projects/linux·blob v3.13 · refs

402 lines · 9.8 KiB

1/*
2 * linux/kernel/time/tick-common.c
3 *
4 * This file contains the base functions to manage periodic tick
5 * related events.
6 *
7 * Copyright(C) 2005-2006, Thomas Gleixner <[email protected]>
8 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
9 * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
10 *
11 * This code is licenced under the GPL version 2. For details see
12 * kernel-base/COPYING.
13 */
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>
22
23#include <asm/irq_regs.h>
24
25#include "tick-internal.h"
26
27/*
28 * Tick devices
29 */
30DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
31/*
32 * Tick next event: keeps track of the tick time
33 */
34ktime_t tick_next_period;
35ktime_t tick_period;
36
37/*
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:
41 *
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.
45 *
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.
50 */
51int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
52
53/*
54 * Debugging: see timer_list.c
55 */
56struct tick_device *tick_get_device(int cpu)
57{
58 return &per_cpu(tick_cpu_device, cpu);
59}
60
61/**
62 * tick_is_oneshot_available - check for a oneshot capable event device
63 */
64int tick_is_oneshot_available(void)
65{
66 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
67
68 if (!dev || !(dev->features & CLOCK_EVT_FEAT_ONESHOT))
69 return 0;
70 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
71 return 1;
72 return tick_broadcast_oneshot_available();
73}
74
75/*
76 * Periodic tick
77 */
78static void tick_periodic(int cpu)
79{
80 if (tick_do_timer_cpu == cpu) {
81 write_seqlock(&jiffies_lock);
82
83 /* Keep track of the next tick event */
84 tick_next_period = ktime_add(tick_next_period, tick_period);
85
86 do_timer(1);
87 write_sequnlock(&jiffies_lock);
88 }
89
90 update_process_times(user_mode(get_irq_regs()));
91 profile_tick(CPU_PROFILING);
92}
93
94/*
95 * Event handler for periodic ticks
96 */
97void tick_handle_periodic(struct clock_event_device *dev)
98{
99 int cpu = smp_processor_id();
100 ktime_t next;
101
102 tick_periodic(cpu);
103
104 if (dev->mode != CLOCK_EVT_MODE_ONESHOT)
105 return;
106 /*
107 * Setup the next period for devices, which do not have
108 * periodic mode:
109 */
110 next = ktime_add(dev->next_event, tick_period);
111 for (;;) {
112 if (!clockevents_program_event(dev, next, false))
113 return;
114 /*
115 * Have to be careful here. If we're in oneshot mode,
116 * before we call tick_periodic() in a loop, we need
117 * to be sure we're using a real hardware clocksource.
118 * Otherwise we could get trapped in an infinite
119 * loop, as the tick_periodic() increments jiffies,
120 * when then will increment time, posibly causing
121 * the loop to trigger again and again.
122 */
123 if (timekeeping_valid_for_hres())
124 tick_periodic(cpu);
125 next = ktime_add(next, tick_period);
126 }
127}
128
129/*
130 * Setup the device for a periodic tick
131 */
132void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
133{
134 tick_set_periodic_handler(dev, broadcast);
135
136 /* Broadcast setup ? */
137 if (!tick_device_is_functional(dev))
138 return;
139
140 if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
141 !tick_broadcast_oneshot_active()) {
142 clockevents_set_mode(dev, CLOCK_EVT_MODE_PERIODIC);
143 } else {
144 unsigned long seq;
145 ktime_t next;
146
147 do {
148 seq = read_seqbegin(&jiffies_lock);
149 next = tick_next_period;
150 } while (read_seqretry(&jiffies_lock, seq));
151
152 clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
153
154 for (;;) {
155 if (!clockevents_program_event(dev, next, false))
156 return;
157 next = ktime_add(next, tick_period);
158 }
159 }
160}
161
162/*
163 * Setup the tick device
164 */
165static void tick_setup_device(struct tick_device *td,
166 struct clock_event_device *newdev, int cpu,
167 const struct cpumask *cpumask)
168{
169 ktime_t next_event;
170 void (*handler)(struct clock_event_device *) = NULL;
171
172 /*
173 * First device setup ?
174 */
175 if (!td->evtdev) {
176 /*
177 * If no cpu took the do_timer update, assign it to
178 * this cpu:
179 */
180 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
181 if (!tick_nohz_full_cpu(cpu))
182 tick_do_timer_cpu = cpu;
183 else
184 tick_do_timer_cpu = TICK_DO_TIMER_NONE;
185 tick_next_period = ktime_get();
186 tick_period = ktime_set(0, NSEC_PER_SEC / HZ);
187 }
188
189 /*
190 * Startup in periodic mode first.
191 */
192 td->mode = TICKDEV_MODE_PERIODIC;
193 } else {
194 handler = td->evtdev->event_handler;
195 next_event = td->evtdev->next_event;
196 td->evtdev->event_handler = clockevents_handle_noop;
197 }
198
199 td->evtdev = newdev;
200
201 /*
202 * When the device is not per cpu, pin the interrupt to the
203 * current cpu:
204 */
205 if (!cpumask_equal(newdev->cpumask, cpumask))
206 irq_set_affinity(newdev->irq, cpumask);
207
208 /*
209 * When global broadcasting is active, check if the current
210 * device is registered as a placeholder for broadcast mode.
211 * This allows us to handle this x86 misfeature in a generic
212 * way. This function also returns !=0 when we keep the
213 * current active broadcast state for this CPU.
