2 * linux/arch/x86-64/kernel/time.c
4 * "High Precision Event Timer" based timekeeping.
6 * Copyright (c) 1991,1992,1995 Linus Torvalds
7 * Copyright (c) 1994 Alan Modra
8 * Copyright (c) 1995 Markus Kuhn
9 * Copyright (c) 1996 Ingo Molnar
10 * Copyright (c) 1998 Andrea Arcangeli
11 * Copyright (c) 2002,2006 Vojtech Pavlik
12 * Copyright (c) 2003 Andi Kleen
13 * RTC support code taken from arch/i386/kernel/timers/time_hpet.c
16 #include <linux/kernel.h>
17 #include <linux/sched.h>
18 #include <linux/interrupt.h>
19 #include <linux/init.h>
20 #include <linux/mc146818rtc.h>
21 #include <linux/time.h>
22 #include <linux/ioport.h>
23 #include <linux/module.h>
24 #include <linux/device.h>
25 #include <linux/sysdev.h>
26 #include <linux/bcd.h>
27 #include <linux/notifier.h>
28 #include <linux/cpu.h>
29 #include <linux/kallsyms.h>
30 #include <linux/acpi.h>
32 #include <acpi/achware.h> /* for PM timer frequency */
33 #include <acpi/acpi_bus.h>
35 #include <asm/8253pit.h>
36 #include <asm/pgtable.h>
37 #include <asm/vsyscall.h>
38 #include <asm/timex.h>
39 #include <asm/proto.h>
41 #include <asm/sections.h>
42 #include <linux/cpufreq.h>
43 #include <linux/hpet.h>
46 #ifdef CONFIG_CPU_FREQ
47 static void cpufreq_delayed_get(void);
49 extern void i8254_timer_resume(void);
50 extern int using_apic_timer;
52 static char *timename = NULL;
54 DEFINE_SPINLOCK(rtc_lock);
55 EXPORT_SYMBOL(rtc_lock);
56 DEFINE_SPINLOCK(i8253_lock);
58 int nohpet __initdata = 0;
59 static int notsc __initdata = 0;
61 #define USEC_PER_TICK (USEC_PER_SEC / HZ)
62 #define NSEC_PER_TICK (NSEC_PER_SEC / HZ)
63 #define FSEC_PER_TICK (FSEC_PER_SEC / HZ)
65 #define NS_SCALE 10 /* 2^10, carefully chosen */
66 #define US_SCALE 32 /* 2^32, arbitralrily chosen */
68 unsigned int cpu_khz; /* TSC clocks / usec, not used here */
69 EXPORT_SYMBOL(cpu_khz);
70 static unsigned long hpet_period; /* fsecs / HPET clock */
71 unsigned long hpet_tick; /* HPET clocks / interrupt */
72 int hpet_use_timer; /* Use counter of hpet for time keeping, otherwise PIT */
73 unsigned long vxtime_hz = PIT_TICK_RATE;
74 int report_lost_ticks; /* command line option */
75 unsigned long long monotonic_base;
77 struct vxtime_data __vxtime __section_vxtime; /* for vsyscalls */
79 volatile unsigned long __jiffies __section_jiffies = INITIAL_JIFFIES;
80 unsigned long __wall_jiffies __section_wall_jiffies = INITIAL_JIFFIES;
81 struct timespec __xtime __section_xtime;
82 struct timezone __sys_tz __section_sys_tz;
85 * do_gettimeoffset() returns microseconds since last timer interrupt was
86 * triggered by hardware. A memory read of HPET is slower than a register read
87 * of TSC, but much more reliable. It's also synchronized to the timer
88 * interrupt. Note that do_gettimeoffset() may return more than hpet_tick, if a
89 * timer interrupt has happened already, but vxtime.trigger wasn't updated yet.
90 * This is not a problem, because jiffies hasn't updated either. They are bound
91 * together by xtime_lock.
94 static inline unsigned int do_gettimeoffset_tsc(void)
98 t = get_cycles_sync();
99 if (t < vxtime.last_tsc)
100 t = vxtime.last_tsc; /* hack */
101 x = ((t - vxtime.last_tsc) * vxtime.tsc_quot) >> US_SCALE;
105 static inline unsigned int do_gettimeoffset_hpet(void)
107 /* cap counter read to one tick to avoid inconsistencies */
108 unsigned long counter = hpet_readl(HPET_COUNTER) - vxtime.last;
109 return (min(counter,hpet_tick) * vxtime.quot) >> US_SCALE;
112 unsigned int (*do_gettimeoffset)(void) = do_gettimeoffset_tsc;
115 * This version of gettimeofday() has microsecond resolution and better than
116 * microsecond precision, as we're using at least a 10 MHz (usually 14.31818
120 void do_gettimeofday(struct timeval *tv)
122 unsigned long seq, t;
123 unsigned int sec, usec;
126 seq = read_seqbegin(&xtime_lock);
129 usec = xtime.tv_nsec / NSEC_PER_USEC;
131 /* i386 does some correction here to keep the clock
132 monotonous even when ntpd is fixing drift.
