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