kernel/time/tick-broadcast.c

   1 /*
   2  * linux/kernel/time/tick-broadcast.c
   3  *
   4  * This file contains functions which emulate a local clock-event
   5  * device via a broadcast event source.
   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
  22 #include "tick-internal.h"
  23
  24 /*
  25  * Broadcast support for broken x86 hardware, where the local apic
  26  * timer stops in C3 state.
  27  */
  28
  29 static struct tick_device tick_broadcast_device;
  30 /* FIXME: Use cpumask_var_t. */
  31 static DECLARE_BITMAP(tick_broadcast_mask, NR_CPUS);
  32 static DECLARE_BITMAP(tmpmask, NR_CPUS);
  33 static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
  34 static int tick_broadcast_force;
  35
  36 #ifdef CONFIG_TICK_ONESHOT
  37 static void tick_broadcast_clear_oneshot(int cpu);
  38 #else
  39 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  40 #endif
  41
  42 /*
  43  * Debugging: see timer_list.c
  44  */
  45 struct tick_device *tick_get_broadcast_device(void)
  46 {
  47         return &tick_broadcast_device;
  48 }
  49
  50 struct cpumask *tick_get_broadcast_mask(void)
  51 {
  52         return to_cpumask(tick_broadcast_mask);
  53 }
  54
  55 /*
  56  * Start the device in periodic mode
  57  */
  58 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  59 {
  60         if (bc)
  61                 tick_setup_periodic(bc, 1);
  62 }
  63
  64 /*
  65  * Check, if the device can be utilized as broadcast device:
  66  */
  67 int tick_check_broadcast_device(struct clock_event_device *dev)
  68 {
  69         struct clock_event_device *cur = tick_broadcast_device.evtdev;
  70
  71         if ((dev->features & CLOCK_EVT_FEAT_DUMMY) ||
  72             (tick_broadcast_device.evtdev &&
  73              tick_broadcast_device.evtdev->rating >= dev->rating) ||
  74              (dev->features & CLOCK_EVT_FEAT_C3STOP))
  75                 return 0;
  76
  77         clockevents_exchange_device(tick_broadcast_device.evtdev, dev);
  78         if (cur)
  79                 cur->event_handler = clockevents_handle_noop;
  80         tick_broadcast_device.evtdev = dev;
  81         if (!cpumask_empty(tick_get_broadcast_mask()))
  82                 tick_broadcast_start_periodic(dev);
  83         return 1;
  84 }
  85
  86 /*
  87  * Check, if the device is the broadcast device
  88  */
  89 int tick_is_broadcast_device(struct clock_event_device *dev)
  90 {
  91         return (dev && tick_broadcast_device.evtdev == dev);
  92 }
  93
  94 /*
  95  * Check, if the device is disfunctional and a place holder, which
  96  * needs to be handled by the broadcast device.
  97  */
  98 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
  99 {
 100         unsigned long flags;
 101         int ret = 0;
 102
 103         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 104
 105         /*
 106          * Devices might be registered with both periodic and oneshot
 107          * mode disabled. This signals, that the device needs to be
 108          * operated from the broadcast device and is a placeholder for
 109          * the cpu local device.
 110          */
 111         if (!tick_device_is_functional(dev)) {
 112                 dev->event_handler = tick_handle_periodic;
 113                 cpumask_set_cpu(cpu, tick_get_broadcast_mask());
 114                 tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
 115                 ret = 1;
 116         } else {
 117                 /*
 118                  * When the new device is not affected by the stop
 119                  * feature and the cpu is marked in the broadcast mask
 120                  * then clear the broadcast bit.
 121                  */
 122                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
 123                         int cpu = smp_processor_id();
 124
 125                         cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
 126                         tick_broadcast_clear_oneshot(cpu);
 127                 }
 128         }
 129         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 130         return ret;
 131 }
 132
 133 /*
 134  * Broadcast the event to the cpus, which are set in the mask (mangled).
