Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
tsc.c
Go to the documentation of this file.
1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
2 
3 #include <linux/kernel.h>
4 #include <linux/sched.h>
5 #include <linux/init.h>
6 #include <linux/module.h>
7 #include <linux/timer.h>
8 #include <linux/acpi_pmtmr.h>
9 #include <linux/cpufreq.h>
10 #include <linux/delay.h>
11 #include <linux/clocksource.h>
12 #include <linux/percpu.h>
13 #include <linux/timex.h>
14 
15 #include <asm/hpet.h>
16 #include <asm/timer.h>
17 #include <asm/vgtod.h>
18 #include <asm/time.h>
19 #include <asm/delay.h>
20 #include <asm/hypervisor.h>
21 #include <asm/nmi.h>
22 #include <asm/x86_init.h>
23 
24 unsigned int __read_mostly cpu_khz; /* TSC clocks / usec, not used here */
26 
27 unsigned int __read_mostly tsc_khz;
29 
30 /*
31  * TSC can be unstable due to cpufreq or due to unsynced TSCs
32  */
33 static int __read_mostly tsc_unstable;
34 
35 /* native_sched_clock() is called before tsc_init(), so
36  we must start with the TSC soft disabled to prevent
37  erroneous rdtsc usage on !cpu_has_tsc processors */
38 static int __read_mostly tsc_disabled = -1;
39 
41 /*
42  * Scheduler clock - returns current time in nanosec units.
43  */
45 {
46  u64 this_offset;
47 
48  /*
49  * Fall back to jiffies if there's no TSC available:
50  * ( But note that we still use it if the TSC is marked
51  * unstable. We do this because unlike Time Of Day,
52  * the scheduler clock tolerates small errors and it's
53  * very important for it to be as fast as the platform
54  * can achieve it. )
55  */
56  if (unlikely(tsc_disabled)) {
57  /* No locking but a rare wrong value is not a big deal: */
58  return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
59  }
60 
61  /* read the Time Stamp Counter: */
62  rdtscll(this_offset);
63 
64  /* return the value in ns */
65  return __cycles_2_ns(this_offset);
66 }
67 
68 /* We need to define a real function for sched_clock, to override the
69  weak default version */
70 #ifdef CONFIG_PARAVIRT
71 unsigned long long sched_clock(void)
72 {
73  return paravirt_sched_clock();
74 }
75 #else
76 unsigned long long
78 #endif
79 
80 int check_tsc_unstable(void)
81 {
82  return tsc_unstable;
83 }
85 
86 #ifdef CONFIG_X86_TSC
87 int __init notsc_setup(char *str)
88 {
89  pr_warn("Kernel compiled with CONFIG_X86_TSC, cannot disable TSC completely\n");
90  tsc_disabled = 1;
91  return 1;
92 }
93 #else
94 /*
95  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
96  * in cpu/common.c
97  */
99 {
100  setup_clear_cpu_cap(X86_FEATURE_TSC);
101  return 1;
102 }
103 #endif
104 
105 __setup("notsc", notsc_setup);
106 
107 static int no_sched_irq_time;
108 
109 static int __init tsc_setup(char *str)
110 {
111  if (!strcmp(str, "reliable"))
113  if (!strncmp(str, "noirqtime", 9))
114  no_sched_irq_time = 1;
115  return 1;
116 }
117 
118 __setup("tsc=", tsc_setup);
119 
120 #define MAX_RETRIES 5
121 #define SMI_TRESHOLD 50000
122 
123 /*
124  * Read TSC and the reference counters. Take care of SMI disturbance
125  */
126 static u64 tsc_read_refs(u64 *p, int hpet)
127 {
128  u64 t1, t2;
129  int i;
130 
131  for (i = 0; i < MAX_RETRIES; i++) {
132  t1 = get_cycles();
133  if (hpet)
134  *p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
135  else
