x86: fix "x86: use cpu_khz for loops_per_jiffy calculation"
[linux-2.6.git] / arch / x86 / kernel / tsc_32.c
1 #include <linux/sched.h>
2 #include <linux/clocksource.h>
3 #include <linux/workqueue.h>
4 #include <linux/delay.h>
5 #include <linux/cpufreq.h>
6 #include <linux/jiffies.h>
7 #include <linux/init.h>
8 #include <linux/dmi.h>
9 #include <linux/percpu.h>
10
11 #include <asm/delay.h>
12 #include <asm/tsc.h>
13 #include <asm/io.h>
14 #include <asm/timer.h>
15
16 #include "mach_timer.h"
17
18 static int tsc_disabled;
19
20 /*
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
24  */
25 unsigned int tsc_khz;
26 EXPORT_SYMBOL_GPL(tsc_khz);
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 completely.\n");
33         tsc_disabled = 1;
34         return 1;
35 }
36 #else
37 /*
38  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
39  * in cpu/common.c
40  */
41 static int __init tsc_setup(char *str)
42 {
43         setup_clear_cpu_cap(X86_FEATURE_TSC);
44         return 1;
45 }
46 #endif
47
48 __setup("notsc", tsc_setup);
49
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
56 int check_tsc_unstable(void)
57 {
58         return tsc_unstable;
59 }
60 EXPORT_SYMBOL_GPL(check_tsc_unstable);
61
62 /* Accelerators 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 precision, 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  */
83
84 DEFINE_PER_CPU(unsigned long, cyc2ns);
85
86 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
87 {
88         unsigned long long tsc_now, ns_now;
89         unsigned long flags, *scale;
90
91         local_irq_save(flags);
92         sched_clock_idle_sleep_event();
93
94         scale = &per_cpu(cyc2ns, cpu);
95
96         rdtscll(tsc_now);
97         ns_now = __cycles_2_ns(tsc_now);
98
99         if (cpu_khz)
100                 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
101
102         /*
103          * Start smoothly with the new frequency:
104          */
105         sched_clock_idle_wakeup_event(0);
106         local_irq_restore(flags);
107 }
108
109 /*
110  * Scheduler clock - returns current time in nanosec units.
111  */
112 unsigned long long native_sched_clock(void)
113 {
114         unsigned long long this_offset;
115
116         /*
117          * Fall back to jiffies if there's no TSC available:
118          * ( But note that we still use it if the TSC is marked
119          *   unstable. We do this because unlike Time Of Day,
120          *   the scheduler clock tolerates small errors and it's
121          *   very important for it to be as fast as the platform
122          *   can achive it. )
123          */
124         if (unlikely(tsc_disabled))
125                 /* No locking but a rare wrong value is not a big deal: */
126                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
127
128         /* read the Time Stamp Counter: */
129         rdtscll(this_offset);
130
131         /* return the value in ns */
132         return cycles_2_ns(this_offset);
133 }
134
135 /* We need to define a real function for sched_clock, to override the
136    weak default version */
137 #ifdef CONFIG_PARAVIRT
138 unsigned long long sched_clock(void)
139 {
140         return paravirt_sched_clock();
141 }
142 #else
143 unsigned long long sched_clock(void)
144         __attribute__((alias("native_sched_clock")));
145 #endif
146
147 unsigned long native_calculate_cpu_khz(void)
148 {
149         unsigned long long start, end;
150         unsigned long count;
151         u64 delta64 = (u64)ULLONG_MAX;
152         int i;
153         unsigned long flags;
154
155         local_irq_save(flags);
156
157         /* run 3 times to ensure the cache is warm and to get an accurate reading */
158         for (i = 0; i < 3; i++) {
159                 mach_prepare_counter();
160                 rdtscll(start);
161                 mach_countup(&count);
162                 rdtscll(end);
163
164                 /*
165                  * Error: ECTCNEVERSET
166                  * The CTC wasn't reliable: we got a hit on the very first read,
167                  * or the CPU was so fast/slow that the quotient wouldn't fit in
168                  * 32 bits..