 135  */
 136 static void tick_do_broadcast(struct cpumask *mask)
 137 {
 138         int cpu = smp_processor_id();
 139         struct tick_device *td;
 140
 141         /*
 142          * Check, if the current cpu is in the mask
 143          */
 144         if (cpumask_test_cpu(cpu, mask)) {
 145                 cpumask_clear_cpu(cpu, mask);
 146                 td = &per_cpu(tick_cpu_device, cpu);
 147                 td->evtdev->event_handler(td->evtdev);
 148         }
 149
 150         if (!cpumask_empty(mask)) {
 151                 /*
 152                  * It might be necessary to actually check whether the devices
 153                  * have different broadcast functions. For now, just use the
 154                  * one of the first device. This works as long as we have this
 155                  * misfeature only on x86 (lapic)
 156                  */
 157                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 158                 td->evtdev->broadcast(mask);
 159         }
 160 }
 161
 162 /*
 163  * Periodic broadcast:
 164  * - invoke the broadcast handlers
 165  */
 166 static void tick_do_periodic_broadcast(void)
 167 {
 168         raw_spin_lock(&tick_broadcast_lock);
 169
 170         cpumask_and(to_cpumask(tmpmask),
 171                     cpu_online_mask, tick_get_broadcast_mask());
 172         tick_do_broadcast(to_cpumask(tmpmask));
 173
 174         raw_spin_unlock(&tick_broadcast_lock);
 175 }
 176
 177 /*
 178  * Event handler for periodic broadcast ticks
 179  */
 180 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 181 {
 182         ktime_t next;
 183
 184         tick_do_periodic_broadcast();
 185
 186         /*
 187          * The device is in periodic mode. No reprogramming necessary:
 188          */
 189         if (dev->mode == CLOCK_EVT_MODE_PERIODIC)
 190                 return;
 191
 192         /*
 193          * Setup the next period for devices, which do not have
 194          * periodic mode. We read dev->next_event first and add to it
 195          * when the event already expired. clockevents_program_event()
 196          * sets dev->next_event only when the event is really
 197          * programmed to the device.
 198          */
 199         for (next = dev->next_event; ;) {
 200                 next = ktime_add(next, tick_period);
 201
 202                 if (!clockevents_program_event(dev, next, false))
 203                         return;
 204                 tick_do_periodic_broadcast();
 205         }
 206 }
 207
 208 /*
 209  * Powerstate information: The system enters/leaves a state, where
 210  * affected devices might stop
 211  */
 212 static void tick_do_broadcast_on_off(unsigned long *reason)
 213 {
 214         struct clock_event_device *bc, *dev;
 215         struct tick_device *td;
 216         unsigned long flags;
 217         int cpu, bc_stopped;
 218
 219         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 220
 221         cpu = smp_processor_id();
 222         td = &per_cpu(tick_cpu_device, cpu);
 223         dev = td->evtdev;
 224         bc = tick_broadcast_device.evtdev;
 225
 226         /*
 227          * Is the device not affected by the powerstate ?
 228          */
 229         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 230                 goto out;
 231
 232         if (!tick_device_is_functional(dev))
 233                 goto out;
 234
 235         bc_stopped = cpumask_empty(tick_get_broadcast_mask());
 236
 237         switch (*reason) {
 238         case CLOCK_EVT_NOTIFY_BROADCAST_ON:
 239         case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
 240                 if (!cpumask_test_cpu(cpu, tick_get_broadcast_mask())) {
 241                         cpumask_set_cpu(cpu, tick_get_broadcast_mask());
 242                         if (tick_broadcast_device.mode ==
 243                             TICKDEV_MODE_PERIODIC)
 244                                 clockevents_shutdown(dev);
 245                 }
 246                 if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
 247                         tick_broadcast_force = 1;
 248                 break;
 249         case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
 250                 if (!tick_broadcast_force &&
 251                     cpumask_test_cpu(cpu, tick_get_broadcast_mask())) {
 252                         cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
 253                         if (tick_broadcast_device.mode ==
 254                             TICKDEV_MODE_PERIODIC)
 255                                 tick_setup_periodic(dev, 0);
 256                 }
 257                 break;
 258         }
 259
 260         if (cpumask_empty(tick_get_broadcast_mask())) {
 261                 if (!bc_stopped)
 262                         clockevents_shutdown(bc);
 263         } else if (bc_stopped) {
 264                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 265                         tick_broadcast_start_periodic(bc);
 266                 else
 267                         tick_broadcast_setup_oneshot(bc);
 268         }
 269 out:
 270         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 271 }
 272
 273 /*
 274  * Powerstate information: The system enters/leaves a state, where
 275  * affected devices might stop.
