Commit | Line | Data |
---|---|---|
c767a54b JP |
1 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
2 | ||
bfc0f594 | 3 | #include <linux/kernel.h> |
0ef95533 AK |
4 | #include <linux/sched.h> |
5 | #include <linux/init.h> | |
6 | #include <linux/module.h> | |
7 | #include <linux/timer.h> | |
bfc0f594 | 8 | #include <linux/acpi_pmtmr.h> |
2dbe06fa | 9 | #include <linux/cpufreq.h> |
8fbbc4b4 AK |
10 | #include <linux/delay.h> |
11 | #include <linux/clocksource.h> | |
12 | #include <linux/percpu.h> | |
08604bd9 | 13 | #include <linux/timex.h> |
10b033d4 | 14 | #include <linux/static_key.h> |
bfc0f594 AK |
15 | |
16 | #include <asm/hpet.h> | |
8fbbc4b4 AK |
17 | #include <asm/timer.h> |
18 | #include <asm/vgtod.h> | |
19 | #include <asm/time.h> | |
20 | #include <asm/delay.h> | |
88b094fb | 21 | #include <asm/hypervisor.h> |
08047c4f | 22 | #include <asm/nmi.h> |
2d826404 | 23 | #include <asm/x86_init.h> |
0ef95533 | 24 | |
f24ade3a | 25 | unsigned int __read_mostly cpu_khz; /* TSC clocks / usec, not used here */ |
0ef95533 | 26 | EXPORT_SYMBOL(cpu_khz); |
f24ade3a IM |
27 | |
28 | unsigned int __read_mostly tsc_khz; | |
0ef95533 AK |
29 | EXPORT_SYMBOL(tsc_khz); |
30 | ||
31 | /* | |
32 | * TSC can be unstable due to cpufreq or due to unsynced TSCs | |
33 | */ | |
f24ade3a | 34 | static int __read_mostly tsc_unstable; |
0ef95533 AK |
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 */ | |
f24ade3a | 39 | static int __read_mostly tsc_disabled = -1; |
0ef95533 | 40 | |
10b033d4 PZ |
41 | static struct static_key __use_tsc = STATIC_KEY_INIT; |
42 | ||
28a00184 | 43 | int tsc_clocksource_reliable; |
57c67da2 | 44 | |
20d1c86a PZ |
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 | ||
68 | struct 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 | ||
74 | static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns); | |
75 | ||
76 | struct 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 | ||
94 | void 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 | */ | |
122 | static 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 | ||
143 | static 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() | |
57c67da2 PZ |
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 | ||
57c67da2 PZ |
179 | #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */ |
180 | ||
20d1c86a PZ |
181 | static void cyc2ns_data_init(struct cyc2ns_data *data) |
182 | { | |
5e3c1afd | 183 | data->cyc2ns_mul = 0; |
20d1c86a PZ |
184 | data->cyc2ns_shift = CYC2NS_SCALE_FACTOR; |
185 | data->cyc2ns_offset = 0; | |
186 | data->__count = 0; | |
187 | } | |
188 | ||
189 | static 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 | ||
57c67da2 PZ |
200 | static inline unsigned long long cycles_2_ns(unsigned long long cyc) |
201 | { | |
20d1c86a PZ |
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 | ||
569d6557 | 212 | preempt_disable_notrace(); |
20d1c86a PZ |
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 | } | |
569d6557 | 232 | preempt_enable_notrace(); |
20d1c86a | 233 | |
57c67da2 PZ |
234 | return ns; |
235 | } | |
236 | ||
237 | static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu) | |
238 | { | |
20d1c86a PZ |
239 | unsigned long long tsc_now, ns_now; |
240 | struct cyc2ns_data *data; | |
241 | unsigned long flags; | |
57c67da2 PZ |
242 | |
243 | local_irq_save(flags); | |
244 | sched_clock_idle_sleep_event(); | |
245 | ||
20d1c86a PZ |
246 | if (!cpu_khz) |
247 | goto done; | |
248 | ||
249 | data = cyc2ns_write_begin(cpu); | |
57c67da2 | 250 | |
4ea1636b | 251 | tsc_now = rdtsc(); |
57c67da2 PZ |
252 | ns_now = cycles_2_ns(tsc_now); |
253 | ||
20d1c86a PZ |
254 | /* |
255 | * Compute a new multiplier as per the above comment and ensure our | |
256 | * time function is continuous; see the comment near struct | |
257 | * cyc2ns_data. | |
258 | */ | |
89171579 MN |
259 | data->cyc2ns_mul = |
260 | DIV_ROUND_CLOSEST(NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR, | |
261 | cpu_khz); | |
20d1c86a PZ |
262 | data->cyc2ns_shift = CYC2NS_SCALE_FACTOR; |
263 | data->cyc2ns_offset = ns_now - | |
264 | mul_u64_u32_shr(tsc_now, data->cyc2ns_mul, CYC2NS_SCALE_FACTOR); | |
265 | ||
266 | cyc2ns_write_end(cpu, data); | |
57c67da2 | 267 | |
20d1c86a | 268 | done: |
57c67da2 PZ |
269 | sched_clock_idle_wakeup_event(0); |
270 | local_irq_restore(flags); | |
271 | } | |
0ef95533 AK |
272 | /* |
273 | * Scheduler clock - returns current time in nanosec units. | |
274 | */ | |
275 | u64 native_sched_clock(void) | |
276 | { | |
20d1c86a | 277 | u64 tsc_now; |
0ef95533 AK |
278 | |
279 | /* | |
280 | * Fall back to jiffies if there's no TSC available: | |
281 | * ( But note that we still use it if the TSC is marked | |
282 | * unstable. We do this because unlike Time Of Day, | |
283 | * the scheduler clock tolerates small errors and it's | |
284 | * very important for it to be as fast as the platform | |
3ad2f3fb | 285 | * can achieve it. ) |
0ef95533 | 286 | */ |
10b033d4 | 287 | if (!static_key_false(&__use_tsc)) { |
0ef95533 AK |
288 | /* No locking but a rare wrong value is not a big deal: */ |
289 | return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ); | |
