Commit | Line | Data |
---|---|---|
c767a54b JP |
1 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
2 | ||
bfc0f594 | 3 | #include <linux/kernel.h> |
0ef95533 | 4 | #include <linux/sched.h> |
e6017571 | 5 | #include <linux/sched/clock.h> |
0ef95533 | 6 | #include <linux/init.h> |
186f4360 | 7 | #include <linux/export.h> |
0ef95533 | 8 | #include <linux/timer.h> |
bfc0f594 | 9 | #include <linux/acpi_pmtmr.h> |
2dbe06fa | 10 | #include <linux/cpufreq.h> |
8fbbc4b4 AK |
11 | #include <linux/delay.h> |
12 | #include <linux/clocksource.h> | |
13 | #include <linux/percpu.h> | |
08604bd9 | 14 | #include <linux/timex.h> |
10b033d4 | 15 | #include <linux/static_key.h> |
bfc0f594 AK |
16 | |
17 | #include <asm/hpet.h> | |
8fbbc4b4 AK |
18 | #include <asm/timer.h> |
19 | #include <asm/vgtod.h> | |
20 | #include <asm/time.h> | |
21 | #include <asm/delay.h> | |
88b094fb | 22 | #include <asm/hypervisor.h> |
08047c4f | 23 | #include <asm/nmi.h> |
2d826404 | 24 | #include <asm/x86_init.h> |
03da3ff1 | 25 | #include <asm/geode.h> |
6731b0d6 | 26 | #include <asm/apic.h> |
655e52d2 | 27 | #include <asm/intel-family.h> |
30c7e5b1 | 28 | #include <asm/i8259.h> |
2647c43c | 29 | #include <asm/uv/uv.h> |
0ef95533 | 30 | |
f24ade3a | 31 | unsigned int __read_mostly cpu_khz; /* TSC clocks / usec, not used here */ |
0ef95533 | 32 | EXPORT_SYMBOL(cpu_khz); |
f24ade3a IM |
33 | |
34 | unsigned int __read_mostly tsc_khz; | |
0ef95533 AK |
35 | EXPORT_SYMBOL(tsc_khz); |
36 | ||
cf7a63ef PT |
37 | #define KHZ 1000 |
38 | ||
0ef95533 AK |
39 | /* |
40 | * TSC can be unstable due to cpufreq or due to unsynced TSCs | |
41 | */ | |
f24ade3a | 42 | static int __read_mostly tsc_unstable; |
0ef95533 | 43 | |
3bbfafb7 | 44 | static DEFINE_STATIC_KEY_FALSE(__use_tsc); |
10b033d4 | 45 | |
28a00184 | 46 | int tsc_clocksource_reliable; |
57c67da2 | 47 | |
f9677e0f CH |
48 | static u32 art_to_tsc_numerator; |
49 | static u32 art_to_tsc_denominator; | |
50 | static u64 art_to_tsc_offset; | |
51 | struct clocksource *art_related_clocksource; | |
52 | ||
20d1c86a | 53 | struct cyc2ns { |
59eaef78 PZ |
54 | struct cyc2ns_data data[2]; /* 0 + 2*16 = 32 */ |
55 | seqcount_t seq; /* 32 + 4 = 36 */ | |
20d1c86a | 56 | |
59eaef78 | 57 | }; /* fits one cacheline */ |
20d1c86a | 58 | |
59eaef78 | 59 | static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns); |
20d1c86a | 60 | |
4907c68a | 61 | void __always_inline cyc2ns_read_begin(struct cyc2ns_data *data) |
20d1c86a | 62 | { |
59eaef78 | 63 | int seq, idx; |
20d1c86a | 64 | |
59eaef78 | 65 | preempt_disable_notrace(); |
20d1c86a | 66 | |
59eaef78 PZ |
67 | do { |
68 | seq = this_cpu_read(cyc2ns.seq.sequence); | |
69 | idx = seq & 1; | |
20d1c86a | 70 | |
59eaef78 PZ |
71 | data->cyc2ns_offset = this_cpu_read(cyc2ns.data[idx].cyc2ns_offset); |
72 | data->cyc2ns_mul = this_cpu_read(cyc2ns.data[idx].cyc2ns_mul); | |
73 | data->cyc2ns_shift = this_cpu_read(cyc2ns.data[idx].cyc2ns_shift); | |
20d1c86a | 74 | |
59eaef78 | 75 | } while (unlikely(seq != this_cpu_read(cyc2ns.seq.sequence))); |
20d1c86a PZ |
76 | } |
77 | ||
4907c68a | 78 | void __always_inline cyc2ns_read_end(void) |
20d1c86a | 79 | { |
59eaef78 | 80 | preempt_enable_notrace(); |
20d1c86a PZ |
81 | } |
82 | ||
83 | /* | |
84 | * Accelerators for sched_clock() | |
57c67da2 PZ |
85 | * convert from cycles(64bits) => nanoseconds (64bits) |
86 | * basic equation: | |
87 | * ns = cycles / (freq / ns_per_sec) | |
88 | * ns = cycles * (ns_per_sec / freq) | |
89 | * ns = cycles * (10^9 / (cpu_khz * 10^3)) | |
90 | * ns = cycles * (10^6 / cpu_khz) | |
91 | * | |
92 | * Then we use scaling math (suggested by george@mvista.com) to get: | |
93 | * ns = cycles * (10^6 * SC / cpu_khz) / SC | |
94 | * ns = cycles * cyc2ns_scale / SC | |
95 | * | |
96 | * And since SC is a constant power of two, we can convert the div | |
b20112ed AH |
97 | * into a shift. The larger SC is, the more accurate the conversion, but |
98 | * cyc2ns_scale needs to be a 32-bit value so that 32-bit multiplication | |
99 | * (64-bit result) can be used. | |
57c67da2 | 100 | * |
b20112ed | 101 | * We can use khz divisor instead of mhz to keep a better precision. |
57c67da2 PZ |
102 | * (mathieu.desnoyers@polymtl.ca) |
103 | * | |
104 | * -johnstul@us.ibm.com "math is hard, lets go shopping!" | |
105 | */ | |
106 | ||
4907c68a | 107 | static __always_inline unsigned long long cycles_2_ns(unsigned long long cyc) |
57c67da2 | 108 | { |
59eaef78 | 109 | struct cyc2ns_data data; |
20d1c86a PZ |
110 | unsigned long long ns; |
111 | ||
59eaef78 | 112 | cyc2ns_read_begin(&data); |
20d1c86a | 113 | |
59eaef78 PZ |
114 | ns = data.cyc2ns_offset; |
115 | ns += mul_u64_u32_shr(cyc, data.cyc2ns_mul, data.cyc2ns_shift); | |
20d1c86a | 116 | |
59eaef78 | 117 | cyc2ns_read_end(); |
20d1c86a | 118 | |
57c67da2 PZ |
119 | return ns; |
120 | } | |
121 | ||
e2a9ca29 | 122 | static void __set_cyc2ns_scale(unsigned long khz, int cpu, unsigned long long tsc_now) |
57c67da2 | 123 | { |
615cd033 | 124 | unsigned long long ns_now; |
59eaef78 PZ |
125 | struct cyc2ns_data data; |
126 | struct cyc2ns *c2n; | |
20d1c86a | 127 | |
57c67da2 PZ |
128 | ns_now = cycles_2_ns(tsc_now); |
129 | ||
20d1c86a PZ |
130 | /* |
131 | * Compute a new multiplier as per the above comment and ensure our | |
132 | * time function is continuous; see the comment near struct | |
133 | * cyc2ns_data. | |
134 | */ | |
59eaef78 | 135 | clocks_calc_mult_shift(&data.cyc2ns_mul, &data.cyc2ns_shift, khz, |
b20112ed AH |
136 | NSEC_PER_MSEC, 0); |
137 | ||
b9511cd7 AH |
138 | /* |
139 | * cyc2ns_shift is exported via arch_perf_update_userpage() where it is | |
140 | * not expected to be greater than 31 due to the original published | |
141 | * conversion algorithm shifting a 32-bit value (now specifies a 64-bit | |
142 | * value) - refer perf_event_mmap_page documentation in perf_event.h. | |
143 | */ | |
59eaef78 PZ |
144 | if (data.cyc2ns_shift == 32) { |
145 | data.cyc2ns_shift = 31; | |
146 | data.cyc2ns_mul >>= 1; | |
b9511cd7 AH |
147 | } |
148 | ||
59eaef78 PZ |
149 | data.cyc2ns_offset = ns_now - |
150 | mul_u64_u32_shr(tsc_now, data.cyc2ns_mul, data.cyc2ns_shift); | |
151 | ||
152 | c2n = per_cpu_ptr(&cyc2ns, cpu); | |
20d1c86a | 153 | |
59eaef78 PZ |
154 | raw_write_seqcount_latch(&c2n->seq); |
155 | c2n->data[0] = data; | |
156 | raw_write_seqcount_latch(&c2n->seq); | |
157 | c2n->data[1] = data; | |
e2a9ca29 PT |
158 | } |
159 | ||
160 | static void set_cyc2ns_scale(unsigned long khz, int cpu, unsigned long long tsc_now) | |
161 | { | |
162 | unsigned long flags; | |
163 | ||
164 | local_irq_save(flags); | |
165 | sched_clock_idle_sleep_event(); | |
166 | ||
167 | if (khz) | |
168 | __set_cyc2ns_scale(khz, cpu, tsc_now); | |
57c67da2 | 169 | |
ac1e843f | 170 | sched_clock_idle_wakeup_event(); |
57c67da2 PZ |
171 | local_irq_restore(flags); |
172 | } | |
615cd033 | 173 | |
e2a9ca29 PT |
174 | /* |
175 | * Initialize cyc2ns for boot cpu | |
176 | */ | |
177 | static void __init cyc2ns_init_boot_cpu(void) | |
178 | { | |
179 | struct cyc2ns *c2n = this_cpu_ptr(&cyc2ns); | |
180 | ||
181 | seqcount_init(&c2n->seq); | |
182 | __set_cyc2ns_scale(tsc_khz, smp_processor_id(), rdtsc()); | |
183 | } | |
184 | ||
185 | /* | |
608008a4 | 186 | * Secondary CPUs do not run through tsc_init(), so set up |
