Loading...
1#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
2
3#include <linux/kernel.h>
4#include <linux/sched.h>
5#include <linux/sched/clock.h>
6#include <linux/init.h>
7#include <linux/export.h>
8#include <linux/timer.h>
9#include <linux/acpi_pmtmr.h>
10#include <linux/cpufreq.h>
11#include <linux/delay.h>
12#include <linux/clocksource.h>
13#include <linux/percpu.h>
14#include <linux/timex.h>
15#include <linux/static_key.h>
16
17#include <asm/hpet.h>
18#include <asm/timer.h>
19#include <asm/vgtod.h>
20#include <asm/time.h>
21#include <asm/delay.h>
22#include <asm/hypervisor.h>
23#include <asm/nmi.h>
24#include <asm/x86_init.h>
25#include <asm/geode.h>
26#include <asm/apic.h>
27#include <asm/intel-family.h>
28#include <asm/i8259.h>
29
30unsigned int __read_mostly cpu_khz; /* TSC clocks / usec, not used here */
31EXPORT_SYMBOL(cpu_khz);
32
33unsigned int __read_mostly tsc_khz;
34EXPORT_SYMBOL(tsc_khz);
35
36/*
37 * TSC can be unstable due to cpufreq or due to unsynced TSCs
38 */
39static int __read_mostly tsc_unstable;
40
41/* native_sched_clock() is called before tsc_init(), so
42 we must start with the TSC soft disabled to prevent
43 erroneous rdtsc usage on !boot_cpu_has(X86_FEATURE_TSC) processors */
44static int __read_mostly tsc_disabled = -1;
45
46static DEFINE_STATIC_KEY_FALSE(__use_tsc);
47
48int tsc_clocksource_reliable;
49
50static u32 art_to_tsc_numerator;
51static u32 art_to_tsc_denominator;
52static u64 art_to_tsc_offset;
53struct clocksource *art_related_clocksource;
54
55struct cyc2ns {
56 struct cyc2ns_data data[2]; /* 0 + 2*16 = 32 */
57 seqcount_t seq; /* 32 + 4 = 36 */
58
59}; /* fits one cacheline */
60
61static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns);
62
63void cyc2ns_read_begin(struct cyc2ns_data *data)
64{
65 int seq, idx;
66
67 preempt_disable_notrace();
68
69 do {
70 seq = this_cpu_read(cyc2ns.seq.sequence);
71 idx = seq & 1;
72
73 data->cyc2ns_offset = this_cpu_read(cyc2ns.data[idx].cyc2ns_offset);
74 data->cyc2ns_mul = this_cpu_read(cyc2ns.data[idx].cyc2ns_mul);
75 data->cyc2ns_shift = this_cpu_read(cyc2ns.data[idx].cyc2ns_shift);
76
77 } while (unlikely(seq != this_cpu_read(cyc2ns.seq.sequence)));
78}
79
80void cyc2ns_read_end(void)
81{
82 preempt_enable_notrace();
83}
84
85/*
86 * Accelerators for sched_clock()
87 * convert from cycles(64bits) => nanoseconds (64bits)
88 * basic equation:
89 * ns = cycles / (freq / ns_per_sec)
90 * ns = cycles * (ns_per_sec / freq)
91 * ns = cycles * (10^9 / (cpu_khz * 10^3))
92 * ns = cycles * (10^6 / cpu_khz)
93 *
94 * Then we use scaling math (suggested by george@mvista.com) to get:
95 * ns = cycles * (10^6 * SC / cpu_khz) / SC
96 * ns = cycles * cyc2ns_scale / SC
97 *
98 * And since SC is a constant power of two, we can convert the div
99 * into a shift. The larger SC is, the more accurate the conversion, but
100 * cyc2ns_scale needs to be a 32-bit value so that 32-bit multiplication
101 * (64-bit result) can be used.
102 *
103 * We can use khz divisor instead of mhz to keep a better precision.
104 * (mathieu.desnoyers@polymtl.ca)
105 *
106 * -johnstul@us.ibm.com "math is hard, lets go shopping!"
107 */
108
109static void cyc2ns_data_init(struct cyc2ns_data *data)
110{
111 data->cyc2ns_mul = 0;
112 data->cyc2ns_shift = 0;
113 data->cyc2ns_offset = 0;
114}
115
116static void __init cyc2ns_init(int cpu)
117{
118 struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
119
120 cyc2ns_data_init(&c2n->data[0]);
121 cyc2ns_data_init(&c2n->data[1]);
122
123 seqcount_init(&c2n->seq);
124}
125
126static inline unsigned long long cycles_2_ns(unsigned long long cyc)
127{
128 struct cyc2ns_data data;
129 unsigned long long ns;
130
131 cyc2ns_read_begin(&data);
132
133 ns = data.cyc2ns_offset;
134 ns += mul_u64_u32_shr(cyc, data.cyc2ns_mul, data.cyc2ns_shift);
135
136 cyc2ns_read_end();
137
138 return ns;
139}
140
141static void set_cyc2ns_scale(unsigned long khz, int cpu, unsigned long long tsc_now)
142{
143 unsigned long long ns_now;
144 struct cyc2ns_data data;
145 struct cyc2ns *c2n;
146 unsigned long flags;
147
148 local_irq_save(flags);
149 sched_clock_idle_sleep_event();
150
151 if (!khz)
152 goto done;
153
154 ns_now = cycles_2_ns(tsc_now);
155
156 /*
157 * Compute a new multiplier as per the above comment and ensure our
158 * time function is continuous; see the comment near struct
159 * cyc2ns_data.
160 */
161 clocks_calc_mult_shift(&data.cyc2ns_mul, &data.cyc2ns_shift, khz,
162 NSEC_PER_MSEC, 0);
163
164 /*
165 * cyc2ns_shift is exported via arch_perf_update_userpage() where it is
166 * not expected to be greater than 31 due to the original published
167 * conversion algorithm shifting a 32-bit value (now specifies a 64-bit
168 * value) - refer perf_event_mmap_page documentation in perf_event.h.
169 */
170 if (data.cyc2ns_shift == 32) {
171 data.cyc2ns_shift = 31;
172 data.cyc2ns_mul >>= 1;
173 }
174
175 data.cyc2ns_offset = ns_now -
176 mul_u64_u32_shr(tsc_now, data.cyc2ns_mul, data.cyc2ns_shift);
177
178 c2n = per_cpu_ptr(&cyc2ns, cpu);
179
180 raw_write_seqcount_latch(&c2n->seq);
181 c2n->data[0] = data;
182 raw_write_seqcount_latch(&c2n->seq);
183 c2n->data[1] = data;
184
185done:
186 sched_clock_idle_wakeup_event();
187 local_irq_restore(flags);
188}
189
190/*
191 * Scheduler clock - returns current time in nanosec units.
192 */
193u64 native_sched_clock(void)
194{
195 if (static_branch_likely(&__use_tsc)) {
196 u64 tsc_now = rdtsc();
197
198 /* return the value in ns */
199 return cycles_2_ns(tsc_now);
200 }
201
202 /*
203 * Fall back to jiffies if there's no TSC available:
204 * ( But note that we still use it if the TSC is marked
205 * unstable. We do this because unlike Time Of Day,
206 * the scheduler clock tolerates small errors and it's
207 * very important for it to be as fast as the platform
208 * can achieve it. )
209 */
210
211 /* No locking but a rare wrong value is not a big deal: */
212 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
213}
214
215/*
216 * Generate a sched_clock if you already have a TSC value.
217 */
218u64 native_sched_clock_from_tsc(u64 tsc)
219{
220 return cycles_2_ns(tsc);
221}
222
223/* We need to define a real function for sched_clock, to override the
224 weak default version */
225#ifdef CONFIG_PARAVIRT
226unsigned long long sched_clock(void)
227{
228 return paravirt_sched_clock();
229}
230
231bool using_native_sched_clock(void)
232{
233 return pv_time_ops.sched_clock == native_sched_clock;
234}
235#else
236unsigned long long
237sched_clock(void) __attribute__((alias("native_sched_clock")));
238
239bool using_native_sched_clock(void) { return true; }
240#endif
241
242int check_tsc_unstable(void)
243{
244 return tsc_unstable;
245}
246EXPORT_SYMBOL_GPL(check_tsc_unstable);
247
248#ifdef CONFIG_X86_TSC
249int __init notsc_setup(char *str)
250{
251 pr_warn("Kernel compiled with CONFIG_X86_TSC, cannot disable TSC completely\n");
252 tsc_disabled = 1;
253 return 1;
254}
255#else
256/*
257 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
258 * in cpu/common.c
259 */
260int __init notsc_setup(char *str)
261{
262 setup_clear_cpu_cap(X86_FEATURE_TSC);
263 return 1;
264}
265#endif
266
267__setup("notsc", notsc_setup);
268
269static int no_sched_irq_time;
270
271static int __init tsc_setup(char *str)
272{
273 if (!strcmp(str, "reliable"))
274 tsc_clocksource_reliable = 1;
275 if (!strncmp(str, "noirqtime", 9))
276 no_sched_irq_time = 1;
277 if (!strcmp(str, "unstable"))
278 mark_tsc_unstable("boot parameter");
279 return 1;
280}
281
282__setup("tsc=", tsc_setup);
283
284#define MAX_RETRIES 5
285#define SMI_TRESHOLD 50000
286
287/*
288 * Read TSC and the reference counters. Take care of SMI disturbance
289 */
290static u64 tsc_read_refs(u64 *p, int hpet)
291{
292 u64 t1, t2;
293 int i;
294
295 for (i = 0; i < MAX_RETRIES; i++) {
296 t1 = get_cycles();
297 if (hpet)
298 *p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
299 else
300 *p = acpi_pm_read_early();
301 t2 = get_cycles();
302 if ((t2 - t1) < SMI_TRESHOLD)
303 return t2;
304 }
305 return ULLONG_MAX;
306}
307
308/*
309 * Calculate the TSC frequency from HPET reference
310 */
311static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
312{
313 u64 tmp;
314
315 if (hpet2 < hpet1)
316 hpet2 += 0x100000000ULL;
317 hpet2 -= hpet1;
318 tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
319 do_div(tmp, 1000000);
320 deltatsc = div64_u64(deltatsc, tmp);
321
322 return (unsigned long) deltatsc;
323}
324
325/*
326 * Calculate the TSC frequency from PMTimer reference
327 */
328static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
329{
330 u64 tmp;
331
332 if (!pm1 && !pm2)
333 return ULONG_MAX;
334
335 if (pm2 < pm1)
336 pm2 += (u64)ACPI_PM_OVRRUN;
337 pm2 -= pm1;
338 tmp = pm2 * 1000000000LL;
339 do_div(tmp, PMTMR_TICKS_PER_SEC);
340 do_div(deltatsc, tmp);
341
342 return (unsigned long) deltatsc;
343}
344
345#define CAL_MS 10
346#define CAL_LATCH (PIT_TICK_RATE / (1000 / CAL_MS))
347#define CAL_PIT_LOOPS 1000
348
349#define CAL2_MS 50
350#define CAL2_LATCH (PIT_TICK_RATE / (1000 / CAL2_MS))
351#define CAL2_PIT_LOOPS 5000
352
353
354/*
355 * Try to calibrate the TSC against the Programmable
356 * Interrupt Timer and return the frequency of the TSC
357 * in kHz.
