2 * linux/kernel/time/tick-common.c
4 * This file contains the base functions to manage periodic tick
7 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
8 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
9 * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
11 * This code is licenced under the GPL version 2. For details see
12 * kernel-base/COPYING.
14 #include <linux/cpu.h>
15 #include <linux/err.h>
16 #include <linux/hrtimer.h>
17 #include <linux/interrupt.h>
18 #include <linux/percpu.h>
19 #include <linux/profile.h>
20 #include <linux/sched.h>
21 #include <linux/module.h>
23 #include <asm/irq_regs.h>
25 #include "tick-internal.h"
30 DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
32 * Tick next event: keeps track of the tick time
34 ktime_t tick_next_period;
38 * tick_do_timer_cpu is a timer core internal variable which holds the CPU NR
39 * which is responsible for calling do_timer(), i.e. the timekeeping stuff. This
40 * variable has two functions:
42 * 1) Prevent a thundering herd issue of a gazillion of CPUs trying to grab the
43 * timekeeping lock all at once. Only the CPU which is assigned to do the
44 * update is handling it.
46 * 2) Hand off the duty in the NOHZ idle case by setting the value to
47 * TICK_DO_TIMER_NONE, i.e. a non existing CPU. So the next cpu which looks
48 * at it will take over and keep the time keeping alive. The handover
49 * procedure also covers cpu hotplug.
51 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
54 * Debugging: see timer_list.c
56 struct tick_device *tick_get_device(int cpu)
58 return &per_cpu(tick_cpu_device, cpu);
62 * tick_is_oneshot_available - check for a oneshot capable event device
64 int tick_is_oneshot_available(void)
66 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
68 if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
70 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
72 return tick_broadcast_oneshot_available();
78 static void tick_periodic(int cpu)
80 if (tick_do_timer_cpu == cpu) {
81 raw_spin_lock(&jiffies_lock);
82 write_seqcount_begin(&jiffies_seq);
84 /* Keep track of the next tick event */
85 tick_next_period = ktime_add(tick_next_period, tick_period);
88 write_seqcount_end(&jiffies_seq);
89 raw_spin_unlock(&jiffies_lock);
93 update_process_times(user_mode(get_irq_regs()));
94 profile_tick(CPU_PROFILING);
98 * Event handler for periodic ticks
100 void tick_handle_periodic(struct clock_event_device *dev)
102 int cpu = smp_processor_id();
103 ktime_t next = dev->next_event;
107 if (dev->state != CLOCK_EVT_STATE_ONESHOT)
111 * Setup the next period for devices, which do not have
114 next = ktime_add(next, tick_period);
116 if (!clockevents_program_event(dev, next, false))
119 * Have to be careful here. If we're in oneshot mode,
120 * before we call tick_periodic() in a loop, we need
121 * to be sure we're using a real hardware clocksource.
122 * Otherwise we could get trapped in an infinite
123 * loop, as the tick_periodic() increments jiffies,
124 * which then will increment time, possibly causing
125 * the loop to trigger again and again.
127 if (timekeeping_valid_for_hres())
133 * Setup the device for a periodic tick
135 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
137 tick_set_periodic_handler(dev, broadcast);
139 /* Broadcast setup ? */
140 if (!tick_device_is_functional(dev))
143 if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
144 !tick_broadcast_oneshot_active()) {
145 clockevents_set_state(dev, CLOCK_EVT_STATE_PERIODIC);
151 seq = read_seqcount_begin(&jiffies_seq);
152 next = tick_next_period;
153 } while (read_seqcount_retry(&jiffies_seq, seq));
155 clockevents_set_state(dev, CLOCK_EVT_STATE_ONESHOT);
158 if (!clockevents_program_event(dev, next, false))
160 next = ktime_add(next, tick_period);
166 * Setup the tick device
168 static void tick_setup_device(struct tick_device *td,
169 struct clock_event_device *newdev, int cpu,
170 const struct cpumask *cpumask)
173 void (*handler)(struct clock_event_device *) = NULL;
176 * First device setup ?
