Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame^] | 1 | /* |
| 2 | * linux/kernel/time/tick-sched.c |
| 3 | * |
| 4 | * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de> |
| 5 | * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar |
| 6 | * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner |
| 7 | * |
| 8 | * No idle tick implementation for low and high resolution timers |
| 9 | * |
| 10 | * Started by: Thomas Gleixner and Ingo Molnar |
| 11 | * |
| 12 | * For licencing details see 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/kernel_stat.h> |
| 19 | #include <linux/percpu.h> |
| 20 | #include <linux/profile.h> |
| 21 | #include <linux/sched.h> |
| 22 | #include <linux/tick.h> |
| 23 | |
| 24 | #include "tick-internal.h" |
| 25 | |
| 26 | /* |
| 27 | * Per cpu nohz control structure |
| 28 | */ |
| 29 | static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); |
| 30 | |
| 31 | /* |
| 32 | * The time, when the last jiffy update happened. Protected by xtime_lock. |
| 33 | */ |
| 34 | static ktime_t last_jiffies_update; |
| 35 | |
| 36 | /* |
| 37 | * Must be called with interrupts disabled ! |
| 38 | */ |
| 39 | static void tick_do_update_jiffies64(ktime_t now) |
| 40 | { |
| 41 | unsigned long ticks = 0; |
| 42 | ktime_t delta; |
| 43 | |
| 44 | /* Reevalute with xtime_lock held */ |
| 45 | write_seqlock(&xtime_lock); |
| 46 | |
| 47 | delta = ktime_sub(now, last_jiffies_update); |
| 48 | if (delta.tv64 >= tick_period.tv64) { |
| 49 | |
| 50 | delta = ktime_sub(delta, tick_period); |
| 51 | last_jiffies_update = ktime_add(last_jiffies_update, |
| 52 | tick_period); |
| 53 | |
| 54 | /* Slow path for long timeouts */ |
| 55 | if (unlikely(delta.tv64 >= tick_period.tv64)) { |
| 56 | s64 incr = ktime_to_ns(tick_period); |
| 57 | |
| 58 | ticks = ktime_divns(delta, incr); |
| 59 | |
| 60 | last_jiffies_update = ktime_add_ns(last_jiffies_update, |
| 61 | incr * ticks); |
| 62 | } |
| 63 | do_timer(++ticks); |
| 64 | } |
| 65 | write_sequnlock(&xtime_lock); |
| 66 | } |
| 67 | |
| 68 | /* |
| 69 | * Initialize and return retrieve the jiffies update. |
| 70 | */ |
| 71 | static ktime_t tick_init_jiffy_update(void) |
| 72 | { |
| 73 | ktime_t period; |
| 74 | |
| 75 | write_seqlock(&xtime_lock); |
| 76 | /* Did we start the jiffies update yet ? */ |
| 77 | if (last_jiffies_update.tv64 == 0) |
| 78 | last_jiffies_update = tick_next_period; |
| 79 | period = last_jiffies_update; |
| 80 | write_sequnlock(&xtime_lock); |
| 81 | return period; |
| 82 | } |
| 83 | |
| 84 | /* |
| 85 | * NOHZ - aka dynamic tick functionality |
| 86 | */ |
| 87 | #ifdef CONFIG_NO_HZ |
| 88 | /* |
| 89 | * NO HZ enabled ? |
| 90 | */ |
| 91 | static int tick_nohz_enabled __read_mostly = 1; |
| 92 | |
| 93 | /* |
| 94 | * Enable / Disable tickless mode |
| 95 | */ |
| 96 | static int __init setup_tick_nohz(char *str) |
| 97 | { |
| 98 | if (!strcmp(str, "off")) |
| 99 | tick_nohz_enabled = 0; |
| 100 | else if (!strcmp(str, "on")) |
| 101 | tick_nohz_enabled = 1; |
| 102 | else |
| 103 | return 0; |
| 104 | return 1; |
| 105 | } |
| 106 | |
| 107 | __setup("nohz=", setup_tick_nohz); |
| 108 | |
| 109 | /** |
| 110 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted |
| 111 | * |
| 112 | * Called from interrupt entry when the CPU was idle |
| 113 | * |
| 114 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies |
| 115 | * must be updated. Otherwise an interrupt handler could use a stale jiffy |
| 116 | * value. We do this unconditionally on any cpu, as we don't know whether the |
| 117 | * cpu, which has the update task assigned is in a long sleep. |
| 118 | */ |
| 119 | void tick_nohz_update_jiffies(void) |
| 120 | { |