133 But they didn't work for me, there is a non monotonic
134 clock anyways with ntp.
135 I dropped all corrections now until a real solution can
136 be found. Note when you fix it here you need to do the same
137 in arch/x86_64/kernel/vsyscall.c and export all needed
138 variables in vmlinux.lds. -AK */
140 t = (jiffies - wall_jiffies) * USEC_PER_TICK +
144 } while (read_seqretry(&xtime_lock, seq));
146 tv->tv_sec = sec + usec / USEC_PER_SEC;
147 tv->tv_usec = usec % USEC_PER_SEC;
150 EXPORT_SYMBOL(do_gettimeofday);
153 * settimeofday() first undoes the correction that gettimeofday would do
154 * on the time, and then saves it. This is ugly, but has been like this for
158 int do_settimeofday(struct timespec *tv)
160 time_t wtm_sec, sec = tv->tv_sec;
161 long wtm_nsec, nsec = tv->tv_nsec;
163 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
166 write_seqlock_irq(&xtime_lock);
168 nsec -= do_gettimeoffset() * NSEC_PER_USEC +
169 (jiffies - wall_jiffies) * NSEC_PER_TICK;
171 wtm_sec = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
172 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
174 set_normalized_timespec(&xtime, sec, nsec);
175 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
179 write_sequnlock_irq(&xtime_lock);
184 EXPORT_SYMBOL(do_settimeofday);
186 unsigned long profile_pc(struct pt_regs *regs)
188 unsigned long pc = instruction_pointer(regs);
190 /* Assume the lock function has either no stack frame or a copy
192 Eflags always has bits 22 and up cleared unlike kernel addresses. */
193 if (!user_mode(regs) && in_lock_functions(pc)) {
194 unsigned long *sp = (unsigned long *)regs->rsp;
202 EXPORT_SYMBOL(profile_pc);
205 * In order to set the CMOS clock precisely, set_rtc_mmss has to be called 500
206 * ms after the second nowtime has started, because when nowtime is written
207 * into the registers of the CMOS clock, it will jump to the next second
208 * precisely 500 ms later. Check the Motorola MC146818A or Dallas DS12887 data
212 static void set_rtc_mmss(unsigned long nowtime)
214 int real_seconds, real_minutes, cmos_minutes;
215 unsigned char control, freq_select;
218 * IRQs are disabled when we're called from the timer interrupt,
219 * no need for spin_lock_irqsave()
222 spin_lock(&rtc_lock);
225 * Tell the clock it's being set and stop it.
228 control = CMOS_READ(RTC_CONTROL);
229 CMOS_WRITE(control | RTC_SET, RTC_CONTROL);
231 freq_select = CMOS_READ(RTC_FREQ_SELECT);
232 CMOS_WRITE(freq_select | RTC_DIV_RESET2, RTC_FREQ_SELECT);
234 cmos_minutes = CMOS_READ(RTC_MINUTES);
235 BCD_TO_BIN(cmos_minutes);
238 * since we're only adjusting minutes and seconds, don't interfere with hour
239 * overflow. This avoids messing with unknown time zones but requires your RTC
240 * not to be off by more than 15 minutes. Since we're calling it only when
241 * our clock is externally synchronized using NTP, this shouldn't be a problem.
244 real_seconds = nowtime % 60;
245 real_minutes = nowtime / 60;
246 if (((abs(real_minutes - cmos_minutes) + 15) / 30) & 1)
247 real_minutes += 30; /* correct for half hour time zone */
250 if (abs(real_minutes - cmos_minutes) >= 30) {
251 printk(KERN_WARNING "time.c: can't update CMOS clock "
252 "from %d to %d\n", cmos_minutes, real_minutes);
254 BIN_TO_BCD(real_seconds);
255 BIN_TO_BCD(real_minutes);
256 CMOS_WRITE(real_seconds, RTC_SECONDS);
257 CMOS_WRITE(real_minutes, RTC_MINUTES);
261 * The following flags have to be released exactly in this order, otherwise the
262 * DS12887 (popular MC146818A clone with integrated battery and quartz) will
263 * not reset the oscillator and will not update precisely 500 ms later. You
264 * won't find this mentioned in the Dallas Semiconductor data sheets, but who
265 * believes data sheets anyway ... -- Markus Kuhn
268 CMOS_WRITE(control, RTC_CONTROL);
269 CMOS_WRITE(freq_select, RTC_FREQ_SELECT);
271 spin_unlock(&rtc_lock);
275 /* monotonic_clock(): returns # of nanoseconds passed since time_init()
276 * Note: This function is required to return accurate
277 * time even in the absence of multiple timer ticks.