 276  */
 277 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
 278 {
 279         if (!cpumask_test_cpu(*oncpu, cpu_online_mask))
 280                 printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
 281                        "offline CPU #%d\n", *oncpu);
 282         else
 283                 tick_do_broadcast_on_off(&reason);
 284 }
 285
 286 /*
 287  * Set the periodic handler depending on broadcast on/off
 288  */
 289 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 290 {
 291         if (!broadcast)
 292                 dev->event_handler = tick_handle_periodic;
 293         else
 294                 dev->event_handler = tick_handle_periodic_broadcast;
 295 }
 296
 297 /*
 298  * Remove a CPU from broadcasting
 299  */
 300 void tick_shutdown_broadcast(unsigned int *cpup)
 301 {
 302         struct clock_event_device *bc;
 303         unsigned long flags;
 304         unsigned int cpu = *cpup;
 305
 306         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 307
 308         bc = tick_broadcast_device.evtdev;
 309         cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
 310
 311         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 312                 if (bc && cpumask_empty(tick_get_broadcast_mask()))
 313                         clockevents_shutdown(bc);
 314         }
 315
 316         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 317 }
 318
 319 void tick_suspend_broadcast(void)
 320 {
 321         struct clock_event_device *bc;
 322         unsigned long flags;
 323
 324         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 325
 326         bc = tick_broadcast_device.evtdev;
 327         if (bc)
 328                 clockevents_shutdown(bc);
 329
 330         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 331 }
 332
 333 int tick_resume_broadcast(void)
 334 {
 335         struct clock_event_device *bc;
 336         unsigned long flags;
 337         int broadcast = 0;
 338
 339         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 340
 341         bc = tick_broadcast_device.evtdev;
 342
 343         if (bc) {
 344                 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
 345
 346                 switch (tick_broadcast_device.mode) {
 347                 case TICKDEV_MODE_PERIODIC:
 348                         if (!cpumask_empty(tick_get_broadcast_mask()))
 349                                 tick_broadcast_start_periodic(bc);
 350                         broadcast = cpumask_test_cpu(smp_processor_id(),
 351                                                      tick_get_broadcast_mask());
 352                         break;
 353                 case TICKDEV_MODE_ONESHOT:
 354                         broadcast = tick_resume_broadcast_oneshot(bc);
 355                         break;
 356                 }
 357         }
 358         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 359
 360         return broadcast;
 361 }
 362
 363
 364 #ifdef CONFIG_TICK_ONESHOT
 365
 366 /* FIXME: use cpumask_var_t. */
 367 static DECLARE_BITMAP(tick_broadcast_oneshot_mask, NR_CPUS);
 368
 369 /*
 370  * Exposed for debugging: see timer_list.c
 371  */
 372 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 373 {
 374         return to_cpumask(tick_broadcast_oneshot_mask);
 375 }
 376
 377 static int tick_broadcast_set_event(ktime_t expires, int force)
 378 {
 379         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 380
 381         return clockevents_program_event(bc, expires, force);
 382 }
 383
 384 int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 385 {
 386         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 387         return 0;
 388 }
 389
 390 /*
 391  * Called from irq_enter() when idle was interrupted to reenable the
 392  * per cpu device.
