[PATCH] x86_64: Explain why HPET T0_CMP register is written twice
[linux-2.6/sactl.git] / arch / x86_64 / kernel / time.c
blobee622ac0376b993689c05732bae8ab5dda7d764b
1 /*
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 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/kallsyms.h>
28 #include <linux/acpi.h>
29 #ifdef CONFIG_ACPI
30 #include <acpi/achware.h> /* for PM timer frequency */
31 #endif
32 #include <asm/8253pit.h>
33 #include <asm/pgtable.h>
34 #include <asm/vsyscall.h>
35 #include <asm/timex.h>
36 #include <asm/proto.h>
37 #include <asm/hpet.h>
38 #include <asm/sections.h>
39 #include <linux/cpufreq.h>
40 #include <linux/hpet.h>
41 #ifdef CONFIG_X86_LOCAL_APIC
42 #include <asm/apic.h>
43 #endif
45 #ifdef CONFIG_CPU_FREQ
46 static void cpufreq_delayed_get(void);
47 #endif
48 extern void i8254_timer_resume(void);
49 extern int using_apic_timer;
51 static char *time_init_gtod(void);
53 DEFINE_SPINLOCK(rtc_lock);
54 DEFINE_SPINLOCK(i8253_lock);
56 int nohpet __initdata = 0;
57 static int notsc __initdata = 0;
59 #define USEC_PER_TICK (USEC_PER_SEC / HZ)
60 #define NSEC_PER_TICK (NSEC_PER_SEC / HZ)
61 #define FSEC_PER_TICK (FSEC_PER_SEC / HZ)
63 #define NS_SCALE 10 /* 2^10, carefully chosen */
64 #define US_SCALE 32 /* 2^32, arbitralrily chosen */
66 unsigned int cpu_khz; /* TSC clocks / usec, not used here */
67 static unsigned long hpet_period; /* fsecs / HPET clock */
68 unsigned long hpet_tick; /* HPET clocks / interrupt */
69 int hpet_use_timer; /* Use counter of hpet for time keeping, otherwise PIT */
70 unsigned long vxtime_hz = PIT_TICK_RATE;
71 int report_lost_ticks; /* command line option */
72 unsigned long long monotonic_base;
74 struct vxtime_data __vxtime __section_vxtime; /* for vsyscalls */
76 volatile unsigned long __jiffies __section_jiffies = INITIAL_JIFFIES;
77 unsigned long __wall_jiffies __section_wall_jiffies = INITIAL_JIFFIES;
78 struct timespec __xtime __section_xtime;
79 struct timezone __sys_tz __section_sys_tz;
82 * do_gettimeoffset() returns microseconds since last timer interrupt was
83 * triggered by hardware. A memory read of HPET is slower than a register read
84 * of TSC, but much more reliable. It's also synchronized to the timer
85 * interrupt. Note that do_gettimeoffset() may return more than hpet_tick, if a
86 * timer interrupt has happened already, but vxtime.trigger wasn't updated yet.
87 * This is not a problem, because jiffies hasn't updated either. They are bound
88 * together by xtime_lock.
91 static inline unsigned int do_gettimeoffset_tsc(void)
93 unsigned long t;
94 unsigned long x;
95 t = get_cycles_sync();
96 if (t < vxtime.last_tsc)
97 t = vxtime.last_tsc; /* hack */
98 x = ((t - vxtime.last_tsc) * vxtime.tsc_quot) >> US_SCALE;
99 return x;
102 static inline unsigned int do_gettimeoffset_hpet(void)
104 /* cap counter read to one tick to avoid inconsistencies */
105 unsigned long counter = hpet_readl(HPET_COUNTER) - vxtime.last;
106 return (min(counter,hpet_tick) * vxtime.quot) >> US_SCALE;
109 unsigned int (*do_gettimeoffset)(void) = do_gettimeoffset_tsc;
112 * This version of gettimeofday() has microsecond resolution and better than
113 * microsecond precision, as we're using at least a 10 MHz (usually 14.31818
114 * MHz) HPET timer.
117 void do_gettimeofday(struct timeval *tv)
119 unsigned long seq, t;
120 unsigned int sec, usec;
122 do {
123 seq = read_seqbegin(&xtime_lock);
125 sec = xtime.tv_sec;
126 usec = xtime.tv_nsec / NSEC_PER_USEC;
128 /* i386 does some correction here to keep the clock
129 monotonous even when ntpd is fixing drift.
130 But they didn't work for me, there is a non monotonic
131 clock anyways with ntp.
132 I dropped all corrections now until a real solution can
133 be found. Note when you fix it here you need to do the same
134 in arch/x86_64/kernel/vsyscall.c and export all needed
135 variables in vmlinux.lds. -AK */
137 t = (jiffies - wall_jiffies) * USEC_PER_TICK +
138 do_gettimeoffset();
139 usec += t;
141 } while (read_seqretry(&xtime_lock, seq));
143 tv->tv_sec = sec + usec / USEC_PER_SEC;
144 tv->tv_usec = usec % USEC_PER_SEC;
147 EXPORT_SYMBOL(do_gettimeofday);
150 * settimeofday() first undoes the correction that gettimeofday would do
151 * on the time, and then saves it. This is ugly, but has been like this for
152 * ages already.