290 | } | |
291 | ||
292 | /* read the Time Stamp Counter: */ | |
4ea1636b | 293 | tsc_now = rdtsc(); |
0ef95533 AK |
294 | |
295 | /* return the value in ns */ | |
20d1c86a | 296 | return cycles_2_ns(tsc_now); |
0ef95533 AK |
297 | } |
298 | ||
299 | /* We need to define a real function for sched_clock, to override the | |
300 | weak default version */ | |
301 | #ifdef CONFIG_PARAVIRT | |
302 | unsigned long long sched_clock(void) | |
303 | { | |
304 | return paravirt_sched_clock(); | |
305 | } | |
306 | #else | |
307 | unsigned long long | |
308 | sched_clock(void) __attribute__((alias("native_sched_clock"))); | |
309 | #endif | |
310 | ||
311 | int check_tsc_unstable(void) | |
312 | { | |
313 | return tsc_unstable; | |
314 | } | |
315 | EXPORT_SYMBOL_GPL(check_tsc_unstable); | |
316 | ||
c73deb6a AH |
317 | int check_tsc_disabled(void) |
318 | { | |
319 | return tsc_disabled; | |
320 | } | |
321 | EXPORT_SYMBOL_GPL(check_tsc_disabled); | |
322 | ||
0ef95533 AK |
323 | #ifdef CONFIG_X86_TSC |
324 | int __init notsc_setup(char *str) | |
325 | { | |
c767a54b | 326 | pr_warn("Kernel compiled with CONFIG_X86_TSC, cannot disable TSC completely\n"); |
0ef95533 AK |
327 | tsc_disabled = 1; |
328 | return 1; | |
329 | } | |
330 | #else | |
331 | /* | |
332 | * disable flag for tsc. Takes effect by clearing the TSC cpu flag | |
333 | * in cpu/common.c | |
334 | */ | |
335 | int __init notsc_setup(char *str) | |
336 | { | |
337 | setup_clear_cpu_cap(X86_FEATURE_TSC); | |
338 | return 1; | |
339 | } | |
340 | #endif | |
341 | ||
342 | __setup("notsc", notsc_setup); | |
bfc0f594 | 343 | |
e82b8e4e VP |
344 | static int no_sched_irq_time; |
345 | ||
395628ef AK |
346 | static int __init tsc_setup(char *str) |
347 | { | |
348 | if (!strcmp(str, "reliable")) | |
349 | tsc_clocksource_reliable = 1; | |
e82b8e4e VP |
350 | if (!strncmp(str, "noirqtime", 9)) |
351 | no_sched_irq_time = 1; | |
395628ef AK |
352 | return 1; |
353 | } | |
354 | ||
355 | __setup("tsc=", tsc_setup); | |
356 | ||
bfc0f594 AK |
357 | #define MAX_RETRIES 5 |
358 | #define SMI_TRESHOLD 50000 | |
359 | ||
360 | /* | |
361 | * Read TSC and the reference counters. Take care of SMI disturbance | |
362 | */ | |
827014be | 363 | static u64 tsc_read_refs(u64 *p, int hpet) |
bfc0f594 AK |
364 | { |
365 | u64 t1, t2; | |
366 | int i; | |
367 | ||
368 | for (i = 0; i < MAX_RETRIES; i++) { | |
369 | t1 = get_cycles(); | |
370 | if (hpet) | |
827014be | 371 | *p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF; |
bfc0f594 | 372 | else |
827014be | 373 | *p = acpi_pm_read_early(); |
bfc0f594 AK |
374 | t2 = get_cycles(); |
375 | if ((t2 - t1) < SMI_TRESHOLD) | |
376 | return t2; | |
377 | } | |
378 | return ULLONG_MAX; | |
379 | } | |
380 | ||
d683ef7a TG |
381 | /* |
382 | * Calculate the TSC frequency from HPET reference | |
bfc0f594 | 383 | */ |
d683ef7a | 384 | static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2) |
bfc0f594 | 385 | { |
d683ef7a | 386 | u64 tmp; |
bfc0f594 | 387 | |
d683ef7a TG |
388 | if (hpet2 < hpet1) |
389 | hpet2 += 0x100000000ULL; | |
390 | hpet2 -= hpet1; | |
391 | tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD)); | |
392 | do_div(tmp, 1000000); | |
393 | do_div(deltatsc, tmp); | |
394 | ||
395 | return (unsigned long) deltatsc; | |
396 | } | |
397 | ||
398 | /* | |
399 | * Calculate the TSC frequency from PMTimer reference | |
400 | */ | |
401 | static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2) | |
402 | { | |
403 | u64 tmp; | |
bfc0f594 | 404 | |
d683ef7a TG |
405 | if (!pm1 && !pm2) |
406 | return ULONG_MAX; | |
407 | ||
408 | if (pm2 < pm1) | |
409 | pm2 += (u64)ACPI_PM_OVRRUN; | |
410 | pm2 -= pm1; | |
411 | tmp = pm2 * 1000000000LL; | |
412 | do_div(tmp, PMTMR_TICKS_PER_SEC); | |
413 | do_div(deltatsc, tmp); | |
414 | ||
415 | return (unsigned long) deltatsc; | |
416 | } | |
417 | ||
a977c400 | 418 | #define CAL_MS 10 |
b7743970 | 419 | #define CAL_LATCH (PIT_TICK_RATE / (1000 / CAL_MS)) |
a977c400 TG |
420 | #define CAL_PIT_LOOPS 1000 |
421 | ||
422 | #define CAL2_MS 50 | |
b7743970 | 423 | #define CAL2_LATCH (PIT_TICK_RATE / (1000 / CAL2_MS)) |
a977c400 TG |
424 | #define CAL2_PIT_LOOPS 5000 |
425 | ||
cce3e057 | 426 | |
ec0c15af LT |
427 | /* |
428 | * Try to calibrate the TSC against the Programmable | |
429 | * Interrupt Timer and return the frequency of the TSC | |
430 | * in kHz. | |
431 | * | |
432 | * Return ULONG_MAX on failure to calibrate. | |
433 | */ | |
a977c400 | 434 | static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin) |
ec0c15af LT |
435 | { |
436 | u64 tsc, t1, t2, delta; | |
437 | unsigned long tscmin, tscmax; | |
438 | int pitcnt; | |
439 | ||
440 | /* Set the Gate high, disable speaker */ | |
441 | outb((inb(0x61) & ~0x02) | 0x01, 0x61); | |
442 | ||
443 | /* | |
444 | * Setup CTC channel 2* for mode 0, (interrupt on terminal | |
445 | * count mode), binary count. Set the latch register to 50ms | |
446 | * (LSB then MSB) to begin countdown. | |
447 | */ | |
448 | outb(0xb0, 0x43); | |
a977c400 TG |
449 | outb(latch & 0xff, 0x42); |
450 | outb(latch >> 8, 0x42); | |
ec0c15af LT |