e2a9ca29 | 187 | * all the scale factors for all CPUs, assuming the same |
c208ac8f | 188 | * speed as the bootup CPU. |
e2a9ca29 PT |
189 | */ |
190 | static void __init cyc2ns_init_secondary_cpus(void) | |
191 | { | |
192 | unsigned int cpu, this_cpu = smp_processor_id(); | |
193 | struct cyc2ns *c2n = this_cpu_ptr(&cyc2ns); | |
194 | struct cyc2ns_data *data = c2n->data; | |
195 | ||
196 | for_each_possible_cpu(cpu) { | |
197 | if (cpu != this_cpu) { | |
198 | seqcount_init(&c2n->seq); | |
199 | c2n = per_cpu_ptr(&cyc2ns, cpu); | |
200 | c2n->data[0] = data[0]; | |
201 | c2n->data[1] = data[1]; | |
202 | } | |
203 | } | |
204 | } | |
205 | ||
0ef95533 AK |
206 | /* |
207 | * Scheduler clock - returns current time in nanosec units. | |
208 | */ | |
209 | u64 native_sched_clock(void) | |
210 | { | |
3bbfafb7 PZ |
211 | if (static_branch_likely(&__use_tsc)) { |
212 | u64 tsc_now = rdtsc(); | |
213 | ||
214 | /* return the value in ns */ | |
215 | return cycles_2_ns(tsc_now); | |
216 | } | |
0ef95533 AK |
217 | |
218 | /* | |
219 | * Fall back to jiffies if there's no TSC available: | |
220 | * ( But note that we still use it if the TSC is marked | |
221 | * unstable. We do this because unlike Time Of Day, | |
222 | * the scheduler clock tolerates small errors and it's | |
223 | * very important for it to be as fast as the platform | |
3ad2f3fb | 224 | * can achieve it. ) |
0ef95533 | 225 | */ |
0ef95533 | 226 | |
3bbfafb7 PZ |
227 | /* No locking but a rare wrong value is not a big deal: */ |
228 | return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ); | |
0ef95533 AK |
229 | } |
230 | ||
a94cab23 AK |
231 | /* |
232 | * Generate a sched_clock if you already have a TSC value. | |
233 | */ | |
234 | u64 native_sched_clock_from_tsc(u64 tsc) | |
235 | { | |
236 | return cycles_2_ns(tsc); | |
237 | } | |
238 | ||
0ef95533 AK |
239 | /* We need to define a real function for sched_clock, to override the |
240 | weak default version */ | |
241 | #ifdef CONFIG_PARAVIRT | |
242 | unsigned long long sched_clock(void) | |
243 | { | |
244 | return paravirt_sched_clock(); | |
245 | } | |
f94c8d11 | 246 | |
698eff63 | 247 | bool using_native_sched_clock(void) |
f94c8d11 | 248 | { |
5c83511b | 249 | return pv_ops.time.sched_clock == native_sched_clock; |
f94c8d11 | 250 | } |
0ef95533 AK |
251 | #else |
252 | unsigned long long | |
253 | sched_clock(void) __attribute__((alias("native_sched_clock"))); | |
f94c8d11 | 254 | |
698eff63 | 255 | bool using_native_sched_clock(void) { return true; } |
0ef95533 AK |
256 | #endif |
257 | ||
258 | int check_tsc_unstable(void) | |
259 | { | |
260 | return tsc_unstable; | |
261 | } | |
262 | EXPORT_SYMBOL_GPL(check_tsc_unstable); | |
263 | ||
264 | #ifdef CONFIG_X86_TSC | |
265 | int __init notsc_setup(char *str) | |
266 | { | |
fe9af81e | 267 | mark_tsc_unstable("boot parameter notsc"); |
0ef95533 AK |
268 | return 1; |
269 | } | |
270 | #else | |
271 | /* | |
272 | * disable flag for tsc. Takes effect by clearing the TSC cpu flag | |
273 | * in cpu/common.c | |
274 | */ | |
275 | int __init notsc_setup(char *str) | |
276 | { | |
277 | setup_clear_cpu_cap(X86_FEATURE_TSC); | |
278 | return 1; | |
279 | } | |
280 | #endif | |
281 | ||
282 | __setup("notsc", notsc_setup); | |
bfc0f594 | 283 | |
e82b8e4e | 284 | static int no_sched_irq_time; |
0f0b7e1c | 285 | static int no_tsc_watchdog; |
e82b8e4e | 286 | |
395628ef AK |
287 | static int __init tsc_setup(char *str) |
288 | { | |
289 | if (!strcmp(str, "reliable")) | |
290 | tsc_clocksource_reliable = 1; | |
e82b8e4e VP |
291 | if (!strncmp(str, "noirqtime", 9)) |
292 | no_sched_irq_time = 1; | |
8309f86c PZ |
293 | if (!strcmp(str, "unstable")) |
294 | mark_tsc_unstable("boot parameter"); | |
0f0b7e1c JL |
295 | if (!strcmp(str, "nowatchdog")) |
296 | no_tsc_watchdog = 1; | |
395628ef AK |
297 | return 1; |
298 | } | |
299 | ||
300 | __setup("tsc=", tsc_setup); | |
301 | ||
a786ef15 DV |
302 | #define MAX_RETRIES 5 |
303 | #define TSC_DEFAULT_THRESHOLD 0x20000 | |
bfc0f594 AK |
304 | |
305 | /* | |
a786ef15 | 306 | * Read TSC and the reference counters. Take care of any disturbances |
bfc0f594 | 307 | */ |
827014be | 308 | static u64 tsc_read_refs(u64 *p, int hpet) |
bfc0f594 AK |
309 | { |
310 | u64 t1, t2; | |
a786ef15 | 311 | u64 thresh = tsc_khz ? tsc_khz >> 5 : TSC_DEFAULT_THRESHOLD; |
bfc0f594 AK |
312 | int i; |
313 | ||
314 | for (i = 0; i < MAX_RETRIES; i++) { | |
315 | t1 = get_cycles(); | |
316 | if (hpet) | |
827014be | 317 | *p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF; |
bfc0f594 | 318 | else |
827014be | 319 | *p = acpi_pm_read_early(); |
bfc0f594 | 320 | t2 = get_cycles(); |
a786ef15 | 321 | if ((t2 - t1) < thresh) |
bfc0f594 AK |
322 | return t2; |
323 | } | |
324 | return ULLONG_MAX; | |
325 | } | |
326 | ||
d683ef7a TG |
327 | /* |
328 | * Calculate the TSC frequency from HPET reference | |
bfc0f594 | 329 | */ |
d683ef7a | 330 | static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2) |
bfc0f594 | 331 | { |
d683ef7a | 332 | u64 tmp; |
bfc0f594 | 333 | |
d683ef7a TG |
334 | if (hpet2 < hpet1) |
335 | hpet2 += 0x100000000ULL; | |
336 | hpet2 -= hpet1; | |
337 | tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD)); | |
338 | do_div(tmp, 1000000); | |
d3878e16 | 339 | deltatsc = div64_u64(deltatsc, tmp); |
d683ef7a TG |
340 | |
341 | return (unsigned long) deltatsc; | |
342 | } | |
343 | ||
344 | /* | |
345 | * Calculate the TSC frequency from PMTimer reference | |
346 | */ | |
347 | static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2) | |
348 | { | |
349 | u64 tmp; | |
bfc0f594 | 350 | |
d683ef7a TG |
351 | if (!pm1 && !pm2) |
352 | return ULONG_MAX; | |
353 | ||
354 | if (pm2 < pm1) | |
355 | pm2 += (u64)ACPI_PM_OVRRUN; | |
356 | pm2 -= pm1; | |
357 | tmp = pm2 * 1000000000LL; | |
358 | do_div(tmp, PMTMR_TICKS_PER_SEC); | |
359 | do_div(deltatsc, tmp); | |
360 | ||
361 | return (unsigned long) deltatsc; | |
362 | } | |
363 | ||
a977c400 | 364 | #define CAL_MS 10 |
b7743970 | 365 | #define CAL_LATCH (PIT_TICK_RATE / (1000 / CAL_MS)) |
a977c400 TG |
366 | #define CAL_PIT_LOOPS 1000 |
367 | ||
368 | #define CAL2_MS 50 | |
b7743970 | 369 | #define CAL2_LATCH (PIT_TICK_RATE / (1000 / CAL2_MS)) |
a977c400 TG |
370 | #define CAL2_PIT_LOOPS 5000 |
371 | ||
cce3e057 | 372 | |
ec0c15af LT |
373 | /* |
374 | * Try to calibrate the TSC against the Programmable | |
375 | * Interrupt Timer and return the frequency of the TSC | |
376 | * in kHz. | |
377 | * | |
378 | * Return ULONG_MAX on failure to calibrate. | |
379 | */ | |
a977c400 | 380 | static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin) |
ec0c15af LT |
381 | { |
382 | u64 tsc, t1, t2, delta; | |
383 | unsigned long tscmin, tscmax; | |
384 | int pitcnt; | |
385 | ||
30c7e5b1 PZ |
386 | if (!has_legacy_pic()) { |
387 | /* | |
388 | * Relies on tsc_early_delay_calibrate() to have given us semi | |
389 | * usable udelay(), wait for the same 50ms we would have with | |
390 | * the PIT loop below. | |
391 | */ | |
392 | udelay(10 * USEC_PER_MSEC); | |
393 | udelay(10 * USEC_PER_MSEC); | |
394 | udelay(10 * USEC_PER_MSEC); | |
395 | udelay(10 * USEC_PER_MSEC); | |
396 | udelay(10 * USEC_PER_MSEC); | |
397 | return ULONG_MAX; | |
398 | } | |
399 | ||
ec0c15af LT |
