1 // SPDX-License-Identifier: GPL-2.0+
3 * Copyright (c) 2012 The Chromium OS Authors.
5 * TSC calibration codes are adapted from Linux kernel
6 * arch/x86/kernel/tsc_msr.c and arch/x86/kernel/tsc.c
10 #include <bootstage.h>
17 #include <asm/global_data.h>
19 #include <asm/i8254.h>
20 #include <asm/ibmpc.h>
22 #include <asm/u-boot-x86.h>
23 #include <linux/delay.h>
25 #define MAX_NUM_FREQS 9
27 #define INTEL_FAM6_SKYLAKE_MOBILE 0x4E
28 #define INTEL_FAM6_ATOM_GOLDMONT 0x5C /* Apollo Lake */
29 #define INTEL_FAM6_SKYLAKE_DESKTOP 0x5E
30 #define INTEL_FAM6_ATOM_GOLDMONT_X 0x5F /* Denverton */
31 #define INTEL_FAM6_KABYLAKE_MOBILE 0x8E
32 #define INTEL_FAM6_KABYLAKE_DESKTOP 0x9E
34 DECLARE_GLOBAL_DATA_PTR;
37 * native_calibrate_tsc
38 * Determine TSC frequency via CPUID, else return 0.
40 static unsigned long native_calibrate_tsc(void)
42 struct cpuid_result tsc_info;
43 unsigned int crystal_freq;
45 if (gd->arch.x86_vendor != X86_VENDOR_INTEL)
48 if (cpuid_eax(0) < 0x15)
51 tsc_info = cpuid(0x15);
53 if (tsc_info.ebx == 0 || tsc_info.eax == 0)
56 crystal_freq = tsc_info.ecx / 1000;
57 if (!CONFIG_IS_ENABLED(X86_TSC_TIMER_NATIVE) && !crystal_freq) {
58 switch (gd->arch.x86_model) {
59 case INTEL_FAM6_SKYLAKE_MOBILE:
60 case INTEL_FAM6_SKYLAKE_DESKTOP:
61 case INTEL_FAM6_KABYLAKE_MOBILE:
62 case INTEL_FAM6_KABYLAKE_DESKTOP:
63 crystal_freq = 24000; /* 24.0 MHz */
65 case INTEL_FAM6_ATOM_GOLDMONT_X:
66 crystal_freq = 25000; /* 25.0 MHz */
68 case INTEL_FAM6_ATOM_GOLDMONT:
69 crystal_freq = 19200; /* 19.2 MHz */
76 return (crystal_freq * tsc_info.ebx / tsc_info.eax) / 1000;
79 static unsigned long cpu_mhz_from_cpuid(void)
81 if (gd->arch.x86_vendor != X86_VENDOR_INTEL)
84 if (cpuid_eax(0) < 0x16)
87 return cpuid_eax(0x16);
91 * According to Intel 64 and IA-32 System Programming Guide,
92 * if MSR_PERF_STAT[31] is set, the maximum resolved bus ratio can be
93 * read in MSR_PLATFORM_ID[12:8], otherwise in MSR_PERF_STAT[44:40].
94 * Unfortunately some Intel Atom SoCs aren't quite compliant to this,
95 * so we need manually differentiate SoC families. This is what the
96 * field msr_plat does.
99 u8 x86_family; /* CPU family */
100 u8 x86_model; /* model */
101 /* 2: use 100MHz, 1: use MSR_PLATFORM_INFO, 0: MSR_IA32_PERF_STATUS */
103 u32 freqs[MAX_NUM_FREQS];
106 static struct freq_desc freq_desc_tables[] = {
108 { 6, 0x27, 0, { 0, 0, 0, 0, 0, 99840, 0, 83200, 0 } },
110 { 6, 0x35, 0, { 0, 133200, 0, 0, 0, 99840, 0, 83200, 0 } },
111 /* TNG - Intel Atom processor Z3400 series */
112 { 6, 0x4a, 1, { 0, 100000, 133300, 0, 0, 0, 0, 0, 0 } },
113 /* VLV2 - Intel Atom processor E3000, Z3600, Z3700 series */
114 { 6, 0x37, 1, { 83300, 100000, 133300, 116700, 80000, 0, 0, 0, 0 } },
115 /* ANN - Intel Atom processor Z3500 series */
116 { 6, 0x5a, 1, { 83300, 100000, 133300, 100000, 0, 0, 0, 0, 0 } },
117 /* AMT - Intel Atom processor X7-Z8000 and X5-Z8000 series */
118 { 6, 0x4c, 1, { 83300, 100000, 133300, 116700,
119 80000, 93300, 90000, 88900, 87500 } },
121 { 6, 0x3a, 2, { 0, 0, 0, 0, 0, 0, 0, 0, 0 } },
124 static int match_cpu(u8 family, u8 model)
128 for (i = 0; i < ARRAY_SIZE(freq_desc_tables); i++) {
129 if ((family == freq_desc_tables[i].x86_family) &&
130 (model == freq_desc_tables[i].x86_model))
137 /* Map CPU reference clock freq ID(0-7) to CPU reference clock freq(KHz) */
138 #define id_to_freq(cpu_index, freq_id) \
139 (freq_desc_tables[cpu_index].freqs[freq_id])
142 * TSC on Intel Atom SoCs capable of determining TSC frequency by MSR is
143 * reliable and the frequency is known (provided by HW).