358 *
359 * Return ULONG_MAX on failure to calibrate.
360 */
361static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
362{
363 u64 tsc, t1, t2, delta;
364 unsigned long tscmin, tscmax;
365 int pitcnt;
366
367 if (!has_legacy_pic()) {
368 /*
369 * Relies on tsc_early_delay_calibrate() to have given us semi
370 * usable udelay(), wait for the same 50ms we would have with
371 * the PIT loop below.
372 */
373 udelay(10 * USEC_PER_MSEC);
374 udelay(10 * USEC_PER_MSEC);
375 udelay(10 * USEC_PER_MSEC);
376 udelay(10 * USEC_PER_MSEC);
377 udelay(10 * USEC_PER_MSEC);
378 return ULONG_MAX;
379 }
380
381 /* Set the Gate high, disable speaker */
382 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
383
384 /*
385 * Setup CTC channel 2* for mode 0, (interrupt on terminal
386 * count mode), binary count. Set the latch register to 50ms
387 * (LSB then MSB) to begin countdown.
388 */
389 outb(0xb0, 0x43);
390 outb(latch & 0xff, 0x42);
391 outb(latch >> 8, 0x42);
392
393 tsc = t1 = t2 = get_cycles();
394
395 pitcnt = 0;
396 tscmax = 0;
397 tscmin = ULONG_MAX;
398 while ((inb(0x61) & 0x20) == 0) {
399 t2 = get_cycles();
400 delta = t2 - tsc;
401 tsc = t2;
402 if ((unsigned long) delta < tscmin)
403 tscmin = (unsigned int) delta;
404 if ((unsigned long) delta > tscmax)
405 tscmax = (unsigned int) delta;
406 pitcnt++;
407 }
408
409 /*
410 * Sanity checks:
411 *
412 * If we were not able to read the PIT more than loopmin
413 * times, then we have been hit by a massive SMI
414 *
415 * If the maximum is 10 times larger than the minimum,
416 * then we got hit by an SMI as well.
417 */
418 if (pitcnt < loopmin || tscmax > 10 * tscmin)
419 return ULONG_MAX;
420
421 /* Calculate the PIT value */
422 delta = t2 - t1;
423 do_div(delta, ms);
424 return delta;
425}
426
427/*
428 * This reads the current MSB of the PIT counter, and
429 * checks if we are running on sufficiently fast and
430 * non-virtualized hardware.
431 *
432 * Our expectations are:
433 *
434 * - the PIT is running at roughly 1.19MHz
435 *
436 * - each IO is going to take about 1us on real hardware,
437 * but we allow it to be much faster (by a factor of 10) or
438 * _slightly_ slower (ie we allow up to a 2us read+counter
439 * update - anything else implies a unacceptably slow CPU
440 * or PIT for the fast calibration to work.
441 *
442 * - with 256 PIT ticks to read the value, we have 214us to
443 * see the same MSB (and overhead like doing a single TSC
444 * read per MSB value etc).
445 *
446 * - We're doing 2 reads per loop (LSB, MSB), and we expect
447 * them each to take about a microsecond on real hardware.
448 * So we expect a count value of around 100. But we'll be
449 * generous, and accept anything over 50.
450 *
451 * - if the PIT is stuck, and we see *many* more reads, we
452 * return early (and the next caller of pit_expect_msb()
453 * then consider it a failure when they don't see the
454 * next expected value).
455 *
456 * These expectations mean that we know that we have seen the
457 * transition from one expected value to another with a fairly
458 * high accuracy, and we didn't miss any events. We can thus
459 * use the TSC value at the transitions to calculate a pretty
460 * good value for the TSC frequencty.
461 */
462static inline int pit_verify_msb(unsigned char val)
463{
464 /* Ignore LSB */
465 inb(0x42);
466 return inb(0x42) == val;
467}
468
469static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
470{
471 int count;
472 u64 tsc = 0, prev_tsc = 0;
473
474 for (count = 0; count < 50000; count++) {
475 if (!pit_verify_msb(val))
476 break;
477 prev_tsc = tsc;
478 tsc = get_cycles();
479 }
480 *deltap = get_cycles() - prev_tsc;
481 *tscp = tsc;
482
483 /*
484 * We require _some_ success, but the quality control
485 * will be based on the error terms on the TSC values.
486 */
487 return count > 5;
488}
489
490/*
491 * How many MSB values do we want to see? We aim for
492 * a maximum error rate of 500ppm (in practice the
493 * real error is much smaller), but refuse to spend
494 * more than 50ms on it.
495 */
496#define MAX_QUICK_PIT_MS 50
497#define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
498
499static unsigned long quick_pit_calibrate(void)
500{
501 int i;
502 u64 tsc, delta;
503 unsigned long d1, d2;
504
505 if (!has_legacy_pic())
506 return 0;
507
508 /* Set the Gate high, disable speaker */
509 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
510
511 /*
512 * Counter 2, mode 0 (one-shot), binary count
513 *
514 * NOTE! Mode 2 decrements by two (and then the
515 * output is flipped each time, giving the same
516 * final output frequency as a decrement-by-one),
517 * so mode 0 is much better when looking at the
518 * individual counts.
519 */
520 outb(0xb0, 0x43);
521
522 /* Start at 0xffff */
523 outb(0xff, 0x42);
524 outb(0xff, 0x42);
525
526 /*
527 * The PIT starts counting at the next edge, so we
528 * need to delay for a microsecond. The easiest way
529 * to do that is to just read back the 16-bit counter
530 * once from the PIT.
531 */
532 pit_verify_msb(0);
533
534 if (pit_expect_msb(0xff, &tsc, &d1)) {
535 for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
536 if (!pit_expect_msb(0xff-i, &delta, &d2))
537 break;
538
539 delta -= tsc;
540
541 /*
542 * Extrapolate the error and fail fast if the error will
543 * never be below 500 ppm.
544 */
545 if (i == 1 &&
546 d1 + d2 >= (delta * MAX_QUICK_PIT_ITERATIONS) >> 11)
547 return 0;
548
549 /*
550 * Iterate until the error is less than 500 ppm
551 */
552 if (d1+d2 >= delta >> 11)
553 continue;
554
555 /*
556 * Check the PIT one more time to verify that
557 * all TSC reads were stable wrt the PIT.
558 *
559 * This also guarantees serialization of the
560 * last cycle read ('d2') in pit_expect_msb.
561 */
562 if (!pit_verify_msb(0xfe - i))
563 break;
564 goto success;
565 }
566 }
567 pr_info("Fast TSC calibration failed\n");
568 return 0;
569
570success:
571 /*
572 * Ok, if we get here, then we've seen the
573 * MSB of the PIT decrement 'i' times, and the
574 * error has shrunk to less than 500 ppm.
575 *
576 * As a result, we can depend on there not being
577 * any odd delays anywhere, and the TSC reads are
578 * reliable (within the error).
579 *
580 * kHz = ticks / time-in-seconds / 1000;
581 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
582 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
583 */
584 delta *= PIT_TICK_RATE;
585 do_div(delta, i*256*1000);
586 pr_info("Fast TSC calibration using PIT\n");
587 return delta;
588}
589
590/**
591 * native_calibrate_tsc
592 * Determine TSC frequency via CPUID, else return 0.
593 */
594unsigned long native_calibrate_tsc(void)
595{
596 unsigned int eax_denominator, ebx_numerator, ecx_hz, edx;
597 unsigned int crystal_khz;
598
599 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
600 return 0;
601
602 if (boot_cpu_data.cpuid_level < 0x15)
603 return 0;
604
605 eax_denominator = ebx_numerator = ecx_hz = edx = 0;
606
607 /* CPUID 15H TSC/Crystal ratio, plus optionally Crystal Hz */
608 cpuid(0x15, &eax_denominator, &ebx_numerator, &ecx_hz, &edx);
609
610 if (ebx_numerator == 0 || eax_denominator == 0)
611 return 0;
612
613 crystal_khz = ecx_hz / 1000;
614
615 if (crystal_khz == 0) {
616 switch (boot_cpu_data.x86_model) {
617 case INTEL_FAM6_SKYLAKE_MOBILE:
618 case INTEL_FAM6_SKYLAKE_DESKTOP:
619 case INTEL_FAM6_KABYLAKE_MOBILE:
620 case INTEL_FAM6_KABYLAKE_DESKTOP:
621 crystal_khz = 24000; /* 24.0 MHz */
622 break;
623 case INTEL_FAM6_ATOM_DENVERTON:
624 crystal_khz = 25000; /* 25.0 MHz */
625 break;
626 case INTEL_FAM6_ATOM_GOLDMONT:
627 crystal_khz = 19200; /* 19.2 MHz */
628 break;
629 }
630 }
631
632 if (crystal_khz == 0)
633 return 0;
634 /*
635 * TSC frequency determined by CPUID is a "hardware reported"
636 * frequency and is the most accurate one so far we have. This
637 * is considered a known frequency.
638 */
639 setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
640
641 /*
642 * For Atom SoCs TSC is the only reliable clocksource.
643 * Mark TSC reliable so no watchdog on it.
644 */
645 if (boot_cpu_data.x86_model == INTEL_FAM6_ATOM_GOLDMONT)
646 setup_force_cpu_cap(X86_FEATURE_TSC_RELIABLE);
647
648 return crystal_khz * ebx_numerator / eax_denominator;
649}
650
651static unsigned long cpu_khz_from_cpuid(void)
652{
653 unsigned int eax_base_mhz, ebx_max_mhz, ecx_bus_mhz, edx;
654
655 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
656 return 0;
657
658 if (boot_cpu_data.cpuid_level < 0x16)
659 return 0;
660
661 eax_base_mhz = ebx_max_mhz = ecx_bus_mhz = edx = 0;
662
663 cpuid(0x16, &eax_base_mhz, &ebx_max_mhz, &ecx_bus_mhz, &edx);
664
665 return eax_base_mhz * 1000;
666}
667
668/**
669 * native_calibrate_cpu - calibrate the cpu on boot
670 */
671unsigned long native_calibrate_cpu(void)
672{
673 u64 tsc1, tsc2, delta, ref1, ref2;
674 unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
675 unsigned long flags, latch, ms, fast_calibrate;
676 int hpet = is_hpet_enabled(), i, loopmin;
677
678 fast_calibrate = cpu_khz_from_cpuid();
679 if (fast_calibrate)
680 return fast_calibrate;
681
682 fast_calibrate = cpu_khz_from_msr();
683 if (fast_calibrate)
684 return fast_calibrate;
685
686 local_irq_save(flags);
687 fast_calibrate = quick_pit_calibrate();
688 local_irq_restore(flags);
689 if (fast_calibrate)
690 return fast_calibrate;
691
692 /*
693 * Run 5 calibration loops to get the lowest frequency value
694 * (the best estimate). We use two different calibration modes
695 * here:
696 *
697 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
698 * load a timeout of 50ms. We read the time right after we
699 * started the timer and wait until the PIT count down reaches
700 * zero. In each wait loop iteration we read the TSC and check
701 * the delta to the previous read. We keep track of the min
702 * and max values of that delta. The delta is mostly defined
703 * by the IO time of the PIT access, so we can detect when a
704 * SMI/SMM disturbance happened between the two reads. If the
705 * maximum time is significantly larger than the minimum time,
706 * then we discard the result and have another try.
