kernel/time/tick-common.c

   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 <tglx@linutronix.de>
   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/tick.h>
  22
  23 #include <asm/irq_regs.h>
  24
  25 #include "tick-internal.h"
  26
  27 /*
  28  * Tick devices
  29  */
  30 DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
  31 /*
  32  * Tick next event: keeps track of the tick time
  33  */
  34 ktime_t tick_next_period;
  35 ktime_t tick_period;
  36 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
  37 static DEFINE_RAW_SPINLOCK(tick_device_lock);
  38
  39 /*
  40  * Debugging: see timer_list.c
  41  */
  42 struct tick_device *tick_get_device(int cpu)
  43 {
  44         return &per_cpu(tick_cpu_device, cpu);
  45 }
  46
  47 /**
  48  * tick_is_oneshot_available - check for a oneshot capable event device
  49  */
  50 int tick_is_oneshot_available(void)
  51 {
  52         struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
  53
  54         return dev && (dev->features & CLOCK_EVT_FEAT_ONESHOT);
  55 }
  56
  57 /*
  58  * Periodic tick
  59  */
  60 static void tick_periodic(int cpu)
  61 {
  62         if (tick_do_timer_cpu == cpu) {
  63                 write_seqlock(&xtime_lock);
  64
  65                 /* Keep track of the next tick event */
  66                 tick_next_period = ktime_add(tick_next_period, tick_period);
  67
  68                 do_timer(1);
  69                 write_sequnlock(&xtime_lock);
  70         }
  71
  72         update_process_times(user_mode(get_irq_regs()));
  73         profile_tick(CPU_PROFILING);
  74 }
  75
  76 /*
  77  * Event handler for periodic ticks
  78  */
  79 void tick_handle_periodic(struct clock_event_device *dev)
  80 {
  81         int cpu = smp_processor_id();
  82         ktime_t next;
  83
  84         tick_periodic(cpu);
  85
  86         if (dev->mode != CLOCK_EVT_MODE_ONESHOT)
  87                 return;
  88         /*
  89          * Setup the next period for devices, which do not have
  90          * periodic mode:
  91          */
  92         next = ktime_add(dev->next_event, tick_period);
  93         for (;;) {
  94                 if (!clockevents_program_event(dev, next, ktime_get()))
  95                         return;
  96                 /*
  97                  * Have to be careful here. If we're in oneshot mode,
  98                  * before we call tick_periodic() in a loop, we need
  99                  * to be sure we're using a real hardware clocksource.
 100                  * Otherwise we could get trapped in an infinite
 101                  * loop, as the tick_periodic() increments jiffies,
 102                  * when then will increment time, posibly causing
 103                  * the loop to trigger again and again.
 104                  */
 105                 if (timekeeping_valid_for_hres())
 106                         tick_periodic(cpu);
 107                 next = ktime_add(next, tick_period);
 108         }
 109 }
 110
 111 /*
 112  * Setup the device for a periodic tick
 113  */
 114 void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
 115 {
 116         tick_set_periodic_handler(dev, broadcast);
 117
 118         /* Broadcast setup ? */
 119         if (!tick_device_is_functional(dev))
 120                 return;
 121
 122         if ((dev->features & CLOCK_EVT_FEAT_PERIODIC) &&
 123             !tick_broadcast_oneshot_active()) {
 124                 clockevents_set_mode(dev, CLOCK_EVT_MODE_PERIODIC);
 125         } else {
 126                 unsigned long seq;
 127                 ktime_t next;
 128
 129                 do {
 130                         seq = read_seqbegin(&xtime_lock);
 131                         next = tick_next_period;
 132                 } while (read_seqretry(&xtime_lock, seq));
 133
 134                 clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
 135
 136                 for (;;) {
 137                         if (!clockevents_program_event(dev, next, ktime_get()))
 138                                 return;
 139                         next = ktime_add(next, tick_period);
 140                 }
 141         }
 142 }
 143
 144 /*
 145  * Setup the tick device
 146  */
 147 static void tick_setup_device(struct tick_device *td,
 148                               struct clock_event_device *newdev, int cpu,
 149                               const struct cpumask *cpumask)
 150 {
 151         ktime_t next_event;
 152         void (*handler)(struct clock_event_device *) = NULL;
 153
 154         /*
 155          * First device setup ?