136  *p = acpi_pm_read_early();
137  t2 = get_cycles();
138  if ((t2 - t1) < SMI_TRESHOLD)
139  return t2;
140  }
141  return ULLONG_MAX;
142 }
143 
144 /*
145  * Calculate the TSC frequency from HPET reference
146  */
147 static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
148 {
149  u64 tmp;
150 
151  if (hpet2 < hpet1)
152  hpet2 += 0x100000000ULL;
153  hpet2 -= hpet1;
154  tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
155  do_div(tmp, 1000000);
156  do_div(deltatsc, tmp);
157 
158  return (unsigned long) deltatsc;
159 }
160 
161 /*
162  * Calculate the TSC frequency from PMTimer reference
163  */
164 static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
165 {
166  u64 tmp;
167 
168  if (!pm1 && !pm2)
169  return ULONG_MAX;
170 
171  if (pm2 < pm1)
172  pm2 += (u64)ACPI_PM_OVRRUN;
173  pm2 -= pm1;
174  tmp = pm2 * 1000000000LL;
176  do_div(deltatsc, tmp);
177 
178  return (unsigned long) deltatsc;
179 }
180 
181 #define CAL_MS 10
182 #define CAL_LATCH (PIT_TICK_RATE / (1000 / CAL_MS))
183 #define CAL_PIT_LOOPS 1000
184 
185 #define CAL2_MS 50
186 #define CAL2_LATCH (PIT_TICK_RATE / (1000 / CAL2_MS))
187 #define CAL2_PIT_LOOPS 5000
188 
189 
190 /*
191  * Try to calibrate the TSC against the Programmable
192  * Interrupt Timer and return the frequency of the TSC
193  * in kHz.
194  *
195  * Return ULONG_MAX on failure to calibrate.
196  */
197 static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
198 {
199  u64 tsc, t1, t2, delta;
200  unsigned long tscmin, tscmax;
201  int pitcnt;
202 
203  /* Set the Gate high, disable speaker */
204  outb((inb(0x61) & ~0x02) | 0x01, 0x61);
205 
206  /*
207  * Setup CTC channel 2* for mode 0, (interrupt on terminal
208  * count mode), binary count. Set the latch register to 50ms
209  * (LSB then MSB) to begin countdown.
210  */
211  outb(0xb0, 0x43);
212  outb(latch & 0xff, 0x42);
213  outb(latch >> 8, 0x42);
214 
215  tsc = t1 = t2 = get_cycles();
216 
217  pitcnt = 0;
218  tscmax = 0;
219  tscmin = ULONG_MAX;
220  while ((inb(0x61) & 0x20) == 0) {
221  t2 = get_cycles();
222  delta = t2 - tsc;
223  tsc = t2;
224  if ((unsigned long) delta < tscmin)
225  tscmin = (unsigned int) delta;
226  if ((unsigned long) delta > tscmax)
227  tscmax = (unsigned int) delta;
228  pitcnt++;
229  }
230 
231  /*
232  * Sanity checks:
233  *
234  * If we were not able to read the PIT more than loopmin
235  * times, then we have been hit by a massive SMI
236  *
237  * If the maximum is 10 times larger than the minimum,
238  * then we got hit by an SMI as well.
239  */
240  if (pitcnt < loopmin || tscmax > 10 * tscmin)
241  return ULONG_MAX;
242 
243  /* Calculate the PIT value */
244  delta = t2 - t1;
245  do_div(delta, ms);
246  return delta;
247 }
248 
249 /*
250  * This reads the current MSB of the PIT counter, and
251  * checks if we are running on sufficiently fast and
252  * non-virtualized hardware.
253  *
254  * Our expectations are:
255  *
256  * - the PIT is running at roughly 1.19MHz
257  *
258  * - each IO is going to take about 1us on real hardware,
259  * but we allow it to be much faster (by a factor of 10) or
260  * _slightly_ slower (ie we allow up to a 2us read+counter
261  * update - anything else implies a unacceptably slow CPU
262  * or PIT for the fast calibration to work.
