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 and blocked time */
40 static DEFINE_PER_CPU(u64, xen_residual_stolen);
41 static DEFINE_PER_CPU(u64, xen_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(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 blocked, 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 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. */
136 stolen = runnable + offline + __this_cpu_read(xen_residual_stolen);
141 ticks = iter_div_u64_rem(stolen, NS_PER_TICK, &stolen);
142 __this_cpu_write(xen_residual_stolen, stolen);
143 account_steal_ticks(ticks);
145 /* Add the appropriate number of ticks of blocked time,
146 including any left-overs from last time. */
147 blocked += __this_cpu_read(xen_residual_blocked);
152 ticks = iter_div_u64_rem(blocked, NS_PER_TICK, &blocked);
153 __this_cpu_write(xen_residual_blocked, blocked);
154 account_idle_ticks(ticks);
157 /* Get the TSC speed from Xen */
158 static unsigned long xen_tsc_khz(void)
160 struct pvclock_vcpu_time_info *info =
161 &HYPERVISOR_shared_info->vcpu_info[0].time;
163 return pvclock_tsc_khz(info);
166 cycle_t xen_clocksource_read(void)
168 struct pvclock_vcpu_time_info *src;
171 src = &get_cpu_var(xen_vcpu)->time;
172 ret = pvclock_clocksource_read(src);
173 put_cpu_var(xen_vcpu);
177 static cycle_t xen_clocksource_get_cycles(struct clocksource *cs)
179 return xen_clocksource_read();
182 static void xen_read_wallclock(struct timespec *ts)
184 struct shared_info *s = HYPERVISOR_shared_info;
185 struct pvclock_wall_clock *wall_clock = &(s->wc);
186 struct pvclock_vcpu_time_info *vcpu_time;
188 vcpu_time = &get_cpu_var(xen_vcpu)->time;
189 pvclock_read_wallclock(wall_clock, vcpu_time, ts);
190 put_cpu_var(xen_vcpu);
193 static unsigned long xen_get_wallclock(void)
197 xen_read_wallclock(&ts);
201 static int xen_set_wallclock(unsigned long now)
203 struct xen_platform_op op;
206 /* do nothing for domU */
207 if (!xen_initial_domain())
210 op.cmd = XENPF_settime;
211 op.u.settime.secs = now;
212 op.u.settime.nsecs = 0;
213 op.u.settime.system_time = xen_clocksource_read();
215 rc = HYPERVISOR_dom0_op(&op);
216 WARN(rc != 0, "XENPF_settime failed: now=%ld\n", now);
221 static struct clocksource xen_clocksource __read_mostly = {
224 .read = xen_clocksource_get_cycles,
226 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
230 Xen clockevent implementation
232 Xen has two clockevent implementations:
234 The old timer_op one works with all released versions of Xen prior
235 to version 3.0.4. This version of the hypervisor provides a
236 single-shot timer with nanosecond resolution. However, sharing the
237 same event channel is a 100Hz tick which is delivered while the
238 vcpu is running. We don't care about or use this tick, but it will
239 cause the core time code to think the timer fired too soon, and
240 will end up resetting it each time. It could be filtered, but
241 doing so has complications when the ktime clocksource is not yet
242 the xen clocksource (ie, at boot time).
244 The new vcpu_op-based timer interface allows the tick timer period
245 to be changed or turned off. The tick timer is not useful as a
246 periodic timer because events are only delivered to running vcpus.
247 The one-shot timer can report when a timeout is in the past, so
248 set_next_event is capable of returning -ETIME when appropriate.
249 This interface is used when available.
254 Get a hypervisor absolute time. In theory we could maintain an
255 offset between the kernel's time and the hypervisor's time, and
256 apply that to a kernel's absolute timeout. Unfortunately the
257 hypervisor and kernel times can drift even if the kernel is using
258 the Xen clocksource, because ntp can warp the kernel's clocksource.