400 | /* Set the Gate high, disable speaker */ |
401 | outb((inb(0x61) & ~0x02) | 0x01, 0x61); | |
402 | ||
403 | /* | |
404 | * Setup CTC channel 2* for mode 0, (interrupt on terminal | |
405 | * count mode), binary count. Set the latch register to 50ms | |
406 | * (LSB then MSB) to begin countdown. | |
407 | */ | |
408 | outb(0xb0, 0x43); | |
a977c400 TG |
409 | outb(latch & 0xff, 0x42); |
410 | outb(latch >> 8, 0x42); | |
ec0c15af LT |
411 | |
412 | tsc = t1 = t2 = get_cycles(); | |
413 | ||
414 | pitcnt = 0; | |
415 | tscmax = 0; | |
416 | tscmin = ULONG_MAX; | |
417 | while ((inb(0x61) & 0x20) == 0) { | |
418 | t2 = get_cycles(); | |
419 | delta = t2 - tsc; | |
420 | tsc = t2; | |
421 | if ((unsigned long) delta < tscmin) | |
422 | tscmin = (unsigned int) delta; | |
423 | if ((unsigned long) delta > tscmax) | |
424 | tscmax = (unsigned int) delta; | |
425 | pitcnt++; | |
426 | } | |
427 | ||
428 | /* | |
429 | * Sanity checks: | |
430 | * | |
a977c400 | 431 | * If we were not able to read the PIT more than loopmin |
ec0c15af LT |
432 | * times, then we have been hit by a massive SMI |
433 | * | |
434 | * If the maximum is 10 times larger than the minimum, | |
435 | * then we got hit by an SMI as well. | |
436 | */ | |
a977c400 | 437 | if (pitcnt < loopmin || tscmax > 10 * tscmin) |
ec0c15af LT |
438 | return ULONG_MAX; |
439 | ||
440 | /* Calculate the PIT value */ | |
441 | delta = t2 - t1; | |
a977c400 | 442 | do_div(delta, ms); |
ec0c15af LT |
443 | return delta; |
444 | } | |
445 | ||
6ac40ed0 LT |
446 | /* |
447 | * This reads the current MSB of the PIT counter, and | |
448 | * checks if we are running on sufficiently fast and | |
449 | * non-virtualized hardware. | |
450 | * | |
451 | * Our expectations are: | |
452 | * | |
453 | * - the PIT is running at roughly 1.19MHz | |
454 | * | |
455 | * - each IO is going to take about 1us on real hardware, | |
456 | * but we allow it to be much faster (by a factor of 10) or | |
457 | * _slightly_ slower (ie we allow up to a 2us read+counter | |
458 | * update - anything else implies a unacceptably slow CPU | |
459 | * or PIT for the fast calibration to work. | |
460 | * | |
461 | * - with 256 PIT ticks to read the value, we have 214us to | |
462 | * see the same MSB (and overhead like doing a single TSC | |
463 | * read per MSB value etc). | |
464 | * | |
465 | * - We're doing 2 reads per loop (LSB, MSB), and we expect | |
466 | * them each to take about a microsecond on real hardware. | |
467 | * So we expect a count value of around 100. But we'll be | |
468 | * generous, and accept anything over 50. | |
469 | * | |
470 | * - if the PIT is stuck, and we see *many* more reads, we | |
471 | * return early (and the next caller of pit_expect_msb() | |
472 | * then consider it a failure when they don't see the | |
473 | * next expected value). | |
474 | * | |
475 | * These expectations mean that we know that we have seen the | |
476 | * transition from one expected value to another with a fairly | |
477 | * high accuracy, and we didn't miss any events. We can thus | |
478 | * use the TSC value at the transitions to calculate a pretty | |
479 | * good value for the TSC frequencty. | |
480 | */ | |
b6e61eef LT |
481 | static inline int pit_verify_msb(unsigned char val) |
482 | { | |
483 | /* Ignore LSB */ | |
484 | inb(0x42); | |
485 | return inb(0x42) == val; | |
486 | } | |
487 | ||
9e8912e0 | 488 | static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap) |
6ac40ed0 | 489 | { |
9e8912e0 | 490 | int count; |
68f30fbe | 491 | u64 tsc = 0, prev_tsc = 0; |
bfc0f594 | 492 | |
6ac40ed0 | 493 | for (count = 0; count < 50000; count++) { |
b6e61eef | 494 | if (!pit_verify_msb(val)) |
6ac40ed0 | 495 | break; |
68f30fbe | 496 | prev_tsc = tsc; |
9e8912e0 | 497 | tsc = get_cycles(); |
6ac40ed0 | 498 | } |
68f30fbe | 499 | *deltap = get_cycles() - prev_tsc; |
9e8912e0 LT |
500 | *tscp = tsc; |
501 | ||
502 | /* | |
503 | * We require _some_ success, but the quality control | |
504 | * will be based on the error terms on the TSC values. | |
505 | */ | |
506 | return count > 5; | |
6ac40ed0 LT |
507 | } |
508 | ||
509 | /* | |
9e8912e0 LT |
510 | * How many MSB values do we want to see? We aim for |
511 | * a maximum error rate of 500ppm (in practice the | |
512 | * real error is much smaller), but refuse to spend | |
68f30fbe | 513 | * more than 50ms on it. |
6ac40ed0 | 514 | */ |
68f30fbe | 515 | #define MAX_QUICK_PIT_MS 50 |
9e8912e0 | 516 | #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256) |
bfc0f594 | 517 | |
6ac40ed0 LT |
518 | static unsigned long quick_pit_calibrate(void) |
519 | { | |
9e8912e0 LT |
520 | int i; |
521 | u64 tsc, delta; | |
522 | unsigned long d1, d2; | |
523 | ||
30c7e5b1 PZ |
524 | if (!has_legacy_pic()) |
525 | return 0; | |
526 | ||
6ac40ed0 | 527 | /* Set the Gate high, disable speaker */ |
bfc0f594 AK |
528 | outb((inb(0x61) & ~0x02) | 0x01, 0x61); |
529 | ||
6ac40ed0 LT |
530 | /* |
531 | * Counter 2, mode 0 (one-shot), binary count | |
532 | * | |
533 | * NOTE! Mode 2 decrements by two (and then the | |
534 | * output is flipped each time, giving the same | |
535 | * final output frequency as a decrement-by-one), | |
536 | * so mode 0 is much better when looking at the | |
537 | * individual counts. | |
538 | */ | |
bfc0f594 | 539 | outb(0xb0, 0x43); |
bfc0f594 | 540 | |
6ac40ed0 LT |
541 | /* Start at 0xffff */ |
542 | outb(0xff, 0x42); | |
543 | outb(0xff, 0x42); | |
544 | ||
a6a80e1d LT |
545 | /* |
546 | * The PIT starts counting at the next edge, so we | |
547 | * need to delay for a microsecond. The easiest way | |
548 | * to do that is to just read back the 16-bit counter | |
549 | * once from the PIT. | |
550 | */ | |
b6e61eef | 551 | pit_verify_msb(0); |
a6a80e1d | 552 | |
9e8912e0 LT |
553 | if (pit_expect_msb(0xff, &tsc, &d1)) { |
554 | for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) { | |
555 | if (!pit_expect_msb(0xff-i, &delta, &d2)) | |
556 | break; | |
557 | ||
5aac644a AH |
558 | delta -= tsc; |
559 | ||
560 | /* | |
561 | * Extrapolate the error and fail fast if the error will | |
562 | * never be below 500 ppm. | |
563 | */ | |
564 | if (i == 1 && | |
565 | d1 + d2 >= (delta * MAX_QUICK_PIT_ITERATIONS) >> 11) | |
566 | return 0; | |
567 | ||
9e8912e0 LT |
568 | /* |
569 | * Iterate until the error is less than 500 ppm | |
570 | */ | |
b6e61eef LT |
571 | if (d1+d2 >= delta >> 11) |
572 | continue; | |
573 | ||
574 | /* | |
575 | * Check the PIT one more time to verify that | |
576 | * all TSC reads were stable wrt the PIT. | |
577 | * | |
578 | * This also guarantees serialization of the | |
579 | * last cycle read ('d2') in pit_expect_msb. | |
580 | */ | |
581 | if (!pit_verify_msb(0xfe - i)) | |
582 | break; | |
583 | goto success; | |
6ac40ed0 | 584 | } |
6ac40ed0 | 585 | } |
52045217 | 586 | pr_info("Fast TSC calibration failed\n"); |
6ac40ed0 | 587 | return 0; |
9e8912e0 LT |
588 | |
589 | success: | |
590 | /* | |
591 | * Ok, if we get here, then we've seen the | |
592 | * MSB of the PIT decrement 'i' times, and the | |
593 | * error has shrunk to less than 500 ppm. | |
594 | * | |
595 | * As a result, we can depend on there not being | |
596 | * any odd delays anywhere, and the TSC reads are | |
68f30fbe | 597 | * reliable (within the error). |