263  *
264  * - with 256 PIT ticks to read the value, we have 214us to
265  * see the same MSB (and overhead like doing a single TSC
266  * read per MSB value etc).
267  *
268  * - We're doing 2 reads per loop (LSB, MSB), and we expect
269  * them each to take about a microsecond on real hardware.
270  * So we expect a count value of around 100. But we'll be
271  * generous, and accept anything over 50.
272  *
273  * - if the PIT is stuck, and we see *many* more reads, we
274  * return early (and the next caller of pit_expect_msb()
275  * then consider it a failure when they don't see the
276  * next expected value).
277  *
278  * These expectations mean that we know that we have seen the
279  * transition from one expected value to another with a fairly
280  * high accuracy, and we didn't miss any events. We can thus
281  * use the TSC value at the transitions to calculate a pretty
282  * good value for the TSC frequencty.
283  */
284 static inline int pit_verify_msb(unsigned char val)
285 {
286  /* Ignore LSB */
287  inb(0x42);
288  return inb(0x42) == val;
289 }
290 
291 static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
292 {
293  int count;
294  u64 tsc = 0, prev_tsc = 0;
295 
296  for (count = 0; count < 50000; count++) {
297  if (!pit_verify_msb(val))
298  break;
299  prev_tsc = tsc;
300  tsc = get_cycles();
301  }
302  *deltap = get_cycles() - prev_tsc;
303  *tscp = tsc;
304 
305  /*
306  * We require _some_ success, but the quality control
307  * will be based on the error terms on the TSC values.
308  */
309  return count > 5;
310 }
311 
312 /*
313  * How many MSB values do we want to see? We aim for
314  * a maximum error rate of 500ppm (in practice the
315  * real error is much smaller), but refuse to spend
316  * more than 50ms on it.
317  */
318 #define MAX_QUICK_PIT_MS 50
319 #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
320 
321 static unsigned long quick_pit_calibrate(void)
322 {
323  int i;
324  u64 tsc, delta;
325  unsigned long d1, d2;
326 
327  /* Set the Gate high, disable speaker */
328  outb((inb(0x61) & ~0x02) | 0x01, 0x61);
329 
330  /*
331  * Counter 2, mode 0 (one-shot), binary count
332  *
333  * NOTE! Mode 2 decrements by two (and then the
334  * output is flipped each time, giving the same
335  * final output frequency as a decrement-by-one),
336  * so mode 0 is much better when looking at the
337  * individual counts.
338  */
339  outb(0xb0, 0x43);
340 
341  /* Start at 0xffff */
342  outb(0xff, 0x42);
343  outb(0xff, 0x42);
344 
345  /*
346  * The PIT starts counting at the next edge, so we
347  * need to delay for a microsecond. The easiest way
348  * to do that is to just read back the 16-bit counter
349  * once from the PIT.
350  */
351  pit_verify_msb(0);
352 
353  if (pit_expect_msb(0xff, &tsc, &d1)) {
354  for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
355  if (!pit_expect_msb(0xff-i, &delta, &d2))
356  break;
357 
358  /*
359  * Iterate until the error is less than 500 ppm
360  */
361  delta -= tsc;
362  if (d1+d2 >= delta >> 11)
363  continue;
364 
365  /*
366  * Check the PIT one more time to verify that
367  * all TSC reads were stable wrt the PIT.
368  *
369  * This also guarantees serialization of the
370  * last cycle read ('d2') in pit_expect_msb.
371  */
372  if (!pit_verify_msb(0xfe - i))
373  break;
374  goto success;
375  }
376  }
377  pr_err("Fast TSC calibration failed\n");
378  return 0;
379 
380 success:
381  /*
382  * Ok, if we get here, then we've seen the
383  * MSB of the PIT decrement 'i' times, and the
384  * error has shrunk to less than 500 ppm.
