diff options
author | Alok Kataria <akataria@vmware.com> | 2008-07-01 14:43:31 -0400 |
---|---|---|
committer | Ingo Molnar <mingo@elte.hu> | 2008-07-09 01:43:26 -0400 |
commit | 2dbe06faf37b39f9ecffc054dd173b2a1dc2adcd (patch) | |
tree | 530104ee7b810420983bcb2674724721ced76d0a /arch | |
parent | bfc0f5947afa5e3a13e55867f4478c8a92c11dca (diff) |
x86: merge the TSC cpu-freq code
Unify the TSC cpufreq code.
Signed-off-by: Alok N Kataria <akataria@vmware.com>
Signed-off-by: Dan Hecht <dhecht@vmware.com>
Cc: Dan Hecht <dhecht@vmware.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Diffstat (limited to 'arch')
-rw-r--r-- | arch/x86/kernel/tsc.c | 114 | ||||
-rw-r--r-- | arch/x86/kernel/tsc_32.c | 113 | ||||
-rw-r--r-- | arch/x86/kernel/tsc_64.c | 114 |
3 files changed, 114 insertions, 227 deletions
diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c index e6ee14533c75..595f78a22212 100644 --- a/arch/x86/kernel/tsc.c +++ b/arch/x86/kernel/tsc.c | |||
@@ -4,6 +4,7 @@ | |||
4 | #include <linux/module.h> | 4 | #include <linux/module.h> |
5 | #include <linux/timer.h> | 5 | #include <linux/timer.h> |
6 | #include <linux/acpi_pmtmr.h> | 6 | #include <linux/acpi_pmtmr.h> |
7 | #include <linux/cpufreq.h> | ||
7 | 8 | ||
8 | #include <asm/hpet.h> | 9 | #include <asm/hpet.h> |
9 | 10 | ||
@@ -215,3 +216,116 @@ int recalibrate_cpu_khz(void) | |||
215 | EXPORT_SYMBOL(recalibrate_cpu_khz); | 216 | EXPORT_SYMBOL(recalibrate_cpu_khz); |
216 | 217 | ||
217 | #endif /* CONFIG_X86_32 */ | 218 | #endif /* CONFIG_X86_32 */ |
219 | |||
220 | /* Accelerators for sched_clock() | ||
221 | * convert from cycles(64bits) => nanoseconds (64bits) | ||
222 | * basic equation: | ||
223 | * ns = cycles / (freq / ns_per_sec) | ||
224 | * ns = cycles * (ns_per_sec / freq) | ||
225 | * ns = cycles * (10^9 / (cpu_khz * 10^3)) | ||
226 | * ns = cycles * (10^6 / cpu_khz) | ||
227 | * | ||
228 | * Then we use scaling math (suggested by george@mvista.com) to get: | ||
229 | * ns = cycles * (10^6 * SC / cpu_khz) / SC | ||
230 | * ns = cycles * cyc2ns_scale / SC | ||
231 | * | ||
232 | * And since SC is a constant power of two, we can convert the div | ||
233 | * into a shift. | ||
234 | * | ||
235 | * We can use khz divisor instead of mhz to keep a better precision, since | ||
236 | * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits. | ||
237 | * (mathieu.desnoyers@polymtl.ca) | ||
238 | * | ||
239 | * -johnstul@us.ibm.com "math is hard, lets go shopping!" | ||
240 | */ | ||
241 | |||
242 | DEFINE_PER_CPU(unsigned long, cyc2ns); | ||
243 | |||
244 | void set_cyc2ns_scale(unsigned long cpu_khz, int cpu) | ||
245 | { | ||
246 | unsigned long long tsc_now, ns_now; | ||
247 | unsigned long flags, *scale; | ||
248 | |||
249 | local_irq_save(flags); | ||
250 | sched_clock_idle_sleep_event(); | ||
251 | |||
252 | scale = &per_cpu(cyc2ns, cpu); | ||
253 | |||
254 | rdtscll(tsc_now); | ||
255 | ns_now = __cycles_2_ns(tsc_now); | ||
256 | |||
257 | if (cpu_khz) | ||
258 | *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz; | ||
259 | |||
260 | sched_clock_idle_wakeup_event(0); | ||
261 | local_irq_restore(flags); | ||
262 | } | ||
263 | |||
264 | #ifdef CONFIG_CPU_FREQ | ||
265 | |||
266 | /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency | ||
267 | * changes. | ||
268 | * | ||
269 | * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's | ||
270 | * not that important because current Opteron setups do not support | ||
271 | * scaling on SMP anyroads. | ||
272 | * | ||
273 | * Should fix up last_tsc too. Currently gettimeofday in the | ||
