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 static cpumask_var_t tick_broadcast_mask;
  32 static cpumask_var_t tmpmask;
  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 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 ((tick_broadcast_device.evtdev &&
  70              tick_broadcast_device.evtdev->rating >= dev->rating) ||
  71              (dev->features & CLOCK_EVT_FEAT_C3STOP))
  72                 return 0;
  73
  74         clockevents_exchange_device(tick_broadcast_device.evtdev, dev);
  75         tick_broadcast_device.evtdev = dev;
  76         if (!cpumask_empty(tick_broadcast_mask))
  77                 tick_broadcast_start_periodic(dev);
  78         return 1;
  79 }
  80
  81 /*
  82  * Check, if the device is the broadcast device
  83  */
  84 int tick_is_broadcast_device(struct clock_event_device *dev)
  85 {
  86         return (dev && tick_broadcast_device.evtdev == dev);
  87 }
  88
  89 static void err_broadcast(const struct cpumask *mask)
  90 {
  91         pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
  92 }
  93
  94 static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
  95 {
  96         if (!dev->broadcast)
  97                 dev->broadcast = tick_broadcast;
  98         if (!dev->broadcast) {
  99                 pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
 100                              dev->name);
 101                 dev->broadcast = err_broadcast;
 102         }
 103 }
 104
 105 /*
 106  * Check, if the device is disfunctional and a place holder, which
 107  * needs to be handled by the broadcast device.
 108  */
 109 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
 110 {
 111         unsigned long flags;
 112         int ret = 0;
 113
 114         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 115
 116         /*
 117          * Devices might be registered with both periodic and oneshot
 118          * mode disabled. This signals, that the device needs to be
 119          * operated from the broadcast device and is a placeholder for
 120          * the cpu local device.
 121          */
 122         if (!tick_device_is_functional(dev)) {
 123                 dev->event_handler = tick_handle_periodic;
 124                 tick_device_setup_broadcast_func(dev);
 125                 cpumask_set_cpu(cpu, tick_broadcast_mask);
 126                 tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
 127                 ret = 1;
 128         } else {
 129                 /*
 130                  * When the new device is not affected by the stop
 131                  * feature and the cpu is marked in the broadcast mask
 132                  * then clear the broadcast bit.
 133                  */
 134                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
 135                         int cpu = smp_processor_id();
 136                         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 137                         tick_broadcast_clear_oneshot(cpu);
 138                 } else {
 139                         tick_device_setup_broadcast_func(dev);
 140                 }
 141         }
 142         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 143         return ret;
 144 }
 145
 146 #ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
 147 int tick_receive_broadcast(void)
 148 {
 149         struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
 150         struct clock_event_device *evt = td->evtdev;
 151
 152         if (!evt)
 153                 return -ENODEV;
 154
 155         if (!evt->event_handler)
 156                 return -EINVAL;
 157
 158         evt->event_handler(evt);
 159         return 0;
 160 }
 161 #endif
 162
 163 /*
 164  * Broadcast the event to the cpus, which are set in the mask (mangled).
 165  */
 166 static void tick_do_broadcast(struct cpumask *mask)
 167 {
 168         int cpu = smp_processor_id();
 169         struct tick_device *td;
 170
 171         /*
 172          * Check, if the current cpu is in the mask
 173          */
 174         if (cpumask_test_cpu(cpu, mask)) {
 175                 cpumask_clear_cpu(cpu, mask);
 176                 td = &per_cpu(tick_cpu_device, cpu);
 177                 td->evtdev->event_handler(td->evtdev);
 178         }
 179
 180         if (!cpumask_empty(mask)) {
 181                 /*
 182                  * It might be necessary to actually check whether the devices
 183                  * have different broadcast functions. For now, just use the
 184                  * one of the first device. This works as long as we have this
 185                  * misfeature only on x86 (lapic)
 186                  */
 187                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
 188                 td->evtdev->broadcast(mask);
 189         }
 190 }
 191
 192 /*
 193  * Periodic broadcast:
 194  * - invoke the broadcast handlers
 195  */
 196 static void tick_do_periodic_broadcast(void)
 197 {
 198         raw_spin_lock(&tick_broadcast_lock);
 199
 200         cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
 201         tick_do_broadcast(tmpmask);
 202
 203         raw_spin_unlock(&tick_broadcast_lock);
 204 }
 205
 206 /*
 207  * Event handler for periodic broadcast ticks
 208  */
 209 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
 210 {
 211         ktime_t next;
 212
 213         tick_do_periodic_broadcast();
 214
 215         /*
 216          * The device is in periodic mode. No reprogramming necessary:
 217          */
 218         if (dev->mode == CLOCK_EVT_MODE_PERIODIC)
 219                 return;
 220
 221         /*
 222          * Setup the next period for devices, which do not have
 223          * periodic mode. We read dev->next_event first and add to it
 224          * when the event already expired. clockevents_program_event()
 225          * sets dev->next_event only when the event is really
 226          * programmed to the device.