385  *
386  * As a result, we can depend on there not being
387  * any odd delays anywhere, and the TSC reads are
388  * reliable (within the error).
389  *
390  * kHz = ticks / time-in-seconds / 1000;
391  * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
392  * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
393  */
394  delta *= PIT_TICK_RATE;
395  do_div(delta, i*256*1000);
396  pr_info("Fast TSC calibration using PIT\n");
397  return delta;
398 }
399 
403 unsigned long native_calibrate_tsc(void)
404 {
405  u64 tsc1, tsc2, delta, ref1, ref2;
406  unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
407  unsigned long flags, latch, ms, fast_calibrate;
408  int hpet = is_hpet_enabled(), i, loopmin;
409 
410  local_irq_save(flags);
411  fast_calibrate = quick_pit_calibrate();
412  local_irq_restore(flags);
413  if (fast_calibrate)
414  return fast_calibrate;
415 
416  /*
417  * Run 5 calibration loops to get the lowest frequency value
418  * (the best estimate). We use two different calibration modes
419  * here:
420  *
421  * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
422  * load a timeout of 50ms. We read the time right after we
423  * started the timer and wait until the PIT count down reaches
424  * zero. In each wait loop iteration we read the TSC and check
425  * the delta to the previous read. We keep track of the min
426  * and max values of that delta. The delta is mostly defined
427  * by the IO time of the PIT access, so we can detect when a
428  * SMI/SMM disturbance happened between the two reads. If the
429  * maximum time is significantly larger than the minimum time,
430  * then we discard the result and have another try.
431  *
432  * 2) Reference counter. If available we use the HPET or the
433  * PMTIMER as a reference to check the sanity of that value.
434  * We use separate TSC readouts and check inside of the
435  * reference read for a SMI/SMM disturbance. We dicard
436  * disturbed values here as well. We do that around the PIT
437  * calibration delay loop as we have to wait for a certain
438  * amount of time anyway.
439  */
440 
441  /* Preset PIT loop values */
442  latch = CAL_LATCH;
443  ms = CAL_MS;
444  loopmin = CAL_PIT_LOOPS;
445 
446  for (i = 0; i < 3; i++) {
447  unsigned long tsc_pit_khz;
448 
449  /*
450  * Read the start value and the reference count of
451  * hpet/pmtimer when available. Then do the PIT
452  * calibration, which will take at least 50ms, and
453  * read the end value.
454  */
455  local_irq_save(flags);
456  tsc1 = tsc_read_refs(&ref1, hpet);
457  tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
458  tsc2 = tsc_read_refs(&ref2, hpet);
459  local_irq_restore(flags);
460 
461  /* Pick the lowest PIT TSC calibration so far */
462  tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
463 
464  /* hpet or pmtimer available ? */
465  if (ref1 == ref2)
466  continue;
467 
468  /* Check, whether the sampling was disturbed by an SMI */
469  if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
470  continue;
471 
472  tsc2 = (tsc2 - tsc1) * 1000000LL;
473  if (hpet)
474  tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
475  else
476  tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
477 
478  tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
479 
480  /* Check the reference deviation */
481  delta = ((u64) tsc_pit_min) * 100;
482  do_div(delta, tsc_ref_min);
483 
484  /*
485  * If both calibration results are inside a 10% window
486  * then we can be sure, that the calibration
487  * succeeded. We break out of the loop right away. We
488  * use the reference value, as it is more precise.
489  */
490  if (delta >= 90 && delta <= 110) {
491  pr_info("PIT calibration matches %s. %d loops\n",
492  hpet ? "HPET" : "PMTIMER", i + 1);
493  return tsc_ref_min;
494  }
495 
496  /*
497  * Check whether PIT failed more than once. This
498  * happens in virtualized environments. We need to
499  * give the virtual PC a slightly longer timeframe for
500  * the HPET/PMTIMER to make the result precise.
