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/sched.h>
8 #include <linux/clocksource.h>
9 #include <linux/workqueue.h>
10 #include <linux/cpufreq.h>
11 #include <linux/jiffies.h>
12 #include <linux/init.h>
13 #include <linux/dmi.h>
15 #include <asm/delay.h>
18 #include <asm/timer.h>
20 #include "mach_timer.h"
22 static int tsc_enabled
;
25 * On some systems the TSC frequency does not
26 * change with the cpu frequency. So we need
27 * an extra value to store the TSC freq
34 static int __init
tsc_setup(char *str
)
36 printk(KERN_WARNING
"notsc: Kernel compiled with CONFIG_X86_TSC, "
37 "cannot disable TSC.\n");
42 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
45 static int __init
tsc_setup(char *str
)
53 __setup("notsc", tsc_setup
);
56 * code to mark and check if the TSC is unstable
57 * due to cpufreq or due to unsynced TSCs
59 static int tsc_unstable
;
61 static inline int check_tsc_unstable(void)
66 /* Accellerators for sched_clock()
67 * convert from cycles(64bits) => nanoseconds (64bits)
69 * ns = cycles / (freq / ns_per_sec)
70 * ns = cycles * (ns_per_sec / freq)
71 * ns = cycles * (10^9 / (cpu_khz * 10^3))
72 * ns = cycles * (10^6 / cpu_khz)
74 * Then we use scaling math (suggested by george@mvista.com) to get:
75 * ns = cycles * (10^6 * SC / cpu_khz) / SC
76 * ns = cycles * cyc2ns_scale / SC
78 * And since SC is a constant power of two, we can convert the div
81 * We can use khz divisor instead of mhz to keep a better percision, since
82 * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
83 * (mathieu.desnoyers@polymtl.ca)
85 * -johnstul@us.ibm.com "math is hard, lets go shopping!"
87 unsigned long cyc2ns_scale __read_mostly
;
89 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
91 static inline void set_cyc2ns_scale(unsigned long cpu_khz
)
93 cyc2ns_scale
= (1000000 << CYC2NS_SCALE_FACTOR
)/cpu_khz
;
97 * Scheduler clock - returns current time in nanosec units.
99 unsigned long long native_sched_clock(void)
101 unsigned long long this_offset
;
104 * Fall back to jiffies if there's no TSC available:
105 * ( But note that we still use it if the TSC is marked
106 * unstable. We do this because unlike Time Of Day,
107 * the scheduler clock tolerates small errors and it's
108 * very important for it to be as fast as the platform
111 if (unlikely(!tsc_enabled
&& !tsc_unstable
))
112 /* No locking but a rare wrong value is not a big deal: */
113 return (jiffies_64
- INITIAL_JIFFIES
) * (1000000000 / HZ
);
115 /* read the Time Stamp Counter: */
116 rdtscll(this_offset
);
118 /* return the value in ns */
119 return cycles_2_ns(this_offset
);
122 /* We need to define a real function for sched_clock, to override the
123 weak default version */
124 #ifdef CONFIG_PARAVIRT
125 unsigned long long sched_clock(void)
127 return paravirt_sched_clock();
130 unsigned long long sched_clock(void)
131 __attribute__((alias("native_sched_clock")));
134 unsigned long native_calculate_cpu_khz(void)
136 unsigned long long start
, end
;
142 local_irq_save(flags
);
144 /* run 3 times to ensure the cache is warm */
145 for (i
= 0; i
< 3; i
++) {
146 mach_prepare_counter();
148 mach_countup(&count
);
152 * Error: ECTCNEVERSET
153 * The CTC wasn't reliable: we got a hit on the very first read,
154 * or the CPU was so fast/slow that the quotient wouldn't fit in
160 delta64
= end
- start
;
162 /* cpu freq too fast: */
163 if (delta64
> (1ULL<<32))
166 /* cpu freq too slow: */
167 if (delta64
<= CALIBRATE_TIME_MSEC
)
170 delta64
+= CALIBRATE_TIME_MSEC
/2; /* round for do_div */
171 do_div(delta64
,CALIBRATE_TIME_MSEC
);
173 local_irq_restore(flags
);
174 return (unsigned long)delta64
;
176 local_irq_restore(flags
);
180 int recalibrate_cpu_khz(void)
183 unsigned long cpu_khz_old
= cpu_khz
;
186 cpu_khz
= calculate_cpu_khz();
188 cpu_data
[0].loops_per_jiffy
=
189 cpufreq_scale(cpu_data
[0].loops_per_jiffy
,
190 cpu_khz_old
, cpu_khz
);
199 EXPORT_SYMBOL(recalibrate_cpu_khz
);
201 #ifdef CONFIG_CPU_FREQ
204 * if the CPU frequency is scaled, TSC-based delays will need a different
205 * loops_per_jiffy value to function properly.
