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 #include <linux/smp.h>
  22
  23 #include "tick-internal.h"
  24
  25 /*
  26  * Broadcast support for broken x86 hardware, where the local apic
  27  * timer stops in C3 state.
  28  */
  29
  30 static struct tick_device tick_broadcast_device;
  31 /* FIXME: Use cpumask_var_t. */
  32 static DECLARE_BITMAP(tick_broadcast_mask, NR_CPUS);
  33 static DECLARE_BITMAP(tmpmask, NR_CPUS);
  34 static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
  35 static int tick_broadcast_force;
  36
  37 #ifdef CONFIG_TICK_ONESHOT
  38 static void tick_broadcast_clear_oneshot(int cpu);
  39 #else
  40 static inline void tick_broadcast_clear_oneshot(int cpu) { }
  41 #endif
  42
  43 /*
  44  * Debugging: see timer_list.c
  45  */
  46 struct tick_device *tick_get_broadcast_device(void)
  47 {
  48         return &tick_broadcast_device;
  49 }
  50
  51 struct cpumask *tick_get_broadcast_mask(void)
  52 {
  53         return to_cpumask(tick_broadcast_mask);
  54 }
  55
  56 /*
  57  * Start the device in periodic mode
  58  */
  59 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  60 {
  61         if (bc)
  62                 tick_setup_periodic(bc, 1);
  63 }
  64
  65 /*
  66  * Check, if the device can be utilized as broadcast device:
  67  */
  68 int tick_check_broadcast_device(struct clock_event_device *dev)
  69 {
  70         if ((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 static void err_broadcast(const struct cpumask *mask)
  91 {
  92         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
  93 }
  94
  95 /*
  96  * Check, if the device is disfunctional and a place holder, which
  97  * needs to be handled by the broadcast device.
  98  */
  99 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 100 {
 101         unsigned long flags;
 102         int ret = 0;
 103
 104         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 105
 106         /*
 107          * Devices might be registered with both periodic and oneshot
 108          * mode disabled. This signals, that the device needs to be
 109          * operated from the broadcast device and is a placeholder for
 110          * the cpu local device.
 111          */
 112         if (!tick_device_is_functional(dev)) {
 113                 dev->event_handler = tick_handle_periodic;
 114                 if (!dev->broadcast)
 115                         dev->broadcast = tick_broadcast;
 116                 if (!dev->broadcast) {
 117                         pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 118                                      dev->name);
 119                         dev->broadcast = err_broadcast;
 120                 }
 121                 cpumask_set_cpu(cpu, tick_get_broadcast_mask());
 122                 tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
 123                 ret = 1;
 124         } else {
 125                 /*
 126                  * When the new device is not affected by the stop
 127                  * feature and the cpu is marked in the broadcast mask
 128                  * then clear the broadcast bit.
 129                  */
 130                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
 131                         int cpu = smp_processor_id();
 132
 133                         cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
 134                         tick_broadcast_clear_oneshot(cpu);
 135                 }
 136         }
 137         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 138         return ret;
 139 }
 140
 141 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 142 int tick_receive_broadcast(void)
 143 {
 144         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 145         struct clock_event_device *evt = td->evtdev;
 146
 147         if (!evt)
 148                 return -ENODEV;
 149
 150         if (!evt->event_handler)
 151                 return -EINVAL;
 152
 153         evt->event_handler(evt);
 154         return 0;
 155 }
 156 #endif
 157
 158 /*
 159  * Broadcast the event to the cpus, which are set in the mask (mangled).
 160  */
 161 static void tick_do_broadcast(struct cpumask *mask)
 162 {
 163         int cpu = smp_processor_id();
 164         struct tick_device *td;
 165
 166         /*
 167          * Check, if the current cpu is in the mask
 168          */
 169         if (cpumask_test_cpu(cpu, mask)) {
 170                 cpumask_clear_cpu(cpu, mask);
 171                 td = &per_cpu(tick_cpu_device, cpu);
 172                 td->evtdev->event_handler(td->evtdev);
 173         }
 174
 175         if (!cpumask_empty(mask)) {
 176                 /*
 177                  * It might be necessary to actually check whether the devices
 178                  * have different broadcast functions. For now, just use the
 179                  * one of the first device. This works as long as we have this
 180                  * misfeature only on x86 (lapic)
 181                  */
 182                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 183                 td->evtdev->broadcast(mask);
 184         }
 185 }
 186
 187 /*
 188  * Periodic broadcast:
 189  * - invoke the broadcast handlers
 190  */
 191 static void tick_do_periodic_broadcast(void)
 192 {
 193         raw_spin_lock(&tick_broadcast_lock);
 194
 195         cpumask_and(to_cpumask(tmpmask),
 196                     cpu_online_mask, tick_get_broadcast_mask());
 197         tick_do_broadcast(to_cpumask(tmpmask));
 198
 199         raw_spin_unlock(&tick_broadcast_lock);
 200 }
 201
 202 /*
 203  * Event handler for periodic broadcast ticks
 204  */
 205 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 206 {
 207         ktime_t next;
 208
 209         tick_do_periodic_broadcast();
 210
 211         /*
 212          * The device is in periodic mode. No reprogramming necessary:
 213          */
 214         if (dev->mode == CLOCK_EVT_MODE_PERIODIC)
 215                 return;
 216
 217         /*
 218          * Setup the next period for devices, which do not have
 219          * periodic mode. We read dev->next_event first and add to it
 220          * when the event already expired. clockevents_program_event()
 221          * sets dev->next_event only when the event is really
 222          * programmed to the device.