260 static s64 get_abs_timeout(unsigned long delta)
262 return xen_clocksource_read() + delta;
265 static void xen_timerop_set_mode(enum clock_event_mode mode,
266 struct clock_event_device *evt)
269 case CLOCK_EVT_MODE_PERIODIC:
274 case CLOCK_EVT_MODE_ONESHOT:
275 case CLOCK_EVT_MODE_RESUME:
278 case CLOCK_EVT_MODE_UNUSED:
279 case CLOCK_EVT_MODE_SHUTDOWN:
280 HYPERVISOR_set_timer_op(0); /* cancel timeout */
285 static int xen_timerop_set_next_event(unsigned long delta,
286 struct clock_event_device *evt)
288 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
290 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0)
293 /* We may have missed the deadline, but there's no real way of
294 knowing for sure. If the event was in the past, then we'll
295 get an immediate interrupt. */
300 static const struct clock_event_device xen_timerop_clockevent = {
302 .features = CLOCK_EVT_FEAT_ONESHOT,
304 .max_delta_ns = 0xffffffff,
305 .min_delta_ns = TIMER_SLOP,
311 .set_mode = xen_timerop_set_mode,
312 .set_next_event = xen_timerop_set_next_event,
317 static void xen_vcpuop_set_mode(enum clock_event_mode mode,
318 struct clock_event_device *evt)
320 int cpu = smp_processor_id();
323 case CLOCK_EVT_MODE_PERIODIC:
324 WARN_ON(1); /* unsupported */
327 case CLOCK_EVT_MODE_ONESHOT:
328 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
332 case CLOCK_EVT_MODE_UNUSED:
333 case CLOCK_EVT_MODE_SHUTDOWN:
334 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, cpu, NULL) ||
335 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
338 case CLOCK_EVT_MODE_RESUME:
343 static int xen_vcpuop_set_next_event(unsigned long delta,
344 struct clock_event_device *evt)
346 int cpu = smp_processor_id();
347 struct vcpu_set_singleshot_timer single;
350 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT);
352 single.timeout_abs_ns = get_abs_timeout(delta);
353 single.flags = VCPU_SSHOTTMR_future;
355 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single);
357 BUG_ON(ret != 0 && ret != -ETIME);
362 static const struct clock_event_device xen_vcpuop_clockevent = {
364 .features = CLOCK_EVT_FEAT_ONESHOT,
366 .max_delta_ns = 0xffffffff,
367 .min_delta_ns = TIMER_SLOP,
373 .set_mode = xen_vcpuop_set_mode,
374 .set_next_event = xen_vcpuop_set_next_event,
377 static const struct clock_event_device *xen_clockevent =
378 &xen_timerop_clockevent;
379 static DEFINE_PER_CPU(struct clock_event_device, xen_clock_events);
381 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id)
383 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events);
387 if (evt->event_handler) {
388 evt->event_handler(evt);
392 do_stolen_accounting();
397 void xen_setup_timer(int cpu)
400 struct clock_event_device *evt;
403 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu);
405 name = kasprintf(GFP_KERNEL, "timer%d", cpu);
407 name = "<timer kasprintf failed>";
409 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt,
410 IRQF_DISABLED|IRQF_PERCPU|
411 IRQF_NOBALANCING|IRQF_TIMER|
415 evt = &per_cpu(xen_clock_events, cpu);
416 memcpy(evt, xen_clockevent, sizeof(*evt));
418 evt->cpumask = cpumask_of(cpu);
422 void xen_teardown_timer(int cpu)
424 struct clock_event_device *evt;
426 evt = &per_cpu(xen_clock_events, cpu);
427 unbind_from_irqhandler(evt->irq, NULL);
430 void xen_setup_cpu_clockevents(void)
432 BUG_ON(preemptible());
434 clockevents_register_device(&__get_cpu_var(xen_clock_events));
437 void xen_timer_resume(void)
443 if (xen_clockevent != &xen_vcpuop_clockevent)
446 for_each_online_cpu(cpu) {
447 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL))
452 static const struct pv_time_ops xen_time_ops __initconst = {
453 .sched_clock = xen_clocksource_read,
456 static void __init xen_time_init(void)
458 int cpu = smp_processor_id();
461 clocksource_register_hz(&xen_clocksource, NSEC_PER_SEC);
463 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) {
464 /* Successfully turned off 100Hz tick, so we have the
465 vcpuop-based timer interface */
466 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n");
467 xen_clockevent = &xen_vcpuop_clockevent;
470 /* Set initial system time with full resolution */
471 xen_read_wallclock(&tp);
472 do_settimeofday(&tp);
474 setup_force_cpu_cap(X86_FEATURE_TSC);
476 xen_setup_runstate_info(cpu);
477 xen_setup_timer(cpu);
478 xen_setup_cpu_clockevents();
481 void __init xen_init_time_ops(void)
483 pv_time_ops = xen_time_ops;
485 x86_init.timers.timer_init = xen_time_init;
486 x86_init.timers.setup_percpu_clockev = x86_init_noop;
487 x86_cpuinit.setup_percpu_clockev = x86_init_noop;
489 x86_platform.calibrate_tsc = xen_tsc_khz;
490 x86_platform.get_wallclock = xen_get_wallclock;
491 x86_platform.set_wallclock = xen_set_wallclock;
494 #ifdef CONFIG_XEN_PVHVM
495 static void xen_hvm_setup_cpu_clockevents(void)
497 int cpu = smp_processor_id();
498 xen_setup_runstate_info(cpu);
499 xen_setup_timer(cpu);
500 xen_setup_cpu_clockevents();
503 void __init xen_hvm_init_time_ops(void)
505 /* vector callback is needed otherwise we cannot receive interrupts
506 * on cpu > 0 and at this point we don't know how many cpus are
508 if (!xen_have_vector_callback)
510 if (!xen_feature(XENFEAT_hvm_safe_pvclock)) {
511 printk(KERN_INFO "Xen doesn't support pvclock on HVM,"
512 "disable pv timer\n");
516 pv_time_ops = xen_time_ops;
517 x86_init.timers.setup_percpu_clockev = xen_time_init;
518 x86_cpuinit.setup_percpu_clockev = xen_hvm_setup_cpu_clockevents;
520 x86_platform.calibrate_tsc = xen_tsc_khz;
521 x86_platform.get_wallclock = xen_get_wallclock;
522 x86_platform.set_wallclock = xen_set_wallclock;