274 | * first tick after the change will be slightly wrong. | ||
275 | */ | ||
276 | |||
277 | static unsigned int ref_freq; | ||
278 | static unsigned long loops_per_jiffy_ref; | ||
279 | static unsigned long tsc_khz_ref; | ||
280 | |||
281 | static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, | ||
282 | void *data) | ||
283 | { | ||
284 | struct cpufreq_freqs *freq = data; | ||
285 | unsigned long *lpj, dummy; | ||
286 | |||
287 | if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC)) | ||
288 | return 0; | ||
289 | |||
290 | lpj = &dummy; | ||
291 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) | ||
292 | #ifdef CONFIG_SMP | ||
293 | lpj = &cpu_data(freq->cpu).loops_per_jiffy; | ||
294 | #else | ||
295 | lpj = &boot_cpu_data.loops_per_jiffy; | ||
296 | #endif | ||
297 | |||
298 | if (!ref_freq) { | ||
299 | ref_freq = freq->old; | ||
300 | loops_per_jiffy_ref = *lpj; | ||
301 | tsc_khz_ref = tsc_khz; | ||
302 | } | ||
303 | if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) || | ||
304 | (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) || | ||
305 | (val == CPUFREQ_RESUMECHANGE)) { | ||
306 | *lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new); | ||
307 | |||
308 | tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new); | ||
309 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) | ||
310 | mark_tsc_unstable("cpufreq changes"); | ||
311 | } | ||
312 | |||
313 | set_cyc2ns_scale(tsc_khz_ref, freq->cpu); | ||
314 | |||
315 | return 0; | ||
316 | } | ||
317 | |||
318 | static struct notifier_block time_cpufreq_notifier_block = { | ||
319 | .notifier_call = time_cpufreq_notifier | ||
320 | }; | ||
321 | |||
322 | static int __init cpufreq_tsc(void) | ||
323 | { | ||
324 | cpufreq_register_notifier(&time_cpufreq_notifier_block, | ||
325 | CPUFREQ_TRANSITION_NOTIFIER); | ||
326 | return 0; | ||
327 | } | ||
328 | |||
329 | core_initcall(cpufreq_tsc); | ||
330 | |||
331 | #endif /* CONFIG_CPU_FREQ */ | ||
diff --git a/arch/x86/kernel/tsc_32.c b/arch/x86/kernel/tsc_32.c index 40c0aafb358d..bbc153d36f84 100644 --- a/arch/x86/kernel/tsc_32.c +++ b/arch/x86/kernel/tsc_32.c | |||
@@ -18,119 +18,6 @@ | |||
18 | extern int tsc_unstable; | 18 | extern int tsc_unstable; |
19 | extern int tsc_disabled; | 19 | extern int tsc_disabled; |
20 | 20 | ||
21 | /* Accelerators for sched_clock() | ||
22 | * convert from cycles(64bits) => nanoseconds (64bits) | ||
23 | * basic equation: | ||
24 | * ns = cycles / (freq / ns_per_sec) | ||
25 | * ns = cycles * (ns_per_sec / freq) | ||
26 | * ns = cycles * (10^9 / (cpu_khz * 10^3)) | ||
27 | * ns = cycles * (10^6 / cpu_khz) | ||
28 | * | ||
29 | * Then we use scaling math (suggested by george@mvista.com) to get: | ||
30 | * ns = cycles * (10^6 * SC / cpu_khz) / SC | ||
31 | * ns = cycles * cyc2ns_scale / SC | ||
32 | * | ||
33 | * And since SC is a constant power of two, we can convert the div | ||
34 | * into a shift. | ||
35 | * | ||
36 | * We can use khz divisor instead of mhz to keep a better precision, since | ||
37 | * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits. | ||
38 | * (mathieu.desnoyers@polymtl.ca) | ||
39 | * | ||
40 | * -johnstul@us.ibm.com "math is hard, lets go shopping!" | ||
41 | */ | ||
42 | |||
43 | DEFINE_PER_CPU(unsigned long, cyc2ns); | ||
44 | |||
45 | void set_cyc2ns_scale(unsigned long cpu_khz, int cpu) | ||
46 | { | ||
47 | unsigned long long tsc_now, ns_now; | ||
48 | unsigned long flags, *scale; | ||
49 | |||
50 | local_irq_save(flags); | ||
51 | sched_clock_idle_sleep_event(); | ||
52 | |||
53 | scale = &per_cpu(cyc2ns, cpu); | ||
