]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - arch/x86_64/kernel/time.c
Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/cpufreq
[mirror_ubuntu-jammy-kernel.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/irq.h>
22 #include <linux/time.h>
23 #include <linux/ioport.h>
24 #include <linux/module.h>
25 #include <linux/device.h>
26 #include <linux/sysdev.h>
27 #include <linux/bcd.h>
28 #include <linux/kallsyms.h>
29 #include <linux/acpi.h>
30 #ifdef CONFIG_ACPI
31 #include <acpi/achware.h> /* for PM timer frequency */
32 #endif
33 #include <asm/8253pit.h>
34 #include <asm/pgtable.h>
35 #include <asm/vsyscall.h>
36 #include <asm/timex.h>
37 #include <asm/proto.h>
38 #include <asm/hpet.h>
39 #include <asm/sections.h>
40 #include <linux/cpufreq.h>
41 #include <linux/hpet.h>
42 #ifdef CONFIG_X86_LOCAL_APIC
43 #include <asm/apic.h>
44 #endif
45
46 u64 jiffies_64 = INITIAL_JIFFIES;
47
48 EXPORT_SYMBOL(jiffies_64);
49
50 #ifdef CONFIG_CPU_FREQ
51 static void cpufreq_delayed_get(void);
52 #endif
53 extern void i8254_timer_resume(void);
54 extern int using_apic_timer;
55
56 DEFINE_SPINLOCK(rtc_lock);
57 DEFINE_SPINLOCK(i8253_lock);
58
59 static int nohpet __initdata = 0;
60 static int notsc __initdata = 0;
61
62 #undef HPET_HACK_ENABLE_DANGEROUS
63
64 unsigned int cpu_khz; /* TSC clocks / usec, not used here */
65 static unsigned long hpet_period; /* fsecs / HPET clock */
66 unsigned long hpet_tick; /* HPET clocks / interrupt */
67 static int hpet_use_timer;
68 unsigned long vxtime_hz = PIT_TICK_RATE;
69 int report_lost_ticks; /* command line option */
70 unsigned long long monotonic_base;
71
72 struct vxtime_data __vxtime __section_vxtime; /* for vsyscalls */
73
74 volatile unsigned long __jiffies __section_jiffies = INITIAL_JIFFIES;
75 unsigned long __wall_jiffies __section_wall_jiffies = INITIAL_JIFFIES;
76 struct timespec __xtime __section_xtime;
77 struct timezone __sys_tz __section_sys_tz;
78
79 static inline void rdtscll_sync(unsigned long *tsc)
80 {
81 #ifdef CONFIG_SMP
82 sync_core();
83 #endif
84 rdtscll(*tsc);
85 }
86
87 /*
88 * do_gettimeoffset() returns microseconds since last timer interrupt was
89 * triggered by hardware. A memory read of HPET is slower than a register read
90 * of TSC, but much more reliable. It's also synchronized to the timer
91 * interrupt. Note that do_gettimeoffset() may return more than hpet_tick, if a
92 * timer interrupt has happened already, but vxtime.trigger wasn't updated yet.
93 * This is not a problem, because jiffies hasn't updated either. They are bound
94 * together by xtime_lock.
95 */
96
97 static inline unsigned int do_gettimeoffset_tsc(void)
98 {
99 unsigned long t;
100 unsigned long x;
101 rdtscll_sync(&t);
102 if (t < vxtime.last_tsc) t = vxtime.last_tsc; /* hack */
103 x = ((t - vxtime.last_tsc) * vxtime.tsc_quot) >> 32;
104 return x;
105 }
106
107 static inline unsigned int do_gettimeoffset_hpet(void)
108 {
109 /* cap counter read to one tick to avoid inconsistencies */
110 unsigned long counter = hpet_readl(HPET_COUNTER) - vxtime.last;
111 return (min(counter,hpet_tick) * vxtime.quot) >> 32;
112 }
113
114 unsigned int (*do_gettimeoffset)(void) = do_gettimeoffset_tsc;
115
116 /*
117 * This version of gettimeofday() has microsecond resolution and better than
118 * microsecond precision, as we're using at least a 10 MHz (usually 14.31818
119 * MHz) HPET timer.
120 */
121
122 void do_gettimeofday(struct timeval *tv)
123 {
124 unsigned long seq, t;
125 unsigned int sec, usec;
126
127 do {
128 seq = read_seqbegin(&xtime_lock);
129
130 sec = xtime.tv_sec;
131 usec = xtime.tv_nsec / 1000;
132
133 /* i386 does some correction here to keep the clock
134 monotonous even when ntpd is fixing drift.
135 But they didn't work for me, there is a non monotonic
136 clock anyways with ntp.
137 I dropped all corrections now until a real solution can
138 be found. Note when you fix it here you need to do the same
139 in arch/x86_64/kernel/vsyscall.c and export all needed
140 variables in vmlinux.lds. -AK */
141
142 t = (jiffies - wall_jiffies) * (1000000L / HZ) +
143 do_gettimeoffset();
144 usec += t;
145
146 } while (read_seqretry(&xtime_lock, seq));
147
148 tv->tv_sec = sec + usec / 1000000;
149 tv->tv_usec = usec % 1000000;
150 }
151
152 EXPORT_SYMBOL(do_gettimeofday);
153
154 /*
155 * settimeofday() first undoes the correction that gettimeofday would do
156 * on the time, and then saves it. This is ugly, but has been like this for
157 * ages already.