451 | |
452 | tsc = t1 = t2 = get_cycles(); | |
453 | ||
454 | pitcnt = 0; | |
455 | tscmax = 0; | |
456 | tscmin = ULONG_MAX; | |
457 | while ((inb(0x61) & 0x20) == 0) { | |
458 | t2 = get_cycles(); | |
459 | delta = t2 - tsc; | |
460 | tsc = t2; | |
461 | if ((unsigned long) delta < tscmin) | |
462 | tscmin = (unsigned int) delta; | |
463 | if ((unsigned long) delta > tscmax) | |
464 | tscmax = (unsigned int) delta; | |
465 | pitcnt++; | |
466 | } | |
467 | ||
468 | /* | |
469 | * Sanity checks: | |
470 | * | |
a977c400 | 471 | * If we were not able to read the PIT more than loopmin |
ec0c15af LT |
472 | * times, then we have been hit by a massive SMI |
473 | * | |
474 | * If the maximum is 10 times larger than the minimum, | |
475 | * then we got hit by an SMI as well. | |
476 | */ | |
a977c400 | 477 | if (pitcnt < loopmin || tscmax > 10 * tscmin) |
ec0c15af LT |
478 | return ULONG_MAX; |
479 | ||
480 | /* Calculate the PIT value */ | |
481 | delta = t2 - t1; | |
a977c400 | 482 | do_div(delta, ms); |
ec0c15af LT |
483 | return delta; |
484 | } | |
485 | ||
6ac40ed0 LT |
486 | /* |
487 | * This reads the current MSB of the PIT counter, and | |
488 | * checks if we are running on sufficiently fast and | |
489 | * non-virtualized hardware. | |
490 | * | |
491 | * Our expectations are: | |
492 | * | |
493 | * - the PIT is running at roughly 1.19MHz | |
494 | * | |
495 | * - each IO is going to take about 1us on real hardware, | |
496 | * but we allow it to be much faster (by a factor of 10) or | |
497 | * _slightly_ slower (ie we allow up to a 2us read+counter | |
498 | * update - anything else implies a unacceptably slow CPU | |
499 | * or PIT for the fast calibration to work. | |
500 | * | |
501 | * - with 256 PIT ticks to read the value, we have 214us to | |
502 | * see the same MSB (and overhead like doing a single TSC | |
503 | * read per MSB value etc). | |
504 | * | |
505 | * - We're doing 2 reads per loop (LSB, MSB), and we expect | |
506 | * them each to take about a microsecond on real hardware. | |
507 | * So we expect a count value of around 100. But we'll be | |
508 | * generous, and accept anything over 50. | |
509 | * | |
510 | * - if the PIT is stuck, and we see *many* more reads, we | |
511 | * return early (and the next caller of pit_expect_msb() | |
512 | * then consider it a failure when they don't see the | |
513 | * next expected value). | |
514 | * | |
515 | * These expectations mean that we know that we have seen the | |
516 | * transition from one expected value to another with a fairly | |
517 | * high accuracy, and we didn't miss any events. We can thus | |
518 | * use the TSC value at the transitions to calculate a pretty | |
519 | * good value for the TSC frequencty. | |
520 | */ | |
b6e61eef LT |
521 | static inline int pit_verify_msb(unsigned char val) |
522 | { | |
523 | /* Ignore LSB */ | |
524 | inb(0x42); | |
525 | return inb(0x42) == val; | |
526 | } | |
527 | ||
9e8912e0 | 528 | static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap) |
6ac40ed0 | 529 | { |
9e8912e0 | 530 | int count; |
68f30fbe | 531 | u64 tsc = 0, prev_tsc = 0; |
bfc0f594 | 532 | |
6ac40ed0 | 533 | for (count = 0; count < 50000; count++) { |
b6e61eef | 534 | if (!pit_verify_msb(val)) |
6ac40ed0 | 535 | break; |
68f30fbe | 536 | prev_tsc = tsc; |
9e8912e0 | 537 | tsc = get_cycles(); |
6ac40ed0 | 538 | } |
68f30fbe | 539 | *deltap = get_cycles() - prev_tsc; |
9e8912e0 LT |
540 | *tscp = tsc; |
541 | ||
542 | /* | |
543 | * We require _some_ success, but the quality control | |
544 | * will be based on the error terms on the TSC values. | |
545 | */ | |
546 | return count > 5; | |
6ac40ed0 LT |
547 | } |
548 | ||
549 | /* | |
9e8912e0 LT |
550 | * How many MSB values do we want to see? We aim for |
551 | * a maximum error rate of 500ppm (in practice the | |
552 | * real error is much smaller), but refuse to spend | |
68f30fbe | 553 | * more than 50ms on it. |
6ac40ed0 | 554 | */ |
68f30fbe | 555 | #define MAX_QUICK_PIT_MS 50 |
9e8912e0 | 556 | #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256) |
bfc0f594 | 557 | |
6ac40ed0 LT |
558 | static unsigned long quick_pit_calibrate(void) |
559 | { | |
9e8912e0 LT |
560 | int i; |
561 | u64 tsc, delta; | |
562 | unsigned long d1, d2; | |
563 | ||
6ac40ed0 | 564 | /* Set the Gate high, disable speaker */ |
bfc0f594 AK |
565 | outb((inb(0x61) & ~0x02) | 0x01, 0x61); |
566 | ||
6ac40ed0 LT |
567 | /* |
568 | * Counter 2, mode 0 (one-shot), binary count | |
569 | * | |
570 | * NOTE! Mode 2 decrements by two (and then the | |
571 | * output is flipped each time, giving the same | |
572 | * final output frequency as a decrement-by-one), | |
573 | * so mode 0 is much better when looking at the | |
574 | * individual counts. | |
575 | */ | |
bfc0f594 | 576 | outb(0xb0, 0x43); |
bfc0f594 | 577 | |
6ac40ed0 LT |
578 | /* Start at 0xffff */ |
579 | outb(0xff, 0x42); | |
580 | outb(0xff, 0x42); | |
581 | ||
a6a80e1d LT |
582 | /* |
583 | * The PIT starts counting at the next edge, so we | |
584 | * need to delay for a microsecond. The easiest way | |
585 | * to do that is to just read back the 16-bit counter | |
