1#include <linux/kernel.h>
  2#include <linux/sched.h>
  3#include <linux/init.h>
  4#include <linux/module.h>
  5#include <linux/timer.h>
  6#include <linux/acpi_pmtmr.h>
  7#include <linux/cpufreq.h>
  8#include <linux/delay.h>
  9#include <linux/clocksource.h>
 10#include <linux/percpu.h>
 11#include <linux/timex.h>
 
 12
 13#include <asm/hpet.h>
 14#include <asm/timer.h>
 15#include <asm/vgtod.h>
 16#include <asm/time.h>
 17#include <asm/delay.h>
 18#include <asm/hypervisor.h>
 19#include <asm/nmi.h>
 20#include <asm/x86_init.h>
 
 21
 22unsigned int __read_mostly cpu_khz;	/* TSC clocks / usec, not used here */
 23EXPORT_SYMBOL(cpu_khz);
 24
 25unsigned int __read_mostly tsc_khz;
 26EXPORT_SYMBOL(tsc_khz);
 27
 28/*
 29 * TSC can be unstable due to cpufreq or due to unsynced TSCs
 30 */
 31static int __read_mostly tsc_unstable;
 32
 33/* native_sched_clock() is called before tsc_init(), so
 34   we must start with the TSC soft disabled to prevent
 35   erroneous rdtsc usage on !cpu_has_tsc processors */
 36static int __read_mostly tsc_disabled = -1;
 37
 38static int tsc_clocksource_reliable;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 39/*
 40 * Scheduler clock - returns current time in nanosec units.
 41 */
 42u64 native_sched_clock(void)
 43{
 44	u64 this_offset;
 
 
 
 
 
 45
 46	/*
 47	 * Fall back to jiffies if there's no TSC available:
 48	 * ( But note that we still use it if the TSC is marked
 49	 *   unstable. We do this because unlike Time Of Day,
 50	 *   the scheduler clock tolerates small errors and it's
 51	 *   very important for it to be as fast as the platform
 52	 *   can achieve it. )
 53	 */
 54	if (unlikely(tsc_disabled)) {
 55		/* No locking but a rare wrong value is not a big deal: */
 56		return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 57	}
 58
 59	/* read the Time Stamp Counter: */
 60	rdtscll(this_offset);
 
 61
 62	/* return the value in ns */
 63	return __cycles_2_ns(this_offset);
 
 
 
 
 64}
 65
 66/* We need to define a real function for sched_clock, to override the
 67   weak default version */
 68#ifdef CONFIG_PARAVIRT
 69unsigned long long sched_clock(void)
 70{
 71	return paravirt_sched_clock();
 72}
 73#else
 74unsigned long long
 75sched_clock(void) __attribute__((alias("native_sched_clock")));
 76#endif
 77
 78int check_tsc_unstable(void)
 79{
 80	return tsc_unstable;
 81}
 82EXPORT_SYMBOL_GPL(check_tsc_unstable);
 83
 
 
 
 
 
