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/features.h>
24 #include <xen/interface/xen.h>
25 #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, xen_runstate);
36 /* snapshots of runstate info */
37 static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate_snapshot);
39 /* unused ns of stolen time */
40 static DEFINE_PER_CPU(u64, xen_residual_stolen);
42 /* return an consistent snapshot of 64-bit time/counter value */
43 static u64 get64(const u64 *p)
47 if (BITS_PER_LONG < 64) {
52 * Read high then low, and then make sure high is
53 * still the same; this will only loop if low wraps
54 * and carries into high.
55 * XXX some clean way to make this endian-proof?
62 } while (p32[1] != h);
64 ret = (((u64)h) << 32) | l;
74 static void get_runstate_snapshot(struct vcpu_runstate_info *res)
77 struct vcpu_runstate_info *state;
79 BUG_ON(preemptible());
81 state = &__get_cpu_var(xen_runstate);
84 * The runstate info is always updated by the hypervisor on
85 * the current CPU, so there's no need to use anything
86 * stronger than a compiler barrier when fetching it.
89 state_time = get64(&state->state_entry_time);
93 } while (get64(&state->state_entry_time) != state_time);
96 /* return true when a vcpu could run but has no real cpu to run on */
97 bool xen_vcpu_stolen(int vcpu)
99 return per_cpu(xen_runstate, vcpu).state == RUNSTATE_runnable;
102 void xen_setup_runstate_info(int cpu)
104 struct vcpu_register_runstate_memory_area area;
106 area.addr.v = &per_cpu(xen_runstate, cpu);
108 if (HYPERVISOR_vcpu_op(VCPUOP_register_runstate_memory_area,
113 static void do_stolen_accounting(void)
115 struct vcpu_runstate_info state;
116 struct vcpu_runstate_info *snap;
117 s64 runnable, offline, stolen;
120 get_runstate_snapshot(&state);
122 WARN_ON(state.state != RUNSTATE_running);
124 snap = &__get_cpu_var(xen_runstate_snapshot);
126 /* work out how much time the VCPU has not been runn*ing* */
127 runnable = state.time[RUNSTATE_runnable] - snap->time[RUNSTATE_runnable];
128 offline = state.time[RUNSTATE_offline] - snap->time[RUNSTATE_offline];
132 /* Add the appropriate number of ticks of stolen time,
133 including any left-overs from last time. */
134 stolen = runnable + offline + __this_cpu_read(xen_residual_stolen);
139 ticks = iter_div_u64_rem(stolen, NS_PER_TICK, &stolen);
140 __this_cpu_write(xen_residual_stolen, stolen);
141 account_steal_ticks(ticks);
144 /* Get the TSC speed from Xen */
145 static unsigned long xen_tsc_khz(void)
147 struct pvclock_vcpu_time_info *info =
148 &HYPERVISOR_shared_info->vcpu_info[0].time;
150 return pvclock_tsc_khz(info);
153 cycle_t xen_clocksource_read(void)
155 struct pvclock_vcpu_time_info *src;
158 preempt_disable_notrace();
159 src = &__get_cpu_var(xen_vcpu)->time;
160 ret = pvclock_clocksource_read(src);
161 preempt_enable_notrace();
165 static cycle_t xen_clocksource_get_cycles(struct clocksource *cs)
167 return xen_clocksource_read();
170 static void xen_read_wallclock(struct timespec *ts)
172 struct shared_info *s = HYPERVISOR_shared_info;
173 struct pvclock_wall_clock *wall_clock = &(s->wc);
174 struct pvclock_vcpu_time_info *vcpu_time;
176 vcpu_time = &get_cpu_var(xen_vcpu)->time;
177 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
178 put_cpu_var(xen_vcpu);
181 static unsigned long xen_get_wallclock(void)
185 xen_read_wallclock(&ts);
189 static int xen_set_wallclock(unsigned long now)
191 struct xen_platform_op op;
194 /* do nothing for domU */
195 if (!xen_initial_domain())
198 op.cmd = XENPF_settime;
199 op.u.settime.secs = now;
200 op.u.settime.nsecs = 0;
201 op.u.settime.system_time = xen_clocksource_read();
203 rc = HYPERVISOR_dom0_op(&op);
204 WARN(rc != 0, "XENPF_settime failed: now=%ld\n", now);
209 static struct clocksource xen_clocksource __read_mostly = {
212 .read = xen_clocksource_get_cycles,
214 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
218 Xen clockevent implementation
220 Xen has two clockevent implementations:
222 The old timer_op one works with all released versions of Xen prior
223 to version 3.0.4. This version of the hypervisor provides a
224 single-shot timer with nanosecond resolution. However, sharing the
225 same event channel is a 100Hz tick which is delivered while the
226 vcpu is running. We don't care about or use this tick, but it will
227 cause the core time code to think the timer fired too soon, and
228 will end up resetting it each time. It could be filtered, but
229 doing so has complications when the ktime clocksource is not yet
230 the xen clocksource (ie, at boot time).
