lh | 9ed821d | 2023-04-07 01:36:19 -0700 | [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 | * Distribute under GPLv2. |
| 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/module.h> |
| 23 | |
| 24 | #include <asm/irq_regs.h> |
| 25 | |
| 26 | #include "tick-internal.h" |
| 27 | |
| 28 | #ifdef CONFIG_SINGLECORE |
| 29 | #ifndef CONFIG_SYSTEM_RECOVERY |
| 30 | #ifndef USE_CPPS_KO |
| 31 | extern void linux_oss_tick_timer_function(void); |
| 32 | #endif |
| 33 | #endif |
| 34 | #endif |
| 35 | |
| 36 | /* |
| 37 | * Per cpu nohz control structure |
| 38 | */ |
| 39 | static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); |
| 40 | |
| 41 | /* |
| 42 | * The time, when the last jiffy update happened. Protected by xtime_lock. |
| 43 | */ |
| 44 | static ktime_t last_jiffies_update; |
| 45 | |
| 46 | struct tick_sched *tick_get_tick_sched(int cpu) |
| 47 | { |
| 48 | return &per_cpu(tick_cpu_sched, cpu); |
| 49 | } |
| 50 | |
| 51 | /* |
| 52 | * Must be called with interrupts disabled ! |
| 53 | */ |
| 54 | static void tick_do_update_jiffies64(ktime_t now) |
| 55 | { |
| 56 | unsigned long ticks = 0; |
| 57 | ktime_t delta; |
| 58 | |
| 59 | /* |
| 60 | * Do a quick check without holding xtime_lock: |
| 61 | */ |
| 62 | delta = ktime_sub(now, last_jiffies_update); |
| 63 | if (delta.tv64 < tick_period.tv64) |
| 64 | return; |
| 65 | |
| 66 | /* Reevalute with xtime_lock held */ |
| 67 | raw_spin_lock(&xtime_lock); |
| 68 | write_seqcount_begin(&xtime_seq); |
| 69 | |
| 70 | delta = ktime_sub(now, last_jiffies_update); |
| 71 | if (delta.tv64 >= tick_period.tv64) { |
| 72 | |
| 73 | delta = ktime_sub(delta, tick_period); |
| 74 | last_jiffies_update = ktime_add(last_jiffies_update, |
| 75 | tick_period); |
| 76 | |
| 77 | /* Slow path for long timeouts */ |
| 78 | if (unlikely(delta.tv64 >= tick_period.tv64)) { |
| 79 | s64 incr = ktime_to_ns(tick_period); |
| 80 | |
| 81 | ticks = ktime_divns(delta, incr); |
| 82 | |
| 83 | last_jiffies_update = ktime_add_ns(last_jiffies_update, |
| 84 | incr * ticks); |
| 85 | } |
| 86 | do_timer(++ticks); |
| 87 | |
| 88 | /* Keep the tick_next_period variable up to date */ |
| 89 | tick_next_period = ktime_add(last_jiffies_update, tick_period); |
| 90 | } |
| 91 | write_seqcount_end(&xtime_seq); |
| 92 | raw_spin_unlock(&xtime_lock); |
| 93 | } |
| 94 | |
| 95 | /* |
| 96 | * Initialize and return retrieve the jiffies update. |
| 97 | */ |
| 98 | static ktime_t tick_init_jiffy_update(void) |
| 99 | { |
| 100 | ktime_t period; |
| 101 | |
| 102 | raw_spin_lock(&xtime_lock); |
| 103 | write_seqcount_begin(&xtime_seq); |
| 104 | /* Did we start the jiffies update yet ? */ |
| 105 | if (last_jiffies_update.tv64 == 0) |
| 106 | last_jiffies_update = tick_next_period; |
| 107 | period = last_jiffies_update; |
| 108 | write_seqcount_end(&xtime_seq); |
| 109 | raw_spin_unlock(&xtime_lock); |
| 110 | return period; |
| 111 | } |
| 112 | |
| 113 | /* |
| 114 | * NOHZ - aka dynamic tick functionality |
| 115 | */ |
| 116 | #ifdef CONFIG_NO_HZ |
| 117 | /* |
| 118 | * NO HZ enabled ? |
| 119 | */ |
| 120 | static int tick_nohz_enabled __read_mostly = 1; |
| 121 | |
| 122 | /* |
| 123 | * Enable / Disable tickless mode |
| 124 | */ |
| 125 | static int __init setup_tick_nohz(char *str) |
| 126 | { |
| 127 | if (!strcmp(str, "off")) |
| 128 | tick_nohz_enabled = 0; |
| 129 | else if (!strcmp(str, "on")) |
| 130 | tick_nohz_enabled = 1; |
| 131 | else |
| 132 | return 0; |
| 133 | return 1; |
| 134 | } |
| 135 | |
| 136 | __setup("nohz=", setup_tick_nohz); |
| 137 | |
| 138 | /** |
| 139 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted |
| 140 | * |
| 141 | * Called from interrupt entry when the CPU was idle |
| 142 | * |
| 143 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies |
| 144 | * must be updated. Otherwise an interrupt handler could use a stale jiffy |
| 145 | * value. We do this unconditionally on any cpu, as we don't know whether the |
