Commit | Line | Data |
---|---|---|
bb29ab26 | 1 | #include <linux/sched.h> |
5d0cf410 | 2 | #include <linux/clocksource.h> |
539eb11e | 3 | #include <linux/workqueue.h> |
4 | #include <linux/cpufreq.h> | |
5 | #include <linux/jiffies.h> | |
6 | #include <linux/init.h> | |
5d0cf410 | 7 | #include <linux/dmi.h> |
539eb11e | 8 | |
5d0cf410 | 9 | #include <asm/delay.h> |
539eb11e | 10 | #include <asm/tsc.h> |
11 | #include <asm/io.h> | |
6cb9a835 | 12 | #include <asm/timer.h> |
539eb11e | 13 | |
14 | #include "mach_timer.h" | |
15 | ||
d9a5c0a4 TG |
16 | static int tsc_enabled; |
17 | ||
539eb11e | 18 | /* |
19 | * On some systems the TSC frequency does not | |
20 | * change with the cpu frequency. So we need | |
21 | * an extra value to store the TSC freq | |
22 | */ | |
23 | unsigned int tsc_khz; | |
d7e28ffe | 24 | EXPORT_SYMBOL_GPL(tsc_khz); |
539eb11e | 25 | |
664c0d3d | 26 | int tsc_disable; |
539eb11e | 27 | |
28 | #ifdef CONFIG_X86_TSC | |
29 | static int __init tsc_setup(char *str) | |
30 | { | |
31 | printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, " | |
32 | "cannot disable TSC.\n"); | |
33 | return 1; | |
34 | } | |
35 | #else | |
36 | /* | |
37 | * disable flag for tsc. Takes effect by clearing the TSC cpu flag | |
38 | * in cpu/common.c | |
39 | */ | |
40 | static int __init tsc_setup(char *str) | |
41 | { | |
42 | tsc_disable = 1; | |
43 | ||
44 | return 1; | |
45 | } | |
46 | #endif | |
47 | ||
48 | __setup("notsc", tsc_setup); | |
49 | ||
539eb11e | 50 | /* |
51 | * code to mark and check if the TSC is unstable | |
52 | * due to cpufreq or due to unsynced TSCs | |
53 | */ | |
54 | static int tsc_unstable; | |
55 | ||
d7e28ffe | 56 | int check_tsc_unstable(void) |
539eb11e | 57 | { |
58 | return tsc_unstable; | |
59 | } | |
d7e28ffe | 60 | EXPORT_SYMBOL_GPL(check_tsc_unstable); |
539eb11e | 61 | |
539eb11e | 62 | /* Accellerators for sched_clock() |
63 | * convert from cycles(64bits) => nanoseconds (64bits) | |
64 | * basic equation: | |
65 | * ns = cycles / (freq / ns_per_sec) | |
66 | * ns = cycles * (ns_per_sec / freq) | |
67 | * ns = cycles * (10^9 / (cpu_khz * 10^3)) | |
68 | * ns = cycles * (10^6 / cpu_khz) | |
69 | * | |
70 | * Then we use scaling math (suggested by george@mvista.com) to get: | |
71 | * ns = cycles * (10^6 * SC / cpu_khz) / SC | |
72 | * ns = cycles * cyc2ns_scale / SC | |
73 | * | |
74 | * And since SC is a constant power of two, we can convert the div | |
75 | * into a shift. | |
76 | * | |
77 | * We can use khz divisor instead of mhz to keep a better percision, since | |
78 | * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits. | |
79 | * (mathieu.desnoyers@polymtl.ca) | |
80 | * | |
81 | * -johnstul@us.ibm.com "math is hard, lets go shopping!" | |
82 | */ | |
688340ea | 83 | unsigned long cyc2ns_scale __read_mostly; |
539eb11e | 84 | |
85 | #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */ | |
86 | ||
87 | static inline void set_cyc2ns_scale(unsigned long cpu_khz) | |
88 | { | |
89 | cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz; | |
90 | } | |
91 | ||
539eb11e | 92 | /* |
93 | * Scheduler clock - returns current time in nanosec units. | |
94 | */ | |
688340ea | 95 | unsigned long long native_sched_clock(void) |
539eb11e | 96 | { |
97 | unsigned long long this_offset; | |
98 | ||
99 | /* | |
f9690982 | 100 | * Fall back to jiffies if there's no TSC available: |