180 * If no cpu took the do_timer update, assign it to
183 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
184 if (!tick_nohz_full_cpu(cpu))
185 tick_do_timer_cpu = cpu;
187 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
188 tick_next_period = ktime_get();
189 tick_period = ktime_set(0, NSEC_PER_SEC / HZ);
193 * Startup in periodic mode first.
195 td->mode = TICKDEV_MODE_PERIODIC;
197 handler = td->evtdev->event_handler;
198 next_event = td->evtdev->next_event;
199 td->evtdev->event_handler = clockevents_handle_noop;
205 * When the device is not per cpu, pin the interrupt to the
208 if (!cpumask_equal(newdev->cpumask, cpumask))
209 irq_set_affinity(newdev->irq, cpumask);
212 * When global broadcasting is active, check if the current
213 * device is registered as a placeholder for broadcast mode.
214 * This allows us to handle this x86 misfeature in a generic
215 * way. This function also returns !=0 when we keep the
216 * current active broadcast state for this CPU.
218 if (tick_device_uses_broadcast(newdev, cpu))
221 if (td->mode == TICKDEV_MODE_PERIODIC)
222 tick_setup_periodic(newdev, 0);
224 tick_setup_oneshot(newdev, handler, next_event);
227 void tick_install_replacement(struct clock_event_device *newdev)
229 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
230 int cpu = smp_processor_id();
232 clockevents_exchange_device(td->evtdev, newdev);
233 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
234 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
235 tick_oneshot_notify();
238 static bool tick_check_percpu(struct clock_event_device *curdev,
239 struct clock_event_device *newdev, int cpu)
241 if (!cpumask_test_cpu(cpu, newdev->cpumask))
243 if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
245 /* Check if irq affinity can be set */
246 if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
248 /* Prefer an existing cpu local device */
249 if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
254 static bool tick_check_preferred(struct clock_event_device *curdev,
255 struct clock_event_device *newdev)
257 /* Prefer oneshot capable device */
258 if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
259 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
261 if (tick_oneshot_mode_active())
266 * Use the higher rated one, but prefer a CPU local device with a lower
267 * rating than a non-CPU local device
270 newdev->rating > curdev->rating ||
271 !cpumask_equal(curdev->cpumask, newdev->cpumask);
275 * Check whether the new device is a better fit than curdev. curdev
278 bool tick_check_replacement(struct clock_event_device *curdev,
279 struct clock_event_device *newdev)
281 if (!tick_check_percpu(curdev, newdev, smp_processor_id()))
284 return tick_check_preferred(curdev, newdev);
288 * Check, if the new registered device should be used. Called with
289 * clockevents_lock held and interrupts disabled.
291 void tick_check_new_device(struct clock_event_device *newdev)
293 struct clock_event_device *curdev;
294 struct tick_device *td;
297 cpu = smp_processor_id();
298 if (!cpumask_test_cpu(cpu, newdev->cpumask))
301 td = &per_cpu(tick_cpu_device, cpu);
304 /* cpu local device ? */
305 if (!tick_check_percpu(curdev, newdev, cpu))
308 /* Preference decision */
309 if (!tick_check_preferred(curdev, newdev))
312 if (!try_module_get(newdev->owner))
316 * Replace the eventually existing device by the new
317 * device. If the current device is the broadcast device, do
318 * not give it back to the clockevents layer !
320 if (tick_is_broadcast_device(curdev)) {
321 clockevents_shutdown(curdev);
324 clockevents_exchange_device(curdev, newdev);
325 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
326 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
327 tick_oneshot_notify();
332 * Can the new device be used as a broadcast device ?
334 tick_install_broadcast_device(newdev);
337 #ifdef CONFIG_HOTPLUG_CPU
339 * Transfer the do_timer job away from a dying cpu.