9e8912e0 LT |
598 | * |
599 | * kHz = ticks / time-in-seconds / 1000; | |
600 | * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000 | |
601 | * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000) | |
602 | */ | |
9e8912e0 LT |
603 | delta *= PIT_TICK_RATE; |
604 | do_div(delta, i*256*1000); | |
c767a54b | 605 | pr_info("Fast TSC calibration using PIT\n"); |
9e8912e0 | 606 | return delta; |
6ac40ed0 | 607 | } |
ec0c15af | 608 | |
bfc0f594 | 609 | /** |
aa297292 LB |
610 | * native_calibrate_tsc |
611 | * Determine TSC frequency via CPUID, else return 0. | |
bfc0f594 | 612 | */ |
e93ef949 | 613 | unsigned long native_calibrate_tsc(void) |
aa297292 LB |
614 | { |
615 | unsigned int eax_denominator, ebx_numerator, ecx_hz, edx; | |
616 | unsigned int crystal_khz; | |
617 | ||
618 | if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) | |
619 | return 0; | |
620 | ||
621 | if (boot_cpu_data.cpuid_level < 0x15) | |
622 | return 0; | |
623 | ||
624 | eax_denominator = ebx_numerator = ecx_hz = edx = 0; | |
625 | ||
626 | /* CPUID 15H TSC/Crystal ratio, plus optionally Crystal Hz */ | |
627 | cpuid(0x15, &eax_denominator, &ebx_numerator, &ecx_hz, &edx); | |
628 | ||
629 | if (ebx_numerator == 0 || eax_denominator == 0) | |
630 | return 0; | |
631 | ||
632 | crystal_khz = ecx_hz / 1000; | |
633 | ||
634 | if (crystal_khz == 0) { | |
635 | switch (boot_cpu_data.x86_model) { | |
655e52d2 PB |
636 | case INTEL_FAM6_SKYLAKE_MOBILE: |
637 | case INTEL_FAM6_SKYLAKE_DESKTOP: | |
6baf3d61 PB |
638 | case INTEL_FAM6_KABYLAKE_MOBILE: |
639 | case INTEL_FAM6_KABYLAKE_DESKTOP: | |
ff4c8663 LB |
640 | crystal_khz = 24000; /* 24.0 MHz */ |
641 | break; | |
f2c4db1b | 642 | case INTEL_FAM6_ATOM_GOLDMONT_X: |
6baf3d61 PB |
643 | crystal_khz = 25000; /* 25.0 MHz */ |
644 | break; | |
655e52d2 | 645 | case INTEL_FAM6_ATOM_GOLDMONT: |
ff4c8663 LB |
646 | crystal_khz = 19200; /* 19.2 MHz */ |
647 | break; | |
aa297292 LB |
648 | } |
649 | } | |
650 | ||
da4ae6c4 LB |
651 | if (crystal_khz == 0) |
652 | return 0; | |
4ca4df0b BG |
653 | /* |
654 | * TSC frequency determined by CPUID is a "hardware reported" | |
655 | * frequency and is the most accurate one so far we have. This | |
656 | * is considered a known frequency. | |
657 | */ | |
658 | setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ); | |
659 | ||
4635fdc6 BG |
660 | /* |
661 | * For Atom SoCs TSC is the only reliable clocksource. | |
662 | * Mark TSC reliable so no watchdog on it. | |
663 | */ | |
664 | if (boot_cpu_data.x86_model == INTEL_FAM6_ATOM_GOLDMONT) | |
665 | setup_force_cpu_cap(X86_FEATURE_TSC_RELIABLE); | |
666 | ||
aa297292 LB |
667 | return crystal_khz * ebx_numerator / eax_denominator; |
668 | } | |
669 | ||
670 | static unsigned long cpu_khz_from_cpuid(void) | |
671 | { | |
672 | unsigned int eax_base_mhz, ebx_max_mhz, ecx_bus_mhz, edx; | |
673 | ||
674 | if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) | |
675 | return 0; | |
676 | ||
677 | if (boot_cpu_data.cpuid_level < 0x16) | |
678 | return 0; | |
679 | ||
680 | eax_base_mhz = ebx_max_mhz = ecx_bus_mhz = edx = 0; | |
681 | ||
682 | cpuid(0x16, &eax_base_mhz, &ebx_max_mhz, &ecx_bus_mhz, &edx); | |
683 | ||
684 | return eax_base_mhz * 1000; | |
685 | } | |
686 | ||
03821f45 PT |
687 | /* |
688 | * calibrate cpu using pit, hpet, and ptimer methods. They are available | |
689 | * later in boot after acpi is initialized. | |
aa297292 | 690 | */ |
03821f45 | 691 | static unsigned long pit_hpet_ptimer_calibrate_cpu(void) |
bfc0f594 | 692 | { |
827014be | 693 | u64 tsc1, tsc2, delta, ref1, ref2; |
fbb16e24 | 694 | unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX; |
03821f45 | 695 | unsigned long flags, latch, ms; |
a977c400 | 696 | int hpet = is_hpet_enabled(), i, loopmin; |
bfc0f594 | 697 | |
fbb16e24 TG |
698 | /* |
699 | * Run 5 calibration loops to get the lowest frequency value | |
700 | * (the best estimate). We use two different calibration modes | |
701 | * here: | |
702 | * | |
703 | * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and | |
704 | * load a timeout of 50ms. We read the time right after we | |
705 | * started the timer and wait until the PIT count down reaches | |
706 | * zero. In each wait loop iteration we read the TSC and check | |
707 | * the delta to the previous read. We keep track of the min | |
708 | * and max values of that delta. The delta is mostly defined | |
a786ef15 DV |
709 | * by the IO time of the PIT access, so we can detect when |
710 | * any disturbance happened between the two reads. If the | |
fbb16e24 TG |
711 | * maximum time is significantly larger than the minimum time, |
712 | * then we discard the result and have another try. | |
713 | * | |
714 | * 2) Reference counter. If available we use the HPET or the | |
715 | * PMTIMER as a reference to check the sanity of that value. | |
716 | * We use separate TSC readouts and check inside of the | |
a786ef15 | 717 | * reference read for any possible disturbance. We dicard |
fbb16e24 TG |
718 | * disturbed values here as well. We do that around the PIT |
719 | * calibration delay loop as we have to wait for a certain | |
720 | * amount of time anyway. | |
721 | */ | |
a977c400 TG |
722 | |
723 | /* Preset PIT loop values */ | |
724 | latch = CAL_LATCH; | |
725 | ms = CAL_MS; | |
726 | loopmin = CAL_PIT_LOOPS; | |
727 | ||
728 | for (i = 0; i < 3; i++) { | |
ec0c15af | 729 | unsigned long tsc_pit_khz; |
fbb16e24 TG |
730 | |
731 | /* | |
732 | * Read the start value and the reference count of | |
ec0c15af LT |
733 | * hpet/pmtimer when available. Then do the PIT |
734 | * calibration, which will take at least 50ms, and | |
735 | * read the end value. | |
fbb16e24 | 736 | */ |
ec0c15af | 737 | local_irq_save(flags); |
827014be | 738 | tsc1 = tsc_read_refs(&ref1, hpet); |
a977c400 | 739 | tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin); |
827014be | 740 | tsc2 = tsc_read_refs(&ref2, hpet); |
fbb16e24 TG |
741 | local_irq_restore(flags); |
742 | ||
ec0c15af LT |
743 | /* Pick the lowest PIT TSC calibration so far */ |
744 | tsc_pit_min = min(tsc_pit_min, tsc_pit_khz); | |
fbb16e24 TG |
745 | |
746 | /* hpet or pmtimer available ? */ | |
62627bec | 747 | if (ref1 == ref2) |
fbb16e24 TG |
748 | continue; |
749 | ||
a786ef15 | 750 | /* Check, whether the sampling was disturbed */ |
fbb16e24 TG |
751 | if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX) |
752 | continue; | |
753 | ||
754 | tsc2 = (tsc2 - tsc1) * 1000000LL; | |
d683ef7a | 755 | if (hpet) |
827014be | 756 | tsc2 = calc_hpet_ref(tsc2, ref1, ref2); |
d683ef7a | 757 | else |
827014be | 758 | tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2); |
fbb16e24 | 759 | |
fbb16e24 | 760 | tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2); |
a977c400 TG |
761 | |
762 | /* Check the reference deviation */ | |
763 | delta = ((u64) tsc_pit_min) * 100; | |
764 | do_div(delta, tsc_ref_min); | |
765 | ||
766 | /* | |
767 | * If both calibration results are inside a 10% window | |
768 | * then we can be sure, that the calibration | |
769 | * succeeded. We break out of the loop right away. We | |
770 | * use the reference value, as it is more precise. | |
771 | */ | |
772 | if (delta >= 90 && delta <= 110) { | |
c767a54b JP |
773 | pr_info("PIT calibration matches %s. %d loops\n", |