| 146 | * cpu, which has the update task assigned is in a long sleep. |
| 147 | */ |
| 148 | static void tick_nohz_update_jiffies(ktime_t now) |
| 149 | { |
| 150 | int cpu = smp_processor_id(); |
| 151 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 152 | unsigned long flags; |
| 153 | |
| 154 | ts->idle_waketime = now; |
| 155 | |
| 156 | local_irq_save(flags); |
| 157 | tick_do_update_jiffies64(now); |
| 158 | local_irq_restore(flags); |
| 159 | |
| 160 | touch_softlockup_watchdog(); |
| 161 | } |
| 162 | |
| 163 | /* |
| 164 | * Updates the per cpu time idle statistics counters |
| 165 | */ |
| 166 | static void |
| 167 | update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) |
| 168 | { |
| 169 | ktime_t delta; |
| 170 | |
| 171 | if (ts->idle_active) { |
| 172 | delta = ktime_sub(now, ts->idle_entrytime); |
| 173 | if (nr_iowait_cpu(cpu) > 0) |
| 174 | ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); |
| 175 | else |
| 176 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); |
| 177 | ts->idle_entrytime = now; |
| 178 | } |
| 179 | |
| 180 | if (last_update_time) |
| 181 | *last_update_time = ktime_to_us(now); |
| 182 | |
| 183 | } |
| 184 | |
| 185 | static void tick_nohz_stop_idle(int cpu, ktime_t now) |
| 186 | { |
| 187 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 188 | |
| 189 | update_ts_time_stats(cpu, ts, now, NULL); |
| 190 | ts->idle_active = 0; |
| 191 | |
| 192 | sched_clock_idle_wakeup_event(0); |
| 193 | } |
| 194 | |
| 195 | static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts) |
| 196 | { |
| 197 | ktime_t now = ktime_get(); |
| 198 | |
| 199 | ts->idle_entrytime = now; |
| 200 | ts->idle_active = 1; |
| 201 | sched_clock_idle_sleep_event(); |
| 202 | return now; |
| 203 | } |
| 204 | |
| 205 | /** |
| 206 | * get_cpu_idle_time_us - get the total idle time of a cpu |
| 207 | * @cpu: CPU number to query |
| 208 | * @last_update_time: variable to store update time in. Do not update |
| 209 | * counters if NULL. |
| 210 | * |
| 211 | * Return the cummulative idle time (since boot) for a given |
| 212 | * CPU, in microseconds. |
| 213 | * |
| 214 | * This time is measured via accounting rather than sampling, |
| 215 | * and is as accurate as ktime_get() is. |
| 216 | * |
| 217 | * This function returns -1 if NOHZ is not enabled. |
| 218 | */ |
| 219 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
| 220 | { |
| 221 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 222 | ktime_t now, idle; |
| 223 | |
| 224 | if (!tick_nohz_enabled) |
| 225 | return -1; |
| 226 | |
| 227 | now = ktime_get(); |
| 228 | if (last_update_time) { |
| 229 | update_ts_time_stats(cpu, ts, now, last_update_time); |
| 230 | idle = ts->idle_sleeptime; |
| 231 | } else { |
| 232 | if (ts->idle_active && !nr_iowait_cpu(cpu)) { |
| 233 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); |
| 234 | |
| 235 | idle = ktime_add(ts->idle_sleeptime, delta); |
| 236 | } else { |
| 237 | idle = ts->idle_sleeptime; |
| 238 | } |
| 239 | } |
| 240 | |
| 241 | return ktime_to_us(idle); |
| 242 | |
| 243 | } |
| 244 | EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); |
| 245 | |
| 246 | /** |
| 247 | * get_cpu_iowait_time_us - get the total iowait time of a cpu |
| 248 | * @cpu: CPU number to query |
| 249 | * @last_update_time: variable to store update time in. Do not update |
| 250 | * counters if NULL. |
| 251 | * |
| 252 | * Return the cummulative iowait time (since boot) for a given |
| 253 | * CPU, in microseconds. |
| 254 | * |
| 255 | * This time is measured via accounting rather than sampling, |
| 256 | * and is as accurate as ktime_get() is. |
| 257 | * |
| 258 | * This function returns -1 if NOHZ is not enabled. |
| 259 | */ |
| 260 | u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) |
| 261 | { |
| 262 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 263 | ktime_t now, iowait; |
| 264 | |
| 265 | if (!tick_nohz_enabled) |
| 266 | return -1; |
| 267 | |
| 268 | now = ktime_get(); |