707 *
708 * 2) Reference counter. If available we use the HPET or the
709 * PMTIMER as a reference to check the sanity of that value.
710 * We use separate TSC readouts and check inside of the
711 * reference read for a SMI/SMM disturbance. We dicard
712 * disturbed values here as well. We do that around the PIT
713 * calibration delay loop as we have to wait for a certain
714 * amount of time anyway.
715 */
716
717 /* Preset PIT loop values */
718 latch = CAL_LATCH;
719 ms = CAL_MS;
720 loopmin = CAL_PIT_LOOPS;
721
722 for (i = 0; i < 3; i++) {
723 unsigned long tsc_pit_khz;
724
725 /*
726 * Read the start value and the reference count of
727 * hpet/pmtimer when available. Then do the PIT
728 * calibration, which will take at least 50ms, and
729 * read the end value.
730 */
731 local_irq_save(flags);
732 tsc1 = tsc_read_refs(&ref1, hpet);
733 tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
734 tsc2 = tsc_read_refs(&ref2, hpet);
735 local_irq_restore(flags);
736
737 /* Pick the lowest PIT TSC calibration so far */
738 tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
739
740 /* hpet or pmtimer available ? */
741 if (ref1 == ref2)
742 continue;
743
744 /* Check, whether the sampling was disturbed by an SMI */
745 if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
746 continue;
747
748 tsc2 = (tsc2 - tsc1) * 1000000LL;
749 if (hpet)
750 tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
751 else
752 tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
753
754 tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
755
756 /* Check the reference deviation */
757 delta = ((u64) tsc_pit_min) * 100;
758 do_div(delta, tsc_ref_min);
759
760 /*
761 * If both calibration results are inside a 10% window
762 * then we can be sure, that the calibration
763 * succeeded. We break out of the loop right away. We
764 * use the reference value, as it is more precise.
765 */
766 if (delta >= 90 && delta <= 110) {
767 pr_info("PIT calibration matches %s. %d loops\n",
768 hpet ? "HPET" : "PMTIMER", i + 1);
769 return tsc_ref_min;
770 }
771
772 /*
773 * Check whether PIT failed more than once. This
774 * happens in virtualized environments. We need to
775 * give the virtual PC a slightly longer timeframe for
776 * the HPET/PMTIMER to make the result precise.
777 */
778 if (i == 1 && tsc_pit_min == ULONG_MAX) {
779 latch = CAL2_LATCH;
780 ms = CAL2_MS;
781 loopmin = CAL2_PIT_LOOPS;
782 }
783 }
784
785 /*
786 * Now check the results.
787 */
788 if (tsc_pit_min == ULONG_MAX) {
789 /* PIT gave no useful value */
790 pr_warn("Unable to calibrate against PIT\n");
791
792 /* We don't have an alternative source, disable TSC */
793 if (!hpet && !ref1 && !ref2) {
794 pr_notice("No reference (HPET/PMTIMER) available\n");
795 return 0;
796 }
797
798 /* The alternative source failed as well, disable TSC */
799 if (tsc_ref_min == ULONG_MAX) {
800 pr_warn("HPET/PMTIMER calibration failed\n");
801 return 0;
802 }
803
804 /* Use the alternative source */
805 pr_info("using %s reference calibration\n",
806 hpet ? "HPET" : "PMTIMER");
807
808 return tsc_ref_min;
809 }
810
811 /* We don't have an alternative source, use the PIT calibration value */
812 if (!hpet && !ref1 && !ref2) {
813 pr_info("Using PIT calibration value\n");
814 return tsc_pit_min;
815 }
816
817 /* The alternative source failed, use the PIT calibration value */
818 if (tsc_ref_min == ULONG_MAX) {
819 pr_warn("HPET/PMTIMER calibration failed. Using PIT calibration.\n");
820 return tsc_pit_min;
821 }
822
823 /*
824 * The calibration values differ too much. In doubt, we use
825 * the PIT value as we know that there are PMTIMERs around
826 * running at double speed. At least we let the user know:
827 */
828 pr_warn("PIT calibration deviates from %s: %lu %lu\n",
829 hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
830 pr_info("Using PIT calibration value\n");
831 return tsc_pit_min;
832}
833
834void recalibrate_cpu_khz(void)
835{
836#ifndef CONFIG_SMP
837 unsigned long cpu_khz_old = cpu_khz;
838
839 if (!boot_cpu_has(X86_FEATURE_TSC))
840 return;
841
842 cpu_khz = x86_platform.calibrate_cpu();
843 tsc_khz = x86_platform.calibrate_tsc();
844 if (tsc_khz == 0)
845 tsc_khz = cpu_khz;
846 else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
847 cpu_khz = tsc_khz;
848 cpu_data(0).loops_per_jiffy = cpufreq_scale(cpu_data(0).loops_per_jiffy,
849 cpu_khz_old, cpu_khz);
850#endif
851}
852
853EXPORT_SYMBOL(recalibrate_cpu_khz);
854
855
856static unsigned long long cyc2ns_suspend;
857
858void tsc_save_sched_clock_state(void)
859{
860 if (!sched_clock_stable())
861 return;
862
863 cyc2ns_suspend = sched_clock();
864}
865
866/*
867 * Even on processors with invariant TSC, TSC gets reset in some the
868 * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to
869 * arbitrary value (still sync'd across cpu's) during resume from such sleep
870 * states. To cope up with this, recompute the cyc2ns_offset for each cpu so
871 * that sched_clock() continues from the point where it was left off during
872 * suspend.
873 */
874void tsc_restore_sched_clock_state(void)
875{
876 unsigned long long offset;
877 unsigned long flags;
878 int cpu;
879
880 if (!sched_clock_stable())
881 return;
882
883 local_irq_save(flags);
884
885 /*
886 * We're coming out of suspend, there's no concurrency yet; don't
887 * bother being nice about the RCU stuff, just write to both
888 * data fields.
889 */
890
891 this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0);
892 this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0);
893
894 offset = cyc2ns_suspend - sched_clock();
895
896 for_each_possible_cpu(cpu) {
897 per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset;
898 per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset;
899 }
900
901 local_irq_restore(flags);
902}
903
904#ifdef CONFIG_CPU_FREQ
905/* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
906 * changes.
907 *
908 * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
909 * not that important because current Opteron setups do not support
910 * scaling on SMP anyroads.
911 *
912 * Should fix up last_tsc too. Currently gettimeofday in the
913 * first tick after the change will be slightly wrong.
914 */
915
916static unsigned int ref_freq;
917static unsigned long loops_per_jiffy_ref;
918static unsigned long tsc_khz_ref;
919
920static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
921 void *data)
922{
923 struct cpufreq_freqs *freq = data;
924 unsigned long *lpj;
925
926 lpj = &boot_cpu_data.loops_per_jiffy;
927#ifdef CONFIG_SMP
928 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
929 lpj = &cpu_data(freq->cpu).loops_per_jiffy;
930#endif
931
932 if (!ref_freq) {
933 ref_freq = freq->old;
934 loops_per_jiffy_ref = *lpj;
935 tsc_khz_ref = tsc_khz;
936 }
937 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
938 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
939 *lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
940
941 tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
942 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
943 mark_tsc_unstable("cpufreq changes");
944
945 set_cyc2ns_scale(tsc_khz, freq->cpu, rdtsc());
946 }
947
948 return 0;
949}
950
951static struct notifier_block time_cpufreq_notifier_block = {
952 .notifier_call = time_cpufreq_notifier
953};
954
955static int __init cpufreq_register_tsc_scaling(void)
956{
957 if (!boot_cpu_has(X86_FEATURE_TSC))
958 return 0;
959 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
960 return 0;
961 cpufreq_register_notifier(&time_cpufreq_notifier_block,
962 CPUFREQ_TRANSITION_NOTIFIER);
963 return 0;
964}
965
966core_initcall(cpufreq_register_tsc_scaling);
967
968#endif /* CONFIG_CPU_FREQ */
969
970#define ART_CPUID_LEAF (0x15)
971#define ART_MIN_DENOMINATOR (1)
972
973
974/*
975 * If ART is present detect the numerator:denominator to convert to TSC
976 */
977static void __init detect_art(void)
978{
979 unsigned int unused[2];
980
981 if (boot_cpu_data.cpuid_level < ART_CPUID_LEAF)
982 return;
983
984 /*
985 * Don't enable ART in a VM, non-stop TSC and TSC_ADJUST required,
986 * and the TSC counter resets must not occur asynchronously.
987 */
988 if (boot_cpu_has(X86_FEATURE_HYPERVISOR) ||
989 !boot_cpu_has(X86_FEATURE_NONSTOP_TSC) ||
990 !boot_cpu_has(X86_FEATURE_TSC_ADJUST) ||
991 tsc_async_resets)
992 return;
993
994 cpuid(ART_CPUID_LEAF, &art_to_tsc_denominator,
995 &art_to_tsc_numerator, unused, unused+1);
996
997 if (art_to_tsc_denominator < ART_MIN_DENOMINATOR)
998 return;
999
1000 rdmsrl(MSR_IA32_TSC_ADJUST, art_to_tsc_offset);
1001
1002 /* Make this sticky over multiple CPU init calls */
1003 setup_force_cpu_cap(X86_FEATURE_ART);
1004}
1005
1006
1007/* clocksource code */
1008
1009static void tsc_resume(struct clocksource *cs)
1010{
1011 tsc_verify_tsc_adjust(true);
1012}
1013
1014/*
1015 * We used to compare the TSC to the cycle_last value in the clocksource
1016 * structure to avoid a nasty time-warp. This can be observed in a
1017 * very small window right after one CPU updated cycle_last under
1018 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
1019 * is smaller than the cycle_last reference value due to a TSC which
1020 * is slighty behind. This delta is nowhere else observable, but in
1021 * that case it results in a forward time jump in the range of hours
1022 * due to the unsigned delta calculation of the time keeping core
1023 * code, which is necessary to support wrapping clocksources like pm
1024 * timer.