158 */
159
160 int do_settimeofday(struct timespec *tv)
161 {
162 time_t wtm_sec, sec = tv->tv_sec;
163 long wtm_nsec, nsec = tv->tv_nsec;
164
165 if ((unsigned long)tv->tv_nsec >= NSEC_PER_SEC)
166 return -EINVAL;
167
168 write_seqlock_irq(&xtime_lock);
169
170 nsec -= do_gettimeoffset() * 1000 +
171 (jiffies - wall_jiffies) * (NSEC_PER_SEC/HZ);
172
173 wtm_sec = wall_to_monotonic.tv_sec + (xtime.tv_sec - sec);
174 wtm_nsec = wall_to_monotonic.tv_nsec + (xtime.tv_nsec - nsec);
175
176 set_normalized_timespec(&xtime, sec, nsec);
177 set_normalized_timespec(&wall_to_monotonic, wtm_sec, wtm_nsec);
178
179 time_adjust = 0; /* stop active adjtime() */
180 time_status |= STA_UNSYNC;
181 time_maxerror = NTP_PHASE_LIMIT;
182 time_esterror = NTP_PHASE_LIMIT;
183
184 write_sequnlock_irq(&xtime_lock);
185 clock_was_set();
186 return 0;
187 }
188
189 EXPORT_SYMBOL(do_settimeofday);
190
191 unsigned long profile_pc(struct pt_regs *regs)
192 {
193 unsigned long pc = instruction_pointer(regs);
194
195 /* Assume the lock function has either no stack frame or only a single word.
196 This checks if the address on the stack looks like a kernel text address.
197 There is a small window for false hits, but in that case the tick
198 is just accounted to the spinlock function.
199 Better would be to write these functions in assembler again
200 and check exactly. */
201 if (in_lock_functions(pc)) {
202 char *v = *(char **)regs->rsp;
203 if ((v >= _stext && v <= _etext) ||
204 (v >= _sinittext && v <= _einittext) ||
205 (v >= (char *)MODULES_VADDR && v <= (char *)MODULES_END))
206 return (unsigned long)v;
207 return ((unsigned long *)regs->rsp)[1];
208 }
209 return pc;
210 }
211 EXPORT_SYMBOL(profile_pc);
212
213 /*
214 * In order to set the CMOS clock precisely, set_rtc_mmss has to be called 500
215 * ms after the second nowtime has started, because when nowtime is written
216 * into the registers of the CMOS clock, it will jump to the next second
217 * precisely 500 ms later. Check the Motorola MC146818A or Dallas DS12887 data
218 * sheet for details.
219 */
220
221 static void set_rtc_mmss(unsigned long nowtime)
222 {
223 int real_seconds, real_minutes, cmos_minutes;
224 unsigned char control, freq_select;
225
226 /*
227 * IRQs are disabled when we're called from the timer interrupt,
228 * no need for spin_lock_irqsave()
229 */
230
231 spin_lock(&rtc_lock);
232
233 /*
234 * Tell the clock it's being set and stop it.
235 */
236
237 control = CMOS_READ(RTC_CONTROL);
238 CMOS_WRITE(control | RTC_SET, RTC_CONTROL);
239
240 freq_select = CMOS_READ(RTC_FREQ_SELECT);
241 CMOS_WRITE(freq_select | RTC_DIV_RESET2, RTC_FREQ_SELECT);
242
243 cmos_minutes = CMOS_READ(RTC_MINUTES);
244 BCD_TO_BIN(cmos_minutes);
245
246 /*
247 * since we're only adjusting minutes and seconds, don't interfere with hour
248 * overflow. This avoids messing with unknown time zones but requires your RTC
249 * not to be off by more than 15 minutes. Since we're calling it only when
250 * our clock is externally synchronized using NTP, this shouldn't be a problem.
251 */
252
253 real_seconds = nowtime % 60;
254 real_minutes = nowtime / 60;
255 if (((abs(real_minutes - cmos_minutes) + 15) / 30) & 1)
256 real_minutes += 30; /* correct for half hour time zone */
257 real_minutes %= 60;
258
259 #if 0
260 /* AMD 8111 is a really bad time keeper and hits this regularly.
261 It probably was an attempt to avoid screwing up DST, but ignore
262 that for now. */
263 if (abs(real_minutes - cmos_minutes) >= 30) {
264 printk(KERN_WARNING "time.c: can't update CMOS clock "
265 "from %d to %d\n", cmos_minutes, real_minutes);
266 } else
267 #endif
268
269 {
270 BIN_TO_BCD(real_seconds);
271 BIN_TO_BCD(real_minutes);
272 CMOS_WRITE(real_seconds, RTC_SECONDS);
273 CMOS_WRITE(real_minutes, RTC_MINUTES);
274 }
275
276 /*
277 * The following flags have to be released exactly in this order, otherwise the
278 * DS12887 (popular MC146818A clone with integrated battery and quartz) will
279 * not reset the oscillator and will not update precisely 500 ms later. You
280 * won't find this mentioned in the Dallas Semiconductor data sheets, but who
281 * believes data sheets anyway ... -- Markus Kuhn
282 */
283
284 CMOS_WRITE(control, RTC_CONTROL);
285 CMOS_WRITE(freq_select, RTC_FREQ_SELECT);
286
287 spin_unlock(&rtc_lock);
288 }
289
290
291 /* monotonic_clock(): returns # of nanoseconds passed since time_init()
292 * Note: This function is required to return accurate
293 * time even in the absence of multiple timer ticks.