155 int do_settimeofday(struct timespec *tv)
157 time_t wtm_sec, sec = tv->tv_sec;
158 long wtm_nsec, nsec = tv->tv_nsec;
160 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
161 return -EINVAL;
163 write_seqlock_irq(&xtime_lock);
165 nsec -= do_gettimeoffset() * NSEC_PER_USEC +
166 (jiffies - wall_jiffies) * NSEC_PER_TICK;
168 wtm_sec = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
169 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
171 set_normalized_timespec(&xtime, sec, nsec);
172 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
174 ntp_clear();
176 write_sequnlock_irq(&xtime_lock);
177 clock_was_set();
178 return 0;
181 EXPORT_SYMBOL(do_settimeofday);
183 unsigned long profile_pc(struct pt_regs *regs)
185 unsigned long pc = instruction_pointer(regs);
187 /* Assume the lock function has either no stack frame or only a single
188 word. This checks if the address on the stack looks like a kernel
189 text address.
190 There is a small window for false hits, but in that case the tick
191 is just accounted to the spinlock function.
192 Better would be to write these functions in assembler again
193 and check exactly. */
194 if (in_lock_functions(pc)) {
195 char *v = *(char **)regs->rsp;
196 if ((v >= _stext && v <= _etext) ||
197 (v >= _sinittext && v <= _einittext) ||
198 (v >= (char *)MODULES_VADDR && v <= (char *)MODULES_END))
199 return (unsigned long)v;
200 return ((unsigned long *)regs->rsp)[1];
202 return pc;
204 EXPORT_SYMBOL(profile_pc);
207 * In order to set the CMOS clock precisely, set_rtc_mmss has to be called 500
208 * ms after the second nowtime has started, because when nowtime is written
209 * into the registers of the CMOS clock, it will jump to the next second
210 * precisely 500 ms later. Check the Motorola MC146818A or Dallas DS12887 data
211 * sheet for details.
214 static void set_rtc_mmss(unsigned long nowtime)
216 int real_seconds, real_minutes, cmos_minutes;
217 unsigned char control, freq_select;
220 * IRQs are disabled when we're called from the timer interrupt,
221 * no need for spin_lock_irqsave()
224 spin_lock(&rtc_lock);
227 * Tell the clock it's being set and stop it.
230 control = CMOS_READ(RTC_CONTROL);
231 CMOS_WRITE(control | RTC_SET, RTC_CONTROL);
233 freq_select = CMOS_READ(RTC_FREQ_SELECT);
234 CMOS_WRITE(freq_select | RTC_DIV_RESET2, RTC_FREQ_SELECT);
236 cmos_minutes = CMOS_READ(RTC_MINUTES);
237 BCD_TO_BIN(cmos_minutes);
240 * since we're only adjusting minutes and seconds, don't interfere with hour
241 * overflow. This avoids messing with unknown time zones but requires your RTC
242 * not to be off by more than 15 minutes. Since we're calling it only when
243 * our clock is externally synchronized using NTP, this shouldn't be a problem.
246 real_seconds = nowtime % 60;
247 real_minutes = nowtime / 60;
248 if (((abs(real_minutes - cmos_minutes) + 15) / 30) & 1)
249 real_minutes += 30; /* correct for half hour time zone */
250 real_minutes %= 60;
252 if (abs(real_minutes - cmos_minutes) >= 30) {
253 printk(KERN_WARNING "time.c: can't update CMOS clock "
254 "from %d to %d\n", cmos_minutes, real_minutes);
255 } else {
256 BIN_TO_BCD(real_seconds);
257 BIN_TO_BCD(real_minutes);
258 CMOS_WRITE(real_seconds, RTC_SECONDS);
259 CMOS_WRITE(real_minutes, RTC_MINUTES);
263 * The following flags have to be released exactly in this order, otherwise the
264 * DS12887 (popular MC146818A clone with integrated battery and quartz) will
265 * not reset the oscillator and will not update precisely 500 ms later. You
266 * won't find this mentioned in the Dallas Semiconductor data sheets, but who
267 * believes data sheets anyway ... -- Markus Kuhn
270 CMOS_WRITE(control, RTC_CONTROL);
271 CMOS_WRITE(freq_select, RTC_FREQ_SELECT);
273 spin_unlock(&rtc_lock);
277 /* monotonic_clock(): returns # of nanoseconds passed since time_init()
278 * Note: This function is required to return accurate
279 * time even in the absence of multiple timer ticks.