| 121 | int cpu = smp_processor_id(); |
| 122 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 123 | unsigned long flags; |
| 124 | ktime_t now; |
| 125 | |
| 126 | if (!ts->tick_stopped) |
| 127 | return; |
| 128 | |
| 129 | cpu_clear(cpu, nohz_cpu_mask); |
| 130 | now = ktime_get(); |
| 131 | |
| 132 | local_irq_save(flags); |
| 133 | tick_do_update_jiffies64(now); |
| 134 | local_irq_restore(flags); |
| 135 | } |
| 136 | |
| 137 | /** |
| 138 | * tick_nohz_stop_sched_tick - stop the idle tick from the idle task |
| 139 | * |
| 140 | * When the next event is more than a tick into the future, stop the idle tick |
| 141 | * Called either from the idle loop or from irq_exit() when an idle period was |
| 142 | * just interrupted by an interrupt which did not cause a reschedule. |
| 143 | */ |
| 144 | void tick_nohz_stop_sched_tick(void) |
| 145 | { |
| 146 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; |
| 147 | struct tick_sched *ts; |
| 148 | ktime_t last_update, expires, now, delta; |
| 149 | int cpu; |
| 150 | |
| 151 | local_irq_save(flags); |
| 152 | |
| 153 | cpu = smp_processor_id(); |
| 154 | ts = &per_cpu(tick_cpu_sched, cpu); |
| 155 | |
| 156 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) |
| 157 | goto end; |
| 158 | |
| 159 | if (need_resched()) |
| 160 | goto end; |
| 161 | |
| 162 | cpu = smp_processor_id(); |
| 163 | BUG_ON(local_softirq_pending()); |
| 164 | |
| 165 | now = ktime_get(); |
| 166 | /* |
| 167 | * When called from irq_exit we need to account the idle sleep time |
| 168 | * correctly. |
| 169 | */ |
| 170 | if (ts->tick_stopped) { |
| 171 | delta = ktime_sub(now, ts->idle_entrytime); |
| 172 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); |
| 173 | } |
| 174 | |
| 175 | ts->idle_entrytime = now; |
| 176 | ts->idle_calls++; |
| 177 | |
| 178 | /* Read jiffies and the time when jiffies were updated last */ |
| 179 | do { |
| 180 | seq = read_seqbegin(&xtime_lock); |
| 181 | last_update = last_jiffies_update; |
| 182 | last_jiffies = jiffies; |
| 183 | } while (read_seqretry(&xtime_lock, seq)); |
| 184 | |
| 185 | /* Get the next timer wheel timer */ |
| 186 | next_jiffies = get_next_timer_interrupt(last_jiffies); |
| 187 | delta_jiffies = next_jiffies - last_jiffies; |
| 188 | |
| 189 | /* |
| 190 | * Do not stop the tick, if we are only one off |
| 191 | * or if the cpu is required for rcu |
| 192 | */ |
| 193 | if (!ts->tick_stopped && (delta_jiffies == 1 || rcu_needs_cpu(cpu))) |
| 194 | goto out; |
| 195 | |
| 196 | /* Schedule the tick, if we are at least one jiffie off */ |
| 197 | if ((long)delta_jiffies >= 1) { |
| 198 | |
| 199 | if (rcu_needs_cpu(cpu)) |
| 200 | delta_jiffies = 1; |
| 201 | else |
| 202 | cpu_set(cpu, nohz_cpu_mask); |
| 203 | /* |
| 204 | * nohz_stop_sched_tick can be called several times before |
| 205 | * the nohz_restart_sched_tick is called. This happens when |
| 206 | * interrupts arrive which do not cause a reschedule. In the |
| 207 | * first call we save the current tick time, so we can restart |
| 208 | * the scheduler tick in nohz_restart_sched_tick. |
| 209 | */ |
| 210 | if (!ts->tick_stopped) { |
| 211 | ts->idle_tick = ts->sched_timer.expires; |
| 212 | ts->tick_stopped = 1; |
| 213 | ts->idle_jiffies = last_jiffies; |
| 214 | } |
| 215 | /* |
| 216 | * calculate the expiry time for the next timer wheel |
| 217 | * timer |
| 218 | */ |
| 219 | expires = ktime_add_ns(last_update, tick_period.tv64 * |
| 220 | delta_jiffies); |
| 221 | ts->idle_expires = expires; |
| 222 | ts->idle_sleeps++; |
| 223 | |
| 224 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
| 225 | hrtimer_start(&ts->sched_timer, expires, |
| 226 | HRTIMER_MODE_ABS); |
