x86, suspend, acpi: correct and add comments about Big Real Mode
[pandora-kernel.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/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 /* native_sched_clock() is called before tsc_init(), so
18    we must start with the TSC soft disabled to prevent
19    erroneous rdtsc usage on !cpu_has_tsc processors */
20 static int tsc_disabled = -1;
21
22 /*
23  * On some systems the TSC frequency does not
24  * change with the cpu frequency. So we need
25  * an extra value to store the TSC freq
26  */
27 unsigned int tsc_khz;
28 EXPORT_SYMBOL_GPL(tsc_khz);
29
30 #ifdef CONFIG_X86_TSC
31 static int __init tsc_setup(char *str)
32 {
33         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
34                "cannot disable TSC completely.\n");
35         tsc_disabled = 1;
36         return 1;
37 }
38 #else
39 /*
40  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
41  * in cpu/common.c
42  */
43 static int __init tsc_setup(char *str)
44 {
45         setup_clear_cpu_cap(X86_FEATURE_TSC);
46         return 1;
47 }
48 #endif
49
50 __setup("notsc", tsc_setup);
51
52 /*
53  * code to mark and check if the TSC is unstable
54  * due to cpufreq or due to unsynced TSCs
55  */
56 static int tsc_unstable;
57
58 int check_tsc_unstable(void)
59 {
60         return tsc_unstable;
61 }
62 EXPORT_SYMBOL_GPL(check_tsc_unstable);
63
64 /* Accelerators for sched_clock()
65  * convert from cycles(64bits) => nanoseconds (64bits)
66  *  basic equation:
67  *              ns = cycles / (freq / ns_per_sec)
68  *              ns = cycles * (ns_per_sec / freq)
69  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
70  *              ns = cycles * (10^6 / cpu_khz)
71  *
72  *      Then we use scaling math (suggested by george@mvista.com) to get:
73  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
74  *              ns = cycles * cyc2ns_scale / SC
75  *
76  *      And since SC is a constant power of two, we can convert the div
77  *  into a shift.
78  *
79  *  We can use khz divisor instead of mhz to keep a better precision, since
80  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
81  *  (mathieu.desnoyers@polymtl.ca)
82  *
83  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
84  */
85
86 DEFINE_PER_CPU(unsigned long, cyc2ns);
87
88 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
89 {
90         unsigned long long tsc_now, ns_now;
91         unsigned long flags, *scale;
92
93         local_irq_save(flags);
94         sched_clock_idle_sleep_event();
95
96         scale = &per_cpu(cyc2ns, cpu);
97
98         rdtscll(tsc_now);
99         ns_now = __cycles_2_ns(tsc_now);
100
101         if (cpu_khz)
102                 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
103
104         /*
105          * Start smoothly with the new frequency:
106          */
107         sched_clock_idle_wakeup_event(0);
108         local_irq_restore(flags);
109 }
110
111 /*
112  * Scheduler clock - returns current time in nanosec units.
113  */
114 unsigned long long native_sched_clock(void)
115 {
116         unsigned long long this_offset;
117
118         /*
119          * Fall back to jiffies if there's no TSC available:
120          * ( But note that we still use it if the TSC is marked
121          *   unstable. We do this because unlike Time Of Day,
122          *   the scheduler clock tolerates small errors and it's
123          *   very important for it to be as fast as the platform
124          *   can achive it. )
125          */
126         if (unlikely(tsc_disabled))
127                 /* No locking but a rare wrong value is not a big deal: */
128                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
129
130         /* read the Time Stamp Counter: */
131         rdtscll(this_offset);
132
133         /* return the value in ns */
134         return cycles_2_ns(this_offset);
135 }
136
137 /* We need to define a real function for sched_clock, to override the
138    weak default version */
139 #ifdef CONFIG_PARAVIRT
140 unsigned long long sched_clock(void)
141 {
142         return paravirt_sched_clock();
143 }
144 #else
145 unsigned long long sched_clock(void)
146         __attribute__((alias("native_sched_clock")));
147 #endif
148
149 unsigned long native_calculate_cpu_khz(void)
150 {
151         unsigned long long start, end;
152         unsigned long count;
153         u64 delta64 = (u64)ULLONG_MAX;
154         int i;
155         unsigned long flags;
156
157         local_irq_save(flags);
158
159         /* run 3 times to ensure the cache is warm and to get an accurate reading */
160         for (i = 0; i < 3; i++) {
161                 mach_prepare_counter();
162                 rdtscll(start);
163                 mach_countup(&count);
164                 rdtscll(end);
165
166                 /*
167                  * Error: ECTCNEVERSET
168                  * The CTC wasn't reliable: we got a hit on the very first read,
169                  * or the CPU was so fast/slow that the quotient wouldn't fit in
170                  * 32 bits..