281 unsigned long long monotonic_clock(void)
283 unsigned long seq;
284 u32 last_offset, this_offset, offset;
285 unsigned long long base;
287 if (vxtime.mode == VXTIME_HPET) {
288 do {
289 seq = read_seqbegin(&xtime_lock);
291 last_offset = vxtime.last;
292 base = monotonic_base;
293 this_offset = hpet_readl(HPET_COUNTER);
294 } while (read_seqretry(&xtime_lock, seq));
295 offset = (this_offset - last_offset);
296 offset *= NSEC_PER_TICK / hpet_tick;
297 } else {
298 do {
299 seq = read_seqbegin(&xtime_lock);
301 last_offset = vxtime.last_tsc;
302 base = monotonic_base;
303 } while (read_seqretry(&xtime_lock, seq));
304 this_offset = get_cycles_sync();
305 /* FIXME: 1000 or 1000000? */
306 offset = (this_offset - last_offset)*1000 / cpu_khz;
308 return base + offset;
310 EXPORT_SYMBOL(monotonic_clock);
312 static noinline void handle_lost_ticks(int lost, struct pt_regs *regs)
314 static long lost_count;
315 static int warned;
316 if (report_lost_ticks) {
317 printk(KERN_WARNING "time.c: Lost %d timer tick(s)! ", lost);
318 print_symbol("rip %s)\n", regs->rip);
321 if (lost_count == 1000 && !warned) {
322 printk(KERN_WARNING "warning: many lost ticks.\n"
323 KERN_WARNING "Your time source seems to be instable or "
324 "some driver is hogging interupts\n");
325 print_symbol("rip %s\n", regs->rip);
326 if (vxtime.mode == VXTIME_TSC && vxtime.hpet_address) {
327 printk(KERN_WARNING "Falling back to HPET\n");
328 if (hpet_use_timer)
329 vxtime.last = hpet_readl(HPET_T0_CMP) -
330 hpet_tick;
331 else
332 vxtime.last = hpet_readl(HPET_COUNTER);
333 vxtime.mode = VXTIME_HPET;
334 do_gettimeoffset = do_gettimeoffset_hpet;
336 /* else should fall back to PIT, but code missing. */
337 warned = 1;
338 } else
339 lost_count++;
341 #ifdef CONFIG_CPU_FREQ
342 /* In some cases the CPU can change frequency without us noticing
343 Give cpufreq a change to catch up. */
344 if ((lost_count+1) % 25 == 0)
345 cpufreq_delayed_get();
346 #endif
349 void main_timer_handler(struct pt_regs *regs)
351 static unsigned long rtc_update = 0;
352 unsigned long tsc;
353 int delay = 0, offset = 0, lost = 0;
356 * Here we are in the timer irq handler. We have irqs locally disabled (so we
357 * don't need spin_lock_irqsave()) but we don't know if the timer_bh is running
358 * on the other CPU, so we need a lock. We also need to lock the vsyscall
359 * variables, because both do_timer() and us change them -arca+vojtech
362 write_seqlock(&xtime_lock);
364 if (vxtime.hpet_address)
365 offset = hpet_readl(HPET_COUNTER);
367 if (hpet_use_timer) {
368 /* if we're using the hpet timer functionality,
369 * we can more accurately know the counter value
370 * when the timer interrupt occured.
372 offset = hpet_readl(HPET_T0_CMP) - hpet_tick;
373 delay = hpet_readl(HPET_COUNTER) - offset;
374 } else if (!pmtmr_ioport) {
375 spin_lock(&i8253_lock);
376 outb_p(0x00, 0x43);
377 delay = inb_p(0x40);
378 delay |= inb(0x40) << 8;
379 spin_unlock(&i8253_lock);
380 delay = LATCH - 1 - delay;
383 tsc = get_cycles_sync();
385 if (vxtime.mode == VXTIME_HPET) {
386 if (offset - vxtime.last > hpet_tick) {
387 lost = (offset - vxtime.last) / hpet_tick - 1;
390 monotonic_base +=
391 (offset - vxtime.last) * NSEC_PER_TICK / hpet_tick;
393 vxtime.last = offset;
394 #ifdef CONFIG_X86_PM_TIMER
395 } else if (vxtime.mode == VXTIME_PMTMR) {
396 lost = pmtimer_mark_offset();
397 #endif
398 } else {
399 offset = (((tsc - vxtime.last_tsc) *
400 vxtime.tsc_quot) >> US_SCALE) - USEC_PER_TICK;
402 if (offset < 0)
403 offset = 0;
405 if (offset > USEC_PER_TICK) {
406 lost = offset / USEC_PER_TICK;
407 offset %= USEC_PER_TICK;
410 /* FIXME: 1000 or 1000000? */
411 monotonic_base += (tsc - vxtime.last_tsc) * 1000000 / cpu_khz;
413 vxtime.last_tsc = tsc - vxtime.quot * delay / vxtime.tsc_quot;
415 if ((((tsc - vxtime.last_tsc) *
416 vxtime.tsc_quot) >> US_SCALE) < offset)
417 vxtime.last_tsc = tsc -
418 (((long) offset << US_SCALE) / vxtime.tsc_quot) - 1;
421 if (lost > 0) {
422 handle_lost_ticks(lost, regs);
423 jiffies += lost;
427 * Do the timer stuff.
430 do_timer(regs);
431 #ifndef CONFIG_SMP
432 update_process_times(user_mode(regs));
433 #endif
436 * In the SMP case we use the local APIC timer interrupt to do the profiling,
437 * except when we simulate SMP mode on a uniprocessor system, in that case we
438 * have to call the local interrupt handler.
441 #ifndef CONFIG_X86_LOCAL_APIC
442 profile_tick(CPU_PROFILING, regs);
443 #else
444 if (!using_apic_timer)
445 smp_local_timer_interrupt(regs);
446 #endif
449 * If we have an externally synchronized Linux clock, then update CMOS clock
450 * accordingly every ~11 minutes. set_rtc_mmss() will be called in the jiffy
451 * closest to exactly 500 ms before the next second. If the update fails, we
452 * don't care, as it'll be updated on the next turn, and the problem (time way
453 * off) isn't likely to go away much sooner anyway.
456 if (ntp_synced() && xtime.tv_sec > rtc_update &&
457 abs(xtime.tv_nsec - 500000000) <= tick_nsec / 2) {
458 set_rtc_mmss(xtime.tv_sec);
459 rtc_update = xtime.tv_sec + 660;
462 write_sequnlock(&xtime_lock);
465 static irqreturn_t timer_interrupt(int irq, void *dev_id, struct pt_regs *regs)
467 if (apic_runs_main_timer > 1)
468 return IRQ_HANDLED;
469 main_timer_handler(regs);
470 #ifdef CONFIG_X86_LOCAL_APIC
471 if (using_apic_timer)
472 smp_send_timer_broadcast_ipi();
473 #endif
474 return IRQ_HANDLED;
477 static unsigned int cyc2ns_scale __read_mostly;
479 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
481 cyc2ns_scale = (NSEC_PER_MSEC << NS_SCALE) / cpu_khz;
484 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
486 return (cyc * cyc2ns_scale) >> NS_SCALE;
489 unsigned long long sched_clock(void)
491 unsigned long a = 0;
493 #if 0
494 /* Don't do a HPET read here. Using TSC always is much faster
495 and HPET may not be mapped yet when the scheduler first runs.