| 227 | /* Check, if the timer was already in the past */ |
| 228 | if (hrtimer_active(&ts->sched_timer)) |
| 229 | goto out; |
| 230 | } else if(!tick_program_event(expires, 0)) |
| 231 | goto out; |
| 232 | /* |
| 233 | * We are past the event already. So we crossed a |
| 234 | * jiffie boundary. Update jiffies and raise the |
| 235 | * softirq. |
| 236 | */ |
| 237 | tick_do_update_jiffies64(ktime_get()); |
| 238 | cpu_clear(cpu, nohz_cpu_mask); |
| 239 | } |
| 240 | raise_softirq_irqoff(TIMER_SOFTIRQ); |
| 241 | out: |
| 242 | ts->next_jiffies = next_jiffies; |
| 243 | ts->last_jiffies = last_jiffies; |
| 244 | end: |
| 245 | local_irq_restore(flags); |
| 246 | } |
| 247 | |
| 248 | /** |
| 249 | * nohz_restart_sched_tick - restart the idle tick from the idle task |
| 250 | * |
| 251 | * Restart the idle tick when the CPU is woken up from idle |
| 252 | */ |
| 253 | void tick_nohz_restart_sched_tick(void) |
| 254 | { |
| 255 | int cpu = smp_processor_id(); |
| 256 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 257 | unsigned long ticks; |
| 258 | ktime_t now, delta; |
| 259 | |
| 260 | if (!ts->tick_stopped) |
| 261 | return; |
| 262 | |
| 263 | /* Update jiffies first */ |
| 264 | now = ktime_get(); |
| 265 | |
| 266 | local_irq_disable(); |
| 267 | tick_do_update_jiffies64(now); |
| 268 | cpu_clear(cpu, nohz_cpu_mask); |
| 269 | |
| 270 | /* Account the idle time */ |
| 271 | delta = ktime_sub(now, ts->idle_entrytime); |
| 272 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); |
| 273 | |
| 274 | /* |
| 275 | * We stopped the tick in idle. Update process times would miss the |
| 276 | * time we slept as update_process_times does only a 1 tick |
| 277 | * accounting. Enforce that this is accounted to idle ! |
| 278 | */ |
| 279 | ticks = jiffies - ts->idle_jiffies; |
| 280 | /* |
| 281 | * We might be one off. Do not randomly account a huge number of ticks! |
| 282 | */ |
| 283 | if (ticks && ticks < LONG_MAX) { |
| 284 | add_preempt_count(HARDIRQ_OFFSET); |
| 285 | account_system_time(current, HARDIRQ_OFFSET, |
| 286 | jiffies_to_cputime(ticks)); |
| 287 | sub_preempt_count(HARDIRQ_OFFSET); |
| 288 | } |
| 289 | |
| 290 | /* |
| 291 | * Cancel the scheduled timer and restore the tick |
| 292 | */ |
| 293 | ts->tick_stopped = 0; |
| 294 | hrtimer_cancel(&ts->sched_timer); |
| 295 | ts->sched_timer.expires = ts->idle_tick; |
| 296 | |
| 297 | while (1) { |
| 298 | /* Forward the time to expire in the future */ |
| 299 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
| 300 | |
| 301 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
| 302 | hrtimer_start(&ts->sched_timer, |
| 303 | ts->sched_timer.expires, |
| 304 | HRTIMER_MODE_ABS); |
| 305 | /* Check, if the timer was already in the past */ |
| 306 | if (hrtimer_active(&ts->sched_timer)) |
| 307 | break; |
| 308 | } else { |
| 309 | if (!tick_program_event(ts->sched_timer.expires, 0)) |
| 310 | break; |
| 311 | } |
| 312 | /* Update jiffies and reread time */ |
| 313 | tick_do_update_jiffies64(now); |
| 314 | now = ktime_get(); |
| 315 | } |
| 316 | local_irq_enable(); |
| 317 | } |
| 318 | |
| 319 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) |
| 320 | { |
| 321 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
| 322 | return tick_program_event(ts->sched_timer.expires, 0); |
| 323 | } |
| 324 | |
| 325 | /* |
| 326 | * The nohz low res interrupt handler |
| 327 | */ |
| 328 | static void tick_nohz_handler(struct clock_event_device *dev) |
| 329 | { |
| 330 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 331 | struct pt_regs *regs = get_irq_regs(); |
| 332 | ktime_t now = ktime_get(); |
| 333 | |
| 334 | dev->next_event.tv64 = KTIME_MAX; |
| 335 | |