207 static unsigned int ref_freq
= 0;
208 static unsigned long loops_per_jiffy_ref
= 0;
209 static unsigned long cpu_khz_ref
= 0;
212 time_cpufreq_notifier(struct notifier_block
*nb
, unsigned long val
, void *data
)
214 struct cpufreq_freqs
*freq
= data
;
218 ref_freq
= freq
->new;
221 ref_freq
= freq
->old
;
222 loops_per_jiffy_ref
= cpu_data
[freq
->cpu
].loops_per_jiffy
;
223 cpu_khz_ref
= cpu_khz
;
226 if ((val
== CPUFREQ_PRECHANGE
&& freq
->old
< freq
->new) ||
227 (val
== CPUFREQ_POSTCHANGE
&& freq
->old
> freq
->new) ||
228 (val
== CPUFREQ_RESUMECHANGE
)) {
229 if (!(freq
->flags
& CPUFREQ_CONST_LOOPS
))
230 cpu_data
[freq
->cpu
].loops_per_jiffy
=
231 cpufreq_scale(loops_per_jiffy_ref
,
232 ref_freq
, freq
->new);
236 if (num_online_cpus() == 1)
237 cpu_khz
= cpufreq_scale(cpu_khz_ref
,
238 ref_freq
, freq
->new);
239 if (!(freq
->flags
& CPUFREQ_CONST_LOOPS
)) {
241 set_cyc2ns_scale(cpu_khz
);
243 * TSC based sched_clock turns
246 mark_tsc_unstable("cpufreq changes");
254 static struct notifier_block time_cpufreq_notifier_block
= {
255 .notifier_call
= time_cpufreq_notifier
258 static int __init
cpufreq_tsc(void)
260 return cpufreq_register_notifier(&time_cpufreq_notifier_block
,
261 CPUFREQ_TRANSITION_NOTIFIER
);
263 core_initcall(cpufreq_tsc
);
267 /* clock source code */
269 static unsigned long current_tsc_khz
= 0;
271 static cycle_t
read_tsc(void)
280 static struct clocksource clocksource_tsc
= {
284 .mask
= CLOCKSOURCE_MASK(64),
285 .mult
= 0, /* to be set */
287 .flags
= CLOCK_SOURCE_IS_CONTINUOUS
|
288 CLOCK_SOURCE_MUST_VERIFY
,
291 void mark_tsc_unstable(char *reason
)
293 sched_clock_unstable_event();
297 printk("Marking TSC unstable due to: %s.\n", reason
);
298 /* Can be called before registration */
299 if (clocksource_tsc
.mult
)
300 clocksource_change_rating(&clocksource_tsc
, 0);
302 clocksource_tsc
.rating
= 0;
305 EXPORT_SYMBOL_GPL(mark_tsc_unstable
);
307 static int __init
dmi_mark_tsc_unstable(struct dmi_system_id
*d
)
309 printk(KERN_NOTICE
"%s detected: marking TSC unstable.\n",
315 /* List of systems that have known TSC problems */
316 static struct dmi_system_id __initdata bad_tsc_dmi_table
[] = {
318 .callback
= dmi_mark_tsc_unstable
,
319 .ident
= "IBM Thinkpad 380XD",
321 DMI_MATCH(DMI_BOARD_VENDOR
, "IBM"),
322 DMI_MATCH(DMI_BOARD_NAME
, "2635FA0"),
329 * Make an educated guess if the TSC is trustworthy and synchronized
332 __cpuinit
int unsynchronized_tsc(void)
334 if (!cpu_has_tsc
|| tsc_unstable
)
337 * Intel systems are normally all synchronized.
338 * Exceptions must mark TSC as unstable:
340 if (boot_cpu_data
.x86_vendor
!= X86_VENDOR_INTEL
) {
341 /* assume multi socket systems are not synchronized: */
342 if (num_possible_cpus() > 1)
349 * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
351 #ifdef CONFIG_MGEODE_LX
352 /* RTSC counts during suspend */
353 #define RTSC_SUSP 0x100
355 static void __init
check_geode_tsc_reliable(void)
359 rdmsrl(MSR_GEODE_BUSCONT_CONF0
, val
);
360 if ((val
& RTSC_SUSP
))
361 clocksource_tsc
.flags
&= ~CLOCK_SOURCE_MUST_VERIFY
;
364 static inline void check_geode_tsc_reliable(void) { }
368 void __init
tsc_init(void)
370 if (!cpu_has_tsc
|| tsc_disable
)
373 cpu_khz
= calculate_cpu_khz();
379 printk("Detected %lu.%03lu MHz processor.\n",
380 (unsigned long)cpu_khz
/ 1000,
381 (unsigned long)cpu_khz
% 1000);
383 set_cyc2ns_scale(cpu_khz
);
386 /* Check and install the TSC clocksource */
387 dmi_check_system(bad_tsc_dmi_table
);
389 unsynchronized_tsc();
390 check_geode_tsc_reliable();
391 current_tsc_khz
= tsc_khz
;
392 clocksource_tsc
.mult
= clocksource_khz2mult(current_tsc_khz
,
393 clocksource_tsc
.shift
);
394 /* lower the rating if we already know its unstable: */
395 if (check_tsc_unstable()) {
396 clocksource_tsc
.rating
= 0;
397 clocksource_tsc
.flags
&= ~CLOCK_SOURCE_IS_CONTINUOUS
;
401 clocksource_register(&clocksource_tsc
);
407 * Set the tsc_disable flag if there's no TSC support, this
408 * makes it a fast flag for the kernel to see whether it
409 * should be using the TSC.