Merge branch 'x86-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6] / arch / x86 / kernel / tsc_32.c
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>
7 #include <linux/dmi.h>
8 #include <linux/percpu.h>
9
10 #include <asm/delay.h>
11 #include <asm/tsc.h>
12 #include <asm/io.h>
13 #include <asm/timer.h>
14
15 #include "mach_timer.h"
16
17 static int tsc_disabled;
18
19 /*
20  * On some systems the TSC frequency does not
21  * change with the cpu frequency. So we need
22  * an extra value to store the TSC freq
23  */
24 unsigned int tsc_khz;
25 EXPORT_SYMBOL_GPL(tsc_khz);
26
27 #ifdef CONFIG_X86_TSC
28 static int __init tsc_setup(char *str)
29 {
30         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
31                "cannot disable TSC completely.\n");
32         tsc_disabled = 1;
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         setup_clear_cpu_cap(X86_FEATURE_TSC);
43         return 1;
44 }
45 #endif
46
47 __setup("notsc", tsc_setup);
48
49 /*
50  * code to mark and check if the TSC is unstable
51  * due to cpufreq or due to unsynced TSCs
52  */
53 static int tsc_unstable;
54
55 int check_tsc_unstable(void)
56 {
57         return tsc_unstable;
58 }
59 EXPORT_SYMBOL_GPL(check_tsc_unstable);
60
61 /* Accelerators for sched_clock()
62  * convert from cycles(64bits) => nanoseconds (64bits)
63  *  basic equation:
64  *              ns = cycles / (freq / ns_per_sec)
65  *              ns = cycles * (ns_per_sec / freq)
66  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
67  *              ns = cycles * (10^6 / cpu_khz)
68  *
69  *      Then we use scaling math (suggested by george@mvista.com) to get:
70  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
71  *              ns = cycles * cyc2ns_scale / SC
72  *
73  *      And since SC is a constant power of two, we can convert the div
74  *  into a shift.
75  *
76  *  We can use khz divisor instead of mhz to keep a better precision, since
77  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
78  *  (mathieu.desnoyers@polymtl.ca)
79  *
80  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
81  */
82
83 DEFINE_PER_CPU(unsigned long, cyc2ns);
84
85 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
86 {
87         unsigned long long tsc_now, ns_now;
88         unsigned long flags, *scale;
89
90         local_irq_save(flags);
91         sched_clock_idle_sleep_event();
92
93         scale = &per_cpu(cyc2ns, cpu);
94
95         rdtscll(tsc_now);
96         ns_now = __cycles_2_ns(tsc_now);
97
98         if (cpu_khz)
99                 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
100
101         /*
102          * Start smoothly with the new frequency:
103          */
104         sched_clock_idle_wakeup_event(0);
105         local_irq_restore(flags);
106 }
107
108 /*
109  * Scheduler clock - returns current time in nanosec units.
110  */
111 unsigned long long native_sched_clock(void)
112 {
113         unsigned long long this_offset;
114
115         /*
116          * Fall back to jiffies if there's no TSC available:
117          * ( But note that we still use it if the TSC is marked
118          *   unstable. We do this because unlike Time Of Day,
119          *   the scheduler clock tolerates small errors and it's
120          *   very important for it to be as fast as the platform
121          *   can achive it. )
122          */
123         if (unlikely(tsc_disabled))
124                 /* No locking but a rare wrong value is not a big deal: */
125                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
126
127         /* read the Time Stamp Counter: */
128         rdtscll(this_offset);
129
130         /* return the value in ns */
131         return cycles_2_ns(this_offset);
132 }
133
134 /* We need to define a real function for sched_clock, to override the
135    weak default version */
136 #ifdef CONFIG_PARAVIRT
137 unsigned long long sched_clock(void)
138 {
139         return paravirt_sched_clock();
140 }
141 #else
142 unsigned long long sched_clock(void)
143         __attribute__((alias("native_sched_clock")));
144 #endif
145
146 unsigned long native_calculate_cpu_khz(void)
147 {
148         unsigned long long start, end;
149         unsigned long count;
150         u64 delta64 = (u64)ULLONG_MAX;
151         int i;
152         unsigned long flags;
153
154         local_irq_save(flags);
155
156         /* run 3 times to ensure the cache is warm and to get an accurate reading */
157         for (i = 0; i < 3; i++) {
158                 mach_prepare_counter();
159                 rdtscll(start);
160                 mach_countup(&count);
161                 rdtscll(end);
162
163                 /*
164                  * Error: ECTCNEVERSET
165                  * The CTC wasn't reliable: we got a hit on the very first read,
166                  * or the CPU was so fast/slow that the quotient wouldn't fit in
167                  * 32 bits..
