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 struct tick_sched *tick_get_tick_sched(int cpu) 37 { 38 return &per_cpu(tick_cpu_sched, cpu); 39 } 40 41 /* 42 * Must be called with interrupts disabled ! 43 */ 44 static void tick_do_update_jiffies64(ktime_t now) 45 { 46 unsigned long ticks = 0; 47 ktime_t delta; 48 49 /* Reevalute with xtime_lock held */ 50 write_seqlock(&xtime_lock); 51 52 delta = ktime_sub(now, last_jiffies_update); 53 if (delta.tv64 >= tick_period.tv64) { 54 55 delta = ktime_sub(delta, tick_period); 56 last_jiffies_update = ktime_add(last_jiffies_update, 57 tick_period); 58 59 /* Slow path for long timeouts */ 60 if (unlikely(delta.tv64 >= tick_period.tv64)) { 61 s64 incr = ktime_to_ns(tick_period); 62 63 ticks = ktime_divns(delta, incr); 64 65 last_jiffies_update = ktime_add_ns(last_jiffies_update, 66 incr * ticks); 67 } 68 do_timer(++ticks); 69 } 70 write_sequnlock(&xtime_lock); 71 } 72 73 /* 74 * Initialize and return retrieve the jiffies update. 75 */ 76 static ktime_t tick_init_jiffy_update(void) 77 { 78 ktime_t period; 79 80 write_seqlock(&xtime_lock); 81 /* Did we start the jiffies update yet ? */ 82 if (last_jiffies_update.tv64 == 0) 83 last_jiffies_update = tick_next_period; 84 period = last_jiffies_update; 85 write_sequnlock(&xtime_lock); 86 return period; 87 } 88 89 /* 90 * NOHZ - aka dynamic tick functionality 91 */ 92 #ifdef CONFIG_NO_HZ 93 /* 94 * NO HZ enabled ? 95 */ 96 static int tick_nohz_enabled __read_mostly = 1; 97 98 /* 99 * Enable / Disable tickless mode 100 */ 101 static int __init setup_tick_nohz(char *str) 102 { 103 if (!strcmp(str, "off")) 104 tick_nohz_enabled = 0; 105 else if (!strcmp(str, "on")) 106 tick_nohz_enabled = 1; 107 else 108 return 0; 109 return 1; 110 } 111 112 __setup("nohz=", setup_tick_nohz); 113 114 /** 115 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 116 * 117 * Called from interrupt entry when the CPU was idle 118 * 119 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 120 * must be updated. Otherwise an interrupt handler could use a stale jiffy 121 * value. We do this unconditionally on any cpu, as we don't know whether the 122 * cpu, which has the update task assigned is in a long sleep. 123 */ 124 void tick_nohz_update_jiffies(void) 125 { 126 int cpu = smp_processor_id(); 127 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 128 unsigned long flags; 129 ktime_t now; 130 131 if (!ts->tick_stopped) 132 return; 133 134 cpu_clear(cpu, nohz_cpu_mask); 135 now = ktime_get(); 136 137 local_irq_save(flags); 138 tick_do_update_jiffies64(now); 139 local_irq_restore(flags); 140 } 141 142 /** 143 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task 144 * 145 * When the next event is more than a tick into the future, stop the idle tick 146 * Called either from the idle loop or from irq_exit() when an idle period was 147 * just interrupted by an interrupt which did not cause a reschedule. 148 */ 149 void tick_nohz_stop_sched_tick(void) 150 { 151 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; 152 struct tick_sched *ts; 153 ktime_t last_update, expires, now, delta; 154 int cpu; 155 156 local_irq_save(flags); 157 158 cpu = smp_processor_id(); 159 ts = &per_cpu(tick_cpu_sched, cpu); 160 161 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) 162 goto end; 163 164 if (need_resched()) 165 goto end; 166 167 cpu = smp_processor_id(); 168 if (unlikely(local_softirq_pending())) 169 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", 170 local_softirq_pending()); 171 172 now = ktime_get(); 173 /* 174 * When called from irq_exit we need to account the idle sleep time 175 * correctly. 