774 | hpet ? "HPET" : "PMTIMER", i + 1); | |
a977c400 | 775 | return tsc_ref_min; |
fbb16e24 TG |
776 | } |
777 | ||
a977c400 TG |
778 | /* |
779 | * Check whether PIT failed more than once. This | |
780 | * happens in virtualized environments. We need to | |
781 | * give the virtual PC a slightly longer timeframe for | |
782 | * the HPET/PMTIMER to make the result precise. | |
783 | */ | |
784 | if (i == 1 && tsc_pit_min == ULONG_MAX) { | |
785 | latch = CAL2_LATCH; | |
786 | ms = CAL2_MS; | |
787 | loopmin = CAL2_PIT_LOOPS; | |
788 | } | |
fbb16e24 | 789 | } |
bfc0f594 AK |
790 | |
791 | /* | |
fbb16e24 | 792 | * Now check the results. |
bfc0f594 | 793 | */ |
fbb16e24 TG |
794 | if (tsc_pit_min == ULONG_MAX) { |
795 | /* PIT gave no useful value */ | |
c767a54b | 796 | pr_warn("Unable to calibrate against PIT\n"); |
fbb16e24 TG |
797 | |
798 | /* We don't have an alternative source, disable TSC */ | |
827014be | 799 | if (!hpet && !ref1 && !ref2) { |
c767a54b | 800 | pr_notice("No reference (HPET/PMTIMER) available\n"); |
fbb16e24 TG |
801 | return 0; |
802 | } | |
803 | ||
804 | /* The alternative source failed as well, disable TSC */ | |
805 | if (tsc_ref_min == ULONG_MAX) { | |
c767a54b | 806 | pr_warn("HPET/PMTIMER calibration failed\n"); |
fbb16e24 TG |
807 | return 0; |
808 | } | |
809 | ||
810 | /* Use the alternative source */ | |
c767a54b JP |
811 | pr_info("using %s reference calibration\n", |
812 | hpet ? "HPET" : "PMTIMER"); | |
fbb16e24 TG |
813 | |
814 | return tsc_ref_min; | |
815 | } | |
bfc0f594 | 816 | |
fbb16e24 | 817 | /* We don't have an alternative source, use the PIT calibration value */ |
827014be | 818 | if (!hpet && !ref1 && !ref2) { |
c767a54b | 819 | pr_info("Using PIT calibration value\n"); |
fbb16e24 | 820 | return tsc_pit_min; |
bfc0f594 AK |
821 | } |
822 | ||
fbb16e24 TG |
823 | /* The alternative source failed, use the PIT calibration value */ |
824 | if (tsc_ref_min == ULONG_MAX) { | |
c767a54b | 825 | pr_warn("HPET/PMTIMER calibration failed. Using PIT calibration.\n"); |
fbb16e24 | 826 | return tsc_pit_min; |
bfc0f594 AK |
827 | } |
828 | ||
fbb16e24 TG |
829 | /* |
830 | * The calibration values differ too much. In doubt, we use | |
831 | * the PIT value as we know that there are PMTIMERs around | |
a977c400 | 832 | * running at double speed. At least we let the user know: |
fbb16e24 | 833 | */ |
c767a54b JP |
834 | pr_warn("PIT calibration deviates from %s: %lu %lu\n", |
835 | hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min); | |
836 | pr_info("Using PIT calibration value\n"); | |
fbb16e24 | 837 | return tsc_pit_min; |
bfc0f594 AK |
838 | } |
839 | ||
03821f45 PT |
840 | /** |
841 | * native_calibrate_cpu_early - can calibrate the cpu early in boot | |
842 | */ | |
843 | unsigned long native_calibrate_cpu_early(void) | |
844 | { | |
845 | unsigned long flags, fast_calibrate = cpu_khz_from_cpuid(); | |
846 | ||
847 | if (!fast_calibrate) | |
848 | fast_calibrate = cpu_khz_from_msr(); | |
849 | if (!fast_calibrate) { | |
850 | local_irq_save(flags); | |
851 | fast_calibrate = quick_pit_calibrate(); | |
852 | local_irq_restore(flags); | |
853 | } | |
854 | return fast_calibrate; | |
855 | } | |
856 | ||
857 | ||
858 | /** | |
859 | * native_calibrate_cpu - calibrate the cpu | |
860 | */ | |
8dbe4385 | 861 | static unsigned long native_calibrate_cpu(void) |
03821f45 PT |
862 | { |
863 | unsigned long tsc_freq = native_calibrate_cpu_early(); | |
864 | ||
865 | if (!tsc_freq) | |
866 | tsc_freq = pit_hpet_ptimer_calibrate_cpu(); | |
867 | ||
868 | return tsc_freq; | |
869 | } | |
870 | ||
af576850 | 871 | void recalibrate_cpu_khz(void) |
bfc0f594 AK |
872 | { |
873 | #ifndef CONFIG_SMP | |
874 | unsigned long cpu_khz_old = cpu_khz; | |
875 | ||
eff4677e | 876 | if (!boot_cpu_has(X86_FEATURE_TSC)) |
af576850 | 877 | return; |
eff4677e | 878 | |
aa297292 | 879 | cpu_khz = x86_platform.calibrate_cpu(); |
eff4677e | 880 | tsc_khz = x86_platform.calibrate_tsc(); |
aa297292 LB |
881 | if (tsc_khz == 0) |
882 | tsc_khz = cpu_khz; | |
ff4c8663 LB |
883 | else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz) |
884 | cpu_khz = tsc_khz; | |
eff4677e BP |
885 | cpu_data(0).loops_per_jiffy = cpufreq_scale(cpu_data(0).loops_per_jiffy, |
886 | cpu_khz_old, cpu_khz); | |
bfc0f594 AK |
887 | #endif |
888 | } | |
889 | ||
890 | EXPORT_SYMBOL(recalibrate_cpu_khz); | |
891 | ||
2dbe06fa | 892 | |
cd7240c0 SS |
893 | static unsigned long long cyc2ns_suspend; |
894 | ||
b74f05d6 | 895 | void tsc_save_sched_clock_state(void) |
cd7240c0 | 896 | { |
35af99e6 | 897 | if (!sched_clock_stable()) |
cd7240c0 SS |
898 | return; |
899 | ||
900 | cyc2ns_suspend = sched_clock(); | |
901 | } | |
902 | ||
903 | /* | |
904 | * Even on processors with invariant TSC, TSC gets reset in some the | |
905 | * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to | |
906 | * arbitrary value (still sync'd across cpu's) during resume from such sleep | |
907 | * states. To cope up with this, recompute the cyc2ns_offset for each cpu so | |
908 | * that sched_clock() continues from the point where it was left off during | |
909 | * suspend. | |
910 | */ | |
b74f05d6 | 911 | void tsc_restore_sched_clock_state(void) |
cd7240c0 SS |
912 | { |
913 | unsigned long long offset; | |
914 | unsigned long flags; | |
915 | int cpu; | |
916 | ||
35af99e6 | 917 | if (!sched_clock_stable()) |
cd7240c0 SS |
918 | return; |
919 | ||
920 | local_irq_save(flags); | |
921 | ||
20d1c86a | 922 | /* |
6a6256f9 | 923 | * We're coming out of suspend, there's no concurrency yet; don't |
20d1c86a PZ |
924 | * bother being nice about the RCU stuff, just write to both |
925 | * data fields. | |
926 | */ | |
927 | ||
928 | this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0); | |
929 | this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0); | |
930 | ||
cd7240c0 SS |
931 | offset = cyc2ns_suspend - sched_clock(); |
932 | ||
20d1c86a PZ |
933 | for_each_possible_cpu(cpu) { |
934 | per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset; | |
935 | per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset; | |
936 | } | |
cd7240c0 SS |
937 | |
938 | local_irq_restore(flags); | |
939 | } | |
940 | ||
2dbe06fa | 941 | #ifdef CONFIG_CPU_FREQ |
c208ac8f RW |
942 | /* |
943 | * Frequency scaling support. Adjust the TSC based timer when the CPU frequency | |
2dbe06fa AK |
944 | * changes. |
945 | * | |
c208ac8f RW |
946 | * NOTE: On SMP the situation is not fixable in general, so simply mark the TSC |
947 | * as unstable and give up in those cases. | |
2dbe06fa AK |
948 | * |
949 | * Should fix up last_tsc too. Currently gettimeofday in the | |
950 | * first tick after the change will be slightly wrong. | |
951 | */ | |
952 | ||
953 | static unsigned int ref_freq; | |
954 | static unsigned long loops_per_jiffy_ref; | |
955 | static unsigned long tsc_khz_ref; | |
956 | ||
957 | static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, | |
958 | void *data) | |
959 | { | |
960 | struct cpufreq_freqs *freq = data; | |
2dbe06fa | 961 | |
c208ac8f RW |
962 | if (num_online_cpus() > 1) { |
963 | mark_tsc_unstable("cpufreq changes on SMP"); | |
964 | return 0; | |
965 | } | |
2dbe06fa AK |
966 | |
967 | if (!ref_freq) { | |
968 | ref_freq = freq->old; | |
