Merge rsync://rsync.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[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
18 #include "mach_timer.h"
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
27 int tsc_disable;
28
29 #ifdef CONFIG_X86_TSC
30 static int __init tsc_setup(char *str)
31 {
32         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
33                                 "cannot disable TSC.\n");
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         tsc_disable = 1;
44
45         return 1;
46 }
47 #endif
48
49 __setup("notsc", tsc_setup);
50
51 /*
52  * code to mark and check if the TSC is unstable
53  * due to cpufreq or due to unsynced TSCs
54  */
55 static int tsc_unstable;
56
57 static inline int check_tsc_unstable(void)
58 {
59         return tsc_unstable;
60 }
61
62 void mark_tsc_unstable(void)
63 {
64         tsc_unstable = 1;
65 }
66 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
67
68 /* Accellerators for sched_clock()
69  * convert from cycles(64bits) => nanoseconds (64bits)
70  *  basic equation:
71  *              ns = cycles / (freq / ns_per_sec)
72  *              ns = cycles * (ns_per_sec / freq)
73  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
74  *              ns = cycles * (10^6 / cpu_khz)
75  *
76  *      Then we use scaling math (suggested by george@mvista.com) to get:
77  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
78  *              ns = cycles * cyc2ns_scale / SC
79  *
80  *      And since SC is a constant power of two, we can convert the div
81  *  into a shift.
82  *
83  *  We can use khz divisor instead of mhz to keep a better percision, since
84  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
85  *  (mathieu.desnoyers@polymtl.ca)
86  *
87  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
88  */
89 static unsigned long cyc2ns_scale __read_mostly;
90
91 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
92
93 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
94 {
95         cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
96 }
97
98 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
99 {
100         return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
101 }
102
103 /*
104  * Scheduler clock - returns current time in nanosec units.
105  */
106 unsigned long long sched_clock(void)
107 {
108         unsigned long long this_offset;
109
110         /*
111          * in the NUMA case we dont use the TSC as they are not
112          * synchronized across all CPUs.
113          */
114 #ifndef CONFIG_NUMA
115         if (!cpu_khz || check_tsc_unstable())
116 #endif
117                 /* no locking but a rare wrong value is not a big deal */
118                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
119
120         /* read the Time Stamp Counter: */
121         rdtscll(this_offset);
122
123         /* return the value in ns */
124         return cycles_2_ns(this_offset);
125 }
126
127 static unsigned long calculate_cpu_khz(void)
128 {
129         unsigned long long start, end;
130         unsigned long count;
131         u64 delta64;
132         int i;
133         unsigned long flags;
134
135         local_irq_save(flags);
136
137         /* run 3 times to ensure the cache is warm */
138         for (i = 0; i < 3; i++) {
139                 mach_prepare_counter();
140                 rdtscll(start);
141                 mach_countup(&count);
142                 rdtscll(end);
143         }
144         /*
145          * Error: ECTCNEVERSET
146          * The CTC wasn't reliable: we got a hit on the very first read,
147          * or the CPU was so fast/slow that the quotient wouldn't fit in
148          * 32 bits..
