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 static cycle_t xen_clocksource_read(void);
35 /* runstate info updated by Xen */
36 static DEFINE_PER_CPU(struct vcpu_runstate_info, runstate);
38 /* snapshots of runstate info */
39 static DEFINE_PER_CPU(struct vcpu_runstate_info, runstate_snapshot);
41 /* unused ns of stolen and blocked time */
42 static DEFINE_PER_CPU(u64, residual_stolen);
43 static DEFINE_PER_CPU(u64, residual_blocked);
45 /* return an consistent snapshot of 64-bit time/counter value */
46 static u64 get64(const u64 *p)
50 if (BITS_PER_LONG < 64) {
55 * Read high then low, and then make sure high is
56 * still the same; this will only loop if low wraps
57 * and carries into high.
58 * XXX some clean way to make this endian-proof?
65 } while (p32[1] != h);
67 ret = (((u64)h) << 32) | l;
77 static void get_runstate_snapshot(struct vcpu_runstate_info *res)
80 struct vcpu_runstate_info *state;
82 BUG_ON(preemptible());
84 state = &__get_cpu_var(runstate);
87 * The runstate info is always updated by the hypervisor on
88 * the current CPU, so there's no need to use anything
89 * stronger than a compiler barrier when fetching it.
92 state_time = get64(&state->state_entry_time);
96 } while (get64(&state->state_entry_time) != state_time);
99 /* return true when a vcpu could run but has no real cpu to run on */
100 bool xen_vcpu_stolen(int vcpu)
102 return per_cpu(runstate, vcpu).state == RUNSTATE_runnable;
105 static void setup_runstate_info(int cpu)
107 struct vcpu_register_runstate_memory_area area;
109 area.addr.v = &per_cpu(runstate, cpu);
111 if (HYPERVISOR_vcpu_op(VCPUOP_register_runstate_memory_area,
116 static void do_stolen_accounting(void)
118 struct vcpu_runstate_info state;
119 struct vcpu_runstate_info *snap;
120 s64 blocked, runnable, offline, stolen;
123 get_runstate_snapshot(&state);
125 WARN_ON(state.state != RUNSTATE_running);
127 snap = &__get_cpu_var(runstate_snapshot);
129 /* work out how much time the VCPU has not been runn*ing* */
130 blocked = state.time[RUNSTATE_blocked] - snap->time[RUNSTATE_blocked];
131 runnable = state.time[RUNSTATE_runnable] - snap->time[RUNSTATE_runnable];
132 offline = state.time[RUNSTATE_offline] - snap->time[RUNSTATE_offline];
136 /* Add the appropriate number of ticks of stolen time,
137 including any left-overs from last time. Passing NULL to
138 account_steal_time accounts the time as stolen. */
139 stolen = runnable + offline + __get_cpu_var(residual_stolen);
144 ticks = iter_div_u64_rem(stolen, NS_PER_TICK, &stolen);
145 __get_cpu_var(residual_stolen) = stolen;
146 account_steal_time(NULL, ticks);
148 /* Add the appropriate number of ticks of blocked time,
149 including any left-overs from last time. Passing idle to
150 account_steal_time accounts the time as idle/wait. */
151 blocked += __get_cpu_var(residual_blocked);
156 ticks = iter_div_u64_rem(blocked, NS_PER_TICK, &blocked);
157 __get_cpu_var(residual_blocked) = blocked;
158 account_steal_time(idle_task(smp_processor_id()), ticks);
162 * Xen sched_clock implementation. Returns the number of unstolen
163 * nanoseconds, which is nanoseconds the VCPU spent in RUNNING+BLOCKED
166 unsigned long long xen_sched_clock(void)
168 struct vcpu_runstate_info state;
174 * Ideally sched_clock should be called on a per-cpu basis
175 * anyway, so preempt should already be disabled, but that's
176 * not current practice at the moment.
180 now = xen_clocksource_read();
182 get_runstate_snapshot(&state);
184 WARN_ON(state.state != RUNSTATE_running);
186 offset = now - state.state_entry_time;
190 ret = state.time[RUNSTATE_blocked] +
191 state.time[RUNSTATE_running] +
200 /* Get the CPU speed from Xen */
201 unsigned long xen_cpu_khz(void)
203 u64 xen_khz = 1000000ULL << 32;
204 const struct pvclock_vcpu_time_info *info =
205 &HYPERVISOR_shared_info->vcpu_info[0].time;
207 do_div(xen_khz, info->tsc_to_system_mul);
208 if (info->tsc_shift < 0)
209 xen_khz <<= -info->tsc_shift;
211 xen_khz >>= info->tsc_shift;
216 static cycle_t xen_clocksource_read(void)
218 struct pvclock_vcpu_time_info *src;
221 src = &get_cpu_var(xen_vcpu)->time;
222 ret = pvclock_clocksource_read(src);
223 put_cpu_var(xen_vcpu);
227 static void xen_read_wallclock(struct timespec *ts)
229 struct shared_info *s = HYPERVISOR_shared_info;
230 struct pvclock_wall_clock *wall_clock = &(s->wc);
231 struct pvclock_vcpu_time_info *vcpu_time;
233 vcpu_time = &get_cpu_var(xen_vcpu)->time;
234 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
235 put_cpu_var(xen_vcpu);
238 unsigned long xen_get_wallclock(void)
242 xen_read_wallclock(&ts);
246 int xen_set_wallclock(unsigned long now)
248 /* do nothing for domU */
252 static struct clocksource xen_clocksource __read_mostly = {
255 .read = xen_clocksource_read,
257 .mult = 1<<XEN_SHIFT, /* time directly in nanoseconds */
259 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
263 Xen clockevent implementation
265 Xen has two clockevent implementations:
267 The old timer_op one works with all released versions of Xen prior
268 to version 3.0.4. This version of the hypervisor provides a
269 single-shot timer with nanosecond resolution. However, sharing the
270 same event channel is a 100Hz tick which is delivered while the
271 vcpu is running. We don't care about or use this tick, but it will
272 cause the core time code to think the timer fired too soon, and
273 will end up resetting it each time. It could be filtered, but
274 doing so has complications when the ktime clocksource is not yet
275 the xen clocksource (ie, at boot time).
