1 #include <linux/sched.h>
2 #include <linux/clocksource.h>
3 #include <linux/workqueue.h>
4 #include <linux/cpufreq.h>
5 #include <linux/jiffies.h>
6 #include <linux/init.h>
8 #include <linux/percpu.h>
10 #include <asm/delay.h>
13 #include <asm/timer.h>
15 #include "mach_timer.h"
17 /* native_sched_clock() is called before tsc_init(), so
18 we must start with the TSC soft disabled to prevent
19 erroneous rdtsc usage on !cpu_has_tsc processors */
20 static int tsc_disabled = -1;
23 * On some systems the TSC frequency does not
24 * change with the cpu frequency. So we need
25 * an extra value to store the TSC freq
28 EXPORT_SYMBOL_GPL(tsc_khz);
31 static int __init tsc_setup(char *str)
33 printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
34 "cannot disable TSC completely.\n");
40 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
43 static int __init tsc_setup(char *str)
45 setup_clear_cpu_cap(X86_FEATURE_TSC);
50 __setup("notsc", tsc_setup);
53 * code to mark and check if the TSC is unstable
54 * due to cpufreq or due to unsynced TSCs
56 static int tsc_unstable;
58 int check_tsc_unstable(void)
62 EXPORT_SYMBOL_GPL(check_tsc_unstable);
64 /* Accelerators for sched_clock()
65 * convert from cycles(64bits) => nanoseconds (64bits)
67 * ns = cycles / (freq / ns_per_sec)
68 * ns = cycles * (ns_per_sec / freq)
69 * ns = cycles * (10^9 / (cpu_khz * 10^3))
70 * ns = cycles * (10^6 / cpu_khz)
72 * Then we use scaling math (suggested by george@mvista.com) to get:
73 * ns = cycles * (10^6 * SC / cpu_khz) / SC
74 * ns = cycles * cyc2ns_scale / SC
76 * And since SC is a constant power of two, we can convert the div
79 * We can use khz divisor instead of mhz to keep a better precision, since
80 * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
81 * (mathieu.desnoyers@polymtl.ca)
83 * -johnstul@us.ibm.com "math is hard, lets go shopping!"
86 DEFINE_PER_CPU(unsigned long, cyc2ns);
88 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
90 unsigned long long tsc_now, ns_now;
91 unsigned long flags, *scale;
93 local_irq_save(flags);
94 sched_clock_idle_sleep_event();
96 scale = &per_cpu(cyc2ns, cpu);
99 ns_now = __cycles_2_ns(tsc_now);
102 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
105 * Start smoothly with the new frequency:
107 sched_clock_idle_wakeup_event(0);
108 local_irq_restore(flags);
112 * Scheduler clock - returns current time in nanosec units.
114 unsigned long long native_sched_clock(void)
116 unsigned long long this_offset;
119 * Fall back to jiffies if there's no TSC available:
120 * ( But note that we still use it if the TSC is marked
121 * unstable. We do this because unlike Time Of Day,
122 * the scheduler clock tolerates small errors and it's
123 * very important for it to be as fast as the platform
126 if (unlikely(tsc_disabled))
127 /* No locking but a rare wrong value is not a big deal: */
128 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
130 /* read the Time Stamp Counter: */
131 rdtscll(this_offset);
133 /* return the value in ns */
134 return cycles_2_ns(this_offset);
137 /* We need to define a real function for sched_clock, to override the
138 weak default version */
139 #ifdef CONFIG_PARAVIRT
140 unsigned long long sched_clock(void)
142 return paravirt_sched_clock();
145 unsigned long long sched_clock(void)
146 __attribute__((alias("native_sched_clock")));
149 unsigned long native_calculate_cpu_khz(void)
151 unsigned long long start, end;
153 u64 delta64 = (u64)ULLONG_MAX;
157 local_irq_save(flags);
159 /* run 3 times to ensure the cache is warm and to get an accurate reading */
160 for (i = 0; i < 3; i++) {
161 mach_prepare_counter();
163 mach_countup(&count);
167 * Error: ECTCNEVERSET
168 * The CTC wasn't reliable: we got a hit on the very first read,
169 * or the CPU was so fast/slow that the quotient wouldn't fit in
175 /* cpu freq too slow: */
176 if ((end - start) <= CALIBRATE_TIME_MSEC)
180 * We want the minimum time of all runs in case one of them
181 * is inaccurate due to SMI or other delay
183 delta64 = min(delta64, (end - start));
186 /* cpu freq too fast (or every run was bad): */
187 if (delta64 > (1ULL<<32))
190 delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
191 do_div(delta64,CALIBRATE_TIME_MSEC);
193 local_irq_restore(flags);
194 return (unsigned long)delta64;
196 local_irq_restore(flags);
200 int recalibrate_cpu_khz(void)
203 unsigned long cpu_khz_old = cpu_khz;
206 cpu_khz = calculate_cpu_khz();
208 cpu_data(0).loops_per_jiffy =
209 cpufreq_scale(cpu_data(0).loops_per_jiffy,
210 cpu_khz_old, cpu_khz);
219 EXPORT_SYMBOL(recalibrate_cpu_khz);
221 #ifdef CONFIG_CPU_FREQ
224 * if the CPU frequency is scaled, TSC-based delays will need a different
225 * loops_per_jiffy value to function properly.
