47d25ad08160bc7646d481996b414203ddc0c315
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / arch / x86_64 / kernel / time.c
1 /*
2 * linux/arch/x86-64/kernel/time.c
3 *
4 * "High Precision Event Timer" based timekeeping.
5 *
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
14 */
15
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
44
45 u64 jiffies_64 = INITIAL_JIFFIES;
46
47 EXPORT_SYMBOL(jiffies_64);
48
49 #ifdef CONFIG_CPU_FREQ
50 static void cpufreq_delayed_get(void);
51 #endif
52 extern void i8254_timer_resume(void);
53 extern int using_apic_timer;
54
55 DEFINE_SPINLOCK(rtc_lock);
56 DEFINE_SPINLOCK(i8253_lock);
57
58 static int nohpet __initdata = 0;
59 static int notsc __initdata = 0;
60
61 #undef HPET_HACK_ENABLE_DANGEROUS
62
63 unsigned int cpu_khz; /* TSC clocks / usec, not used here */
64 static unsigned long hpet_period; /* fsecs / HPET clock */
65 unsigned long hpet_tick; /* HPET clocks / interrupt */
66 static int hpet_use_timer;
67 unsigned long vxtime_hz = PIT_TICK_RATE;
68 int report_lost_ticks; /* command line option */
69 unsigned long long monotonic_base;
70
71 struct vxtime_data __vxtime __section_vxtime; /* for vsyscalls */
72
73 volatile unsigned long __jiffies __section_jiffies = INITIAL_JIFFIES;
74 unsigned long __wall_jiffies __section_wall_jiffies = INITIAL_JIFFIES;
75 struct timespec __xtime __section_xtime;
76 struct timezone __sys_tz __section_sys_tz;
77
78 static inline void rdtscll_sync(unsigned long *tsc)
79 {
80 #ifdef CONFIG_SMP
81 sync_core();
82 #endif
83 rdtscll(*tsc);
84 }
85
86 /*
87 * do_gettimeoffset() returns microseconds since last timer interrupt was
88 * triggered by hardware. A memory read of HPET is slower than a register read
89 * of TSC, but much more reliable. It's also synchronized to the timer
90 * interrupt. Note that do_gettimeoffset() may return more than hpet_tick, if a
91 * timer interrupt has happened already, but vxtime.trigger wasn't updated yet.
92 * This is not a problem, because jiffies hasn't updated either. They are bound
93 * together by xtime_lock.
94 */
95
96 static inline unsigned int do_gettimeoffset_tsc(void)
97 {
98 unsigned long t;
99 unsigned long x;
100 rdtscll_sync(&t);
101 if (t < vxtime.last_tsc) t = vxtime.last_tsc; /* hack */
102 x = ((t - vxtime.last_tsc) * vxtime.tsc_quot) >> 32;
103 return x;
104 }
105
106 static inline unsigned int do_gettimeoffset_hpet(void)
107 {
108 /* cap counter read to one tick to avoid inconsistencies */
109 unsigned long counter = hpet_readl(HPET_COUNTER) - vxtime.last;
110 return (min(counter,hpet_tick) * vxtime.quot) >> 32;
111 }
112
113 unsigned int (*do_gettimeoffset)(void) = do_gettimeoffset_tsc;
114
115 /*
116 * This version of gettimeofday() has microsecond resolution and better than
117 * microsecond precision, as we're using at least a 10 MHz (usually 14.31818
118 * MHz) HPET timer.
119 */
120
121 void do_gettimeofday(struct timeval *tv)
122 {
123 unsigned long seq, t;
124 unsigned int sec, usec;
125
126 do {
127 seq = read_seqbegin(&xtime_lock);
128
129 sec = xtime.tv_sec;
130 usec = xtime.tv_nsec / 1000;
131
132 /* i386 does some correction here to keep the clock
133 monotonous even when ntpd is fixing drift.
134 But they didn't work for me, there is a non monotonic
135 clock anyways with ntp.
136 I dropped all corrections now until a real solution can
137 be found. Note when you fix it here you need to do the same
138 in arch/x86_64/kernel/vsyscall.c and export all needed
139 variables in vmlinux.lds. -AK */
140
141 t = (jiffies - wall_jiffies) * (1000000L / HZ) +
142 do_gettimeoffset();
143 usec += t;
144
145 } while (read_seqretry(&xtime_lock, seq));
146
147 tv->tv_sec = sec + usec / 1000000;
148 tv->tv_usec = usec % 1000000;
149 }
150
151 EXPORT_SYMBOL(do_gettimeofday);
152
153 /*
154 * settimeofday() first undoes the correction that gettimeofday would do
155 * on the time, and then saves it. This is ugly, but has been like this for
156 * ages already.
157 */
158
159 int do_settimeofday(struct timespec *tv)
160 {
161 time_t wtm_sec, sec = tv->tv_sec;
162 long wtm_nsec, nsec = tv->tv_nsec;
163
164 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
165 return -EINVAL;
166
167 write_seqlock_irq(&xtime_lock);
168
169 nsec -= do_gettimeoffset() * 1000 +
170 (jiffies - wall_jiffies) * (NSEC_PER_SEC/HZ);
171
172 wtm_sec = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
173 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
174
175 set_normalized_timespec(&xtime, sec, nsec);
176 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
177
178 ntp_clear();
179
180 write_sequnlock_irq(&xtime_lock);
181 clock_was_set();
182 return 0;
183 }
184
185 EXPORT_SYMBOL(do_settimeofday);
186
187 unsigned long profile_pc(struct pt_regs *regs)
188 {
189 unsigned long pc = instruction_pointer(regs);
190
191 /* Assume the lock function has either no stack frame or only a single word.