149          */
150         if (count <= 1)
151                 goto err;
152
153         delta64 = end - start;
154
155         /* cpu freq too fast: */
156         if (delta64 > (1ULL<<32))
157                 goto err;
158
159         /* cpu freq too slow: */
160         if (delta64 <= CALIBRATE_TIME_MSEC)
161                 goto err;
162
163         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
164         do_div(delta64,CALIBRATE_TIME_MSEC);
165
166         local_irq_restore(flags);
167         return (unsigned long)delta64;
168 err:
169         local_irq_restore(flags);
170         return 0;
171 }
172
173 int recalibrate_cpu_khz(void)
174 {
175 #ifndef CONFIG_SMP
176         unsigned long cpu_khz_old = cpu_khz;
177
178         if (cpu_has_tsc) {
179                 cpu_khz = calculate_cpu_khz();
180                 tsc_khz = cpu_khz;
181                 cpu_data[0].loops_per_jiffy =
182                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
183                                         cpu_khz_old, cpu_khz);
184                 return 0;
185         } else
186                 return -ENODEV;
187 #else
188         return -ENODEV;
189 #endif
190 }
191
192 EXPORT_SYMBOL(recalibrate_cpu_khz);
193
194 void __init tsc_init(void)
195 {
196         if (!cpu_has_tsc || tsc_disable)
197                 return;
198
199         cpu_khz = calculate_cpu_khz();
200         tsc_khz = cpu_khz;
201
202         if (!cpu_khz)
203                 return;
204
205         printk("Detected %lu.%03lu MHz processor.\n",
206                                 (unsigned long)cpu_khz / 1000,
207                                 (unsigned long)cpu_khz % 1000);
208
209         set_cyc2ns_scale(cpu_khz);
210         use_tsc_delay();
211 }
212
213 #ifdef CONFIG_CPU_FREQ
214
215 static unsigned int cpufreq_delayed_issched = 0;
216 static unsigned int cpufreq_init = 0;
217 static struct work_struct cpufreq_delayed_get_work;
218
219 static void handle_cpufreq_delayed_get(struct work_struct *work)
220 {
221         unsigned int cpu;
222
223         for_each_online_cpu(cpu)
224                 cpufreq_get(cpu);
225
226         cpufreq_delayed_issched = 0;
227 }
228
229 /*
230  * if we notice cpufreq oddness, schedule a call to cpufreq_get() as it tries
231  * to verify the CPU frequency the timing core thinks the CPU is running
232  * at is still correct.
233  */
234 static inline void cpufreq_delayed_get(void)
235 {
236         if (cpufreq_init && !cpufreq_delayed_issched) {
237                 cpufreq_delayed_issched = 1;
238                 printk(KERN_DEBUG "Checking if CPU frequency changed.\n");
239                 schedule_work(&cpufreq_delayed_get_work);
240         }
241 }
242
243 /*
244  * if the CPU frequency is scaled, TSC-based delays will need a different
245  * loops_per_jiffy value to function properly.
246  */
247 static unsigned int ref_freq = 0;
248 static unsigned long loops_per_jiffy_ref = 0;
249 static unsigned long cpu_khz_ref = 0;
250
251 static int
252 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
253 {
254         struct cpufreq_freqs *freq = data;
255
256         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
257                 write_seqlock_irq(&xtime_lock);
258
259         if (!ref_freq) {
260                 if (!freq->old){
261                         ref_freq = freq->new;
262                         goto end;
263                 }
264                 ref_freq = freq->old;
265                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
266                 cpu_khz_ref = cpu_khz;
267         }
268
269         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
270             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
271             (val == CPUFREQ_RESUMECHANGE)) {
272                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
273                         cpu_data[freq->cpu].loops_per_jiffy =
274                                 cpufreq_scale(loops_per_jiffy_ref,
275                                                 ref_freq, freq->new);
276
277                 if (cpu_khz) {
278
279                         if (num_online_cpus() == 1)
280                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
281                                                 ref_freq, freq->new);
282                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
283                                 tsc_khz = cpu_khz;
284                                 set_cyc2ns_scale(cpu_khz);
285                                 /*
286                                  * TSC based sched_clock turns
287                                  * to junk w/ cpufreq
288                                  */
289                                 mark_tsc_unstable();
290                         }
291                 }
292         }
293 end:
294         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
295                 write_sequnlock_irq(&xtime_lock);
296
297         return 0;
298 }
299
300 static struct notifier_block time_cpufreq_notifier_block = {
301         .notifier_call  = time_cpufreq_notifier
302 };
303
304 static int __init cpufreq_tsc(void)
305 {
306         int ret;
307
308         INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get);
309         ret = cpufreq_register_notifier(&time_cpufreq_notifier_block,
310                                         CPUFREQ_TRANSITION_NOTIFIER);