279 unsigned long long monotonic_clock(void)
282 u32 last_offset, this_offset, offset;
283 unsigned long long base;
285 if (vxtime.mode == VXTIME_HPET) {
287 seq = read_seqbegin(&xtime_lock);
289 last_offset = vxtime.last;
290 base = monotonic_base;
291 this_offset = hpet_readl(HPET_COUNTER);
292 } while (read_seqretry(&xtime_lock, seq));
293 offset = (this_offset - last_offset);
294 offset *= NSEC_PER_TICK / hpet_tick;
297 seq = read_seqbegin(&xtime_lock);
299 last_offset = vxtime.last_tsc;
300 base = monotonic_base;
301 } while (read_seqretry(&xtime_lock, seq));
302 this_offset = get_cycles_sync();
303 /* FIXME: 1000 or 1000000? */
304 offset = (this_offset - last_offset)*1000 / cpu_khz;
306 return base + offset;
308 EXPORT_SYMBOL(monotonic_clock);
310 static noinline void handle_lost_ticks(int lost, struct pt_regs *regs)
312 static long lost_count;
314 if (report_lost_ticks) {
315 printk(KERN_WARNING "time.c: Lost %d timer tick(s)! ", lost);
316 print_symbol("rip %s)\n", regs->rip);
319 if (lost_count == 1000 && !warned) {
320 printk(KERN_WARNING "warning: many lost ticks.\n"
321 KERN_WARNING "Your time source seems to be instable or "
322 "some driver is hogging interupts\n");
323 print_symbol("rip %s\n", regs->rip);
324 if (vxtime.mode == VXTIME_TSC && vxtime.hpet_address) {
325 printk(KERN_WARNING "Falling back to HPET\n");
327 vxtime.last = hpet_readl(HPET_T0_CMP) -
330 vxtime.last = hpet_readl(HPET_COUNTER);
331 vxtime.mode = VXTIME_HPET;
332 do_gettimeoffset = do_gettimeoffset_hpet;
334 /* else should fall back to PIT, but code missing. */
339 #ifdef CONFIG_CPU_FREQ
340 /* In some cases the CPU can change frequency without us noticing
341 Give cpufreq a change to catch up. */
342 if ((lost_count+1) % 25 == 0)
343 cpufreq_delayed_get();
347 void main_timer_handler(struct pt_regs *regs)
349 static unsigned long rtc_update = 0;
351 int delay = 0, offset = 0, lost = 0;
354 * Here we are in the timer irq handler. We have irqs locally disabled (so we
355 * don't need spin_lock_irqsave()) but we don't know if the timer_bh is running
356 * on the other CPU, so we need a lock. We also need to lock the vsyscall
357 * variables, because both do_timer() and us change them -arca+vojtech
360 write_seqlock(&xtime_lock);
362 if (vxtime.hpet_address)
363 offset = hpet_readl(HPET_COUNTER);
365 if (hpet_use_timer) {
366 /* if we're using the hpet timer functionality,
367 * we can more accurately know the counter value
368 * when the timer interrupt occured.
370 offset = hpet_readl(HPET_T0_CMP) - hpet_tick;
371 delay = hpet_readl(HPET_COUNTER) - offset;
372 } else if (!pmtmr_ioport) {
373 spin_lock(&i8253_lock);
376 delay |= inb(0x40) << 8;
377 spin_unlock(&i8253_lock);
378 delay = LATCH - 1 - delay;
381 tsc = get_cycles_sync();
383 if (vxtime.mode == VXTIME_HPET) {
384 if (offset - vxtime.last > hpet_tick) {
385 lost = (offset - vxtime.last) / hpet_tick - 1;
389 (offset - vxtime.last) * NSEC_PER_TICK / hpet_tick;
391 vxtime.last = offset;
392 #ifdef CONFIG_X86_PM_TIMER
393 } else if (vxtime.mode == VXTIME_PMTMR) {
394 lost = pmtimer_mark_offset();
397 offset = (((tsc - vxtime.last_tsc) *
398 vxtime.tsc_quot) >> US_SCALE) - USEC_PER_TICK;
403 if (offset > USEC_PER_TICK) {
404 lost = offset / USEC_PER_TICK;
405 offset %= USEC_PER_TICK;
408 /* FIXME: 1000 or 1000000? */
409 monotonic_base += (tsc - vxtime.last_tsc) * 1000000 / cpu_khz;
411 vxtime.last_tsc = tsc - vxtime.quot * delay / vxtime.tsc_quot;
413 if ((((tsc - vxtime.last_tsc) *
414 vxtime.tsc_quot) >> US_SCALE) < offset)
415 vxtime.last_tsc = tsc -
416 (((long) offset << US_SCALE) / vxtime.tsc_quot) - 1;
420 handle_lost_ticks(lost, regs);
425 * Do the timer stuff.
430 update_process_times(user_mode(regs));
434 * In the SMP case we use the local APIC timer interrupt to do the profiling,
435 * except when we simulate SMP mode on a uniprocessor system, in that case we
436 * have to call the local interrupt handler.
439 if (!using_apic_timer)
440 smp_local_timer_interrupt(regs);
443 * If we have an externally synchronized Linux clock, then update CMOS clock
444 * accordingly every ~11 minutes. set_rtc_mmss() will be called in the jiffy
445 * closest to exactly 500 ms before the next second. If the update fails, we
446 * don't care, as it'll be updated on the next turn, and the problem (time way
447 * off) isn't likely to go away much sooner anyway.