294 */
295 unsigned long long monotonic_clock(void)
296 {
297 unsigned long seq;
298 u32 last_offset, this_offset, offset;
299 unsigned long long base;
300
301 if (vxtime.mode == VXTIME_HPET) {
302 do {
303 seq = read_seqbegin(&xtime_lock);
304
305 last_offset = vxtime.last;
306 base = monotonic_base;
307 this_offset = hpet_readl(HPET_COUNTER);
308
309 } while (read_seqretry(&xtime_lock, seq));
310 offset = (this_offset - last_offset);
311 offset *=(NSEC_PER_SEC/HZ)/hpet_tick;
312 return base + offset;
313 }else{
314 do {
315 seq = read_seqbegin(&xtime_lock);
316
317 last_offset = vxtime.last_tsc;
318 base = monotonic_base;
319 } while (read_seqretry(&xtime_lock, seq));
320 sync_core();
321 rdtscll(this_offset);
322 offset = (this_offset - last_offset)*1000/cpu_khz;
323 return base + offset;
324 }
325
326
327 }
328 EXPORT_SYMBOL(monotonic_clock);
329
330 static noinline void handle_lost_ticks(int lost, struct pt_regs *regs)
331 {
332 static long lost_count;
333 static int warned;
334
335 if (report_lost_ticks) {
336 printk(KERN_WARNING "time.c: Lost %d timer "
337 "tick(s)! ", lost);
338 print_symbol("rip %s)\n", regs->rip);
339 }
340
341 if (lost_count == 1000 && !warned) {
342 printk(KERN_WARNING
343 "warning: many lost ticks.\n"
344 KERN_WARNING "Your time source seems to be instable or "
345 "some driver is hogging interupts\n");
346 print_symbol("rip %s\n", regs->rip);
347 if (vxtime.mode == VXTIME_TSC && vxtime.hpet_address) {
348 printk(KERN_WARNING "Falling back to HPET\n");
349 vxtime.last = hpet_readl(HPET_T0_CMP) - hpet_tick;
350 vxtime.mode = VXTIME_HPET;
351 do_gettimeoffset = do_gettimeoffset_hpet;
352 }
353 /* else should fall back to PIT, but code missing. */
354 warned = 1;
355 } else
356 lost_count++;
357
358 #ifdef CONFIG_CPU_FREQ
359 /* In some cases the CPU can change frequency without us noticing
360 (like going into thermal throttle)
361 Give cpufreq a change to catch up. */
362 if ((lost_count+1) % 25 == 0) {
363 cpufreq_delayed_get();
364 }
365 #endif
366 }
367
368 static irqreturn_t timer_interrupt(int irq, void *dev_id, struct pt_regs *regs)
369 {
370 static unsigned long rtc_update = 0;
371 unsigned long tsc;
372 int delay, offset = 0, lost = 0;
373
374 /*
375 * Here we are in the timer irq handler. We have irqs locally disabled (so we
376 * don't need spin_lock_irqsave()) but we don't know if the timer_bh is running
377 * on the other CPU, so we need a lock. We also need to lock the vsyscall
378 * variables, because both do_timer() and us change them -arca+vojtech
379 */
380
381 write_seqlock(&xtime_lock);
382
383 if (vxtime.hpet_address)
384 offset = hpet_readl(HPET_COUNTER);
385
386 if (hpet_use_timer) {
387 /* if we're using the hpet timer functionality,
388 * we can more accurately know the counter value
389 * when the timer interrupt occured.
390 */
391 offset = hpet_readl(HPET_T0_CMP) - hpet_tick;
392 delay = hpet_readl(HPET_COUNTER) - offset;
393 } else {
394 spin_lock(&i8253_lock);
395 outb_p(0x00, 0x43);
396 delay = inb_p(0x40);
397 delay |= inb(0x40) << 8;
398 spin_unlock(&i8253_lock);
399 delay = LATCH - 1 - delay;
400 }
401
402 rdtscll_sync(&tsc);
403
404 if (vxtime.mode == VXTIME_HPET) {
405 if (offset - vxtime.last > hpet_tick) {
406 lost = (offset - vxtime.last) / hpet_tick - 1;
407 }
408
409 monotonic_base +=
410 (offset - vxtime.last)*(NSEC_PER_SEC/HZ) / hpet_tick;
411
412 vxtime.last = offset;
413 #ifdef CONFIG_X86_PM_TIMER
414 } else if (vxtime.mode == VXTIME_PMTMR) {
415 lost = pmtimer_mark_offset();
416 #endif
417 } else {
418 offset = (((tsc - vxtime.last_tsc) *
419 vxtime.tsc_quot) >> 32) - (USEC_PER_SEC / HZ);
420
421 if (offset < 0)
422 offset = 0;
423
424 if (offset > (USEC_PER_SEC / HZ)) {
425 lost = offset / (USEC_PER_SEC / HZ);
426 offset %= (USEC_PER_SEC / HZ);
427 }
428
429 monotonic_base += (tsc - vxtime.last_tsc)*1000000/cpu_khz ;
430
431 vxtime.last_tsc = tsc - vxtime.quot * delay / vxtime.tsc_quot;
432
433 if ((((tsc - vxtime.last_tsc) *
434 vxtime.tsc_quot) >> 32) < offset)
435 vxtime.last_tsc = tsc -
436 (((long) offset << 32) / vxtime.tsc_quot) - 1;
437 }
438
439 if (lost > 0) {
440 handle_lost_ticks(lost, regs);
441 jiffies += lost;
442 }
443
444 /*
445 * Do the timer stuff.
446 */
447
448 do_timer(regs);
449 #ifndef CONFIG_SMP
450 update_process_times(user_mode(regs));
451 #endif
452
453 /*
454 * In the SMP case we use the local APIC timer interrupt to do the profiling,
455 * except when we simulate SMP mode on a uniprocessor system, in that case we
456 * have to call the local interrupt handler.
457 */
458
459 #ifndef CONFIG_X86_LOCAL_APIC
460 profile_tick(CPU_PROFILING, regs);
461 #else
462 if (!using_apic_timer)
463 smp_local_timer_interrupt(regs);
464 #endif
465
466 /*
467 * If we have an externally synchronized Linux clock, then update CMOS clock
468 * accordingly every ~11 minutes. set_rtc_mmss() will be called in the jiffy
469 * closest to exactly 500 ms before the next second. If the update fails, we
470 * don't care, as it'll be updated on the next turn, and the problem (time way
471 * off) isn't likely to go away much sooner anyway.
