Merge git://git.kernel.org/pub/scm/linux/kernel/git/lethal/sh-2.6.24
[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
9 #include <asm/delay.h>
10 #include <asm/tsc.h>
11 #include <asm/io.h>
12 #include <asm/timer.h>
13
14 #include "mach_timer.h"
15
16 static int tsc_enabled;
17
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;
24 EXPORT_SYMBOL_GPL(tsc_khz);
25
26 int tsc_disable;
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
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 unsigned long cyc2ns_scale __read_mostly;
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
92 /*
93  * Scheduler clock - returns current time in nanosec units.
94  */
95 unsigned long long native_sched_clock(void)
96 {
97         unsigned long long this_offset;
98
99         /*
100          * Fall back to jiffies if there's no TSC available:
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. )
106          */
107         if (unlikely(!tsc_enabled && !tsc_unstable))
108                 /* No locking but a rare wrong value is not a big deal: */
109                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
110
111         /* read the Time Stamp Counter: */
112         rdtscll(this_offset);
113
114         /* return the value in ns */
115         return cycles_2_ns(this_offset);
116 }
117
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
130 unsigned long native_calculate_cpu_khz(void)
131 {
132         unsigned long long start, end;
133         unsigned long count;
134         u64 delta64 = (u64)ULLONG_MAX;
135         int i;
136         unsigned long flags;
137
138         local_irq_save(flags);
139
140         /* run 3 times to ensure the cache is warm and to get an accurate reading */
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                         continue;
155
156                 /* cpu freq too slow: */
157                 if ((end - start) <= CALIBRATE_TIME_MSEC)
158                         continue;
159
160                 /*
161                  * We want the minimum time of all runs in case one of them
162                  * is inaccurate due to SMI or other delay
163                  */
164                 delta64 = min(delta64, (end - start));
165         }
166
167         /* cpu freq too fast (or every run was bad): */
168         if (delta64 > (1ULL<<32))
169                 goto err;
170
171         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
172         do_div(delta64,CALIBRATE_TIME_MSEC);
173
174         local_irq_restore(flags);
175         return (unsigned long)delta64;
176 err:
177         local_irq_restore(flags);
178         return 0;
179 }
180
181 int recalibrate_cpu_khz(void)
182 {
183 #ifndef CONFIG_SMP
184         unsigned long cpu_khz_old = cpu_khz;
185
186         if (cpu_has_tsc) {
187                 cpu_khz = calculate_cpu_khz();
188                 tsc_khz = cpu_khz;
189                 cpu_data(0).loops_per_jiffy =
190                         cpufreq_scale(cpu_data(0).loops_per_jiffy,
191                                         cpu_khz_old, cpu_khz);
192                 return 0;
193         } else
194                 return -ENODEV;
195 #else
196         return -ENODEV;
197 #endif
198 }
199
200 EXPORT_SYMBOL(recalibrate_cpu_khz);
201
202 #ifdef CONFIG_CPU_FREQ
203
204 /*
205  * if the CPU frequency is scaled, TSC-based delays will need a different
206  * loops_per_jiffy value to function properly.