54 | |||
55 | rdtscll(tsc_now); | ||
56 | ns_now = __cycles_2_ns(tsc_now); | ||
57 | |||
58 | if (cpu_khz) | ||
59 | *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz; | ||
60 | |||
61 | /* | ||
62 | * Start smoothly with the new frequency: | ||
63 | */ | ||
64 | sched_clock_idle_wakeup_event(0); | ||
65 | local_irq_restore(flags); | ||
66 | } | ||
67 | |||
68 | #ifdef CONFIG_CPU_FREQ | ||
69 | |||
70 | /* | ||
71 | * if the CPU frequency is scaled, TSC-based delays will need a different | ||
72 | * loops_per_jiffy value to function properly. | ||
73 | */ | ||
74 | static unsigned int ref_freq; | ||
75 | static unsigned long loops_per_jiffy_ref; | ||
76 | static unsigned long cpu_khz_ref; | ||
77 | |||
78 | static int | ||
79 | time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data) | ||
80 | { | ||
81 | struct cpufreq_freqs *freq = data; | ||
82 | |||
83 | if (!ref_freq) { | ||
84 | if (!freq->old){ | ||
85 | ref_freq = freq->new; | ||
86 | return 0; | ||
87 | } | ||
88 | ref_freq = freq->old; | ||
89 | loops_per_jiffy_ref = cpu_data(freq->cpu).loops_per_jiffy; | ||
90 | cpu_khz_ref = cpu_khz; | ||
91 | } | ||
92 | |||
93 | if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) || | ||
94 | (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) || | ||
95 | (val == CPUFREQ_RESUMECHANGE)) { | ||
96 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) | ||
97 | cpu_data(freq->cpu).loops_per_jiffy = | ||
98 | cpufreq_scale(loops_per_jiffy_ref, | ||
99 | ref_freq, freq->new); | ||
100 | |||
101 | if (cpu_khz) { | ||
102 | |||
103 | if (num_online_cpus() == 1) | ||
104 | cpu_khz = cpufreq_scale(cpu_khz_ref, | ||
105 | ref_freq, freq->new); | ||
106 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) { | ||
107 | tsc_khz = cpu_khz; | ||
108 | set_cyc2ns_scale(cpu_khz, freq->cpu); | ||
109 | /* | ||
110 | * TSC based sched_clock turns | ||
111 | * to junk w/ cpufreq | ||
112 | */ | ||
113 | mark_tsc_unstable("cpufreq changes"); | ||
114 | } | ||
115 | } | ||
116 | } | ||
117 | |||
118 | return 0; | ||
119 | } | ||
120 | |||
121 | static struct notifier_block time_cpufreq_notifier_block = { | ||
122 | .notifier_call = time_cpufreq_notifier | ||
123 | }; | ||
124 | |||
125 | static int __init cpufreq_tsc(void) | ||
126 | { | ||
127 | return cpufreq_register_notifier(&time_cpufreq_notifier_block, | ||
128 | CPUFREQ_TRANSITION_NOTIFIER); | ||
129 | } | ||
130 | core_initcall(cpufreq_tsc); | ||
131 | |||
132 | #endif | ||
133 | |||
134 | /* clock source code */ | 21 | /* clock source code */ |
135 | 22 | ||
136 | static struct clocksource clocksource_tsc; | 23 | static struct clocksource clocksource_tsc; |
diff --git a/arch/x86/kernel/tsc_64.c b/arch/x86/kernel/tsc_64.c index c852ff9bd5d4..80a274b018c2 100644 --- a/arch/x86/kernel/tsc_64.c +++ b/arch/x86/kernel/tsc_64.c | |||
@@ -16,120 +16,6 @@ | |||
16 | extern int tsc_unstable; | 16 | extern int tsc_unstable; |
17 | extern int tsc_disabled; | 17 | extern int tsc_disabled; |
18 | 18 | ||
19 | /* Accelerators for sched_clock() | ||
20 | * convert from cycles(64bits) => nanoseconds (64bits) | ||
21 | * basic equation: | ||
22 | * ns = cycles / (freq / ns_per_sec) | ||
23 | * ns = cycles * (ns_per_sec / freq) | ||
24 | * ns = cycles * (10^9 / (cpu_khz * 10^3)) | ||
25 | * ns = cycles * (10^6 / cpu_khz) | ||
26 | * | ||
27 | * Then we use scaling math (suggested by george@mvista.com) to get: | ||
28 | * ns = cycles * (10^6 * SC / cpu_khz) / SC | ||
29 | * ns = cycles * cyc2ns_scale / SC | ||
30 | * | ||
31 | * And since SC is a constant power of two, we can convert the div | ||