472 */
473
474 if ((~time_status & STA_UNSYNC) && xtime.tv_sec > rtc_update &&
475 abs(xtime.tv_nsec - 500000000) <= tick_nsec / 2) {
476 set_rtc_mmss(xtime.tv_sec);
477 rtc_update = xtime.tv_sec + 660;
478 }
479
480 write_sequnlock(&xtime_lock);
481
482 return IRQ_HANDLED;
483 }
484
485 static unsigned int cyc2ns_scale;
486 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
487
488 static inline void set_cyc2ns_scale(unsigned long cpu_mhz)
489 {
490 cyc2ns_scale = (1000 << CYC2NS_SCALE_FACTOR)/cpu_mhz;
491 }
492
493 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
494 {
495 return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
496 }
497
498 unsigned long long sched_clock(void)
499 {
500 unsigned long a = 0;
501
502 #if 0
503 /* Don't do a HPET read here. Using TSC always is much faster
504 and HPET may not be mapped yet when the scheduler first runs.
505 Disadvantage is a small drift between CPUs in some configurations,
506 but that should be tolerable. */
507 if (__vxtime.mode == VXTIME_HPET)
508 return (hpet_readl(HPET_COUNTER) * vxtime.quot) >> 32;
509 #endif
510
511 /* Could do CPU core sync here. Opteron can execute rdtsc speculatively,
512 which means it is not completely exact and may not be monotonous between
513 CPUs. But the errors should be too small to matter for scheduling
514 purposes. */
515
516 rdtscll(a);
517 return cycles_2_ns(a);
518 }
519
520 unsigned long get_cmos_time(void)
521 {
522 unsigned int timeout, year, mon, day, hour, min, sec;
523 unsigned char last, this;
524 unsigned long flags;
525
526 /*
527 * The Linux interpretation of the CMOS clock register contents: When the
528 * Update-In-Progress (UIP) flag goes from 1 to 0, the RTC registers show the
529 * second which has precisely just started. Waiting for this can take up to 1
530 * second, we timeout approximately after 2.4 seconds on a machine with
531 * standard 8.3 MHz ISA bus.
532 */
533
534 spin_lock_irqsave(&rtc_lock, flags);
535
536 timeout = 1000000;
537 last = this = 0;
538
539 while (timeout && last && !this) {
540 last = this;
541 this = CMOS_READ(RTC_FREQ_SELECT) & RTC_UIP;
542 timeout--;
543 }
544
545 /*
546 * Here we are safe to assume the registers won't change for a whole second, so
547 * we just go ahead and read them.
548 */
549
550 sec = CMOS_READ(RTC_SECONDS);
551 min = CMOS_READ(RTC_MINUTES);
552 hour = CMOS_READ(RTC_HOURS);
553 day = CMOS_READ(RTC_DAY_OF_MONTH);
554 mon = CMOS_READ(RTC_MONTH);
555 year = CMOS_READ(RTC_YEAR);
556
557 spin_unlock_irqrestore(&rtc_lock, flags);
558
559 /*
560 * We know that x86-64 always uses BCD format, no need to check the config
561 * register.
562 */
563
564 BCD_TO_BIN(sec);
565 BCD_TO_BIN(min);
566 BCD_TO_BIN(hour);
567 BCD_TO_BIN(day);
568 BCD_TO_BIN(mon);
569 BCD_TO_BIN(year);
570
571 /*
572 * x86-64 systems only exists since 2002.
573 * This will work up to Dec 31, 2100
574 */
575 year += 2000;
576
577 return mktime(year, mon, day, hour, min, sec);
578 }
579
580 #ifdef CONFIG_CPU_FREQ
581
582 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
583 changes.
584
585 RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
586 not that important because current Opteron setups do not support
587 scaling on SMP anyroads.
588
589 Should fix up last_tsc too. Currently gettimeofday in the
590 first tick after the change will be slightly wrong. */
591
592 #include <linux/workqueue.h>
593
594 static unsigned int cpufreq_delayed_issched = 0;
595 static unsigned int cpufreq_init = 0;
596 static struct work_struct cpufreq_delayed_get_work;
597
598 static void handle_cpufreq_delayed_get(void *v)
599 {
600 unsigned int cpu;
601 for_each_online_cpu(cpu) {
602 cpufreq_get(cpu);
603 }
604 cpufreq_delayed_issched = 0;
605 }
606
607 /* if we notice lost ticks, schedule a call to cpufreq_get() as it tries
608 * to verify the CPU frequency the timing core thinks the CPU is running
609 * at is still correct.
610 */
611 static void cpufreq_delayed_get(void)
612 {
613 static int warned;
614 if (cpufreq_init && !cpufreq_delayed_issched) {
615 cpufreq_delayed_issched = 1;
616 if (!warned) {
617 warned = 1;
618 printk(KERN_DEBUG "Losing some ticks... checking if CPU frequency changed.\n");
619 }
620 schedule_work(&cpufreq_delayed_get_work);
621 }
622 }
623
624 static unsigned int ref_freq = 0;
625 static unsigned long loops_per_jiffy_ref = 0;
626
627 static unsigned long cpu_khz_ref = 0;
628
629 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
630 void *data)
631 {
632 struct cpufreq_freqs *freq = data;
633 unsigned long *lpj, dummy;
634
635 if (cpu_has(&cpu_data[freq->cpu], X86_FEATURE_CONSTANT_TSC))
636 return 0;
637
638 lpj = &dummy;
639 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
640 #ifdef CONFIG_SMP
641 lpj = &cpu_data[freq->cpu].loops_per_jiffy;
642 #else
643 lpj = &boot_cpu_data.loops_per_jiffy;
644 #endif
645
646 if (!ref_freq) {
647 ref_freq = freq->old;
648 loops_per_jiffy_ref = *lpj;
649 cpu_khz_ref = cpu_khz;
650 }
651 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
652 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
653 (val == CPUFREQ_RESUMECHANGE)) {
654 *lpj =
655 cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
656
657 cpu_khz = cpufreq_scale(cpu_khz_ref, ref_freq, freq->new);
658 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
659 vxtime.tsc_quot = (1000L << 32) / cpu_khz;
660 }
661
662 set_cyc2ns_scale(cpu_khz_ref / 1000);
663
664 return 0;
665 }
666
667 static struct notifier_block time_cpufreq_notifier_block = {
668 .notifier_call = time_cpufreq_notifier
669 };
670
671 static int __init cpufreq_tsc(void)
672 {
673 INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get, NULL);
674 if (!cpufreq_register_notifier(&time_cpufreq_notifier_block,
675 CPUFREQ_TRANSITION_NOTIFIER))
676 cpufreq_init = 1;
677 return 0;
678 }
679
680 core_initcall(cpufreq_tsc);
681
682 #endif
683
684 /*
685 * calibrate_tsc() calibrates the processor TSC in a very simple way, comparing
686 * it to the HPET timer of known frequency.