192 This checks if the address on the stack looks like a kernel text address.
193 There is a small window for false hits, but in that case the tick
194 is just accounted to the spinlock function.
195 Better would be to write these functions in assembler again
196 and check exactly. */
197 if (in_lock_functions(pc)) {
198 char *v = *(char **)regs->rsp;
199 if ((v >= _stext && v <= _etext) ||
200 (v >= _sinittext && v <= _einittext) ||
201 (v >= (char *)MODULES_VADDR && v <= (char *)MODULES_END))
202 return (unsigned long)v;
203 return ((unsigned long *)regs->rsp)[1];
204 }
205 return pc;
206 }
207 EXPORT_SYMBOL(profile_pc);
208
209 /*
210 * In order to set the CMOS clock precisely, set_rtc_mmss has to be called 500
211 * ms after the second nowtime has started, because when nowtime is written
212 * into the registers of the CMOS clock, it will jump to the next second
213 * precisely 500 ms later. Check the Motorola MC146818A or Dallas DS12887 data
214 * sheet for details.
215 */
216
217 static void set_rtc_mmss(unsigned long nowtime)
218 {
219 int real_seconds, real_minutes, cmos_minutes;
220 unsigned char control, freq_select;
221
222 /*
223 * IRQs are disabled when we're called from the timer interrupt,
224 * no need for spin_lock_irqsave()
225 */
226
227 spin_lock(&rtc_lock);
228
229 /*
230 * Tell the clock it's being set and stop it.
231 */
232
233 control = CMOS_READ(RTC_CONTROL);
234 CMOS_WRITE(control | RTC_SET, RTC_CONTROL);
235
236 freq_select = CMOS_READ(RTC_FREQ_SELECT);
237 CMOS_WRITE(freq_select | RTC_DIV_RESET2, RTC_FREQ_SELECT);
238
239 cmos_minutes = CMOS_READ(RTC_MINUTES);
240 BCD_TO_BIN(cmos_minutes);
241
242 /*
243 * since we're only adjusting minutes and seconds, don't interfere with hour
244 * overflow. This avoids messing with unknown time zones but requires your RTC
245 * not to be off by more than 15 minutes. Since we're calling it only when
246 * our clock is externally synchronized using NTP, this shouldn't be a problem.
247 */
248
249 real_seconds = nowtime % 60;
250 real_minutes = nowtime / 60;
251 if (((abs(real_minutes - cmos_minutes) + 15) / 30) & 1)
252 real_minutes += 30; /* correct for half hour time zone */
253 real_minutes %= 60;
254
255 #if 0
256 /* AMD 8111 is a really bad time keeper and hits this regularly.
257 It probably was an attempt to avoid screwing up DST, but ignore
258 that for now. */
259 if (abs(real_minutes - cmos_minutes) >= 30) {
260 printk(KERN_WARNING "time.c: can't update CMOS clock "
261 "from %d to %d\n", cmos_minutes, real_minutes);
262 } else
263 #endif
264
265 {
266 BIN_TO_BCD(real_seconds);
267 BIN_TO_BCD(real_minutes);
268 CMOS_WRITE(real_seconds, RTC_SECONDS);
269 CMOS_WRITE(real_minutes, RTC_MINUTES);
270 }
271
272 /*
273 * The following flags have to be released exactly in this order, otherwise the
274 * DS12887 (popular MC146818A clone with integrated battery and quartz) will
275 * not reset the oscillator and will not update precisely 500 ms later. You
276 * won't find this mentioned in the Dallas Semiconductor data sheets, but who
277 * believes data sheets anyway ... -- Markus Kuhn
278 */
279
280 CMOS_WRITE(control, RTC_CONTROL);
281 CMOS_WRITE(freq_select, RTC_FREQ_SELECT);
282
283 spin_unlock(&rtc_lock);
284 }
285
286
287 /* monotonic_clock(): returns # of nanoseconds passed since time_init()
288 * Note: This function is required to return accurate
289 * time even in the absence of multiple timer ticks.
290 */
291 unsigned long long monotonic_clock(void)
292 {
293 unsigned long seq;
294 u32 last_offset, this_offset, offset;
295 unsigned long long base;
296
297 if (vxtime.mode == VXTIME_HPET) {
298 do {
299 seq = read_seqbegin(&xtime_lock);
300
301 last_offset = vxtime.last;
302 base = monotonic_base;
303 this_offset = hpet_readl(HPET_COUNTER);
304
305 } while (read_seqretry(&xtime_lock, seq));
306 offset = (this_offset - last_offset);
307 offset *=(NSEC_PER_SEC/HZ)/hpet_tick;
308 return base + offset;
309 }else{
310 do {
311 seq = read_seqbegin(&xtime_lock);
312
313 last_offset = vxtime.last_tsc;
314 base = monotonic_base;
315 } while (read_seqretry(&xtime_lock, seq));
316 sync_core();
317 rdtscll(this_offset);
318 offset = (this_offset - last_offset)*1000/cpu_khz;
319 return base + offset;
320 }
321
322
323 }
324 EXPORT_SYMBOL(monotonic_clock);
325
326 static noinline void handle_lost_ticks(int lost, struct pt_regs *regs)
327 {
328 static long lost_count;
329 static int warned;
330
331 if (report_lost_ticks) {
332 printk(KERN_WARNING "time.c: Lost %d timer "
333 "tick(s)! ", lost);
334 print_symbol("rip %s)\n", regs->rip);
335 }
336
337 if (lost_count == 1000 && !warned) {
338 printk(KERN_WARNING
339 "warning: many lost ticks.\n"
340 KERN_WARNING "Your time source seems to be instable or "
341 "some driver is hogging interupts\n");
342 print_symbol("rip %s\n", regs->rip);
343 if (vxtime.mode == VXTIME_TSC && vxtime.hpet_address) {
344 printk(KERN_WARNING "Falling back to HPET\n");
345 vxtime.last = hpet_readl(HPET_T0_CMP) - hpet_tick;
346 vxtime.mode = VXTIME_HPET;
347 do_gettimeoffset = do_gettimeoffset_hpet;
348 }
349 /* else should fall back to PIT, but code missing. */
350 warned = 1;
351 } else
352 lost_count++;
353
354 #ifdef CONFIG_CPU_FREQ
355 /* In some cases the CPU can change frequency without us noticing
356 (like going into thermal throttle)
357 Give cpufreq a change to catch up. */
358 if ((lost_count+1) % 25 == 0) {
359 cpufreq_delayed_get();
360 }
361 #endif
362 }
363
364 static irqreturn_t timer_interrupt(int irq, void *dev_id, struct pt_regs *regs)
365 {
366 static unsigned long rtc_update = 0;
367 unsigned long tsc;
368 int delay, offset = 0, lost = 0;
369
370 /*
371 * Here we are in the timer irq handler. We have irqs locally disabled (so we
372 * don't need spin_lock_irqsave()) but we don't know if the timer_bh is running
373 * on the other CPU, so we need a lock. We also need to lock the vsyscall
374 * variables, because both do_timer() and us change them -arca+vojtech
375 */
376
377 write_seqlock(&xtime_lock);
378
379 if (vxtime.hpet_address)
380 offset = hpet_readl(HPET_COUNTER);
381
382 if (hpet_use_timer) {
383 /* if we're using the hpet timer functionality,
384 * we can more accurately know the counter value
385 * when the timer interrupt occured.