496 Disadvantage is a small drift between CPUs in some configurations,
497 but that should be tolerable. */
498 if (__vxtime.mode == VXTIME_HPET)
499 return (hpet_readl(HPET_COUNTER) * vxtime.quot) >> US_SCALE;
500 #endif
502 /* Could do CPU core sync here. Opteron can execute rdtsc speculatively,
503 which means it is not completely exact and may not be monotonous between
504 CPUs. But the errors should be too small to matter for scheduling
505 purposes. */
507 rdtscll(a);
508 return cycles_2_ns(a);
511 static unsigned long get_cmos_time(void)
513 unsigned int year, mon, day, hour, min, sec;
514 unsigned long flags;
515 unsigned extyear = 0;
517 spin_lock_irqsave(&rtc_lock, flags);
519 do {
520 sec = CMOS_READ(RTC_SECONDS);
521 min = CMOS_READ(RTC_MINUTES);
522 hour = CMOS_READ(RTC_HOURS);
523 day = CMOS_READ(RTC_DAY_OF_MONTH);
524 mon = CMOS_READ(RTC_MONTH);
525 year = CMOS_READ(RTC_YEAR);
526 #ifdef CONFIG_ACPI
527 if (acpi_fadt.revision >= FADT2_REVISION_ID &&
528 acpi_fadt.century)
529 extyear = CMOS_READ(acpi_fadt.century);
530 #endif
531 } while (sec != CMOS_READ(RTC_SECONDS));
533 spin_unlock_irqrestore(&rtc_lock, flags);
536 * We know that x86-64 always uses BCD format, no need to check the
537 * config register.
540 BCD_TO_BIN(sec);
541 BCD_TO_BIN(min);
542 BCD_TO_BIN(hour);
543 BCD_TO_BIN(day);
544 BCD_TO_BIN(mon);
545 BCD_TO_BIN(year);
547 if (extyear) {
548 BCD_TO_BIN(extyear);
549 year += extyear;
550 printk(KERN_INFO "Extended CMOS year: %d\n", extyear);
551 } else {
553 * x86-64 systems only exists since 2002.
554 * This will work up to Dec 31, 2100
556 year += 2000;
559 return mktime(year, mon, day, hour, min, sec);
562 #ifdef CONFIG_CPU_FREQ
564 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
565 changes.
567 RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
568 not that important because current Opteron setups do not support
569 scaling on SMP anyroads.
571 Should fix up last_tsc too. Currently gettimeofday in the
572 first tick after the change will be slightly wrong. */
574 #include <linux/workqueue.h>
576 static unsigned int cpufreq_delayed_issched = 0;
577 static unsigned int cpufreq_init = 0;
578 static struct work_struct cpufreq_delayed_get_work;
580 static void handle_cpufreq_delayed_get(void *v)
582 unsigned int cpu;
583 for_each_online_cpu(cpu) {
584 cpufreq_get(cpu);
586 cpufreq_delayed_issched = 0;
589 /* if we notice lost ticks, schedule a call to cpufreq_get() as it tries
590 * to verify the CPU frequency the timing core thinks the CPU is running
591 * at is still correct.
593 static void cpufreq_delayed_get(void)
595 static int warned;
596 if (cpufreq_init && !cpufreq_delayed_issched) {
597 cpufreq_delayed_issched = 1;
598 if (!warned) {
599 warned = 1;
600 printk(KERN_DEBUG
601 "Losing some ticks... checking if CPU frequency changed.\n");
603 schedule_work(&cpufreq_delayed_get_work);
607 static unsigned int ref_freq = 0;
608 static unsigned long loops_per_jiffy_ref = 0;
610 static unsigned long cpu_khz_ref = 0;
612 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
613 void *data)
615 struct cpufreq_freqs *freq = data;
616 unsigned long *lpj, dummy;
618 if (cpu_has(&cpu_data[freq->cpu], X86_FEATURE_CONSTANT_TSC))
619 return 0;
621 lpj = &dummy;
622 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
623 #ifdef CONFIG_SMP
624 lpj = &cpu_data[freq->cpu].loops_per_jiffy;
625 #else
626 lpj = &boot_cpu_data.loops_per_jiffy;
627 #endif
629 if (!ref_freq) {
630 ref_freq = freq->old;
631 loops_per_jiffy_ref = *lpj;
632 cpu_khz_ref = cpu_khz;
634 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
635 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
636 (val == CPUFREQ_RESUMECHANGE)) {
637 *lpj =
638 cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
640 cpu_khz = cpufreq_scale(cpu_khz_ref, ref_freq, freq->new);
641 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
642 vxtime.tsc_quot = (USEC_PER_MSEC << US_SCALE) / cpu_khz;
645 set_cyc2ns_scale(cpu_khz_ref);
647 return 0;
650 static struct notifier_block time_cpufreq_notifier_block = {
651 .notifier_call = time_cpufreq_notifier
654 static int __init cpufreq_tsc(void)
656 INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get, NULL);
657 if (!cpufreq_register_notifier(&time_cpufreq_notifier_block,
658 CPUFREQ_TRANSITION_NOTIFIER))
659 cpufreq_init = 1;
660 return 0;
663 core_initcall(cpufreq_tsc);
665 #endif
668 * calibrate_tsc() calibrates the processor TSC in a very simple way, comparing
669 * it to the HPET timer of known frequency.
