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 | * |
Pavel Machek | b10db7f | 2008-01-30 13:30:00 +0100 | [diff] [blame] | 12 | * Distribute under GPLv2. |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 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 | |
David S. Miller | 9e203bc | 2007-02-24 22:10:13 -0800 | [diff] [blame] | 24 | #include <asm/irq_regs.h> |
| 25 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 26 | #include "tick-internal.h" |
| 27 | |
| 28 | /* |
| 29 | * Per cpu nohz control structure |
| 30 | */ |
| 31 | static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); |
| 32 | |
| 33 | /* |
| 34 | * The time, when the last jiffy update happened. Protected by xtime_lock. |
| 35 | */ |
| 36 | static ktime_t last_jiffies_update; |
| 37 | |
Ingo Molnar | 289f480 | 2007-02-16 01:28:15 -0800 | [diff] [blame] | 38 | struct tick_sched *tick_get_tick_sched(int cpu) |
| 39 | { |
| 40 | return &per_cpu(tick_cpu_sched, cpu); |
| 41 | } |
| 42 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 43 | /* |
| 44 | * Must be called with interrupts disabled ! |
| 45 | */ |
| 46 | static void tick_do_update_jiffies64(ktime_t now) |
| 47 | { |
| 48 | unsigned long ticks = 0; |
| 49 | ktime_t delta; |
| 50 | |
| 51 | /* Reevalute with xtime_lock held */ |
| 52 | write_seqlock(&xtime_lock); |
| 53 | |
| 54 | delta = ktime_sub(now, last_jiffies_update); |
| 55 | if (delta.tv64 >= tick_period.tv64) { |
| 56 | |
| 57 | delta = ktime_sub(delta, tick_period); |
| 58 | last_jiffies_update = ktime_add(last_jiffies_update, |
| 59 | tick_period); |
| 60 | |
| 61 | /* Slow path for long timeouts */ |
| 62 | if (unlikely(delta.tv64 >= tick_period.tv64)) { |
| 63 | s64 incr = ktime_to_ns(tick_period); |
| 64 | |
| 65 | ticks = ktime_divns(delta, incr); |
| 66 | |
| 67 | last_jiffies_update = ktime_add_ns(last_jiffies_update, |
| 68 | incr * ticks); |
| 69 | } |
| 70 | do_timer(++ticks); |
| 71 | } |
| 72 | write_sequnlock(&xtime_lock); |
| 73 | } |
| 74 | |
| 75 | /* |
| 76 | * Initialize and return retrieve the jiffies update. |
| 77 | */ |
| 78 | static ktime_t tick_init_jiffy_update(void) |
| 79 | { |
| 80 | ktime_t period; |
| 81 | |
| 82 | write_seqlock(&xtime_lock); |
| 83 | /* Did we start the jiffies update yet ? */ |
| 84 | if (last_jiffies_update.tv64 == 0) |
| 85 | last_jiffies_update = tick_next_period; |
| 86 | period = last_jiffies_update; |
| 87 | write_sequnlock(&xtime_lock); |
| 88 | return period; |
| 89 | } |
| 90 | |
| 91 | /* |
| 92 | * NOHZ - aka dynamic tick functionality |
| 93 | */ |
| 94 | #ifdef CONFIG_NO_HZ |
| 95 | /* |
| 96 | * NO HZ enabled ? |
| 97 | */ |
| 98 | static int tick_nohz_enabled __read_mostly = 1; |
| 99 | |
| 100 | /* |
| 101 | * Enable / Disable tickless mode |
| 102 | */ |
| 103 | static int __init setup_tick_nohz(char *str) |
| 104 | { |
| 105 | if (!strcmp(str, "off")) |
| 106 | tick_nohz_enabled = 0; |
| 107 | else if (!strcmp(str, "on")) |
| 108 | tick_nohz_enabled = 1; |
| 109 | else |
| 110 | return 0; |
| 111 | return 1; |
| 112 | } |
| 113 | |
| 114 | __setup("nohz=", setup_tick_nohz); |
| 115 | |
| 116 | /** |
| 117 | * tick_nohz_update_jiffies - update jiffies when idle was interrupted |
| 118 | * |
| 119 | * Called from interrupt entry when the CPU was idle |
| 120 | * |
| 121 | * In case the sched_tick was stopped on this CPU, we have to check if jiffies |
