Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/avi/kvm
[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 static int tsc_enabled;
22
23 /*
24  * On some systems the TSC frequency does not
25  * change with the cpu frequency. So we need
26  * an extra value to store the TSC freq
27  */
28 unsigned int tsc_khz;
29
30 int tsc_disable;
31
32 #ifdef CONFIG_X86_TSC
33 static int __init tsc_setup(char *str)
34 {
35         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
36                                 "cannot disable TSC.\n");
37         return 1;
38 }
39 #else
40 /*
41  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
42  * in cpu/common.c
43  */
44 static int __init tsc_setup(char *str)
45 {
46         tsc_disable = 1;
47
48         return 1;
49 }
50 #endif
51
52 __setup("notsc", tsc_setup);
53
54 /*
55  * code to mark and check if the TSC is unstable
56  * due to cpufreq or due to unsynced TSCs
57  */
58 static int tsc_unstable;
59
60 static inline int check_tsc_unstable(void)
61 {
62         return tsc_unstable;
63 }
64
65 /* Accellerators for sched_clock()
66  * convert from cycles(64bits) => nanoseconds (64bits)
67  *  basic equation:
68  *              ns = cycles / (freq / ns_per_sec)
69  *              ns = cycles * (ns_per_sec / freq)
70  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
71  *              ns = cycles * (10^6 / cpu_khz)
72  *
73  *      Then we use scaling math (suggested by george@mvista.com) to get:
74  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
75  *              ns = cycles * cyc2ns_scale / SC
76  *
77  *      And since SC is a constant power of two, we can convert the div
78  *  into a shift.
79  *
80  *  We can use khz divisor instead of mhz to keep a better percision, since
81  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
82  *  (mathieu.desnoyers@polymtl.ca)
83  *
84  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
85  */
86 static unsigned long cyc2ns_scale __read_mostly;
87
88 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
89
90 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
91 {
92         cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
93 }
94
95 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
96 {
97         return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
98 }
99
100 /*
101  * Scheduler clock - returns current time in nanosec units.
102  */
103 unsigned long long sched_clock(void)
104 {
105         unsigned long long this_offset;
106
107         /*
108          * Fall back to jiffies if there's no TSC available:
109          */
110         if (unlikely(!tsc_enabled))
111                 /* No locking but a rare wrong value is not a big deal: */
112                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
113
114         /* read the Time Stamp Counter: */
115         get_scheduled_cycles(this_offset);
116
117         /* return the value in ns */
118         return cycles_2_ns(this_offset);
119 }
120
121 unsigned long native_calculate_cpu_khz(void)
122 {
123         unsigned long long start, end;
124         unsigned long count;
125         u64 delta64;
126         int i;
127         unsigned long flags;
128
129         local_irq_save(flags);
130
131         /* run 3 times to ensure the cache is warm */
132         for (i = 0; i < 3; i++) {
133                 mach_prepare_counter();
134                 rdtscll(start);
135                 mach_countup(&count);
136                 rdtscll(end);
137         }
138         /*
139          * Error: ECTCNEVERSET
140          * The CTC wasn't reliable: we got a hit on the very first read,
141          * or the CPU was so fast/slow that the quotient wouldn't fit in
142          * 32 bits..
143          */
144         if (count <= 1)
145                 goto err;
146
147         delta64 = end - start;
148
149         /* cpu freq too fast: */
150         if (delta64 > (1ULL<<32))
151                 goto err;
152
153         /* cpu freq too slow: */
154         if (delta64 <= CALIBRATE_TIME_MSEC)
155                 goto err;
156
157         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
158         do_div(delta64,CALIBRATE_TIME_MSEC);
159
160         local_irq_restore(flags);
161         return (unsigned long)delta64;
162 err:
163         local_irq_restore(flags);
164         return 0;
165 }
166
167 int recalibrate_cpu_khz(void)
168 {
169 #ifndef CONFIG_SMP
170         unsigned long cpu_khz_old = cpu_khz;
171
172         if (cpu_has_tsc) {
173                 cpu_khz = calculate_cpu_khz();
174                 tsc_khz = cpu_khz;
175                 cpu_data[0].loops_per_jiffy =
176                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
177                                         cpu_khz_old, cpu_khz);
178                 return 0;
179         } else
180                 return -ENODEV;
181 #else
182         return -ENODEV;
183 #endif
184 }
185
186 EXPORT_SYMBOL(recalibrate_cpu_khz);
187
188 #ifdef CONFIG_CPU_FREQ
189
190 /*
191  * if the CPU frequency is scaled, TSC-based delays will need a different
192  * loops_per_jiffy value to function properly.