 
 84#ifdef CONFIG_X86_TSC
 85int __init notsc_setup(char *str)
 86{
 87	printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
 88			"cannot disable TSC completely.\n");
 89	tsc_disabled = 1;
 90	return 1;
 91}
 92#else
 93/*
 94 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
 95 * in cpu/common.c
 96 */
 97int __init notsc_setup(char *str)
 98{
 99	setup_clear_cpu_cap(X86_FEATURE_TSC);
100	return 1;
101}
102#endif
103
104__setup("notsc", notsc_setup);
105
106static int no_sched_irq_time;
107
108static int __init tsc_setup(char *str)
109{
110	if (!strcmp(str, "reliable"))
111		tsc_clocksource_reliable = 1;
112	if (!strncmp(str, "noirqtime", 9))
113		no_sched_irq_time = 1;
114	return 1;
115}
116
117__setup("tsc=", tsc_setup);
118
119#define MAX_RETRIES     5
120#define SMI_TRESHOLD    50000
121
122/*
123 * Read TSC and the reference counters. Take care of SMI disturbance
124 */
125static u64 tsc_read_refs(u64 *p, int hpet)
126{
127	u64 t1, t2;
128	int i;
129
130	for (i = 0; i < MAX_RETRIES; i++) {
131		t1 = get_cycles();
132		if (hpet)
133			*p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
134		else
135			*p = acpi_pm_read_early();
136		t2 = get_cycles();
137		if ((t2 - t1) < SMI_TRESHOLD)
138			return t2;
139	}
140	return ULLONG_MAX;
141}
142
143/*
144 * Calculate the TSC frequency from HPET reference
145 */
146static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
147{
148	u64 tmp;
149
150	if (hpet2 < hpet1)
151		hpet2 += 0x100000000ULL;
152	hpet2 -= hpet1;
153	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
154	do_div(tmp, 1000000);
155	do_div(deltatsc, tmp);
156
157	return (unsigned long) deltatsc;
158}
159
160/*
161 * Calculate the TSC frequency from PMTimer reference
162 */
163static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
164{
165	u64 tmp;
166
167	if (!pm1 && !pm2)
168		return ULONG_MAX;
169
170	if (pm2 < pm1)
171		pm2 += (u64)ACPI_PM_OVRRUN;
172	pm2 -= pm1;
173	tmp = pm2 * 1000000000LL;
174	do_div(tmp, PMTMR_TICKS_PER_SEC);
175	do_div(deltatsc, tmp);
176
177	return (unsigned long) deltatsc;
178}
179
180#define CAL_MS		10
181#define CAL_LATCH	(CLOCK_TICK_RATE / (1000 / CAL_MS))
182#define CAL_PIT_LOOPS	1000
183
184#define CAL2_MS		50
185#define CAL2_LATCH	(CLOCK_TICK_RATE / (1000 / CAL2_MS))
186#define CAL2_PIT_LOOPS	5000
187
188
189/*
190 * Try to calibrate the TSC against the Programmable
191 * Interrupt Timer and return the frequency of the TSC
192 * in kHz.
193 *
194 * Return ULONG_MAX on failure to calibrate.
195 */
196static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
197{
198	u64 tsc, t1, t2, delta;
199	unsigned long tscmin, tscmax;
200	int pitcnt;
201
202	/* Set the Gate high, disable speaker */
203	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
204
205	/*
206	 * Setup CTC channel 2* for mode 0, (interrupt on terminal
207	 * count mode), binary count. Set the latch register to 50ms
208	 * (LSB then MSB) to begin countdown.
209	 */
210	outb(0xb0, 0x43);
211	outb(latch & 0xff, 0x42);
212	outb(latch >> 8, 0x42);
213
214	tsc = t1 = t2 = get_cycles();
215
216	pitcnt = 0;
217	tscmax = 0;
218	tscmin = ULONG_MAX;
219	while ((inb(0x61) & 0x20) == 0) {
220		t2 = get_cycles();
221		delta = t2 - tsc;
222		tsc = t2;
223		if ((unsigned long) delta < tscmin)
224			tscmin = (unsigned int) delta;
225		if ((unsigned long) delta > tscmax)
226			tscmax = (unsigned int) delta;
227		pitcnt++;
228	}
229
230	/*
231	 * Sanity checks:
232	 *
233	 * If we were not able to read the PIT more than loopmin
234	 * times, then we have been hit by a massive SMI
235	 *
236	 * If the maximum is 10 times larger than the minimum,
237	 * then we got hit by an SMI as well.
238	 */
239	if (pitcnt < loopmin || tscmax > 10 * tscmin)
240		return ULONG_MAX;
241
242	/* Calculate the PIT value */
243	delta = t2 - t1;
244	do_div(delta, ms);
245	return delta;
246}
247
248/*
249 * This reads the current MSB of the PIT counter, and
250 * checks if we are running on sufficiently fast and
251 * non-virtualized hardware.
252 *
253 * Our expectations are:
254 *
255 *  - the PIT is running at roughly 1.19MHz
256 *
257 *  - each IO is going to take about 1us on real hardware,
258 *    but we allow it to be much faster (by a factor of 10) or
259 *    _slightly_ slower (ie we allow up to a 2us read+counter
260 *    update - anything else implies a unacceptably slow CPU
261 *    or PIT for the fast calibration to work.
262 *
263 *  - with 256 PIT ticks to read the value, we have 214us to
264 *    see the same MSB (and overhead like doing a single TSC
265 *    read per MSB value etc).
266 *
267 *  - We're doing 2 reads per loop (LSB, MSB), and we expect
268 *    them each to take about a microsecond on real hardware.
269 *    So we expect a count value of around 100. But we'll be
270 *    generous, and accept anything over 50.
271 *
272 *  - if the PIT is stuck, and we see *many* more reads, we
273 *    return early (and the next caller of pit_expect_msb()
274 *    then consider it a failure when they don't see the
275 *    next expected value).
276 *
277 * These expectations mean that we know that we have seen the
278 * transition from one expected value to another with a fairly
279 * high accuracy, and we didn't miss any events. We can thus
280 * use the TSC value at the transitions to calculate a pretty
281 * good value for the TSC frequencty.
282 */
283static inline int pit_verify_msb(unsigned char val)
284{
285	/* Ignore LSB */
286	inb(0x42);
287	return inb(0x42) == val;
288}
289
290static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
291{
292	int count;
293	u64 tsc = 0;
294
295	for (count = 0; count < 50000; count++) {
296		if (!pit_verify_msb(val))
297			break;
 
