2 * This code largely moved from arch/i386/kernel/timer/timer_tsc.c
3 * which was originally moved from arch/i386/kernel/time.c.
4 * See comments there for proper credits.
7 #include <linux/clocksource.h>
8 #include <linux/workqueue.h>
9 #include <linux/cpufreq.h>
10 #include <linux/jiffies.h>
11 #include <linux/init.h>
12 #include <linux/dmi.h>
14 #include <asm/delay.h>
18 #include "mach_timer.h"
21 * On some systems the TSC frequency does not
22 * change with the cpu frequency. So we need
23 * an extra value to store the TSC freq
26 unsigned long long (*custom_sched_clock)(void);
31 static int __init tsc_setup(char *str)
33 printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
34 "cannot disable TSC.\n");
39 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
42 static int __init tsc_setup(char *str)
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 static inline int check_tsc_unstable(void)
63 void mark_tsc_unstable(void)
67 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
69 /* Accellerators for sched_clock()
70 * convert from cycles(64bits) => nanoseconds (64bits)
72 * ns = cycles / (freq / ns_per_sec)
73 * ns = cycles * (ns_per_sec / freq)
74 * ns = cycles * (10^9 / (cpu_khz * 10^3))
75 * ns = cycles * (10^6 / cpu_khz)
77 * Then we use scaling math (suggested by george@mvista.com) to get:
78 * ns = cycles * (10^6 * SC / cpu_khz) / SC
79 * ns = cycles * cyc2ns_scale / SC
81 * And since SC is a constant power of two, we can convert the div
84 * We can use khz divisor instead of mhz to keep a better percision, since
85 * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
86 * (mathieu.desnoyers@polymtl.ca)
88 * -johnstul@us.ibm.com "math is hard, lets go shopping!"
90 static unsigned long cyc2ns_scale __read_mostly;
92 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
94 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
96 cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
99 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
101 return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
105 * Scheduler clock - returns current time in nanosec units.
107 unsigned long long sched_clock(void)
109 unsigned long long this_offset;
111 if (unlikely(custom_sched_clock))
112 return (*custom_sched_clock)();
115 * Fall back to jiffies if there's no TSC available:
117 if (unlikely(tsc_disable))
118 /* No locking but a rare wrong value is not a big deal: */
119 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
121 /* read the Time Stamp Counter: */
122 rdtscll(this_offset);
124 /* return the value in ns */
125 return cycles_2_ns(this_offset);
128 static unsigned long calculate_cpu_khz(void)
130 unsigned long long start, end;
136 local_irq_save(flags);
138 /* run 3 times to ensure the cache is warm */
139 for (i = 0; i < 3; i++) {
140 mach_prepare_counter();
142 mach_countup(&count);
146 * Error: ECTCNEVERSET
147 * The CTC wasn't reliable: we got a hit on the very first read,
148 * or the CPU was so fast/slow that the quotient wouldn't fit in
154 delta64 = end - start;
156 /* cpu freq too fast: */
157 if (delta64 > (1ULL<<32))
160 /* cpu freq too slow: */
161 if (delta64 <= CALIBRATE_TIME_MSEC)
164 delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
165 do_div(delta64,CALIBRATE_TIME_MSEC);
167 local_irq_restore(flags);
168 return (unsigned long)delta64;
170 local_irq_restore(flags);
174 int recalibrate_cpu_khz(void)
177 unsigned long cpu_khz_old = cpu_khz;
180 cpu_khz = calculate_cpu_khz();
182 cpu_data[0].loops_per_jiffy =
183 cpufreq_scale(cpu_data[0].loops_per_jiffy,
184 cpu_khz_old, cpu_khz);
193 EXPORT_SYMBOL(recalibrate_cpu_khz);
195 void __init tsc_init(void)
197 if (!cpu_has_tsc || tsc_disable)
200 cpu_khz = calculate_cpu_khz();
206 printk("Detected %lu.%03lu MHz processor.\n",
207 (unsigned long)cpu_khz / 1000,
208 (unsigned long)cpu_khz % 1000);
210 set_cyc2ns_scale(cpu_khz);
216 * Set the tsc_disable flag if there's no TSC support, this
217 * makes it a fast flag for the kernel to see whether it
218 * should be using the TSC.
223 #ifdef CONFIG_CPU_FREQ
225 static unsigned int cpufreq_delayed_issched = 0;
226 static unsigned int cpufreq_init = 0;
227 static struct work_struct cpufreq_delayed_get_work;
229 static void handle_cpufreq_delayed_get(struct work_struct *work)
233 for_each_online_cpu(cpu)
236 cpufreq_delayed_issched = 0;
240 * if we notice cpufreq oddness, schedule a call to cpufreq_get() as it tries
241 * to verify the CPU frequency the timing core thinks the CPU is running
242 * at is still correct.