450 if (ntp_synced() && xtime.tv_sec > rtc_update &&
451 abs(xtime.tv_nsec - 500000000) <= tick_nsec / 2) {
452 set_rtc_mmss(xtime.tv_sec);
453 rtc_update = xtime.tv_sec + 660;
456 write_sequnlock(&xtime_lock);
459 static irqreturn_t timer_interrupt(int irq, void *dev_id, struct pt_regs *regs)
461 if (apic_runs_main_timer > 1)
463 main_timer_handler(regs);
464 if (using_apic_timer)
465 smp_send_timer_broadcast_ipi();
469 static unsigned int cyc2ns_scale __read_mostly;
471 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
473 cyc2ns_scale = (NSEC_PER_MSEC << NS_SCALE) / cpu_khz;
476 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
478 return (cyc * cyc2ns_scale) >> NS_SCALE;
481 unsigned long long sched_clock(void)
486 /* Don't do a HPET read here. Using TSC always is much faster
487 and HPET may not be mapped yet when the scheduler first runs.
488 Disadvantage is a small drift between CPUs in some configurations,
489 but that should be tolerable. */
490 if (__vxtime.mode == VXTIME_HPET)
491 return (hpet_readl(HPET_COUNTER) * vxtime.quot) >> US_SCALE;
494 /* Could do CPU core sync here. Opteron can execute rdtsc speculatively,
495 which means it is not completely exact and may not be monotonous between
496 CPUs. But the errors should be too small to matter for scheduling
500 return cycles_2_ns(a);
503 static unsigned long get_cmos_time(void)
505 unsigned int year, mon, day, hour, min, sec;
507 unsigned extyear = 0;
509 spin_lock_irqsave(&rtc_lock, flags);
512 sec = CMOS_READ(RTC_SECONDS);
513 min = CMOS_READ(RTC_MINUTES);
514 hour = CMOS_READ(RTC_HOURS);
515 day = CMOS_READ(RTC_DAY_OF_MONTH);
516 mon = CMOS_READ(RTC_MONTH);
517 year = CMOS_READ(RTC_YEAR);
519 if (acpi_fadt.revision >= FADT2_REVISION_ID &&
521 extyear = CMOS_READ(acpi_fadt.century);
523 } while (sec != CMOS_READ(RTC_SECONDS));
525 spin_unlock_irqrestore(&rtc_lock, flags);
528 * We know that x86-64 always uses BCD format, no need to check the
542 printk(KERN_INFO "Extended CMOS year: %d\n", extyear);
545 * x86-64 systems only exists since 2002.
546 * This will work up to Dec 31, 2100
551 return mktime(year, mon, day, hour, min, sec);
554 #ifdef CONFIG_CPU_FREQ
556 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
559 RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
560 not that important because current Opteron setups do not support
561 scaling on SMP anyroads.
563 Should fix up last_tsc too. Currently gettimeofday in the
564 first tick after the change will be slightly wrong. */
566 #include <linux/workqueue.h>
568 static unsigned int cpufreq_delayed_issched = 0;
569 static unsigned int cpufreq_init = 0;
570 static struct work_struct cpufreq_delayed_get_work;
572 static void handle_cpufreq_delayed_get(void *v)
575 for_each_online_cpu(cpu) {
578 cpufreq_delayed_issched = 0;
581 /* if we notice lost ticks, schedule a call to cpufreq_get() as it tries
582 * to verify the CPU frequency the timing core thinks the CPU is running
583 * at is still correct.
585 static void cpufreq_delayed_get(void)
588 if (cpufreq_init && !cpufreq_delayed_issched) {
589 cpufreq_delayed_issched = 1;
593 "Losing some ticks... checking if CPU frequency changed.\n");
595 schedule_work(&cpufreq_delayed_get_work);
599 static unsigned int ref_freq = 0;
600 static unsigned long loops_per_jiffy_ref = 0;
602 static unsigned long cpu_khz_ref = 0;
604 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
607 struct cpufreq_freqs *freq = data;
608 unsigned long *lpj, dummy;
610 if (cpu_has(&cpu_data[freq->cpu], X86_FEATURE_CONSTANT_TSC))
614 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
616 lpj = &cpu_data[freq->cpu].loops_per_jiffy;
618 lpj = &boot_cpu_data.loops_per_jiffy;
622 ref_freq = freq->old;
623 loops_per_jiffy_ref = *lpj;
624 cpu_khz_ref = cpu_khz;
626 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
627 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
628 (val == CPUFREQ_RESUMECHANGE)) {
630 cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
632 cpu_khz = cpufreq_scale(cpu_khz_ref, ref_freq, freq->new);
633 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
634 vxtime.tsc_quot = (USEC_PER_MSEC << US_SCALE) / cpu_khz;
637 set_cyc2ns_scale(cpu_khz_ref);
642 static struct notifier_block time_cpufreq_notifier_block = {
643 .notifier_call = time_cpufreq_notifier
646 static int __init cpufreq_tsc(void)
648 INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get, NULL);
649 if (!cpufreq_register_notifier(&time_cpufreq_notifier_block,
650 CPUFREQ_TRANSITION_NOTIFIER))
655 core_initcall(cpufreq_tsc);
660 * calibrate_tsc() calibrates the processor TSC in a very simple way, comparing
661 * it to the HPET timer of known frequency.