298		tsc = get_cycles();
299	}
300	*deltap = get_cycles() - tsc;
301	*tscp = tsc;
302
303	/*
304	 * We require _some_ success, but the quality control
305	 * will be based on the error terms on the TSC values.
306	 */
307	return count > 5;
308}
309
310/*
311 * How many MSB values do we want to see? We aim for
312 * a maximum error rate of 500ppm (in practice the
313 * real error is much smaller), but refuse to spend
314 * more than 25ms on it.
315 */
316#define MAX_QUICK_PIT_MS 25
317#define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
318
319static unsigned long quick_pit_calibrate(void)
320{
321	int i;
322	u64 tsc, delta;
323	unsigned long d1, d2;
324
325	/* Set the Gate high, disable speaker */
326	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
327
328	/*
329	 * Counter 2, mode 0 (one-shot), binary count
330	 *
331	 * NOTE! Mode 2 decrements by two (and then the
332	 * output is flipped each time, giving the same
333	 * final output frequency as a decrement-by-one),
334	 * so mode 0 is much better when looking at the
335	 * individual counts.
336	 */
337	outb(0xb0, 0x43);
338
339	/* Start at 0xffff */
340	outb(0xff, 0x42);
341	outb(0xff, 0x42);
342
343	/*
344	 * The PIT starts counting at the next edge, so we
345	 * need to delay for a microsecond. The easiest way
346	 * to do that is to just read back the 16-bit counter
347	 * once from the PIT.
348	 */
349	pit_verify_msb(0);
350
351	if (pit_expect_msb(0xff, &tsc, &d1)) {
352		for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
353			if (!pit_expect_msb(0xff-i, &delta, &d2))
354				break;
355
 
 
 
 
 
 
 
 
 