341 * Called with interrupts disabled. Not locking required. If
342 * tick_do_timer_cpu is owned by this cpu, nothing can change it.
344 void tick_handover_do_timer(void)
346 if (tick_do_timer_cpu == smp_processor_id()) {
347 int cpu = cpumask_first(cpu_online_mask);
349 tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
355 * Shutdown an event device on a given cpu:
357 * This is called on a life CPU, when a CPU is dead. So we cannot
358 * access the hardware device itself.
359 * We just set the mode and remove it from the lists.
361 void tick_shutdown(unsigned int cpu)
363 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
364 struct clock_event_device *dev = td->evtdev;
366 td->mode = TICKDEV_MODE_PERIODIC;
369 * Prevent that the clock events layer tries to call
370 * the set mode function!
372 dev->state = CLOCK_EVT_STATE_DETACHED;
373 dev->mode = CLOCK_EVT_MODE_UNUSED;
374 clockevents_exchange_device(dev, NULL);
375 dev->event_handler = clockevents_handle_noop;
382 * tick_suspend_local - Suspend the local tick device
384 * Called from the local cpu for freeze with interrupts disabled.
386 * No locks required. Nothing can change the per cpu device.
388 void tick_suspend_local(void)
390 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
392 clockevents_shutdown(td->evtdev);
396 * tick_resume_local - Resume the local tick device
398 * Called from the local CPU for unfreeze or XEN resume magic.
400 * No locks required. Nothing can change the per cpu device.
402 void tick_resume_local(void)
404 struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
405 bool broadcast = tick_resume_check_broadcast();
407 clockevents_tick_resume(td->evtdev);
409 if (td->mode == TICKDEV_MODE_PERIODIC)
410 tick_setup_periodic(td->evtdev, 0);
412 tick_resume_oneshot();
417 * tick_suspend - Suspend the tick and the broadcast device
419 * Called from syscore_suspend() via timekeeping_suspend with only one
420 * CPU online and interrupts disabled or from tick_unfreeze() under
423 * No locks required. Nothing can change the per cpu device.
425 void tick_suspend(void)
427 tick_suspend_local();
428 tick_suspend_broadcast();
432 * tick_resume - Resume the tick and the broadcast device
434 * Called from syscore_resume() via timekeeping_resume with only one
435 * CPU online and interrupts disabled.
437 * No locks required. Nothing can change the per cpu device.
439 void tick_resume(void)
441 tick_resume_broadcast();
445 static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
446 static unsigned int tick_freeze_depth;
449 * tick_freeze - Suspend the local tick and (possibly) timekeeping.
451 * Check if this is the last online CPU executing the function and if so,
452 * suspend timekeeping. Otherwise suspend the local tick.
454 * Call with interrupts disabled. Must be balanced with %tick_unfreeze().
455 * Interrupts must not be enabled before the subsequent %tick_unfreeze().
457 void tick_freeze(void)
459 raw_spin_lock(&tick_freeze_lock);
462 if (tick_freeze_depth == num_online_cpus())
463 timekeeping_suspend();
465 tick_suspend_local();
467 raw_spin_unlock(&tick_freeze_lock);
471 * tick_unfreeze - Resume the local tick and (possibly) timekeeping.
473 * Check if this is the first CPU executing the function and if so, resume
474 * timekeeping. Otherwise resume the local tick.
476 * Call with interrupts disabled. Must be balanced with %tick_freeze().
477 * Interrupts must not be enabled after the preceding %tick_freeze().
479 void tick_unfreeze(void)
481 raw_spin_lock(&tick_freeze_lock);
483 if (tick_freeze_depth == num_online_cpus())
484 timekeeping_resume();
490 raw_spin_unlock(&tick_freeze_lock);
494 * tick_init - initialize the tick control
496 void __init tick_init(void)
498 tick_broadcast_init();