 223          */
 224         for (next = dev->next_event; ;) {
 225                 next = ktime_add(next, tick_period);
 226
 227                 if (!clockevents_program_event(dev, next, false))
 228                         return;
 229                 tick_do_periodic_broadcast();
 230         }
 231 }
 232
 233 /*
 234  * Powerstate information: The system enters/leaves a state, where
 235  * affected devices might stop
 236  */
 237 static void tick_do_broadcast_on_off(unsigned long *reason)
 238 {
 239         struct clock_event_device *bc, *dev;
 240         struct tick_device *td;
 241         unsigned long flags;
 242         int cpu, bc_stopped;
 243
 244         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 245
 246         cpu = smp_processor_id();
 247         td = &per_cpu(tick_cpu_device, cpu);
 248         dev = td->evtdev;
 249         bc = tick_broadcast_device.evtdev;
 250
 251         /*
 252          * Is the device not affected by the powerstate ?
 253          */
 254         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 255                 goto out;
 256
 257         if (!tick_device_is_functional(dev))
 258                 goto out;
 259
 260         bc_stopped = cpumask_empty(tick_get_broadcast_mask());
 261
 262         switch (*reason) {
 263         case CLOCK_EVT_NOTIFY_BROADCAST_ON:
 264         case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
 265                 if (!cpumask_test_cpu(cpu, tick_get_broadcast_mask())) {
 266                         cpumask_set_cpu(cpu, tick_get_broadcast_mask());
 267                         if (tick_broadcast_device.mode ==
 268                             TICKDEV_MODE_PERIODIC)
 269                                 clockevents_shutdown(dev);
 270                 }
 271                 if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
 272                         tick_broadcast_force = 1;
 273                 break;
 274         case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
 275                 if (!tick_broadcast_force &&
 276                     cpumask_test_cpu(cpu, tick_get_broadcast_mask())) {
 277                         cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
 278                         if (tick_broadcast_device.mode ==
 279                             TICKDEV_MODE_PERIODIC)
 280                                 tick_setup_periodic(dev, 0);
 281                 }
 282                 break;
 283         }
 284
 285         if (cpumask_empty(tick_get_broadcast_mask())) {
 286                 if (!bc_stopped)
 287                         clockevents_shutdown(bc);
 288         } else if (bc_stopped) {
 289                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 290                         tick_broadcast_start_periodic(bc);
 291                 else
 292                         tick_broadcast_setup_oneshot(bc);
 293         }
 294 out:
 295         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 296 }
 297
 298 /*
 299  * Powerstate information: The system enters/leaves a state, where
 300  * affected devices might stop.
 301  */
 302 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
 303 {
 304         if (!cpumask_test_cpu(*oncpu, cpu_online_mask))
 305                 printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
 306                        "offline CPU #%d\n", *oncpu);
 307         else
 308                 tick_do_broadcast_on_off(&reason);
 309 }
 310
 311 /*
 312  * Set the periodic handler depending on broadcast on/off
 313  */
 314 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 315 {
 316         if (!broadcast)
 317                 dev->event_handler = tick_handle_periodic;
 318         else
 319                 dev->event_handler = tick_handle_periodic_broadcast;
 320 }
 321
 322 /*
 323  * Remove a CPU from broadcasting
 324  */
 325 void tick_shutdown_broadcast(unsigned int *cpup)
 326 {
 327         struct clock_event_device *bc;
 328         unsigned long flags;
 329         unsigned int cpu = *cpup;
 330
 331         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 332
 333         bc = tick_broadcast_device.evtdev;
 334         cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
 335
 336         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 337                 if (bc && cpumask_empty(tick_get_broadcast_mask()))
 338                         clockevents_shutdown(bc);
 339         }
 340
 341         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 342 }
 343
 344 void tick_suspend_broadcast(void)
 345 {
 346         struct clock_event_device *bc;
 347         unsigned long flags;
 348
 349         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 350
 351         bc = tick_broadcast_device.evtdev;
 352         if (bc)
 353                 clockevents_shutdown(bc);
 354
 355         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 356 }
 357
 358 int tick_resume_broadcast(void)
 359 {
 360         struct clock_event_device *bc;
 361         unsigned long flags;