386 */
387 offset = hpet_readl(HPET_T0_CMP) - hpet_tick;
388 delay = hpet_readl(HPET_COUNTER) - offset;
389 } else {
390 spin_lock(&i8253_lock);
391 outb_p(0x00, 0x43);
392 delay = inb_p(0x40);
393 delay |= inb(0x40) << 8;
394 spin_unlock(&i8253_lock);
395 delay = LATCH - 1 - delay;
396 }
397
398 rdtscll_sync(&tsc);
399
400 if (vxtime.mode == VXTIME_HPET) {
401 if (offset - vxtime.last > hpet_tick) {
402 lost = (offset - vxtime.last) / hpet_tick - 1;
403 }
404
405 monotonic_base +=
406 (offset - vxtime.last)*(NSEC_PER_SEC/HZ) / hpet_tick;
407
408 vxtime.last = offset;
409 #ifdef CONFIG_X86_PM_TIMER
410 } else if (vxtime.mode == VXTIME_PMTMR) {
411 lost = pmtimer_mark_offset();
412 #endif
413 } else {
414 offset = (((tsc - vxtime.last_tsc) *
415 vxtime.tsc_quot) >> 32) - (USEC_PER_SEC / HZ);
416
417 if (offset < 0)
418 offset = 0;
419
420 if (offset > (USEC_PER_SEC / HZ)) {
421 lost = offset / (USEC_PER_SEC / HZ);
422 offset %= (USEC_PER_SEC / HZ);
423 }
424
425 monotonic_base += (tsc - vxtime.last_tsc)*1000000/cpu_khz ;
426
427 vxtime.last_tsc = tsc - vxtime.quot * delay / vxtime.tsc_quot;
428
429 if ((((tsc - vxtime.last_tsc) *
430 vxtime.tsc_quot) >> 32) < offset)
431 vxtime.last_tsc = tsc -
432 (((long) offset << 32) / vxtime.tsc_quot) - 1;
433 }
434
435 if (lost > 0) {
436 handle_lost_ticks(lost, regs);
437 jiffies += lost;
438 }
439
440 /*
441 * Do the timer stuff.
442 */
443
444 do_timer(regs);
445 #ifndef CONFIG_SMP
446 update_process_times(user_mode(regs));
447 #endif
448
449 /*
450 * In the SMP case we use the local APIC timer interrupt to do the profiling,
451 * except when we simulate SMP mode on a uniprocessor system, in that case we
452 * have to call the local interrupt handler.
453 */
454
455 #ifndef CONFIG_X86_LOCAL_APIC
456 profile_tick(CPU_PROFILING, regs);
457 #else
458 if (!using_apic_timer)
459 smp_local_timer_interrupt(regs);
460 #endif
461
462 /*
463 * If we have an externally synchronized Linux clock, then update CMOS clock
464 * accordingly every ~11 minutes. set_rtc_mmss() will be called in the jiffy
465 * closest to exactly 500 ms before the next second. If the update fails, we
466 * don't care, as it'll be updated on the next turn, and the problem (time way
467 * off) isn't likely to go away much sooner anyway.
468 */
469
470 if (ntp_synced() && xtime.tv_sec > rtc_update &&
471 abs(xtime.tv_nsec - 500000000) <= tick_nsec / 2) {
472 set_rtc_mmss(xtime.tv_sec);
473 rtc_update = xtime.tv_sec + 660;
474 }
475
476 write_sequnlock(&xtime_lock);
477
478 return IRQ_HANDLED;
479 }
480
481 static unsigned int cyc2ns_scale;
482 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
483
484 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
485 {
486 cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
487 }
488
489 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
490 {
491 return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
492 }
493
494 unsigned long long sched_clock(void)
495 {
496 unsigned long a = 0;
497
498 #if 0
499 /* Don't do a HPET read here. Using TSC always is much faster
500 and HPET may not be mapped yet when the scheduler first runs.