| 269 | if (last_update_time) { |
| 270 | update_ts_time_stats(cpu, ts, now, last_update_time); |
| 271 | iowait = ts->iowait_sleeptime; |
| 272 | } else { |
| 273 | if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { |
| 274 | ktime_t delta = ktime_sub(now, ts->idle_entrytime); |
| 275 | |
| 276 | iowait = ktime_add(ts->iowait_sleeptime, delta); |
| 277 | } else { |
| 278 | iowait = ts->iowait_sleeptime; |
| 279 | } |
| 280 | } |
| 281 | |
| 282 | return ktime_to_us(iowait); |
| 283 | } |
| 284 | EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); |
| 285 | |
| 286 | static void tick_nohz_stop_sched_tick(struct tick_sched *ts) |
| 287 | { |
| 288 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; |
| 289 | ktime_t last_update, expires, now; |
| 290 | struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; |
| 291 | u64 time_delta; |
| 292 | int cpu; |
| 293 | |
| 294 | cpu = smp_processor_id(); |
| 295 | ts = &per_cpu(tick_cpu_sched, cpu); |
| 296 | |
| 297 | now = tick_nohz_start_idle(cpu, ts); |
| 298 | |
| 299 | /* |
| 300 | * If this cpu is offline and it is the one which updates |
| 301 | * jiffies, then give up the assignment and let it be taken by |
| 302 | * the cpu which runs the tick timer next. If we don't drop |
| 303 | * this here the jiffies might be stale and do_timer() never |
| 304 | * invoked. |
| 305 | */ |
| 306 | if (unlikely(!cpu_online(cpu))) { |
| 307 | if (cpu == tick_do_timer_cpu) |
| 308 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
| 309 | } |
| 310 | |
| 311 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { |
| 312 | ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ }; |
| 313 | return; |
| 314 | } |
| 315 | |
| 316 | if (need_resched()) |
| 317 | return; |
| 318 | |
| 319 | if (unlikely(local_softirq_pending() && cpu_online(cpu))) { |
| 320 | softirq_check_pending_idle(); |
| 321 | return; |
| 322 | } |
| 323 | |
| 324 | ts->idle_calls++; |
| 325 | /* Read jiffies and the time when jiffies were updated last */ |
| 326 | do { |
| 327 | seq = read_seqcount_begin(&xtime_seq); |
| 328 | last_update = last_jiffies_update; |
| 329 | last_jiffies = jiffies; |
| 330 | time_delta = timekeeping_max_deferment(); |
| 331 | } while (read_seqcount_retry(&xtime_seq, seq)); |
| 332 | |
| 333 | if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) || |
| 334 | arch_needs_cpu(cpu)) { |
| 335 | next_jiffies = last_jiffies + 1; |
| 336 | delta_jiffies = 1; |
| 337 | } else { |
| 338 | /* Get the next timer wheel timer */ |
| 339 | next_jiffies = get_next_timer_interrupt(last_jiffies); |
| 340 | delta_jiffies = next_jiffies - last_jiffies; |
| 341 | } |
| 342 | /* |
| 343 | * Do not stop the tick, if we are only one off |
| 344 | * or if the cpu is required for rcu |
| 345 | */ |
| 346 | if (!ts->tick_stopped && delta_jiffies == 1) |
| 347 | goto out; |
| 348 | |
| 349 | /* Schedule the tick, if we are at least one jiffie off */ |
| 350 | if ((long)delta_jiffies >= 1) { |
| 351 | |
| 352 | /* |
| 353 | * If this cpu is the one which updates jiffies, then |
| 354 | * give up the assignment and let it be taken by the |
| 355 | * cpu which runs the tick timer next, which might be |
| 356 | * this cpu as well. If we don't drop this here the |
| 357 | * jiffies might be stale and do_timer() never |
| 358 | * invoked. Keep track of the fact that it was the one |
| 359 | * which had the do_timer() duty last. If this cpu is |
| 360 | * the one which had the do_timer() duty last, we |
| 361 | * limit the sleep time to the timekeeping |
| 362 | * max_deferement value which we retrieved |
| 363 | * above. Otherwise we can sleep as long as we want. |
| 364 | */ |
| 365 | if (cpu == tick_do_timer_cpu) { |
| 366 | tick_do_timer_cpu = TICK_DO_TIMER_NONE; |
| 367 | ts->do_timer_last = 1; |
| 368 | } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { |
| 369 | time_delta = KTIME_MAX; |
| 370 | ts->do_timer_last = 0; |
| 371 | } else if (!ts->do_timer_last) { |