 362         int broadcast = 0;
 363
 364         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 365
 366         bc = tick_broadcast_device.evtdev;
 367
 368         if (bc) {
 369                 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
 370
 371                 switch (tick_broadcast_device.mode) {
 372                 case TICKDEV_MODE_PERIODIC:
 373                         if (!cpumask_empty(tick_get_broadcast_mask()))
 374                                 tick_broadcast_start_periodic(bc);
 375                         broadcast = cpumask_test_cpu(smp_processor_id(),
 376                                                      tick_get_broadcast_mask());
 377                         break;
 378                 case TICKDEV_MODE_ONESHOT:
 379                         if (!cpumask_empty(tick_get_broadcast_mask()))
 380                                 broadcast = tick_resume_broadcast_oneshot(bc);
 381                         break;
 382                 }
 383         }
 384         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 385
 386         return broadcast;
 387 }
 388
 389
 390 #ifdef CONFIG_TICK_ONESHOT
 391
 392 /* FIXME: use cpumask_var_t. */
 393 static DECLARE_BITMAP(tick_broadcast_oneshot_mask, NR_CPUS);
 394
 395 /*
 396  * Exposed for debugging: see timer_list.c
 397  */
 398 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 399 {
 400         return to_cpumask(tick_broadcast_oneshot_mask);
 401 }
 402
 403 static int tick_broadcast_set_event(ktime_t expires, int force)
 404 {
 405         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 406
 407         if (bc->mode != CLOCK_EVT_MODE_ONESHOT)
 408                 clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 409
 410         return clockevents_program_event(bc, expires, force);
 411 }
 412
 413 int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 414 {
 415         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 416         return 0;
 417 }
 418
 419 /*
 420  * Called from irq_enter() when idle was interrupted to reenable the
 421  * per cpu device.
 422  */
 423 void tick_check_oneshot_broadcast(int cpu)
 424 {
 425         if (cpumask_test_cpu(cpu, to_cpumask(tick_broadcast_oneshot_mask))) {
 426                 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
 427
 428                 clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_ONESHOT);
 429         }
 430 }
 431
 432 /*
 433  * Handle oneshot mode broadcasting
 434  */
 435 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 436 {
 437         struct tick_device *td;
 438         ktime_t now, next_event;
 439         int cpu;
 440
 441         raw_spin_lock(&tick_broadcast_lock);
 442 again:
 443         dev->next_event.tv64 = KTIME_MAX;
 444         next_event.tv64 = KTIME_MAX;
 445         cpumask_clear(to_cpumask(tmpmask));
 446         now = ktime_get();
 447         /* Find all expired events */
 448         for_each_cpu(cpu, tick_get_broadcast_oneshot_mask()) {
 449                 td = &per_cpu(tick_cpu_device, cpu);
 450                 if (td->evtdev->next_event.tv64 <= now.tv64)
 451                         cpumask_set_cpu(cpu, to_cpumask(tmpmask));
 452                 else if (td->evtdev->next_event.tv64 < next_event.tv64)
 453                         next_event.tv64 = td->evtdev->next_event.tv64;
 454         }
 455
 456         /*
 457          * Wakeup the cpus which have an expired event.
 458          */
 459         tick_do_broadcast(to_cpumask(tmpmask));
 460
 461         /*
 462          * Two reasons for reprogram:
 463          *
 464          * - The global event did not expire any CPU local
 465          * events. This happens in dyntick mode, as the maximum PIT
 466          * delta is quite small.
 467          *
 468          * - There are pending events on sleeping CPUs which were not
 469          * in the event mask
 470          */
 471         if (next_event.tv64 != KTIME_MAX) {
 472                 /*
 473                  * Rearm the broadcast device. If event expired,
 474                  * repeat the above
 475                  */
 476                 if (tick_broadcast_set_event(next_event, 0))
 477                         goto again;
 478         }
 479         raw_spin_unlock(&tick_broadcast_lock);
 480 }
 481
 482 /*
 483  * Powerstate information: The system enters/leaves a state, where
 484  * affected devices might stop
 485  */
 486 void tick_broadcast_oneshot_control(unsigned long reason)
 487 {
 488         struct clock_event_device *bc, *dev;
 489         struct tick_device *td;
 490         unsigned long flags;
 491         int cpu;
 492
 493         /*
 494          * Periodic mode does not care about the enter/exit of power
 495          * states
 496          */
 497         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 498                 return;
 499
 500         /*
 501          * We are called with preemtion disabled from the depth of the
 502          * idle code, so we can't be moved away.