 
356			/*
357			 * Iterate until the error is less than 500 ppm
358			 */
359			delta -= tsc;
360			if (d1+d2 >= delta >> 11)
361				continue;
362
363			/*
364			 * Check the PIT one more time to verify that
365			 * all TSC reads were stable wrt the PIT.
366			 *
367			 * This also guarantees serialization of the
368			 * last cycle read ('d2') in pit_expect_msb.
369			 */
370			if (!pit_verify_msb(0xfe - i))
371				break;
372			goto success;
373		}
374	}
375	printk("Fast TSC calibration failed\n");
376	return 0;
377
378success:
379	/*
380	 * Ok, if we get here, then we've seen the
381	 * MSB of the PIT decrement 'i' times, and the
382	 * error has shrunk to less than 500 ppm.
383	 *
384	 * As a result, we can depend on there not being
385	 * any odd delays anywhere, and the TSC reads are
386	 * reliable (within the error). We also adjust the
387	 * delta to the middle of the error bars, just
388	 * because it looks nicer.
389	 *
390	 * kHz = ticks / time-in-seconds / 1000;
391	 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
392	 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
393	 */
394	delta += (long)(d2 - d1)/2;
395	delta *= PIT_TICK_RATE;
396	do_div(delta, i*256*1000);
397	printk("Fast TSC calibration using PIT\n");
398	return delta;
399}
400
401/**
402 * native_calibrate_tsc - calibrate the tsc on boot
403 */
404unsigned long native_calibrate_tsc(void)
405{
406	u64 tsc1, tsc2, delta, ref1, ref2;
407	unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
408	unsigned long flags, latch, ms, fast_calibrate;
409	int hpet = is_hpet_enabled(), i, loopmin;
410
 
 
 
 
 
 
 