1025 *
1026 * This sanity check is now done in the core timekeeping code.
1027 * checking the result of read_tsc() - cycle_last for being negative.
1028 * That works because CLOCKSOURCE_MASK(64) does not mask out any bit.
1029 */
1030static u64 read_tsc(struct clocksource *cs)
1031{
1032 return (u64)rdtsc_ordered();
1033}
1034
1035static void tsc_cs_mark_unstable(struct clocksource *cs)
1036{
1037 if (tsc_unstable)
1038 return;
1039
1040 tsc_unstable = 1;
1041 if (using_native_sched_clock())
1042 clear_sched_clock_stable();
1043 disable_sched_clock_irqtime();
1044 pr_info("Marking TSC unstable due to clocksource watchdog\n");
1045}
1046
1047static void tsc_cs_tick_stable(struct clocksource *cs)
1048{
1049 if (tsc_unstable)
1050 return;
1051
1052 if (using_native_sched_clock())
1053 sched_clock_tick_stable();
1054}
1055
1056/*
1057 * .mask MUST be CLOCKSOURCE_MASK(64). See comment above read_tsc()
1058 */
1059static struct clocksource clocksource_tsc_early = {
1060 .name = "tsc-early",
1061 .rating = 299,
1062 .read = read_tsc,
1063 .mask = CLOCKSOURCE_MASK(64),
1064 .flags = CLOCK_SOURCE_IS_CONTINUOUS |
1065 CLOCK_SOURCE_MUST_VERIFY,
1066 .archdata = { .vclock_mode = VCLOCK_TSC },
1067 .resume = tsc_resume,
1068 .mark_unstable = tsc_cs_mark_unstable,
1069 .tick_stable = tsc_cs_tick_stable,
1070 .list = LIST_HEAD_INIT(clocksource_tsc_early.list),
1071};
1072
1073/*
1074 * Must mark VALID_FOR_HRES early such that when we unregister tsc_early
1075 * this one will immediately take over. We will only register if TSC has
1076 * been found good.
1077 */
1078static struct clocksource clocksource_tsc = {
1079 .name = "tsc",
1080 .rating = 300,
1081 .read = read_tsc,
1082 .mask = CLOCKSOURCE_MASK(64),
1083 .flags = CLOCK_SOURCE_IS_CONTINUOUS |
1084 CLOCK_SOURCE_VALID_FOR_HRES |
1085 CLOCK_SOURCE_MUST_VERIFY,
1086 .archdata = { .vclock_mode = VCLOCK_TSC },
1087 .resume = tsc_resume,
1088 .mark_unstable = tsc_cs_mark_unstable,
1089 .tick_stable = tsc_cs_tick_stable,
1090 .list = LIST_HEAD_INIT(clocksource_tsc.list),
1091};
1092
1093void mark_tsc_unstable(char *reason)
1094{
1095 if (tsc_unstable)
1096 return;
1097
1098 tsc_unstable = 1;
1099 if (using_native_sched_clock())
1100 clear_sched_clock_stable();
1101 disable_sched_clock_irqtime();
1102 pr_info("Marking TSC unstable due to %s\n", reason);
1103
1104 clocksource_mark_unstable(&clocksource_tsc_early);
1105 clocksource_mark_unstable(&clocksource_tsc);
1106}
1107
1108EXPORT_SYMBOL_GPL(mark_tsc_unstable);
1109
1110static void __init check_system_tsc_reliable(void)
1111{
1112#if defined(CONFIG_MGEODEGX1) || defined(CONFIG_MGEODE_LX) || defined(CONFIG_X86_GENERIC)
1113 if (is_geode_lx()) {
1114 /* RTSC counts during suspend */
1115#define RTSC_SUSP 0x100
1116 unsigned long res_low, res_high;
1117
1118 rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
1119 /* Geode_LX - the OLPC CPU has a very reliable TSC */
1120 if (res_low & RTSC_SUSP)
1121 tsc_clocksource_reliable = 1;
1122 }
1123#endif
1124 if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
1125 tsc_clocksource_reliable = 1;
1126}
1127
1128/*
1129 * Make an educated guess if the TSC is trustworthy and synchronized
1130 * over all CPUs.
1131 */
1132int unsynchronized_tsc(void)
1133{
1134 if (!boot_cpu_has(X86_FEATURE_TSC) || tsc_unstable)
1135 return 1;
1136
1137#ifdef CONFIG_SMP
1138 if (apic_is_clustered_box())
1139 return 1;
1140#endif
1141
1142 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1143 return 0;
1144
1145 if (tsc_clocksource_reliable)
1146 return 0;
1147 /*
1148 * Intel systems are normally all synchronized.
1149 * Exceptions must mark TSC as unstable:
1150 */
1151 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
1152 /* assume multi socket systems are not synchronized: */
1153 if (num_possible_cpus() > 1)
1154 return 1;
1155 }
1156
1157 return 0;
1158}
1159
1160/*
1161 * Convert ART to TSC given numerator/denominator found in detect_art()
1162 */
1163struct system_counterval_t convert_art_to_tsc(u64 art)
1164{
1165 u64 tmp, res, rem;
1166
1167 rem = do_div(art, art_to_tsc_denominator);
1168
1169 res = art * art_to_tsc_numerator;
1170 tmp = rem * art_to_tsc_numerator;
1171
1172 do_div(tmp, art_to_tsc_denominator);
1173 res += tmp + art_to_tsc_offset;
1174
1175 return (struct system_counterval_t) {.cs = art_related_clocksource,
1176 .cycles = res};
1177}
1178EXPORT_SYMBOL(convert_art_to_tsc);
1179
1180/**
1181 * convert_art_ns_to_tsc() - Convert ART in nanoseconds to TSC.
1182 * @art_ns: ART (Always Running Timer) in unit of nanoseconds
1183 *
1184 * PTM requires all timestamps to be in units of nanoseconds. When user
1185 * software requests a cross-timestamp, this function converts system timestamp
1186 * to TSC.
1187 *
1188 * This is valid when CPU feature flag X86_FEATURE_TSC_KNOWN_FREQ is set
1189 * indicating the tsc_khz is derived from CPUID[15H]. Drivers should check
1190 * that this flag is set before conversion to TSC is attempted.
1191 *
1192 * Return:
1193 * struct system_counterval_t - system counter value with the pointer to the
1194 * corresponding clocksource
1195 * @cycles: System counter value
1196 * @cs: Clocksource corresponding to system counter value. Used
1197 * by timekeeping code to verify comparibility of two cycle
1198 * values.
1199 */
1200
1201struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns)
1202{
1203 u64 tmp, res, rem;
1204
1205 rem = do_div(art_ns, USEC_PER_SEC);
1206
1207 res = art_ns * tsc_khz;
1208 tmp = rem * tsc_khz;
1209
1210 do_div(tmp, USEC_PER_SEC);
1211 res += tmp;
1212
1213 return (struct system_counterval_t) { .cs = art_related_clocksource,
1214 .cycles = res};
1215}
1216EXPORT_SYMBOL(convert_art_ns_to_tsc);
1217
1218
1219static void tsc_refine_calibration_work(struct work_struct *work);
1220static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);
1221/**
1222 * tsc_refine_calibration_work - Further refine tsc freq calibration
1223 * @work - ignored.
1224 *
1225 * This functions uses delayed work over a period of a
1226 * second to further refine the TSC freq value. Since this is
1227 * timer based, instead of loop based, we don't block the boot
1228 * process while this longer calibration is done.
1229 *
1230 * If there are any calibration anomalies (too many SMIs, etc),
1231 * or the refined calibration is off by 1% of the fast early
1232 * calibration, we throw out the new calibration and use the
1233 * early calibration.
1234 */
1235static void tsc_refine_calibration_work(struct work_struct *work)
1236{
1237 static u64 tsc_start = -1, ref_start;
1238 static int hpet;
1239 u64 tsc_stop, ref_stop, delta;
1240 unsigned long freq;
1241 int cpu;
1242
1243 /* Don't bother refining TSC on unstable systems */
1244 if (tsc_unstable)
1245 goto unreg;
1246
1247 /*
1248 * Since the work is started early in boot, we may be
1249 * delayed the first time we expire. So set the workqueue
1250 * again once we know timers are working.
1251 */
1252 if (tsc_start == -1) {
1253 /*
1254 * Only set hpet once, to avoid mixing hardware
1255 * if the hpet becomes enabled later.
1256 */
1257 hpet = is_hpet_enabled();
1258 schedule_delayed_work(&tsc_irqwork, HZ);
1259 tsc_start = tsc_read_refs(&ref_start, hpet);
1260 return;
1261 }
1262
1263 tsc_stop = tsc_read_refs(&ref_stop, hpet);
1264
1265 /* hpet or pmtimer available ? */
1266 if (ref_start == ref_stop)
1267 goto out;
1268
1269 /* Check, whether the sampling was disturbed by an SMI */
1270 if (tsc_start == ULLONG_MAX || tsc_stop == ULLONG_MAX)
1271 goto out;
1272
1273 delta = tsc_stop - tsc_start;
1274 delta *= 1000000LL;
1275 if (hpet)
1276 freq = calc_hpet_ref(delta, ref_start, ref_stop);
1277 else
1278 freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
1279
1280 /* Make sure we're within 1% */
1281 if (abs(tsc_khz - freq) > tsc_khz/100)
1282 goto out;
1283
1284 tsc_khz = freq;
1285 pr_info("Refined TSC clocksource calibration: %lu.%03lu MHz\n",
1286 (unsigned long)tsc_khz / 1000,
1287 (unsigned long)tsc_khz % 1000);
1288
1289 /* Inform the TSC deadline clockevent devices about the recalibration */
1290 lapic_update_tsc_freq();
1291
1292 /* Update the sched_clock() rate to match the clocksource one */
1293 for_each_possible_cpu(cpu)
1294 set_cyc2ns_scale(tsc_khz, cpu, tsc_stop);
1295
1296out:
1297 if (tsc_unstable)
1298 goto unreg;
1299
1300 if (boot_cpu_has(X86_FEATURE_ART))
1301 art_related_clocksource = &clocksource_tsc;
1302 clocksource_register_khz(&clocksource_tsc, tsc_khz);
1303unreg:
1304 clocksource_unregister(&clocksource_tsc_early);
1305}
1306
1307
1308static int __init init_tsc_clocksource(void)
1309{
1310 if (!boot_cpu_has(X86_FEATURE_TSC) || tsc_disabled > 0 || !tsc_khz)
1311 return 0;
1312
1313 if (tsc_unstable)
1314 goto unreg;
1315
1316 if (tsc_clocksource_reliable)
1317 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
1318
1319 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3))
1320 clocksource_tsc.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP;
1321
1322 /*
1323 * When TSC frequency is known (retrieved via MSR or CPUID), we skip
1324 * the refined calibration and directly register it as a clocksource.