| 336 | /* Check, if the jiffies need an update */ |
| 337 | tick_do_update_jiffies64(now); |
| 338 | |
| 339 | /* |
| 340 | * When we are idle and the tick is stopped, we have to touch |
| 341 | * the watchdog as we might not schedule for a really long |
| 342 | * time. This happens on complete idle SMP systems while |
| 343 | * waiting on the login prompt. We also increment the "start |
| 344 | * of idle" jiffy stamp so the idle accounting adjustment we |
| 345 | * do when we go busy again does not account too much ticks. |
| 346 | */ |
| 347 | if (ts->tick_stopped) { |
| 348 | touch_softlockup_watchdog(); |
| 349 | ts->idle_jiffies++; |
| 350 | } |
| 351 | |
| 352 | update_process_times(user_mode(regs)); |
| 353 | profile_tick(CPU_PROFILING); |
| 354 | |
| 355 | /* Do not restart, when we are in the idle loop */ |
| 356 | if (ts->tick_stopped) |
| 357 | return; |
| 358 | |
| 359 | while (tick_nohz_reprogram(ts, now)) { |
| 360 | now = ktime_get(); |
| 361 | tick_do_update_jiffies64(now); |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | /** |
| 366 | * tick_nohz_switch_to_nohz - switch to nohz mode |
| 367 | */ |
| 368 | static void tick_nohz_switch_to_nohz(void) |
| 369 | { |
| 370 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 371 | ktime_t next; |
| 372 | |
| 373 | if (!tick_nohz_enabled) |
| 374 | return; |
| 375 | |
| 376 | local_irq_disable(); |
| 377 | if (tick_switch_to_oneshot(tick_nohz_handler)) { |
| 378 | local_irq_enable(); |
| 379 | return; |
| 380 | } |
| 381 | |
| 382 | ts->nohz_mode = NOHZ_MODE_LOWRES; |
| 383 | |
| 384 | /* |
| 385 | * Recycle the hrtimer in ts, so we can share the |
| 386 | * hrtimer_forward with the highres code. |
| 387 | */ |
| 388 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); |
| 389 | /* Get the next period */ |
| 390 | next = tick_init_jiffy_update(); |
| 391 | |
| 392 | for (;;) { |
| 393 | ts->sched_timer.expires = next; |
| 394 | if (!tick_program_event(next, 0)) |
| 395 | break; |
| 396 | next = ktime_add(next, tick_period); |
| 397 | } |
| 398 | local_irq_enable(); |
| 399 | |
| 400 | printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", |
| 401 | smp_processor_id()); |
| 402 | } |
| 403 | |
| 404 | #else |
| 405 | |
| 406 | static inline void tick_nohz_switch_to_nohz(void) { } |
| 407 | |
| 408 | #endif /* NO_HZ */ |
| 409 | |
| 410 | /* |
| 411 | * High resolution timer specific code |
| 412 | */ |
| 413 | #ifdef CONFIG_HIGH_RES_TIMERS |
| 414 | /* |
| 415 | * We rearm the timer until we get disabled by the idle code |
| 416 | * Called with interrupts disabled and timer->base->cpu_base->lock held. |
| 417 | */ |
| 418 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) |
| 419 | { |
| 420 | struct tick_sched *ts = |
| 421 | container_of(timer, struct tick_sched, sched_timer); |
| 422 | struct hrtimer_cpu_base *base = timer->base->cpu_base; |
| 423 | struct pt_regs *regs = get_irq_regs(); |
| 424 | ktime_t now = ktime_get(); |
| 425 | |
| 426 | /* Check, if the jiffies need an update */ |
| 427 | tick_do_update_jiffies64(now); |
| 428 | |
| 429 | /* |
| 430 | * Do not call, when we are not in irq context and have |
| 431 | * no valid regs pointer |
| 432 | */ |
| 433 | if (regs) { |
| 434 | /* |
| 435 | * When we are idle and the tick is stopped, we have to touch |
| 436 | * the watchdog as we might not schedule for a really long |
| 437 | * time. This happens on complete idle SMP systems while |
| 438 | * waiting on the login prompt. We also increment the "start of |
| 439 | * idle" jiffy stamp so the idle accounting adjustment we do |
| 440 | * when we go busy again does not account too much ticks. |
| 441 | */ |
| 442 | if (ts->tick_stopped) { |