672 #define TICK_COUNT 100000000
674 static unsigned int __init hpet_calibrate_tsc(void)
676 int tsc_start, hpet_start;
677 int tsc_now, hpet_now;
678 unsigned long flags;
680 local_irq_save(flags);
681 local_irq_disable();
683 hpet_start = hpet_readl(HPET_COUNTER);
684 rdtscl(tsc_start);
686 do {
687 local_irq_disable();
688 hpet_now = hpet_readl(HPET_COUNTER);
689 tsc_now = get_cycles_sync();
690 local_irq_restore(flags);
691 } while ((tsc_now - tsc_start) < TICK_COUNT &&
692 (hpet_now - hpet_start) < TICK_COUNT);
694 return (tsc_now - tsc_start) * 1000000000L
695 / ((hpet_now - hpet_start) * hpet_period / 1000);
700 * pit_calibrate_tsc() uses the speaker output (channel 2) of
701 * the PIT. This is better than using the timer interrupt output,
702 * because we can read the value of the speaker with just one inb(),
703 * where we need three i/o operations for the interrupt channel.
704 * We count how many ticks the TSC does in 50 ms.
707 static unsigned int __init pit_calibrate_tsc(void)
709 unsigned long start, end;
710 unsigned long flags;
712 spin_lock_irqsave(&i8253_lock, flags);
714 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
716 outb(0xb0, 0x43);
717 outb((PIT_TICK_RATE / (1000 / 50)) & 0xff, 0x42);
718 outb((PIT_TICK_RATE / (1000 / 50)) >> 8, 0x42);
719 start = get_cycles_sync();
720 while ((inb(0x61) & 0x20) == 0);
721 end = get_cycles_sync();
723 spin_unlock_irqrestore(&i8253_lock, flags);
725 return (end - start) / 50;
728 #ifdef CONFIG_HPET
729 static __init int late_hpet_init(void)
731 struct hpet_data hd;
732 unsigned int ntimer;
734 if (!vxtime.hpet_address)
735 return 0;
737 memset(&hd, 0, sizeof (hd));
739 ntimer = hpet_readl(HPET_ID);
740 ntimer = (ntimer & HPET_ID_NUMBER) >> HPET_ID_NUMBER_SHIFT;
741 ntimer++;
744 * Register with driver.
745 * Timer0 and Timer1 is used by platform.
747 hd.hd_phys_address = vxtime.hpet_address;
748 hd.hd_address = (void __iomem *)fix_to_virt(FIX_HPET_BASE);
749 hd.hd_nirqs = ntimer;
750 hd.hd_flags = HPET_DATA_PLATFORM;
751 hpet_reserve_timer(&hd, 0);
752 #ifdef CONFIG_HPET_EMULATE_RTC
753 hpet_reserve_timer(&hd, 1);
754 #endif
755 hd.hd_irq[0] = HPET_LEGACY_8254;
756 hd.hd_irq[1] = HPET_LEGACY_RTC;
757 if (ntimer > 2) {
758 struct hpet *hpet;
759 struct hpet_timer *timer;
760 int i;
762 hpet = (struct hpet *) fix_to_virt(FIX_HPET_BASE);
763 timer = &hpet->hpet_timers[2];
764 for (i = 2; i < ntimer; timer++, i++)
765 hd.hd_irq[i] = (timer->hpet_config &
766 Tn_INT_ROUTE_CNF_MASK) >>
767 Tn_INT_ROUTE_CNF_SHIFT;
771 hpet_alloc(&hd);
772 return 0;
774 fs_initcall(late_hpet_init);
775 #endif
777 static int hpet_timer_stop_set_go(unsigned long tick)
779 unsigned int cfg;
782 * Stop the timers and reset the main counter.
785 cfg = hpet_readl(HPET_CFG);
786 cfg &= ~(HPET_CFG_ENABLE | HPET_CFG_LEGACY);
787 hpet_writel(cfg, HPET_CFG);
788 hpet_writel(0, HPET_COUNTER);
789 hpet_writel(0, HPET_COUNTER + 4);
792 * Set up timer 0, as periodic with first interrupt to happen at hpet_tick,
793 * and period also hpet_tick.
795 if (hpet_use_timer) {
796 hpet_writel(HPET_TN_ENABLE | HPET_TN_PERIODIC | HPET_TN_SETVAL |
797 HPET_TN_32BIT, HPET_T0_CFG);
798 hpet_writel(hpet_tick, HPET_T0_CMP); /* next interrupt */
799 hpet_writel(hpet_tick, HPET_T0_CMP); /* period */
800 cfg |= HPET_CFG_LEGACY;
803 * Go!
806 cfg |= HPET_CFG_ENABLE;
807 hpet_writel(cfg, HPET_CFG);
809 return 0;
812 static int hpet_init(void)
814 unsigned int id;
816 if (!vxtime.hpet_address)
817 return -1;
818 set_fixmap_nocache(FIX_HPET_BASE, vxtime.hpet_address);
819 __set_fixmap(VSYSCALL_HPET, vxtime.hpet_address, PAGE_KERNEL_VSYSCALL_NOCACHE);
822 * Read the period, compute tick and quotient.