| 122 | * must be updated. Otherwise an interrupt handler could use a stale jiffy |
| 123 | * value. We do this unconditionally on any cpu, as we don't know whether the |
| 124 | * cpu, which has the update task assigned is in a long sleep. |
| 125 | */ |
| 126 | void tick_nohz_update_jiffies(void) |
| 127 | { |
| 128 | int cpu = smp_processor_id(); |
| 129 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 130 | unsigned long flags; |
| 131 | ktime_t now; |
| 132 | |
| 133 | if (!ts->tick_stopped) |
| 134 | return; |
| 135 | |
Thomas Gleixner | d393820 | 2007-11-28 15:52:56 +0100 | [diff] [blame] | 136 | touch_softlockup_watchdog(); |
| 137 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 138 | cpu_clear(cpu, nohz_cpu_mask); |
| 139 | now = ktime_get(); |
| 140 | |
| 141 | local_irq_save(flags); |
| 142 | tick_do_update_jiffies64(now); |
| 143 | local_irq_restore(flags); |
| 144 | } |
| 145 | |
Venki Pallipadi | 6378ddb | 2008-01-30 13:30:04 +0100 | [diff] [blame^] | 146 | void tick_nohz_stop_idle(int cpu) |
| 147 | { |
| 148 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 149 | |
| 150 | if (ts->idle_active) { |
| 151 | ktime_t now, delta; |
| 152 | now = ktime_get(); |
| 153 | delta = ktime_sub(now, ts->idle_entrytime); |
| 154 | ts->idle_lastupdate = now; |
| 155 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); |
| 156 | ts->idle_active = 0; |
| 157 | } |
| 158 | } |
| 159 | |
| 160 | static ktime_t tick_nohz_start_idle(int cpu) |
| 161 | { |
| 162 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 163 | ktime_t now, delta; |
| 164 | |
| 165 | now = ktime_get(); |
| 166 | if (ts->idle_active) { |
| 167 | delta = ktime_sub(now, ts->idle_entrytime); |
| 168 | ts->idle_lastupdate = now; |
| 169 | ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); |
| 170 | } |
| 171 | ts->idle_entrytime = now; |
| 172 | ts->idle_active = 1; |
| 173 | return now; |
| 174 | } |
| 175 | |
| 176 | u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) |
| 177 | { |
| 178 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 179 | |
| 180 | *last_update_time = ktime_to_us(ts->idle_lastupdate); |
| 181 | return ktime_to_us(ts->idle_sleeptime); |
| 182 | } |
| 183 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 184 | /** |
| 185 | * tick_nohz_stop_sched_tick - stop the idle tick from the idle task |
| 186 | * |
| 187 | * When the next event is more than a tick into the future, stop the idle tick |
| 188 | * Called either from the idle loop or from irq_exit() when an idle period was |
| 189 | * just interrupted by an interrupt which did not cause a reschedule. |
| 190 | */ |
| 191 | void tick_nohz_stop_sched_tick(void) |
| 192 | { |
| 193 | unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; |
Peter Zijlstra | 48d5e25 | 2008-01-25 21:08:31 +0100 | [diff] [blame] | 194 | unsigned long rt_jiffies; |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 195 | struct tick_sched *ts; |
Venki Pallipadi | 6378ddb | 2008-01-30 13:30:04 +0100 | [diff] [blame^] | 196 | ktime_t last_update, expires, now; |
Len Brown | 4f86d3a | 2007-10-03 18:58:00 -0400 | [diff] [blame] | 197 | struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 198 | int cpu; |
| 199 | |
| 200 | local_irq_save(flags); |
| 201 | |
| 202 | cpu = smp_processor_id(); |