501 Disadvantage is a small drift between CPUs in some configurations,
502 but that should be tolerable. */
503 if (__vxtime.mode == VXTIME_HPET)
504 return (hpet_readl(HPET_COUNTER) * vxtime.quot) >> 32;
505 #endif
506
507 /* Could do CPU core sync here. Opteron can execute rdtsc speculatively,
508 which means it is not completely exact and may not be monotonous between
509 CPUs. But the errors should be too small to matter for scheduling
510 purposes. */
511
512 rdtscll(a);
513 return cycles_2_ns(a);
514 }
515
516 unsigned long get_cmos_time(void)
517 {
518 unsigned int timeout, year, mon, day, hour, min, sec;
519 unsigned char last, this;
520 unsigned long flags;
521
522 /*
523 * The Linux interpretation of the CMOS clock register contents: When the
524 * Update-In-Progress (UIP) flag goes from 1 to 0, the RTC registers show the
525 * second which has precisely just started. Waiting for this can take up to 1
526 * second, we timeout approximately after 2.4 seconds on a machine with
527 * standard 8.3 MHz ISA bus.
528 */
529
530 spin_lock_irqsave(&rtc_lock, flags);
531
532 timeout = 1000000;
533 last = this = 0;
534
535 while (timeout && last && !this) {
536 last = this;
537 this = CMOS_READ(RTC_FREQ_SELECT) & RTC_UIP;
538 timeout--;
539 }
540
541 /*
542 * Here we are safe to assume the registers won't change for a whole second, so
543 * we just go ahead and read them.
544 */
545
546 sec = CMOS_READ(RTC_SECONDS);
547 min = CMOS_READ(RTC_MINUTES);
548 hour = CMOS_READ(RTC_HOURS);
549 day = CMOS_READ(RTC_DAY_OF_MONTH);
550 mon = CMOS_READ(RTC_MONTH);
551 year = CMOS_READ(RTC_YEAR);
552
553 spin_unlock_irqrestore(&rtc_lock, flags);
554
555 /*
556 * We know that x86-64 always uses BCD format, no need to check the config
557 * register.
558 */
559
560 BCD_TO_BIN(sec);
561 BCD_TO_BIN(min);
562 BCD_TO_BIN(hour);
563 BCD_TO_BIN(day);
564 BCD_TO_BIN(mon);
565 BCD_TO_BIN(year);
566
567 /*
568 * x86-64 systems only exists since 2002.
569 * This will work up to Dec 31, 2100
570 */
571 year += 2000;
572
573 return mktime(year, mon, day, hour, min, sec);
574 }
575
576 #ifdef CONFIG_CPU_FREQ
577
578 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
579 changes.
580
581 RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
582 not that important because current Opteron setups do not support
583 scaling on SMP anyroads.
584
585 Should fix up last_tsc too. Currently gettimeofday in the
586 first tick after the change will be slightly wrong. */
587
588 #include <linux/workqueue.h>
589
590 static unsigned int cpufreq_delayed_issched = 0;
591 static unsigned int cpufreq_init = 0;
592 static struct work_struct cpufreq_delayed_get_work;
593
594 static void handle_cpufreq_delayed_get(void *v)
595 {
596 unsigned int cpu;
597 for_each_online_cpu(cpu) {
598 cpufreq_get(cpu);
599 }
600 cpufreq_delayed_issched = 0;
601 }
602
603 /* if we notice lost ticks, schedule a call to cpufreq_get() as it tries
604 * to verify the CPU frequency the timing core thinks the CPU is running
605 * at is still correct.
606 */
607 static void cpufreq_delayed_get(void)
608 {
609 static int warned;
610 if (cpufreq_init && !cpufreq_delayed_issched) {
611 cpufreq_delayed_issched = 1;
612 if (!warned) {
613 warned = 1;
614 printk(KERN_DEBUG "Losing some ticks... checking if CPU frequency changed.\n");
615 }
616 schedule_work(&cpufreq_delayed_get_work);
617 }
618 }
619
620 static unsigned int ref_freq = 0;
621 static unsigned long loops_per_jiffy_ref = 0;
622
623 static unsigned long cpu_khz_ref = 0;
624
625 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
626 void *data)
627 {
628 struct cpufreq_freqs *freq = data;
629 unsigned long *lpj, dummy;
630
631 if (cpu_has(&cpu_data[freq->cpu], X86_FEATURE_CONSTANT_TSC))
632 return 0;
633
634 lpj = &dummy;
635 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
636 #ifdef CONFIG_SMP
637 lpj = &cpu_data[freq->cpu].loops_per_jiffy;
638 #else
639 lpj = &boot_cpu_data.loops_per_jiffy;
640 #endif
641
642 if (!ref_freq) {
643 ref_freq = freq->old;
644 loops_per_jiffy_ref = *lpj;
645 cpu_khz_ref = cpu_khz;
646 }
647 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
648 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
649 (val == CPUFREQ_RESUMECHANGE)) {
650 *lpj =
651 cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
652
653 cpu_khz = cpufreq_scale(cpu_khz_ref, ref_freq, freq->new);
654 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
655 vxtime.tsc_quot = (1000L << 32) / cpu_khz;
656 }
657
658 set_cyc2ns_scale(cpu_khz_ref);
659
660 return 0;
661 }
662
663 static struct notifier_block time_cpufreq_notifier_block = {
664 .notifier_call = time_cpufreq_notifier
665 };
666
667 static int __init cpufreq_tsc(void)
668 {
669 INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get, NULL);
670 if (!cpufreq_register_notifier(&time_cpufreq_notifier_block,
671 CPUFREQ_TRANSITION_NOTIFIER))
672 cpufreq_init = 1;
673 return 0;
674 }
675
676 core_initcall(cpufreq_tsc);
677
678 #endif
679
680 /*
681 * calibrate_tsc() calibrates the processor TSC in a very simple way, comparing
682 * it to the HPET timer of known frequency.