232 The new vcpu_op-based timer interface allows the tick timer period
233 to be changed or turned off. The tick timer is not useful as a
234 periodic timer because events are only delivered to running vcpus.
235 The one-shot timer can report when a timeout is in the past, so
236 set_next_event is capable of returning -ETIME when appropriate.
237 This interface is used when available.
242 Get a hypervisor absolute time. In theory we could maintain an
243 offset between the kernel's time and the hypervisor's time, and
244 apply that to a kernel's absolute timeout. Unfortunately the
245 hypervisor and kernel times can drift even if the kernel is using
246 the Xen clocksource, because ntp can warp the kernel's clocksource.
248 static s64 get_abs_timeout(unsigned long delta)
250 return xen_clocksource_read() + delta;
253 static void xen_timerop_set_mode(enum clock_event_mode mode,
254 struct clock_event_device *evt)
257 case CLOCK_EVT_MODE_PERIODIC:
262 case CLOCK_EVT_MODE_ONESHOT:
263 case CLOCK_EVT_MODE_RESUME:
266 case CLOCK_EVT_MODE_UNUSED:
267 case CLOCK_EVT_MODE_SHUTDOWN:
268 HYPERVISOR_set_timer_op(0); /* cancel timeout */
273 static int xen_timerop_set_next_event(unsigned long delta,
274 struct clock_event_device *evt)
276 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
278 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
281 /* We may have missed the deadline, but there's no real way of
282 knowing for sure. If the event was in the past, then we'll
283 get an immediate interrupt. */
288 static const struct clock_event_device xen_timerop_clockevent = {
290 .features = CLOCK_EVT_FEAT_ONESHOT,
292 .max_delta_ns = 0xffffffff,
293 .min_delta_ns = TIMER_SLOP,
299 .set_mode = xen_timerop_set_mode,
300 .set_next_event = xen_timerop_set_next_event,
305 static void xen_vcpuop_set_mode(enum clock_event_mode mode,
306 struct clock_event_device *evt)
308 int cpu = smp_processor_id();
311 case CLOCK_EVT_MODE_PERIODIC:
312 WARN_ON(1); /* unsupported */
315 case CLOCK_EVT_MODE_ONESHOT:
316 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
320 case CLOCK_EVT_MODE_UNUSED:
321 case CLOCK_EVT_MODE_SHUTDOWN:
322 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, cpu, NULL) ||
323 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
326 case CLOCK_EVT_MODE_RESUME:
331 static int xen_vcpuop_set_next_event(unsigned long delta,
332 struct clock_event_device *evt)
334 int cpu = smp_processor_id();
335 struct vcpu_set_singleshot_timer single;
338 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
340 single.timeout_abs_ns = get_abs_timeout(delta);
341 single.flags = VCPU_SSHOTTMR_future;
343 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single);
345 BUG_ON(ret != 0 && ret != -ETIME);
350 static const struct clock_event_device xen_vcpuop_clockevent = {
352 .features = CLOCK_EVT_FEAT_ONESHOT,
354 .max_delta_ns = 0xffffffff,
355 .min_delta_ns = TIMER_SLOP,
361 .set_mode = xen_vcpuop_set_mode,
362 .set_next_event = xen_vcpuop_set_next_event,
365 static const struct clock_event_device *xen_clockevent =
366 &xen_timerop_clockevent;
367 static DEFINE_PER_CPU(struct clock_event_device, xen_clock_events);
369 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
371 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events);