 503          */
 504         cpu = smp_processor_id();
 505         td = &per_cpu(tick_cpu_device, cpu);
 506         dev = td->evtdev;
 507
 508         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
 509                 return;
 510
 511         bc = tick_broadcast_device.evtdev;
 512
 513         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 514         if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
 515                 if (!cpumask_test_cpu(cpu, tick_get_broadcast_oneshot_mask())) {
 516                         cpumask_set_cpu(cpu, tick_get_broadcast_oneshot_mask());
 517                         clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
 518                         if (dev->next_event.tv64 < bc->next_event.tv64)
 519                                 tick_broadcast_set_event(dev->next_event, 1);
 520                 }
 521         } else {
 522                 if (cpumask_test_cpu(cpu, tick_get_broadcast_oneshot_mask())) {
 523                         cpumask_clear_cpu(cpu,
 524                                           tick_get_broadcast_oneshot_mask());
 525                         clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
 526                         if (dev->next_event.tv64 != KTIME_MAX)
 527                                 tick_program_event(dev->next_event, 1);
 528                 }
 529         }
 530         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 531 }
 532
 533 /*
 534  * Reset the one shot broadcast for a cpu
 535  *
 536  * Called with tick_broadcast_lock held
 537  */
 538 static void tick_broadcast_clear_oneshot(int cpu)
 539 {
 540         cpumask_clear_cpu(cpu, tick_get_broadcast_oneshot_mask());
 541 }
 542
 543 static void tick_broadcast_init_next_event(struct cpumask *mask,
 544                                            ktime_t expires)
 545 {
 546         struct tick_device *td;
 547         int cpu;
 548
 549         for_each_cpu(cpu, mask) {
 550                 td = &per_cpu(tick_cpu_device, cpu);
 551                 if (td->evtdev)
 552                         td->evtdev->next_event = expires;
 553         }
 554 }
 555
 556 /**
 557  * tick_broadcast_setup_oneshot - setup the broadcast device
 558  */
 559 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 560 {
 561         int cpu = smp_processor_id();
 562
 563         /* Set it up only once ! */
 564         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 565                 int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
 566
 567                 bc->event_handler = tick_handle_oneshot_broadcast;
 568
 569                 /* Take the do_timer update */
 570                 tick_do_timer_cpu = cpu;
 571
 572                 /*
 573                  * We must be careful here. There might be other CPUs
 574                  * waiting for periodic broadcast. We need to set the
 575                  * oneshot_mask bits for those and program the
 576                  * broadcast device to fire.
 577                  */
 578                 cpumask_copy(to_cpumask(tmpmask), tick_get_broadcast_mask());
 579                 cpumask_clear_cpu(cpu, to_cpumask(tmpmask));
 580                 cpumask_or(tick_get_broadcast_oneshot_mask(),
 581                            tick_get_broadcast_oneshot_mask(),
 582                            to_cpumask(tmpmask));
 583
 584                 if (was_periodic && !cpumask_empty(to_cpumask(tmpmask))) {
 585                         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 586                         tick_broadcast_init_next_event(to_cpumask(tmpmask),
 587                                                        tick_next_period);
 588                         tick_broadcast_set_event(tick_next_period, 1);
 589                 } else
 590                         bc->next_event.tv64 = KTIME_MAX;
 591         } else {
 592                 /*
 593                  * The first cpu which switches to oneshot mode sets
 594                  * the bit for all other cpus which are in the general
 595                  * (periodic) broadcast mask. So the bit is set and
 596                  * would prevent the first broadcast enter after this
 597                  * to program the bc device.
 598                  */
 599                 tick_broadcast_clear_oneshot(cpu);
 600         }
 601 }
 602
 603 /*
 604  * Select oneshot operating mode for the broadcast device
 605  */
 606 void tick_broadcast_switch_to_oneshot(void)
 607 {
 608         struct clock_event_device *bc;
 609         unsigned long flags;
 610
 611         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 612
 613         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 614         bc = tick_broadcast_device.evtdev;
 615         if (bc)
 616                 tick_broadcast_setup_oneshot(bc);
 617
 618         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 619 }
 620
 621
 622 /*
 623  * Remove a dead CPU from broadcasting
 624  */
 625 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
 626 {
 627         unsigned long flags;
 628         unsigned int cpu = *cpup;
 629
 630         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 631
 632         /*
 633          * Clear the broadcast mask flag for the dead cpu, but do not
 634          * stop the broadcast device!
 635          */
 636         cpumask_clear_cpu(cpu, tick_get_broadcast_oneshot_mask());
 637
 638         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 639 }
 640
 641 /*
 642  * Check, whether the broadcast device is in one shot mode
 643  */
 644 int tick_broadcast_oneshot_active(void)
 645 {
 646         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 647 }
 648
 649 /*
 650  * Check whether the broadcast device supports oneshot.
 651  */
 652 bool tick_broadcast_oneshot_available(void)
 653 {
 654         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 655
 656         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 657 }
 658
 659 #endif