683 */
684
685 #define TICK_COUNT 100000000
686
687 static unsigned int __init hpet_calibrate_tsc(void)
688 {
689 int tsc_start, hpet_start;
690 int tsc_now, hpet_now;
691 unsigned long flags;
692
693 local_irq_save(flags);
694 local_irq_disable();
695
696 hpet_start = hpet_readl(HPET_COUNTER);
697 rdtscl(tsc_start);
698
699 do {
700 local_irq_disable();
701 hpet_now = hpet_readl(HPET_COUNTER);
702 sync_core();
703 rdtscl(tsc_now);
704 local_irq_restore(flags);
705 } while ((tsc_now - tsc_start) < TICK_COUNT &&
706 (hpet_now - hpet_start) < TICK_COUNT);
707
708 return (tsc_now - tsc_start) * 1000000000L
709 / ((hpet_now - hpet_start) * hpet_period / 1000);
710 }
711
712
713 /*
714 * pit_calibrate_tsc() uses the speaker output (channel 2) of
715 * the PIT. This is better than using the timer interrupt output,
716 * because we can read the value of the speaker with just one inb(),
717 * where we need three i/o operations for the interrupt channel.
718 * We count how many ticks the TSC does in 50 ms.
719 */
720
721 static unsigned int __init pit_calibrate_tsc(void)
722 {
723 unsigned long start, end;
724 unsigned long flags;
725
726 spin_lock_irqsave(&i8253_lock, flags);
727
728 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
729
730 outb(0xb0, 0x43);
731 outb((PIT_TICK_RATE / (1000 / 50)) & 0xff, 0x42);
732 outb((PIT_TICK_RATE / (1000 / 50)) >> 8, 0x42);
733 rdtscll(start);
734 sync_core();
735 while ((inb(0x61) & 0x20) == 0);
736 sync_core();
737 rdtscll(end);
738
739 spin_unlock_irqrestore(&i8253_lock, flags);
740
741 return (end - start) / 50;
742 }
743
744 #ifdef CONFIG_HPET
745 static __init int late_hpet_init(void)
746 {
747 struct hpet_data hd;
748 unsigned int ntimer;
749
750 if (!vxtime.hpet_address)
751 return -1;
752
753 memset(&hd, 0, sizeof (hd));
754
755 ntimer = hpet_readl(HPET_ID);
756 ntimer = (ntimer & HPET_ID_NUMBER) >> HPET_ID_NUMBER_SHIFT;
757 ntimer++;
758
759 /*
760 * Register with driver.
761 * Timer0 and Timer1 is used by platform.
762 */
763 hd.hd_phys_address = vxtime.hpet_address;
764 hd.hd_address = (void *)fix_to_virt(FIX_HPET_BASE);
765 hd.hd_nirqs = ntimer;
766 hd.hd_flags = HPET_DATA_PLATFORM;
767 hpet_reserve_timer(&hd, 0);
768 #ifdef CONFIG_HPET_EMULATE_RTC
769 hpet_reserve_timer(&hd, 1);
770 #endif
771 hd.hd_irq[0] = HPET_LEGACY_8254;
772 hd.hd_irq[1] = HPET_LEGACY_RTC;
773 if (ntimer > 2) {
774 struct hpet *hpet;
775 struct hpet_timer *timer;
776 int i;
777
778 hpet = (struct hpet *) fix_to_virt(FIX_HPET_BASE);
779
780 for (i = 2, timer = &hpet->hpet_timers[2]; i < ntimer;
781 timer++, i++)
782 hd.hd_irq[i] = (timer->hpet_config &
783 Tn_INT_ROUTE_CNF_MASK) >>
784 Tn_INT_ROUTE_CNF_SHIFT;
785
786 }
787
788 hpet_alloc(&hd);
789 return 0;
790 }
791 fs_initcall(late_hpet_init);
792 #endif
793
794 static int hpet_timer_stop_set_go(unsigned long tick)
795 {
796 unsigned int cfg;
797
798 /*
799 * Stop the timers and reset the main counter.
800 */
801
802 cfg = hpet_readl(HPET_CFG);
803 cfg &= ~(HPET_CFG_ENABLE | HPET_CFG_LEGACY);
804 hpet_writel(cfg, HPET_CFG);
805 hpet_writel(0, HPET_COUNTER);
806 hpet_writel(0, HPET_COUNTER + 4);
807
808 /*
809 * Set up timer 0, as periodic with first interrupt to happen at hpet_tick,
810 * and period also hpet_tick.
811 */
812 if (hpet_use_timer) {
813 hpet_writel(HPET_TN_ENABLE | HPET_TN_PERIODIC | HPET_TN_SETVAL |
814 HPET_TN_32BIT, HPET_T0_CFG);
815 hpet_writel(hpet_tick, HPET_T0_CMP);
816 hpet_writel(hpet_tick, HPET_T0_CMP); /* AK: why twice? */
817 cfg |= HPET_CFG_LEGACY;
818 }
819 /*
820 * Go!
821 */
822
823 cfg |= HPET_CFG_ENABLE;
824 hpet_writel(cfg, HPET_CFG);
825
826 return 0;
827 }
828
829 static int hpet_init(void)
830 {
831 unsigned int id;
832
833 if (!vxtime.hpet_address)
834 return -1;
835 set_fixmap_nocache(FIX_HPET_BASE, vxtime.hpet_address);
836 __set_fixmap(VSYSCALL_HPET, vxtime.hpet_address, PAGE_KERNEL_VSYSCALL_NOCACHE);
837
838 /*
839 * Read the period, compute tick and quotient.
