real way of
325 knowing for sure. If the event was in the past, then we'll
326 get an immediate interrupt. */
331 static const struct clock_event_device xen_timerop_clockevent = {
333 .features = CLOCK_EVT_FEAT_ONESHOT,
335 .max_delta_ns = 0xffffffff,
336 .min_delta_ns = TIMER_SLOP,
342 .set_mode = xen_timerop_set_mode,
343 .set_next_event = xen_timerop_set_next_event,
348 static void xen_vcpuop_set_mode(enum clock_event_mode mode,
349 struct clock_event_device *evt)
351 int cpu = smp_processor_id();
354 case CLOCK_EVT_MODE_PERIODIC:
355 WARN_ON(1); /* unsupported */
358 case CLOCK_EVT_MODE_ONESHOT:
359 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
363 case CLOCK_EVT_MODE_UNUSED:
364 case CLOCK_EVT_MODE_SHUTDOWN:
365 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, cpu, NULL) ||
366 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
369 case CLOCK_EVT_MODE_RESUME:
374 static int xen_vcpuop_set_next_event(unsigned long delta,
375 struct clock_event_device *evt)
377 int cpu = smp_processor_id();
378 struct vcpu_set_singleshot_timer single;
381 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
383 single.timeout_abs_ns = get_abs_timeout(delta);
384 single.flags = VCPU_SSHOTTMR_future;
386 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single);
388 BUG_ON(ret != 0 && ret != -ETIME);
393 static const struct clock_event_device xen_vcpuop_clockevent = {
395 .features = CLOCK_EVT_FEAT_ONESHOT,
397 .max_delta_ns = 0xffffffff,
398 .min_delta_ns = TIMER_SLOP,
404 .set_mode = xen_vcpuop_set_mode,
405 .set_next_event = xen_vcpuop_set_next_event,
408 static const struct clock_event_device *xen_clockevent =
409 &xen_timerop_clockevent;
410 static DEFINE_PER_CPU(struct clock_event_device, xen_clock_events);
412 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
414 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events);
418 if (evt->event_handler) {
419 evt->event_handler(evt);
423 do_stolen_accounting();
428 void xen_setup_timer(int cpu)
431 struct clock_event_device *evt;
434 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
436 name = kasprintf(GFP_KERNEL, "timer%d", cpu);
438 name = "<timer kasprintf failed>";
440 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
441 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
444 evt = &per_cpu(xen_clock_events, cpu);
445 memcpy(evt, xen_clockevent, sizeof(*evt));
447 evt->cpumask = cpumask_of_cpu(cpu);
450 setup_runstate_info(cpu);
453 void xen_teardown_timer(int cpu)
455 struct clock_event_device *evt;
457 evt = &per_cpu(xen_clock_events, cpu);
458 unbind_from_irqhandler(evt->irq, NULL);
461 void xen_setup_cpu_clockevents(void)
463 BUG_ON(preemptible());
465 clockevents_register_device(&__get_cpu_var(xen_clock_events));
468 void xen_timer_resume(void)
472 if (xen_clockevent != &xen_vcpuop_clockevent)
475 for_each_online_cpu(cpu) {
476 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
481 __init void xen_time_init(void)
483 int cpu = smp_processor_id();
485 clocksource_register(&xen_clocksource);
487 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) {
488 /* Successfully turned off 100Hz tick, so we have the
489 vcpuop-based timer interface */
490 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
491 xen_clockevent = &xen_vcpuop_clockevent;
494 /* Set initial system time with full resolution */
495 xen_read_wallclock(&xtime);
496 set_normalized_timespec(&wall_to_monotonic,
497 -xtime.tv_sec, -xtime.tv_nsec);
499 setup_force_cpu_cap(X86_FEATURE_TSC);
501 xen_setup_timer(cpu);
502 xen_setup_cpu_clockevents();