bb29ab26 IM |
101 | * ( But note that we still use it if the TSC is marked |
102 | * unstable. We do this because unlike Time Of Day, | |
103 | * the scheduler clock tolerates small errors and it's | |
104 | * very important for it to be as fast as the platform | |
105 | * can achive it. ) | |
539eb11e | 106 | */ |
bb29ab26 | 107 | if (unlikely(!tsc_enabled && !tsc_unstable)) |
f9690982 | 108 | /* No locking but a rare wrong value is not a big deal: */ |
539eb11e | 109 | return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ); |
110 | ||
111 | /* read the Time Stamp Counter: */ | |
688340ea | 112 | rdtscll(this_offset); |
539eb11e | 113 | |
114 | /* return the value in ns */ | |
115 | return cycles_2_ns(this_offset); | |
116 | } | |
117 | ||
688340ea JF |
118 | /* We need to define a real function for sched_clock, to override the |
119 | weak default version */ | |
120 | #ifdef CONFIG_PARAVIRT | |
121 | unsigned long long sched_clock(void) | |
122 | { | |
123 | return paravirt_sched_clock(); | |
124 | } | |
125 | #else | |
126 | unsigned long long sched_clock(void) | |
127 | __attribute__((alias("native_sched_clock"))); | |
128 | #endif | |
129 | ||
1182d852 | 130 | unsigned long native_calculate_cpu_khz(void) |
539eb11e | 131 | { |
132 | unsigned long long start, end; | |
133 | unsigned long count; | |
134 | u64 delta64; | |
135 | int i; | |
136 | unsigned long flags; | |
137 | ||
138 | local_irq_save(flags); | |
139 | ||
140 | /* run 3 times to ensure the cache is warm */ | |
141 | for (i = 0; i < 3; i++) { | |
142 | mach_prepare_counter(); | |
143 | rdtscll(start); | |
144 | mach_countup(&count); | |
145 | rdtscll(end); | |
146 | } | |
147 | /* | |
148 | * Error: ECTCNEVERSET | |
149 | * The CTC wasn't reliable: we got a hit on the very first read, | |
150 | * or the CPU was so fast/slow that the quotient wouldn't fit in | |
151 | * 32 bits.. | |
152 | */ | |
153 | if (count <= 1) | |
154 | goto err; | |
155 | ||
156 | delta64 = end - start; | |
157 | ||
158 | /* cpu freq too fast: */ | |
159 | if (delta64 > (1ULL<<32)) | |
160 | goto err; | |
161 | ||
162 | /* cpu freq too slow: */ | |
163 | if (delta64 <= CALIBRATE_TIME_MSEC) | |
164 | goto err; | |
165 | ||
166 | delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */ | |
167 | do_div(delta64,CALIBRATE_TIME_MSEC); | |
168 | ||
169 | local_irq_restore(flags); | |
170 | return (unsigned long)delta64; | |
171 | err: | |
172 | local_irq_restore(flags); | |
173 | return 0; | |
174 | } | |
175 | ||
176 | int recalibrate_cpu_khz(void) | |
177 | { | |
178 | #ifndef CONFIG_SMP | |
179 | unsigned long cpu_khz_old = cpu_khz; | |
180 | ||
181 | if (cpu_has_tsc) { | |
182 | cpu_khz = calculate_cpu_khz(); | |
183 | tsc_khz = cpu_khz; | |
184 | cpu_data[0].loops_per_jiffy = | |
185 | cpufreq_scale(cpu_data[0].loops_per_jiffy, | |
186 | cpu_khz_old, cpu_khz); | |
187 | return 0; | |
188 | } else | |
189 | return -ENODEV; | |
190 | #else | |
191 | return -ENODEV; | |
192 | #endif | |
193 | } | |
194 | ||
195 | EXPORT_SYMBOL(recalibrate_cpu_khz); | |
196 | ||
539eb11e | 197 | #ifdef CONFIG_CPU_FREQ |
198 | ||
539eb11e | 199 | /* |
200 | * if the CPU frequency is scaled, TSC-based delays will need a different | |
201 | * loops_per_jiffy value to function properly. | |
202 | */ | |
203 | static unsigned int ref_freq = 0; | |