c208ac8f | 969 | loops_per_jiffy_ref = boot_cpu_data.loops_per_jiffy; |
2dbe06fa AK |
970 | tsc_khz_ref = tsc_khz; |
971 | } | |
c208ac8f | 972 | |
2dbe06fa | 973 | if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) || |
c208ac8f RW |
974 | (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) { |
975 | boot_cpu_data.loops_per_jiffy = | |
976 | cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new); | |
2dbe06fa AK |
977 | |
978 | tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new); | |
979 | if (!(freq->flags & CPUFREQ_CONST_LOOPS)) | |
980 | mark_tsc_unstable("cpufreq changes"); | |
2dbe06fa | 981 | |
df24014a | 982 | set_cyc2ns_scale(tsc_khz, freq->policy->cpu, rdtsc()); |
3896c329 | 983 | } |
2dbe06fa AK |
984 | |
985 | return 0; | |
986 | } | |
987 | ||
988 | static struct notifier_block time_cpufreq_notifier_block = { | |
989 | .notifier_call = time_cpufreq_notifier | |
990 | }; | |
991 | ||
a841cca7 | 992 | static int __init cpufreq_register_tsc_scaling(void) |
2dbe06fa | 993 | { |
59e21e3d | 994 | if (!boot_cpu_has(X86_FEATURE_TSC)) |
060700b5 LT |
995 | return 0; |
996 | if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) | |
997 | return 0; | |
2dbe06fa AK |
998 | cpufreq_register_notifier(&time_cpufreq_notifier_block, |
999 | CPUFREQ_TRANSITION_NOTIFIER); | |
1000 | return 0; | |
1001 | } | |
1002 | ||
a841cca7 | 1003 | core_initcall(cpufreq_register_tsc_scaling); |
2dbe06fa AK |
1004 | |
1005 | #endif /* CONFIG_CPU_FREQ */ | |
8fbbc4b4 | 1006 | |
f9677e0f CH |
1007 | #define ART_CPUID_LEAF (0x15) |
1008 | #define ART_MIN_DENOMINATOR (1) | |
1009 | ||
1010 | ||
1011 | /* | |
1012 | * If ART is present detect the numerator:denominator to convert to TSC | |
1013 | */ | |
120fc3fb | 1014 | static void __init detect_art(void) |
f9677e0f CH |
1015 | { |
1016 | unsigned int unused[2]; | |
1017 | ||
1018 | if (boot_cpu_data.cpuid_level < ART_CPUID_LEAF) | |
1019 | return; | |
1020 | ||
6c66350d | 1021 | /* |
1022 | * Don't enable ART in a VM, non-stop TSC and TSC_ADJUST required, | |
1023 | * and the TSC counter resets must not occur asynchronously. | |
1024 | */ | |
f9677e0f CH |
1025 | if (boot_cpu_has(X86_FEATURE_HYPERVISOR) || |
1026 | !boot_cpu_has(X86_FEATURE_NONSTOP_TSC) || | |
6c66350d | 1027 | !boot_cpu_has(X86_FEATURE_TSC_ADJUST) || |
1028 | tsc_async_resets) | |
f9677e0f CH |
1029 | return; |
1030 | ||
7b3d2f6e TG |
1031 | cpuid(ART_CPUID_LEAF, &art_to_tsc_denominator, |
1032 | &art_to_tsc_numerator, unused, unused+1); | |
1033 | ||
1034 | if (art_to_tsc_denominator < ART_MIN_DENOMINATOR) | |
f9677e0f CH |
1035 | return; |
1036 | ||
7b3d2f6e TG |
1037 | rdmsrl(MSR_IA32_TSC_ADJUST, art_to_tsc_offset); |
1038 | ||
f9677e0f CH |
1039 | /* Make this sticky over multiple CPU init calls */ |
1040 | setup_force_cpu_cap(X86_FEATURE_ART); | |
1041 | } | |
1042 | ||
1043 | ||
8fbbc4b4 AK |
1044 | /* clocksource code */ |
1045 | ||
6a369583 TG |
1046 | static void tsc_resume(struct clocksource *cs) |
1047 | { | |
1048 | tsc_verify_tsc_adjust(true); | |
1049 | } | |
1050 | ||
8fbbc4b4 | 1051 | /* |
09ec5442 | 1052 | * We used to compare the TSC to the cycle_last value in the clocksource |
8fbbc4b4 AK |
1053 | * structure to avoid a nasty time-warp. This can be observed in a |
1054 | * very small window right after one CPU updated cycle_last under | |
1055 | * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which | |
1056 | * is smaller than the cycle_last reference value due to a TSC which | |
1057 | * is slighty behind. This delta is nowhere else observable, but in | |
1058 | * that case it results in a forward time jump in the range of hours | |
1059 | * due to the unsigned delta calculation of the time keeping core | |
1060 | * code, which is necessary to support wrapping clocksources like pm | |
1061 | * timer. | |
09ec5442 TG |
1062 | * |
1063 | * This sanity check is now done in the core timekeeping code. | |
1064 | * checking the result of read_tsc() - cycle_last for being negative. | |
1065 | * That works because CLOCKSOURCE_MASK(64) does not mask out any bit. | |
8fbbc4b4 | 1066 | */ |
a5a1d1c2 | 1067 | static u64 read_tsc(struct clocksource *cs) |
8fbbc4b4 | 1068 | { |
a5a1d1c2 | 1069 | return (u64)rdtsc_ordered(); |
1be39679 MS |
1070 | } |
1071 | ||
12907fbb TG |
1072 | static void tsc_cs_mark_unstable(struct clocksource *cs) |
1073 | { | |
1074 | if (tsc_unstable) | |
1075 | return; | |
f94c8d11 | 1076 | |
12907fbb | 1077 | tsc_unstable = 1; |
f94c8d11 PZ |
1078 | if (using_native_sched_clock()) |
1079 | clear_sched_clock_stable(); | |
12907fbb TG |
1080 | disable_sched_clock_irqtime(); |
1081 | pr_info("Marking TSC unstable due to clocksource watchdog\n"); | |
1082 | } | |
1083 | ||
b421b22b PZ |
1084 | static void tsc_cs_tick_stable(struct clocksource *cs) |
1085 | { | |
1086 | if (tsc_unstable) | |
1087 | return; | |
1088 | ||
1089 | if (using_native_sched_clock()) | |
1090 | sched_clock_tick_stable(); | |
1091 | } | |
1092 | ||
09ec5442 TG |
1093 | /* |
1094 | * .mask MUST be CLOCKSOURCE_MASK(64). See comment above read_tsc() | |
1095 | */ | |
aa83c457 PZ |
1096 | static struct clocksource clocksource_tsc_early = { |
1097 | .name = "tsc-early", | |
1098 | .rating = 299, | |
1099 | .read = read_tsc, | |
1100 | .mask = CLOCKSOURCE_MASK(64), | |
1101 | .flags = CLOCK_SOURCE_IS_CONTINUOUS | | |
1102 | CLOCK_SOURCE_MUST_VERIFY, | |
1103 | .archdata = { .vclock_mode = VCLOCK_TSC }, | |
1104 | .resume = tsc_resume, | |
1105 | .mark_unstable = tsc_cs_mark_unstable, | |
1106 | .tick_stable = tsc_cs_tick_stable, | |
e3b4f790 | 1107 | .list = LIST_HEAD_INIT(clocksource_tsc_early.list), |
aa83c457 PZ |
1108 | }; |
1109 | ||
1110 | /* | |
1111 | * Must mark VALID_FOR_HRES early such that when we unregister tsc_early | |
1112 | * this one will immediately take over. We will only register if TSC has | |
1113 | * been found good. | |
1114 | */ | |
8fbbc4b4 AK |
1115 | static struct clocksource clocksource_tsc = { |
1116 | .name = "tsc", | |
1117 | .rating = 300, | |
1118 | .read = read_tsc, | |
1119 | .mask = CLOCKSOURCE_MASK(64), | |
8fbbc4b4 | 1120 | .flags = CLOCK_SOURCE_IS_CONTINUOUS | |
aa83c457 | 1121 | CLOCK_SOURCE_VALID_FOR_HRES | |
8fbbc4b4 | 1122 | CLOCK_SOURCE_MUST_VERIFY, |
98d0ac38 | 1123 | .archdata = { .vclock_mode = VCLOCK_TSC }, |
6a369583 | 1124 | .resume = tsc_resume, |
12907fbb | 1125 | .mark_unstable = tsc_cs_mark_unstable, |
b421b22b | 1126 | .tick_stable = tsc_cs_tick_stable, |
e3b4f790 | 1127 | .list = LIST_HEAD_INIT(clocksource_tsc.list), |
8fbbc4b4 AK |
1128 | }; |
1129 | ||
1130 | void mark_tsc_unstable(char *reason) | |
1131 | { | |
f94c8d11 PZ |
1132 | if (tsc_unstable) |
1133 | return; | |
1134 | ||
1135 | tsc_unstable = 1; | |
1136 | if (using_native_sched_clock()) | |
35af99e6 | 1137 | clear_sched_clock_stable(); |
f94c8d11 PZ |
1138 | disable_sched_clock_irqtime(); |
1139 | pr_info("Marking TSC unstable due to %s\n", reason); | |
e3b4f790 PZ |
1140 | |
1141 | clocksource_mark_unstable(&clocksource_tsc_early); | |
1142 | clocksource_mark_unstable(&clocksource_tsc); | |
8fbbc4b4 AK |
1143 | } |
1144 | ||
1145 | EXPORT_SYMBOL_GPL(mark_tsc_unstable); | |
1146 | ||
395628ef AK |
1147 | static void __init check_system_tsc_reliable(void) |
1148 | { | |
03da3ff1 DW |