664 #define TICK_COUNT 100000000
666 static unsigned int __init hpet_calibrate_tsc(void)
668 int tsc_start, hpet_start;
669 int tsc_now, hpet_now;
672 local_irq_save(flags);
675 hpet_start = hpet_readl(HPET_COUNTER);
680 hpet_now = hpet_readl(HPET_COUNTER);
681 tsc_now = get_cycles_sync();
682 local_irq_restore(flags);
683 } while ((tsc_now - tsc_start) < TICK_COUNT &&
684 (hpet_now - hpet_start) < TICK_COUNT);
686 return (tsc_now - tsc_start) * 1000000000L
687 / ((hpet_now - hpet_start) * hpet_period / 1000);
692 * pit_calibrate_tsc() uses the speaker output (channel 2) of
693 * the PIT. This is better than using the timer interrupt output,
694 * because we can read the value of the speaker with just one inb(),
695 * where we need three i/o operations for the interrupt channel.
696 * We count how many ticks the TSC does in 50 ms.
699 static unsigned int __init pit_calibrate_tsc(void)
701 unsigned long start, end;
704 spin_lock_irqsave(&i8253_lock, flags);
706 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
709 outb((PIT_TICK_RATE / (1000 / 50)) & 0xff, 0x42);
710 outb((PIT_TICK_RATE / (1000 / 50)) >> 8, 0x42);
711 start = get_cycles_sync();
712 while ((inb(0x61) & 0x20) == 0);
713 end = get_cycles_sync();
715 spin_unlock_irqrestore(&i8253_lock, flags);
717 return (end - start) / 50;
721 static __init int late_hpet_init(void)
726 if (!vxtime.hpet_address)
729 memset(&hd, 0, sizeof (hd));
731 ntimer = hpet_readl(HPET_ID);
732 ntimer = (ntimer & HPET_ID_NUMBER) >> HPET_ID_NUMBER_SHIFT;
736 * Register with driver.
737 * Timer0 and Timer1 is used by platform.
739 hd.hd_phys_address = vxtime.hpet_address;
740 hd.hd_address = (void __iomem *)fix_to_virt(FIX_HPET_BASE);
741 hd.hd_nirqs = ntimer;
742 hd.hd_flags = HPET_DATA_PLATFORM;
743 hpet_reserve_timer(&hd, 0);
744 #ifdef CONFIG_HPET_EMULATE_RTC
745 hpet_reserve_timer(&hd, 1);
747 hd.hd_irq[0] = HPET_LEGACY_8254;
748 hd.hd_irq[1] = HPET_LEGACY_RTC;
751 struct hpet_timer *timer;
754 hpet = (struct hpet *) fix_to_virt(FIX_HPET_BASE);
755 timer = &hpet->hpet_timers[2];
756 for (i = 2; i < ntimer; timer++, i++)
757 hd.hd_irq[i] = (timer->hpet_config &
758 Tn_INT_ROUTE_CNF_MASK) >>
759 Tn_INT_ROUTE_CNF_SHIFT;
766 fs_initcall(late_hpet_init);
769 static int hpet_timer_stop_set_go(unsigned long tick)
774 * Stop the timers and reset the main counter.
777 cfg = hpet_readl(HPET_CFG);
778 cfg &= ~(HPET_CFG_ENABLE | HPET_CFG_LEGACY);
779 hpet_writel(cfg, HPET_CFG);
780 hpet_writel(0, HPET_COUNTER);
781 hpet_writel(0, HPET_COUNTER + 4);
784 * Set up timer 0, as periodic with first interrupt to happen at hpet_tick,
785 * and period also hpet_tick.
787 if (hpet_use_timer) {
788 hpet_writel(HPET_TN_ENABLE | HPET_TN_PERIODIC | HPET_TN_SETVAL |
789 HPET_TN_32BIT, HPET_T0_CFG);
790 hpet_writel(hpet_tick, HPET_T0_CMP); /* next interrupt */
791 hpet_writel(hpet_tick, HPET_T0_CMP); /* period */
792 cfg |= HPET_CFG_LEGACY;
798 cfg |= HPET_CFG_ENABLE;
799 hpet_writel(cfg, HPET_CFG);
804 static int hpet_init(void)
808 if (!vxtime.hpet_address)
810 set_fixmap_nocache(FIX_HPET_BASE, vxtime.hpet_address);
811 __set_fixmap(VSYSCALL_HPET, vxtime.hpet_address, PAGE_KERNEL_VSYSCALL_NOCACHE);
814 * Read the period, compute tick and quotient.