 393  */
 394 void tick_check_oneshot_broadcast(int cpu)
 395 {
 396         if (cpumask_test_cpu(cpu, to_cpumask(tick_broadcast_oneshot_mask))) {
 397                 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
 398
 399                 clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_ONESHOT);
 400         }
 401 }
 402
 403 /*
 404  * Handle oneshot mode broadcasting
 405  */
 406 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 407 {
 408         struct tick_device *td;
 409         ktime_t now, next_event;
 410         int cpu;
 411
 412         raw_spin_lock(&tick_broadcast_lock);
 413 again:
 414         dev->next_event.tv64 = KTIME_MAX;
 415         next_event.tv64 = KTIME_MAX;
 416         cpumask_clear(to_cpumask(tmpmask));
 417         now = ktime_get();
 418         /* Find all expired events */
 419         for_each_cpu(cpu, tick_get_broadcast_oneshot_mask()) {
 420                 td = &per_cpu(tick_cpu_device, cpu);
 421                 if (td->evtdev->next_event.tv64 <= now.tv64)
 422                         cpumask_set_cpu(cpu, to_cpumask(tmpmask));
 423                 else if (td->evtdev->next_event.tv64 < next_event.tv64)
 424                         next_event.tv64 = td->evtdev->next_event.tv64;
 425         }
 426
 427         /*
 428          * Wakeup the cpus which have an expired event.
 429          */
 430         tick_do_broadcast(to_cpumask(tmpmask));
 431
 432         /*
 433          * Two reasons for reprogram:
 434          *
 435          * - The global event did not expire any CPU local
 436          * events. This happens in dyntick mode, as the maximum PIT
 437          * delta is quite small.
 438          *
 439          * - There are pending events on sleeping CPUs which were not
 440          * in the event mask
 441          */
 442         if (next_event.tv64 != KTIME_MAX) {
 443                 /*
 444                  * Rearm the broadcast device. If event expired,
 445                  * repeat the above
 446                  */
 447                 if (tick_broadcast_set_event(next_event, 0))
 448                         goto again;
 449         }
 450         raw_spin_unlock(&tick_broadcast_lock);
 451 }
 452
 453 /*
 454  * Powerstate information: The system enters/leaves a state, where
 455  * affected devices might stop
 456  */
 457 void tick_broadcast_oneshot_control(unsigned long reason)
 458 {
 459         struct clock_event_device *bc, *dev;
 460         struct tick_device *td;
 461         unsigned long flags;
 462         int cpu;
 463
 464         /*
 465          * Periodic mode does not care about the enter/exit of power
 466          * states
 467          */
 468         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 469                 return;
 470
 471         /*
 472          * We are called with preemtion disabled from the depth of the
 473          * idle code, so we can't be moved away.