311         if (!ret)
312                 cpufreq_init = 1;
313
314         return ret;
315 }
316
317 core_initcall(cpufreq_tsc);
318
319 #endif
320
321 /* clock source code */
322
323 static unsigned long current_tsc_khz = 0;
324 static int tsc_update_callback(void);
325
326 static cycle_t read_tsc(void)
327 {
328         cycle_t ret;
329
330         rdtscll(ret);
331
332         return ret;
333 }
334
335 static struct clocksource clocksource_tsc = {
336         .name                   = "tsc",
337         .rating                 = 300,
338         .read                   = read_tsc,
339         .mask                   = CLOCKSOURCE_MASK(64),
340         .mult                   = 0, /* to be set */
341         .shift                  = 22,
342         .update_callback        = tsc_update_callback,
343         .is_continuous          = 1,
344 };
345
346 static int tsc_update_callback(void)
347 {
348         int change = 0;
349
350         /* check to see if we should switch to the safe clocksource: */
351         if (clocksource_tsc.rating != 0 && check_tsc_unstable()) {
352                 clocksource_tsc.rating = 0;
353                 clocksource_reselect();
354                 change = 1;
355         }
356
357         /* only update if tsc_khz has changed: */
358         if (current_tsc_khz != tsc_khz) {
359                 current_tsc_khz = tsc_khz;
360                 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
361                                                         clocksource_tsc.shift);
362                 change = 1;
363         }
364
365         return change;
366 }
367
368 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
369 {
370         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
371                        d->ident);
372         mark_tsc_unstable();
373         return 0;
374 }
375
376 /* List of systems that have known TSC problems */
377 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
378         {
379          .callback = dmi_mark_tsc_unstable,
380          .ident = "IBM Thinkpad 380XD",
381          .matches = {
382                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
383                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
384                      },
385          },
386          {}
387 };
388
389 #define TSC_FREQ_CHECK_INTERVAL (10*MSEC_PER_SEC) /* 10sec in MS */
390 static struct timer_list verify_tsc_freq_timer;
391
392 /* XXX - Probably should add locking */
393 static void verify_tsc_freq(unsigned long unused)
394 {
395         static u64 last_tsc;
396         static unsigned long last_jiffies;
397
398         u64 now_tsc, interval_tsc;
399         unsigned long now_jiffies, interval_jiffies;
400
401
402         if (check_tsc_unstable())
403                 return;
404
405         rdtscll(now_tsc);
406         now_jiffies = jiffies;
407
408         if (!last_jiffies) {
409                 goto out;
410         }
411
412         interval_jiffies = now_jiffies - last_jiffies;
413         interval_tsc = now_tsc - last_tsc;
414         interval_tsc *= HZ;
415         do_div(interval_tsc, cpu_khz*1000);
416
417         if (interval_tsc < (interval_jiffies * 3 / 4)) {
418                 printk("TSC appears to be running slowly. "
419                         "Marking it as unstable\n");
420                 mark_tsc_unstable();
421                 return;
422         }
423
424 out:
425         last_tsc = now_tsc;
426         last_jiffies = now_jiffies;
427         /* set us up to go off on the next interval: */
428         mod_timer(&verify_tsc_freq_timer,
429                 jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL));
430 }
431
432 /*
433  * Make an educated guess if the TSC is trustworthy and synchronized
434  * over all CPUs.
435  */
436 static __init int unsynchronized_tsc(void)
437 {
438         /*
439          * Intel systems are normally all synchronized.
440          * Exceptions must mark TSC as unstable:
441          */
442         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
443                 return 0;
444
445         /* assume multi socket systems are not synchronized: */
446         return num_possible_cpus() > 1;
447 }
448
449 static int __init init_tsc_clocksource(void)
450 {
451
452         if (cpu_has_tsc && tsc_khz && !tsc_disable) {
453                 /* check blacklist */
454                 dmi_check_system(bad_tsc_dmi_table);
455
456                 if (unsynchronized_tsc()) /* mark unstable if unsynced */
457                         mark_tsc_unstable();
458                 current_tsc_khz = tsc_khz;
459                 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
460                                                         clocksource_tsc.shift);
461                 /* lower the rating if we already know its unstable: */
462                 if (check_tsc_unstable())
463                         clocksource_tsc.rating = 0;
464
465                 init_timer(&verify_tsc_freq_timer);
466                 verify_tsc_freq_timer.function = verify_tsc_freq;
467                 verify_tsc_freq_timer.expires =
468                         jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL);
469                 add_timer(&verify_tsc_freq_timer);
470
471                 return clocksource_register(&clocksource_tsc);
472         }
473
474         return 0;
475 }
476
477 module_init(init_tsc_clocksource);