375 if (evt->event_handler) {
376 evt->event_handler(evt);
380 do_stolen_accounting();
385 void xen_setup_timer(int cpu)
388 struct clock_event_device *evt;
391 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
393 name = kasprintf(GFP_KERNEL, "timer%d", cpu);
395 name = "<timer kasprintf failed>";
397 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
398 IRQF_DISABLED|IRQF_PERCPU|
399 IRQF_NOBALANCING|IRQF_TIMER|
400 IRQF_FORCE_RESUME|IRQF_EARLY_RESUME,
403 evt = &per_cpu(xen_clock_events, cpu);
404 memcpy(evt, xen_clockevent, sizeof(*evt));
406 evt->cpumask = cpumask_of(cpu);
410 void xen_teardown_timer(int cpu)
412 struct clock_event_device *evt;
414 evt = &per_cpu(xen_clock_events, cpu);
415 unbind_from_irqhandler(evt->irq, NULL);
418 void xen_setup_cpu_clockevents(void)
420 BUG_ON(preemptible());
422 clockevents_register_device(&__get_cpu_var(xen_clock_events));
425 void xen_timer_resume(void)
431 if (xen_clockevent != &xen_vcpuop_clockevent)
434 for_each_online_cpu(cpu) {
435 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
440 static const struct pv_time_ops xen_time_ops __initconst = {
441 .sched_clock = xen_clocksource_read,
444 static void __init xen_time_init(void)
446 int cpu = smp_processor_id();
449 clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
451 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) {
452 /* Successfully turned off 100Hz tick, so we have the
453 vcpuop-based timer interface */
454 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
455 xen_clockevent = &xen_vcpuop_clockevent;
458 /* Set initial system time with full resolution */
459 xen_read_wallclock(&tp);
460 do_settimeofday(&tp);
462 setup_force_cpu_cap(X86_FEATURE_TSC);
464 xen_setup_runstate_info(cpu);
465 xen_setup_timer(cpu);
466 xen_setup_cpu_clockevents();
469 void __init xen_init_time_ops(void)
471 pv_time_ops = xen_time_ops;
473 x86_init.timers.timer_init = xen_time_init;
474 x86_init.timers.setup_percpu_clockev = x86_init_noop;
475 x86_cpuinit.setup_percpu_clockev = x86_init_noop;
477 x86_platform.calibrate_tsc = xen_tsc_khz;
478 x86_platform.get_wallclock = xen_get_wallclock;
479 x86_platform.set_wallclock = xen_set_wallclock;
482 #ifdef CONFIG_XEN_PVHVM
483 static void xen_hvm_setup_cpu_clockevents(void)
485 int cpu = smp_processor_id();
486 xen_setup_runstate_info(cpu);
488 * xen_setup_timer(cpu) - snprintf is bad in atomic context. Hence
489 * doing it xen_hvm_cpu_notify (which gets called by smp_init during
490 * early bootup and also during CPU hotplug events).
492 xen_setup_cpu_clockevents();
495 void __init xen_hvm_init_time_ops(void)
497 /* vector callback is needed otherwise we cannot receive interrupts
498 * on cpu > 0 and at this point we don't know how many cpus are
500 if (!xen_have_vector_callback)
502 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
503 printk(KERN_INFO "Xen doesn't support pvclock on HVM,"
504 "disable pv timer\n");
508 pv_time_ops = xen_time_ops;
509 x86_init.timers.setup_percpu_clockev = xen_time_init;
510 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
512 x86_platform.calibrate_tsc = xen_tsc_khz;
513 x86_platform.get_wallclock = xen_get_wallclock;
514 x86_platform.set_wallclock = xen_set_wallclock;