Merge git://git.kernel.org/pub/scm/linux/kernel/git/joern/misc
[linux-2.6] / arch / i386 / kernel / tsc.c
1 /*
2  * This code largely moved from arch/i386/kernel/timer/timer_tsc.c
3  * which was originally moved from arch/i386/kernel/time.c.
4  * See comments there for proper credits.
5  */
6
7 #include <linux/clocksource.h>
8 #include <linux/workqueue.h>
9 #include <linux/cpufreq.h>
10 #include <linux/jiffies.h>
11 #include <linux/init.h>
12 #include <linux/dmi.h>
13
14 #include <asm/delay.h>
15 #include <asm/tsc.h>
16 #include <asm/io.h>
17 #include <asm/timer.h>
18
19 #include "mach_timer.h"
20
21 /*
22  * On some systems the TSC frequency does not
23  * change with the cpu frequency. So we need
24  * an extra value to store the TSC freq
25  */
26 unsigned int tsc_khz;
27
28 int tsc_disable;
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.\n");
35         return 1;
36 }
37 #else
38 /*
39  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
40  * in cpu/common.c
41  */
42 static int __init tsc_setup(char *str)
43 {
44         tsc_disable = 1;
45
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 static inline int check_tsc_unstable(void)
59 {
60         return tsc_unstable;
61 }
62
63 /* Accellerators for sched_clock()
64  * convert from cycles(64bits) => nanoseconds (64bits)
65  *  basic equation:
66  *              ns = cycles / (freq / ns_per_sec)
67  *              ns = cycles * (ns_per_sec / freq)
68  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
69  *              ns = cycles * (10^6 / cpu_khz)
70  *
71  *      Then we use scaling math (suggested by george@mvista.com) to get:
72  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
73  *              ns = cycles * cyc2ns_scale / SC
74  *
75  *      And since SC is a constant power of two, we can convert the div
76  *  into a shift.
77  *
78  *  We can use khz divisor instead of mhz to keep a better percision, since
79  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
80  *  (mathieu.desnoyers@polymtl.ca)
81  *
82  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
83  */
84 static unsigned long cyc2ns_scale __read_mostly;
85
86 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
87
88 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
89 {
90         cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
91 }
92
93 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
94 {
95         return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
96 }
97
98 /*
99  * Scheduler clock - returns current time in nanosec units.
100  */
101 unsigned long long sched_clock(void)
102 {
103         unsigned long long this_offset;
104
105         /*
106          * Fall back to jiffies if there's no TSC available:
107          */
108         if (unlikely(tsc_disable))
109                 /* No locking but a rare wrong value is not a big deal: */
110                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
111
112         /* read the Time Stamp Counter: */
113         get_scheduled_cycles(this_offset);
114
115         /* return the value in ns */
116         return cycles_2_ns(this_offset);
117 }
118
119 unsigned long native_calculate_cpu_khz(void)
120 {
121         unsigned long long start, end;
122         unsigned long count;
123         u64 delta64;
124         int i;
125         unsigned long flags;
126
127         local_irq_save(flags);
128
129         /* run 3 times to ensure the cache is warm */
130         for (i = 0; i < 3; i++) {
131                 mach_prepare_counter();
132                 rdtscll(start);
133                 mach_countup(&count);
134                 rdtscll(end);
135         }
136         /*
137          * Error: ECTCNEVERSET
138          * The CTC wasn't reliable: we got a hit on the very first read,
139          * or the CPU was so fast/slow that the quotient wouldn't fit in
140          * 32 bits..
141          */
142         if (count <= 1)
143                 goto err;
144
145         delta64 = end - start;
146
147         /* cpu freq too fast: */
148         if (delta64 > (1ULL<<32))
149                 goto err;
150
151         /* cpu freq too slow: */
152         if (delta64 <= CALIBRATE_TIME_MSEC)
153                 goto err;
154
155         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
156         do_div(delta64,CALIBRATE_TIME_MSEC);
157
158         local_irq_restore(flags);
159         return (unsigned long)delta64;
160 err:
161         local_irq_restore(flags);
162         return 0;
163 }
164
165 int recalibrate_cpu_khz(void)
166 {
167 #ifndef CONFIG_SMP
168         unsigned long cpu_khz_old = cpu_khz;
169
170         if (cpu_has_tsc) {
171                 cpu_khz = calculate_cpu_khz();
172                 tsc_khz = cpu_khz;
173                 cpu_data[0].loops_per_jiffy =
174                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
175                                         cpu_khz_old, cpu_khz);
176                 return 0;
177         } else
178                 return -ENODEV;
179 #else
180         return -ENODEV;
181 #endif
182 }
183
184 EXPORT_SYMBOL(recalibrate_cpu_khz);
185
186 #ifdef CONFIG_CPU_FREQ
187
188 /*
189  * if the CPU frequency is scaled, TSC-based delays will need a different
190  * loops_per_jiffy value to function properly.