193  */
194 static unsigned int ref_freq = 0;
195 static unsigned long loops_per_jiffy_ref = 0;
196 static unsigned long cpu_khz_ref = 0;
197
198 static int
199 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
200 {
201         struct cpufreq_freqs *freq = data;
202
203         if (!ref_freq) {
204                 if (!freq->old){
205                         ref_freq = freq->new;
206                         return 0;
207                 }
208                 ref_freq = freq->old;
209                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
210                 cpu_khz_ref = cpu_khz;
211         }
212
213         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
214             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
215             (val == CPUFREQ_RESUMECHANGE)) {
216                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
217                         cpu_data[freq->cpu].loops_per_jiffy =
218                                 cpufreq_scale(loops_per_jiffy_ref,
219                                                 ref_freq, freq->new);
220
221                 if (cpu_khz) {
222
223                         if (num_online_cpus() == 1)
224                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
225                                                 ref_freq, freq->new);
226                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
227                                 tsc_khz = cpu_khz;
228                                 set_cyc2ns_scale(cpu_khz);
229                                 /*
230                                  * TSC based sched_clock turns
231                                  * to junk w/ cpufreq
232                                  */
233                                 mark_tsc_unstable("cpufreq changes");
234                         }
235                 }
236         }
237
238         return 0;
239 }
240
241 static struct notifier_block time_cpufreq_notifier_block = {
242         .notifier_call  = time_cpufreq_notifier
243 };
244
245 static int __init cpufreq_tsc(void)
246 {
247         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
248                                          CPUFREQ_TRANSITION_NOTIFIER);
249 }
250 core_initcall(cpufreq_tsc);
251
252 #endif
253
254 /* clock source code */
255
256 static unsigned long current_tsc_khz = 0;
257
258 static cycle_t read_tsc(void)
259 {
260         cycle_t ret;
261
262         rdtscll(ret);
263
264         return ret;
265 }
266
267 static struct clocksource clocksource_tsc = {
268         .name                   = "tsc",
269         .rating                 = 300,
270         .read                   = read_tsc,
271         .mask                   = CLOCKSOURCE_MASK(64),
272         .mult                   = 0, /* to be set */
273         .shift                  = 22,
274         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
275                                   CLOCK_SOURCE_MUST_VERIFY,
276 };
277
278 void mark_tsc_unstable(char *reason)
279 {
280         if (!tsc_unstable) {
281                 tsc_unstable = 1;
282                 tsc_enabled = 0;
283                 printk("Marking TSC unstable due to: %s.\n", reason);
284                 /* Can be called before registration */
285                 if (clocksource_tsc.mult)
286                         clocksource_change_rating(&clocksource_tsc, 0);
287                 else
288                         clocksource_tsc.rating = 0;
289         }
290 }
291 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
292
293 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
294 {
295         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
296                        d->ident);
297         tsc_unstable = 1;
298         return 0;
299 }
300
301 /* List of systems that have known TSC problems */
302 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
303         {
304          .callback = dmi_mark_tsc_unstable,
305          .ident = "IBM Thinkpad 380XD",
306          .matches = {
307                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
308                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
309                      },
310          },
311          {}
312 };
313
314 /*
315  * Make an educated guess if the TSC is trustworthy and synchronized
316  * over all CPUs.
317  */
318 __cpuinit int unsynchronized_tsc(void)
319 {
320         if (!cpu_has_tsc || tsc_unstable)
321                 return 1;
322         /*
323          * Intel systems are normally all synchronized.
324          * Exceptions must mark TSC as unstable:
325          */
326         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
327                 /* assume multi socket systems are not synchronized: */
328                 if (num_possible_cpus() > 1)
329                         tsc_unstable = 1;
330         }
331         return tsc_unstable;
332 }
333
334 /*
335  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
336  */
337 #ifdef CONFIG_MGEODE_LX
338 /* RTSC counts during suspend */
339 #define RTSC_SUSP 0x100
340
341 static void __init check_geode_tsc_reliable(void)
342 {
343         unsigned long val;
344
345         rdmsrl(MSR_GEODE_BUSCONT_CONF0, val);
346         if ((val & RTSC_SUSP))
347                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
348 }
349 #else
350 static inline void check_geode_tsc_reliable(void) { }
351 #endif
352
353
354 void __init tsc_init(void)
355 {
356         if (!cpu_has_tsc || tsc_disable)
357                 goto out_no_tsc;
358
359         cpu_khz = calculate_cpu_khz();
360         tsc_khz = cpu_khz;
361
362         if (!cpu_khz)
363                 goto out_no_tsc;
364
365         printk("Detected %lu.%03lu MHz processor.\n",
366                                 (unsigned long)cpu_khz / 1000,
367                                 (unsigned long)cpu_khz % 1000);
368
369         set_cyc2ns_scale(cpu_khz);
370         use_tsc_delay();
371
372         /* Check and install the TSC clocksource */
373         dmi_check_system(bad_tsc_dmi_table);
374
375         unsynchronized_tsc();
376         check_geode_tsc_reliable();
377         current_tsc_khz = tsc_khz;
378         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
379                                                         clocksource_tsc.shift);
380         /* lower the rating if we already know its unstable: */
381         if (check_tsc_unstable()) {
382                 clocksource_tsc.rating = 0;
383                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
384         } else
385                 tsc_enabled = 1;
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 }