586 | * once from the PIT. | |
587 | */ | |
b6e61eef | 588 | pit_verify_msb(0); |
a6a80e1d | 589 | |
9e8912e0 LT |
590 | if (pit_expect_msb(0xff, &tsc, &d1)) { |
591 | for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) { | |
592 | if (!pit_expect_msb(0xff-i, &delta, &d2)) | |
593 | break; | |
594 | ||
5aac644a AH |
595 | delta -= tsc; |
596 | ||
597 | /* | |
598 | * Extrapolate the error and fail fast if the error will | |
599 | * never be below 500 ppm. | |
600 | */ | |
601 | if (i == 1 && | |
602 | d1 + d2 >= (delta * MAX_QUICK_PIT_ITERATIONS) >> 11) | |
603 | return 0; | |
604 | ||
9e8912e0 LT |
605 | /* |
606 | * Iterate until the error is less than 500 ppm | |
607 | */ | |
b6e61eef LT |
608 | if (d1+d2 >= delta >> 11) |
609 | continue; | |
610 | ||
611 | /* | |
612 | * Check the PIT one more time to verify that | |
613 | * all TSC reads were stable wrt the PIT. | |
614 | * | |
615 | * This also guarantees serialization of the | |
616 | * last cycle read ('d2') in pit_expect_msb. | |
617 | */ | |
618 | if (!pit_verify_msb(0xfe - i)) | |
619 | break; | |
620 | goto success; | |
6ac40ed0 | 621 | } |
6ac40ed0 | 622 | } |
52045217 | 623 | pr_info("Fast TSC calibration failed\n"); |
6ac40ed0 | 624 | return 0; |
9e8912e0 LT |
625 | |
626 | success: | |
627 | /* | |
628 | * Ok, if we get here, then we've seen the | |
629 | * MSB of the PIT decrement 'i' times, and the | |
630 | * error has shrunk to less than 500 ppm. | |
631 | * | |
632 | * As a result, we can depend on there not being | |
633 | * any odd delays anywhere, and the TSC reads are | |
68f30fbe | 634 | * reliable (within the error). |
9e8912e0 LT |
635 | * |
636 | * kHz = ticks / time-in-seconds / 1000; | |
637 | * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000 | |
638 | * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000) | |
639 | */ | |
9e8912e0 LT |
640 | delta *= PIT_TICK_RATE; |
641 | do_div(delta, i*256*1000); | |
c767a54b | 642 | pr_info("Fast TSC calibration using PIT\n"); |
9e8912e0 | 643 | return delta; |
6ac40ed0 | 644 | } |
ec0c15af | 645 | |
bfc0f594 | 646 | /** |
e93ef949 | 647 | * native_calibrate_tsc - calibrate the tsc on boot |
bfc0f594 | 648 | */ |
e93ef949 | 649 | unsigned long native_calibrate_tsc(void) |
bfc0f594 | 650 | { |
827014be | 651 | u64 tsc1, tsc2, delta, ref1, ref2; |
fbb16e24 | 652 | unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX; |
2d826404 | 653 | unsigned long flags, latch, ms, fast_calibrate; |
a977c400 | 654 | int hpet = is_hpet_enabled(), i, loopmin; |
bfc0f594 | 655 | |
7da7c156 BG |
656 | /* Calibrate TSC using MSR for Intel Atom SoCs */ |
657 | local_irq_save(flags); | |
5f0e0309 | 658 | fast_calibrate = try_msr_calibrate_tsc(); |
7da7c156 | 659 | local_irq_restore(flags); |
5f0e0309 | 660 | if (fast_calibrate) |
7da7c156 | 661 | return fast_calibrate; |
7da7c156 | 662 | |
6ac40ed0 LT |
663 | local_irq_save(flags); |
664 | fast_calibrate = quick_pit_calibrate(); | |
bfc0f594 | 665 | local_irq_restore(flags); |
6ac40ed0 LT |
666 | if (fast_calibrate) |
667 | return fast_calibrate; | |
bfc0f594 | 668 | |
fbb16e24 TG |
669 | /* |
670 | * Run 5 calibration loops to get the lowest frequency value | |
671 | * (the best estimate). We use two different calibration modes | |
672 | * here: | |
673 | * | |
674 | * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and | |
675 | * load a timeout of 50ms. We read the time right after we | |
676 | * started the timer and wait until the PIT count down reaches | |
677 | * zero. In each wait loop iteration we read the TSC and check | |
678 | * the delta to the previous read. We keep track of the min | |
679 | * and max values of that delta. The delta is mostly defined | |
680 | * by the IO time of the PIT access, so we can detect when a | |
0d2eb44f | 681 | * SMI/SMM disturbance happened between the two reads. If the |
fbb16e24 TG |
682 | * maximum time is significantly larger than the minimum time, |
683 | * then we discard the result and have another try. | |
684 | * | |
685 | * 2) Reference counter. If available we use the HPET or the | |
686 | * PMTIMER as a reference to check the sanity of that value. | |
687 | * We use separate TSC readouts and check inside of the | |
688 | * reference read for a SMI/SMM disturbance. We dicard | |
689 | * disturbed values here as well. We do that around the PIT | |
690 | * calibration delay loop as we have to wait for a certain | |
691 | * amount of time anyway. | |
692 | */ | |
a977c400 TG |
693 | |
694 | /* Preset PIT loop values */ | |
695 | latch = CAL_LATCH; | |
696 | ms = CAL_MS; | |
697 | loopmin = CAL_PIT_LOOPS; | |
698 | ||
699 | for (i = 0; i < 3; i++) { | |
ec0c15af | 700 | unsigned long tsc_pit_khz; |
fbb16e24 TG |
701 | |
702 | /* | |
703 | * Read the start value and the reference count of | |
ec0c15af LT |
704 | * hpet/pmtimer when available. Then do the PIT |
705 | * calibration, which will take at least 50ms, and | |
706 | * read the end value. | |
fbb16e24 | 707 | */ |
ec0c15af | 708 | local_irq_save(flags); |
827014be | 709 | tsc1 = tsc_read_refs(&ref1, hpet); |
a977c400 | 710 | tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin); |