687 */
688
689 #define TICK_COUNT 100000000
690
691 static unsigned int __init hpet_calibrate_tsc(void)
692 {
693 int tsc_start, hpet_start;
694 int tsc_now, hpet_now;
695 unsigned long flags;
696
697 local_irq_save(flags);
698 local_irq_disable();
699
700 hpet_start = hpet_readl(HPET_COUNTER);
701 rdtscl(tsc_start);
702
703 do {
704 local_irq_disable();
705 hpet_now = hpet_readl(HPET_COUNTER);
706 sync_core();
707 rdtscl(tsc_now);
708 local_irq_restore(flags);
709 } while ((tsc_now - tsc_start) < TICK_COUNT &&
710 (hpet_now - hpet_start) < TICK_COUNT);
711
712 return (tsc_now - tsc_start) * 1000000000L
713 / ((hpet_now - hpet_start) * hpet_period / 1000);
714 }
715
716
717 /*
718 * pit_calibrate_tsc() uses the speaker output (channel 2) of
719 * the PIT. This is better than using the timer interrupt output,
720 * because we can read the value of the speaker with just one inb(),
721 * where we need three i/o operations for the interrupt channel.
722 * We count how many ticks the TSC does in 50 ms.
723 */
724
725 static unsigned int __init pit_calibrate_tsc(void)
726 {
727 unsigned long start, end;
728 unsigned long flags;
729
730 spin_lock_irqsave(&i8253_lock, flags);
731
732 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
733
734 outb(0xb0, 0x43);
735 outb((PIT_TICK_RATE / (1000 / 50)) & 0xff, 0x42);
736 outb((PIT_TICK_RATE / (1000 / 50)) >> 8, 0x42);
737 rdtscll(start);
738 sync_core();
739 while ((inb(0x61) & 0x20) == 0);
740 sync_core();
741 rdtscll(end);
742
743 spin_unlock_irqrestore(&i8253_lock, flags);
744
745 return (end - start) / 50;
746 }
747
748 #ifdef CONFIG_HPET
749 static __init int late_hpet_init(void)
750 {
751 struct hpet_data hd;
752 unsigned int ntimer;
753
754 if (!vxtime.hpet_address)
755 return -1;
756
757 memset(&hd, 0, sizeof (hd));
758
759 ntimer = hpet_readl(HPET_ID);
760 ntimer = (ntimer & HPET_ID_NUMBER) >> HPET_ID_NUMBER_SHIFT;
761 ntimer++;
762
763 /*
764 * Register with driver.
765 * Timer0 and Timer1 is used by platform.
766 */
767 hd.hd_phys_address = vxtime.hpet_address;
768 hd.hd_address = (void *)fix_to_virt(FIX_HPET_BASE);
769 hd.hd_nirqs = ntimer;
770 hd.hd_flags = HPET_DATA_PLATFORM;
771 hpet_reserve_timer(&hd, 0);
772 #ifdef CONFIG_HPET_EMULATE_RTC
773 hpet_reserve_timer(&hd, 1);
774 #endif
775 hd.hd_irq[0] = HPET_LEGACY_8254;
776 hd.hd_irq[1] = HPET_LEGACY_RTC;
777 if (ntimer > 2) {
778 struct hpet *hpet;
779 struct hpet_timer *timer;
780 int i;
781
782 hpet = (struct hpet *) fix_to_virt(FIX_HPET_BASE);
783
784 for (i = 2, timer = &hpet->hpet_timers[2]; i < ntimer;
785 timer++, i++)
786 hd.hd_irq[i] = (timer->hpet_config &
787 Tn_INT_ROUTE_CNF_MASK) >>
788 Tn_INT_ROUTE_CNF_SHIFT;
789
790 }
791
792 hpet_alloc(&hd);
793 return 0;
794 }
795 fs_initcall(late_hpet_init);
796 #endif
797
798 static int hpet_timer_stop_set_go(unsigned long tick)
799 {
800 unsigned int cfg;
801
802 /*
803 * Stop the timers and reset the main counter.
804 */
805
806 cfg = hpet_readl(HPET_CFG);
807 cfg &= ~(HPET_CFG_ENABLE | HPET_CFG_LEGACY);
808 hpet_writel(cfg, HPET_CFG);
809 hpet_writel(0, HPET_COUNTER);
810 hpet_writel(0, HPET_COUNTER + 4);
811
812 /*
813 * Set up timer 0, as periodic with first interrupt to happen at hpet_tick,
814 * and period also hpet_tick.
815 */
816 if (hpet_use_timer) {
817 hpet_writel(HPET_TN_ENABLE | HPET_TN_PERIODIC | HPET_TN_SETVAL |
818 HPET_TN_32BIT, HPET_T0_CFG);
819 hpet_writel(hpet_tick, HPET_T0_CMP);
820 hpet_writel(hpet_tick, HPET_T0_CMP); /* AK: why twice? */
821 cfg |= HPET_CFG_LEGACY;
822 }
823 /*
824 * Go!