Venki Pallipadi | 6378ddb | 2008-01-30 13:30:04 +0100 | [diff] [blame^] | 203 | now = tick_nohz_start_idle(cpu); |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 204 | ts = &per_cpu(tick_cpu_sched, cpu); |
| 205 | |
Thomas Gleixner | 5e41d0d | 2007-09-16 15:36:43 +0200 | [diff] [blame] | 206 | /* |
| 207 | * If this cpu is offline and it is the one which updates |
| 208 | * jiffies, then give up the assignment and let it be taken by |
| 209 | * the cpu which runs the tick timer next. If we don't drop |
| 210 | * this here the jiffies might be stale and do_timer() never |
| 211 | * invoked. |
| 212 | */ |
| 213 | if (unlikely(!cpu_online(cpu))) { |
| 214 | if (cpu == tick_do_timer_cpu) |
| 215 | tick_do_timer_cpu = -1; |
| 216 | } |
| 217 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 218 | if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) |
| 219 | goto end; |
| 220 | |
| 221 | if (need_resched()) |
| 222 | goto end; |
| 223 | |
| 224 | cpu = smp_processor_id(); |
Thomas Gleixner | 3528231 | 2007-05-23 13:57:37 -0700 | [diff] [blame] | 225 | if (unlikely(local_softirq_pending())) { |
| 226 | static int ratelimit; |
| 227 | |
| 228 | if (ratelimit < 10) { |
| 229 | printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", |
| 230 | local_softirq_pending()); |
| 231 | ratelimit++; |
| 232 | } |
| 233 | } |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 234 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 235 | ts->idle_calls++; |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 236 | /* Read jiffies and the time when jiffies were updated last */ |
| 237 | do { |
| 238 | seq = read_seqbegin(&xtime_lock); |
| 239 | last_update = last_jiffies_update; |
| 240 | last_jiffies = jiffies; |
| 241 | } while (read_seqretry(&xtime_lock, seq)); |
| 242 | |
| 243 | /* Get the next timer wheel timer */ |
| 244 | next_jiffies = get_next_timer_interrupt(last_jiffies); |
| 245 | delta_jiffies = next_jiffies - last_jiffies; |
| 246 | |
Peter Zijlstra | 48d5e25 | 2008-01-25 21:08:31 +0100 | [diff] [blame] | 247 | rt_jiffies = rt_needs_cpu(cpu); |
| 248 | if (rt_jiffies && rt_jiffies < delta_jiffies) |
| 249 | delta_jiffies = rt_jiffies; |
| 250 | |
Ingo Molnar | 6ba9b34 | 2007-02-19 18:11:56 +0000 | [diff] [blame] | 251 | if (rcu_needs_cpu(cpu)) |
| 252 | delta_jiffies = 1; |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 253 | /* |
| 254 | * Do not stop the tick, if we are only one off |
| 255 | * or if the cpu is required for rcu |
| 256 | */ |
Ingo Molnar | 6ba9b34 | 2007-02-19 18:11:56 +0000 | [diff] [blame] | 257 | if (!ts->tick_stopped && delta_jiffies == 1) |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 258 | goto out; |
| 259 | |
| 260 | /* Schedule the tick, if we are at least one jiffie off */ |
| 261 | if ((long)delta_jiffies >= 1) { |
| 262 | |
Ingo Molnar | 6ba9b34 | 2007-02-19 18:11:56 +0000 | [diff] [blame] | 263 | if (delta_jiffies > 1) |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 264 | cpu_set(cpu, nohz_cpu_mask); |
| 265 | /* |
| 266 | * nohz_stop_sched_tick can be called several times before |
| 267 | * the nohz_restart_sched_tick is called. This happens when |
| 268 | * interrupts arrive which do not cause a reschedule. In the |
| 269 | * first call we save the current tick time, so we can restart |
| 270 | * the scheduler tick in nohz_restart_sched_tick. |
| 271 | */ |
| 272 | if (!ts->tick_stopped) { |