214 */
215 if (tick_device_uses_broadcast(newdev, cpu))
216 return;
217
218 if (td->mode == TICKDEV_MODE_PERIODIC)
219 tick_setup_periodic(newdev, 0);
220 else
221 tick_setup_oneshot(newdev, handler, next_event);
222}
223
224void tick_install_replacement(struct clock_event_device *newdev)
225{
226 struct tick_device *td = &__get_cpu_var(tick_cpu_device);
227 int cpu = smp_processor_id();
228
229 clockevents_exchange_device(td->evtdev, newdev);
230 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
231 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
232 tick_oneshot_notify();
233}
234
235static bool tick_check_percpu(struct clock_event_device *curdev,
236 struct clock_event_device *newdev, int cpu)
237{
238 if (!cpumask_test_cpu(cpu, newdev->cpumask))
239 return false;
240 if (cpumask_equal(newdev->cpumask, cpumask_of(cpu)))
241 return true;
242 /* Check if irq affinity can be set */
243 if (newdev->irq >= 0 && !irq_can_set_affinity(newdev->irq))
244 return false;
245 /* Prefer an existing cpu local device */
246 if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
247 return false;
248 return true;
249}
250
251static bool tick_check_preferred(struct clock_event_device *curdev,
252 struct clock_event_device *newdev)
253{
254 /* Prefer oneshot capable device */
255 if (!(newdev->features & CLOCK_EVT_FEAT_ONESHOT)) {
256 if (curdev && (curdev->features & CLOCK_EVT_FEAT_ONESHOT))
257 return false;
258 if (tick_oneshot_mode_active())
259 return false;
260 }
261
262 /*
263 * Use the higher rated one, but prefer a CPU local device with a lower
264 * rating than a non-CPU local device
265 */
266 return !curdev ||
267 newdev->rating > curdev->rating ||
268 !cpumask_equal(curdev->cpumask, newdev->cpumask);
269}
270
271/*
272 * Check whether the new device is a better fit than curdev. curdev
273 * can be NULL !
274 */
275bool tick_check_replacement(struct clock_event_device *curdev,
276 struct clock_event_device *newdev)
277{
278 if (tick_check_percpu(curdev, newdev, smp_processor_id()))
279 return false;
280
281 return tick_check_preferred(curdev, newdev);
282}
283
284/*
285 * Check, if the new registered device should be used. Called with
286 * clockevents_lock held and interrupts disabled.
287 */
288void tick_check_new_device(struct clock_event_device *newdev)
289{
290 struct clock_event_device *curdev;
291 struct tick_device *td;
292 int cpu;
293
294 cpu = smp_processor_id();
295 if (!cpumask_test_cpu(cpu, newdev->cpumask))
296 goto out_bc;
297
298 td = &per_cpu(tick_cpu_device, cpu);
299 curdev = td->evtdev;
300
301 /* cpu local device ? */
302 if (!tick_check_percpu(curdev, newdev, cpu))
303 goto out_bc;
304
305 /* Preference decision */
306 if (!tick_check_preferred(curdev, newdev))
307 goto out_bc;
308
309 if (!try_module_get(newdev->owner))
310 return;
311
312 /*
313 * Replace the eventually existing device by the new
314 * device. If the current device is the broadcast device, do
315 * not give it back to the clockevents layer !
316 */
317 if (tick_is_broadcast_device(curdev)) {
318 clockevents_shutdown(curdev);
319 curdev = NULL;
320 }
321 clockevents_exchange_device(curdev, newdev);
322 tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
323 if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
324 tick_oneshot_notify();
325 return;
326
327out_bc:
328 /*
329 * Can the new device be used as a broadcast device ?
330 */
331 tick_install_broadcast_device(newdev);
332}
333
334/*
335 * Transfer the do_timer job away from a dying cpu.
336 *
337 * Called with interrupts disabled.
338 */
339void tick_handover_do_timer(int *cpup)
340{
341 if (*cpup == tick_do_timer_cpu) {
342 int cpu = cpumask_first(cpu_online_mask);
343
344 tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
345 TICK_DO_TIMER_NONE;
346 }
347}
348
349/*
350 * Shutdown an event device on a given cpu:
351 *
352 * This is called on a life CPU, when a CPU is dead. So we cannot
353 * access the hardware device itself.
354 * We just set the mode and remove it from the lists.
355 */
356void tick_shutdown(unsigned int *cpup)
357{
358 struct tick_device *td = &per_cpu(tick_cpu_device, *cpup);
359 struct clock_event_device *dev = td->evtdev;
360
361 td->mode = TICKDEV_MODE_PERIODIC;
362 if (dev) {
363 /*
364 * Prevent that the clock events layer tries to call
365 * the set mode function!
366 */
367 dev->mode = CLOCK_EVT_MODE_UNUSED;
368 clockevents_exchange_device(dev, NULL);
369 dev->event_handler = clockevents_handle_noop;
370 td->evtdev = NULL;
371 }
372}
373
374void tick_suspend(void)
375{
376 struct tick_device *td = &__get_cpu_var(tick_cpu_device);
377
378 clockevents_shutdown(td->evtdev);
379}
380
381void tick_resume(void)
382{
383 struct tick_device *td = &__get_cpu_var(tick_cpu_device);
384 int broadcast = tick_resume_broadcast();
385
386 clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_RESUME);
387
388 if (!broadcast) {
389 if (td->mode == TICKDEV_MODE_PERIODIC)
390 tick_setup_periodic(td->evtdev, 0);
391 else
392 tick_resume_oneshot();
393 }
394}
395
396/**
397 * tick_init - initialize the tick control
398 */
399void __init tick_init(void)
400{
401 tick_broadcast_init();
402}