825 id = hpet_readl(HPET_ID);
827 if (!(id & HPET_ID_VENDOR) || !(id & HPET_ID_NUMBER))
828 return -1;
830 hpet_period = hpet_readl(HPET_PERIOD);
831 if (hpet_period < 100000 || hpet_period > 100000000)
832 return -1;
834 hpet_tick = (FSEC_PER_TICK + hpet_period / 2) / hpet_period;
836 hpet_use_timer = (id & HPET_ID_LEGSUP);
838 return hpet_timer_stop_set_go(hpet_tick);
841 static int hpet_reenable(void)
843 return hpet_timer_stop_set_go(hpet_tick);
846 #define PIT_MODE 0x43
847 #define PIT_CH0 0x40
849 static void __init __pit_init(int val, u8 mode)
851 unsigned long flags;
853 spin_lock_irqsave(&i8253_lock, flags);
854 outb_p(mode, PIT_MODE);
855 outb_p(val & 0xff, PIT_CH0); /* LSB */
856 outb_p(val >> 8, PIT_CH0); /* MSB */
857 spin_unlock_irqrestore(&i8253_lock, flags);
860 void __init pit_init(void)
862 __pit_init(LATCH, 0x34); /* binary, mode 2, LSB/MSB, ch 0 */
865 void __init pit_stop_interrupt(void)
867 __pit_init(0, 0x30); /* mode 0 */
870 void __init stop_timer_interrupt(void)
872 char *name;
873 if (vxtime.hpet_address) {
874 name = "HPET";
875 hpet_timer_stop_set_go(0);
876 } else {
877 name = "PIT";
878 pit_stop_interrupt();
880 printk(KERN_INFO "timer: %s interrupt stopped.\n", name);
883 int __init time_setup(char *str)
885 report_lost_ticks = 1;
886 return 1;
889 static struct irqaction irq0 = {
890 timer_interrupt, SA_INTERRUPT, CPU_MASK_NONE, "timer", NULL, NULL
893 void __init time_init(void)
895 char *timename;
896 char *gtod;
898 if (nohpet)
899 vxtime.hpet_address = 0;
901 xtime.tv_sec = get_cmos_time();
902 xtime.tv_nsec = 0;
904 set_normalized_timespec(&wall_to_monotonic,
905 -xtime.tv_sec, -xtime.tv_nsec);
907 if (!hpet_init())
908 vxtime_hz = (FSEC_PER_SEC + hpet_period / 2) / hpet_period;
909 else
910 vxtime.hpet_address = 0;
912 if (hpet_use_timer) {
913 /* set tick_nsec to use the proper rate for HPET */
914 tick_nsec = TICK_NSEC_HPET;
915 cpu_khz = hpet_calibrate_tsc();
916 timename = "HPET";
917 #ifdef CONFIG_X86_PM_TIMER
918 } else if (pmtmr_ioport && !vxtime.hpet_address) {
919 vxtime_hz = PM_TIMER_FREQUENCY;
920 timename = "PM";
921 pit_init();
922 cpu_khz = pit_calibrate_tsc();
923 #endif
924 } else {
925 pit_init();
926 cpu_khz = pit_calibrate_tsc();
927 timename = "PIT";
930 vxtime.mode = VXTIME_TSC;
931 gtod = time_init_gtod();
933 printk(KERN_INFO "time.c: Using %ld.%06ld MHz WALL %s GTOD %s timer.\n",
934 vxtime_hz / 1000000, vxtime_hz % 1000000, timename, gtod);
935 printk(KERN_INFO "time.c: Detected %d.%03d MHz processor.\n",
936 cpu_khz / 1000, cpu_khz % 1000);
937 vxtime.quot = (USEC_PER_SEC << US_SCALE) / vxtime_hz;
938 vxtime.tsc_quot = (USEC_PER_MSEC << US_SCALE) / cpu_khz;
939 vxtime.last_tsc = get_cycles_sync();
940 setup_irq(0, &irq0);
942 set_cyc2ns_scale(cpu_khz);
946 * Make an educated guess if the TSC is trustworthy and synchronized
947 * over all CPUs.
949 __cpuinit int unsynchronized_tsc(void)
951 #ifdef CONFIG_SMP
952 if (apic_is_clustered_box())
953 return 1;
954 /* Intel systems are normally all synchronized. Exceptions
955 are handled in the check above. */
956 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
957 return 0;
958 #endif
959 /* Assume multi socket systems are not synchronized */
960 return num_present_cpus() > 1;
964 * Decide what mode gettimeofday should use.
966 __init static char *time_init_gtod(void)
968 char *timetype;
970 if (unsynchronized_tsc())
971 notsc = 1;
972 if (vxtime.hpet_address && notsc) {
973 timetype = hpet_use_timer ? "HPET" : "PIT/HPET";
974 if (hpet_use_timer)
975 vxtime.last = hpet_readl(HPET_T0_CMP) - hpet_tick;
976 else
977 vxtime.last = hpet_readl(HPET_COUNTER);
978 vxtime.mode = VXTIME_HPET;
979 do_gettimeoffset = do_gettimeoffset_hpet;
980 #ifdef CONFIG_X86_PM_TIMER
981 /* Using PM for gettimeofday is quite slow, but we have no other
982 choice because the TSC is too unreliable on some systems. */
983 } else if (pmtmr_ioport && !vxtime.hpet_address && notsc) {
984 timetype = "PM";
985 do_gettimeoffset = do_gettimeoffset_pm;
986 vxtime.mode = VXTIME_PMTMR;
987 sysctl_vsyscall = 0;
988 printk(KERN_INFO "Disabling vsyscall due to use of PM timer\n");
989 #endif
990 } else {
991 timetype = hpet_use_timer ? "HPET/TSC" : "PIT/TSC";
992 vxtime.mode = VXTIME_TSC;
994 return timetype;
997 __setup("report_lost_ticks", time_setup);
999 static long clock_cmos_diff;
1000 static unsigned long sleep_start;
1003 * sysfs support for the timer.