Siddha, Suresh B | 46cb4b7 | 2007-05-08 00:32:51 -0700 | [diff] [blame] | 273 | if (select_nohz_load_balancer(1)) { |
| 274 | /* |
| 275 | * sched tick not stopped! |
| 276 | */ |
| 277 | cpu_clear(cpu, nohz_cpu_mask); |
| 278 | goto out; |
| 279 | } |
| 280 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 281 | ts->idle_tick = ts->sched_timer.expires; |
| 282 | ts->tick_stopped = 1; |
| 283 | ts->idle_jiffies = last_jiffies; |
| 284 | } |
Thomas Gleixner | d3ed782 | 2007-05-08 00:30:03 -0700 | [diff] [blame] | 285 | |
| 286 | /* |
| 287 | * If this cpu is the one which updates jiffies, then |
| 288 | * give up the assignment and let it be taken by the |
| 289 | * cpu which runs the tick timer next, which might be |
| 290 | * this cpu as well. If we don't drop this here the |
| 291 | * jiffies might be stale and do_timer() never |
| 292 | * invoked. |
| 293 | */ |
| 294 | if (cpu == tick_do_timer_cpu) |
| 295 | tick_do_timer_cpu = -1; |
| 296 | |
Thomas Gleixner | eaad084 | 2007-05-29 23:47:39 +0200 | [diff] [blame] | 297 | ts->idle_sleeps++; |
| 298 | |
| 299 | /* |
| 300 | * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that |
| 301 | * there is no timer pending or at least extremly far |
| 302 | * into the future (12 days for HZ=1000). In this case |
| 303 | * we simply stop the tick timer: |
| 304 | */ |
| 305 | if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) { |
| 306 | ts->idle_expires.tv64 = KTIME_MAX; |
| 307 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) |
| 308 | hrtimer_cancel(&ts->sched_timer); |
| 309 | goto out; |
| 310 | } |
| 311 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 312 | /* |
| 313 | * calculate the expiry time for the next timer wheel |
| 314 | * timer |
| 315 | */ |
| 316 | expires = ktime_add_ns(last_update, tick_period.tv64 * |
| 317 | delta_jiffies); |
| 318 | ts->idle_expires = expires; |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 319 | |
| 320 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
| 321 | hrtimer_start(&ts->sched_timer, expires, |
| 322 | HRTIMER_MODE_ABS); |
| 323 | /* Check, if the timer was already in the past */ |
| 324 | if (hrtimer_active(&ts->sched_timer)) |
| 325 | goto out; |
Pavel Machek | 4c9dc64 | 2008-01-30 13:30:00 +0100 | [diff] [blame] | 326 | } else if (!tick_program_event(expires, 0)) |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 327 | goto out; |
| 328 | /* |
| 329 | * We are past the event already. So we crossed a |
| 330 | * jiffie boundary. Update jiffies and raise the |
| 331 | * softirq. |
| 332 | */ |
| 333 | tick_do_update_jiffies64(ktime_get()); |
| 334 | cpu_clear(cpu, nohz_cpu_mask); |
| 335 | } |
| 336 | raise_softirq_irqoff(TIMER_SOFTIRQ); |
| 337 | out: |
| 338 | ts->next_jiffies = next_jiffies; |
| 339 | ts->last_jiffies = last_jiffies; |
Len Brown | 4f86d3a | 2007-10-03 18:58:00 -0400 | [diff] [blame] | 340 | ts->sleep_length = ktime_sub(dev->next_event, now); |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 341 | end: |
| 342 | local_irq_restore(flags); |
| 343 | } |
| 344 | |
| 345 | /** |
Len Brown | 4f86d3a | 2007-10-03 18:58:00 -0400 | [diff] [blame] | 346 | * tick_nohz_get_sleep_length - return the length of the current sleep |
| 347 | * |
| 348 | * Called from power state control code with interrupts disabled |
| 349 | */ |
| 350 | ktime_t tick_nohz_get_sleep_length(void) |