| 372 | time_delta = KTIME_MAX; |
| 373 | } |
| 374 | |
| 375 | /* |
| 376 | * calculate the expiry time for the next timer wheel |
| 377 | * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals |
| 378 | * that there is no timer pending or at least extremely |
| 379 | * far into the future (12 days for HZ=1000). In this |
| 380 | * case we set the expiry to the end of time. |
| 381 | */ |
| 382 | if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { |
| 383 | /* |
| 384 | * Calculate the time delta for the next timer event. |
| 385 | * If the time delta exceeds the maximum time delta |
| 386 | * permitted by the current clocksource then adjust |
| 387 | * the time delta accordingly to ensure the |
| 388 | * clocksource does not wrap. |
| 389 | */ |
| 390 | time_delta = min_t(u64, time_delta, |
| 391 | tick_period.tv64 * delta_jiffies); |
| 392 | } |
| 393 | |
| 394 | if (time_delta < KTIME_MAX) |
| 395 | expires = ktime_add_ns(last_update, time_delta); |
| 396 | else |
| 397 | expires.tv64 = KTIME_MAX; |
| 398 | |
| 399 | /* Skip reprogram of event if its not changed */ |
| 400 | if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) |
| 401 | goto out; |
| 402 | |
| 403 | /* |
| 404 | * nohz_stop_sched_tick can be called several times before |
| 405 | * the nohz_restart_sched_tick is called. This happens when |
| 406 | * interrupts arrive which do not cause a reschedule. In the |
| 407 | * first call we save the current tick time, so we can restart |
| 408 | * the scheduler tick in nohz_restart_sched_tick. |
| 409 | */ |
| 410 | if (!ts->tick_stopped) { |
| 411 | select_nohz_load_balancer(1); |
| 412 | calc_load_enter_idle(); |
| 413 | |
| 414 | ts->idle_tick = hrtimer_get_expires(&ts->sched_timer); |
| 415 | ts->tick_stopped = 1; |
| 416 | ts->idle_jiffies = last_jiffies; |
| 417 | } |
| 418 | |
| 419 | ts->idle_sleeps++; |
| 420 | |
| 421 | /* Mark expires */ |
| 422 | ts->idle_expires = expires; |
| 423 | |
| 424 | /* |
| 425 | * If the expiration time == KTIME_MAX, then |
| 426 | * in this case we simply stop the tick timer. |
| 427 | */ |
| 428 | if (unlikely(expires.tv64 == KTIME_MAX)) { |
| 429 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
| 430 | hrtimer_cancel(&ts->sched_timer); |
| 431 | goto out; |
| 432 | } |
| 433 | |
| 434 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
| 435 | hrtimer_start(&ts->sched_timer, expires, |
| 436 | HRTIMER_MODE_ABS_PINNED); |
| 437 | /* Check, if the timer was already in the past */ |
| 438 | if (hrtimer_active(&ts->sched_timer)) |
| 439 | goto out; |
| 440 | } else if (!tick_program_event(expires, 0)) |
| 441 | goto out; |
| 442 | /* |
| 443 | * We are past the event already. So we crossed a |
| 444 | * jiffie boundary. Update jiffies and raise the |
| 445 | * softirq. |
| 446 | */ |
| 447 | tick_do_update_jiffies64(ktime_get()); |
| 448 | } |
| 449 | raise_softirq_irqoff(TIMER_SOFTIRQ); |
| 450 | out: |
| 451 | ts->next_jiffies = next_jiffies; |
| 452 | ts->last_jiffies = last_jiffies; |
| 453 | ts->sleep_length = ktime_sub(dev->next_event, now); |
| 454 | } |
| 455 | |
| 456 | /** |
| 457 | * tick_nohz_idle_enter - stop the idle tick from the idle task |
| 458 | * |
| 459 | * When the next event is more than a tick into the future, stop the idle tick |
| 460 | * Called when we start the idle loop. |
| 461 | * |
| 462 | * The arch is responsible of calling: |
| 463 | * |
| 464 | * - rcu_idle_enter() after its last use of RCU before the CPU is put |
| 465 | * to sleep. |
| 466 | * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. |
| 467 | */ |
| 468 | void tick_nohz_idle_enter(void) |
| 469 | { |
| 470 | struct tick_sched *ts; |
| 471 | |
| 472 | WARN_ON_ONCE(irqs_disabled()); |
| 473 | |
| 474 | /* |
| 475 | * Update the idle state in the scheduler domain hierarchy |
| 476 | * when tick_nohz_stop_sched_tick() is called from the idle loop. |
| 477 | * State will be updated to busy during the first busy tick after |
| 478 | * exiting idle. |
| 479 | */ |