168                  */
169                 if (count <= 1)
170                         continue;
171
172                 /* cpu freq too slow: */
173                 if ((end - start) <= CALIBRATE_TIME_MSEC)
174                         continue;
175
176                 /*
177                  * We want the minimum time of all runs in case one of them
178                  * is inaccurate due to SMI or other delay
179                  */
180                 delta64 = min(delta64, (end - start));
181         }
182
183         /* cpu freq too fast (or every run was bad): */
184         if (delta64 > (1ULL<<32))
185                 goto err;
186
187         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
188         do_div(delta64,CALIBRATE_TIME_MSEC);
189
190         local_irq_restore(flags);
191         return (unsigned long)delta64;
192 err:
193         local_irq_restore(flags);
194         return 0;
195 }
196
197 int recalibrate_cpu_khz(void)
198 {
199 #ifndef CONFIG_SMP
200         unsigned long cpu_khz_old = cpu_khz;
201
202         if (cpu_has_tsc) {
203                 cpu_khz = calculate_cpu_khz();
204                 tsc_khz = cpu_khz;
205                 cpu_data(0).loops_per_jiffy =
206                         cpufreq_scale(cpu_data(0).loops_per_jiffy,
207                                         cpu_khz_old, cpu_khz);
208                 return 0;
209         } else
210                 return -ENODEV;
211 #else
212         return -ENODEV;
213 #endif
214 }
215
216 EXPORT_SYMBOL(recalibrate_cpu_khz);
217
218 #ifdef CONFIG_CPU_FREQ
219
220 /*
221  * if the CPU frequency is scaled, TSC-based delays will need a different
222  * loops_per_jiffy value to function properly.
223  */
224 static unsigned int ref_freq;
225 static unsigned long loops_per_jiffy_ref;
226 static unsigned long cpu_khz_ref;
227
228 static int
229 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
230 {
231         struct cpufreq_freqs *freq = data;
232
233         if (!ref_freq) {
234                 if (!freq->old){
235                         ref_freq = freq->new;
236                         return 0;
237                 }
238                 ref_freq = freq->old;
239                 loops_per_jiffy_ref = cpu_data(freq->cpu).loops_per_jiffy;
240                 cpu_khz_ref = cpu_khz;
241         }
242
243         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
244             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
245             (val == CPUFREQ_RESUMECHANGE)) {
246                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
247                         cpu_data(freq->cpu).loops_per_jiffy =
248                                 cpufreq_scale(loops_per_jiffy_ref,
249                                                 ref_freq, freq->new);
250
251                 if (cpu_khz) {
252
253                         if (num_online_cpus() == 1)
254                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
255                                                 ref_freq, freq->new);
256                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
257                                 tsc_khz = cpu_khz;
258                                 set_cyc2ns_scale(cpu_khz, freq->cpu);
259                                 /*
260                                  * TSC based sched_clock turns
261                                  * to junk w/ cpufreq
262                                  */
263                                 mark_tsc_unstable("cpufreq changes");
264                         }
265                 }
266         }
267
268         return 0;
269 }
270
271 static struct notifier_block time_cpufreq_notifier_block = {
272         .notifier_call  = time_cpufreq_notifier
273 };
274
275 static int __init cpufreq_tsc(void)
276 {
277         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
278                                          CPUFREQ_TRANSITION_NOTIFIER);
279 }
280 core_initcall(cpufreq_tsc);
281
282 #endif
283
284 /* clock source code */
285
286 static unsigned long current_tsc_khz;
287 static struct clocksource clocksource_tsc;
288
289 /*
290  * We compare the TSC to the cycle_last value in the clocksource
291  * structure to avoid a nasty time-warp issue. This can be observed in
292  * a very small window right after one CPU updated cycle_last under
293  * xtime lock and the other CPU reads a TSC value which is smaller
294  * than the cycle_last reference value due to a TSC which is slighty
295  * behind. This delta is nowhere else observable, but in that case it
296  * results in a forward time jump in the range of hours due to the
297  * unsigned delta calculation of the time keeping core code, which is
298  * necessary to support wrapping clocksources like pm timer.