32 | * into a shift. | ||
33 | * | ||
34 | * We can use khz divisor instead of mhz to keep a better precision, since | ||
35 | * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits. | ||
36 | * (mathieu.desnoyers@polymtl.ca) | ||
37 | * | ||
38 | * -johnstul@us.ibm.com "math is hard, lets go shopping!" | ||
39 | */ | ||
40 | |||
41 | DEFINE_PER_CPU(unsigned long, cyc2ns); | ||
42 | |||
43 | void set_cyc2ns_scale(unsigned long cpu_khz, int cpu) | ||
44 | { | ||
45 | unsigned long long tsc_now, ns_now; | ||
46 | unsigned long flags, *scale; | ||
47 | |||
48 | local_irq_save(flags); | ||
49 | sched_clock_idle_sleep_event(); | ||
50 | |||
51 | scale = &per_cpu(cyc2ns, cpu); | ||
52 | |||
53 | rdtscll(tsc_now); | ||
54 | ns_now = __cycles_2_ns(tsc_now); | ||
55 | |||
56 | if (cpu_khz) | ||
57 | *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz; | ||
58 | |||
59 | sched_clock_idle_wakeup_event(0); | ||
60 | local_irq_restore(flags); | ||
61 | } | ||
62 | |||
63 | #ifdef CONFIG_CPU_FREQ | ||
64 | |||
65 | /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency | ||
66 | * changes. | ||
67 | * | ||
68 | * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's | ||
69 | * not that important because current Opteron setups do not support | ||
70 | * scaling on SMP anyroads. | ||
71 | * | ||
72 | * Should fix up last_tsc too. Currently gettimeofday in the | ||
73 | * first tick after the change will be slightly wrong. | ||
74 | */ | ||
75 | |||
76 | static unsigned int ref_freq; | ||
77 | static unsigned long loops_per_jiffy_ref; | ||
78 | static unsigned long tsc_khz_ref; | ||
79 | |||
80 | static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, | ||
81 | void *data) | ||
82 | { | ||
83 | struct cpufreq_freqs *freq = data; | ||
84 | unsigned long *lpj, dummy; | ||
85 | |||
86 | if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC)) | ||
87 | return 0; | ||
88 | |||
89 | lpj = &dummy; | ||
90 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) | ||
91 | #ifdef CONFIG_SMP | ||
92 | lpj = &cpu_data(freq->cpu).loops_per_jiffy; | ||
93 | #else | ||
94 | lpj = &boot_cpu_data.loops_per_jiffy; | ||
95 | #endif | ||
96 | |||
97 | if (!ref_freq) { | ||
98 | ref_freq = freq->old; | ||
99 | loops_per_jiffy_ref = *lpj; | ||
100 | tsc_khz_ref = tsc_khz; | ||
101 | } | ||
102 | if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) || | ||
103 | (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) || | ||
104 | (val == CPUFREQ_RESUMECHANGE)) { | ||
105 | *lpj = | ||
106 | cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new); | ||
107 | |||
108 | tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new); | ||
109 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) | ||
110 | mark_tsc_unstable("cpufreq changes"); | ||
111 | } | ||
112 | |||
113 | set_cyc2ns_scale(tsc_khz_ref, freq->cpu); | ||
114 | |||
115 | return 0; | ||
116 | } | ||
117 | |||
118 | static struct notifier_block time_cpufreq_notifier_block = { | ||
119 | .notifier_call = time_cpufreq_notifier | ||
120 | }; | ||
121 | |||
122 | static int __init cpufreq_tsc(void) | ||
123 | { | ||
124 | cpufreq_register_notifier(&time_cpufreq_notifier_block, | ||
125 | CPUFREQ_TRANSITION_NOTIFIER); | ||
126 | return 0; | ||
127 | } | ||
128 | |||
129 | core_initcall(cpufreq_tsc); | ||
130 | |||
131 | #endif | ||
132 | |||
133 | /* | 19 | /* |
134 | * Make an educated guess if the TSC is trustworthy and synchronized | 20 | * Make an educated guess if the TSC is trustworthy and synchronized |
135 | * over all CPUs. | 21 | * over all CPUs. |