| 480 | set_cpu_sd_state_idle(); |
| 481 | |
| 482 | local_irq_disable(); |
| 483 | |
| 484 | ts = &__get_cpu_var(tick_cpu_sched); |
| 485 | /* |
| 486 | * set ts->inidle unconditionally. even if the system did not |
| 487 | * switch to nohz mode the cpu frequency governers rely on the |
| 488 | * update of the idle time accounting in tick_nohz_start_idle(). |
| 489 | */ |
| 490 | ts->inidle = 1; |
| 491 | tick_nohz_stop_sched_tick(ts); |
| 492 | |
| 493 | local_irq_enable(); |
| 494 | } |
| 495 | |
| 496 | /** |
| 497 | * tick_nohz_irq_exit - update next tick event from interrupt exit |
| 498 | * |
| 499 | * When an interrupt fires while we are idle and it doesn't cause |
| 500 | * a reschedule, it may still add, modify or delete a timer, enqueue |
| 501 | * an RCU callback, etc... |
| 502 | * So we need to re-calculate and reprogram the next tick event. |
| 503 | */ |
| 504 | void tick_nohz_irq_exit(void) |
| 505 | { |
| 506 | unsigned long flags; |
| 507 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 508 | |
| 509 | if (!ts->inidle) |
| 510 | return; |
| 511 | |
| 512 | local_irq_save(flags); |
| 513 | |
| 514 | tick_nohz_stop_sched_tick(ts); |
| 515 | |
| 516 | local_irq_restore(flags); |
| 517 | } |
| 518 | |
| 519 | /** |
| 520 | * tick_nohz_get_sleep_length - return the length of the current sleep |
| 521 | * |
| 522 | * Called from power state control code with interrupts disabled |
| 523 | */ |
| 524 | ktime_t tick_nohz_get_sleep_length(void) |
| 525 | { |
| 526 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 527 | |
| 528 | return ts->sleep_length; |
| 529 | } |
| 530 | |
| 531 | static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) |
| 532 | { |
| 533 | hrtimer_cancel(&ts->sched_timer); |
| 534 | hrtimer_set_expires(&ts->sched_timer, ts->idle_tick); |
| 535 | |
| 536 | while (1) { |
| 537 | /* Forward the time to expire in the future */ |
| 538 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
| 539 | |
| 540 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
| 541 | hrtimer_start_expires(&ts->sched_timer, |
| 542 | HRTIMER_MODE_ABS_PINNED); |
| 543 | /* Check, if the timer was already in the past */ |
| 544 | if (hrtimer_active(&ts->sched_timer)) |
| 545 | break; |
| 546 | } else { |
| 547 | if (!tick_program_event( |
| 548 | hrtimer_get_expires(&ts->sched_timer), 0)) |
| 549 | break; |
| 550 | } |
| 551 | /* Reread time and update jiffies */ |
| 552 | now = ktime_get(); |
| 553 | tick_do_update_jiffies64(now); |
| 554 | } |
| 555 | } |
| 556 | |
| 557 | /** |
| 558 | * tick_nohz_idle_exit - restart the idle tick from the idle task |
| 559 | * |
| 560 | * Restart the idle tick when the CPU is woken up from idle |
| 561 | * This also exit the RCU extended quiescent state. The CPU |
| 562 | * can use RCU again after this function is called. |
| 563 | */ |
| 564 | void tick_nohz_idle_exit(void) |
| 565 | { |
| 566 | int cpu = smp_processor_id(); |
| 567 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 568 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
| 569 | unsigned long ticks; |
| 570 | #endif |
| 571 | ktime_t now; |
| 572 | |
| 573 | local_irq_disable(); |
| 574 | |
| 575 | WARN_ON_ONCE(!ts->inidle); |
| 576 | |
| 577 | ts->inidle = 0; |
| 578 | |
| 579 | if (ts->idle_active || ts->tick_stopped) |
| 580 | now = ktime_get(); |
| 581 | |
| 582 | if (ts->idle_active) |
| 583 | tick_nohz_stop_idle(cpu, now); |
| 584 | |
| 585 | if (!ts->tick_stopped) { |
| 586 | local_irq_enable(); |
| 587 | return; |
| 588 | } |
| 589 | |
| 590 | /* Update jiffies first */ |
| 591 | select_nohz_load_balancer(0); |
| 592 | tick_do_update_jiffies64(now); |
| 593 | update_cpu_load_nohz(); |
| 594 | |
| 595 | #ifndef CONFIG_VIRT_CPU_ACCOUNTING |
| 596 | /* |
| 597 | * We stopped the tick in idle. Update process times would miss the |
| 598 | * time we slept as update_process_times does only a 1 tick |
| 599 | * accounting. Enforce that this is accounted to idle ! |
| 600 | */ |