1325 */
1326 if (boot_cpu_has(X86_FEATURE_TSC_KNOWN_FREQ)) {
1327 if (boot_cpu_has(X86_FEATURE_ART))
1328 art_related_clocksource = &clocksource_tsc;
1329 clocksource_register_khz(&clocksource_tsc, tsc_khz);
1330unreg:
1331 clocksource_unregister(&clocksource_tsc_early);
1332 return 0;
1333 }
1334
1335 schedule_delayed_work(&tsc_irqwork, 0);
1336 return 0;
1337}
1338/*
1339 * We use device_initcall here, to ensure we run after the hpet
1340 * is fully initialized, which may occur at fs_initcall time.
1341 */
1342device_initcall(init_tsc_clocksource);
1343
1344void __init tsc_early_delay_calibrate(void)
1345{
1346 unsigned long lpj;
1347
1348 if (!boot_cpu_has(X86_FEATURE_TSC))
1349 return;
1350
1351 cpu_khz = x86_platform.calibrate_cpu();
1352 tsc_khz = x86_platform.calibrate_tsc();
1353
1354 tsc_khz = tsc_khz ? : cpu_khz;
1355 if (!tsc_khz)
1356 return;
1357
1358 lpj = tsc_khz * 1000;
1359 do_div(lpj, HZ);
1360 loops_per_jiffy = lpj;
1361}
1362
1363void __init tsc_init(void)
1364{
1365 u64 lpj, cyc;
1366 int cpu;
1367
1368 if (!boot_cpu_has(X86_FEATURE_TSC)) {
1369 setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1370 return;
1371 }
1372
1373 cpu_khz = x86_platform.calibrate_cpu();
1374 tsc_khz = x86_platform.calibrate_tsc();
1375
1376 /*
1377 * Trust non-zero tsc_khz as authorative,
1378 * and use it to sanity check cpu_khz,
1379 * which will be off if system timer is off.
1380 */
1381 if (tsc_khz == 0)
1382 tsc_khz = cpu_khz;
1383 else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
1384 cpu_khz = tsc_khz;
1385
1386 if (!tsc_khz) {
1387 mark_tsc_unstable("could not calculate TSC khz");
1388 setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1389 return;
1390 }
1391
1392 pr_info("Detected %lu.%03lu MHz processor\n",
1393 (unsigned long)cpu_khz / 1000,
1394 (unsigned long)cpu_khz % 1000);
1395
1396 if (cpu_khz != tsc_khz) {
1397 pr_info("Detected %lu.%03lu MHz TSC",
1398 (unsigned long)tsc_khz / 1000,
1399 (unsigned long)tsc_khz % 1000);
1400 }
1401
1402 /* Sanitize TSC ADJUST before cyc2ns gets initialized */
1403 tsc_store_and_check_tsc_adjust(true);
1404
1405 /*
1406 * Secondary CPUs do not run through tsc_init(), so set up
1407 * all the scale factors for all CPUs, assuming the same
1408 * speed as the bootup CPU. (cpufreq notifiers will fix this
1409 * up if their speed diverges)
1410 */
1411 cyc = rdtsc();
1412 for_each_possible_cpu(cpu) {
1413 cyc2ns_init(cpu);
1414 set_cyc2ns_scale(tsc_khz, cpu, cyc);
1415 }
1416
1417 if (tsc_disabled > 0)
1418 return;
1419
1420 /* now allow native_sched_clock() to use rdtsc */
1421
1422 tsc_disabled = 0;
1423 static_branch_enable(&__use_tsc);
1424
1425 if (!no_sched_irq_time)
1426 enable_sched_clock_irqtime();
1427
1428 lpj = ((u64)tsc_khz * 1000);
1429 do_div(lpj, HZ);
1430 lpj_fine = lpj;
1431
1432 use_tsc_delay();
1433
1434 check_system_tsc_reliable();
1435
1436 if (unsynchronized_tsc()) {
1437 mark_tsc_unstable("TSCs unsynchronized");
1438 return;
1439 }
1440
1441 clocksource_register_khz(&clocksource_tsc_early, tsc_khz);
1442 detect_art();
1443}
1444
1445#ifdef CONFIG_SMP
1446/*
1447 * If we have a constant TSC and are using the TSC for the delay loop,
1448 * we can skip clock calibration if another cpu in the same socket has already
1449 * been calibrated. This assumes that CONSTANT_TSC applies to all
1450 * cpus in the socket - this should be a safe assumption.
1451 */
1452unsigned long calibrate_delay_is_known(void)
1453{
1454 int sibling, cpu = smp_processor_id();
1455 int constant_tsc = cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC);
1456 const struct cpumask *mask = topology_core_cpumask(cpu);
1457
1458 if (tsc_disabled || !constant_tsc || !mask)
1459 return 0;
1460
1461 sibling = cpumask_any_but(mask, cpu);
1462 if (sibling < nr_cpu_ids)
1463 return cpu_data(sibling).loops_per_jiffy;
1464 return 0;
1465}
1466#endif
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#include <linux/static_key.h>
15
16#include <asm/hpet.h>
17#include <asm/timer.h>
18#include <asm/vgtod.h>
19#include <asm/time.h>
20#include <asm/delay.h>
21#include <asm/hypervisor.h>
22#include <asm/nmi.h>
23#include <asm/x86_init.h>
24
25unsigned int __read_mostly cpu_khz; /* TSC clocks / usec, not used here */
26EXPORT_SYMBOL(cpu_khz);
27
28unsigned int __read_mostly tsc_khz;
29EXPORT_SYMBOL(tsc_khz);
30
31/*
32 * TSC can be unstable due to cpufreq or due to unsynced TSCs
33 */
34static int __read_mostly tsc_unstable;
35
36/* native_sched_clock() is called before tsc_init(), so
37 we must start with the TSC soft disabled to prevent
38 erroneous rdtsc usage on !cpu_has_tsc processors */
39static int __read_mostly tsc_disabled = -1;
40
41static struct static_key __use_tsc = STATIC_KEY_INIT;
42
43int tsc_clocksource_reliable;
44
45/*
46 * Use a ring-buffer like data structure, where a writer advances the head by
47 * writing a new data entry and a reader advances the tail when it observes a
48 * new entry.
49 *
50 * Writers are made to wait on readers until there's space to write a new
51 * entry.
52 *
53 * This means that we can always use an {offset, mul} pair to compute a ns
54 * value that is 'roughly' in the right direction, even if we're writing a new
55 * {offset, mul} pair during the clock read.
56 *
57 * The down-side is that we can no longer guarantee strict monotonicity anymore
58 * (assuming the TSC was that to begin with), because while we compute the
59 * intersection point of the two clock slopes and make sure the time is
60 * continuous at the point of switching; we can no longer guarantee a reader is
61 * strictly before or after the switch point.
62 *
63 * It does mean a reader no longer needs to disable IRQs in order to avoid
64 * CPU-Freq updates messing with his times, and similarly an NMI reader will
65 * no longer run the risk of hitting half-written state.
66 */
67
68struct cyc2ns {
69 struct cyc2ns_data data[2]; /* 0 + 2*24 = 48 */
70 struct cyc2ns_data *head; /* 48 + 8 = 56 */
71 struct cyc2ns_data *tail; /* 56 + 8 = 64 */
72}; /* exactly fits one cacheline */
73
74static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns);
75
76struct cyc2ns_data *cyc2ns_read_begin(void)
77{
78 struct cyc2ns_data *head;
79
80 preempt_disable();
81
82 head = this_cpu_read(cyc2ns.head);
83 /*
84 * Ensure we observe the entry when we observe the pointer to it.
85 * matches the wmb from cyc2ns_write_end().
86 */
87 smp_read_barrier_depends();
88 head->__count++;
89 barrier();
90
91 return head;
92}
93
94void cyc2ns_read_end(struct cyc2ns_data *head)
95{
96 barrier();
97 /*
98 * If we're the outer most nested read; update the tail pointer
99 * when we're done. This notifies possible pending writers
100 * that we've observed the head pointer and that the other
101 * entry is now free.
102 */
103 if (!--head->__count) {
104 /*
105 * x86-TSO does not reorder writes with older reads;
106 * therefore once this write becomes visible to another
107 * cpu, we must be finished reading the cyc2ns_data.
108 *
109 * matches with cyc2ns_write_begin().
110 */
111 this_cpu_write(cyc2ns.tail, head);
112 }
113 preempt_enable();
114}
115
116/*
117 * Begin writing a new @data entry for @cpu.
118 *
119 * Assumes some sort of write side lock; currently 'provided' by the assumption
120 * that cpufreq will call its notifiers sequentially.
121 */
122static struct cyc2ns_data *cyc2ns_write_begin(int cpu)
123{
124 struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
125 struct cyc2ns_data *data = c2n->data;
126
127 if (data == c2n->head)
128 data++;
129
130 /* XXX send an IPI to @cpu in order to guarantee a read? */
131
132 /*
133 * When we observe the tail write from cyc2ns_read_end(),
134 * the cpu must be done with that entry and its safe
135 * to start writing to it.
136 */
137 while (c2n->tail == data)
138 cpu_relax();
139
140 return data;
141}
142
143static void cyc2ns_write_end(int cpu, struct cyc2ns_data *data)
144{
145 struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
146
147 /*
148 * Ensure the @data writes are visible before we publish the
149 * entry. Matches the data-depencency in cyc2ns_read_begin().
150 */
151 smp_wmb();
152
153 ACCESS_ONCE(c2n->head) = data;
154}
155
156/*
157 * Accelerators for sched_clock()
158 * convert from cycles(64bits) => nanoseconds (64bits)
159 * basic equation:
160 * ns = cycles / (freq / ns_per_sec)
161 * ns = cycles * (ns_per_sec / freq)
162 * ns = cycles * (10^9 / (cpu_khz * 10^3))
163 * ns = cycles * (10^6 / cpu_khz)
164 *
165 * Then we use scaling math (suggested by george@mvista.com) to get:
166 * ns = cycles * (10^6 * SC / cpu_khz) / SC
167 * ns = cycles * cyc2ns_scale / SC
168 *
169 * And since SC is a constant power of two, we can convert the div
170 * into a shift.
171 *
172 * We can use khz divisor instead of mhz to keep a better precision, since
173 * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
174 * (mathieu.desnoyers@polymtl.ca)
175 *
176 * -johnstul@us.ibm.com "math is hard, lets go shopping!"