 156          */
 157         if (!td->evtdev) {
 158                 /*
 159                  * If no cpu took the do_timer update, assign it to
 160                  * this cpu:
 161                  */
 162                 if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
 163                         tick_do_timer_cpu = cpu;
 164                         tick_next_period = ktime_get();
 165                         tick_period = ktime_set(0, NSEC_PER_SEC / HZ);
 166                 }
 167
 168                 /*
 169                  * Startup in periodic mode first.
 170                  */
 171                 td->mode = TICKDEV_MODE_PERIODIC;
 172         } else {
 173                 handler = td->evtdev->event_handler;
 174                 next_event = td->evtdev->next_event;
 175                 td->evtdev->event_handler = clockevents_handle_noop;
 176         }
 177
 178         td->evtdev = newdev;
 179
 180         /*
 181          * When the device is not per cpu, pin the interrupt to the
 182          * current cpu:
 183          */
 184         if (!cpumask_equal(newdev->cpumask, cpumask))
 185                 irq_set_affinity(newdev->irq, cpumask);
 186
 187         /*
 188          * When global broadcasting is active, check if the current
 189          * device is registered as a placeholder for broadcast mode.
 190          * This allows us to handle this x86 misfeature in a generic
 191          * way.
 192          */
 193         if (tick_device_uses_broadcast(newdev, cpu))
 194                 return;
 195
 196         if (td->mode == TICKDEV_MODE_PERIODIC)
 197                 tick_setup_periodic(newdev, 0);
 198         else
 199                 tick_setup_oneshot(newdev, handler, next_event);
 200 }
 201
 202 /*
 203  * Check, if the new registered device should be used.
 204  */
 205 static int tick_check_new_device(struct clock_event_device *newdev)
 206 {
 207         struct clock_event_device *curdev;
 208         struct tick_device *td;
 209         int cpu, ret = NOTIFY_OK;
 210         unsigned long flags;
 211
 212         raw_spin_lock_irqsave(&tick_device_lock, flags);
 213
 214         cpu = smp_processor_id();
 215         if (!cpumask_test_cpu(cpu, newdev->cpumask))
 216                 goto out_bc;
 217
 218         td = &per_cpu(tick_cpu_device, cpu);
 219         curdev = td->evtdev;
 220
 221         /* cpu local device ? */
 222         if (!cpumask_equal(newdev->cpumask, cpumask_of(cpu))) {
 223
 224                 /*
 225                  * If the cpu affinity of the device interrupt can not
 226                  * be set, ignore it.
 227                  */
 228                 if (!irq_can_set_affinity(newdev->irq))
 229                         goto out_bc;
 230
 231                 /*
 232                  * If we have a cpu local device already, do not replace it
 233                  * by a non cpu local device
 234                  */
 235                 if (curdev && cpumask_equal(curdev->cpumask, cpumask_of(cpu)))
 236                         goto out_bc;
 237         }
 238
 239         /*
 240          * If we have an active device, then check the rating and the oneshot
 241          * feature.
 242          */
 243         if (curdev) {
 244                 /*
 245                  * Prefer one shot capable devices !
 246                  */
 247                 if ((curdev->features & CLOCK_EVT_FEAT_ONESHOT) &&
 248                     !(newdev->features & CLOCK_EVT_FEAT_ONESHOT))
 249                         goto out_bc;
 250                 /*
 251                  * Check the rating
 252                  */
 253                 if (curdev->rating >= newdev->rating)
 254                         goto out_bc;
 255         }
 256
 257         /*
 258          * Replace the eventually existing device by the new
 259          * device. If the current device is the broadcast device, do
 260          * not give it back to the clockevents layer !
 261          */
 262         if (tick_is_broadcast_device(curdev)) {
 263                 clockevents_shutdown(curdev);
 264                 curdev = NULL;
 265         }
 266         clockevents_exchange_device(curdev, newdev);
 267         tick_setup_device(td, newdev, cpu, cpumask_of(cpu));
 268         if (newdev->features & CLOCK_EVT_FEAT_ONESHOT)
 269                 tick_oneshot_notify();
 270
 271         raw_spin_unlock_irqrestore(&tick_device_lock, flags);
 272         return NOTIFY_STOP;
 273
 274 out_bc:
 275         /*
 276          * Can the new device be used as a broadcast device ?
 277          */
 278         if (tick_check_broadcast_device(newdev))
 279                 ret = NOTIFY_STOP;
 280
 281         raw_spin_unlock_irqrestore(&tick_device_lock, flags);
 282
 283         return ret;
 284 }
 285
 286 /*
 287  * Transfer the do_timer job away from a dying cpu.
 288  *
 289  * Called with interrupts disabled.