| 601 | ticks = jiffies - ts->idle_jiffies; |
| 602 | /* |
| 603 | * We might be one off. Do not randomly account a huge number of ticks! |
| 604 | */ |
| 605 | if (ticks && ticks < LONG_MAX) |
| 606 | account_idle_ticks(ticks); |
| 607 | #endif |
| 608 | |
| 609 | calc_load_exit_idle(); |
| 610 | touch_softlockup_watchdog(); |
| 611 | /* |
| 612 | * Cancel the scheduled timer and restore the tick |
| 613 | */ |
| 614 | ts->tick_stopped = 0; |
| 615 | ts->idle_exittime = now; |
| 616 | |
| 617 | tick_nohz_restart(ts, now); |
| 618 | |
| 619 | local_irq_enable(); |
| 620 | } |
| 621 | |
| 622 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) |
| 623 | { |
| 624 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
| 625 | return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); |
| 626 | } |
| 627 | |
| 628 | /* |
| 629 | * The nohz low res interrupt handler |
| 630 | */ |
| 631 | static void tick_nohz_handler(struct clock_event_device *dev) |
| 632 | { |
| 633 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 634 | struct pt_regs *regs = get_irq_regs(); |
| 635 | int cpu = smp_processor_id(); |
| 636 | ktime_t now = ktime_get(); |
| 637 | |
| 638 | dev->next_event.tv64 = KTIME_MAX; |
| 639 | |
| 640 | /* |
| 641 | * Check if the do_timer duty was dropped. We don't care about |
| 642 | * concurrency: This happens only when the cpu in charge went |
| 643 | * into a long sleep. If two cpus happen to assign themself to |
| 644 | * this duty, then the jiffies update is still serialized by |
| 645 | * xtime_lock. |
| 646 | */ |
| 647 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
| 648 | tick_do_timer_cpu = cpu; |
| 649 | |
| 650 | /* Check, if the jiffies need an update */ |
| 651 | if (tick_do_timer_cpu == cpu) |
| 652 | tick_do_update_jiffies64(now); |
| 653 | |
| 654 | /* |
| 655 | * When we are idle and the tick is stopped, we have to touch |
| 656 | * the watchdog as we might not schedule for a really long |
| 657 | * time. This happens on complete idle SMP systems while |
| 658 | * waiting on the login prompt. We also increment the "start |
| 659 | * of idle" jiffy stamp so the idle accounting adjustment we |
| 660 | * do when we go busy again does not account too much ticks. |
| 661 | */ |
| 662 | if (ts->tick_stopped) { |
| 663 | touch_softlockup_watchdog(); |
| 664 | ts->idle_jiffies++; |
| 665 | } |
| 666 | |
| 667 | update_process_times(user_mode(regs)); |
| 668 | profile_tick(CPU_PROFILING); |
| 669 | |
| 670 | while (tick_nohz_reprogram(ts, now)) { |
| 671 | now = ktime_get(); |
| 672 | tick_do_update_jiffies64(now); |
| 673 | } |
| 674 | } |
| 675 | |
| 676 | /** |
| 677 | * tick_nohz_switch_to_nohz - switch to nohz mode |
| 678 | */ |
| 679 | static void tick_nohz_switch_to_nohz(void) |
| 680 | { |
| 681 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 682 | ktime_t next; |
| 683 | |
| 684 | if (!tick_nohz_enabled) |
| 685 | return; |
| 686 | |
| 687 | local_irq_disable(); |
| 688 | if (tick_switch_to_oneshot(tick_nohz_handler)) { |
| 689 | local_irq_enable(); |
| 690 | return; |
| 691 | } |
| 692 | |
| 693 | ts->nohz_mode = NOHZ_MODE_LOWRES; |
| 694 | |
| 695 | /* |
| 696 | * Recycle the hrtimer in ts, so we can share the |
| 697 | * hrtimer_forward with the highres code. |
| 698 | */ |
| 699 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); |
| 700 | /* Get the next period */ |
| 701 | next = tick_init_jiffy_update(); |
| 702 | |
| 703 | for (;;) { |
| 704 | hrtimer_set_expires(&ts->sched_timer, next); |
| 705 | if (!tick_program_event(next, 0)) |
| 706 | break; |
| 707 | next = ktime_add(next, tick_period); |
| 708 | } |
| 709 | local_irq_enable(); |
| 710 | } |
| 711 | |
| 712 | /* |
| 713 | * When NOHZ is enabled and the tick is stopped, we need to kick the |
| 714 | * tick timer from irq_enter() so that the jiffies update is kept |
| 715 | * alive during long running softirqs. That's ugly as hell, but |
| 716 | * correctness is key even if we need to fix the offending softirq in |