827014be | 711 | tsc2 = tsc_read_refs(&ref2, hpet); |
fbb16e24 TG |
712 | local_irq_restore(flags); |
713 | ||
ec0c15af LT |
714 | /* Pick the lowest PIT TSC calibration so far */ |
715 | tsc_pit_min = min(tsc_pit_min, tsc_pit_khz); | |
fbb16e24 TG |
716 | |
717 | /* hpet or pmtimer available ? */ | |
62627bec | 718 | if (ref1 == ref2) |
fbb16e24 TG |
719 | continue; |
720 | ||
721 | /* Check, whether the sampling was disturbed by an SMI */ | |
722 | if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX) | |
723 | continue; | |
724 | ||
725 | tsc2 = (tsc2 - tsc1) * 1000000LL; | |
d683ef7a | 726 | if (hpet) |
827014be | 727 | tsc2 = calc_hpet_ref(tsc2, ref1, ref2); |
d683ef7a | 728 | else |
827014be | 729 | tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2); |
fbb16e24 | 730 | |
fbb16e24 | 731 | tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2); |
a977c400 TG |
732 | |
733 | /* Check the reference deviation */ | |
734 | delta = ((u64) tsc_pit_min) * 100; | |
735 | do_div(delta, tsc_ref_min); | |
736 | ||
737 | /* | |
738 | * If both calibration results are inside a 10% window | |
739 | * then we can be sure, that the calibration | |
740 | * succeeded. We break out of the loop right away. We | |
741 | * use the reference value, as it is more precise. | |
742 | */ | |
743 | if (delta >= 90 && delta <= 110) { | |
c767a54b JP |
744 | pr_info("PIT calibration matches %s. %d loops\n", |
745 | hpet ? "HPET" : "PMTIMER", i + 1); | |
a977c400 | 746 | return tsc_ref_min; |
fbb16e24 TG |
747 | } |
748 | ||
a977c400 TG |
749 | /* |
750 | * Check whether PIT failed more than once. This | |
751 | * happens in virtualized environments. We need to | |
752 | * give the virtual PC a slightly longer timeframe for | |
753 | * the HPET/PMTIMER to make the result precise. | |
754 | */ | |
755 | if (i == 1 && tsc_pit_min == ULONG_MAX) { | |
756 | latch = CAL2_LATCH; | |
757 | ms = CAL2_MS; | |
758 | loopmin = CAL2_PIT_LOOPS; | |
759 | } | |
fbb16e24 | 760 | } |
bfc0f594 AK |
761 | |
762 | /* | |
fbb16e24 | 763 | * Now check the results. |
bfc0f594 | 764 | */ |
fbb16e24 TG |
765 | if (tsc_pit_min == ULONG_MAX) { |
766 | /* PIT gave no useful value */ | |
c767a54b | 767 | pr_warn("Unable to calibrate against PIT\n"); |
fbb16e24 TG |
768 | |
769 | /* We don't have an alternative source, disable TSC */ | |
827014be | 770 | if (!hpet && !ref1 && !ref2) { |
c767a54b | 771 | pr_notice("No reference (HPET/PMTIMER) available\n"); |
fbb16e24 TG |
772 | return 0; |
773 | } | |
774 | ||
775 | /* The alternative source failed as well, disable TSC */ | |
776 | if (tsc_ref_min == ULONG_MAX) { | |
c767a54b | 777 | pr_warn("HPET/PMTIMER calibration failed\n"); |
fbb16e24 TG |
778 | return 0; |
779 | } | |
780 | ||
781 | /* Use the alternative source */ | |
c767a54b JP |
782 | pr_info("using %s reference calibration\n", |
783 | hpet ? "HPET" : "PMTIMER"); | |
fbb16e24 TG |
784 | |
785 | return tsc_ref_min; | |
786 | } | |
bfc0f594 | 787 | |
fbb16e24 | 788 | /* We don't have an alternative source, use the PIT calibration value */ |
827014be | 789 | if (!hpet && !ref1 && !ref2) { |
c767a54b | 790 | pr_info("Using PIT calibration value\n"); |
fbb16e24 | 791 | return tsc_pit_min; |
bfc0f594 AK |
792 | } |
793 | ||
fbb16e24 TG |
794 | /* The alternative source failed, use the PIT calibration value */ |
795 | if (tsc_ref_min == ULONG_MAX) { | |
c767a54b | 796 | pr_warn("HPET/PMTIMER calibration failed. Using PIT calibration.\n"); |
fbb16e24 | 797 | return tsc_pit_min; |
bfc0f594 AK |
798 | } |
799 | ||
fbb16e24 TG |
800 | /* |
801 | * The calibration values differ too much. In doubt, we use | |
802 | * the PIT value as we know that there are PMTIMERs around | |
a977c400 | 803 | * running at double speed. At least we let the user know: |
fbb16e24 | 804 | */ |
c767a54b JP |
805 | pr_warn("PIT calibration deviates from %s: %lu %lu\n", |
806 | hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min); | |
807 | pr_info("Using PIT calibration value\n"); | |
fbb16e24 | 808 | return tsc_pit_min; |
bfc0f594 AK |
809 | } |
810 | ||
bfc0f594 AK |
811 | int recalibrate_cpu_khz(void) |
812 | { | |
813 | #ifndef CONFIG_SMP | |
814 | unsigned long cpu_khz_old = cpu_khz; | |
815 | ||
816 | if (cpu_has_tsc) { | |
2d826404 | 817 | tsc_khz = x86_platform.calibrate_tsc(); |
e93ef949 | 818 | cpu_khz = tsc_khz; |
bfc0f594 AK |
819 | cpu_data(0).loops_per_jiffy = |
820 | cpufreq_scale(cpu_data(0).loops_per_jiffy, | |
821 | cpu_khz_old, cpu_khz); | |
822 | return 0; | |
823 | } else | |
824 | return -ENODEV; | |
825 | #else | |
826 | return -ENODEV; | |
827 | #endif | |
828 | } | |
829 | ||
830 | EXPORT_SYMBOL(recalibrate_cpu_khz); | |
831 | ||
2dbe06fa | 832 | |
cd7240c0 SS |
833 | static unsigned long long cyc2ns_suspend; |
834 | ||
b74f05d6 | 835 | void tsc_save_sched_clock_state(void) |
cd7240c0 | 836 | { |
35af99e6 | 837 | if (!sched_clock_stable()) |
cd7240c0 SS |
838 | return; |
839 | ||
840 | cyc2ns_suspend = sched_clock(); | |
841 | } | |
842 | ||
843 | /* | |
844 | * Even on processors with invariant TSC, TSC gets reset in some the | |
845 | * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to | |