1149 | #if defined(CONFIG_MGEODEGX1) || defined(CONFIG_MGEODE_LX) || defined(CONFIG_X86_GENERIC) |
1150 | if (is_geode_lx()) { | |
1151 | /* RTSC counts during suspend */ | |
8fbbc4b4 | 1152 | #define RTSC_SUSP 0x100 |
03da3ff1 | 1153 | unsigned long res_low, res_high; |
8fbbc4b4 | 1154 | |
03da3ff1 DW |
1155 | rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high); |
1156 | /* Geode_LX - the OLPC CPU has a very reliable TSC */ | |
1157 | if (res_low & RTSC_SUSP) | |
1158 | tsc_clocksource_reliable = 1; | |
1159 | } | |
8fbbc4b4 | 1160 | #endif |
395628ef AK |
1161 | if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE)) |
1162 | tsc_clocksource_reliable = 1; | |
1163 | } | |
8fbbc4b4 AK |
1164 | |
1165 | /* | |
1166 | * Make an educated guess if the TSC is trustworthy and synchronized | |
1167 | * over all CPUs. | |
1168 | */ | |
148f9bb8 | 1169 | int unsynchronized_tsc(void) |
8fbbc4b4 | 1170 | { |
59e21e3d | 1171 | if (!boot_cpu_has(X86_FEATURE_TSC) || tsc_unstable) |
8fbbc4b4 AK |
1172 | return 1; |
1173 | ||
3e5095d1 | 1174 | #ifdef CONFIG_SMP |
8fbbc4b4 AK |
1175 | if (apic_is_clustered_box()) |
1176 | return 1; | |
1177 | #endif | |
1178 | ||
1179 | if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) | |
1180 | return 0; | |
d3b8f889 | 1181 | |
1182 | if (tsc_clocksource_reliable) | |
1183 | return 0; | |
8fbbc4b4 AK |
1184 | /* |
1185 | * Intel systems are normally all synchronized. | |
1186 | * Exceptions must mark TSC as unstable: | |
1187 | */ | |
1188 | if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) { | |
1189 | /* assume multi socket systems are not synchronized: */ | |
1190 | if (num_possible_cpus() > 1) | |
d3b8f889 | 1191 | return 1; |
8fbbc4b4 AK |
1192 | } |
1193 | ||
d3b8f889 | 1194 | return 0; |
8fbbc4b4 AK |
1195 | } |
1196 | ||
f9677e0f CH |
1197 | /* |
1198 | * Convert ART to TSC given numerator/denominator found in detect_art() | |
1199 | */ | |
a5a1d1c2 | 1200 | struct system_counterval_t convert_art_to_tsc(u64 art) |
f9677e0f CH |
1201 | { |
1202 | u64 tmp, res, rem; | |
1203 | ||
1204 | rem = do_div(art, art_to_tsc_denominator); | |
1205 | ||
1206 | res = art * art_to_tsc_numerator; | |
1207 | tmp = rem * art_to_tsc_numerator; | |
1208 | ||
1209 | do_div(tmp, art_to_tsc_denominator); | |
1210 | res += tmp + art_to_tsc_offset; | |
1211 | ||
1212 | return (struct system_counterval_t) {.cs = art_related_clocksource, | |
1213 | .cycles = res}; | |
1214 | } | |
1215 | EXPORT_SYMBOL(convert_art_to_tsc); | |
08ec0c58 | 1216 | |
fc804f65 RJ |
1217 | /** |
1218 | * convert_art_ns_to_tsc() - Convert ART in nanoseconds to TSC. | |
1219 | * @art_ns: ART (Always Running Timer) in unit of nanoseconds | |
1220 | * | |
1221 | * PTM requires all timestamps to be in units of nanoseconds. When user | |
1222 | * software requests a cross-timestamp, this function converts system timestamp | |
1223 | * to TSC. | |
1224 | * | |
1225 | * This is valid when CPU feature flag X86_FEATURE_TSC_KNOWN_FREQ is set | |
1226 | * indicating the tsc_khz is derived from CPUID[15H]. Drivers should check | |
1227 | * that this flag is set before conversion to TSC is attempted. | |
1228 | * | |
1229 | * Return: | |
1230 | * struct system_counterval_t - system counter value with the pointer to the | |
1231 | * corresponding clocksource | |
1232 | * @cycles: System counter value | |
1233 | * @cs: Clocksource corresponding to system counter value. Used | |
1234 | * by timekeeping code to verify comparibility of two cycle | |
1235 | * values. | |
1236 | */ | |
1237 | ||
1238 | struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns) | |
1239 | { | |
1240 | u64 tmp, res, rem; | |
1241 | ||
1242 | rem = do_div(art_ns, USEC_PER_SEC); | |
1243 | ||
1244 | res = art_ns * tsc_khz; | |
1245 | tmp = rem * tsc_khz; | |
1246 | ||
1247 | do_div(tmp, USEC_PER_SEC); | |
1248 | res += tmp; | |
1249 | ||
1250 | return (struct system_counterval_t) { .cs = art_related_clocksource, | |
1251 | .cycles = res}; | |
1252 | } | |
1253 | EXPORT_SYMBOL(convert_art_ns_to_tsc); | |
1254 | ||
1255 | ||
08ec0c58 JS |
1256 | static void tsc_refine_calibration_work(struct work_struct *work); |
1257 | static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work); | |
1258 | /** | |
1259 | * tsc_refine_calibration_work - Further refine tsc freq calibration | |
1260 | * @work - ignored. | |
1261 | * | |
1262 | * This functions uses delayed work over a period of a | |
1263 | * second to further refine the TSC freq value. Since this is | |
1264 | * timer based, instead of loop based, we don't block the boot | |
1265 | * process while this longer calibration is done. | |
1266 | * | |
0d2eb44f | 1267 | * If there are any calibration anomalies (too many SMIs, etc), |
08ec0c58 JS |
1268 | * or the refined calibration is off by 1% of the fast early |
1269 | * calibration, we throw out the new calibration and use the | |
1270 | * early calibration. | |
1271 | */ | |
1272 | static void tsc_refine_calibration_work(struct work_struct *work) | |
1273 | { | |
a786ef15 | 1274 | static u64 tsc_start = ULLONG_MAX, ref_start; |
08ec0c58 JS |
1275 | static int hpet; |
1276 | u64 tsc_stop, ref_stop, delta; | |
1277 | unsigned long freq; | |
aa7b630e | 1278 | int cpu; |
08ec0c58 JS |
1279 | |
1280 | /* Don't bother refining TSC on unstable systems */ | |
aa83c457 | 1281 | if (tsc_unstable) |
e9088add | 1282 | goto unreg; |
08ec0c58 JS |
1283 | |
1284 | /* | |
1285 | * Since the work is started early in boot, we may be | |
1286 | * delayed the first time we expire. So set the workqueue | |
1287 | * again once we know timers are working. | |
1288 | */ | |
a786ef15 DV |
1289 | if (tsc_start == ULLONG_MAX) { |
1290 | restart: | |
08ec0c58 JS |
1291 | /* |
1292 | * Only set hpet once, to avoid mixing hardware | |
1293 | * if the hpet becomes enabled later. | |
1294 | */ | |
1295 | hpet = is_hpet_enabled(); | |
08ec0c58 | 1296 | tsc_start = tsc_read_refs(&ref_start, hpet); |
a786ef15 | 1297 | schedule_delayed_work(&tsc_irqwork, HZ); |
08ec0c58 JS |
1298 | return; |
1299 | } | |
1300 | ||
1301 | tsc_stop = tsc_read_refs(&ref_stop, hpet); | |
1302 | ||
1303 | /* hpet or pmtimer available ? */ | |
62627bec | 1304 | if (ref_start == ref_stop) |
08ec0c58 JS |
1305 | goto out; |
1306 | ||
a786ef15 DV |
1307 | /* Check, whether the sampling was disturbed */ |
1308 | if (tsc_stop == ULLONG_MAX) | |
1309 | goto restart; | |
08ec0c58 JS |
1310 | |
1311 | delta = tsc_stop - tsc_start; | |
1312 | delta *= 1000000LL; | |
1313 | if (hpet) | |
1314 | freq = calc_hpet_ref(delta, ref_start, ref_stop); | |
1315 | else | |
1316 | freq = calc_pmtimer_ref(delta, ref_start, ref_stop); | |
1317 | ||
1318 | /* Make sure we're within 1% */ | |
1319 | if (abs(tsc_khz - freq) > tsc_khz/100) | |
1320 | goto out; | |
1321 | ||
1322 | tsc_khz = freq; | |
c767a54b JP |
1323 | pr_info("Refined TSC clocksource calibration: %lu.%03lu MHz\n", |
1324 | (unsigned long)tsc_khz / 1000, | |
1325 | (unsigned long)tsc_khz % 1000); | |
08ec0c58 | 1326 | |
6731b0d6 NS |
1327 | /* Inform the TSC deadline clockevent devices about the recalibration */ |
1328 | lapic_update_tsc_freq(); | |
1329 | ||
aa7b630e PZ |
1330 | /* Update the sched_clock() rate to match the clocksource one */ |
1331 | for_each_possible_cpu(cpu) | |
5c3c2ea6 | 1332 | set_cyc2ns_scale(tsc_khz, cpu, tsc_stop); |
aa7b630e | 1333 | |