| 717 | * the first place. |
| 718 | * |
| 719 | * Note, this is different to tick_nohz_restart. We just kick the |
| 720 | * timer and do not touch the other magic bits which need to be done |
| 721 | * when idle is left. |
| 722 | */ |
| 723 | static void tick_nohz_kick_tick(int cpu, ktime_t now) |
| 724 | { |
| 725 | #if 0 |
| 726 | /* Switch back to 2.6.27 behaviour */ |
| 727 | |
| 728 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 729 | ktime_t delta; |
| 730 | |
| 731 | /* |
| 732 | * Do not touch the tick device, when the next expiry is either |
| 733 | * already reached or less/equal than the tick period. |
| 734 | */ |
| 735 | delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); |
| 736 | if (delta.tv64 <= tick_period.tv64) |
| 737 | return; |
| 738 | |
| 739 | tick_nohz_restart(ts, now); |
| 740 | #endif |
| 741 | } |
| 742 | |
| 743 | static inline void tick_check_nohz(int cpu) |
| 744 | { |
| 745 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 746 | ktime_t now; |
| 747 | |
| 748 | if (!ts->idle_active && !ts->tick_stopped) |
| 749 | return; |
| 750 | now = ktime_get(); |
| 751 | if (ts->idle_active) |
| 752 | tick_nohz_stop_idle(cpu, now); |
| 753 | if (ts->tick_stopped) { |
| 754 | tick_nohz_update_jiffies(now); |
| 755 | tick_nohz_kick_tick(cpu, now); |
| 756 | } |
| 757 | } |
| 758 | |
| 759 | #else |
| 760 | |
| 761 | static inline void tick_nohz_switch_to_nohz(void) { } |
| 762 | static inline void tick_check_nohz(int cpu) { } |
| 763 | |
| 764 | #endif /* NO_HZ */ |
| 765 | |
| 766 | /* |
| 767 | * Called from irq_enter to notify about the possible interruption of idle() |
| 768 | */ |
| 769 | void tick_check_idle(int cpu) |
| 770 | { |
| 771 | tick_check_oneshot_broadcast(cpu); |
| 772 | tick_check_nohz(cpu); |
| 773 | } |
| 774 | |
| 775 | /* |
| 776 | * High resolution timer specific code |
| 777 | */ |
| 778 | #ifdef CONFIG_HIGH_RES_TIMERS |
| 779 | /* |
| 780 | * We rearm the timer until we get disabled by the idle code. |
| 781 | * Called with interrupts disabled and timer->base->cpu_base->lock held. |
| 782 | */ |
| 783 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) |
| 784 | { |
| 785 | struct tick_sched *ts = |
| 786 | container_of(timer, struct tick_sched, sched_timer); |
| 787 | struct pt_regs *regs = get_irq_regs(); |
| 788 | ktime_t now = ktime_get(); |
| 789 | int cpu = smp_processor_id(); |
| 790 | |
| 791 | #ifdef CONFIG_NO_HZ |
| 792 | /* |
| 793 | * Check if the do_timer duty was dropped. We don't care about |
| 794 | * concurrency: This happens only when the cpu in charge went |
| 795 | * into a long sleep. If two cpus happen to assign themself to |
| 796 | * this duty, then the jiffies update is still serialized by |
| 797 | * xtime_lock. |
| 798 | */ |
| 799 | if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) |
| 800 | tick_do_timer_cpu = cpu; |
| 801 | #endif |
| 802 | |
| 803 | /* Check, if the jiffies need an update */ |
| 804 | if (tick_do_timer_cpu == cpu) |
| 805 | tick_do_update_jiffies64(now); |
| 806 | |
| 807 | /* |
| 808 | * Do not call, when we are not in irq context and have |
| 809 | * no valid regs pointer |
| 810 | */ |
| 811 | if (regs) { |
| 812 | /* |
| 813 | * When we are idle and the tick is stopped, we have to touch |
| 814 | * the watchdog as we might not schedule for a really long |
| 815 | * time. This happens on complete idle SMP systems while |
| 816 | * waiting on the login prompt. We also increment the "start of |
| 817 | * idle" jiffy stamp so the idle accounting adjustment we do |
| 818 | * when we go busy again does not account too much ticks. |
| 819 | */ |
| 820 | if (ts->tick_stopped) { |
| 821 | touch_softlockup_watchdog(); |
| 822 | ts->idle_jiffies++; |
| 823 | } |
| 824 | update_process_times(user_mode(regs)); |
| 825 | profile_tick(CPU_PROFILING); |
| 826 | } |
| 827 | #ifdef CONFIG_SINGLECORE |
| 828 | #ifndef CONFIG_SYSTEM_RECOVERY |
| 829 | #ifndef CONFIG_SYSTEM_CAP |
| 830 | #ifdef USE_CPPS_KO |