177 */
178
179#define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
180
181static void cyc2ns_data_init(struct cyc2ns_data *data)
182{
183 data->cyc2ns_mul = 0;
184 data->cyc2ns_shift = CYC2NS_SCALE_FACTOR;
185 data->cyc2ns_offset = 0;
186 data->__count = 0;
187}
188
189static void cyc2ns_init(int cpu)
190{
191 struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
192
193 cyc2ns_data_init(&c2n->data[0]);
194 cyc2ns_data_init(&c2n->data[1]);
195
196 c2n->head = c2n->data;
197 c2n->tail = c2n->data;
198}
199
200static inline unsigned long long cycles_2_ns(unsigned long long cyc)
201{
202 struct cyc2ns_data *data, *tail;
203 unsigned long long ns;
204
205 /*
206 * See cyc2ns_read_*() for details; replicated in order to avoid
207 * an extra few instructions that came with the abstraction.
208 * Notable, it allows us to only do the __count and tail update
209 * dance when its actually needed.
210 */
211
212 preempt_disable_notrace();
213 data = this_cpu_read(cyc2ns.head);
214 tail = this_cpu_read(cyc2ns.tail);
215
216 if (likely(data == tail)) {
217 ns = data->cyc2ns_offset;
218 ns += mul_u64_u32_shr(cyc, data->cyc2ns_mul, CYC2NS_SCALE_FACTOR);
219 } else {
220 data->__count++;
221
222 barrier();
223
224 ns = data->cyc2ns_offset;
225 ns += mul_u64_u32_shr(cyc, data->cyc2ns_mul, CYC2NS_SCALE_FACTOR);
226
227 barrier();
228
229 if (!--data->__count)
230 this_cpu_write(cyc2ns.tail, data);
231 }
232 preempt_enable_notrace();
233
234 return ns;
235}
236
237/* XXX surely we already have this someplace in the kernel?! */
238#define DIV_ROUND(n, d) (((n) + ((d) / 2)) / (d))
239
240static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
241{
242 unsigned long long tsc_now, ns_now;
243 struct cyc2ns_data *data;
244 unsigned long flags;
245
246 local_irq_save(flags);
247 sched_clock_idle_sleep_event();
248
249 if (!cpu_khz)
250 goto done;
251
252 data = cyc2ns_write_begin(cpu);
253
254 rdtscll(tsc_now);
255 ns_now = cycles_2_ns(tsc_now);
256
257 /*
258 * Compute a new multiplier as per the above comment and ensure our
259 * time function is continuous; see the comment near struct
260 * cyc2ns_data.
261 */
262 data->cyc2ns_mul = DIV_ROUND(NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR, cpu_khz);
263 data->cyc2ns_shift = CYC2NS_SCALE_FACTOR;
264 data->cyc2ns_offset = ns_now -
265 mul_u64_u32_shr(tsc_now, data->cyc2ns_mul, CYC2NS_SCALE_FACTOR);
266
267 cyc2ns_write_end(cpu, data);
268
269done:
270 sched_clock_idle_wakeup_event(0);
271 local_irq_restore(flags);
272}
273/*
274 * Scheduler clock - returns current time in nanosec units.
275 */
276u64 native_sched_clock(void)
277{
278 u64 tsc_now;
279
280 /*
281 * Fall back to jiffies if there's no TSC available:
282 * ( But note that we still use it if the TSC is marked
283 * unstable. We do this because unlike Time Of Day,
284 * the scheduler clock tolerates small errors and it's
285 * very important for it to be as fast as the platform
286 * can achieve it. )
287 */
288 if (!static_key_false(&__use_tsc)) {
289 /* No locking but a rare wrong value is not a big deal: */
290 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
291 }
292
293 /* read the Time Stamp Counter: */
294 rdtscll(tsc_now);
295
296 /* return the value in ns */
297 return cycles_2_ns(tsc_now);
298}
299
300/* We need to define a real function for sched_clock, to override the
301 weak default version */
302#ifdef CONFIG_PARAVIRT
303unsigned long long sched_clock(void)
304{
305 return paravirt_sched_clock();
306}
307#else
308unsigned long long
309sched_clock(void) __attribute__((alias("native_sched_clock")));
310#endif
311
312unsigned long long native_read_tsc(void)
313{
314 return __native_read_tsc();
315}
316EXPORT_SYMBOL(native_read_tsc);
317
318int check_tsc_unstable(void)
319{
320 return tsc_unstable;
321}
322EXPORT_SYMBOL_GPL(check_tsc_unstable);
323
324int check_tsc_disabled(void)
325{
326 return tsc_disabled;
327}
328EXPORT_SYMBOL_GPL(check_tsc_disabled);
329
330#ifdef CONFIG_X86_TSC
331int __init notsc_setup(char *str)
332{
333 pr_warn("Kernel compiled with CONFIG_X86_TSC, cannot disable TSC completely\n");
334 tsc_disabled = 1;
335 return 1;
336}
337#else
338/*
339 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
340 * in cpu/common.c
341 */
342int __init notsc_setup(char *str)
343{
344 setup_clear_cpu_cap(X86_FEATURE_TSC);
345 return 1;
346}
347#endif
348
349__setup("notsc", notsc_setup);
350
351static int no_sched_irq_time;
352
353static int __init tsc_setup(char *str)
354{
355 if (!strcmp(str, "reliable"))
356 tsc_clocksource_reliable = 1;
357 if (!strncmp(str, "noirqtime", 9))
358 no_sched_irq_time = 1;
359 return 1;
360}
361
362__setup("tsc=", tsc_setup);
363
364#define MAX_RETRIES 5
365#define SMI_TRESHOLD 50000
366
367/*
368 * Read TSC and the reference counters. Take care of SMI disturbance
369 */
370static u64 tsc_read_refs(u64 *p, int hpet)
371{
372 u64 t1, t2;
373 int i;
374
375 for (i = 0; i < MAX_RETRIES; i++) {
376 t1 = get_cycles();
377 if (hpet)
378 *p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
379 else
380 *p = acpi_pm_read_early();
381 t2 = get_cycles();
382 if ((t2 - t1) < SMI_TRESHOLD)
383 return t2;
384 }
385 return ULLONG_MAX;
386}
387
388/*
389 * Calculate the TSC frequency from HPET reference
390 */
391static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
392{
393 u64 tmp;
394
395 if (hpet2 < hpet1)
396 hpet2 += 0x100000000ULL;
397 hpet2 -= hpet1;
398 tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
399 do_div(tmp, 1000000);
400 do_div(deltatsc, tmp);
401
402 return (unsigned long) deltatsc;
403}
404
405/*
406 * Calculate the TSC frequency from PMTimer reference
407 */
408static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
409{
410 u64 tmp;
411
412 if (!pm1 && !pm2)
413 return ULONG_MAX;
414
415 if (pm2 < pm1)
416 pm2 += (u64)ACPI_PM_OVRRUN;
417 pm2 -= pm1;
418 tmp = pm2 * 1000000000LL;
419 do_div(tmp, PMTMR_TICKS_PER_SEC);
420 do_div(deltatsc, tmp);
421
422 return (unsigned long) deltatsc;
423}
424
425#define CAL_MS 10
426#define CAL_LATCH (PIT_TICK_RATE / (1000 / CAL_MS))
427#define CAL_PIT_LOOPS 1000
428
429#define CAL2_MS 50
430#define CAL2_LATCH (PIT_TICK_RATE / (1000 / CAL2_MS))
431#define CAL2_PIT_LOOPS 5000
432
433
434/*
435 * Try to calibrate the TSC against the Programmable
436 * Interrupt Timer and return the frequency of the TSC
437 * in kHz.
438 *
439 * Return ULONG_MAX on failure to calibrate.
440 */
441static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
442{
443 u64 tsc, t1, t2, delta;
444 unsigned long tscmin, tscmax;
445 int pitcnt;
446
447 /* Set the Gate high, disable speaker */
448 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
449
450 /*
451 * Setup CTC channel 2* for mode 0, (interrupt on terminal
452 * count mode), binary count. Set the latch register to 50ms
453 * (LSB then MSB) to begin countdown.
454 */
455 outb(0xb0, 0x43);
456 outb(latch & 0xff, 0x42);
457 outb(latch >> 8, 0x42);
458
459 tsc = t1 = t2 = get_cycles();
460
461 pitcnt = 0;
462 tscmax = 0;
463 tscmin = ULONG_MAX;
464 while ((inb(0x61) & 0x20) == 0) {
465 t2 = get_cycles();
466 delta = t2 - tsc;
467 tsc = t2;
468 if ((unsigned long) delta < tscmin)
469 tscmin = (unsigned int) delta;
470 if ((unsigned long) delta > tscmax)
471 tscmax = (unsigned int) delta;
472 pitcnt++;
473 }
474
475 /*
476 * Sanity checks:
477 *
478 * If we were not able to read the PIT more than loopmin
479 * times, then we have been hit by a massive SMI
480 *
481 * If the maximum is 10 times larger than the minimum,
482 * then we got hit by an SMI as well.
483 */
484 if (pitcnt < loopmin || tscmax > 10 * tscmin)
485 return ULONG_MAX;
486
487 /* Calculate the PIT value */
488 delta = t2 - t1;
489 do_div(delta, ms);
490 return delta;
491}
492
493/*
494 * This reads the current MSB of the PIT counter, and
495 * checks if we are running on sufficiently fast and
496 * non-virtualized hardware.
497 *
498 * Our expectations are:
499 *
500 * - the PIT is running at roughly 1.19MHz
501 *
502 * - each IO is going to take about 1us on real hardware,
503 * but we allow it to be much faster (by a factor of 10) or
504 * _slightly_ slower (ie we allow up to a 2us read+counter
505 * update - anything else implies a unacceptably slow CPU
506 * or PIT for the fast calibration to work.
507 *
508 * - with 256 PIT ticks to read the value, we have 214us to
509 * see the same MSB (and overhead like doing a single TSC
510 * read per MSB value etc).
511 *
512 * - We're doing 2 reads per loop (LSB, MSB), and we expect
513 * them each to take about a microsecond on real hardware.
514 * So we expect a count value of around 100. But we'll be
515 * generous, and accept anything over 50.
516 *
517 * - if the PIT is stuck, and we see *many* more reads, we
518 * return early (and the next caller of pit_expect_msb()
519 * then consider it a failure when they don't see the
520 * next expected value).
521 *
522 * These expectations mean that we know that we have seen the
523 * transition from one expected value to another with a fairly
524 * high accuracy, and we didn't miss any events. We can thus
525 * use the TSC value at the transitions to calculate a pretty
526 * good value for the TSC frequencty.
527 */
528static inline int pit_verify_msb(unsigned char val)
529{
530 /* Ignore LSB */
531 inb(0x42);
532 return inb(0x42) == val;
533}
534
535static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
536{
537 int count;
538 u64 tsc = 0, prev_tsc = 0;
539
540 for (count = 0; count < 50000; count++) {
541 if (!pit_verify_msb(val))
542 break;
543 prev_tsc = tsc;
544 tsc = get_cycles();
545 }
546 *deltap = get_cycles() - prev_tsc;
547 *tscp = tsc;
548
549 /*
550 * We require _some_ success, but the quality control
551 * will be based on the error terms on the TSC values.