846 | * arbitrary value (still sync'd across cpu's) during resume from such sleep | |
847 | * states. To cope up with this, recompute the cyc2ns_offset for each cpu so | |
848 | * that sched_clock() continues from the point where it was left off during | |
849 | * suspend. | |
850 | */ | |
b74f05d6 | 851 | void tsc_restore_sched_clock_state(void) |
cd7240c0 SS |
852 | { |
853 | unsigned long long offset; | |
854 | unsigned long flags; | |
855 | int cpu; | |
856 | ||
35af99e6 | 857 | if (!sched_clock_stable()) |
cd7240c0 SS |
858 | return; |
859 | ||
860 | local_irq_save(flags); | |
861 | ||
20d1c86a PZ |
862 | /* |
863 | * We're comming out of suspend, there's no concurrency yet; don't | |
864 | * bother being nice about the RCU stuff, just write to both | |
865 | * data fields. | |
866 | */ | |
867 | ||
868 | this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0); | |
869 | this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0); | |
870 | ||
cd7240c0 SS |
871 | offset = cyc2ns_suspend - sched_clock(); |
872 | ||
20d1c86a PZ |
873 | for_each_possible_cpu(cpu) { |
874 | per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset; | |
875 | per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset; | |
876 | } | |
cd7240c0 SS |
877 | |
878 | local_irq_restore(flags); | |
879 | } | |
880 | ||
2dbe06fa AK |
881 | #ifdef CONFIG_CPU_FREQ |
882 | ||
883 | /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency | |
884 | * changes. | |
885 | * | |
886 | * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's | |
887 | * not that important because current Opteron setups do not support | |
888 | * scaling on SMP anyroads. | |
889 | * | |
890 | * Should fix up last_tsc too. Currently gettimeofday in the | |
891 | * first tick after the change will be slightly wrong. | |
892 | */ | |
893 | ||
894 | static unsigned int ref_freq; | |
895 | static unsigned long loops_per_jiffy_ref; | |
896 | static unsigned long tsc_khz_ref; | |
897 | ||
898 | static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, | |
899 | void *data) | |
900 | { | |
901 | struct cpufreq_freqs *freq = data; | |
931db6a3 | 902 | unsigned long *lpj; |
2dbe06fa AK |
903 | |
904 | if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC)) | |
905 | return 0; | |
906 | ||
931db6a3 | 907 | lpj = &boot_cpu_data.loops_per_jiffy; |
2dbe06fa | 908 | #ifdef CONFIG_SMP |
931db6a3 | 909 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) |
2dbe06fa | 910 | lpj = &cpu_data(freq->cpu).loops_per_jiffy; |
2dbe06fa AK |
911 | #endif |
912 | ||
913 | if (!ref_freq) { | |
914 | ref_freq = freq->old; | |
915 | loops_per_jiffy_ref = *lpj; | |
916 | tsc_khz_ref = tsc_khz; | |
917 | } | |
918 | if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) || | |
0b443ead | 919 | (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) { |
878f4f53 | 920 | *lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new); |
2dbe06fa AK |
921 | |
922 | tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new); | |
923 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) | |
924 | mark_tsc_unstable("cpufreq changes"); | |
2dbe06fa | 925 | |
3896c329 PZ |
926 | set_cyc2ns_scale(tsc_khz, freq->cpu); |
927 | } | |
2dbe06fa AK |
928 | |
929 | return 0; | |
930 | } | |
931 | ||
932 | static struct notifier_block time_cpufreq_notifier_block = { | |
933 | .notifier_call = time_cpufreq_notifier | |
934 | }; | |
935 | ||
936 | static int __init cpufreq_tsc(void) | |
937 | { | |
060700b5 LT |
938 | if (!cpu_has_tsc) |
939 | return 0; | |
940 | if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) | |
941 | return 0; | |
2dbe06fa AK |
942 | cpufreq_register_notifier(&time_cpufreq_notifier_block, |
943 | CPUFREQ_TRANSITION_NOTIFIER); | |
944 | return 0; | |
945 | } | |
946 | ||
947 | core_initcall(cpufreq_tsc); | |
948 | ||
949 | #endif /* CONFIG_CPU_FREQ */ | |
8fbbc4b4 AK |
950 | |
951 | /* clocksource code */ | |
952 | ||
953 | static struct clocksource clocksource_tsc; | |
954 | ||
955 | /* | |
09ec5442 | 956 | * We used to compare the TSC to the cycle_last value in the clocksource |
8fbbc4b4 AK |
957 | * structure to avoid a nasty time-warp. This can be observed in a |
958 | * very small window right after one CPU updated cycle_last under | |
959 | * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which | |
960 | * is smaller than the cycle_last reference value due to a TSC which | |
961 | * is slighty behind. This delta is nowhere else observable, but in | |
962 | * that case it results in a forward time jump in the range of hours | |
963 | * due to the unsigned delta calculation of the time keeping core | |
964 | * code, which is necessary to support wrapping clocksources like pm | |
965 | * timer. | |
09ec5442 TG |
966 | * |
967 | * This sanity check is now done in the core timekeeping code. | |
968 | * checking the result of read_tsc() - cycle_last for being negative. | |
969 | * That works because CLOCKSOURCE_MASK(64) does not mask out any bit. | |
8fbbc4b4 | 970 | */ |
8e19608e | 971 | static cycle_t read_tsc(struct clocksource *cs) |
8fbbc4b4 | 972 | { |