169                  */
170                 if (count <= 1)
171                         continue;
172
173                 /* cpu freq too slow: */
174                 if ((end - start) <= CALIBRATE_TIME_MSEC)
175                         continue;
176
177                 /*
178                  * We want the minimum time of all runs in case one of them
179                  * is inaccurate due to SMI or other delay
180                  */
181                 delta64 = min(delta64, (end - start));
182         }
183
184         /* cpu freq too fast (or every run was bad): */
185         if (delta64 > (1ULL<<32))
186                 goto err;
187
188         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
189         do_div(delta64,CALIBRATE_TIME_MSEC);
190
191         local_irq_restore(flags);
192         return (unsigned long)delta64;
193 err:
194         local_irq_restore(flags);
195         return 0;
196 }
197
198 int recalibrate_cpu_khz(void)
199 {
200 #ifndef CONFIG_SMP
201         unsigned long cpu_khz_old = cpu_khz;
202
203         if (cpu_has_tsc) {
204                 cpu_khz = calculate_cpu_khz();
205                 tsc_khz = cpu_khz;
206                 cpu_data(0).loops_per_jiffy =
207                         cpufreq_scale(cpu_data(0).loops_per_jiffy,
208                                         cpu_khz_old, cpu_khz);
209                 return 0;
210         } else
211                 return -ENODEV;
212 #else
213         return -ENODEV;
214 #endif
215 }
216
217 EXPORT_SYMBOL(recalibrate_cpu_khz);
218
219 #ifdef CONFIG_CPU_FREQ
220
221 /*
222  * if the CPU frequency is scaled, TSC-based delays will need a different
223  * loops_per_jiffy value to function properly.
224  */
225 static unsigned int ref_freq;
226 static unsigned long loops_per_jiffy_ref;
227 static unsigned long cpu_khz_ref;
228
229 static int
230 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
231 {
232         struct cpufreq_freqs *freq = data;
233
234         if (!ref_freq) {
235                 if (!freq->old){
236                         ref_freq = freq->new;
237                         return 0;
238                 }
239                 ref_freq = freq->old;
240                 loops_per_jiffy_ref = cpu_data(freq->cpu).loops_per_jiffy;
241                 cpu_khz_ref = cpu_khz;
242         }
243
244         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
245             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
246             (val == CPUFREQ_RESUMECHANGE)) {
247                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
248                         cpu_data(freq->cpu).loops_per_jiffy =
249                                 cpufreq_scale(loops_per_jiffy_ref,
250                                                 ref_freq, freq->new);
251
252                 if (cpu_khz) {
253
254                         if (num_online_cpus() == 1)
255                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
256                                                 ref_freq, freq->new);
257                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
258                                 tsc_khz = cpu_khz;
259                                 set_cyc2ns_scale(cpu_khz, freq->cpu);
260                                 /*
261                                  * TSC based sched_clock turns
262                                  * to junk w/ cpufreq
263                                  */
264                                 mark_tsc_unstable("cpufreq changes");
265                         }
266                 }
267         }
268
269         return 0;
270 }
271
272 static struct notifier_block time_cpufreq_notifier_block = {
273         .notifier_call  = time_cpufreq_notifier
274 };
275
276 static int __init cpufreq_tsc(void)
277 {
278         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
279                                          CPUFREQ_TRANSITION_NOTIFIER);
280 }
281 core_initcall(cpufreq_tsc);
282
283 #endif
284
285 /* clock source code */
286
287 static unsigned long current_tsc_khz;
288 static struct clocksource clocksource_tsc;
289
290 /*
291  * We compare the TSC to the cycle_last value in the clocksource
292  * structure to avoid a nasty time-warp issue. This can be observed in
293  * a very small window right after one CPU updated cycle_last under
294  * xtime lock and the other CPU reads a TSC value which is smaller
295  * than the cycle_last reference value due to a TSC which is slighty
296  * behind. This delta is nowhere else observable, but in that case it
297  * results in a forward time jump in the range of hours due to the
298  * unsigned delta calculation of the time keeping core code, which is
299  * necessary to support wrapping clocksources like pm timer.
300  */
301 static cycle_t read_tsc(void)
302 {
303         cycle_t ret;
304
305         rdtscll(ret);
306
307         return ret >= clocksource_tsc.cycle_last ?