299  */
300 static cycle_t read_tsc(void)
301 {
302         cycle_t ret;
303
304         rdtscll(ret);
305
306         return ret >= clocksource_tsc.cycle_last ?
307                 ret : clocksource_tsc.cycle_last;
308 }
309
310 static struct clocksource clocksource_tsc = {
311         .name                   = "tsc",
312         .rating                 = 300,
313         .read                   = read_tsc,
314         .mask                   = CLOCKSOURCE_MASK(64),
315         .mult                   = 0, /* to be set */
316         .shift                  = 22,
317         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
318                                   CLOCK_SOURCE_MUST_VERIFY,
319 };
320
321 void mark_tsc_unstable(char *reason)
322 {
323         if (!tsc_unstable) {
324                 tsc_unstable = 1;
325                 printk("Marking TSC unstable due to: %s.\n", reason);
326                 /* Can be called before registration */
327                 if (clocksource_tsc.mult)
328                         clocksource_change_rating(&clocksource_tsc, 0);
329                 else
330                         clocksource_tsc.rating = 0;
331         }
332 }
333 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
334
335 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
336 {
337         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
338                d->ident);
339         tsc_unstable = 1;
340         return 0;
341 }
342
343 /* List of systems that have known TSC problems */
344 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
345         {
346          .callback = dmi_mark_tsc_unstable,
347          .ident = "IBM Thinkpad 380XD",
348          .matches = {
349                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
350                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
351                      },
352          },
353          {}
354 };
355
356 /*
357  * Make an educated guess if the TSC is trustworthy and synchronized
358  * over all CPUs.
359  */
360 __cpuinit int unsynchronized_tsc(void)
361 {
362         if (!cpu_has_tsc || tsc_unstable)
363                 return 1;
364
365         /* Anything with constant TSC should be synchronized */
366         if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
367                 return 0;
368
369         /*
370          * Intel systems are normally all synchronized.
371          * Exceptions must mark TSC as unstable:
372          */
373         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
374                 /* assume multi socket systems are not synchronized: */
375                 if (num_possible_cpus() > 1)
376                         tsc_unstable = 1;
377         }
378         return tsc_unstable;
379 }
380
381 /*
382  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
383  */
384 #ifdef CONFIG_MGEODE_LX
385 /* RTSC counts during suspend */
386 #define RTSC_SUSP 0x100
387
388 static void __init check_geode_tsc_reliable(void)
389 {
390         unsigned long res_low, res_high;
391
392         rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
393         if (res_low & RTSC_SUSP)
394                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
395 }
396 #else
397 static inline void check_geode_tsc_reliable(void) { }
398 #endif
399
400
401 void __init tsc_init(void)
402 {
403         int cpu;
404
405         if (!cpu_has_tsc || tsc_disabled) {
406                 /* Disable the TSC in case of !cpu_has_tsc */
407                 tsc_disabled = 1;
408                 return;
409         }
410
411         cpu_khz = calculate_cpu_khz();
412         tsc_khz = cpu_khz;
413
414         if (!cpu_khz) {
415                 mark_tsc_unstable("could not calculate TSC khz");
416                 /*
417                  * We need to disable the TSC completely in this case
418                  * to prevent sched_clock() from using it.
419                  */
420                 tsc_disabled = 1;
421                 return;
422         }
423
424         printk("Detected %lu.%03lu MHz processor.\n",
425                                 (unsigned long)cpu_khz / 1000,
426                                 (unsigned long)cpu_khz % 1000);
427
428         /*
429          * Secondary CPUs do not run through tsc_init(), so set up
430          * all the scale factors for all CPUs, assuming the same
431          * speed as the bootup CPU. (cpufreq notifiers will fix this
432          * up if their speed diverges)
433          */
434         for_each_possible_cpu(cpu)
435                 set_cyc2ns_scale(cpu_khz, cpu);
436
437         use_tsc_delay();
438
439         /* Check and install the TSC clocksource */
440         dmi_check_system(bad_tsc_dmi_table);
441
442         unsynchronized_tsc();
443         check_geode_tsc_reliable();
444         current_tsc_khz = tsc_khz;
445         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
446                                                         clocksource_tsc.shift);
447         /* lower the rating if we already know its unstable: */
448         if (check_tsc_unstable()) {
449                 clocksource_tsc.rating = 0;
450                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
451         }
452         clocksource_register(&clocksource_tsc);
453 }