227 static unsigned int ref_freq;
228 static unsigned long loops_per_jiffy_ref;
229 static unsigned long cpu_khz_ref;
232 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
234 struct cpufreq_freqs *freq = data;
238 ref_freq = freq->new;
241 ref_freq = freq->old;
242 loops_per_jiffy_ref = cpu_data(freq->cpu).loops_per_jiffy;
243 cpu_khz_ref = cpu_khz;
246 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
247 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
248 (val == CPUFREQ_RESUMECHANGE)) {
249 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
250 cpu_data(freq->cpu).loops_per_jiffy =
251 cpufreq_scale(loops_per_jiffy_ref,
252 ref_freq, freq->new);
256 if (num_online_cpus() == 1)
257 cpu_khz = cpufreq_scale(cpu_khz_ref,
258 ref_freq, freq->new);
259 if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
261 set_cyc2ns_scale(cpu_khz, freq->cpu);
263 * TSC based sched_clock turns
266 mark_tsc_unstable("cpufreq changes");
274 static struct notifier_block time_cpufreq_notifier_block = {
275 .notifier_call = time_cpufreq_notifier
278 static int __init cpufreq_tsc(void)
280 return cpufreq_register_notifier(&time_cpufreq_notifier_block,
281 CPUFREQ_TRANSITION_NOTIFIER);
283 core_initcall(cpufreq_tsc);
287 /* clock source code */
289 static struct clocksource clocksource_tsc;
292 * We compare the TSC to the cycle_last value in the clocksource
293 * structure to avoid a nasty time-warp issue. This can be observed in
294 * a very small window right after one CPU updated cycle_last under
295 * xtime lock and the other CPU reads a TSC value which is smaller
296 * than the cycle_last reference value due to a TSC which is slighty
297 * behind. This delta is nowhere else observable, but in that case it
298 * results in a forward time jump in the range of hours due to the
299 * unsigned delta calculation of the time keeping core code, which is
300 * necessary to support wrapping clocksources like pm timer.
302 static cycle_t read_tsc(void)
308 return ret >= clocksource_tsc.cycle_last ?
309 ret : clocksource_tsc.cycle_last;
312 static struct clocksource clocksource_tsc = {
316 .mask = CLOCKSOURCE_MASK(64),
317 .mult = 0, /* to be set */
319 .flags = CLOCK_SOURCE_IS_CONTINUOUS |
320 CLOCK_SOURCE_MUST_VERIFY,
323 void mark_tsc_unstable(char *reason)
327 printk("Marking TSC unstable due to: %s.\n", reason);
328 /* Can be called before registration */
329 if (clocksource_tsc.mult)
330 clocksource_change_rating(&clocksource_tsc, 0);
332 clocksource_tsc.rating = 0;
335 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
337 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
339 printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
345 /* List of systems that have known TSC problems */
346 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
348 .callback = dmi_mark_tsc_unstable,
349 .ident = "IBM Thinkpad 380XD",
351 DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
352 DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
359 * Make an educated guess if the TSC is trustworthy and synchronized
362 __cpuinit int unsynchronized_tsc(void)
364 if (!cpu_has_tsc || tsc_unstable)
367 /* Anything with constant TSC should be synchronized */
368 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
372 * Intel systems are normally all synchronized.
373 * Exceptions must mark TSC as unstable:
375 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
376 /* assume multi socket systems are not synchronized: */
377 if (num_possible_cpus() > 1)
384 * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
386 #ifdef CONFIG_MGEODE_LX
387 /* RTSC counts during suspend */
388 #define RTSC_SUSP 0x100
390 static void __init check_geode_tsc_reliable(void)
392 unsigned long res_low, res_high;
394 rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
395 if (res_low & RTSC_SUSP)
396 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
399 static inline void check_geode_tsc_reliable(void) { }
403 void __init tsc_init(void)
407 if (!cpu_has_tsc || tsc_disabled > 0)
410 cpu_khz = calculate_cpu_khz();
414 mark_tsc_unstable("could not calculate TSC khz");
418 /* now allow native_sched_clock() to use rdtsc */
421 printk("Detected %lu.%03lu MHz processor.\n",
422 (unsigned long)cpu_khz / 1000,
423 (unsigned long)cpu_khz % 1000);
426 * Secondary CPUs do not run through tsc_init(), so set up
427 * all the scale factors for all CPUs, assuming the same
428 * speed as the bootup CPU. (cpufreq notifiers will fix this
429 * up if their speed diverges)
431 for_each_possible_cpu(cpu)
432 set_cyc2ns_scale(cpu_khz, cpu);
436 /* Check and install the TSC clocksource */
437 dmi_check_system(bad_tsc_dmi_table);
439 unsynchronized_tsc();
440 check_geode_tsc_reliable();
441 clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
442 clocksource_tsc.shift);
443 /* lower the rating if we already know its unstable: */
444 if (check_tsc_unstable()) {
445 clocksource_tsc.rating = 0;
446 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
448 clocksource_register(&clocksource_tsc);