825 */
826
827 cfg |= HPET_CFG_ENABLE;
828 hpet_writel(cfg, HPET_CFG);
829
830 return 0;
831 }
832
833 static int hpet_init(void)
834 {
835 unsigned int id;
836
837 if (!vxtime.hpet_address)
838 return -1;
839 set_fixmap_nocache(FIX_HPET_BASE, vxtime.hpet_address);
840 __set_fixmap(VSYSCALL_HPET, vxtime.hpet_address, PAGE_KERNEL_VSYSCALL_NOCACHE);
841
842 /*
843 * Read the period, compute tick and quotient.
844 */
845
846 id = hpet_readl(HPET_ID);
847
848 if (!(id & HPET_ID_VENDOR) || !(id & HPET_ID_NUMBER))
849 return -1;
850
851 hpet_period = hpet_readl(HPET_PERIOD);
852 if (hpet_period < 100000 || hpet_period > 100000000)
853 return -1;
854
855 hpet_tick = (1000000000L * (USEC_PER_SEC / HZ) + hpet_period / 2) /
856 hpet_period;
857
858 hpet_use_timer = (id & HPET_ID_LEGSUP);
859
860 return hpet_timer_stop_set_go(hpet_tick);
861 }
862
863 static int hpet_reenable(void)
864 {
865 return hpet_timer_stop_set_go(hpet_tick);
866 }
867
868 void __init pit_init(void)
869 {
870 unsigned long flags;
871
872 spin_lock_irqsave(&i8253_lock, flags);
873 outb_p(0x34, 0x43); /* binary, mode 2, LSB/MSB, ch 0 */
874 outb_p(LATCH & 0xff, 0x40); /* LSB */
875 outb_p(LATCH >> 8, 0x40); /* MSB */
876 spin_unlock_irqrestore(&i8253_lock, flags);
877 }
878
879 int __init time_setup(char *str)
880 {
881 report_lost_ticks = 1;
882 return 1;
883 }
884
885 static struct irqaction irq0 = {
886 timer_interrupt, SA_INTERRUPT, CPU_MASK_NONE, "timer", NULL, NULL
887 };
888
889 extern void __init config_acpi_tables(void);
890
891 void __init time_init(void)
892 {
893 char *timename;
894
895 #ifdef HPET_HACK_ENABLE_DANGEROUS
896 if (!vxtime.hpet_address) {
897 printk(KERN_WARNING "time.c: WARNING: Enabling HPET base "
898 "manually!\n");
899 outl(0x800038a0, 0xcf8);
900 outl(0xff000001, 0xcfc);
901 outl(0x800038a0, 0xcf8);
902 vxtime.hpet_address = inl(0xcfc) & 0xfffffffe;
903 printk(KERN_WARNING "time.c: WARNING: Enabled HPET "
904 "at %#lx.\n", vxtime.hpet_address);
905 }
906 #endif
907 if (nohpet)
908 vxtime.hpet_address = 0;
909
910 xtime.tv_sec = get_cmos_time();
911 xtime.tv_nsec = 0;
912
913 set_normalized_timespec(&wall_to_monotonic,
914 -xtime.tv_sec, -xtime.tv_nsec);
915
916 if (!hpet_init())
917 vxtime_hz = (1000000000000000L + hpet_period / 2) /
918 hpet_period;
919
920 if (hpet_use_timer) {
921 cpu_khz = hpet_calibrate_tsc();
922 timename = "HPET";
923 #ifdef CONFIG_X86_PM_TIMER
924 } else if (pmtmr_ioport) {
925 vxtime_hz = PM_TIMER_FREQUENCY;
926 timename = "PM";
927 pit_init();
928 cpu_khz = pit_calibrate_tsc();
929 #endif
930 } else {
931 pit_init();
932 cpu_khz = pit_calibrate_tsc();
933 timename = "PIT";
934 }
935
936 printk(KERN_INFO "time.c: Using %ld.%06ld MHz %s timer.\n",
937 vxtime_hz / 1000000, vxtime_hz % 1000000, timename);
938 printk(KERN_INFO "time.c: Detected %d.%03d MHz processor.\n",
939 cpu_khz / 1000, cpu_khz % 1000);
940 vxtime.mode = VXTIME_TSC;
941 vxtime.quot = (1000000L << 32) / vxtime_hz;
942 vxtime.tsc_quot = (1000L << 32) / cpu_khz;
943 vxtime.hz = vxtime_hz;
944 rdtscll_sync(&vxtime.last_tsc);
945 setup_irq(0, &irq0);
946
947 set_cyc2ns_scale(cpu_khz / 1000);
948
949 #ifndef CONFIG_SMP
950 time_init_gtod();
951 #endif
952 }
953
954 /*
955 * Make an educated guess if the TSC is trustworthy and synchronized
956 * over all CPUs.
957 */
958 static __init int unsynchronized_tsc(void)
959 {
960 #ifdef CONFIG_SMP
961 if (oem_force_hpet_timer())
962 return 1;
963 /* Intel systems are normally all synchronized. Exceptions
964 are handled in the OEM check above. */
965 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
966 return 0;
967 /* All in a single socket - should be synchronized */
968 if (cpus_weight(cpu_core_map[0]) == num_online_cpus())
969 return 0;
970 #endif
971 /* Assume multi socket systems are not synchronized */
972 return num_online_cpus() > 1;
973 }
974
975 /*
976 * Decide after all CPUs are booted what mode gettimeofday should use.