1006 static int timer_suspend(struct sys_device *dev, pm_message_t state)
1009 * Estimate time zone so that set_time can update the clock
1011 long cmos_time = get_cmos_time();
1013 clock_cmos_diff = -cmos_time;
1014 clock_cmos_diff += get_seconds();
1015 sleep_start = cmos_time;
1016 return 0;
1019 static int timer_resume(struct sys_device *dev)
1021 unsigned long flags;
1022 unsigned long sec;
1023 unsigned long ctime = get_cmos_time();
1024 unsigned long sleep_length = (ctime - sleep_start) * HZ;
1026 if (vxtime.hpet_address)
1027 hpet_reenable();
1028 else
1029 i8254_timer_resume();
1031 sec = ctime + clock_cmos_diff;
1032 write_seqlock_irqsave(&xtime_lock,flags);
1033 xtime.tv_sec = sec;
1034 xtime.tv_nsec = 0;
1035 if (vxtime.mode == VXTIME_HPET) {
1036 if (hpet_use_timer)
1037 vxtime.last = hpet_readl(HPET_T0_CMP) - hpet_tick;
1038 else
1039 vxtime.last = hpet_readl(HPET_COUNTER);
1040 #ifdef CONFIG_X86_PM_TIMER
1041 } else if (vxtime.mode == VXTIME_PMTMR) {
1042 pmtimer_resume();
1043 #endif
1044 } else
1045 vxtime.last_tsc = get_cycles_sync();
1046 write_sequnlock_irqrestore(&xtime_lock,flags);
1047 jiffies += sleep_length;
1048 wall_jiffies += sleep_length;
1049 monotonic_base += sleep_length * (NSEC_PER_SEC/HZ);
1050 touch_softlockup_watchdog();
1051 return 0;
1054 static struct sysdev_class timer_sysclass = {
1055 .resume = timer_resume,
1056 .suspend = timer_suspend,
1057 set_kset_name("timer"),
1060 /* XXX this driverfs stuff should probably go elsewhere later -john */
1061 static struct sys_device device_timer = {
1062 .id = 0,
1063 .cls = &timer_sysclass,
1066 static int time_init_device(void)
1068 int error = sysdev_class_register(&timer_sysclass);
1069 if (!error)
1070 error = sysdev_register(&device_timer);
1071 return error;
1074 device_initcall(time_init_device);
1076 #ifdef CONFIG_HPET_EMULATE_RTC
1077 /* HPET in LegacyReplacement Mode eats up RTC interrupt line. When, HPET
1078 * is enabled, we support RTC interrupt functionality in software.
1079 * RTC has 3 kinds of interrupts:
1080 * 1) Update Interrupt - generate an interrupt, every sec, when RTC clock
1081 * is updated
1082 * 2) Alarm Interrupt - generate an interrupt at a specific time of day
1083 * 3) Periodic Interrupt - generate periodic interrupt, with frequencies
1084 * 2Hz-8192Hz (2Hz-64Hz for non-root user) (all freqs in powers of 2)
1085 * (1) and (2) above are implemented using polling at a frequency of
1086 * 64 Hz. The exact frequency is a tradeoff between accuracy and interrupt
1087 * overhead. (DEFAULT_RTC_INT_FREQ)
1088 * For (3), we use interrupts at 64Hz or user specified periodic
1089 * frequency, whichever is higher.
1091 #include <linux/rtc.h>
1093 #define DEFAULT_RTC_INT_FREQ 64
1094 #define RTC_NUM_INTS 1
1096 static unsigned long UIE_on;
1097 static unsigned long prev_update_sec;
1099 static unsigned long AIE_on;
1100 static struct rtc_time alarm_time;
1102 static unsigned long PIE_on;
1103 static unsigned long PIE_freq = DEFAULT_RTC_INT_FREQ;
1104 static unsigned long PIE_count;
1106 static unsigned long hpet_rtc_int_freq; /* RTC interrupt frequency */
1107 static unsigned int hpet_t1_cmp; /* cached comparator register */
1109 int is_hpet_enabled(void)
1111 return vxtime.hpet_address != 0;
1115 * Timer 1 for RTC, we do not use periodic interrupt feature,
1116 * even if HPET supports periodic interrupts on Timer 1.
1117 * The reason being, to set up a periodic interrupt in HPET, we need to
1118 * stop the main counter. And if we do that everytime someone diables/enables
1119 * RTC, we will have adverse effect on main kernel timer running on Timer 0.
1120 * So, for the time being, simulate the periodic interrupt in software.
1122 * hpet_rtc_timer_init() is called for the first time and during subsequent
1123 * interuppts reinit happens through hpet_rtc_timer_reinit().
1125 int hpet_rtc_timer_init(void)
1127 unsigned int cfg, cnt;
1128 unsigned long flags;
1130 if (!is_hpet_enabled())
1131 return 0;
1133 * Set the counter 1 and enable the interrupts.