840 */
841
842 id = hpet_readl(HPET_ID);
843
844 if (!(id & HPET_ID_VENDOR) || !(id & HPET_ID_NUMBER))
845 return -1;
846
847 hpet_period = hpet_readl(HPET_PERIOD);
848 if (hpet_period < 100000 || hpet_period > 100000000)
849 return -1;
850
851 hpet_tick = (1000000000L * (USEC_PER_SEC / HZ) + hpet_period / 2) /
852 hpet_period;
853
854 hpet_use_timer = (id & HPET_ID_LEGSUP);
855
856 return hpet_timer_stop_set_go(hpet_tick);
857 }
858
859 static int hpet_reenable(void)
860 {
861 return hpet_timer_stop_set_go(hpet_tick);
862 }
863
864 void __init pit_init(void)
865 {
866 unsigned long flags;
867
868 spin_lock_irqsave(&i8253_lock, flags);
869 outb_p(0x34, 0x43); /* binary, mode 2, LSB/MSB, ch 0 */
870 outb_p(LATCH & 0xff, 0x40); /* LSB */
871 outb_p(LATCH >> 8, 0x40); /* MSB */
872 spin_unlock_irqrestore(&i8253_lock, flags);
873 }
874
875 int __init time_setup(char *str)
876 {
877 report_lost_ticks = 1;
878 return 1;
879 }
880
881 static struct irqaction irq0 = {
882 timer_interrupt, SA_INTERRUPT, CPU_MASK_NONE, "timer", NULL, NULL
883 };
884
885 extern void __init config_acpi_tables(void);
886
887 void __init time_init(void)
888 {
889 char *timename;
890
891 #ifdef HPET_HACK_ENABLE_DANGEROUS
892 if (!vxtime.hpet_address) {
893 printk(KERN_WARNING "time.c: WARNING: Enabling HPET base "
894 "manually!\n");
895 outl(0x800038a0, 0xcf8);
896 outl(0xff000001, 0xcfc);
897 outl(0x800038a0, 0xcf8);
898 vxtime.hpet_address = inl(0xcfc) & 0xfffffffe;
899 printk(KERN_WARNING "time.c: WARNING: Enabled HPET "
900 "at %#lx.\n", vxtime.hpet_address);
901 }
902 #endif
903 if (nohpet)
904 vxtime.hpet_address = 0;
905
906 xtime.tv_sec = get_cmos_time();
907 xtime.tv_nsec = 0;
908
909 set_normalized_timespec(&wall_to_monotonic,
910 -xtime.tv_sec, -xtime.tv_nsec);
911
912 if (!hpet_init())
913 vxtime_hz = (1000000000000000L + hpet_period / 2) /
914 hpet_period;
915
916 if (hpet_use_timer) {
917 cpu_khz = hpet_calibrate_tsc();
918 timename = "HPET";
919 #ifdef CONFIG_X86_PM_TIMER
920 } else if (pmtmr_ioport) {
921 vxtime_hz = PM_TIMER_FREQUENCY;
922 timename = "PM";
923 pit_init();
924 cpu_khz = pit_calibrate_tsc();
925 #endif
926 } else {
927 pit_init();
928 cpu_khz = pit_calibrate_tsc();
929 timename = "PIT";
930 }
931
932 printk(KERN_INFO "time.c: Using %ld.%06ld MHz %s timer.\n",
933 vxtime_hz / 1000000, vxtime_hz % 1000000, timename);
934 printk(KERN_INFO "time.c: Detected %d.%03d MHz processor.\n",
935 cpu_khz / 1000, cpu_khz % 1000);
936 vxtime.mode = VXTIME_TSC;
937 vxtime.quot = (1000000L << 32) / vxtime_hz;
938 vxtime.tsc_quot = (1000L << 32) / cpu_khz;
939 rdtscll_sync(&vxtime.last_tsc);
940 setup_irq(0, &irq0);
941
942 set_cyc2ns_scale(cpu_khz);
943
944 #ifndef CONFIG_SMP
945 time_init_gtod();
946 #endif
947 }
948
949 /*
950 * Make an educated guess if the TSC is trustworthy and synchronized
951 * over all CPUs.
952 */
953 static __init int unsynchronized_tsc(void)
954 {
955 #ifdef CONFIG_SMP
956 if (oem_force_hpet_timer())
957 return 1;
958 /* Intel systems are normally all synchronized. Exceptions
959 are handled in the OEM check above. */
960 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
961 return 0;
962 #endif
963 /* Assume multi socket systems are not synchronized */
964 return num_online_cpus() > 1;
965 }
966
967 /*
968 * Decide after all CPUs are booted what mode gettimeofday should use.