| 831 | if(cpps_callbacks.linux_oss_tick_timer_function) |
| 832 | cpps_callbacks.linux_oss_tick_timer_function(); |
| 833 | #else |
| 834 | linux_oss_tick_timer_function(); |
| 835 | #endif |
| 836 | #endif |
| 837 | #endif |
| 838 | #endif |
| 839 | hrtimer_forward(timer, now, tick_period); |
| 840 | |
| 841 | return HRTIMER_RESTART; |
| 842 | } |
| 843 | |
| 844 | static int sched_skew_tick; |
| 845 | |
| 846 | static int __init skew_tick(char *str) |
| 847 | { |
| 848 | get_option(&str, &sched_skew_tick); |
| 849 | |
| 850 | return 0; |
| 851 | } |
| 852 | early_param("skew_tick", skew_tick); |
| 853 | |
| 854 | /** |
| 855 | * tick_setup_sched_timer - setup the tick emulation timer |
| 856 | */ |
| 857 | void tick_setup_sched_timer(void) |
| 858 | { |
| 859 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 860 | ktime_t now = ktime_get(); |
| 861 | |
| 862 | /* |
| 863 | * Emulate tick processing via per-CPU hrtimers: |
| 864 | */ |
| 865 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); |
| 866 | ts->sched_timer.irqsafe = 1; |
| 867 | ts->sched_timer.function = tick_sched_timer; |
| 868 | |
| 869 | /* Get the next period (per cpu) */ |
| 870 | hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); |
| 871 | |
| 872 | /* Offset the tick to avert xtime_lock contention. */ |
| 873 | if (sched_skew_tick) { |
| 874 | u64 offset = ktime_to_ns(tick_period) >> 1; |
| 875 | do_div(offset, num_possible_cpus()); |
| 876 | offset *= smp_processor_id(); |
| 877 | hrtimer_add_expires_ns(&ts->sched_timer, offset); |
| 878 | } |
| 879 | |
| 880 | for (;;) { |
| 881 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
| 882 | hrtimer_start_expires(&ts->sched_timer, |
| 883 | HRTIMER_MODE_ABS_PINNED); |
| 884 | /* Check, if the timer was already in the past */ |
| 885 | if (hrtimer_active(&ts->sched_timer)) |
| 886 | break; |
| 887 | now = ktime_get(); |
| 888 | } |
| 889 | |
| 890 | #ifdef CONFIG_NO_HZ |
| 891 | if (tick_nohz_enabled) |
| 892 | ts->nohz_mode = NOHZ_MODE_HIGHRES; |
| 893 | #endif |
| 894 | } |
| 895 | #endif /* HIGH_RES_TIMERS */ |
| 896 | |
| 897 | #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS |
| 898 | void tick_cancel_sched_timer(int cpu) |
| 899 | { |
| 900 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 901 | |
| 902 | # ifdef CONFIG_HIGH_RES_TIMERS |
| 903 | if (ts->sched_timer.base) |
| 904 | hrtimer_cancel(&ts->sched_timer); |
| 905 | # endif |
| 906 | |
| 907 | memset(ts, 0, sizeof(*ts)); |
| 908 | } |
| 909 | #endif |
| 910 | |
| 911 | /** |
| 912 | * Async notification about clocksource changes |
| 913 | */ |
| 914 | void tick_clock_notify(void) |
| 915 | { |
| 916 | int cpu; |
| 917 | |
| 918 | for_each_possible_cpu(cpu) |
| 919 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); |
| 920 | } |
| 921 | |
| 922 | /* |
| 923 | * Async notification about clock event changes |
| 924 | */ |
| 925 | void tick_oneshot_notify(void) |
| 926 | { |
| 927 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 928 | |
| 929 | set_bit(0, &ts->check_clocks); |
| 930 | } |
| 931 | |
| 932 | /** |
| 933 | * Check, if a change happened, which makes oneshot possible. |
| 934 | * |
| 935 | * Called cyclic from the hrtimer softirq (driven by the timer |
| 936 | * softirq) allow_nohz signals, that we can switch into low-res nohz |
| 937 | * mode, because high resolution timers are disabled (either compile |
| 938 | * or runtime). |
| 939 | */ |
| 940 | int tick_check_oneshot_change(int allow_nohz) |
| 941 | { |
| 942 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 943 | |
| 944 | if (!test_and_clear_bit(0, &ts->check_clocks)) |
| 945 | return 0; |
| 946 | |
| 947 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) |
| 948 | return 0; |
| 949 | |
| 950 | if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) |
| 951 | return 0; |
| 952 | |
| 953 | if (!allow_nohz) |
| 954 | return 1; |
| 955 | |
| 956 | tick_nohz_switch_to_nohz(); |
| 957 | return 0; |
| 958 | } |