411	local_irq_save(flags);
412	fast_calibrate = quick_pit_calibrate();
413	local_irq_restore(flags);
414	if (fast_calibrate)
415		return fast_calibrate;
416
417	/*
418	 * Run 5 calibration loops to get the lowest frequency value
419	 * (the best estimate). We use two different calibration modes
420	 * here:
421	 *
422	 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
423	 * load a timeout of 50ms. We read the time right after we
424	 * started the timer and wait until the PIT count down reaches
425	 * zero. In each wait loop iteration we read the TSC and check
426	 * the delta to the previous read. We keep track of the min
427	 * and max values of that delta. The delta is mostly defined
428	 * by the IO time of the PIT access, so we can detect when a
429	 * SMI/SMM disturbance happened between the two reads. If the
430	 * maximum time is significantly larger than the minimum time,
431	 * then we discard the result and have another try.
432	 *
433	 * 2) Reference counter. If available we use the HPET or the
434	 * PMTIMER as a reference to check the sanity of that value.
435	 * We use separate TSC readouts and check inside of the
436	 * reference read for a SMI/SMM disturbance. We dicard
437	 * disturbed values here as well. We do that around the PIT
438	 * calibration delay loop as we have to wait for a certain
439	 * amount of time anyway.
440	 */
441
442	/* Preset PIT loop values */
443	latch = CAL_LATCH;
444	ms = CAL_MS;
445	loopmin = CAL_PIT_LOOPS;
446
447	for (i = 0; i < 3; i++) {
448		unsigned long tsc_pit_khz;
449
450		/*
451		 * Read the start value and the reference count of
452		 * hpet/pmtimer when available. Then do the PIT
453		 * calibration, which will take at least 50ms, and
454		 * read the end value.
455		 */
456		local_irq_save(flags);
457		tsc1 = tsc_read_refs(&ref1, hpet);
458		tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
459		tsc2 = tsc_read_refs(&ref2, hpet);
460		local_irq_restore(flags);
461
462		/* Pick the lowest PIT TSC calibration so far */
463		tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
464
465		/* hpet or pmtimer available ? */
466		if (ref1 == ref2)
467			continue;
468
469		/* Check, whether the sampling was disturbed by an SMI */
470		if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
471			continue;
472
473		tsc2 = (tsc2 - tsc1) * 1000000LL;
474		if (hpet)
475			tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
476		else
477			tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
478
479		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
480
481		/* Check the reference deviation */
482		delta = ((u64) tsc_pit_min) * 100;
483		do_div(delta, tsc_ref_min);
484
485		/*
486		 * If both calibration results are inside a 10% window
487		 * then we can be sure, that the calibration
488		 * succeeded. We break out of the loop right away. We
489		 * use the reference value, as it is more precise.
490		 */
491		if (delta >= 90 && delta <= 110) {
492			printk(KERN_INFO
493			       "TSC: PIT calibration matches %s. %d loops\n",
494			       hpet ? "HPET" : "PMTIMER", i + 1);
495			return tsc_ref_min;
496		}
497
498		/*
499		 * Check whether PIT failed more than once. This
500		 * happens in virtualized environments. We need to
501		 * give the virtual PC a slightly longer timeframe for
502		 * the HPET/PMTIMER to make the result precise.
503		 */
504		if (i == 1 && tsc_pit_min == ULONG_MAX) {
505			latch = CAL2_LATCH;
506			ms = CAL2_MS;
507			loopmin = CAL2_PIT_LOOPS;
508		}
509	}
510
511	/*
512	 * Now check the results.
513	 */
514	if (tsc_pit_min == ULONG_MAX) {
515		/* PIT gave no useful value */
516		printk(KERN_WARNING "TSC: Unable to calibrate against PIT\n");
517
518		/* We don't have an alternative source, disable TSC */
519		if (!hpet && !ref1 && !ref2) {
520			printk("TSC: No reference (HPET/PMTIMER) available\n");
521			return 0;
522		}
523
524		/* The alternative source failed as well, disable TSC */
525		if (tsc_ref_min == ULONG_MAX) {
526			printk(KERN_WARNING "TSC: HPET/PMTIMER calibration "
527			       "failed.\n");
528			return 0;
529		}
530
531		/* Use the alternative source */
532		printk(KERN_INFO "TSC: using %s reference calibration\n",
533		       hpet ? "HPET" : "PMTIMER");
534
535		return tsc_ref_min;
536	}
537
538	/* We don't have an alternative source, use the PIT calibration value */
539	if (!hpet && !ref1 && !ref2) {
540		printk(KERN_INFO "TSC: Using PIT calibration value\n");
541		return tsc_pit_min;
542	}
543
544	/* The alternative source failed, use the PIT calibration value */
545	if (tsc_ref_min == ULONG_MAX) {
546		printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. "
547		       "Using PIT calibration\n");
548		return tsc_pit_min;
549	}
550
551	/*
552	 * The calibration values differ too much. In doubt, we use
553	 * the PIT value as we know that there are PMTIMERs around
554	 * running at double speed. At least we let the user know:
555	 */
556	printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n",
557	       hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
558	printk(KERN_INFO "TSC: Using PIT calibration value\n");
559	return tsc_pit_min;
560}
561
562int recalibrate_cpu_khz(void)
563{
564#ifndef CONFIG_SMP
565	unsigned long cpu_khz_old = cpu_khz;
566
567	if (cpu_has_tsc) {
568		tsc_khz = x86_platform.calibrate_tsc();
569		cpu_khz = tsc_khz;
570		cpu_data(0).loops_per_jiffy =
571			cpufreq_scale(cpu_data(0).loops_per_jiffy,
572					cpu_khz_old, cpu_khz);
573		return 0;
574	} else
575		return -ENODEV;
576#else
577	return -ENODEV;
578#endif
579}
580
581EXPORT_SYMBOL(recalibrate_cpu_khz);
582
583
584/* Accelerators for sched_clock()
585 * convert from cycles(64bits) => nanoseconds (64bits)
586 *  basic equation:
587 *              ns = cycles / (freq / ns_per_sec)
588 *              ns = cycles * (ns_per_sec / freq)
589 *              ns = cycles * (10^9 / (cpu_khz * 10^3))
590 *              ns = cycles * (10^6 / cpu_khz)
591 *
592 *      Then we use scaling math (suggested by george@mvista.com) to get:
593 *              ns = cycles * (10^6 * SC / cpu_khz) / SC
594 *              ns = cycles * cyc2ns_scale / SC
595 *
596 *      And since SC is a constant power of two, we can convert the div
597 *  into a shift.
598 *
599 *  We can use khz divisor instead of mhz to keep a better precision, since
600 *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
601 *  (mathieu.desnoyers@polymtl.ca)
602 *
603 *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
604 */
605
606DEFINE_PER_CPU(unsigned long, cyc2ns);
607DEFINE_PER_CPU(unsigned long long, cyc2ns_offset);
608
609static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
610{
611	unsigned long long tsc_now, ns_now, *offset;
612	unsigned long flags, *scale;
613
614	local_irq_save(flags);
615	sched_clock_idle_sleep_event();
616
617	scale = &per_cpu(cyc2ns, cpu);
618	offset = &per_cpu(cyc2ns_offset, cpu);
619
620	rdtscll(tsc_now);
621	ns_now = __cycles_2_ns(tsc_now);
622
623	if (cpu_khz) {
624		*scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
625		*offset = ns_now - (tsc_now * *scale >> CYC2NS_SCALE_FACTOR);
626	}
627
628	sched_clock_idle_wakeup_event(0);
629	local_irq_restore(flags);
630}
631
632static unsigned long long cyc2ns_suspend;
633
634void save_sched_clock_state(void)
635{
636	if (!sched_clock_stable)
637		return;
638
639	cyc2ns_suspend = sched_clock();
640}
641
642/*
643 * Even on processors with invariant TSC, TSC gets reset in some the
644 * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to
645 * arbitrary value (still sync'd across cpu's) during resume from such sleep
646 * states. To cope up with this, recompute the cyc2ns_offset for each cpu so
647 * that sched_clock() continues from the point where it was left off during
648 * suspend.
649 */
650void restore_sched_clock_state(void)
651{
652	unsigned long long offset;
653	unsigned long flags;
654	int cpu;
655
656	if (!sched_clock_stable)
657		return;
658
659	local_irq_save(flags);
660
661	__this_cpu_write(cyc2ns_offset, 0);
 