817 id = hpet_readl(HPET_ID);
819 if (!(id & HPET_ID_VENDOR) || !(id & HPET_ID_NUMBER))
822 hpet_period = hpet_readl(HPET_PERIOD);
823 if (hpet_period < 100000 || hpet_period > 100000000)
826 hpet_tick = (FSEC_PER_TICK + hpet_period / 2) / hpet_period;
828 hpet_use_timer = (id & HPET_ID_LEGSUP);
830 return hpet_timer_stop_set_go(hpet_tick);
833 static int hpet_reenable(void)
835 return hpet_timer_stop_set_go(hpet_tick);
838 #define PIT_MODE 0x43
841 static void __init __pit_init(int val, u8 mode)
845 spin_lock_irqsave(&i8253_lock, flags);
846 outb_p(mode, PIT_MODE);
847 outb_p(val & 0xff, PIT_CH0); /* LSB */
848 outb_p(val >> 8, PIT_CH0); /* MSB */
849 spin_unlock_irqrestore(&i8253_lock, flags);
852 void __init pit_init(void)
854 __pit_init(LATCH, 0x34); /* binary, mode 2, LSB/MSB, ch 0 */
857 void __init pit_stop_interrupt(void)
859 __pit_init(0, 0x30); /* mode 0 */
862 void __init stop_timer_interrupt(void)
865 if (vxtime.hpet_address) {
867 hpet_timer_stop_set_go(0);
870 pit_stop_interrupt();
872 printk(KERN_INFO "timer: %s interrupt stopped.\n", name);
875 int __init time_setup(char *str)
877 report_lost_ticks = 1;
881 static struct irqaction irq0 = {
882 timer_interrupt, IRQF_DISABLED, CPU_MASK_NONE, "timer", NULL, NULL
886 time_cpu_notifier(struct notifier_block *nb, unsigned long action, void *hcpu)
888 unsigned cpu = (unsigned long) hcpu;
889 if (action == CPU_ONLINE)
890 vsyscall_set_cpu(cpu);
894 void __init time_init(void)
897 vxtime.hpet_address = 0;
899 xtime.tv_sec = get_cmos_time();
902 set_normalized_timespec(&wall_to_monotonic,
903 -xtime.tv_sec, -xtime.tv_nsec);
906 vxtime_hz = (FSEC_PER_SEC + hpet_period / 2) / hpet_period;
908 vxtime.hpet_address = 0;
910 if (hpet_use_timer) {
911 /* set tick_nsec to use the proper rate for HPET */
912 tick_nsec = TICK_NSEC_HPET;
913 cpu_khz = hpet_calibrate_tsc();
915 #ifdef CONFIG_X86_PM_TIMER
916 } else if (pmtmr_ioport && !vxtime.hpet_address) {
917 vxtime_hz = PM_TIMER_FREQUENCY;
920 cpu_khz = pit_calibrate_tsc();
924 cpu_khz = pit_calibrate_tsc();
928 vxtime.mode = VXTIME_TSC;
929 vxtime.quot = (USEC_PER_SEC << US_SCALE) / vxtime_hz;
930 vxtime.tsc_quot = (USEC_PER_MSEC << US_SCALE) / cpu_khz;
931 vxtime.last_tsc = get_cycles_sync();
934 set_cyc2ns_scale(cpu_khz);
936 hotcpu_notifier(time_cpu_notifier, 0);
937 time_cpu_notifier(NULL, CPU_ONLINE, (void *)(long)smp_processor_id());
945 * Make an educated guess if the TSC is trustworthy and synchronized
948 __cpuinit int unsynchronized_tsc(void)
951 if (apic_is_clustered_box())
954 /* Most intel systems have synchronized TSCs except for
955 multi node systems */
956 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) {
958 /* But TSC doesn't tick in C3 so don't use it there */
959 if (acpi_fadt.length > 0 && acpi_fadt.plvl3_lat < 100)
965 /* Assume multi socket systems are not synchronized */
966 return num_present_cpus() > 1;
970 * Decide what mode gettimeofday should use.
972 void time_init_gtod(void)
976 if (unsynchronized_tsc())
979 if (cpu_has(&boot_cpu_data, X86_FEATURE_RDTSCP))
980 vgetcpu_mode = VGETCPU_RDTSCP;
982 vgetcpu_mode = VGETCPU_LSL;
984 if (vxtime.hpet_address && notsc) {
985 timetype = hpet_use_timer ? "HPET" : "PIT/HPET";
987 vxtime.last = hpet_readl(HPET_T0_CMP) - hpet_tick;
989 vxtime.last = hpet_readl(HPET_COUNTER);
990 vxtime.mode = VXTIME_HPET;
991 do_gettimeoffset = do_gettimeoffset_hpet;
992 #ifdef CONFIG_X86_PM_TIMER
993 /* Using PM for gettimeofday is quite slow, but we have no other
994 choice because the TSC is too unreliable on some systems. */
995 } else if (pmtmr_ioport && !vxtime.hpet_address && notsc) {
997 do_gettimeoffset = do_gettimeoffset_pm;
998 vxtime.mode = VXTIME_PMTMR;
1000 printk(KERN_INFO "Disabling vsyscall due to use of PM timer\n");
1003 timetype = hpet_use_timer ? "HPET/TSC" : "PIT/TSC";
1004 vxtime.mode = VXTIME_TSC;
1007 printk(KERN_INFO "time.c: Using %ld.%06ld MHz WALL %s GTOD %s timer.\n",
1008 vxtime_hz / 1000000, vxtime_hz % 1000000, timename, timetype);
1009 printk(KERN_INFO "time.c: Detected %d.%03d MHz processor.\n",
1010 cpu_khz / 1000, cpu_khz % 1000);
1011 vxtime.quot = (USEC_PER_SEC << US_SCALE) / vxtime_hz;
1012 vxtime.tsc_quot = (USEC_PER_MSEC << US_SCALE) / cpu_khz;
1013 vxtime.last_tsc = get_cycles_sync();
1015 set_cyc2ns_scale(cpu_khz);
1018 __setup("report_lost_ticks", time_setup);
1020 static long clock_cmos_diff;
1021 static unsigned long sleep_start;
1024 * sysfs support for the timer.