 227          */
 228         for (next = dev->next_event; ;) {
 229                 next = ktime_add(next, tick_period);
 230
 231                 if (!clockevents_program_event(dev, next, false))
 232                         return;
 233                 tick_do_periodic_broadcast();
 234         }
 235 }
 236
 237 /*
 238  * Powerstate information: The system enters/leaves a state, where
 239  * affected devices might stop
 240  */
 241 static void tick_do_broadcast_on_off(unsigned long *reason)
 242 {
 243         struct clock_event_device *bc, *dev;
 244         struct tick_device *td;
 245         unsigned long flags;
 246         int cpu, bc_stopped;
 247
 248         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 249
 250         cpu = smp_processor_id();
 251         td = &per_cpu(tick_cpu_device, cpu);
 252         dev = td->evtdev;
 253         bc = tick_broadcast_device.evtdev;
 254
 255         /*
 256          * Is the device not affected by the powerstate ?
 257          */
 258         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
 259                 goto out;
 260
 261         if (!tick_device_is_functional(dev))
 262                 goto out;
 263
 264         bc_stopped = cpumask_empty(tick_broadcast_mask);
 265
 266         switch (*reason) {
 267         case CLOCK_EVT_NOTIFY_BROADCAST_ON:
 268         case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
 269                 if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
 270                         if (tick_broadcast_device.mode ==
 271                             TICKDEV_MODE_PERIODIC)
 272                                 clockevents_shutdown(dev);
 273                 }
 274                 if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
 275                         tick_broadcast_force = 1;
 276                 break;
 277         case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
 278                 if (!tick_broadcast_force &&
 279                     cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
 280                         if (tick_broadcast_device.mode ==
 281                             TICKDEV_MODE_PERIODIC)
 282                                 tick_setup_periodic(dev, 0);
 283                 }
 284                 break;
 285         }
 286
 287         if (cpumask_empty(tick_broadcast_mask)) {
 288                 if (!bc_stopped)
 289                         clockevents_shutdown(bc);
 290         } else if (bc_stopped) {
 291                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
 292                         tick_broadcast_start_periodic(bc);
 293                 else
 294                         tick_broadcast_setup_oneshot(bc);
 295         }
 296 out:
 297         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 298 }
 299
 300 /*
 301  * Powerstate information: The system enters/leaves a state, where
 302  * affected devices might stop.
 303  */
 304 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
 305 {
 306         if (!cpumask_test_cpu(*oncpu, cpu_online_mask))
 307                 printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
 308                        "offline CPU #%d\n", *oncpu);
 309         else
 310                 tick_do_broadcast_on_off(&reason);
 311 }
 312
 313 /*
 314  * Set the periodic handler depending on broadcast on/off
 315  */
 316 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
 317 {
 318         if (!broadcast)
 319                 dev->event_handler = tick_handle_periodic;
 320         else
 321                 dev->event_handler = tick_handle_periodic_broadcast;
 322 }
 323
 324 /*
 325  * Remove a CPU from broadcasting
 326  */
 327 void tick_shutdown_broadcast(unsigned int *cpup)
 328 {
 329         struct clock_event_device *bc;
 330         unsigned long flags;
 331         unsigned int cpu = *cpup;
 332
 333         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 334
 335         bc = tick_broadcast_device.evtdev;
 336         cpumask_clear_cpu(cpu, tick_broadcast_mask);
 337
 338         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
 339                 if (bc && cpumask_empty(tick_broadcast_mask))
 340                         clockevents_shutdown(bc);
 341         }
 342
 343         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 344 }
 345
 346 void tick_suspend_broadcast(void)
 347 {
 348         struct clock_event_device *bc;
 349         unsigned long flags;
 350
 351         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 352
 353         bc = tick_broadcast_device.evtdev;
 354         if (bc)
 355                 clockevents_shutdown(bc);
 356
 357         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 358 }
 359
 360 int tick_resume_broadcast(void)
 361 {
 362         struct clock_event_device *bc;
 363         unsigned long flags;
 364         int broadcast = 0;
 365
 366         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 367
 368         bc = tick_broadcast_device.evtdev;
 369
 370         if (bc) {
 371                 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
 372
 373                 switch (tick_broadcast_device.mode) {
 374                 case TICKDEV_MODE_PERIODIC:
 375                         if (!cpumask_empty(tick_broadcast_mask))
 376                                 tick_broadcast_start_periodic(bc);
 377                         broadcast = cpumask_test_cpu(smp_processor_id(),
 378                                                      tick_broadcast_mask);
 379                         break;
 380                 case TICKDEV_MODE_ONESHOT:
 381                         if (!cpumask_empty(tick_broadcast_mask))
 382                                 broadcast = tick_resume_broadcast_oneshot(bc);
 383                         break;
 384                 }
 385         }
 386         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 387
 388         return broadcast;
 389 }
 390
 391
 392 #ifdef CONFIG_TICK_ONESHOT
 393
 394 static cpumask_var_t tick_broadcast_oneshot_mask;
 395
 396 /*
 397  * Exposed for debugging: see timer_list.c
 398  */
 399 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 400 {