1135 if (PIE_on && (PIE_freq > DEFAULT_RTC_INT_FREQ))
1136 hpet_rtc_int_freq = PIE_freq;
1137 else
1138 hpet_rtc_int_freq = DEFAULT_RTC_INT_FREQ;
1140 local_irq_save(flags);
1141 cnt = hpet_readl(HPET_COUNTER);
1142 cnt += ((hpet_tick*HZ)/hpet_rtc_int_freq);
1143 hpet_writel(cnt, HPET_T1_CMP);
1144 hpet_t1_cmp = cnt;
1145 local_irq_restore(flags);
1147 cfg = hpet_readl(HPET_T1_CFG);
1148 cfg &= ~HPET_TN_PERIODIC;
1149 cfg |= HPET_TN_ENABLE | HPET_TN_32BIT;
1150 hpet_writel(cfg, HPET_T1_CFG);
1152 return 1;
1155 static void hpet_rtc_timer_reinit(void)
1157 unsigned int cfg, cnt;
1159 if (unlikely(!(PIE_on | AIE_on | UIE_on))) {
1160 cfg = hpet_readl(HPET_T1_CFG);
1161 cfg &= ~HPET_TN_ENABLE;
1162 hpet_writel(cfg, HPET_T1_CFG);
1163 return;
1166 if (PIE_on && (PIE_freq > DEFAULT_RTC_INT_FREQ))
1167 hpet_rtc_int_freq = PIE_freq;
1168 else
1169 hpet_rtc_int_freq = DEFAULT_RTC_INT_FREQ;
1171 /* It is more accurate to use the comparator value than current count.*/
1172 cnt = hpet_t1_cmp;
1173 cnt += hpet_tick*HZ/hpet_rtc_int_freq;
1174 hpet_writel(cnt, HPET_T1_CMP);
1175 hpet_t1_cmp = cnt;
1179 * The functions below are called from rtc driver.
1180 * Return 0 if HPET is not being used.
1181 * Otherwise do the necessary changes and return 1.
1183 int hpet_mask_rtc_irq_bit(unsigned long bit_mask)
1185 if (!is_hpet_enabled())
1186 return 0;
1188 if (bit_mask & RTC_UIE)
1189 UIE_on = 0;
1190 if (bit_mask & RTC_PIE)
1191 PIE_on = 0;
1192 if (bit_mask & RTC_AIE)
1193 AIE_on = 0;
1195 return 1;
1198 int hpet_set_rtc_irq_bit(unsigned long bit_mask)
1200 int timer_init_reqd = 0;
1202 if (!is_hpet_enabled())
1203 return 0;
1205 if (!(PIE_on | AIE_on | UIE_on))
1206 timer_init_reqd = 1;
1208 if (bit_mask & RTC_UIE) {
1209 UIE_on = 1;
1211 if (bit_mask & RTC_PIE) {
1212 PIE_on = 1;
1213 PIE_count = 0;
1215 if (bit_mask & RTC_AIE) {
1216 AIE_on = 1;
1219 if (timer_init_reqd)
1220 hpet_rtc_timer_init();
1222 return 1;
1225 int hpet_set_alarm_time(unsigned char hrs, unsigned char min, unsigned char sec)
1227 if (!is_hpet_enabled())
1228 return 0;
1230 alarm_time.tm_hour = hrs;
1231 alarm_time.tm_min = min;
1232 alarm_time.tm_sec = sec;
1234 return 1;
1237 int hpet_set_periodic_freq(unsigned long freq)
1239 if (!is_hpet_enabled())
1240 return 0;
1242 PIE_freq = freq;
1243 PIE_count = 0;
1245 return 1;
1248 int hpet_rtc_dropped_irq(void)
1250 if (!is_hpet_enabled())
1251 return 0;
1253 return 1;
1256 irqreturn_t hpet_rtc_interrupt(int irq, void *dev_id, struct pt_regs *regs)
1258 struct rtc_time curr_time;
1259 unsigned long rtc_int_flag = 0;
1260 int call_rtc_interrupt = 0;
1262 hpet_rtc_timer_reinit();
1264 if (UIE_on | AIE_on) {
1265 rtc_get_rtc_time(&curr_time);
1267 if (UIE_on) {
1268 if (curr_time.tm_sec != prev_update_sec) {
1269 /* Set update int info, call real rtc int routine */
1270 call_rtc_interrupt = 1;
1271 rtc_int_flag = RTC_UF;
1272 prev_update_sec = curr_time.tm_sec;
1275 if (PIE_on) {
1276 PIE_count++;
1277 if (PIE_count >= hpet_rtc_int_freq/PIE_freq) {
1278 /* Set periodic int info, call real rtc int routine */
1279 call_rtc_interrupt = 1;
1280 rtc_int_flag |= RTC_PF;
1281 PIE_count = 0;
1284 if (AIE_on) {
1285 if ((curr_time.tm_sec == alarm_time.tm_sec) &&
1286 (curr_time.tm_min == alarm_time.tm_min) &&
1287 (curr_time.tm_hour == alarm_time.tm_hour)) {
1288 /* Set alarm int info, call real rtc int routine */
1289 call_rtc_interrupt = 1;
1290 rtc_int_flag |= RTC_AF;
1293 if (call_rtc_interrupt) {
1294 rtc_int_flag |= (RTC_IRQF | (RTC_NUM_INTS << 8));
1295 rtc_interrupt(rtc_int_flag, dev_id, regs);
1297 return IRQ_HANDLED;
1299 #endif
1301 static int __init nohpet_setup(char *s)
1303 nohpet = 1;
1304 return 1;
1307 __setup("nohpet", nohpet_setup);
1309 int __init notsc_setup(char *s)
1311 notsc = 1;
1312 return 1;
1315 __setup("notsc", notsc_setup);