207  */
208 static unsigned int ref_freq = 0;
209 static unsigned long loops_per_jiffy_ref = 0;
210 static unsigned long cpu_khz_ref = 0;
211
212 static int
213 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
214 {
215         struct cpufreq_freqs *freq = data;
216
217         if (!ref_freq) {
218                 if (!freq->old){
219                         ref_freq = freq->new;
220                         return 0;
221                 }
222                 ref_freq = freq->old;
223                 loops_per_jiffy_ref = cpu_data(freq->cpu).loops_per_jiffy;
224                 cpu_khz_ref = cpu_khz;
225         }
226
227         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
228             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
229             (val == CPUFREQ_RESUMECHANGE)) {
230                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
231                         cpu_data(freq->cpu).loops_per_jiffy =
232                                 cpufreq_scale(loops_per_jiffy_ref,
233                                                 ref_freq, freq->new);
234
235                 if (cpu_khz) {
236
237                         if (num_online_cpus() == 1)
238                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
239                                                 ref_freq, freq->new);
240                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
241                                 tsc_khz = cpu_khz;
242                                 set_cyc2ns_scale(cpu_khz);
243                                 /*
244                                  * TSC based sched_clock turns
245                                  * to junk w/ cpufreq
246                                  */
247                                 mark_tsc_unstable("cpufreq changes");
248                         }
249                 }
250         }
251
252         return 0;
253 }
254
255 static struct notifier_block time_cpufreq_notifier_block = {
256         .notifier_call  = time_cpufreq_notifier
257 };
258
259 static int __init cpufreq_tsc(void)
260 {
261         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
262                                          CPUFREQ_TRANSITION_NOTIFIER);
263 }
264 core_initcall(cpufreq_tsc);
265
266 #endif
267
268 /* clock source code */
269
270 static unsigned long current_tsc_khz = 0;
271
272 static cycle_t read_tsc(void)
273 {
274         cycle_t ret;
275
276         rdtscll(ret);
277
278         return ret;
279 }
280
281 static struct clocksource clocksource_tsc = {
282         .name                   = "tsc",
283         .rating                 = 300,
284         .read                   = read_tsc,
285         .mask                   = CLOCKSOURCE_MASK(64),
286         .mult                   = 0, /* to be set */
287         .shift                  = 22,
288         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
289                                   CLOCK_SOURCE_MUST_VERIFY,
290 };
291
292 void mark_tsc_unstable(char *reason)
293 {
294         if (!tsc_unstable) {
295                 tsc_unstable = 1;
296                 tsc_enabled = 0;
297                 printk("Marking TSC unstable due to: %s.\n", reason);
298                 /* Can be called before registration */
299                 if (clocksource_tsc.mult)
300                         clocksource_change_rating(&clocksource_tsc, 0);
301                 else
302                         clocksource_tsc.rating = 0;
303         }
304 }
305 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
306
307 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
308 {
309         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
310                        d->ident);
311         tsc_unstable = 1;
312         return 0;
313 }
314
315 /* List of systems that have known TSC problems */
316 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
317         {
318          .callback = dmi_mark_tsc_unstable,
319          .ident = "IBM Thinkpad 380XD",
320          .matches = {
321                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
322                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
323                      },
324          },
325          {}
326 };
327
328 /*
329  * Make an educated guess if the TSC is trustworthy and synchronized
330  * over all CPUs.
331  */
332 __cpuinit int unsynchronized_tsc(void)
333 {
334         if (!cpu_has_tsc || tsc_unstable)
335                 return 1;
336         /*
337          * Intel systems are normally all synchronized.
338          * Exceptions must mark TSC as unstable:
339          */
340         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
341                 /* assume multi socket systems are not synchronized: */
342                 if (num_possible_cpus() > 1)
343                         tsc_unstable = 1;
344         }
345         return tsc_unstable;
346 }
347
348 /*
349  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
350  */
351 #ifdef CONFIG_MGEODE_LX
352 /* RTSC counts during suspend */
353 #define RTSC_SUSP 0x100
354
355 static void __init check_geode_tsc_reliable(void)
356 {
357         unsigned long res_low, res_high;
358
359         rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
360         if (res_low & RTSC_SUSP)
361                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
362 }
363 #else
364 static inline void check_geode_tsc_reliable(void) { }
365 #endif
366
367
368 void __init tsc_init(void)
369 {
370         if (!cpu_has_tsc || tsc_disable)
371                 goto out_no_tsc;
372
373         cpu_khz = calculate_cpu_khz();
374         tsc_khz = cpu_khz;
375
376         if (!cpu_khz)
377                 goto out_no_tsc;
378
379         printk("Detected %lu.%03lu MHz processor.\n",
380                                 (unsigned long)cpu_khz / 1000,
381                                 (unsigned long)cpu_khz % 1000);
382
383         set_cyc2ns_scale(cpu_khz);
384         use_tsc_delay();
385
386         /* Check and install the TSC clocksource */
387         dmi_check_system(bad_tsc_dmi_table);
388
389         unsynchronized_tsc();
390         check_geode_tsc_reliable();
391         current_tsc_khz = tsc_khz;
392         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
393                                                         clocksource_tsc.shift);
394         /* lower the rating if we already know its unstable: */
395         if (check_tsc_unstable()) {
396                 clocksource_tsc.rating = 0;
397                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
398         } else
399                 tsc_enabled = 1;
400
401         clocksource_register(&clocksource_tsc);
402
403         return;
404
405 out_no_tsc:
406         /*
407          * Set the tsc_disable flag if there's no TSC support, this
408          * makes it a fast flag for the kernel to see whether it
409          * should be using the TSC.
410          */
411         tsc_disable = 1;
412 }