27c63405 | 973 | return (cycle_t)rdtsc_ordered(); |
1be39679 MS |
974 | } |
975 | ||
09ec5442 TG |
976 | /* |
977 | * .mask MUST be CLOCKSOURCE_MASK(64). See comment above read_tsc() | |
978 | */ | |
8fbbc4b4 AK |
979 | static struct clocksource clocksource_tsc = { |
980 | .name = "tsc", | |
981 | .rating = 300, | |
982 | .read = read_tsc, | |
983 | .mask = CLOCKSOURCE_MASK(64), | |
8fbbc4b4 AK |
984 | .flags = CLOCK_SOURCE_IS_CONTINUOUS | |
985 | CLOCK_SOURCE_MUST_VERIFY, | |
98d0ac38 | 986 | .archdata = { .vclock_mode = VCLOCK_TSC }, |
8fbbc4b4 AK |
987 | }; |
988 | ||
989 | void mark_tsc_unstable(char *reason) | |
990 | { | |
991 | if (!tsc_unstable) { | |
992 | tsc_unstable = 1; | |
35af99e6 | 993 | clear_sched_clock_stable(); |
e82b8e4e | 994 | disable_sched_clock_irqtime(); |
c767a54b | 995 | pr_info("Marking TSC unstable due to %s\n", reason); |
8fbbc4b4 AK |
996 | /* Change only the rating, when not registered */ |
997 | if (clocksource_tsc.mult) | |
7285dd7f TG |
998 | clocksource_mark_unstable(&clocksource_tsc); |
999 | else { | |
1000 | clocksource_tsc.flags |= CLOCK_SOURCE_UNSTABLE; | |
8fbbc4b4 | 1001 | clocksource_tsc.rating = 0; |
7285dd7f | 1002 | } |
8fbbc4b4 AK |
1003 | } |
1004 | } | |
1005 | ||
1006 | EXPORT_SYMBOL_GPL(mark_tsc_unstable); | |
1007 | ||
395628ef AK |
1008 | static void __init check_system_tsc_reliable(void) |
1009 | { | |
8fbbc4b4 | 1010 | #ifdef CONFIG_MGEODE_LX |
395628ef | 1011 | /* RTSC counts during suspend */ |
8fbbc4b4 | 1012 | #define RTSC_SUSP 0x100 |
8fbbc4b4 AK |
1013 | unsigned long res_low, res_high; |
1014 | ||
1015 | rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high); | |
00097c4f | 1016 | /* Geode_LX - the OLPC CPU has a very reliable TSC */ |
8fbbc4b4 | 1017 | if (res_low & RTSC_SUSP) |
395628ef | 1018 | tsc_clocksource_reliable = 1; |
8fbbc4b4 | 1019 | #endif |
395628ef AK |
1020 | if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE)) |
1021 | tsc_clocksource_reliable = 1; | |
1022 | } | |
8fbbc4b4 AK |
1023 | |
1024 | /* | |
1025 | * Make an educated guess if the TSC is trustworthy and synchronized | |
1026 | * over all CPUs. | |
1027 | */ | |
148f9bb8 | 1028 | int unsynchronized_tsc(void) |
8fbbc4b4 AK |
1029 | { |
1030 | if (!cpu_has_tsc || tsc_unstable) | |
1031 | return 1; | |
1032 | ||
3e5095d1 | 1033 | #ifdef CONFIG_SMP |
8fbbc4b4 AK |
1034 | if (apic_is_clustered_box()) |
1035 | return 1; | |
1036 | #endif | |
1037 | ||
1038 | if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) | |
1039 | return 0; | |
d3b8f889 | 1040 | |
1041 | if (tsc_clocksource_reliable) | |
1042 | return 0; | |
8fbbc4b4 AK |
1043 | /* |
1044 | * Intel systems are normally all synchronized. | |
1045 | * Exceptions must mark TSC as unstable: | |
1046 | */ | |
1047 | if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) { | |
1048 | /* assume multi socket systems are not synchronized: */ | |
1049 | if (num_possible_cpus() > 1) | |
d3b8f889 | 1050 | return 1; |
8fbbc4b4 AK |
1051 | } |
1052 | ||
d3b8f889 | 1053 | return 0; |
8fbbc4b4 AK |
1054 | } |
1055 | ||
08ec0c58 JS |
1056 | |
1057 | static void tsc_refine_calibration_work(struct work_struct *work); | |
1058 | static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work); | |
1059 | /** | |
1060 | * tsc_refine_calibration_work - Further refine tsc freq calibration | |
1061 | * @work - ignored. | |
1062 | * | |
1063 | * This functions uses delayed work over a period of a | |
1064 | * second to further refine the TSC freq value. Since this is | |
1065 | * timer based, instead of loop based, we don't block the boot | |
1066 | * process while this longer calibration is done. | |
1067 | * | |
0d2eb44f | 1068 | * If there are any calibration anomalies (too many SMIs, etc), |
08ec0c58 JS |
1069 | * or the refined calibration is off by 1% of the fast early |
1070 | * calibration, we throw out the new calibration and use the | |
1071 | * early calibration. | |
1072 | */ | |
1073 | static void tsc_refine_calibration_work(struct work_struct *work) | |
1074 | { | |
1075 | static u64 tsc_start = -1, ref_start; | |
1076 | static int hpet; | |
1077 | u64 tsc_stop, ref_stop, delta; | |
1078 | unsigned long freq; | |
1079 | ||
1080 | /* Don't bother refining TSC on unstable systems */ | |
1081 | if (check_tsc_unstable()) | |
1082 | goto out; | |
1083 | ||
1084 | /* | |
1085 | * Since the work is started early in boot, we may be | |
1086 | * delayed the first time we expire. So set the workqueue | |
1087 | * again once we know timers are working. | |
1088 | */ | |
1089 | if (tsc_start == -1) { | |
1090 | /* | |
1091 | * Only set hpet once, to avoid mixing hardware | |
1092 | * if the hpet becomes enabled later. | |
1093 | */ | |
1094 | hpet = is_hpet_enabled(); | |
1095 | schedule_delayed_work(&tsc_irqwork, HZ); | |
1096 | tsc_start = tsc_read_refs(&ref_start, hpet); | |
1097 | return; | |
1098 | } | |
1099 | ||
1100 | tsc_stop = tsc_read_refs(&ref_stop, hpet); | |
1101 | ||
1102 | /* hpet or pmtimer available ? */ | |
62627bec | 1103 | if (ref_start == ref_stop) |
08ec0c58 JS |
1104 | goto out; |
1105 | ||
1106 | /* Check, whether the sampling was disturbed by an SMI */ | |