 290  */
 291 static void tick_handover_do_timer(int *cpup)
 292 {
 293         if (*cpup == tick_do_timer_cpu) {
 294                 int cpu = cpumask_first(cpu_online_mask);
 295
 296                 tick_do_timer_cpu = (cpu < nr_cpu_ids) ? cpu :
 297                         TICK_DO_TIMER_NONE;
 298         }
 299 }
 300
 301 /*
 302  * Shutdown an event device on a given cpu:
 303  *
 304  * This is called on a life CPU, when a CPU is dead. So we cannot
 305  * access the hardware device itself.
 306  * We just set the mode and remove it from the lists.
 307  */
 308 static void tick_shutdown(unsigned int *cpup)
 309 {
 310         struct tick_device *td = &per_cpu(tick_cpu_device, *cpup);
 311         struct clock_event_device *dev = td->evtdev;
 312         unsigned long flags;
 313
 314         raw_spin_lock_irqsave(&tick_device_lock, flags);
 315         td->mode = TICKDEV_MODE_PERIODIC;
 316         if (dev) {
 317                 /*
 318                  * Prevent that the clock events layer tries to call
 319                  * the set mode function!
 320                  */
 321                 dev->mode = CLOCK_EVT_MODE_UNUSED;
 322                 clockevents_exchange_device(dev, NULL);
 323                 td->evtdev = NULL;
 324         }
 325         raw_spin_unlock_irqrestore(&tick_device_lock, flags);
 326 }
 327
 328 static void tick_suspend(void)
 329 {
 330         struct tick_device *td = &__get_cpu_var(tick_cpu_device);
 331         unsigned long flags;
 332
 333         raw_spin_lock_irqsave(&tick_device_lock, flags);
 334         clockevents_shutdown(td->evtdev);
 335         raw_spin_unlock_irqrestore(&tick_device_lock, flags);
 336 }
 337
 338 static void tick_resume(void)
 339 {
 340         struct tick_device *td = &__get_cpu_var(tick_cpu_device);
 341         unsigned long flags;
 342         int broadcast = tick_resume_broadcast();
 343
 344         raw_spin_lock_irqsave(&tick_device_lock, flags);
 345         clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_RESUME);
 346
 347         if (!broadcast) {
 348                 if (td->mode == TICKDEV_MODE_PERIODIC)
 349                         tick_setup_periodic(td->evtdev, 0);
 350                 else
 351                         tick_resume_oneshot();
 352         }
 353         raw_spin_unlock_irqrestore(&tick_device_lock, flags);
 354 }
 355
 356 /*
 357  * Notification about clock event devices
 358  */
 359 static int tick_notify(struct notifier_block *nb, unsigned long reason,
 360                                void *dev)
 361 {
 362         switch (reason) {
 363
 364         case CLOCK_EVT_NOTIFY_ADD:
 365                 return tick_check_new_device(dev);
 366
 367         case CLOCK_EVT_NOTIFY_BROADCAST_ON:
 368         case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
 369         case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
 370                 tick_broadcast_on_off(reason, dev);
 371                 break;
 372
 373         case CLOCK_EVT_NOTIFY_BROADCAST_ENTER:
 374         case CLOCK_EVT_NOTIFY_BROADCAST_EXIT:
 375                 tick_broadcast_oneshot_control(reason);
 376                 break;
 377
 378         case CLOCK_EVT_NOTIFY_CPU_DYING:
 379                 tick_handover_do_timer(dev);
 380                 break;
 381
 382         case CLOCK_EVT_NOTIFY_CPU_DEAD:
 383                 tick_shutdown_broadcast_oneshot(dev);
 384                 tick_shutdown_broadcast(dev);
 385                 tick_shutdown(dev);
 386                 break;
 387
 388         case CLOCK_EVT_NOTIFY_SUSPEND:
 389                 tick_suspend();
 390                 tick_suspend_broadcast();
 391                 break;
 392
 393         case CLOCK_EVT_NOTIFY_RESUME:
 394                 tick_resume();
 395                 break;
 396
 397         default:
 398                 break;
 399         }
 400
 401         return NOTIFY_OK;
 402 }
 403
 404 static struct notifier_block tick_notifier = {
 405         .notifier_call = tick_notify,
 406 };
 407
 408 /**
 409  * tick_init - initialize the tick control
 410  *
 411  * Register the notifier with the clockevents framework
 412  */
 413 void __init tick_init(void)
 414 {
 415         clockevents_register_notifier(&tick_notifier);
 416 }