| 351 | { |
| 352 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 353 | |
| 354 | return ts->sleep_length; |
| 355 | } |
| 356 | |
Len Brown | 4f86d3a | 2007-10-03 18:58:00 -0400 | [diff] [blame] | 357 | /** |
Li Zefan | 8dce39c | 2007-11-05 14:51:10 -0800 | [diff] [blame] | 358 | * tick_nohz_restart_sched_tick - restart the idle tick from the idle task |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 359 | * |
| 360 | * Restart the idle tick when the CPU is woken up from idle |
| 361 | */ |
| 362 | void tick_nohz_restart_sched_tick(void) |
| 363 | { |
| 364 | int cpu = smp_processor_id(); |
| 365 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 366 | unsigned long ticks; |
Venki Pallipadi | 6378ddb | 2008-01-30 13:30:04 +0100 | [diff] [blame^] | 367 | ktime_t now; |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 368 | |
| 369 | local_irq_disable(); |
Venki Pallipadi | 6378ddb | 2008-01-30 13:30:04 +0100 | [diff] [blame^] | 370 | tick_nohz_stop_idle(cpu); |
| 371 | |
| 372 | if (!ts->tick_stopped) { |
| 373 | local_irq_enable(); |
| 374 | return; |
| 375 | } |
| 376 | |
| 377 | /* Update jiffies first */ |
Siddha, Suresh B | 46cb4b7 | 2007-05-08 00:32:51 -0700 | [diff] [blame] | 378 | select_nohz_load_balancer(0); |
Venki Pallipadi | 6378ddb | 2008-01-30 13:30:04 +0100 | [diff] [blame^] | 379 | now = ktime_get(); |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 380 | tick_do_update_jiffies64(now); |
| 381 | cpu_clear(cpu, nohz_cpu_mask); |
| 382 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 383 | /* |
| 384 | * We stopped the tick in idle. Update process times would miss the |
| 385 | * time we slept as update_process_times does only a 1 tick |
| 386 | * accounting. Enforce that this is accounted to idle ! |
| 387 | */ |
| 388 | ticks = jiffies - ts->idle_jiffies; |
| 389 | /* |
| 390 | * We might be one off. Do not randomly account a huge number of ticks! |
| 391 | */ |
| 392 | if (ticks && ticks < LONG_MAX) { |
| 393 | add_preempt_count(HARDIRQ_OFFSET); |
| 394 | account_system_time(current, HARDIRQ_OFFSET, |
| 395 | jiffies_to_cputime(ticks)); |
| 396 | sub_preempt_count(HARDIRQ_OFFSET); |
| 397 | } |
| 398 | |
| 399 | /* |
| 400 | * Cancel the scheduled timer and restore the tick |
| 401 | */ |
| 402 | ts->tick_stopped = 0; |
| 403 | hrtimer_cancel(&ts->sched_timer); |
| 404 | ts->sched_timer.expires = ts->idle_tick; |
| 405 | |
| 406 | while (1) { |
| 407 | /* Forward the time to expire in the future */ |
| 408 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
| 409 | |
| 410 | if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { |
| 411 | hrtimer_start(&ts->sched_timer, |
| 412 | ts->sched_timer.expires, |
| 413 | HRTIMER_MODE_ABS); |
| 414 | /* Check, if the timer was already in the past */ |
| 415 | if (hrtimer_active(&ts->sched_timer)) |
| 416 | break; |
| 417 | } else { |
| 418 | if (!tick_program_event(ts->sched_timer.expires, 0)) |
| 419 | break; |
| 420 | } |
| 421 | /* Update jiffies and reread time */ |
| 422 | tick_do_update_jiffies64(now); |
| 423 | now = ktime_get(); |
| 424 | } |
| 425 | local_irq_enable(); |
| 426 | } |
| 427 | |
| 428 | static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) |
| 429 | { |
| 430 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
| 431 | return tick_program_event(ts->sched_timer.expires, 0); |
| 432 | } |