171                  */
172                 if (count <= 1)
173                         continue;
174
175                 /* cpu freq too slow: */
176                 if ((end - start) <= CALIBRATE_TIME_MSEC)
177                         continue;
178
179                 /*
180                  * We want the minimum time of all runs in case one of them
181                  * is inaccurate due to SMI or other delay
182                  */
183                 delta64 = min(delta64, (end - start));
184         }
185
186         /* cpu freq too fast (or every run was bad): */
187         if (delta64 > (1ULL<<32))
188                 goto err;
189
190         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
191         do_div(delta64,CALIBRATE_TIME_MSEC);
192
193         local_irq_restore(flags);
194         return (unsigned long)delta64;
195 err:
196         local_irq_restore(flags);
197         return 0;
198 }
199
200 int recalibrate_cpu_khz(void)
201 {
202 #ifndef CONFIG_SMP
203         unsigned long cpu_khz_old = cpu_khz;
204
205         if (cpu_has_tsc) {
206                 cpu_khz = calculate_cpu_khz();
207                 tsc_khz = cpu_khz;
208                 cpu_data(0).loops_per_jiffy =
209                         cpufreq_scale(cpu_data(0).loops_per_jiffy,
210                                         cpu_khz_old, cpu_khz);
211                 return 0;
212         } else
213                 return -ENODEV;
214 #else
215         return -ENODEV;
216 #endif
217 }
218
219 EXPORT_SYMBOL(recalibrate_cpu_khz);
220
221 #ifdef CONFIG_CPU_FREQ
222
223 /*
224  * if the CPU frequency is scaled, TSC-based delays will need a different
225  * loops_per_jiffy value to function properly.
226  */
227 static unsigned int ref_freq;
228 static unsigned long loops_per_jiffy_ref;
229 static unsigned long cpu_khz_ref;
230
231 static int
232 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
233 {
234         struct cpufreq_freqs *freq = data;
235
236         if (!ref_freq) {
237                 if (!freq->old){
238                         ref_freq = freq->new;
239                         return 0;
240                 }
241                 ref_freq = freq->old;
242                 loops_per_jiffy_ref = cpu_data(freq->cpu).loops_per_jiffy;
243                 cpu_khz_ref = cpu_khz;
244         }
245
246         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
247             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
248             (val == CPUFREQ_RESUMECHANGE)) {
249                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
250                         cpu_data(freq->cpu).loops_per_jiffy =
251                                 cpufreq_scale(loops_per_jiffy_ref,
252                                                 ref_freq, freq->new);
253
254                 if (cpu_khz) {
255
256                         if (num_online_cpus() == 1)
257                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
258                                                 ref_freq, freq->new);
259                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
260                                 tsc_khz = cpu_khz;
261                                 set_cyc2ns_scale(cpu_khz, freq->cpu);
262                                 /*
263                                  * TSC based sched_clock turns
264                                  * to junk w/ cpufreq
265                                  */
266                                 mark_tsc_unstable("cpufreq changes");
267                         }
268                 }
269         }
270
271         return 0;
272 }
273
274 static struct notifier_block time_cpufreq_notifier_block = {
275         .notifier_call  = time_cpufreq_notifier
276 };
277
278 static int __init cpufreq_tsc(void)
279 {
280         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
281                                          CPUFREQ_TRANSITION_NOTIFIER);
282 }
283 core_initcall(cpufreq_tsc);
284
285 #endif
286
287 /* clock source code */
288
289 static unsigned long current_tsc_khz;
290 static struct clocksource clocksource_tsc;
291
292 /*
293  * We compare the TSC to the cycle_last value in the clocksource
294  * structure to avoid a nasty time-warp issue. This can be observed in
295  * a very small window right after one CPU updated cycle_last under
296  * xtime lock and the other CPU reads a TSC value which is smaller
297  * than the cycle_last reference value due to a TSC which is slighty
298  * behind. This delta is nowhere else observable, but in that case it
299  * results in a forward time jump in the range of hours due to the
300  * unsigned delta calculation of the time keeping core code, which is
301  * necessary to support wrapping clocksources like pm timer.