 
 
 
 
 
 
 
662	offset = cyc2ns_suspend - sched_clock();
663
664	for_each_possible_cpu(cpu)
665		per_cpu(cyc2ns_offset, cpu) = offset;
 
 
666
667	local_irq_restore(flags);
668}
669
670#ifdef CONFIG_CPU_FREQ
671
672/* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
673 * changes.
674 *
675 * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
676 * not that important because current Opteron setups do not support
677 * scaling on SMP anyroads.
678 *
679 * Should fix up last_tsc too. Currently gettimeofday in the
680 * first tick after the change will be slightly wrong.
681 */
682
683static unsigned int  ref_freq;
684static unsigned long loops_per_jiffy_ref;
685static unsigned long tsc_khz_ref;
686
687static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
688				void *data)
689{
690	struct cpufreq_freqs *freq = data;
691	unsigned long *lpj;
692
693	if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
694		return 0;
695
696	lpj = &boot_cpu_data.loops_per_jiffy;
697#ifdef CONFIG_SMP
698	if (!(freq->flags & CPUFREQ_CONST_LOOPS))
699		lpj = &cpu_data(freq->cpu).loops_per_jiffy;
700#endif
701
702	if (!ref_freq) {
703		ref_freq = freq->old;
704		loops_per_jiffy_ref = *lpj;
705		tsc_khz_ref = tsc_khz;
706	}
707	if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
708			(val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
709			(val == CPUFREQ_RESUMECHANGE)) {
710		*lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
711
712		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
713		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
714			mark_tsc_unstable("cpufreq changes");
715	}
716
717	set_cyc2ns_scale(tsc_khz, freq->cpu);
 
718
719	return 0;
720}
721
722static struct notifier_block time_cpufreq_notifier_block = {
723	.notifier_call  = time_cpufreq_notifier
724};
725
726static int __init cpufreq_tsc(void)
727{
728	if (!cpu_has_tsc)
729		return 0;
730	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
731		return 0;
732	cpufreq_register_notifier(&time_cpufreq_notifier_block,
733				CPUFREQ_TRANSITION_NOTIFIER);
734	return 0;
735}
736
737core_initcall(cpufreq_tsc);
738
739#endif /* CONFIG_CPU_FREQ */
740
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
741/* clocksource code */
742
743static struct clocksource clocksource_tsc;
744
745/*
746 * We compare the TSC to the cycle_last value in the clocksource
747 * structure to avoid a nasty time-warp. This can be observed in a
748 * very small window right after one CPU updated cycle_last under
749 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
750 * is smaller than the cycle_last reference value due to a TSC which
751 * is slighty behind. This delta is nowhere else observable, but in
752 * that case it results in a forward time jump in the range of hours
753 * due to the unsigned delta calculation of the time keeping core
754 * code, which is necessary to support wrapping clocksources like pm
755 * timer.
 