204 | static unsigned long loops_per_jiffy_ref = 0; | |
205 | static unsigned long cpu_khz_ref = 0; | |
206 | ||
207 | static int | |
208 | time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data) | |
209 | { | |
210 | struct cpufreq_freqs *freq = data; | |
211 | ||
539eb11e | 212 | if (!ref_freq) { |
213 | if (!freq->old){ | |
214 | ref_freq = freq->new; | |
df3624aa | 215 | return 0; |
539eb11e | 216 | } |
217 | ref_freq = freq->old; | |
218 | loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy; | |
219 | cpu_khz_ref = cpu_khz; | |
220 | } | |
221 | ||
222 | if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) || | |
223 | (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) || | |
224 | (val == CPUFREQ_RESUMECHANGE)) { | |
225 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) | |
226 | cpu_data[freq->cpu].loops_per_jiffy = | |
227 | cpufreq_scale(loops_per_jiffy_ref, | |
228 | ref_freq, freq->new); | |
229 | ||
230 | if (cpu_khz) { | |
231 | ||
232 | if (num_online_cpus() == 1) | |
233 | cpu_khz = cpufreq_scale(cpu_khz_ref, | |
234 | ref_freq, freq->new); | |
235 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) { | |
236 | tsc_khz = cpu_khz; | |
237 | set_cyc2ns_scale(cpu_khz); | |
238 | /* | |
239 | * TSC based sched_clock turns | |
240 | * to junk w/ cpufreq | |
241 | */ | |
5a90cf20 | 242 | mark_tsc_unstable("cpufreq changes"); |
539eb11e | 243 | } |
244 | } | |
245 | } | |
539eb11e | 246 | |
247 | return 0; | |
248 | } | |
249 | ||
250 | static struct notifier_block time_cpufreq_notifier_block = { | |
251 | .notifier_call = time_cpufreq_notifier | |
252 | }; | |
253 | ||
254 | static int __init cpufreq_tsc(void) | |
255 | { | |
26a08eb3 TG |
256 | return cpufreq_register_notifier(&time_cpufreq_notifier_block, |
257 | CPUFREQ_TRANSITION_NOTIFIER); | |
539eb11e | 258 | } |
539eb11e | 259 | core_initcall(cpufreq_tsc); |
260 | ||
261 | #endif | |
5d0cf410 | 262 | |
263 | /* clock source code */ | |
264 | ||
265 | static unsigned long current_tsc_khz = 0; | |
5d0cf410 | 266 | |
267 | static cycle_t read_tsc(void) | |
268 | { | |
269 | cycle_t ret; | |
270 | ||
271 | rdtscll(ret); | |
272 | ||
273 | return ret; | |
274 | } | |
275 | ||
276 | static struct clocksource clocksource_tsc = { | |
277 | .name = "tsc", | |
278 | .rating = 300, | |
279 | .read = read_tsc, | |
7f9f303a | 280 | .mask = CLOCKSOURCE_MASK(64), |
5d0cf410 | 281 | .mult = 0, /* to be set */ |
282 | .shift = 22, | |
73b08d2a TG |
283 | .flags = CLOCK_SOURCE_IS_CONTINUOUS | |
284 | CLOCK_SOURCE_MUST_VERIFY, | |
5d0cf410 | 285 | }; |
286 | ||
5a90cf20 | 287 | void mark_tsc_unstable(char *reason) |
5d0cf410 | 288 | { |
7e69f2b1 TG |
289 | if (!tsc_unstable) { |
290 | tsc_unstable = 1; | |
d9a5c0a4 | 291 | tsc_enabled = 0; |
5a90cf20 | 292 | printk("Marking TSC unstable due to: %s.\n", reason); |
7e69f2b1 TG |
293 | /* Can be called before registration */ |
294 | if (clocksource_tsc.mult) | |
295 | clocksource_change_rating(&clocksource_tsc, 0); | |
296 | else | |
297 | clocksource_tsc.rating = 0; | |
5d0cf410 | 298 | } |
5d0cf410 | 299 | } |
7e69f2b1 | 300 | EXPORT_SYMBOL_GPL(mark_tsc_unstable); |
5d0cf410 | 301 | |
1855256c | 302 | static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d) |
5d0cf410 | 303 | { |
304 | printk(KERN_NOTICE "%s detected: marking TSC unstable.\n", | |
305 | d->ident); | |
7e69f2b1 | 306 | tsc_unstable = 1; |