176 */ 177 if (ts->tick_stopped) { 178 delta = ktime_sub(now, ts->idle_entrytime); 179 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 180 } 181 182 ts->idle_entrytime = now; 183 ts->idle_calls++; 184 185 /* Read jiffies and the time when jiffies were updated last */ 186 do { 187 seq = read_seqbegin(&xtime_lock); 188 last_update = last_jiffies_update; 189 last_jiffies = jiffies; 190 } while (read_seqretry(&xtime_lock, seq)); 191 192 /* Get the next timer wheel timer */ 193 next_jiffies = get_next_timer_interrupt(last_jiffies); 194 delta_jiffies = next_jiffies - last_jiffies; 195 196 if (rcu_needs_cpu(cpu)) 197 delta_jiffies = 1; 198 /* 199 * Do not stop the tick, if we are only one off 200 * or if the cpu is required for rcu 201 */ 202 if (!ts->tick_stopped && delta_jiffies == 1) 203 goto out; 204 205 /* Schedule the tick, if we are at least one jiffie off */ 206 if ((long)delta_jiffies >= 1) { 207 208 if (delta_jiffies > 1) 209 cpu_set(cpu, nohz_cpu_mask); 210 /* 211 * nohz_stop_sched_tick can be called several times before 212 * the nohz_restart_sched_tick is called. This happens when 213 * interrupts arrive which do not cause a reschedule. In the 214 * first call we save the current tick time, so we can restart 215 * the scheduler tick in nohz_restart_sched_tick. 216 */ 217 if (!ts->tick_stopped) { 218 ts->idle_tick = ts->sched_timer.expires; 219 ts->tick_stopped = 1; 220 ts->idle_jiffies = last_jiffies; 221 } 222 /* 223 * calculate the expiry time for the next timer wheel 224 * timer 225 */ 226 expires = ktime_add_ns(last_update, tick_period.tv64 * 227 delta_jiffies); 228 ts->idle_expires = expires; 229 ts->idle_sleeps++; 230 231 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 232 hrtimer_start(&ts->sched_timer, expires, 233 HRTIMER_MODE_ABS); 234 /* Check, if the timer was already in the past */ 235 if (hrtimer_active(&ts->sched_timer)) 236 goto out; 237 } else if(!tick_program_event(expires, 0)) 238 goto out; 239 /* 240 * We are past the event already. So we crossed a 241 * jiffie boundary. Update jiffies and raise the 242 * softirq. 243 */ 244 tick_do_update_jiffies64(ktime_get()); 245 cpu_clear(cpu, nohz_cpu_mask); 246 } 247 raise_softirq_irqoff(TIMER_SOFTIRQ); 248 out: 249 ts->next_jiffies = next_jiffies; 250 ts->last_jiffies = last_jiffies; 251 end: 252 local_irq_restore(flags); 253 } 254 255 /** 256 * nohz_restart_sched_tick - restart the idle tick from the idle task 257 * 258 * Restart the idle tick when the CPU is woken up from idle 259 */ 260 void tick_nohz_restart_sched_tick(void) 261 { 262 int cpu = smp_processor_id(); 263 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 264 unsigned long ticks; 265 ktime_t now, delta; 266 267 if (!ts->tick_stopped) 268 return; 269 270 /* Update jiffies first */ 271 now = ktime_get(); 272 273 local_irq_disable(); 274 tick_do_update_jiffies64(now); 275 cpu_clear(cpu, nohz_cpu_mask); 276 277 /* Account the idle time */ 278 delta = ktime_sub(now, ts->idle_entrytime); 279 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 280 281 /* 282 * We stopped the tick in idle. Update process times would miss the 283 * time we slept as update_process_times does only a 1 tick 284 * accounting. Enforce that this is accounted to idle ! 285 */ 286 ticks = jiffies - ts->idle_jiffies; 287 /* 288 * We might be one off. Do not randomly account a huge number of ticks! 