244 static inline void cpufreq_delayed_get(void)
246 if (cpufreq_init && !cpufreq_delayed_issched) {
247 cpufreq_delayed_issched = 1;
248 printk(KERN_DEBUG "Checking if CPU frequency changed.\n");
249 schedule_work(&cpufreq_delayed_get_work);
254 * if the CPU frequency is scaled, TSC-based delays will need a different
255 * loops_per_jiffy value to function properly.
257 static unsigned int ref_freq = 0;
258 static unsigned long loops_per_jiffy_ref = 0;
259 static unsigned long cpu_khz_ref = 0;
262 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
264 struct cpufreq_freqs *freq = data;
266 if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
267 write_seqlock_irq(&xtime_lock);
271 ref_freq = freq->new;
274 ref_freq = freq->old;
275 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
276 cpu_khz_ref = cpu_khz;
279 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
280 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
281 (val == CPUFREQ_RESUMECHANGE)) {
282 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
283 cpu_data[freq->cpu].loops_per_jiffy =
284 cpufreq_scale(loops_per_jiffy_ref,
285 ref_freq, freq->new);
289 if (num_online_cpus() == 1)
290 cpu_khz = cpufreq_scale(cpu_khz_ref,
291 ref_freq, freq->new);
292 if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
294 set_cyc2ns_scale(cpu_khz);
296 * TSC based sched_clock turns
304 if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
305 write_sequnlock_irq(&xtime_lock);
310 static struct notifier_block time_cpufreq_notifier_block = {
311 .notifier_call = time_cpufreq_notifier
314 static int __init cpufreq_tsc(void)
318 INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get);
319 ret = cpufreq_register_notifier(&time_cpufreq_notifier_block,
320 CPUFREQ_TRANSITION_NOTIFIER);
327 core_initcall(cpufreq_tsc);
331 /* clock source code */
333 static unsigned long current_tsc_khz = 0;
334 static int tsc_update_callback(void);
336 static cycle_t read_tsc(void)
345 static struct clocksource clocksource_tsc = {
349 .mask = CLOCKSOURCE_MASK(64),
350 .mult = 0, /* to be set */
352 .update_callback = tsc_update_callback,
356 static int tsc_update_callback(void)
360 /* check to see if we should switch to the safe clocksource: */
361 if (clocksource_tsc.rating != 0 && check_tsc_unstable()) {
362 clocksource_tsc.rating = 0;
363 clocksource_reselect();
367 /* only update if tsc_khz has changed: */
368 if (current_tsc_khz != tsc_khz) {
369 current_tsc_khz = tsc_khz;
370 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
371 clocksource_tsc.shift);
378 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
380 printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
386 /* List of systems that have known TSC problems */
387 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
389 .callback = dmi_mark_tsc_unstable,
390 .ident = "IBM Thinkpad 380XD",
392 DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
393 DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
399 #define TSC_FREQ_CHECK_INTERVAL (10*MSEC_PER_SEC) /* 10sec in MS */
400 static struct timer_list verify_tsc_freq_timer;
402 /* XXX - Probably should add locking */
403 static void verify_tsc_freq(unsigned long unused)
406 static unsigned long last_jiffies;
408 u64 now_tsc, interval_tsc;
409 unsigned long now_jiffies, interval_jiffies;
412 if (check_tsc_unstable())
416 now_jiffies = jiffies;
422 interval_jiffies = now_jiffies - last_jiffies;
423 interval_tsc = now_tsc - last_tsc;
425 do_div(interval_tsc, cpu_khz*1000);
427 if (interval_tsc < (interval_jiffies * 3 / 4)) {
428 printk("TSC appears to be running slowly. "
429 "Marking it as unstable\n");
436 last_jiffies = now_jiffies;
437 /* set us up to go off on the next interval: */
438 mod_timer(&verify_tsc_freq_timer,
439 jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL));
443 * Make an educated guess if the TSC is trustworthy and synchronized
446 static __init int unsynchronized_tsc(void)
449 * Intel systems are normally all synchronized.
450 * Exceptions must mark TSC as unstable:
452 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
455 /* assume multi socket systems are not synchronized: */
456 return num_possible_cpus() > 1;
459 static int __init init_tsc_clocksource(void)
462 if (cpu_has_tsc && tsc_khz && !tsc_disable) {
463 /* check blacklist */
464 dmi_check_system(bad_tsc_dmi_table);
466 if (unsynchronized_tsc()) /* mark unstable if unsynced */
468 current_tsc_khz = tsc_khz;
469 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
470 clocksource_tsc.shift);
471 /* lower the rating if we already know its unstable: */
472 if (check_tsc_unstable())
473 clocksource_tsc.rating = 0;
475 init_timer(&verify_tsc_freq_timer);
476 verify_tsc_freq_timer.function = verify_tsc_freq;
477 verify_tsc_freq_timer.expires =
478 jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL);
479 add_timer(&verify_tsc_freq_timer);
481 return clocksource_register(&clocksource_tsc);
487 module_init(init_tsc_clocksource);