 401         return tick_broadcast_oneshot_mask;
 402 }
 403
 404 static int tick_broadcast_set_event(struct clock_event_device *bc,
 405                                     ktime_t expires, int force)
 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, 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(tmpmask);
 446         now = ktime_get();
 447         /* Find all expired events */
 448         for_each_cpu(cpu, tick_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, 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(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(dev, 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_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
 516                         clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
 517                         if (dev->next_event.tv64 < bc->next_event.tv64)
 518                                 tick_broadcast_set_event(bc, dev->next_event, 1);
 519                 }
 520         } else {
 521                 if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
 522                         clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
 523                         if (dev->next_event.tv64 != KTIME_MAX)
 524                                 tick_program_event(dev->next_event, 1);
 525                 }
 526         }
 527         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 528 }
 529
 530 /*
 531  * Reset the one shot broadcast for a cpu
 532  *
 533  * Called with tick_broadcast_lock held
 534  */
 535 static void tick_broadcast_clear_oneshot(int cpu)
 536 {
 537         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 538 }
 539
 540 static void tick_broadcast_init_next_event(struct cpumask *mask,
 541                                            ktime_t expires)
 542 {
 543         struct tick_device *td;
 544         int cpu;
 545
 546         for_each_cpu(cpu, mask) {
 547                 td = &per_cpu(tick_cpu_device, cpu);
 548                 if (td->evtdev)
 549                         td->evtdev->next_event = expires;
 550         }
 551 }
 552
 553 /**
 554  * tick_broadcast_setup_oneshot - setup the broadcast device
 555  */
 556 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 557 {
 558         int cpu = smp_processor_id();
 559
 560         /* Set it up only once ! */
 561         if (bc->event_handler != tick_handle_oneshot_broadcast) {
 562                 int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
 563
 564                 bc->event_handler = tick_handle_oneshot_broadcast;
 565
 566                 /* Take the do_timer update */
 567                 tick_do_timer_cpu = cpu;
 568
 569                 /*
 570                  * We must be careful here. There might be other CPUs
 571                  * waiting for periodic broadcast. We need to set the
 572                  * oneshot_mask bits for those and program the
 573                  * broadcast device to fire.
 574                  */
 575                 cpumask_copy(tmpmask, tick_broadcast_mask);
 576                 cpumask_clear_cpu(cpu, tmpmask);
 577                 cpumask_or(tick_broadcast_oneshot_mask,
 578                            tick_broadcast_oneshot_mask, tmpmask);
 579
 580                 if (was_periodic && !cpumask_empty(tmpmask)) {
 581                         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
 582                         tick_broadcast_init_next_event(tmpmask,
 583                                                        tick_next_period);
 584                         tick_broadcast_set_event(bc, tick_next_period, 1);
 585                 } else
 586                         bc->next_event.tv64 = KTIME_MAX;
 587         } else {
 588                 /*
 589                  * The first cpu which switches to oneshot mode sets
 590                  * the bit for all other cpus which are in the general
 591                  * (periodic) broadcast mask. So the bit is set and
 592                  * would prevent the first broadcast enter after this
 593                  * to program the bc device.
 594                  */
 595                 tick_broadcast_clear_oneshot(cpu);
 596         }
 597 }
 598
 599 /*
 600  * Select oneshot operating mode for the broadcast device
 601  */
 602 void tick_broadcast_switch_to_oneshot(void)
 603 {
 604         struct clock_event_device *bc;
 605         unsigned long flags;
 606
 607         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 608
 609         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
 610         bc = tick_broadcast_device.evtdev;
 611         if (bc)
 612                 tick_broadcast_setup_oneshot(bc);
 613
 614         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 615 }
 616
 617
 618 /*
 619  * Remove a dead CPU from broadcasting
 620  */
 621 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
 622 {
 623         unsigned long flags;
 624         unsigned int cpu = *cpup;
 625
 626         raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 627
 628         /*
 629          * Clear the broadcast mask flag for the dead cpu, but do not
 630          * stop the broadcast device!
 631          */
 632         cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 633
 634         raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 635 }
 636
 637 /*
 638  * Check, whether the broadcast device is in one shot mode
 639  */
 640 int tick_broadcast_oneshot_active(void)
 641 {
 642         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 643 }
 644
 645 /*
 646  * Check whether the broadcast device supports oneshot.
 647  */
 648 bool tick_broadcast_oneshot_available(void)
 649 {
 650         struct clock_event_device *bc = tick_broadcast_device.evtdev;
 651
 652         return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 653 }
 654
 655 #endif
 656
 657 void __init tick_broadcast_init(void)
 658 {
 659         alloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
 660         alloc_cpumask_var(&tmpmask, GFP_NOWAIT);
 661 #ifdef CONFIG_TICK_ONESHOT
 662         alloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
 663 #endif
 664 }