969 */
970 void __init time_init_gtod(void)
971 {
972 char *timetype;
973
974 if (unsynchronized_tsc())
975 notsc = 1;
976 if (vxtime.hpet_address && notsc) {
977 timetype = hpet_use_timer ? "HPET" : "PIT/HPET";
978 vxtime.last = hpet_readl(HPET_T0_CMP) - hpet_tick;
979 vxtime.mode = VXTIME_HPET;
980 do_gettimeoffset = do_gettimeoffset_hpet;
981 #ifdef CONFIG_X86_PM_TIMER
982 /* Using PM for gettimeofday is quite slow, but we have no other
983 choice because the TSC is too unreliable on some systems. */
984 } else if (pmtmr_ioport && !vxtime.hpet_address && notsc) {
985 timetype = "PM";
986 do_gettimeoffset = do_gettimeoffset_pm;
987 vxtime.mode = VXTIME_PMTMR;
988 sysctl_vsyscall = 0;
989 printk(KERN_INFO "Disabling vsyscall due to use of PM timer\n");
990 #endif
991 } else {
992 timetype = hpet_use_timer ? "HPET/TSC" : "PIT/TSC";
993 vxtime.mode = VXTIME_TSC;
994 }
995
996 printk(KERN_INFO "time.c: Using %s based timekeeping.\n", timetype);
997 }
998
999 __setup("report_lost_ticks", time_setup);
1000
1001 static long clock_cmos_diff;
1002 static unsigned long sleep_start;
1003
1004 static int timer_suspend(struct sys_device *dev, pm_message_t state)
1005 {
1006 /*
1007 * Estimate time zone so that set_time can update the clock
1008 */
1009 long cmos_time = get_cmos_time();
1010
1011 clock_cmos_diff = -cmos_time;
1012 clock_cmos_diff += get_seconds();
1013 sleep_start = cmos_time;
1014 return 0;
1015 }
1016
1017 static int timer_resume(struct sys_device *dev)
1018 {
1019 unsigned long flags;
1020 unsigned long sec;
1021 unsigned long ctime = get_cmos_time();
1022 unsigned long sleep_length = (ctime - sleep_start) * HZ;
1023
1024 if (vxtime.hpet_address)
1025 hpet_reenable();
1026 else
1027 i8254_timer_resume();
1028
1029 sec = ctime + clock_cmos_diff;
1030 write_seqlock_irqsave(&xtime_lock,flags);
1031 xtime.tv_sec = sec;
1032 xtime.tv_nsec = 0;
1033 write_sequnlock_irqrestore(&xtime_lock,flags);
1034 jiffies += sleep_length;
1035 wall_jiffies += sleep_length;
1036 touch_softlockup_watchdog();
1037 return 0;
1038 }
1039
1040 static struct sysdev_class timer_sysclass = {
1041 .resume = timer_resume,
1042 .suspend = timer_suspend,
1043 set_kset_name("timer"),
1044 };
1045
1046
1047 /* XXX this driverfs stuff should probably go elsewhere later -john */
1048 static struct sys_device device_timer = {
1049 .id = 0,
1050 .cls = &timer_sysclass,
1051 };
1052
1053 static int time_init_device(void)
1054 {
1055 int error = sysdev_class_register(&timer_sysclass);
1056 if (!error)
1057 error = sysdev_register(&device_timer);
1058 return error;
1059 }
1060
1061 device_initcall(time_init_device);
1062
1063 #ifdef CONFIG_HPET_EMULATE_RTC
1064 /* HPET in LegacyReplacement Mode eats up RTC interrupt line. When, HPET
1065 * is enabled, we support RTC interrupt functionality in software.
1066 * RTC has 3 kinds of interrupts:
1067 * 1) Update Interrupt - generate an interrupt, every sec, when RTC clock
1068 * is updated
1069 * 2) Alarm Interrupt - generate an interrupt at a specific time of day
1070 * 3) Periodic Interrupt - generate periodic interrupt, with frequencies
1071 * 2Hz-8192Hz (2Hz-64Hz for non-root user) (all freqs in powers of 2)
1072 * (1) and (2) above are implemented using polling at a frequency of
1073 * 64 Hz. The exact frequency is a tradeoff between accuracy and interrupt
1074 * overhead. (DEFAULT_RTC_INT_FREQ)
1075 * For (3), we use interrupts at 64Hz or user specified periodic
1076 * frequency, whichever is higher.
1077 */
1078 #include <linux/rtc.h>
1079
1080 extern irqreturn_t rtc_interrupt(int irq, void *dev_id, struct pt_regs *regs);
1081
1082 #define DEFAULT_RTC_INT_FREQ 64
1083 #define RTC_NUM_INTS 1
1084
1085 static unsigned long UIE_on;
1086 static unsigned long prev_update_sec;
1087
1088 static unsigned long AIE_on;
1089 static struct rtc_time alarm_time;
1090
1091 static unsigned long PIE_on;
1092 static unsigned long PIE_freq = DEFAULT_RTC_INT_FREQ;
1093 static unsigned long PIE_count;
1094
1095 static unsigned long hpet_rtc_int_freq; /* RTC interrupt frequency */
1096
1097 int is_hpet_enabled(void)
1098 {
1099 return vxtime.hpet_address != 0;
1100 }
1101
1102 /*
1103 * Timer 1 for RTC, we do not use periodic interrupt feature,
1104 * even if HPET supports periodic interrupts on Timer 1.
1105 * The reason being, to set up a periodic interrupt in HPET, we need to
1106 * stop the main counter. And if we do that everytime someone diables/enables
1107 * RTC, we will have adverse effect on main kernel timer running on Timer 0.
1108 * So, for the time being, simulate the periodic interrupt in software.
1109 *
1110 * hpet_rtc_timer_init() is called for the first time and during subsequent
1111 * interuppts reinit happens through hpet_rtc_timer_reinit().
1112 */
1113 int hpet_rtc_timer_init(void)
1114 {
1115 unsigned int cfg, cnt;
1116 unsigned long flags;
1117
1118 if (!is_hpet_enabled())
1119 return 0;
1120 /*
1121 * Set the counter 1 and enable the interrupts.