| 433 | |
| 434 | /* |
| 435 | * The nohz low res interrupt handler |
| 436 | */ |
| 437 | static void tick_nohz_handler(struct clock_event_device *dev) |
| 438 | { |
| 439 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 440 | struct pt_regs *regs = get_irq_regs(); |
Thomas Gleixner | d3ed782 | 2007-05-08 00:30:03 -0700 | [diff] [blame] | 441 | int cpu = smp_processor_id(); |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 442 | ktime_t now = ktime_get(); |
| 443 | |
| 444 | dev->next_event.tv64 = KTIME_MAX; |
| 445 | |
Thomas Gleixner | d3ed782 | 2007-05-08 00:30:03 -0700 | [diff] [blame] | 446 | /* |
| 447 | * Check if the do_timer duty was dropped. We don't care about |
| 448 | * concurrency: This happens only when the cpu in charge went |
| 449 | * into a long sleep. If two cpus happen to assign themself to |
| 450 | * this duty, then the jiffies update is still serialized by |
| 451 | * xtime_lock. |
| 452 | */ |
| 453 | if (unlikely(tick_do_timer_cpu == -1)) |
| 454 | tick_do_timer_cpu = cpu; |
| 455 | |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 456 | /* Check, if the jiffies need an update */ |
Thomas Gleixner | d3ed782 | 2007-05-08 00:30:03 -0700 | [diff] [blame] | 457 | if (tick_do_timer_cpu == cpu) |
| 458 | tick_do_update_jiffies64(now); |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 459 | |
| 460 | /* |
| 461 | * When we are idle and the tick is stopped, we have to touch |
| 462 | * the watchdog as we might not schedule for a really long |
| 463 | * time. This happens on complete idle SMP systems while |
| 464 | * waiting on the login prompt. We also increment the "start |
| 465 | * of idle" jiffy stamp so the idle accounting adjustment we |
| 466 | * do when we go busy again does not account too much ticks. |
| 467 | */ |
| 468 | if (ts->tick_stopped) { |
| 469 | touch_softlockup_watchdog(); |
| 470 | ts->idle_jiffies++; |
| 471 | } |
| 472 | |
| 473 | update_process_times(user_mode(regs)); |
| 474 | profile_tick(CPU_PROFILING); |
| 475 | |
| 476 | /* Do not restart, when we are in the idle loop */ |
| 477 | if (ts->tick_stopped) |
| 478 | return; |
| 479 | |
| 480 | while (tick_nohz_reprogram(ts, now)) { |
| 481 | now = ktime_get(); |
| 482 | tick_do_update_jiffies64(now); |
| 483 | } |
| 484 | } |
| 485 | |
| 486 | /** |
| 487 | * tick_nohz_switch_to_nohz - switch to nohz mode |
| 488 | */ |
| 489 | static void tick_nohz_switch_to_nohz(void) |
| 490 | { |
| 491 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 492 | ktime_t next; |
| 493 | |
| 494 | if (!tick_nohz_enabled) |
| 495 | return; |
| 496 | |
| 497 | local_irq_disable(); |
| 498 | if (tick_switch_to_oneshot(tick_nohz_handler)) { |
| 499 | local_irq_enable(); |
| 500 | return; |
| 501 | } |
| 502 | |
| 503 | ts->nohz_mode = NOHZ_MODE_LOWRES; |
| 504 | |
| 505 | /* |
| 506 | * Recycle the hrtimer in ts, so we can share the |
| 507 | * hrtimer_forward with the highres code. |
| 508 | */ |
| 509 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); |
| 510 | /* Get the next period */ |
| 511 | next = tick_init_jiffy_update(); |
| 512 | |
| 513 | for (;;) { |
| 514 | ts->sched_timer.expires = next; |
| 515 | if (!tick_program_event(next, 0)) |
| 516 | break; |
| 517 | next = ktime_add(next, tick_period); |
| 518 | } |
| 519 | local_irq_enable(); |
| 520 | |
| 521 | printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", |