289 */ 290 if (ticks && ticks < LONG_MAX) { 291 add_preempt_count(HARDIRQ_OFFSET); 292 account_system_time(current, HARDIRQ_OFFSET, 293 jiffies_to_cputime(ticks)); 294 sub_preempt_count(HARDIRQ_OFFSET); 295 } 296 297 /* 298 * Cancel the scheduled timer and restore the tick 299 */ 300 ts->tick_stopped = 0; 301 hrtimer_cancel(&ts->sched_timer); 302 ts->sched_timer.expires = ts->idle_tick; 303 304 while (1) { 305 /* Forward the time to expire in the future */ 306 hrtimer_forward(&ts->sched_timer, now, tick_period); 307 308 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 309 hrtimer_start(&ts->sched_timer, 310 ts->sched_timer.expires, 311 HRTIMER_MODE_ABS); 312 /* Check, if the timer was already in the past */ 313 if (hrtimer_active(&ts->sched_timer)) 314 break; 315 } else { 316 if (!tick_program_event(ts->sched_timer.expires, 0)) 317 break; 318 } 319 /* Update jiffies and reread time */ 320 tick_do_update_jiffies64(now); 321 now = ktime_get(); 322 } 323 local_irq_enable(); 324 } 325 326 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 327 { 328 hrtimer_forward(&ts->sched_timer, now, tick_period); 329 return tick_program_event(ts->sched_timer.expires, 0); 330 } 331 332 /* 333 * The nohz low res interrupt handler 334 */ 335 static void tick_nohz_handler(struct clock_event_device *dev) 336 { 337 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 338 struct pt_regs *regs = get_irq_regs(); 339 ktime_t now = ktime_get(); 340 341 dev->next_event.tv64 = KTIME_MAX; 342 343 /* Check, if the jiffies need an update */ 344 tick_do_update_jiffies64(now); 345 346 /* 347 * When we are idle and the tick is stopped, we have to touch 348 * the watchdog as we might not schedule for a really long 349 * time. This happens on complete idle SMP systems while 350 * waiting on the login prompt. We also increment the "start 351 * of idle" jiffy stamp so the idle accounting adjustment we 352 * do when we go busy again does not account too much ticks. 353 */ 354 if (ts->tick_stopped) { 355 touch_softlockup_watchdog(); 356 ts->idle_jiffies++; 357 } 358 359 update_process_times(user_mode(regs)); 360 profile_tick(CPU_PROFILING); 361 362 /* Do not restart, when we are in the idle loop */ 363 if (ts->tick_stopped) 364 return; 365 366 while (tick_nohz_reprogram(ts, now)) { 367 now = ktime_get(); 368 tick_do_update_jiffies64(now); 369 } 370 } 371 372 /** 373 * tick_nohz_switch_to_nohz - switch to nohz mode 374 */ 375 static void tick_nohz_switch_to_nohz(void) 376 { 377 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 378 ktime_t next; 379 380 if (!tick_nohz_enabled) 381 return; 382 383 local_irq_disable(); 384 if (tick_switch_to_oneshot(tick_nohz_handler)) { 385 local_irq_enable(); 386 return; 387 } 388 389 ts->nohz_mode = NOHZ_MODE_LOWRES; 390 391 /* 392 * Recycle the hrtimer in ts, so we can share the 393 * hrtimer_forward with the highres code. 394 */ 395 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 396 /* Get the next period */ 397 next = tick_init_jiffy_update(); 398 399 for (;;) { 400 ts->sched_timer.expires = next; 401 if (!tick_program_event(next, 0)) 402 break; 403 next = ktime_add(next, tick_period); 404 } 405 local_irq_enable(); 406 407 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", 408 smp_processor_id()); 409 } 410 411 #else 412 413 static inline void tick_nohz_switch_to_nohz(void) { } 414 415 #endif /* NO_HZ */ 416 417 /* 418 * High resolution timer specific code 419 */ 420 #ifdef CONFIG_HIGH_RES_TIMERS 421 /* 422 * We rearm the timer until we get disabled by the idle code 423 * Called with interrupts disabled and timer->base->cpu_base->lock held. 424 */ 425 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 426 { 427 struct tick_sched *ts = 428 container_of(timer, struct tick_sched, sched_timer); 429 struct hrtimer_cpu_base *base = timer->base->cpu_base; 430 struct pt_regs *regs = get_irq_regs(); 431 ktime_t now = ktime_get(); 432 433 /* Check, if