5d0cf410 | 307 | return 0; |
308 | } | |
309 | ||
310 | /* List of systems that have known TSC problems */ | |
311 | static struct dmi_system_id __initdata bad_tsc_dmi_table[] = { | |
312 | { | |
313 | .callback = dmi_mark_tsc_unstable, | |
314 | .ident = "IBM Thinkpad 380XD", | |
315 | .matches = { | |
316 | DMI_MATCH(DMI_BOARD_VENDOR, "IBM"), | |
317 | DMI_MATCH(DMI_BOARD_NAME, "2635FA0"), | |
318 | }, | |
319 | }, | |
320 | {} | |
321 | }; | |
322 | ||
5d0cf410 | 323 | /* |
324 | * Make an educated guess if the TSC is trustworthy and synchronized | |
325 | * over all CPUs. | |
326 | */ | |
95492e46 | 327 | __cpuinit int unsynchronized_tsc(void) |
5d0cf410 | 328 | { |
95492e46 IM |
329 | if (!cpu_has_tsc || tsc_unstable) |
330 | return 1; | |
5d0cf410 | 331 | /* |
332 | * Intel systems are normally all synchronized. | |
333 | * Exceptions must mark TSC as unstable: | |
334 | */ | |
7e69f2b1 TG |
335 | if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) { |
336 | /* assume multi socket systems are not synchronized: */ | |
337 | if (num_possible_cpus() > 1) | |
338 | tsc_unstable = 1; | |
339 | } | |
340 | return tsc_unstable; | |
5d0cf410 | 341 | } |
342 | ||
07190a08 MT |
343 | /* |
344 | * Geode_LX - the OLPC CPU has a possibly a very reliable TSC | |
345 | */ | |
346 | #ifdef CONFIG_MGEODE_LX | |
347 | /* RTSC counts during suspend */ | |
348 | #define RTSC_SUSP 0x100 | |
349 | ||
350 | static void __init check_geode_tsc_reliable(void) | |
351 | { | |
352 | unsigned long val; | |
353 | ||
354 | rdmsrl(MSR_GEODE_BUSCONT_CONF0, val); | |
355 | if ((val & RTSC_SUSP)) | |
356 | clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY; | |
357 | } | |
358 | #else | |
359 | static inline void check_geode_tsc_reliable(void) { } | |
360 | #endif | |
361 | ||
6bb74df4 | 362 | |
363 | void __init tsc_init(void) | |
5d0cf410 | 364 | { |
6bb74df4 | 365 | if (!cpu_has_tsc || tsc_disable) |
366 | goto out_no_tsc; | |
5d0cf410 | 367 | |
6bb74df4 | 368 | cpu_khz = calculate_cpu_khz(); |
369 | tsc_khz = cpu_khz; | |
5d0cf410 | 370 | |
6bb74df4 | 371 | if (!cpu_khz) |
372 | goto out_no_tsc; | |
5d0cf410 | 373 | |
6bb74df4 | 374 | printk("Detected %lu.%03lu MHz processor.\n", |
375 | (unsigned long)cpu_khz / 1000, | |
376 | (unsigned long)cpu_khz % 1000); | |
377 | ||
378 | set_cyc2ns_scale(cpu_khz); | |
379 | use_tsc_delay(); | |
380 | ||
381 | /* Check and install the TSC clocksource */ | |
382 | dmi_check_system(bad_tsc_dmi_table); | |
383 | ||
384 | unsynchronized_tsc(); | |
385 | check_geode_tsc_reliable(); | |
386 | current_tsc_khz = tsc_khz; | |
387 | clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz, | |
388 | clocksource_tsc.shift); | |
389 | /* lower the rating if we already know its unstable: */ | |
390 | if (check_tsc_unstable()) { | |
391 | clocksource_tsc.rating = 0; | |
392 | clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS; | |
d9a5c0a4 TG |
393 | } else |
394 | tsc_enabled = 1; | |
395 | ||
6bb74df4 | 396 | clocksource_register(&clocksource_tsc); |
5d0cf410 | 397 | |
6bb74df4 | 398 | return; |
5d0cf410 | 399 | |
6bb74df4 | 400 | out_no_tsc: |
401 | /* | |
402 | * Set the tsc_disable flag if there's no TSC support, this | |
403 | * makes it a fast flag for the kernel to see whether it | |
404 | * should be using the TSC. | |
405 | */ | |
406 | tsc_disable = 1; | |
407 | } |