308                 ret : clocksource_tsc.cycle_last;
309 }
310
311 static struct clocksource clocksource_tsc = {
312         .name                   = "tsc",
313         .rating                 = 300,
314         .read                   = read_tsc,
315         .mask                   = CLOCKSOURCE_MASK(64),
316         .mult                   = 0, /* to be set */
317         .shift                  = 22,
318         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
319                                   CLOCK_SOURCE_MUST_VERIFY,
320 };
321
322 void mark_tsc_unstable(char *reason)
323 {
324         if (!tsc_unstable) {
325                 tsc_unstable = 1;
326                 printk("Marking TSC unstable due to: %s.\n", reason);
327                 /* Can be called before registration */
328                 if (clocksource_tsc.mult)
329                         clocksource_change_rating(&clocksource_tsc, 0);
330                 else
331                         clocksource_tsc.rating = 0;
332         }
333 }
334 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
335
336 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
337 {
338         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
339                d->ident);
340         tsc_unstable = 1;
341         return 0;
342 }
343
344 /* List of systems that have known TSC problems */
345 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
346         {
347          .callback = dmi_mark_tsc_unstable,
348          .ident = "IBM Thinkpad 380XD",
349          .matches = {
350                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
351                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
352                      },
353          },
354          {}
355 };
356
357 /*
358  * Make an educated guess if the TSC is trustworthy and synchronized
359  * over all CPUs.
360  */
361 __cpuinit int unsynchronized_tsc(void)
362 {
363         if (!cpu_has_tsc || tsc_unstable)
364                 return 1;
365
366         /* Anything with constant TSC should be synchronized */
367         if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
368                 return 0;
369
370         /*
371          * Intel systems are normally all synchronized.
372          * Exceptions must mark TSC as unstable:
373          */
374         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
375                 /* assume multi socket systems are not synchronized: */
376                 if (num_possible_cpus() > 1)
377                         tsc_unstable = 1;
378         }
379         return tsc_unstable;
380 }
381
382 /*
383  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
384  */
385 #ifdef CONFIG_MGEODE_LX
386 /* RTSC counts during suspend */
387 #define RTSC_SUSP 0x100
388
389 static void __init check_geode_tsc_reliable(void)
390 {
391         unsigned long res_low, res_high;
392
393         rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
394         if (res_low & RTSC_SUSP)
395                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
396 }
397 #else
398 static inline void check_geode_tsc_reliable(void) { }
399 #endif
400
401
402 void __init tsc_init(void)
403 {
404         int cpu;
405         u64 lpj;
406
407         if (!cpu_has_tsc || tsc_disabled) {
408                 /* Disable the TSC in case of !cpu_has_tsc */
409                 tsc_disabled = 1;
410                 return;
411         }
412
413         cpu_khz = calculate_cpu_khz();
414         tsc_khz = cpu_khz;
415
416         if (!cpu_khz) {
417                 mark_tsc_unstable("could not calculate TSC khz");
418                 /*
419                  * We need to disable the TSC completely in this case
420                  * to prevent sched_clock() from using it.
421                  */
422                 tsc_disabled = 1;
423                 return;
424         }
425
426         lpj = ((u64)tsc_khz * 1000);
427         do_div(lpj, HZ);
428         lpj_tsc = lpj;
429
430         printk("Detected %lu.%03lu MHz processor.\n",
431                                 (unsigned long)cpu_khz / 1000,
432                                 (unsigned long)cpu_khz % 1000);
433
434         /*
435          * Secondary CPUs do not run through tsc_init(), so set up
436          * all the scale factors for all CPUs, assuming the same
437          * speed as the bootup CPU. (cpufreq notifiers will fix this
438          * up if their speed diverges)
439          */
440         for_each_possible_cpu(cpu)
441                 set_cyc2ns_scale(cpu_khz, cpu);
442
443         use_tsc_delay();
444
445         /* Check and install the TSC clocksource */
446         dmi_check_system(bad_tsc_dmi_table);
447
448         unsynchronized_tsc();
449         check_geode_tsc_reliable();
450         current_tsc_khz = tsc_khz;
451         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
452                                                         clocksource_tsc.shift);
453         /* lower the rating if we already know its unstable: */
454         if (check_tsc_unstable()) {
455                 clocksource_tsc.rating = 0;
456                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
457         }
458         clocksource_register(&clocksource_tsc);
459 }