1027 static int timer_suspend(struct sys_device *dev, pm_message_t state)
1030 * Estimate time zone so that set_time can update the clock
1032 long cmos_time = get_cmos_time();
1034 clock_cmos_diff = -cmos_time;
1035 clock_cmos_diff += get_seconds();
1036 sleep_start = cmos_time;
1040 static int timer_resume(struct sys_device *dev)
1042 unsigned long flags;
1044 unsigned long ctime = get_cmos_time();
1045 unsigned long sleep_length = (ctime - sleep_start) * HZ;
1047 if (vxtime.hpet_address)
1050 i8254_timer_resume();
1052 sec = ctime + clock_cmos_diff;
1053 write_seqlock_irqsave(&xtime_lock,flags);
1056 if (vxtime.mode == VXTIME_HPET) {
1058 vxtime.last = hpet_readl(HPET_T0_CMP) - hpet_tick;
1060 vxtime.last = hpet_readl(HPET_COUNTER);
1061 #ifdef CONFIG_X86_PM_TIMER
1062 } else if (vxtime.mode == VXTIME_PMTMR) {
1066 vxtime.last_tsc = get_cycles_sync();
1067 write_sequnlock_irqrestore(&xtime_lock,flags);
1068 jiffies += sleep_length;
1069 wall_jiffies += sleep_length;
1070 monotonic_base += sleep_length * (NSEC_PER_SEC/HZ);
1071 touch_softlockup_watchdog();
1075 static struct sysdev_class timer_sysclass = {
1076 .resume = timer_resume,
1077 .suspend = timer_suspend,
1078 set_kset_name("timer"),
1081 /* XXX this driverfs stuff should probably go elsewhere later -john */
1082 static struct sys_device device_timer = {
1084 .cls = &timer_sysclass,
1087 static int time_init_device(void)
1089 int error = sysdev_class_register(&timer_sysclass);
1091 error = sysdev_register(&device_timer);
1095 device_initcall(time_init_device);
1097 #ifdef CONFIG_HPET_EMULATE_RTC
1098 /* HPET in LegacyReplacement Mode eats up RTC interrupt line. When, HPET
1099 * is enabled, we support RTC interrupt functionality in software.
1100 * RTC has 3 kinds of interrupts:
1101 * 1) Update Interrupt - generate an interrupt, every sec, when RTC clock
1103 * 2) Alarm Interrupt - generate an interrupt at a specific time of day
1104 * 3) Periodic Interrupt - generate periodic interrupt, with frequencies
1105 * 2Hz-8192Hz (2Hz-64Hz for non-root user) (all freqs in powers of 2)
1106 * (1) and (2) above are implemented using polling at a frequency of
1107 * 64 Hz. The exact frequency is a tradeoff between accuracy and interrupt
1108 * overhead. (DEFAULT_RTC_INT_FREQ)
1109 * For (3), we use interrupts at 64Hz or user specified periodic
1110 * frequency, whichever is higher.
1112 #include <linux/rtc.h>
1114 #define DEFAULT_RTC_INT_FREQ 64
1115 #define RTC_NUM_INTS 1
1117 static unsigned long UIE_on;
1118 static unsigned long prev_update_sec;
1120 static unsigned long AIE_on;
1121 static struct rtc_time alarm_time;
1123 static unsigned long PIE_on;
1124 static unsigned long PIE_freq = DEFAULT_RTC_INT_FREQ;
1125 static unsigned long PIE_count;
1127 static unsigned long hpet_rtc_int_freq; /* RTC interrupt frequency */
1128 static unsigned int hpet_t1_cmp; /* cached comparator register */
1130 int is_hpet_enabled(void)
1132 return vxtime.hpet_address != 0;
1136 * Timer 1 for RTC, we do not use periodic interrupt feature,
1137 * even if HPET supports periodic interrupts on Timer 1.
1138 * The reason being, to set up a periodic interrupt in HPET, we need to
1139 * stop the main counter. And if we do that everytime someone diables/enables
1140 * RTC, we will have adverse effect on main kernel timer running on Timer 0.
1141 * So, for the time being, simulate the periodic interrupt in software.
1143 * hpet_rtc_timer_init() is called for the first time and during subsequent
1144 * interuppts reinit happens through hpet_rtc_timer_reinit().