977 */
978 void __init time_init_gtod(void)
979 {
980 char *timetype;
981
982 if (unsynchronized_tsc())
983 notsc = 1;
984 if (vxtime.hpet_address && notsc) {
985 timetype = hpet_use_timer ? "HPET" : "PIT/HPET";
986 vxtime.last = hpet_readl(HPET_T0_CMP) - hpet_tick;
987 vxtime.mode = VXTIME_HPET;
988 do_gettimeoffset = do_gettimeoffset_hpet;
989 #ifdef CONFIG_X86_PM_TIMER
990 /* Using PM for gettimeofday is quite slow, but we have no other
991 choice because the TSC is too unreliable on some systems. */
992 } else if (pmtmr_ioport && !vxtime.hpet_address && notsc) {
993 timetype = "PM";
994 do_gettimeoffset = do_gettimeoffset_pm;
995 vxtime.mode = VXTIME_PMTMR;
996 sysctl_vsyscall = 0;
997 printk(KERN_INFO "Disabling vsyscall due to use of PM timer\n");
998 #endif
999 } else {
1000 timetype = hpet_use_timer ? "HPET/TSC" : "PIT/TSC";
1001 vxtime.mode = VXTIME_TSC;
1002 }
1003
1004 printk(KERN_INFO "time.c: Using %s based timekeeping.\n", timetype);
1005 }
1006
1007 __setup("report_lost_ticks", time_setup);
1008
1009 static long clock_cmos_diff;
1010 static unsigned long sleep_start;
1011
1012 static int timer_suspend(struct sys_device *dev, pm_message_t state)
1013 {
1014 /*
1015 * Estimate time zone so that set_time can update the clock
1016 */
1017 long cmos_time = get_cmos_time();
1018
1019 clock_cmos_diff = -cmos_time;
1020 clock_cmos_diff += get_seconds();
1021 sleep_start = cmos_time;
1022 return 0;
1023 }
1024
1025 static int timer_resume(struct sys_device *dev)
1026 {
1027 unsigned long flags;
1028 unsigned long sec;
1029 unsigned long ctime = get_cmos_time();
1030 unsigned long sleep_length = (ctime - sleep_start) * HZ;
1031
1032 if (vxtime.hpet_address)
1033 hpet_reenable();
1034 else
1035 i8254_timer_resume();
1036
1037 sec = ctime + clock_cmos_diff;
1038 write_seqlock_irqsave(&xtime_lock,flags);
1039 xtime.tv_sec = sec;
1040 xtime.tv_nsec = 0;
1041 write_sequnlock_irqrestore(&xtime_lock,flags);
1042 jiffies += sleep_length;
1043 wall_jiffies += sleep_length;
1044 return 0;
1045 }
1046
1047 static struct sysdev_class timer_sysclass = {
1048 .resume = timer_resume,
1049 .suspend = timer_suspend,
1050 set_kset_name("timer"),
1051 };
1052
1053
1054 /* XXX this driverfs stuff should probably go elsewhere later -john */
1055 static struct sys_device device_timer = {
1056 .id = 0,
1057 .cls = &timer_sysclass,
1058 };
1059
1060 static int time_init_device(void)
1061 {
1062 int error = sysdev_class_register(&timer_sysclass);
1063 if (!error)
1064 error = sysdev_register(&device_timer);
1065 return error;
1066 }
1067
1068 device_initcall(time_init_device);
1069
1070 #ifdef CONFIG_HPET_EMULATE_RTC
1071 /* HPET in LegacyReplacement Mode eats up RTC interrupt line. When, HPET
1072 * is enabled, we support RTC interrupt functionality in software.
1073 * RTC has 3 kinds of interrupts:
1074 * 1) Update Interrupt - generate an interrupt, every sec, when RTC clock
1075 * is updated
1076 * 2) Alarm Interrupt - generate an interrupt at a specific time of day
1077 * 3) Periodic Interrupt - generate periodic interrupt, with frequencies
1078 * 2Hz-8192Hz (2Hz-64Hz for non-root user) (all freqs in powers of 2)
1079 * (1) and (2) above are implemented using polling at a frequency of
1080 * 64 Hz. The exact frequency is a tradeoff between accuracy and interrupt
1081 * overhead. (DEFAULT_RTC_INT_FREQ)
1082 * For (3), we use interrupts at 64Hz or user specified periodic
1083 * frequency, whichever is higher.
1084 */
1085 #include <linux/rtc.h>
1086
1087 extern irqreturn_t rtc_interrupt(int irq, void *dev_id, struct pt_regs *regs);
1088
1089 #define DEFAULT_RTC_INT_FREQ 64
1090 #define RTC_NUM_INTS 1
1091
1092 static unsigned long UIE_on;
1093 static unsigned long prev_update_sec;
1094
1095 static unsigned long AIE_on;
1096 static struct rtc_time alarm_time;
1097
1098 static unsigned long PIE_on;
1099 static unsigned long PIE_freq = DEFAULT_RTC_INT_FREQ;
1100 static unsigned long PIE_count;
1101
1102 static unsigned long hpet_rtc_int_freq; /* RTC interrupt frequency */
1103
1104 int is_hpet_enabled(void)
1105 {
1106 return vxtime.hpet_address != 0;
1107 }
1108
1109 /*
1110 * Timer 1 for RTC, we do not use periodic interrupt feature,
1111 * even if HPET supports periodic interrupts on Timer 1.
1112 * The reason being, to set up a periodic interrupt in HPET, we need to
1113 * stop the main counter. And if we do that everytime someone diables/enables
1114 * RTC, we will have adverse effect on main kernel timer running on Timer 0.
1115 * So, for the time being, simulate the periodic interrupt in software.
1116 *
1117 * hpet_rtc_timer_init() is called for the first time and during subsequent
1118 * interuppts reinit happens through hpet_rtc_timer_reinit().
1119 */
1120 int hpet_rtc_timer_init(void)
1121 {
1122 unsigned int cfg, cnt;
1123 unsigned long flags;
1124
1125 if (!is_hpet_enabled())
1126 return 0;
1127 /*
1128 * Set the counter 1 and enable the interrupts.