the jiffies need an update */ 434 tick_do_update_jiffies64(now); 435 436 /* 437 * Do not call, when we are not in irq context and have 438 * no valid regs pointer 439 */ 440 if (regs) { 441 /* 442 * When we are idle and the tick is stopped, we have to touch 443 * the watchdog as we might not schedule for a really long 444 * time. This happens on complete idle SMP systems while 445 * waiting on the login prompt. We also increment the "start of 446 * idle" jiffy stamp so the idle accounting adjustment we do 447 * when we go busy again does not account too much ticks. 448 */ 449 if (ts->tick_stopped) { 450 touch_softlockup_watchdog(); 451 ts->idle_jiffies++; 452 } 453 /* 454 * update_process_times() might take tasklist_lock, hence 455 * drop the base lock. sched-tick hrtimers are per-CPU and 456 * never accessible by userspace APIs, so this is safe to do. 457 */ 458 spin_unlock(&base->lock); 459 update_process_times(user_mode(regs)); 460 profile_tick(CPU_PROFILING); 461 spin_lock(&base->lock); 462 } 463 464 /* Do not restart, when we are in the idle loop */ 465 if (ts->tick_stopped) 466 return HRTIMER_NORESTART; 467 468 hrtimer_forward(timer, now, tick_period); 469 470 return HRTIMER_RESTART; 471 } 472 473 /** 474 * tick_setup_sched_timer - setup the tick emulation timer 475 */ 476 void tick_setup_sched_timer(void) 477 { 478 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 479 ktime_t now = ktime_get(); 480 481 /* 482 * Emulate tick processing via per-CPU hrtimers: 483 */ 484 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 485 ts->sched_timer.function = tick_sched_timer; 486 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; 487 488 /* Get the next period */ 489 ts->sched_timer.expires = tick_init_jiffy_update(); 490 491 for (;;) { 492 hrtimer_forward(&ts->sched_timer, now, tick_period); 493 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires, 494 HRTIMER_MODE_ABS); 495 /* Check, if the timer was already in the past */ 496 if (hrtimer_active(&ts->sched_timer)) 497 break; 498 now = ktime_get(); 499 } 500 501 #ifdef CONFIG_NO_HZ 502 if (tick_nohz_enabled) 503 ts->nohz_mode = NOHZ_MODE_HIGHRES; 504 #endif 505 } 506 507 void tick_cancel_sched_timer(int cpu) 508 { 509 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 510 511 if (ts->sched_timer.base) 512 hrtimer_cancel(&ts->sched_timer); 513 ts->tick_stopped = 0; 514 ts->nohz_mode = NOHZ_MODE_INACTIVE; 515 } 516 #endif /* HIGH_RES_TIMERS */ 517 518 /** 519 * Async notification about clocksource changes 520 */ 521 void tick_clock_notify(void) 522 { 523 int cpu; 524 525 for_each_possible_cpu(cpu) 526 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 527 } 528 529 /* 530 * Async notification about clock event changes 531 */ 532 void tick_oneshot_notify(void) 533 { 534 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 535 536 set_bit(0, &ts->check_clocks); 537 } 538 539 /** 540 * Check, if a change happened, which makes oneshot possible. 541 * 542 * Called cyclic from the hrtimer softirq (driven by the timer 543 * softirq) allow_nohz signals, that we can switch into low-res nohz 544 * mode, because high resolution timers are disabled (either compile 545 * or runtime). 546 */ 547 int tick_check_oneshot_change(int allow_nohz) 548 { 549 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 550 551 if (!test_and_clear_bit(0, &ts->check_clocks)) 552 return 0; 553 554 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 555 return 0; 556 557 if (!timekeeping_is_continuous() || !tick_is_oneshot_available()) 558 return 0; 559 560 if (!allow_nohz) 561 return 1; 562 563 tick_nohz_switch_to_nohz(); 564 return 0; 565 } 566