08ec0c58 | 1334 | out: |
aa83c457 | 1335 | if (tsc_unstable) |
e9088add | 1336 | goto unreg; |
aa83c457 | 1337 | |
f9677e0f CH |
1338 | if (boot_cpu_has(X86_FEATURE_ART)) |
1339 | art_related_clocksource = &clocksource_tsc; | |
08ec0c58 | 1340 | clocksource_register_khz(&clocksource_tsc, tsc_khz); |
e9088add | 1341 | unreg: |
aa83c457 | 1342 | clocksource_unregister(&clocksource_tsc_early); |
08ec0c58 JS |
1343 | } |
1344 | ||
1345 | ||
1346 | static int __init init_tsc_clocksource(void) | |
8fbbc4b4 | 1347 | { |
fe9af81e | 1348 | if (!boot_cpu_has(X86_FEATURE_TSC) || !tsc_khz) |
a8760eca TG |
1349 | return 0; |
1350 | ||
e9088add PZ |
1351 | if (tsc_unstable) |
1352 | goto unreg; | |
aa83c457 | 1353 | |
0f0b7e1c | 1354 | if (tsc_clocksource_reliable || no_tsc_watchdog) |
395628ef | 1355 | clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY; |
57779dc2 | 1356 | |
82f9c080 FT |
1357 | if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3)) |
1358 | clocksource_tsc.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP; | |
1359 | ||
57779dc2 | 1360 | /* |
47c95a46 BG |
1361 | * When TSC frequency is known (retrieved via MSR or CPUID), we skip |
1362 | * the refined calibration and directly register it as a clocksource. | |
57779dc2 | 1363 | */ |
984feceb | 1364 | if (boot_cpu_has(X86_FEATURE_TSC_KNOWN_FREQ)) { |
44fee88c PZ |
1365 | if (boot_cpu_has(X86_FEATURE_ART)) |
1366 | art_related_clocksource = &clocksource_tsc; | |
57779dc2 | 1367 | clocksource_register_khz(&clocksource_tsc, tsc_khz); |
e9088add | 1368 | unreg: |
aa83c457 | 1369 | clocksource_unregister(&clocksource_tsc_early); |
57779dc2 AK |
1370 | return 0; |
1371 | } | |
1372 | ||
08ec0c58 JS |
1373 | schedule_delayed_work(&tsc_irqwork, 0); |
1374 | return 0; | |
8fbbc4b4 | 1375 | } |
08ec0c58 JS |
1376 | /* |
1377 | * We use device_initcall here, to ensure we run after the hpet | |
1378 | * is fully initialized, which may occur at fs_initcall time. | |
1379 | */ | |
1380 | device_initcall(init_tsc_clocksource); | |
8fbbc4b4 | 1381 | |
8dbe4385 | 1382 | static bool __init determine_cpu_tsc_frequencies(bool early) |
8fbbc4b4 | 1383 | { |
cf7a63ef PT |
1384 | /* Make sure that cpu and tsc are not already calibrated */ |
1385 | WARN_ON(cpu_khz || tsc_khz); | |
8fbbc4b4 | 1386 | |
8dbe4385 PT |
1387 | if (early) { |
1388 | cpu_khz = x86_platform.calibrate_cpu(); | |
1389 | tsc_khz = x86_platform.calibrate_tsc(); | |
1390 | } else { | |
1391 | /* We should not be here with non-native cpu calibration */ | |
1392 | WARN_ON(x86_platform.calibrate_cpu != native_calibrate_cpu); | |
1393 | cpu_khz = pit_hpet_ptimer_calibrate_cpu(); | |
1394 | } | |
ff4c8663 LB |
1395 | |
1396 | /* | |
608008a4 | 1397 | * Trust non-zero tsc_khz as authoritative, |
ff4c8663 LB |
1398 | * and use it to sanity check cpu_khz, |
1399 | * which will be off if system timer is off. | |
1400 | */ | |
aa297292 LB |
1401 | if (tsc_khz == 0) |
1402 | tsc_khz = cpu_khz; | |
ff4c8663 LB |
1403 | else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz) |
1404 | cpu_khz = tsc_khz; | |
8fbbc4b4 | 1405 | |
cf7a63ef PT |
1406 | if (tsc_khz == 0) |
1407 | return false; | |
8fbbc4b4 | 1408 | |
c767a54b | 1409 | pr_info("Detected %lu.%03lu MHz processor\n", |
cf7a63ef PT |
1410 | (unsigned long)cpu_khz / KHZ, |
1411 | (unsigned long)cpu_khz % KHZ); | |
8fbbc4b4 | 1412 | |
4b5b2127 LB |
1413 | if (cpu_khz != tsc_khz) { |
1414 | pr_info("Detected %lu.%03lu MHz TSC", | |
cf7a63ef PT |
1415 | (unsigned long)tsc_khz / KHZ, |
1416 | (unsigned long)tsc_khz % KHZ); | |
1417 | } | |
1418 | return true; | |
1419 | } | |
1420 | ||
1421 | static unsigned long __init get_loops_per_jiffy(void) | |
1422 | { | |
17f6bac2 | 1423 | u64 lpj = (u64)tsc_khz * KHZ; |
cf7a63ef PT |
1424 | |
1425 | do_div(lpj, HZ); | |
1426 | return lpj; | |
1427 | } | |
1428 | ||
608008a4 DL |
1429 | static void __init tsc_enable_sched_clock(void) |
1430 | { | |
1431 | /* Sanitize TSC ADJUST before cyc2ns gets initialized */ | |
1432 | tsc_store_and_check_tsc_adjust(true); | |
1433 | cyc2ns_init_boot_cpu(); | |
1434 | static_branch_enable(&__use_tsc); | |
1435 | } | |
1436 | ||
cf7a63ef PT |
1437 | void __init tsc_early_init(void) |
1438 | { | |
1439 | if (!boot_cpu_has(X86_FEATURE_TSC)) | |
1440 | return; | |
2647c43c MT |
1441 | /* Don't change UV TSC multi-chassis synchronization */ |
1442 | if (is_early_uv_system()) | |
1443 | return; | |
8dbe4385 | 1444 | if (!determine_cpu_tsc_frequencies(true)) |
cf7a63ef PT |
1445 | return; |
1446 | loops_per_jiffy = get_loops_per_jiffy(); | |
e2a9ca29 | 1447 | |
608008a4 | 1448 | tsc_enable_sched_clock(); |
cf7a63ef PT |
1449 | } |
1450 | ||
1451 | void __init tsc_init(void) | |
1452 | { | |
8dbe4385 PT |
1453 | /* |
1454 | * native_calibrate_cpu_early can only calibrate using methods that are | |
1455 | * available early in boot. | |
1456 | */ | |
1457 | if (x86_platform.calibrate_cpu == native_calibrate_cpu_early) | |
1458 | x86_platform.calibrate_cpu = native_calibrate_cpu; | |
1459 | ||
cf7a63ef PT |
1460 | if (!boot_cpu_has(X86_FEATURE_TSC)) { |
1461 | setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER); | |
1462 | return; | |
1463 | } | |
1464 | ||
1465 | if (!tsc_khz) { | |
1466 | /* We failed to determine frequencies earlier, try again */ | |
8dbe4385 | 1467 | if (!determine_cpu_tsc_frequencies(false)) { |
cf7a63ef PT |
1468 | mark_tsc_unstable("could not calculate TSC khz"); |
1469 | setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER); | |
1470 | return; | |
1471 | } | |
608008a4 | 1472 | tsc_enable_sched_clock(); |
4b5b2127 LB |
1473 | } |
1474 | ||
e2a9ca29 | 1475 | cyc2ns_init_secondary_cpus(); |
8fbbc4b4 | 1476 | |
e82b8e4e VP |
1477 | if (!no_sched_irq_time) |
1478 | enable_sched_clock_irqtime(); | |
1479 | ||
cf7a63ef | 1480 | lpj_fine = get_loops_per_jiffy(); |
8fbbc4b4 | 1481 | use_tsc_delay(); |
8fbbc4b4 | 1482 | |
a1272dd5 ZD |
1483 | check_system_tsc_reliable(); |
1484 | ||
aa83c457 | 1485 | if (unsynchronized_tsc()) { |
8fbbc4b4 | 1486 | mark_tsc_unstable("TSCs unsynchronized"); |
aa83c457 PZ |
1487 | return; |
1488 | } | |
8fbbc4b4 | 1489 | |
aa83c457 | 1490 | clocksource_register_khz(&clocksource_tsc_early, tsc_khz); |
f9677e0f | 1491 | detect_art(); |
8fbbc4b4 AK |
1492 | } |
1493 | ||
b565201c JS |
1494 | #ifdef CONFIG_SMP |
1495 | /* | |
1496 | * If we have a constant TSC and are using the TSC for the delay loop, | |
1497 | * we can skip clock calibration if another cpu in the same socket has already | |
1498 | * been calibrated. This assumes that CONSTANT_TSC applies to all | |
1499 | * cpus in the socket - this should be a safe assumption. | |
1500 | */ | |
148f9bb8 | 1501 | unsigned long calibrate_delay_is_known(void) |
b565201c | 1502 | { |
c25323c0 | 1503 | int sibling, cpu = smp_processor_id(); |
76ce7cfe PT |
1504 | int constant_tsc = cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC); |
1505 | const struct cpumask *mask = topology_core_cpumask(cpu); | |
b565201c | 1506 | |
fe9af81e | 1507 | if (!constant_tsc || !mask) |
f508a5ba TG |
1508 | return 0; |
1509 | ||
1510 | sibling = cpumask_any_but(mask, cpu); | |
c25323c0 TG |
1511 | if (sibling < nr_cpu_ids) |
1512 | return cpu_data(sibling).loops_per_jiffy; | |
b565201c JS |
1513 | return 0; |
1514 | } | |
1515 | #endif |