191  */
192 static unsigned int ref_freq = 0;
193 static unsigned long loops_per_jiffy_ref = 0;
194 static unsigned long cpu_khz_ref = 0;
195
196 static int
197 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
198 {
199         struct cpufreq_freqs *freq = data;
200
201         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
202                 write_seqlock_irq(&xtime_lock);
203
204         if (!ref_freq) {
205                 if (!freq->old){
206                         ref_freq = freq->new;
207                         goto end;
208                 }
209                 ref_freq = freq->old;
210                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
211                 cpu_khz_ref = cpu_khz;
212         }
213
214         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
215             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
216             (val == CPUFREQ_RESUMECHANGE)) {
217                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
218                         cpu_data[freq->cpu].loops_per_jiffy =
219                                 cpufreq_scale(loops_per_jiffy_ref,
220                                                 ref_freq, freq->new);
221
222                 if (cpu_khz) {
223
224                         if (num_online_cpus() == 1)
225                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
226                                                 ref_freq, freq->new);
227                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
228                                 tsc_khz = cpu_khz;
229                                 set_cyc2ns_scale(cpu_khz);
230                                 /*
231                                  * TSC based sched_clock turns
232                                  * to junk w/ cpufreq
233                                  */
234                                 mark_tsc_unstable();
235                         }
236                 }
237         }
238 end:
239         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
240                 write_sequnlock_irq(&xtime_lock);
241
242         return 0;
243 }
244
245 static struct notifier_block time_cpufreq_notifier_block = {
246         .notifier_call  = time_cpufreq_notifier
247 };
248
249 static int __init cpufreq_tsc(void)
250 {
251         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
252                                          CPUFREQ_TRANSITION_NOTIFIER);
253 }
254 core_initcall(cpufreq_tsc);
255
256 #endif
257
258 /* clock source code */
259
260 static unsigned long current_tsc_khz = 0;
261
262 static cycle_t read_tsc(void)
263 {
264         cycle_t ret;
265
266         rdtscll(ret);
267
268         return ret;
269 }
270
271 static struct clocksource clocksource_tsc = {
272         .name                   = "tsc",
273         .rating                 = 300,
274         .read                   = read_tsc,
275         .mask                   = CLOCKSOURCE_MASK(64),
276         .mult                   = 0, /* to be set */
277         .shift                  = 22,
278         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
279                                   CLOCK_SOURCE_MUST_VERIFY,
280 };
281
282 void mark_tsc_unstable(void)
283 {
284         if (!tsc_unstable) {
285                 tsc_unstable = 1;
286                 /* Can be called before registration */
287                 if (clocksource_tsc.mult)
288                         clocksource_change_rating(&clocksource_tsc, 0);
289                 else
290                         clocksource_tsc.rating = 0;
291         }
292 }
293 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
294
295 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
296 {
297         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
298                        d->ident);
299         tsc_unstable = 1;
300         return 0;
301 }
302
303 /* List of systems that have known TSC problems */
304 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
305         {
306          .callback = dmi_mark_tsc_unstable,
307          .ident = "IBM Thinkpad 380XD",
308          .matches = {
309                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
310                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
311                      },
312          },
313          {}
314 };
315
316 /*
317  * Make an educated guess if the TSC is trustworthy and synchronized
318  * over all CPUs.
319  */
320 __cpuinit int unsynchronized_tsc(void)
321 {
322         if (!cpu_has_tsc || tsc_unstable)
323                 return 1;
324         /*
325          * Intel systems are normally all synchronized.
326          * Exceptions must mark TSC as unstable:
327          */
328         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
329                 /* assume multi socket systems are not synchronized: */
330                 if (num_possible_cpus() > 1)
331                         tsc_unstable = 1;
332         }
333         return tsc_unstable;
334 }
335
336 /*
337  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
338  */
339 #ifdef CONFIG_MGEODE_LX
340 /* RTSC counts during suspend */
341 #define RTSC_SUSP 0x100
342
343 static void __init check_geode_tsc_reliable(void)
344 {
345         unsigned long val;
346
347         rdmsrl(MSR_GEODE_BUSCONT_CONF0, val);
348         if ((val & RTSC_SUSP))
349                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
350 }
351 #else
352 static inline void check_geode_tsc_reliable(void) { }
353 #endif
354
355
356 void __init tsc_init(void)
357 {
358         if (!cpu_has_tsc || tsc_disable)
359                 goto out_no_tsc;
360
361         cpu_khz = calculate_cpu_khz();
362         tsc_khz = cpu_khz;
363
364         if (!cpu_khz)
365                 goto out_no_tsc;
366
367         printk("Detected %lu.%03lu MHz processor.\n",
368                                 (unsigned long)cpu_khz / 1000,
369                                 (unsigned long)cpu_khz % 1000);
370
371         set_cyc2ns_scale(cpu_khz);
372         use_tsc_delay();
373
374         /* Check and install the TSC clocksource */
375         dmi_check_system(bad_tsc_dmi_table);
376
377         unsynchronized_tsc();
378         check_geode_tsc_reliable();
379         current_tsc_khz = tsc_khz;
380         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
381                                                         clocksource_tsc.shift);
382         /* lower the rating if we already know its unstable: */
383         if (check_tsc_unstable()) {
384                 clocksource_tsc.rating = 0;
385                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
386         }
387         clocksource_register(&clocksource_tsc);
388
389         return;
390
391 out_no_tsc:
392         /*
393          * Set the tsc_disable flag if there's no TSC support, this
394          * makes it a fast flag for the kernel to see whether it
395          * should be using the TSC.
396          */
397         tsc_disable = 1;
398 }