501  */
502  if (i == 1 && tsc_pit_min == ULONG_MAX) {
503  latch = CAL2_LATCH;
504  ms = CAL2_MS;
505  loopmin = CAL2_PIT_LOOPS;
506  }
507  }
508 
509  /*
510  * Now check the results.
511  */
512  if (tsc_pit_min == ULONG_MAX) {
513  /* PIT gave no useful value */
514  pr_warn("Unable to calibrate against PIT\n");
515 
516  /* We don't have an alternative source, disable TSC */
517  if (!hpet && !ref1 && !ref2) {
518  pr_notice("No reference (HPET/PMTIMER) available\n");
519  return 0;
520  }
521 
522  /* The alternative source failed as well, disable TSC */
523  if (tsc_ref_min == ULONG_MAX) {
524  pr_warn("HPET/PMTIMER calibration failed\n");
525  return 0;
526  }
527 
528  /* Use the alternative source */
529  pr_info("using %s reference calibration\n",
530  hpet ? "HPET" : "PMTIMER");
531 
532  return tsc_ref_min;
533  }
534 
535  /* We don't have an alternative source, use the PIT calibration value */
536  if (!hpet && !ref1 && !ref2) {
537  pr_info("Using PIT calibration value\n");
538  return tsc_pit_min;
539  }
540 
541  /* The alternative source failed, use the PIT calibration value */
542  if (tsc_ref_min == ULONG_MAX) {
543  pr_warn("HPET/PMTIMER calibration failed. Using PIT calibration.\n");
544  return tsc_pit_min;
545  }
546 
547  /*
548  * The calibration values differ too much. In doubt, we use
549  * the PIT value as we know that there are PMTIMERs around
550  * running at double speed. At least we let the user know:
551  */
552  pr_warn("PIT calibration deviates from %s: %lu %lu\n",
553  hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
554  pr_info("Using PIT calibration value\n");
555  return tsc_pit_min;
556 }
557 
559 {
560 #ifndef CONFIG_SMP
561  unsigned long cpu_khz_old = cpu_khz;
562 
563  if (cpu_has_tsc) {
564  tsc_khz = x86_platform.calibrate_tsc();
565  cpu_khz = tsc_khz;
566  cpu_data(0).loops_per_jiffy =
567  cpufreq_scale(cpu_data(0).loops_per_jiffy,
568  cpu_khz_old, cpu_khz);
569  return 0;
570  } else
571  return -ENODEV;
572 #else
573  return -ENODEV;
574 #endif
575 }
576 
578 
579 
580 /* Accelerators for sched_clock()
581  * convert from cycles(64bits) => nanoseconds (64bits)
582  * basic equation:
583  * ns = cycles / (freq / ns_per_sec)
584  * ns = cycles * (ns_per_sec / freq)
585  * ns = cycles * (10^9 / (cpu_khz * 10^3))
586  * ns = cycles * (10^6 / cpu_khz)
587  *
588  * Then we use scaling math (suggested by [email protected]) to get:
589  * ns = cycles * (10^6 * SC / cpu_khz) / SC
590  * ns = cycles * cyc2ns_scale / SC
591  *
592  * And since SC is a constant power of two, we can convert the div
593  * into a shift.
594  *
595  * We can use khz divisor instead of mhz to keep a better precision, since
596  * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
598  *
599  * [email protected] "math is hard, lets go shopping!"