1107 | if (tsc_start == ULLONG_MAX || tsc_stop == ULLONG_MAX) | |
1108 | goto out; | |
1109 | ||
1110 | delta = tsc_stop - tsc_start; | |
1111 | delta *= 1000000LL; | |
1112 | if (hpet) | |
1113 | freq = calc_hpet_ref(delta, ref_start, ref_stop); | |
1114 | else | |
1115 | freq = calc_pmtimer_ref(delta, ref_start, ref_stop); | |
1116 | ||
1117 | /* Make sure we're within 1% */ | |
1118 | if (abs(tsc_khz - freq) > tsc_khz/100) | |
1119 | goto out; | |
1120 | ||
1121 | tsc_khz = freq; | |
c767a54b JP |
1122 | pr_info("Refined TSC clocksource calibration: %lu.%03lu MHz\n", |
1123 | (unsigned long)tsc_khz / 1000, | |
1124 | (unsigned long)tsc_khz % 1000); | |
08ec0c58 JS |
1125 | |
1126 | out: | |
1127 | clocksource_register_khz(&clocksource_tsc, tsc_khz); | |
1128 | } | |
1129 | ||
1130 | ||
1131 | static int __init init_tsc_clocksource(void) | |
8fbbc4b4 | 1132 | { |
29fe359c | 1133 | if (!cpu_has_tsc || tsc_disabled > 0 || !tsc_khz) |
a8760eca TG |
1134 | return 0; |
1135 | ||
395628ef AK |
1136 | if (tsc_clocksource_reliable) |
1137 | clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY; | |
8fbbc4b4 AK |
1138 | /* lower the rating if we already know its unstable: */ |
1139 | if (check_tsc_unstable()) { | |
1140 | clocksource_tsc.rating = 0; | |
1141 | clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS; | |
1142 | } | |
57779dc2 | 1143 | |
82f9c080 FT |
1144 | if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3)) |
1145 | clocksource_tsc.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP; | |
1146 | ||
57779dc2 AK |
1147 | /* |
1148 | * Trust the results of the earlier calibration on systems | |
1149 | * exporting a reliable TSC. | |
1150 | */ | |
1151 | if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE)) { | |
1152 | clocksource_register_khz(&clocksource_tsc, tsc_khz); | |
1153 | return 0; | |
1154 | } | |
1155 | ||
08ec0c58 JS |
1156 | schedule_delayed_work(&tsc_irqwork, 0); |
1157 | return 0; | |
8fbbc4b4 | 1158 | } |
08ec0c58 JS |
1159 | /* |
1160 | * We use device_initcall here, to ensure we run after the hpet | |
1161 | * is fully initialized, which may occur at fs_initcall time. | |
1162 | */ | |
1163 | device_initcall(init_tsc_clocksource); | |
8fbbc4b4 AK |
1164 | |
1165 | void __init tsc_init(void) | |
1166 | { | |
1167 | u64 lpj; | |
1168 | int cpu; | |
1169 | ||
845b3944 TG |
1170 | x86_init.timers.tsc_pre_init(); |
1171 | ||
b47dcbdc AL |
1172 | if (!cpu_has_tsc) { |
1173 | setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER); | |
8fbbc4b4 | 1174 | return; |
b47dcbdc | 1175 | } |
8fbbc4b4 | 1176 | |
2d826404 | 1177 | tsc_khz = x86_platform.calibrate_tsc(); |
e93ef949 | 1178 | cpu_khz = tsc_khz; |
8fbbc4b4 | 1179 | |
e93ef949 | 1180 | if (!tsc_khz) { |
8fbbc4b4 | 1181 | mark_tsc_unstable("could not calculate TSC khz"); |
b47dcbdc | 1182 | setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER); |
8fbbc4b4 AK |
1183 | return; |
1184 | } | |
1185 | ||
c767a54b JP |
1186 | pr_info("Detected %lu.%03lu MHz processor\n", |
1187 | (unsigned long)cpu_khz / 1000, | |
1188 | (unsigned long)cpu_khz % 1000); | |
8fbbc4b4 AK |
1189 | |
1190 | /* | |
1191 | * Secondary CPUs do not run through tsc_init(), so set up | |
1192 | * all the scale factors for all CPUs, assuming the same | |
1193 | * speed as the bootup CPU. (cpufreq notifiers will fix this | |
1194 | * up if their speed diverges) | |
1195 | */ | |
20d1c86a PZ |
1196 | for_each_possible_cpu(cpu) { |
1197 | cyc2ns_init(cpu); | |
8fbbc4b4 | 1198 | set_cyc2ns_scale(cpu_khz, cpu); |
20d1c86a | 1199 | } |
8fbbc4b4 AK |
1200 | |
1201 | if (tsc_disabled > 0) | |
1202 | return; | |
1203 | ||
1204 | /* now allow native_sched_clock() to use rdtsc */ | |
10b033d4 | 1205 | |
8fbbc4b4 | 1206 | tsc_disabled = 0; |
10b033d4 | 1207 | static_key_slow_inc(&__use_tsc); |
8fbbc4b4 | 1208 | |
e82b8e4e VP |
1209 | if (!no_sched_irq_time) |
1210 | enable_sched_clock_irqtime(); | |
1211 | ||
70de9a97 AK |
1212 | lpj = ((u64)tsc_khz * 1000); |
1213 | do_div(lpj, HZ); | |
1214 | lpj_fine = lpj; | |
1215 | ||
8fbbc4b4 | 1216 | use_tsc_delay(); |
8fbbc4b4 AK |
1217 | |
1218 | if (unsynchronized_tsc()) | |
1219 | mark_tsc_unstable("TSCs unsynchronized"); | |
1220 | ||
395628ef | 1221 | check_system_tsc_reliable(); |
8fbbc4b4 AK |
1222 | } |
1223 | ||
b565201c JS |
1224 | #ifdef CONFIG_SMP |
1225 | /* | |
1226 | * If we have a constant TSC and are using the TSC for the delay loop, | |
1227 | * we can skip clock calibration if another cpu in the same socket has already | |
1228 | * been calibrated. This assumes that CONSTANT_TSC applies to all | |
1229 | * cpus in the socket - this should be a safe assumption. | |
1230 | */ | |
148f9bb8 | 1231 | unsigned long calibrate_delay_is_known(void) |
b565201c JS |
1232 | { |
1233 | int i, cpu = smp_processor_id(); | |
1234 | ||
1235 | if (!tsc_disabled && !cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC)) | |
1236 | return 0; | |
1237 | ||
1238 | for_each_online_cpu(i) | |
1239 | if (cpu_data(i).phys_proc_id == cpu_data(cpu).phys_proc_id) | |
1240 | return cpu_data(i).loops_per_jiffy; | |
1241 | return 0; | |
1242 | } | |
1243 | #endif |