 
 
 
756 */
757static cycle_t read_tsc(struct clocksource *cs)
758{
759	cycle_t ret = (cycle_t)get_cycles();
760
761	return ret >= clocksource_tsc.cycle_last ?
762		ret : clocksource_tsc.cycle_last;
763}
764
765static void resume_tsc(struct clocksource *cs)
766{
767	clocksource_tsc.cycle_last = 0;
768}
769
 
 
 
770static struct clocksource clocksource_tsc = {
771	.name                   = "tsc",
772	.rating                 = 300,
773	.read                   = read_tsc,
774	.resume			= resume_tsc,
775	.mask                   = CLOCKSOURCE_MASK(64),
776	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
777				  CLOCK_SOURCE_MUST_VERIFY,
778#ifdef CONFIG_X86_64
779	.archdata               = { .vclock_mode = VCLOCK_TSC },
780#endif
781};
782
783void mark_tsc_unstable(char *reason)
784{
785	if (!tsc_unstable) {
786		tsc_unstable = 1;
787		sched_clock_stable = 0;
788		disable_sched_clock_irqtime();
789		printk(KERN_INFO "Marking TSC unstable due to %s\n", reason);
790		/* Change only the rating, when not registered */
791		if (clocksource_tsc.mult)
792			clocksource_mark_unstable(&clocksource_tsc);
793		else {
794			clocksource_tsc.flags |= CLOCK_SOURCE_UNSTABLE;
795			clocksource_tsc.rating = 0;
796		}
797	}
798}
799
800EXPORT_SYMBOL_GPL(mark_tsc_unstable);
801
802static void __init check_system_tsc_reliable(void)
803{
804#ifdef CONFIG_MGEODE_LX
805	/* RTSC counts during suspend */
 
806#define RTSC_SUSP 0x100
807	unsigned long res_low, res_high;
808
809	rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
810	/* Geode_LX - the OLPC CPU has a very reliable TSC */
811	if (res_low & RTSC_SUSP)
812		tsc_clocksource_reliable = 1;
 
813#endif
814	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
815		tsc_clocksource_reliable = 1;
816}
817
818/*
819 * Make an educated guess if the TSC is trustworthy and synchronized
820 * over all CPUs.
821 */
822__cpuinit int unsynchronized_tsc(void)
823{
824	if (!cpu_has_tsc || tsc_unstable)
825		return 1;
826
827#ifdef CONFIG_SMP
828	if (apic_is_clustered_box())
829		return 1;
830#endif
831
832	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
833		return 0;
834
835	if (tsc_clocksource_reliable)
836		return 0;
837	/*
838	 * Intel systems are normally all synchronized.
839	 * Exceptions must mark TSC as unstable:
840	 */
841	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
842		/* assume multi socket systems are not synchronized: */
843		if (num_possible_cpus() > 1)
844			return 1;
845	}
846
847	return 0;
848}
849
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
850
851static void tsc_refine_calibration_work(struct work_struct *work);
852static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);
853/**
854 * tsc_refine_calibration_work - Further refine tsc freq calibration
855 * @work - ignored.
856 *
857 * This functions uses delayed work over a period of a
858 * second to further refine the TSC freq value. Since this is
859 * timer based, instead of loop based, we don't block the boot
860 * process while this longer calibration is done.
861 *
862 * If there are any calibration anomalies (too many SMIs, etc),
863 * or the refined calibration is off by 1% of the fast early
864 * calibration, we throw out the new calibration and use the
865 * early calibration.
866 */
867static void tsc_refine_calibration_work(struct work_struct *work)
868{
869	static u64 tsc_start = -1, ref_start;
870	static int hpet;
871	u64 tsc_stop, ref_stop, delta;
872	unsigned long freq;
873
874	/* Don't bother refining TSC on unstable systems */
875	if (check_tsc_unstable())
876		goto out;
877
878	/*
879	 * Since the work is started early in boot, we may be
880	 * delayed the first time we expire. So set the workqueue
881	 * again once we know timers are working.
882	 */
883	if (tsc_start == -1) {
884		/*
885		 * Only set hpet once, to avoid mixing hardware
886		 * if the hpet becomes enabled later.
887		 */
888		hpet = is_hpet_enabled();
889		schedule_delayed_work(&tsc_irqwork, HZ);
890		tsc_start = tsc_read_refs(&ref_start, hpet);
891		return;
892	}
893
894	tsc_stop = tsc_read_refs(&ref_stop, hpet);
895
896	/* hpet or pmtimer available ? */
897	if (ref_start == ref_stop)
898		goto out;
899
900	/* Check, whether the sampling was disturbed by an SMI */
901	if (tsc_start == ULLONG_MAX || tsc_stop == ULLONG_MAX)
902		goto out;
903
904	delta = tsc_stop - tsc_start;
905	delta *= 1000000LL;
906	if (hpet)
907		freq = calc_hpet_ref(delta, ref_start, ref_stop);
908	else
909		freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
910
911	/* Make sure we're within 1% */
912	if (abs(tsc_khz - freq) > tsc_khz/100)
913		goto out;
914
915	tsc_khz = freq;
916	printk(KERN_INFO "Refined TSC clocksource calibration: "
917		"%lu.%03lu MHz.\n", (unsigned long)tsc_khz / 1000,
918					(unsigned long)tsc_khz % 1000);
919
920out:
 