302  */
303 static cycle_t read_tsc(void)
304 {
305         cycle_t ret;
306
307         rdtscll(ret);
308
309         return ret >= clocksource_tsc.cycle_last ?
310                 ret : clocksource_tsc.cycle_last;
311 }
312
313 static struct clocksource clocksource_tsc = {
314         .name                   = "tsc",
315         .rating                 = 300,
316         .read                   = read_tsc,
317         .mask                   = CLOCKSOURCE_MASK(64),
318         .mult                   = 0, /* to be set */
319         .shift                  = 22,
320         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
321                                   CLOCK_SOURCE_MUST_VERIFY,
322 };
323
324 void mark_tsc_unstable(char *reason)
325 {
326         if (!tsc_unstable) {
327                 tsc_unstable = 1;
328                 printk("Marking TSC unstable due to: %s.\n", reason);
329                 /* Can be called before registration */
330                 if (clocksource_tsc.mult)
331                         clocksource_change_rating(&clocksource_tsc, 0);
332                 else
333                         clocksource_tsc.rating = 0;
334         }
335 }
336 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
337
338 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
339 {
340         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
341                d->ident);
342         tsc_unstable = 1;
343         return 0;
344 }
345
346 /* List of systems that have known TSC problems */
347 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
348         {
349          .callback = dmi_mark_tsc_unstable,
350          .ident = "IBM Thinkpad 380XD",
351          .matches = {
352                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
353                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
354                      },
355          },
356          {}
357 };
358
359 /*
360  * Make an educated guess if the TSC is trustworthy and synchronized
361  * over all CPUs.
362  */
363 __cpuinit int unsynchronized_tsc(void)
364 {
365         if (!cpu_has_tsc || tsc_unstable)
366                 return 1;
367
368         /* Anything with constant TSC should be synchronized */
369         if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
370                 return 0;
371
372         /*
373          * Intel systems are normally all synchronized.
374          * Exceptions must mark TSC as unstable:
375          */
376         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
377                 /* assume multi socket systems are not synchronized: */
378                 if (num_possible_cpus() > 1)
379                         tsc_unstable = 1;
380         }
381         return tsc_unstable;
382 }
383
384 /*
385  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
386  */
387 #ifdef CONFIG_MGEODE_LX
388 /* RTSC counts during suspend */
389 #define RTSC_SUSP 0x100
390
391 static void __init check_geode_tsc_reliable(void)
392 {
393         unsigned long res_low, res_high;
394
395         rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
396         if (res_low & RTSC_SUSP)
397                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
398 }
399 #else
400 static inline void check_geode_tsc_reliable(void) { }
401 #endif
402
403
404 void __init tsc_init(void)
405 {
406         int cpu;
407
408         if (!cpu_has_tsc || tsc_disabled > 0)
409                 return;
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                 return;
417         }
418
419         /* now allow native_sched_clock() to use rdtsc */
420         tsc_disabled = 0;
421
422         printk("Detected %lu.%03lu MHz processor.\n",
423                                 (unsigned long)cpu_khz / 1000,
424                                 (unsigned long)cpu_khz % 1000);
425
426         /*
427          * Secondary CPUs do not run through tsc_init(), so set up
428          * all the scale factors for all CPUs, assuming the same
429          * speed as the bootup CPU. (cpufreq notifiers will fix this
430          * up if their speed diverges)
431          */
432         for_each_possible_cpu(cpu)
433                 set_cyc2ns_scale(cpu_khz, cpu);
434
435         use_tsc_delay();
436
437         /* Check and install the TSC clocksource */
438         dmi_check_system(bad_tsc_dmi_table);
439
440         unsynchronized_tsc();
441         check_geode_tsc_reliable();
442         current_tsc_khz = tsc_khz;
443         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
444                                                         clocksource_tsc.shift);
445         /* lower the rating if we already know its unstable: */
446         if (check_tsc_unstable()) {
447                 clocksource_tsc.rating = 0;
448                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
449         }
450         clocksource_register(&clocksource_tsc);
451 }