| 522 | smp_processor_id()); |
| 523 | } |
| 524 | |
| 525 | #else |
| 526 | |
| 527 | static inline void tick_nohz_switch_to_nohz(void) { } |
| 528 | |
| 529 | #endif /* NO_HZ */ |
| 530 | |
| 531 | /* |
| 532 | * High resolution timer specific code |
| 533 | */ |
| 534 | #ifdef CONFIG_HIGH_RES_TIMERS |
| 535 | /* |
Pavel Machek | 4c9dc64 | 2008-01-30 13:30:00 +0100 | [diff] [blame] | 536 | * We rearm the timer until we get disabled by the idle code. |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 537 | * Called with interrupts disabled and timer->base->cpu_base->lock held. |
| 538 | */ |
| 539 | static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) |
| 540 | { |
| 541 | struct tick_sched *ts = |
| 542 | container_of(timer, struct tick_sched, sched_timer); |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 543 | struct pt_regs *regs = get_irq_regs(); |
| 544 | ktime_t now = ktime_get(); |
Thomas Gleixner | d3ed782 | 2007-05-08 00:30:03 -0700 | [diff] [blame] | 545 | int cpu = smp_processor_id(); |
| 546 | |
| 547 | #ifdef CONFIG_NO_HZ |
| 548 | /* |
| 549 | * Check if the do_timer duty was dropped. We don't care about |
| 550 | * concurrency: This happens only when the cpu in charge went |
| 551 | * into a long sleep. If two cpus happen to assign themself to |
| 552 | * this duty, then the jiffies update is still serialized by |
| 553 | * xtime_lock. |
| 554 | */ |
| 555 | if (unlikely(tick_do_timer_cpu == -1)) |
| 556 | tick_do_timer_cpu = cpu; |
| 557 | #endif |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 558 | |
| 559 | /* Check, if the jiffies need an update */ |
Thomas Gleixner | d3ed782 | 2007-05-08 00:30:03 -0700 | [diff] [blame] | 560 | if (tick_do_timer_cpu == cpu) |
| 561 | tick_do_update_jiffies64(now); |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 562 | |
| 563 | /* |
| 564 | * Do not call, when we are not in irq context and have |
| 565 | * no valid regs pointer |
| 566 | */ |
| 567 | if (regs) { |
| 568 | /* |
| 569 | * When we are idle and the tick is stopped, we have to touch |
| 570 | * the watchdog as we might not schedule for a really long |
| 571 | * time. This happens on complete idle SMP systems while |
| 572 | * waiting on the login prompt. We also increment the "start of |
| 573 | * idle" jiffy stamp so the idle accounting adjustment we do |
| 574 | * when we go busy again does not account too much ticks. |
| 575 | */ |
| 576 | if (ts->tick_stopped) { |
| 577 | touch_softlockup_watchdog(); |
| 578 | ts->idle_jiffies++; |
| 579 | } |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 580 | update_process_times(user_mode(regs)); |
| 581 | profile_tick(CPU_PROFILING); |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 582 | } |
| 583 | |
| 584 | /* Do not restart, when we are in the idle loop */ |
| 585 | if (ts->tick_stopped) |
| 586 | return HRTIMER_NORESTART; |
| 587 | |
| 588 | hrtimer_forward(timer, now, tick_period); |
| 589 | |
| 590 | return HRTIMER_RESTART; |
| 591 | } |
| 592 | |
| 593 | /** |
| 594 | * tick_setup_sched_timer - setup the tick emulation timer |
| 595 | */ |
| 596 | void tick_setup_sched_timer(void) |
| 597 | { |
| 598 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 599 | ktime_t now = ktime_get(); |
john stultz | 3704540 | 2007-07-21 04:37:35 -0700 | [diff] [blame] | 600 | u64 offset; |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 601 | |