552 */
553 return count > 5;
554}
555
556/*
557 * How many MSB values do we want to see? We aim for
558 * a maximum error rate of 500ppm (in practice the
559 * real error is much smaller), but refuse to spend
560 * more than 50ms on it.
561 */
562#define MAX_QUICK_PIT_MS 50
563#define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
564
565static unsigned long quick_pit_calibrate(void)
566{
567 int i;
568 u64 tsc, delta;
569 unsigned long d1, d2;
570
571 /* Set the Gate high, disable speaker */
572 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
573
574 /*
575 * Counter 2, mode 0 (one-shot), binary count
576 *
577 * NOTE! Mode 2 decrements by two (and then the
578 * output is flipped each time, giving the same
579 * final output frequency as a decrement-by-one),
580 * so mode 0 is much better when looking at the
581 * individual counts.
582 */
583 outb(0xb0, 0x43);
584
585 /* Start at 0xffff */
586 outb(0xff, 0x42);
587 outb(0xff, 0x42);
588
589 /*
590 * The PIT starts counting at the next edge, so we
591 * need to delay for a microsecond. The easiest way
592 * to do that is to just read back the 16-bit counter
593 * once from the PIT.
594 */
595 pit_verify_msb(0);
596
597 if (pit_expect_msb(0xff, &tsc, &d1)) {
598 for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
599 if (!pit_expect_msb(0xff-i, &delta, &d2))
600 break;
601
602 /*
603 * Iterate until the error is less than 500 ppm
604 */
605 delta -= tsc;
606 if (d1+d2 >= delta >> 11)
607 continue;
608
609 /*
610 * Check the PIT one more time to verify that
611 * all TSC reads were stable wrt the PIT.
612 *
613 * This also guarantees serialization of the
614 * last cycle read ('d2') in pit_expect_msb.
615 */
616 if (!pit_verify_msb(0xfe - i))
617 break;
618 goto success;
619 }
620 }
621 pr_err("Fast TSC calibration failed\n");
622 return 0;
623
624success:
625 /*
626 * Ok, if we get here, then we've seen the
627 * MSB of the PIT decrement 'i' times, and the
628 * error has shrunk to less than 500 ppm.
629 *
630 * As a result, we can depend on there not being
631 * any odd delays anywhere, and the TSC reads are
632 * reliable (within the error).
633 *
634 * kHz = ticks / time-in-seconds / 1000;
635 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
636 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
637 */
638 delta *= PIT_TICK_RATE;
639 do_div(delta, i*256*1000);
640 pr_info("Fast TSC calibration using PIT\n");
641 return delta;
642}
643
644/**
645 * native_calibrate_tsc - calibrate the tsc on boot
646 */
647unsigned long native_calibrate_tsc(void)
648{
649 u64 tsc1, tsc2, delta, ref1, ref2;
650 unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
651 unsigned long flags, latch, ms, fast_calibrate;
652 int hpet = is_hpet_enabled(), i, loopmin;
653
654 /* Calibrate TSC using MSR for Intel Atom SoCs */
655 local_irq_save(flags);
656 fast_calibrate = try_msr_calibrate_tsc();
657 local_irq_restore(flags);
658 if (fast_calibrate)
659 return fast_calibrate;
660
661 local_irq_save(flags);
662 fast_calibrate = quick_pit_calibrate();
663 local_irq_restore(flags);
664 if (fast_calibrate)
665 return fast_calibrate;
666
667 /*
668 * Run 5 calibration loops to get the lowest frequency value
669 * (the best estimate). We use two different calibration modes
670 * here:
671 *
672 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
673 * load a timeout of 50ms. We read the time right after we
674 * started the timer and wait until the PIT count down reaches
675 * zero. In each wait loop iteration we read the TSC and check
676 * the delta to the previous read. We keep track of the min
677 * and max values of that delta. The delta is mostly defined
678 * by the IO time of the PIT access, so we can detect when a
679 * SMI/SMM disturbance happened between the two reads. If the
680 * maximum time is significantly larger than the minimum time,
681 * then we discard the result and have another try.
682 *
683 * 2) Reference counter. If available we use the HPET or the
684 * PMTIMER as a reference to check the sanity of that value.
685 * We use separate TSC readouts and check inside of the
686 * reference read for a SMI/SMM disturbance. We dicard
687 * disturbed values here as well. We do that around the PIT
688 * calibration delay loop as we have to wait for a certain
689 * amount of time anyway.
690 */
691
692 /* Preset PIT loop values */
693 latch = CAL_LATCH;
694 ms = CAL_MS;
695 loopmin = CAL_PIT_LOOPS;
696
697 for (i = 0; i < 3; i++) {
698 unsigned long tsc_pit_khz;
699
700 /*
701 * Read the start value and the reference count of
702 * hpet/pmtimer when available. Then do the PIT
703 * calibration, which will take at least 50ms, and
704 * read the end value.
705 */
706 local_irq_save(flags);
707 tsc1 = tsc_read_refs(&ref1, hpet);
708 tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
709 tsc2 = tsc_read_refs(&ref2, hpet);
710 local_irq_restore(flags);
711
712 /* Pick the lowest PIT TSC calibration so far */
713 tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
714
715 /* hpet or pmtimer available ? */
716 if (ref1 == ref2)
717 continue;
718
719 /* Check, whether the sampling was disturbed by an SMI */
720 if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
721 continue;
722
723 tsc2 = (tsc2 - tsc1) * 1000000LL;
724 if (hpet)
725 tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
726 else
727 tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
728
729 tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
730
731 /* Check the reference deviation */
732 delta = ((u64) tsc_pit_min) * 100;
733 do_div(delta, tsc_ref_min);
734
735 /*
736 * If both calibration results are inside a 10% window
737 * then we can be sure, that the calibration
738 * succeeded. We break out of the loop right away. We
739 * use the reference value, as it is more precise.
740 */
741 if (delta >= 90 && delta <= 110) {
742 pr_info("PIT calibration matches %s. %d loops\n",
743 hpet ? "HPET" : "PMTIMER", i + 1);
744 return tsc_ref_min;
745 }
746
747 /*
748 * Check whether PIT failed more than once. This
749 * happens in virtualized environments. We need to
750 * give the virtual PC a slightly longer timeframe for
751 * the HPET/PMTIMER to make the result precise.
752 */
753 if (i == 1 && tsc_pit_min == ULONG_MAX) {
754 latch = CAL2_LATCH;
755 ms = CAL2_MS;
756 loopmin = CAL2_PIT_LOOPS;
757 }
758 }
759
760 /*
761 * Now check the results.
762 */
763 if (tsc_pit_min == ULONG_MAX) {
764 /* PIT gave no useful value */
765 pr_warn("Unable to calibrate against PIT\n");
766
767 /* We don't have an alternative source, disable TSC */
768 if (!hpet && !ref1 && !ref2) {
769 pr_notice("No reference (HPET/PMTIMER) available\n");
770 return 0;
771 }
772
773 /* The alternative source failed as well, disable TSC */
774 if (tsc_ref_min == ULONG_MAX) {
775 pr_warn("HPET/PMTIMER calibration failed\n");
776 return 0;
777 }
778
779 /* Use the alternative source */
780 pr_info("using %s reference calibration\n",
781 hpet ? "HPET" : "PMTIMER");
782
783 return tsc_ref_min;
784 }
785
786 /* We don't have an alternative source, use the PIT calibration value */
787 if (!hpet && !ref1 && !ref2) {
788 pr_info("Using PIT calibration value\n");
789 return tsc_pit_min;
790 }
791
792 /* The alternative source failed, use the PIT calibration value */
793 if (tsc_ref_min == ULONG_MAX) {
794 pr_warn("HPET/PMTIMER calibration failed. Using PIT calibration.\n");
795 return tsc_pit_min;
796 }
797
798 /*
799 * The calibration values differ too much. In doubt, we use
800 * the PIT value as we know that there are PMTIMERs around
801 * running at double speed. At least we let the user know:
802 */
803 pr_warn("PIT calibration deviates from %s: %lu %lu\n",
804 hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
805 pr_info("Using PIT calibration value\n");
806 return tsc_pit_min;
807}
808
809int recalibrate_cpu_khz(void)
810{
811#ifndef CONFIG_SMP
812 unsigned long cpu_khz_old = cpu_khz;
813
814 if (cpu_has_tsc) {
815 tsc_khz = x86_platform.calibrate_tsc();
816 cpu_khz = tsc_khz;
817 cpu_data(0).loops_per_jiffy =
818 cpufreq_scale(cpu_data(0).loops_per_jiffy,
819 cpu_khz_old, cpu_khz);
820 return 0;
821 } else
822 return -ENODEV;
823#else
824 return -ENODEV;
825#endif
826}
827
828EXPORT_SYMBOL(recalibrate_cpu_khz);
829
830
831static unsigned long long cyc2ns_suspend;
832
833void tsc_save_sched_clock_state(void)
834{
835 if (!sched_clock_stable())
836 return;
837
838 cyc2ns_suspend = sched_clock();
839}
840
841/*
842 * Even on processors with invariant TSC, TSC gets reset in some the
843 * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to
844 * arbitrary value (still sync'd across cpu's) during resume from such sleep
845 * states. To cope up with this, recompute the cyc2ns_offset for each cpu so
846 * that sched_clock() continues from the point where it was left off during
847 * suspend.
848 */
849void tsc_restore_sched_clock_state(void)
850{
851 unsigned long long offset;
852 unsigned long flags;
853 int cpu;
854
855 if (!sched_clock_stable())
856 return;
857
858 local_irq_save(flags);
859
860 /*
861 * We're comming out of suspend, there's no concurrency yet; don't
862 * bother being nice about the RCU stuff, just write to both
863 * data fields.
864 */
865
866 this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0);
867 this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0);
868
869 offset = cyc2ns_suspend - sched_clock();
870
871 for_each_possible_cpu(cpu) {
872 per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset;
873 per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset;
874 }
875
876 local_irq_restore(flags);
877}
878
879#ifdef CONFIG_CPU_FREQ
880
881/* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
882 * changes.
883 *
884 * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
885 * not that important because current Opteron setups do not support
886 * scaling on SMP anyroads.
887 *
888 * Should fix up last_tsc too. Currently gettimeofday in the
889 * first tick after the change will be slightly wrong.