| 443 | touch_softlockup_watchdog(); |
| 444 | ts->idle_jiffies++; |
| 445 | } |
| 446 | /* |
| 447 | * update_process_times() might take tasklist_lock, hence |
| 448 | * drop the base lock. sched-tick hrtimers are per-CPU and |
| 449 | * never accessible by userspace APIs, so this is safe to do. |
| 450 | */ |
| 451 | spin_unlock(&base->lock); |
| 452 | update_process_times(user_mode(regs)); |
| 453 | profile_tick(CPU_PROFILING); |
| 454 | spin_lock(&base->lock); |
| 455 | } |
| 456 | |
| 457 | /* Do not restart, when we are in the idle loop */ |
| 458 | if (ts->tick_stopped) |
| 459 | return HRTIMER_NORESTART; |
| 460 | |
| 461 | hrtimer_forward(timer, now, tick_period); |
| 462 | |
| 463 | return HRTIMER_RESTART; |
| 464 | } |
| 465 | |
| 466 | /** |
| 467 | * tick_setup_sched_timer - setup the tick emulation timer |
| 468 | */ |
| 469 | void tick_setup_sched_timer(void) |
| 470 | { |
| 471 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 472 | ktime_t now = ktime_get(); |
| 473 | |
| 474 | /* |
| 475 | * Emulate tick processing via per-CPU hrtimers: |
| 476 | */ |
| 477 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); |
| 478 | ts->sched_timer.function = tick_sched_timer; |
| 479 | ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; |
| 480 | |
| 481 | /* Get the next period */ |
| 482 | ts->sched_timer.expires = tick_init_jiffy_update(); |
| 483 | |
| 484 | for (;;) { |
| 485 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
| 486 | hrtimer_start(&ts->sched_timer, ts->sched_timer.expires, |
| 487 | HRTIMER_MODE_ABS); |
| 488 | /* Check, if the timer was already in the past */ |
| 489 | if (hrtimer_active(&ts->sched_timer)) |
| 490 | break; |
| 491 | now = ktime_get(); |
| 492 | } |
| 493 | |
| 494 | #ifdef CONFIG_NO_HZ |
| 495 | if (tick_nohz_enabled) |
| 496 | ts->nohz_mode = NOHZ_MODE_HIGHRES; |
| 497 | #endif |
| 498 | } |
| 499 | |
| 500 | void tick_cancel_sched_timer(int cpu) |
| 501 | { |
| 502 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 503 | |
| 504 | if (ts->sched_timer.base) |
| 505 | hrtimer_cancel(&ts->sched_timer); |
| 506 | ts->tick_stopped = 0; |
| 507 | ts->nohz_mode = NOHZ_MODE_INACTIVE; |
| 508 | } |
| 509 | #endif /* HIGH_RES_TIMERS */ |
| 510 | |
| 511 | /** |
| 512 | * Async notification about clocksource changes |
| 513 | */ |
| 514 | void tick_clock_notify(void) |
| 515 | { |
| 516 | int cpu; |
| 517 | |
| 518 | for_each_possible_cpu(cpu) |
| 519 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); |
| 520 | } |
| 521 | |
| 522 | /* |
| 523 | * Async notification about clock event changes |
| 524 | */ |
| 525 | void tick_oneshot_notify(void) |
| 526 | { |
| 527 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 528 | |
| 529 | set_bit(0, &ts->check_clocks); |
| 530 | } |
| 531 | |
| 532 | /** |
| 533 | * Check, if a change happened, which makes oneshot possible. |
| 534 | * |
| 535 | * Called cyclic from the hrtimer softirq (driven by the timer |
| 536 | * softirq) allow_nohz signals, that we can switch into low-res nohz |
| 537 | * mode, because high resolution timers are disabled (either compile |
| 538 | * or runtime). |
| 539 | */ |
| 540 | int tick_check_oneshot_change(int allow_nohz) |
| 541 | { |
| 542 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 543 | |
| 544 | if (!test_and_clear_bit(0, &ts->check_clocks)) |
| 545 | return 0; |
| 546 | |
| 547 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) |
| 548 | return 0; |
| 549 | |
| 550 | if (!timekeeping_is_continuous() || !tick_is_oneshot_available()) |
| 551 | return 0; |
| 552 | |
| 553 | if (!allow_nohz) |
| 554 | return 1; |
| 555 | |
| 556 | tick_nohz_switch_to_nohz(); |
| 557 | return 0; |
| 558 | } |