1146 int hpet_rtc_timer_init(void)
1148 unsigned int cfg, cnt;
1149 unsigned long flags;
1151 if (!is_hpet_enabled())
1154 * Set the counter 1 and enable the interrupts.
1156 if (PIE_on && (PIE_freq > DEFAULT_RTC_INT_FREQ))
1157 hpet_rtc_int_freq = PIE_freq;
1159 hpet_rtc_int_freq = DEFAULT_RTC_INT_FREQ;
1161 local_irq_save(flags);
1162 cnt = hpet_readl(HPET_COUNTER);
1163 cnt += ((hpet_tick*HZ)/hpet_rtc_int_freq);
1164 hpet_writel(cnt, HPET_T1_CMP);
1166 local_irq_restore(flags);
1168 cfg = hpet_readl(HPET_T1_CFG);
1169 cfg &= ~HPET_TN_PERIODIC;
1170 cfg |= HPET_TN_ENABLE | HPET_TN_32BIT;
1171 hpet_writel(cfg, HPET_T1_CFG);
1176 static void hpet_rtc_timer_reinit(void)
1178 unsigned int cfg, cnt;
1180 if (unlikely(!(PIE_on | AIE_on | UIE_on))) {
1181 cfg = hpet_readl(HPET_T1_CFG);
1182 cfg &= ~HPET_TN_ENABLE;
1183 hpet_writel(cfg, HPET_T1_CFG);
1187 if (PIE_on && (PIE_freq > DEFAULT_RTC_INT_FREQ))
1188 hpet_rtc_int_freq = PIE_freq;
1190 hpet_rtc_int_freq = DEFAULT_RTC_INT_FREQ;
1192 /* It is more accurate to use the comparator value than current count.*/
1194 cnt += hpet_tick*HZ/hpet_rtc_int_freq;
1195 hpet_writel(cnt, HPET_T1_CMP);
1200 * The functions below are called from rtc driver.
1201 * Return 0 if HPET is not being used.
1202 * Otherwise do the necessary changes and return 1.
1204 int hpet_mask_rtc_irq_bit(unsigned long bit_mask)
1206 if (!is_hpet_enabled())
1209 if (bit_mask & RTC_UIE)
1211 if (bit_mask & RTC_PIE)
1213 if (bit_mask & RTC_AIE)
1219 int hpet_set_rtc_irq_bit(unsigned long bit_mask)
1221 int timer_init_reqd = 0;
1223 if (!is_hpet_enabled())
1226 if (!(PIE_on | AIE_on | UIE_on))
1227 timer_init_reqd = 1;
1229 if (bit_mask & RTC_UIE) {
1232 if (bit_mask & RTC_PIE) {
1236 if (bit_mask & RTC_AIE) {
1240 if (timer_init_reqd)
1241 hpet_rtc_timer_init();
1246 int hpet_set_alarm_time(unsigned char hrs, unsigned char min, unsigned char sec)
1248 if (!is_hpet_enabled())
1251 alarm_time.tm_hour = hrs;
1252 alarm_time.tm_min = min;
1253 alarm_time.tm_sec = sec;
1258 int hpet_set_periodic_freq(unsigned long freq)
1260 if (!is_hpet_enabled())
1269 int hpet_rtc_dropped_irq(void)
1271 if (!is_hpet_enabled())
1277 irqreturn_t hpet_rtc_interrupt(int irq, void *dev_id, struct pt_regs *regs)
1279 struct rtc_time curr_time;
1280 unsigned long rtc_int_flag = 0;
1281 int call_rtc_interrupt = 0;
1283 hpet_rtc_timer_reinit();
1285 if (UIE_on | AIE_on) {
1286 rtc_get_rtc_time(&curr_time);
1289 if (curr_time.tm_sec != prev_update_sec) {
1290 /* Set update int info, call real rtc int routine */
1291 call_rtc_interrupt = 1;
1292 rtc_int_flag = RTC_UF;
1293 prev_update_sec = curr_time.tm_sec;
1298 if (PIE_count >= hpet_rtc_int_freq/PIE_freq) {
1299 /* Set periodic int info, call real rtc int routine */
1300 call_rtc_interrupt = 1;
1301 rtc_int_flag |= RTC_PF;
1306 if ((curr_time.tm_sec == alarm_time.tm_sec) &&
1307 (curr_time.tm_min == alarm_time.tm_min) &&
1308 (curr_time.tm_hour == alarm_time.tm_hour)) {
1309 /* Set alarm int info, call real rtc int routine */
1310 call_rtc_interrupt = 1;
1311 rtc_int_flag |= RTC_AF;
1314 if (call_rtc_interrupt) {
1315 rtc_int_flag |= (RTC_IRQF | (RTC_NUM_INTS << 8));
1316 rtc_interrupt(rtc_int_flag, dev_id, regs);
1322 static int __init nohpet_setup(char *s)
1328 __setup("nohpet", nohpet_setup);
1330 int __init notsc_setup(char *s)
1336 __setup("notsc", notsc_setup);