277 The new vcpu_op-based timer interface allows the tick timer period
278 to be changed or turned off. The tick timer is not useful as a
279 periodic timer because events are only delivered to running vcpus.
280 The one-shot timer can report when a timeout is in the past, so
281 set_next_event is capable of returning -ETIME when appropriate.
282 This interface is used when available.
287 Get a hypervisor absolute time. In theory we could maintain an
288 offset between the kernel's time and the hypervisor's time, and
289 apply that to a kernel's absolute timeout. Unfortunately the
290 hypervisor and kernel times can drift even if the kernel is using
291 the Xen clocksource, because ntp can warp the kernel's clocksource.
293 static s64 get_abs_timeout(unsigned long delta)
295 return xen_clocksource_read() + delta;
298 static void xen_timerop_set_mode(enum clock_event_mode mode,
299 struct clock_event_device *evt)
302 case CLOCK_EVT_MODE_PERIODIC:
307 case CLOCK_EVT_MODE_ONESHOT:
308 case CLOCK_EVT_MODE_RESUME:
311 case CLOCK_EVT_MODE_UNUSED:
312 case CLOCK_EVT_MODE_SHUTDOWN:
313 HYPERVISOR_set_timer_op(0); /* cancel timeout */
318 static int xen_timerop_set_next_event(unsigned long delta,
319 struct clock_event_device *evt)
321 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
323 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
326 /* We may have missed the deadline, but there's no real way of
327 knowing for sure. If the event was in the past, then we'll
328 get an immediate interrupt. */
333 static const struct clock_event_device xen_timerop_clockevent = {
335 .features = CLOCK_EVT_FEAT_ONESHOT,
337 .max_delta_ns = 0xffffffff,
338 .min_delta_ns = TIMER_SLOP,
344 .set_mode = xen_timerop_set_mode,
345 .set_next_event = xen_timerop_set_next_event,
350 static void xen_vcpuop_set_mode(enum clock_event_mode mode,
351 struct clock_event_device *evt)
353 int cpu = smp_processor_id();
356 case CLOCK_EVT_MODE_PERIODIC:
357 WARN_ON(1); /* unsupported */
360 case CLOCK_EVT_MODE_ONESHOT:
361 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
365 case CLOCK_EVT_MODE_UNUSED:
366 case CLOCK_EVT_MODE_SHUTDOWN:
367 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, cpu, NULL) ||
368 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
371 case CLOCK_EVT_MODE_RESUME:
376 static int xen_vcpuop_set_next_event(unsigned long delta,
377 struct clock_event_device *evt)
379 int cpu = smp_processor_id();
380 struct vcpu_set_singleshot_timer single;
383 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
385 single.timeout_abs_ns = get_abs_timeout(delta);
386 single.flags = VCPU_SSHOTTMR_future;
388 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single);
390 BUG_ON(ret != 0 && ret != -ETIME);
395 static const struct clock_event_device xen_vcpuop_clockevent = {
397 .features = CLOCK_EVT_FEAT_ONESHOT,
399 .max_delta_ns = 0xffffffff,
400 .min_delta_ns = TIMER_SLOP,
406 .set_mode = xen_vcpuop_set_mode,
407 .set_next_event = xen_vcpuop_set_next_event,
410 static const struct clock_event_device *xen_clockevent =
411 &xen_timerop_clockevent;
412 static DEFINE_PER_CPU(struct clock_event_device, xen_clock_events);
414 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
416 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events);
420 if (evt->event_handler) {
421 evt->event_handler(evt);
425 do_stolen_accounting();
430 void xen_setup_timer(int cpu)
433 struct clock_event_device *evt;
436 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
438 name = kasprintf(GFP_KERNEL, "timer%d", cpu);
440 name = "<timer kasprintf failed>";
442 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
443 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
446 evt = &per_cpu(xen_clock_events, cpu);
447 memcpy(evt, xen_clockevent, sizeof(*evt));
449 evt->cpumask = cpumask_of_cpu(cpu);
452 setup_runstate_info(cpu);
455 void xen_setup_cpu_clockevents(void)
457 BUG_ON(preemptible());
459 clockevents_register_device(&__get_cpu_var(xen_clock_events));
462 void xen_timer_resume(void)
466 if (xen_clockevent != &xen_vcpuop_clockevent)
469 for_each_online_cpu(cpu) {
470 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
475 __init void xen_time_init(void)
477 int cpu = smp_processor_id();
479 clocksource_register(&xen_clocksource);
481 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) {
482 /* Successfully turned off 100Hz tick, so we have the
483 vcpuop-based timer interface */
484 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
485 xen_clockevent = &xen_vcpuop_clockevent;
488 /* Set initial system time with full resolution */
489 xen_read_wallclock(&xtime);
490 set_normalized_timespec(&wall_to_monotonic,
491 -xtime.tv_sec, -xtime.tv_nsec);
493 setup_force_cpu_cap(X86_FEATURE_TSC);
495 xen_setup_timer(cpu);
496 xen_setup_cpu_clockevents();