| 602 | /* |
| 603 | * Emulate tick processing via per-CPU hrtimers: |
| 604 | */ |
| 605 | hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); |
| 606 | ts->sched_timer.function = tick_sched_timer; |
| 607 | ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; |
| 608 | |
john stultz | 3704540 | 2007-07-21 04:37:35 -0700 | [diff] [blame] | 609 | /* Get the next period (per cpu) */ |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 610 | ts->sched_timer.expires = tick_init_jiffy_update(); |
john stultz | 3704540 | 2007-07-21 04:37:35 -0700 | [diff] [blame] | 611 | offset = ktime_to_ns(tick_period) >> 1; |
john stultz | b2d9323 | 2007-10-16 23:27:18 -0700 | [diff] [blame] | 612 | do_div(offset, num_possible_cpus()); |
john stultz | 3704540 | 2007-07-21 04:37:35 -0700 | [diff] [blame] | 613 | offset *= smp_processor_id(); |
| 614 | ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset); |
Thomas Gleixner | 79bf2bb | 2007-02-16 01:28:03 -0800 | [diff] [blame] | 615 | |
| 616 | for (;;) { |
| 617 | hrtimer_forward(&ts->sched_timer, now, tick_period); |
| 618 | hrtimer_start(&ts->sched_timer, ts->sched_timer.expires, |
| 619 | HRTIMER_MODE_ABS); |
| 620 | /* Check, if the timer was already in the past */ |
| 621 | if (hrtimer_active(&ts->sched_timer)) |
| 622 | break; |
| 623 | now = ktime_get(); |
| 624 | } |
| 625 | |
| 626 | #ifdef CONFIG_NO_HZ |
| 627 | if (tick_nohz_enabled) |
| 628 | ts->nohz_mode = NOHZ_MODE_HIGHRES; |
| 629 | #endif |
| 630 | } |
| 631 | |
| 632 | void tick_cancel_sched_timer(int cpu) |
| 633 | { |
| 634 | struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); |
| 635 | |
| 636 | if (ts->sched_timer.base) |
| 637 | hrtimer_cancel(&ts->sched_timer); |
| 638 | ts->tick_stopped = 0; |
| 639 | ts->nohz_mode = NOHZ_MODE_INACTIVE; |
| 640 | } |
| 641 | #endif /* HIGH_RES_TIMERS */ |
| 642 | |
| 643 | /** |
| 644 | * Async notification about clocksource changes |
| 645 | */ |
| 646 | void tick_clock_notify(void) |
| 647 | { |
| 648 | int cpu; |
| 649 | |
| 650 | for_each_possible_cpu(cpu) |
| 651 | set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); |
| 652 | } |
| 653 | |
| 654 | /* |
| 655 | * Async notification about clock event changes |
| 656 | */ |
| 657 | void tick_oneshot_notify(void) |
| 658 | { |
| 659 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 660 | |
| 661 | set_bit(0, &ts->check_clocks); |
| 662 | } |
| 663 | |
| 664 | /** |
| 665 | * Check, if a change happened, which makes oneshot possible. |
| 666 | * |
| 667 | * Called cyclic from the hrtimer softirq (driven by the timer |
| 668 | * softirq) allow_nohz signals, that we can switch into low-res nohz |
| 669 | * mode, because high resolution timers are disabled (either compile |
| 670 | * or runtime). |
| 671 | */ |
| 672 | int tick_check_oneshot_change(int allow_nohz) |
| 673 | { |
| 674 | struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); |
| 675 | |
| 676 | if (!test_and_clear_bit(0, &ts->check_clocks)) |
| 677 | return 0; |
| 678 | |
| 679 | if (ts->nohz_mode != NOHZ_MODE_INACTIVE) |
| 680 | return 0; |
| 681 | |
| 682 | if (!timekeeping_is_continuous() || !tick_is_oneshot_available()) |
| 683 | return 0; |
| 684 | |
| 685 | if (!allow_nohz) |
| 686 | return 1; |
| 687 | |
| 688 | tick_nohz_switch_to_nohz(); |
| 689 | return 0; |
| 690 | } |