 474          */
 475         cpu = smp_processor_id();
 476         td = &per_cpu(tick_cpu_device, cpu);
 477         dev = td->evtdev;
 478
 479         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 480                 return;
 481
 482         bc = tick_broadcast_device.evtdev;
 483
 484         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 485         if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
 486                 if (!cpumask_test_cpu(cpu, tick_get_broadcast_oneshot_mask())) {
 487                         cpumask_set_cpu(cpu, tick_get_broadcast_oneshot_mask());
 488                         clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
 489                         if (dev->next_event.tv64 < bc->next_event.tv64)
 490                                 tick_broadcast_set_event(dev->next_event, 1);
 491                 }
 492         } else {
 493                 if (cpumask_test_cpu(cpu, tick_get_broadcast_oneshot_mask())) {
 494                         cpumask_clear_cpu(cpu,
 495                                           tick_get_broadcast_oneshot_mask());
 496                         clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
 497                         if (dev->next_event.tv64 != KTIME_MAX)
 498                                 tick_program_event(dev->next_event, 1);
 499                 }
 500         }
 501         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 502 }
 503
 504 /*
 505  * Reset the one shot broadcast for a cpu
 506  *
 507  * Called with tick_broadcast_lock held
 508  */
 509 static void tick_broadcast_clear_oneshot(int cpu)
 510 {
 511         cpumask_clear_cpu(cpu, tick_get_broadcast_oneshot_mask());
 512 }
 513
 514 static void tick_broadcast_init_next_event(struct cpumask *mask,
 515                                            ktime_t expires)
 516 {
 517         struct tick_device *td;
 518         int cpu;
 519
 520         for_each_cpu(cpu, mask) {
 521                 td = &per_cpu(tick_cpu_device, cpu);
 522                 if (td->evtdev)
 523                         td->evtdev->next_event = expires;
 524         }
 525 }
 526
 527 /**
 528  * tick_broadcast_setup_oneshot - setup the broadcast device
 529  */
 530 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 531 {
 532         int cpu = smp_processor_id();
 533
 534         /* Set it up only once ! */
 535         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 536                 int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
 537
 538                 bc->event_handler = tick_handle_oneshot_broadcast;
 539                 clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 540
 541                 /* Take the do_timer update */
 542                 tick_do_timer_cpu = cpu;
 543
 544                 /*
 545                  * We must be careful here. There might be other CPUs
 546                  * waiting for periodic broadcast. We need to set the
 547                  * oneshot_mask bits for those and program the
 548                  * broadcast device to fire.
 549                  */
 550                 cpumask_copy(to_cpumask(tmpmask), tick_get_broadcast_mask());
 551                 cpumask_clear_cpu(cpu, to_cpumask(tmpmask));
 552                 cpumask_or(tick_get_broadcast_oneshot_mask(),
 553                            tick_get_broadcast_oneshot_mask(),
 554                            to_cpumask(tmpmask));
 555
 556                 if (was_periodic && !cpumask_empty(to_cpumask(tmpmask))) {
 557                         tick_broadcast_init_next_event(to_cpumask(tmpmask),
 558                                                        tick_next_period);
 559                         tick_broadcast_set_event(tick_next_period, 1);
 560                 } else
 561                         bc->next_event.tv64 = KTIME_MAX;
 562         } else {
 563                 /*
 564                  * The first cpu which switches to oneshot mode sets
 565                  * the bit for all other cpus which are in the general
 566                  * (periodic) broadcast mask. So the bit is set and
 567                  * would prevent the first broadcast enter after this
 568                  * to program the bc device.
 569                  */
 570                 tick_broadcast_clear_oneshot(cpu);
 571         }
 572 }
 573
 574 /*
 575  * Select oneshot operating mode for the broadcast device
 576  */
 577 void tick_broadcast_switch_to_oneshot(void)
 578 {
 579         struct clock_event_device *bc;
 580         unsigned long flags;
 581
 582         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 583
 584         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 585         bc = tick_broadcast_device.evtdev;
 586         if (bc)
 587                 tick_broadcast_setup_oneshot(bc);
 588         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 589 }
 590
 591
 592 /*
 593  * Remove a dead CPU from broadcasting
 594  */
 595 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
 596 {
 597         unsigned long flags;
 598         unsigned int cpu = *cpup;
 599
 600         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 601
 602         /*
 603          * Clear the broadcast mask flag for the dead cpu, but do not
 604          * stop the broadcast device!
 605          */
 606         cpumask_clear_cpu(cpu, tick_get_broadcast_oneshot_mask());
 607
 608         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 609 }
 610
 611 /*
 612  * Check, whether the broadcast device is in one shot mode
 613  */
 614 int tick_broadcast_oneshot_active(void)
 615 {
 616         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 617 }
 618
 619 /*
 620  * Check whether the broadcast device supports oneshot.
 621  */
 622 bool tick_broadcast_oneshot_available(void)
 623 {
 624         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 625
 626         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 627 }
 628
 629 #endif