1122 */
1123 if (PIE_on && (PIE_freq > DEFAULT_RTC_INT_FREQ))
1124 hpet_rtc_int_freq = PIE_freq;
1125 else
1126 hpet_rtc_int_freq = DEFAULT_RTC_INT_FREQ;
1127
1128 local_irq_save(flags);
1129 cnt = hpet_readl(HPET_COUNTER);
1130 cnt += ((hpet_tick*HZ)/hpet_rtc_int_freq);
1131 hpet_writel(cnt, HPET_T1_CMP);
1132 local_irq_restore(flags);
1133
1134 cfg = hpet_readl(HPET_T1_CFG);
1135 cfg |= HPET_TN_ENABLE | HPET_TN_SETVAL | HPET_TN_32BIT;
1136 hpet_writel(cfg, HPET_T1_CFG);
1137
1138 return 1;
1139 }
1140
1141 static void hpet_rtc_timer_reinit(void)
1142 {
1143 unsigned int cfg, cnt;
1144
1145 if (!(PIE_on | AIE_on | UIE_on))
1146 return;
1147
1148 if (PIE_on && (PIE_freq > DEFAULT_RTC_INT_FREQ))
1149 hpet_rtc_int_freq = PIE_freq;
1150 else
1151 hpet_rtc_int_freq = DEFAULT_RTC_INT_FREQ;
1152
1153 /* It is more accurate to use the comparator value than current count.*/
1154 cnt = hpet_readl(HPET_T1_CMP);
1155 cnt += hpet_tick*HZ/hpet_rtc_int_freq;
1156 hpet_writel(cnt, HPET_T1_CMP);
1157
1158 cfg = hpet_readl(HPET_T1_CFG);
1159 cfg |= HPET_TN_ENABLE | HPET_TN_SETVAL | HPET_TN_32BIT;
1160 hpet_writel(cfg, HPET_T1_CFG);
1161
1162 return;
1163 }
1164
1165 /*
1166 * The functions below are called from rtc driver.
1167 * Return 0 if HPET is not being used.
1168 * Otherwise do the necessary changes and return 1.
1169 */
1170 int hpet_mask_rtc_irq_bit(unsigned long bit_mask)
1171 {
1172 if (!is_hpet_enabled())
1173 return 0;
1174
1175 if (bit_mask & RTC_UIE)
1176 UIE_on = 0;
1177 if (bit_mask & RTC_PIE)
1178 PIE_on = 0;
1179 if (bit_mask & RTC_AIE)
1180 AIE_on = 0;
1181
1182 return 1;
1183 }
1184
1185 int hpet_set_rtc_irq_bit(unsigned long bit_mask)
1186 {
1187 int timer_init_reqd = 0;
1188
1189 if (!is_hpet_enabled())
1190 return 0;
1191
1192 if (!(PIE_on | AIE_on | UIE_on))
1193 timer_init_reqd = 1;
1194
1195 if (bit_mask & RTC_UIE) {
1196 UIE_on = 1;
1197 }
1198 if (bit_mask & RTC_PIE) {
1199 PIE_on = 1;
1200 PIE_count = 0;
1201 }
1202 if (bit_mask & RTC_AIE) {
1203 AIE_on = 1;
1204 }
1205
1206 if (timer_init_reqd)
1207 hpet_rtc_timer_init();
1208
1209 return 1;
1210 }
1211
1212 int hpet_set_alarm_time(unsigned char hrs, unsigned char min, unsigned char sec)
1213 {
1214 if (!is_hpet_enabled())
1215 return 0;
1216
1217 alarm_time.tm_hour = hrs;
1218 alarm_time.tm_min = min;
1219 alarm_time.tm_sec = sec;
1220
1221 return 1;
1222 }
1223
1224 int hpet_set_periodic_freq(unsigned long freq)
1225 {
1226 if (!is_hpet_enabled())
1227 return 0;
1228
1229 PIE_freq = freq;
1230 PIE_count = 0;
1231
1232 return 1;
1233 }
1234
1235 int hpet_rtc_dropped_irq(void)
1236 {
1237 if (!is_hpet_enabled())
1238 return 0;
1239
1240 return 1;
1241 }
1242
1243 irqreturn_t hpet_rtc_interrupt(int irq, void *dev_id, struct pt_regs *regs)
1244 {
1245 struct rtc_time curr_time;
1246 unsigned long rtc_int_flag = 0;
1247 int call_rtc_interrupt = 0;
1248
1249 hpet_rtc_timer_reinit();
1250
1251 if (UIE_on | AIE_on) {
1252 rtc_get_rtc_time(&curr_time);
1253 }
1254 if (UIE_on) {
1255 if (curr_time.tm_sec != prev_update_sec) {
1256 /* Set update int info, call real rtc int routine */
1257 call_rtc_interrupt = 1;
1258 rtc_int_flag = RTC_UF;
1259 prev_update_sec = curr_time.tm_sec;
1260 }
1261 }
1262 if (PIE_on) {
1263 PIE_count++;
1264 if (PIE_count >= hpet_rtc_int_freq/PIE_freq) {
1265 /* Set periodic int info, call real rtc int routine */
1266 call_rtc_interrupt = 1;
1267 rtc_int_flag |= RTC_PF;
1268 PIE_count = 0;
1269 }
1270 }
1271 if (AIE_on) {
1272 if ((curr_time.tm_sec == alarm_time.tm_sec) &&
1273 (curr_time.tm_min == alarm_time.tm_min) &&
1274 (curr_time.tm_hour == alarm_time.tm_hour)) {
1275 /* Set alarm int info, call real rtc int routine */
1276 call_rtc_interrupt = 1;
1277 rtc_int_flag |= RTC_AF;
1278 }
1279 }
1280 if (call_rtc_interrupt) {
1281 rtc_int_flag |= (RTC_IRQF | (RTC_NUM_INTS << 8));
1282 rtc_interrupt(rtc_int_flag, dev_id, regs);
1283 }
1284 return IRQ_HANDLED;
1285 }
1286 #endif
1287
1288
1289
1290 static int __init nohpet_setup(char *s)
1291 {
1292 nohpet = 1;
1293 return 0;
1294 }
1295
1296 __setup("nohpet", nohpet_setup);
1297
1298
1299 static int __init notsc_setup(char *s)
1300 {
1301 notsc = 1;
1302 return 0;
1303 }
1304
1305 __setup("notsc", notsc_setup);
1306
1307