600  */
601 
602 DEFINE_PER_CPU(unsigned long, cyc2ns);
603 DEFINE_PER_CPU(unsigned long long, cyc2ns_offset);
604 
605 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
606 {
607  unsigned long long tsc_now, ns_now, *offset;
608  unsigned long flags, *scale;
609 
610  local_irq_save(flags);
611  sched_clock_idle_sleep_event();
612 
613  scale = &per_cpu(cyc2ns, cpu);
614  offset = &per_cpu(cyc2ns_offset, cpu);
615 
616  rdtscll(tsc_now);
617  ns_now = __cycles_2_ns(tsc_now);
618 
619  if (cpu_khz) {
620  *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
621  *offset = ns_now - mult_frac(tsc_now, *scale,
622  (1UL << CYC2NS_SCALE_FACTOR));
623  }
624 
625  sched_clock_idle_wakeup_event(0);
626  local_irq_restore(flags);
627 }
628 
629 static unsigned long long cyc2ns_suspend;
630 
632 {
633  if (!sched_clock_stable)
634  return;
635 
636  cyc2ns_suspend = sched_clock();
637 }
638 
639 /*
640  * Even on processors with invariant TSC, TSC gets reset in some the
641  * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to
642  * arbitrary value (still sync'd across cpu's) during resume from such sleep
643  * states. To cope up with this, recompute the cyc2ns_offset for each cpu so
644  * that sched_clock() continues from the point where it was left off during
645  * suspend.
646  */
648 {
649  unsigned long long offset;
650  unsigned long flags;
651  int cpu;
652 
653  if (!sched_clock_stable)
654  return;
655 
656  local_irq_save(flags);
657 
658  __this_cpu_write(cyc2ns_offset, 0);
659  offset = cyc2ns_suspend - sched_clock();
660 
662  per_cpu(cyc2ns_offset, cpu) = offset;
663 
664  local_irq_restore(flags);
665 }
666 
667 #ifdef CONFIG_CPU_FREQ
668 
669 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
670  * changes.
671  *
672  * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
673  * not that important because current Opteron setups do not support
674  * scaling on SMP anyroads.
675  *
676  * Should fix up last_tsc too. Currently gettimeofday in the
677  * first tick after the change will be slightly wrong.
678  */
679 
680 static unsigned int ref_freq;
681 static unsigned long loops_per_jiffy_ref;
682 static unsigned long tsc_khz_ref;
683 
684 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
685  void *data)
686 {
687  struct cpufreq_freqs *freq = data;
688  unsigned long *lpj;
689 
690  if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
691  return 0;
692 
693  lpj = &boot_cpu_data.loops_per_jiffy;
694 #ifdef CONFIG_SMP
695  if (!(freq->flags & CPUFREQ_CONST_LOOPS))
696  lpj = &cpu_data(freq->cpu).loops_per_jiffy;
697 #endif
698 
699  if (!ref_freq) {
700  ref_freq = freq->old;
701  loops_per_jiffy_ref = *lpj;
702  tsc_khz_ref = tsc_khz;
703  }
704  if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
705  (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
706  (val == CPUFREQ_RESUMECHANGE)) {
707  *lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
708 
709  tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
710  if (!(freq->flags & CPUFREQ_CONST_LOOPS))
711  mark_tsc_unstable("cpufreq changes");
712  }
713 
714  set_cyc2ns_scale(tsc_khz, freq->cpu);
715 
716  return 0;
717 }
718 
719 static struct notifier_block time_cpufreq_notifier_block = {
720  .notifier_call = time_cpufreq_notifier
721 };
722 
723 static int __init cpufreq_tsc(void)
724 {
725  if (!cpu_has_tsc)
726  return 0;
727  if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
728  return 0;
729  cpufreq_register_notifier(&time_cpufreq_notifier_block,
731  return 0;
732 }
733 
734 core_initcall(cpufreq_tsc);
735 
736 #endif /* CONFIG_CPU_FREQ */
737 
738 /* clocksource code */
739 
740 static struct clocksource clocksource_tsc;
741 
742 /*
743  * We compare the TSC to the cycle_last value in the clocksource
744  * structure to avoid a nasty time-warp. This can be observed in a
745  * very small window right after one CPU updated cycle_last under
746  * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
747  * is smaller than the cycle_last reference value due to a TSC which
748  * is slighty behind. This delta is nowhere else observable, but in
749  * that case it results in a forward time jump in the range of hours
750  * due to the unsigned delta calculation of the time keeping core
751  * code, which is necessary to support wrapping clocksources like pm
752  * timer.