1129 */
1130 if (PIE_on && (PIE_freq > DEFAULT_RTC_INT_FREQ))
1131 hpet_rtc_int_freq = PIE_freq;
1132 else
1133 hpet_rtc_int_freq = DEFAULT_RTC_INT_FREQ;
1134
1135 local_irq_save(flags);
1136 cnt = hpet_readl(HPET_COUNTER);
1137 cnt += ((hpet_tick*HZ)/hpet_rtc_int_freq);
1138 hpet_writel(cnt, HPET_T1_CMP);
1139 local_irq_restore(flags);
1140
1141 cfg = hpet_readl(HPET_T1_CFG);
1142 cfg |= HPET_TN_ENABLE | HPET_TN_SETVAL | HPET_TN_32BIT;
1143 hpet_writel(cfg, HPET_T1_CFG);
1144
1145 return 1;
1146 }
1147
1148 static void hpet_rtc_timer_reinit(void)
1149 {
1150 unsigned int cfg, cnt;
1151
1152 if (!(PIE_on | AIE_on | UIE_on))
1153 return;
1154
1155 if (PIE_on && (PIE_freq > DEFAULT_RTC_INT_FREQ))
1156 hpet_rtc_int_freq = PIE_freq;
1157 else
1158 hpet_rtc_int_freq = DEFAULT_RTC_INT_FREQ;
1159
1160 /* It is more accurate to use the comparator value than current count.*/
1161 cnt = hpet_readl(HPET_T1_CMP);
1162 cnt += hpet_tick*HZ/hpet_rtc_int_freq;
1163 hpet_writel(cnt, HPET_T1_CMP);
1164
1165 cfg = hpet_readl(HPET_T1_CFG);
1166 cfg |= HPET_TN_ENABLE | HPET_TN_SETVAL | HPET_TN_32BIT;
1167 hpet_writel(cfg, HPET_T1_CFG);
1168
1169 return;
1170 }
1171
1172 /*
1173 * The functions below are called from rtc driver.
1174 * Return 0 if HPET is not being used.
1175 * Otherwise do the necessary changes and return 1.
1176 */
1177 int hpet_mask_rtc_irq_bit(unsigned long bit_mask)
1178 {
1179 if (!is_hpet_enabled())
1180 return 0;
1181
1182 if (bit_mask & RTC_UIE)
1183 UIE_on = 0;
1184 if (bit_mask & RTC_PIE)
1185 PIE_on = 0;
1186 if (bit_mask & RTC_AIE)
1187 AIE_on = 0;
1188
1189 return 1;
1190 }
1191
1192 int hpet_set_rtc_irq_bit(unsigned long bit_mask)
1193 {
1194 int timer_init_reqd = 0;
1195
1196 if (!is_hpet_enabled())
1197 return 0;
1198
1199 if (!(PIE_on | AIE_on | UIE_on))
1200 timer_init_reqd = 1;
1201
1202 if (bit_mask & RTC_UIE) {
1203 UIE_on = 1;
1204 }
1205 if (bit_mask & RTC_PIE) {
1206 PIE_on = 1;
1207 PIE_count = 0;
1208 }
1209 if (bit_mask & RTC_AIE) {
1210 AIE_on = 1;
1211 }
1212
1213 if (timer_init_reqd)
1214 hpet_rtc_timer_init();
1215
1216 return 1;
1217 }
1218
1219 int hpet_set_alarm_time(unsigned char hrs, unsigned char min, unsigned char sec)
1220 {
1221 if (!is_hpet_enabled())
1222 return 0;
1223
1224 alarm_time.tm_hour = hrs;
1225 alarm_time.tm_min = min;
1226 alarm_time.tm_sec = sec;
1227
1228 return 1;
1229 }
1230
1231 int hpet_set_periodic_freq(unsigned long freq)
1232 {
1233 if (!is_hpet_enabled())
1234 return 0;
1235
1236 PIE_freq = freq;
1237 PIE_count = 0;
1238
1239 return 1;
1240 }
1241
1242 int hpet_rtc_dropped_irq(void)
1243 {
1244 if (!is_hpet_enabled())
1245 return 0;
1246
1247 return 1;
1248 }
1249
1250 irqreturn_t hpet_rtc_interrupt(int irq, void *dev_id, struct pt_regs *regs)
1251 {
1252 struct rtc_time curr_time;
1253 unsigned long rtc_int_flag = 0;
1254 int call_rtc_interrupt = 0;
1255
1256 hpet_rtc_timer_reinit();
1257
1258 if (UIE_on | AIE_on) {
1259 rtc_get_rtc_time(&curr_time);
1260 }
1261 if (UIE_on) {
1262 if (curr_time.tm_sec != prev_update_sec) {
1263 /* Set update int info, call real rtc int routine */
1264 call_rtc_interrupt = 1;
1265 rtc_int_flag = RTC_UF;
1266 prev_update_sec = curr_time.tm_sec;
1267 }
1268 }
1269 if (PIE_on) {
1270 PIE_count++;
1271 if (PIE_count >= hpet_rtc_int_freq/PIE_freq) {
1272 /* Set periodic int info, call real rtc int routine */
1273 call_rtc_interrupt = 1;
1274 rtc_int_flag |= RTC_PF;
1275 PIE_count = 0;
1276 }
1277 }
1278 if (AIE_on) {
1279 if ((curr_time.tm_sec == alarm_time.tm_sec) &&
1280 (curr_time.tm_min == alarm_time.tm_min) &&
1281 (curr_time.tm_hour == alarm_time.tm_hour)) {
1282 /* Set alarm int info, call real rtc int routine */
1283 call_rtc_interrupt = 1;
1284 rtc_int_flag |= RTC_AF;
1285 }
1286 }
1287 if (call_rtc_interrupt) {
1288 rtc_int_flag |= (RTC_IRQF | (RTC_NUM_INTS << 8));
1289 rtc_interrupt(rtc_int_flag, dev_id, regs);
1290 }
1291 return IRQ_HANDLED;
1292 }
1293 #endif
1294
1295
1296
1297 static int __init nohpet_setup(char *s)
1298 {
1299 nohpet = 1;
1300 return 0;
1301 }
1302
1303 __setup("nohpet", nohpet_setup);
1304
1305
1306 static int __init notsc_setup(char *s)
1307 {
1308 notsc = 1;
1309 return 0;
1310 }
1311
1312 __setup("notsc", notsc_setup);
1313
1314