890 */
891
892static unsigned int ref_freq;
893static unsigned long loops_per_jiffy_ref;
894static unsigned long tsc_khz_ref;
895
896static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
897 void *data)
898{
899 struct cpufreq_freqs *freq = data;
900 unsigned long *lpj;
901
902 if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
903 return 0;
904
905 lpj = &boot_cpu_data.loops_per_jiffy;
906#ifdef CONFIG_SMP
907 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
908 lpj = &cpu_data(freq->cpu).loops_per_jiffy;
909#endif
910
911 if (!ref_freq) {
912 ref_freq = freq->old;
913 loops_per_jiffy_ref = *lpj;
914 tsc_khz_ref = tsc_khz;
915 }
916 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
917 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
918 *lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
919
920 tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
921 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
922 mark_tsc_unstable("cpufreq changes");
923 }
924
925 set_cyc2ns_scale(tsc_khz, freq->cpu);
926
927 return 0;
928}
929
930static struct notifier_block time_cpufreq_notifier_block = {
931 .notifier_call = time_cpufreq_notifier
932};
933
934static int __init cpufreq_tsc(void)
935{
936 if (!cpu_has_tsc)
937 return 0;
938 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
939 return 0;
940 cpufreq_register_notifier(&time_cpufreq_notifier_block,
941 CPUFREQ_TRANSITION_NOTIFIER);
942 return 0;
943}
944
945core_initcall(cpufreq_tsc);
946
947#endif /* CONFIG_CPU_FREQ */
948
949/* clocksource code */
950
951static struct clocksource clocksource_tsc;
952
953/*
954 * We compare the TSC to the cycle_last value in the clocksource
955 * structure to avoid a nasty time-warp. This can be observed in a
956 * very small window right after one CPU updated cycle_last under
957 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
958 * is smaller than the cycle_last reference value due to a TSC which
959 * is slighty behind. This delta is nowhere else observable, but in
960 * that case it results in a forward time jump in the range of hours
961 * due to the unsigned delta calculation of the time keeping core
962 * code, which is necessary to support wrapping clocksources like pm
963 * timer.
964 */
965static cycle_t read_tsc(struct clocksource *cs)
966{
967 cycle_t ret = (cycle_t)get_cycles();
968
969 return ret >= clocksource_tsc.cycle_last ?
970 ret : clocksource_tsc.cycle_last;
971}
972
973static void resume_tsc(struct clocksource *cs)
974{
975 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3))
976 clocksource_tsc.cycle_last = 0;
977}
978
979static struct clocksource clocksource_tsc = {
980 .name = "tsc",
981 .rating = 300,
982 .read = read_tsc,
983 .resume = resume_tsc,
984 .mask = CLOCKSOURCE_MASK(64),
985 .flags = CLOCK_SOURCE_IS_CONTINUOUS |
986 CLOCK_SOURCE_MUST_VERIFY,
987 .archdata = { .vclock_mode = VCLOCK_TSC },
988};
989
990void mark_tsc_unstable(char *reason)
991{
992 if (!tsc_unstable) {
993 tsc_unstable = 1;
994 clear_sched_clock_stable();
995 disable_sched_clock_irqtime();
996 pr_info("Marking TSC unstable due to %s\n", reason);
997 /* Change only the rating, when not registered */
998 if (clocksource_tsc.mult)
999 clocksource_mark_unstable(&clocksource_tsc);
1000 else {
1001 clocksource_tsc.flags |= CLOCK_SOURCE_UNSTABLE;
1002 clocksource_tsc.rating = 0;
1003 }
1004 }
1005}
1006
1007EXPORT_SYMBOL_GPL(mark_tsc_unstable);
1008
1009static void __init check_system_tsc_reliable(void)
1010{
1011#ifdef CONFIG_MGEODE_LX
1012 /* RTSC counts during suspend */
1013#define RTSC_SUSP 0x100
1014 unsigned long res_low, res_high;
1015
1016 rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
1017 /* Geode_LX - the OLPC CPU has a very reliable TSC */
1018 if (res_low & RTSC_SUSP)
1019 tsc_clocksource_reliable = 1;
1020#endif
1021 if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
1022 tsc_clocksource_reliable = 1;
1023}
1024
1025/*
1026 * Make an educated guess if the TSC is trustworthy and synchronized
1027 * over all CPUs.
1028 */
1029int unsynchronized_tsc(void)
1030{
1031 if (!cpu_has_tsc || tsc_unstable)
1032 return 1;
1033
1034#ifdef CONFIG_SMP
1035 if (apic_is_clustered_box())
1036 return 1;
1037#endif
1038
1039 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1040 return 0;
1041
1042 if (tsc_clocksource_reliable)
1043 return 0;
1044 /*
1045 * Intel systems are normally all synchronized.
1046 * Exceptions must mark TSC as unstable:
1047 */
1048 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
1049 /* assume multi socket systems are not synchronized: */
1050 if (num_possible_cpus() > 1)
1051 return 1;
1052 }
1053
1054 return 0;
1055}
1056
1057
1058static void tsc_refine_calibration_work(struct work_struct *work);
1059static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);
1060/**
1061 * tsc_refine_calibration_work - Further refine tsc freq calibration
1062 * @work - ignored.
1063 *
1064 * This functions uses delayed work over a period of a
1065 * second to further refine the TSC freq value. Since this is
1066 * timer based, instead of loop based, we don't block the boot
1067 * process while this longer calibration is done.
1068 *
1069 * If there are any calibration anomalies (too many SMIs, etc),
1070 * or the refined calibration is off by 1% of the fast early
1071 * calibration, we throw out the new calibration and use the
1072 * early calibration.
1073 */
1074static void tsc_refine_calibration_work(struct work_struct *work)
1075{
1076 static u64 tsc_start = -1, ref_start;
1077 static int hpet;
1078 u64 tsc_stop, ref_stop, delta;
1079 unsigned long freq;
1080
1081 /* Don't bother refining TSC on unstable systems */
1082 if (check_tsc_unstable())
1083 goto out;
1084
1085 /*
1086 * Since the work is started early in boot, we may be
1087 * delayed the first time we expire. So set the workqueue
1088 * again once we know timers are working.
1089 */
1090 if (tsc_start == -1) {
1091 /*
1092 * Only set hpet once, to avoid mixing hardware
1093 * if the hpet becomes enabled later.
1094 */
1095 hpet = is_hpet_enabled();
1096 schedule_delayed_work(&tsc_irqwork, HZ);
1097 tsc_start = tsc_read_refs(&ref_start, hpet);
1098 return;
1099 }
1100
1101 tsc_stop = tsc_read_refs(&ref_stop, hpet);
1102
1103 /* hpet or pmtimer available ? */
1104 if (ref_start == ref_stop)
1105 goto out;
1106
1107 /* Check, whether the sampling was disturbed by an SMI */
1108 if (tsc_start == ULLONG_MAX || tsc_stop == ULLONG_MAX)
1109 goto out;
1110
1111 delta = tsc_stop - tsc_start;
1112 delta *= 1000000LL;
1113 if (hpet)
1114 freq = calc_hpet_ref(delta, ref_start, ref_stop);
1115 else
1116 freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
1117
1118 /* Make sure we're within 1% */
1119 if (abs(tsc_khz - freq) > tsc_khz/100)
1120 goto out;
1121
1122 tsc_khz = freq;
1123 pr_info("Refined TSC clocksource calibration: %lu.%03lu MHz\n",
1124 (unsigned long)tsc_khz / 1000,
1125 (unsigned long)tsc_khz % 1000);
1126
1127out:
1128 clocksource_register_khz(&clocksource_tsc, tsc_khz);
1129}
1130
1131
1132static int __init init_tsc_clocksource(void)
1133{
1134 if (!cpu_has_tsc || tsc_disabled > 0 || !tsc_khz)
1135 return 0;
1136
1137 if (tsc_clocksource_reliable)
1138 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
1139 /* lower the rating if we already know its unstable: */
1140 if (check_tsc_unstable()) {
1141 clocksource_tsc.rating = 0;
1142 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
1143 }
1144
1145 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3))
1146 clocksource_tsc.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP;
1147
1148 /*
1149 * Trust the results of the earlier calibration on systems
1150 * exporting a reliable TSC.
1151 */
1152 if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE)) {
1153 clocksource_register_khz(&clocksource_tsc, tsc_khz);
1154 return 0;
1155 }
1156
1157 schedule_delayed_work(&tsc_irqwork, 0);
1158 return 0;
1159}
1160/*
1161 * We use device_initcall here, to ensure we run after the hpet
1162 * is fully initialized, which may occur at fs_initcall time.
1163 */
1164device_initcall(init_tsc_clocksource);
1165
1166void __init tsc_init(void)
1167{
1168 u64 lpj;
1169 int cpu;
1170
1171 x86_init.timers.tsc_pre_init();
1172
1173 if (!cpu_has_tsc)
1174 return;
1175
1176 tsc_khz = x86_platform.calibrate_tsc();
1177 cpu_khz = tsc_khz;
1178
1179 if (!tsc_khz) {
1180 mark_tsc_unstable("could not calculate TSC khz");
1181 return;
1182 }
1183
1184 pr_info("Detected %lu.%03lu MHz processor\n",
1185 (unsigned long)cpu_khz / 1000,
1186 (unsigned long)cpu_khz % 1000);
1187
1188 /*
1189 * Secondary CPUs do not run through tsc_init(), so set up
1190 * all the scale factors for all CPUs, assuming the same
1191 * speed as the bootup CPU. (cpufreq notifiers will fix this
1192 * up if their speed diverges)
1193 */
1194 for_each_possible_cpu(cpu) {
1195 cyc2ns_init(cpu);
1196 set_cyc2ns_scale(cpu_khz, cpu);
1197 }
1198
1199 if (tsc_disabled > 0)
1200 return;
1201
1202 /* now allow native_sched_clock() to use rdtsc */
1203
1204 tsc_disabled = 0;
1205 static_key_slow_inc(&__use_tsc);
1206
1207 if (!no_sched_irq_time)
1208 enable_sched_clock_irqtime();
1209
1210 lpj = ((u64)tsc_khz * 1000);
1211 do_div(lpj, HZ);
1212 lpj_fine = lpj;
1213
1214 use_tsc_delay();
1215
1216 if (unsynchronized_tsc())
1217 mark_tsc_unstable("TSCs unsynchronized");
1218
1219 check_system_tsc_reliable();
1220}
1221
1222#ifdef CONFIG_SMP
1223/*
1224 * If we have a constant TSC and are using the TSC for the delay loop,
1225 * we can skip clock calibration if another cpu in the same socket has already
1226 * been calibrated. This assumes that CONSTANT_TSC applies to all
1227 * cpus in the socket - this should be a safe assumption.
1228 */
1229unsigned long calibrate_delay_is_known(void)
1230{
1231 int i, cpu = smp_processor_id();
1232
1233 if (!tsc_disabled && !cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC))
1234 return 0;
1235
1236 for_each_online_cpu(i)
1237 if (cpu_data(i).phys_proc_id == cpu_data(cpu).phys_proc_id)
1238 return cpu_data(i).loops_per_jiffy;
1239 return 0;
1240}
1241#endif