753  */
754 static cycle_t read_tsc(struct clocksource *cs)
755 {
757 
758  return ret >= clocksource_tsc.cycle_last ?
759  ret : clocksource_tsc.cycle_last;
760 }
761 
762 static void resume_tsc(struct clocksource *cs)
763 {
764  clocksource_tsc.cycle_last = 0;
765 }
766 
767 static struct clocksource clocksource_tsc = {
768  .name = "tsc",
769  .rating = 300,
770  .read = read_tsc,
771  .resume = resume_tsc,
772  .mask = CLOCKSOURCE_MASK(64),
773  .flags = CLOCK_SOURCE_IS_CONTINUOUS |
775 #ifdef CONFIG_X86_64
776  .archdata = { .vclock_mode = VCLOCK_TSC },
777 #endif
778 };
779 
781 {
782  if (!tsc_unstable) {
783  tsc_unstable = 1;
784  sched_clock_stable = 0;
785  disable_sched_clock_irqtime();
786  pr_info("Marking TSC unstable due to %s\n", reason);
787  /* Change only the rating, when not registered */
788  if (clocksource_tsc.mult)
789  clocksource_mark_unstable(&clocksource_tsc);
790  else {
791  clocksource_tsc.flags |= CLOCK_SOURCE_UNSTABLE;
792  clocksource_tsc.rating = 0;
793  }
794  }
795 }
796 
798 
799 static void __init check_system_tsc_reliable(void)
800 {
801 #ifdef CONFIG_MGEODE_LX
802  /* RTSC counts during suspend */
803 #define RTSC_SUSP 0x100
804  unsigned long res_low, res_high;
805 
806  rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
807  /* Geode_LX - the OLPC CPU has a very reliable TSC */
808  if (res_low & RTSC_SUSP)
810 #endif
811  if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
813 }
814 
815 /*
816  * Make an educated guess if the TSC is trustworthy and synchronized
817  * over all CPUs.
818  */
820 {
821  if (!cpu_has_tsc || tsc_unstable)
822  return 1;
823 
824 #ifdef CONFIG_SMP
825  if (apic_is_clustered_box())
826  return 1;
827 #endif
828 
829  if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
830  return 0;
831 
833  return 0;
834  /*
835  * Intel systems are normally all synchronized.
836  * Exceptions must mark TSC as unstable:
837  */
838  if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
839  /* assume multi socket systems are not synchronized: */
840  if (num_possible_cpus() > 1)
841  return 1;
842  }
843 
844  return 0;
845 }
846 
847 
848 static void tsc_refine_calibration_work(struct work_struct *work);
849 static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);
864 static void tsc_refine_calibration_work(struct work_struct *work)
865 {
866  static u64 tsc_start = -1, ref_start;
867  static int hpet;
868  u64 tsc_stop, ref_stop, delta;
869  unsigned long freq;
870 
871  /* Don't bother refining TSC on unstable systems */
872  if (check_tsc_unstable())
873  goto out;
874 
875  /*
876  * Since the work is started early in boot, we may be
877  * delayed the first time we expire. So set the workqueue
878  * again once we know timers are working.
879  */
880  if (tsc_start == -1) {
881  /*
882  * Only set hpet once, to avoid mixing hardware
883  * if the hpet becomes enabled later.