 
921	clocksource_register_khz(&clocksource_tsc, tsc_khz);
922}
923
924
925static int __init init_tsc_clocksource(void)
926{
927	if (!cpu_has_tsc || tsc_disabled > 0 || !tsc_khz)
928		return 0;
929
930	if (tsc_clocksource_reliable)
931		clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
932	/* lower the rating if we already know its unstable: */
933	if (check_tsc_unstable()) {
934		clocksource_tsc.rating = 0;
935		clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
936	}
 
 
 
 
 
 
 
 
 
 
 
 
 
937	schedule_delayed_work(&tsc_irqwork, 0);
938	return 0;
939}
940/*
941 * We use device_initcall here, to ensure we run after the hpet
942 * is fully initialized, which may occur at fs_initcall time.
943 */
944device_initcall(init_tsc_clocksource);
945
946void __init tsc_init(void)
947{
948	u64 lpj;
949	int cpu;
950
951	x86_init.timers.tsc_pre_init();
952
953	if (!cpu_has_tsc)
954		return;
 
955
956	tsc_khz = x86_platform.calibrate_tsc();
957	cpu_khz = tsc_khz;
958
959	if (!tsc_khz) {
960		mark_tsc_unstable("could not calculate TSC khz");
 
961		return;
962	}
963
964	printk("Detected %lu.%03lu MHz processor.\n",
965			(unsigned long)cpu_khz / 1000,
966			(unsigned long)cpu_khz % 1000);
967
968	/*
969	 * Secondary CPUs do not run through tsc_init(), so set up
970	 * all the scale factors for all CPUs, assuming the same
971	 * speed as the bootup CPU. (cpufreq notifiers will fix this
972	 * up if their speed diverges)
973	 */
974	for_each_possible_cpu(cpu)
 
975		set_cyc2ns_scale(cpu_khz, cpu);
 
976
977	if (tsc_disabled > 0)
978		return;
979
980	/* now allow native_sched_clock() to use rdtsc */
 
981	tsc_disabled = 0;
 
982
983	if (!no_sched_irq_time)
984		enable_sched_clock_irqtime();
985
986	lpj = ((u64)tsc_khz * 1000);
987	do_div(lpj, HZ);
988	lpj_fine = lpj;
989
990	use_tsc_delay();
991
992	if (unsynchronized_tsc())
993		mark_tsc_unstable("TSCs unsynchronized");
994
995	check_system_tsc_reliable();
 
 
996}
997