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>
17 #include <linux/slab.h>
19 #include <asm/pvclock.h>
20 #include <asm/xen/hypervisor.h>
21 #include <asm/xen/hypercall.h>
23 #include <xen/events.h>
24 #include <xen/features.h>
25 #include <xen/interface/xen.h>
26 #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 time */
41 static DEFINE_PER_CPU(u64, xen_residual_stolen);
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(xen_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(xen_runstate, vcpu).state == RUNSTATE_runnable;
103 void xen_setup_runstate_info(int cpu)
105 struct vcpu_register_runstate_memory_area area;
107 area.addr.v = &per_cpu(xen_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 runnable, offline, stolen;
121 get_runstate_snapshot(&state);
123 WARN_ON(state.state != RUNSTATE_running);
125 snap = &__get_cpu_var(xen_runstate_snapshot);
127 /* work out how much time the VCPU has not been runn*ing* */
128 runnable = state.time[RUNSTATE_runnable] - snap->time[RUNSTATE_runnable];
129 offline = state.time[RUNSTATE_offline] - snap->time[RUNSTATE_offline];
133 /* Add the appropriate number of ticks of stolen time,
134 including any left-overs from last time. */
135 stolen = runnable + offline + __this_cpu_read(xen_residual_stolen);
140 ticks = iter_div_u64_rem(stolen, NS_PER_TICK, &stolen);
141 __this_cpu_write(xen_residual_stolen, stolen);
142 account_steal_ticks(ticks);
145 /* Get the TSC speed from Xen */
146 static unsigned long xen_tsc_khz(void)
148 struct pvclock_vcpu_time_info *info =
149 &HYPERVISOR_shared_info->vcpu_info[0].time;
151 return pvclock_tsc_khz(info);
154 cycle_t xen_clocksource_read(void)
156 struct pvclock_vcpu_time_info *src;
159 preempt_disable_notrace();
160 src = &__get_cpu_var(xen_vcpu)->time;
161 ret = pvclock_clocksource_read(src);
162 preempt_enable_notrace();
166 static cycle_t xen_clocksource_get_cycles(struct clocksource *cs)
168 return xen_clocksource_read();
171 static void xen_read_wallclock(struct timespec *ts)
173 struct shared_info *s = HYPERVISOR_shared_info;
174 struct pvclock_wall_clock *wall_clock = &(s->wc);
175 struct pvclock_vcpu_time_info *vcpu_time;
177 vcpu_time = &get_cpu_var(xen_vcpu)->time;
178 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
179 put_cpu_var(xen_vcpu);
182 static unsigned long xen_get_wallclock(void)
186 xen_read_wallclock(&ts);
190 static int xen_set_wallclock(unsigned long now)
192 struct xen_platform_op op;
195 /* do nothing for domU */
196 if (!xen_initial_domain())
199 op.cmd = XENPF_settime;
200 op.u.settime.secs = now;
201 op.u.settime.nsecs = 0;
202 op.u.settime.system_time = xen_clocksource_read();
204 rc = HYPERVISOR_dom0_op(&op);
205 WARN(rc != 0, "XENPF_settime failed: now=%ld\n", now);
210 static struct clocksource xen_clocksource __read_mostly = {
213 .read = xen_clocksource_get_cycles,
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;
369 struct xen_clock_event_device {
370 struct clock_event_device evt;
373 static DEFINE_PER_CPU(struct xen_clock_event_device, xen_clock_events) = { .evt.irq = -1 };
375 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
377 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events).evt;
381 if (evt->event_handler) {
382 evt->event_handler(evt);
386 do_stolen_accounting();
391 void xen_teardown_timer(int cpu)
393 struct clock_event_device *evt;
395 evt = &per_cpu(xen_clock_events, cpu).evt;
398 unbind_from_irqhandler(evt->irq, NULL);
400 kfree(per_cpu(xen_clock_events, cpu).name);
401 per_cpu(xen_clock_events, cpu).name = NULL;
405 void xen_setup_timer(int cpu)
408 struct clock_event_device *evt;
411 evt = &per_cpu(xen_clock_events, cpu).evt;
412 WARN(evt->irq >= 0, "IRQ%d for CPU%d is already allocated\n", evt->irq, cpu);
414 xen_teardown_timer(cpu);
416 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
418 name = kasprintf(GFP_KERNEL, "timer%d", cpu);
420 name = "<timer kasprintf failed>";
422 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
423 IRQF_DISABLED|IRQF_PERCPU|
424 IRQF_NOBALANCING|IRQF_TIMER|
428 memcpy(evt, xen_clockevent, sizeof(*evt));
430 evt->cpumask = cpumask_of(cpu);
432 per_cpu(xen_clock_events, cpu).name = name;
436 void xen_setup_cpu_clockevents(void)
438 BUG_ON(preemptible());
440 clockevents_register_device(&__get_cpu_var(xen_clock_events).evt);
443 void xen_timer_resume(void)
449 if (xen_clockevent != &xen_vcpuop_clockevent)
452 for_each_online_cpu(cpu) {
453 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
458 static const struct pv_time_ops xen_time_ops __initconst = {
459 .sched_clock = xen_clocksource_read,
462 static void __init xen_time_init(void)
464 int cpu = smp_processor_id();
467 clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
469 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) {
470 /* Successfully turned off 100Hz tick, so we have the
471 vcpuop-based timer interface */
472 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
473 xen_clockevent = &xen_vcpuop_clockevent;
476 /* Set initial system time with full resolution */
477 xen_read_wallclock(&tp);
478 do_settimeofday(&tp);
480 setup_force_cpu_cap(X86_FEATURE_TSC);
482 xen_setup_runstate_info(cpu);
483 xen_setup_timer(cpu);
484 xen_setup_cpu_clockevents();
487 void __init xen_init_time_ops(void)
489 pv_time_ops = xen_time_ops;
491 x86_init.timers.timer_init = xen_time_init;
492 x86_init.timers.setup_percpu_clockev = x86_init_noop;
493 x86_cpuinit.setup_percpu_clockev = x86_init_noop;
495 x86_platform.calibrate_tsc = xen_tsc_khz;
496 x86_platform.get_wallclock = xen_get_wallclock;
497 x86_platform.set_wallclock = xen_set_wallclock;
500 #ifdef CONFIG_XEN_PVHVM
501 static void xen_hvm_setup_cpu_clockevents(void)
503 int cpu = smp_processor_id();
504 xen_setup_runstate_info(cpu);
506 * xen_setup_timer(cpu) - snprintf is bad in atomic context. Hence
507 * doing it xen_hvm_cpu_notify (which gets called by smp_init during
508 * early bootup and also during CPU hotplug events).
510 xen_setup_cpu_clockevents();
513 void __init xen_hvm_init_time_ops(void)
515 /* vector callback is needed otherwise we cannot receive interrupts
516 * on cpu > 0 and at this point we don't know how many cpus are
518 if (!xen_have_vector_callback)
520 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
521 printk(KERN_INFO "Xen doesn't support pvclock on HVM,"
522 "disable pv timer\n");
526 pv_time_ops = xen_time_ops;
527 x86_init.timers.setup_percpu_clockev = xen_time_init;
528 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
530 x86_platform.calibrate_tsc = xen_tsc_khz;
531 x86_platform.get_wallclock = xen_get_wallclock;
532 x86_platform.set_wallclock = xen_set_wallclock;