884  */
885  hpet = is_hpet_enabled();
886  schedule_delayed_work(&tsc_irqwork, HZ);
887  tsc_start = tsc_read_refs(&ref_start, hpet);
888  return;
889  }
890 
891  tsc_stop = tsc_read_refs(&ref_stop, hpet);
892 
893  /* hpet or pmtimer available ? */
894  if (ref_start == ref_stop)
895  goto out;
896 
897  /* Check, whether the sampling was disturbed by an SMI */
898  if (tsc_start == ULLONG_MAX || tsc_stop == ULLONG_MAX)
899  goto out;
900 
901  delta = tsc_stop - tsc_start;
902  delta *= 1000000LL;
903  if (hpet)
904  freq = calc_hpet_ref(delta, ref_start, ref_stop);
905  else
906  freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
907 
908  /* Make sure we're within 1% */
909  if (abs(tsc_khz - freq) > tsc_khz/100)
910  goto out;
911 
912  tsc_khz = freq;
913  pr_info("Refined TSC clocksource calibration: %lu.%03lu MHz\n",
914  (unsigned long)tsc_khz / 1000,
915  (unsigned long)tsc_khz % 1000);
916 
917 out:
918  clocksource_register_khz(&clocksource_tsc, tsc_khz);
919 }
920 
921 
922 static int __init init_tsc_clocksource(void)
923 {
924  if (!cpu_has_tsc || tsc_disabled > 0 || !tsc_khz)
925  return 0;
926 
928  clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
929  /* lower the rating if we already know its unstable: */
930  if (check_tsc_unstable()) {
931  clocksource_tsc.rating = 0;
932  clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
933  }
934 
935  /*
936  * Trust the results of the earlier calibration on systems
937  * exporting a reliable TSC.
938  */
939  if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE)) {
940  clocksource_register_khz(&clocksource_tsc, tsc_khz);
941  return 0;
942  }
943 
944  schedule_delayed_work(&tsc_irqwork, 0);
945  return 0;
946 }
947 /*
948  * We use device_initcall here, to ensure we run after the hpet
949  * is fully initialized, which may occur at fs_initcall time.
950  */
951 device_initcall(init_tsc_clocksource);
952 
953 void __init tsc_init(void)
954 {
955  u64 lpj;
956  int cpu;
957 
958  x86_init.timers.tsc_pre_init();
959 
960  if (!cpu_has_tsc)
961  return;
962 
963  tsc_khz = x86_platform.calibrate_tsc();
964  cpu_khz = tsc_khz;
965 
966  if (!tsc_khz) {
967  mark_tsc_unstable("could not calculate TSC khz");
968  return;
969  }
970 
971  pr_info("Detected %lu.%03lu MHz processor\n",
972  (unsigned long)cpu_khz / 1000,
973  (unsigned long)cpu_khz % 1000);
974 
975  /*
976  * Secondary CPUs do not run through tsc_init(), so set up
977  * all the scale factors for all CPUs, assuming the same
978  * speed as the bootup CPU. (cpufreq notifiers will fix this
979  * up if their speed diverges)
980  */
982  set_cyc2ns_scale(cpu_khz, cpu);
983 
984  if (tsc_disabled > 0)
985  return;
986 
987  /* now allow native_sched_clock() to use rdtsc */
988  tsc_disabled = 0;
989 
990  if (!no_sched_irq_time)
991  enable_sched_clock_irqtime();
992 
993  lpj = ((u64)tsc_khz * 1000);
994  do_div(lpj, HZ);
995  lpj_fine = lpj;
996 
997  use_tsc_delay();
998 
999  if (unsynchronized_tsc())
1000  mark_tsc_unstable("TSCs unsynchronized");
1001 
1002  check_system_tsc_reliable();
1003 }
1004 
1005 #ifdef CONFIG_SMP
1006 /*
1007  * If we have a constant TSC and are using the TSC for the delay loop,
1008  * we can skip clock calibration if another cpu in the same socket has already
1009  * been calibrated. This assumes that CONSTANT_TSC applies to all
1010  * cpus in the socket - this should be a safe assumption.
1011  */
1012 unsigned long __cpuinit calibrate_delay_is_known(void)
1013 {
1014  int i, cpu = smp_processor_id();
1015 
1016  if (!tsc_disabled && !cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC))
1017  return 0;
1018 
1022  return 0;
1023 }
1024 #endif