145 * On these platforms PIT/HPET is generally not available so calibration won't
146 * work at all and there is no other clocksource to act as a watchdog for the
147 * TSC, so we have no other choice than to trust it.
149 * Returns the TSC frequency in MHz or 0 if HW does not provide it.
151 static unsigned long __maybe_unused cpu_mhz_from_msr(void)
153 u32 lo, hi, ratio, freq_id, freq;
157 if (gd->arch.x86_vendor != X86_VENDOR_INTEL)
160 cpu_index = match_cpu(gd->arch.x86, gd->arch.x86_model);
164 if (freq_desc_tables[cpu_index].msr_plat) {
165 rdmsr(MSR_PLATFORM_INFO, lo, hi);
166 ratio = (lo >> 8) & 0xff;
168 rdmsr(MSR_IA32_PERF_STATUS, lo, hi);
169 ratio = (hi >> 8) & 0x1f;
171 debug("Maximum core-clock to bus-clock ratio: 0x%x\n", ratio);
173 if (freq_desc_tables[cpu_index].msr_plat == 2) {
174 /* TODO: Figure out how best to deal with this */
176 debug("Using frequency: %u KHz\n", freq);
178 /* Get FSB FREQ ID */
179 rdmsr(MSR_FSB_FREQ, lo, hi);
181 freq = id_to_freq(cpu_index, freq_id);
182 debug("Resolved frequency ID: %u, frequency: %u KHz\n",
186 /* TSC frequency = maximum resolved freq * maximum resolved bus ratio */
187 res = freq * ratio / 1000;
188 debug("TSC runs at %lu MHz\n", res);
194 * This reads the current MSB of the PIT counter, and
195 * checks if we are running on sufficiently fast and
196 * non-virtualized hardware.
198 * Our expectations are:
200 * - the PIT is running at roughly 1.19MHz
202 * - each IO is going to take about 1us on real hardware,
203 * but we allow it to be much faster (by a factor of 10) or
204 * _slightly_ slower (ie we allow up to a 2us read+counter
205 * update - anything else implies a unacceptably slow CPU
206 * or PIT for the fast calibration to work.
208 * - with 256 PIT ticks to read the value, we have 214us to
209 * see the same MSB (and overhead like doing a single TSC
210 * read per MSB value etc).
212 * - We're doing 2 reads per loop (LSB, MSB), and we expect
213 * them each to take about a microsecond on real hardware.
214 * So we expect a count value of around 100. But we'll be
215 * generous, and accept anything over 50.
217 * - if the PIT is stuck, and we see *many* more reads, we
218 * return early (and the next caller of pit_expect_msb()
219 * then consider it a failure when they don't see the
220 * next expected value).
222 * These expectations mean that we know that we have seen the
223 * transition from one expected value to another with a fairly
224 * high accuracy, and we didn't miss any events. We can thus
225 * use the TSC value at the transitions to calculate a pretty
226 * good value for the TSC frequencty.
228 static inline int pit_verify_msb(unsigned char val)
232 return inb(0x42) == val;
235 static inline int pit_expect_msb(unsigned char val, u64 *tscp,
236 unsigned long *deltap)
239 u64 tsc = 0, prev_tsc = 0;
241 for (count = 0; count < 50000; count++) {
242 if (!pit_verify_msb(val))
247 *deltap = rdtsc() - prev_tsc;
251 * We require _some_ success, but the quality control
252 * will be based on the error terms on the TSC values.
258 * How many MSB values do we want to see? We aim for
259 * a maximum error rate of 500ppm (in practice the
260 * real error is much smaller), but refuse to spend
261 * more than 50ms on it.
263 #define MAX_QUICK_PIT_MS 50
264 #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
266 static unsigned long __maybe_unused quick_pit_calibrate(void)
270 unsigned long d1, d2;
272 /* Set the Gate high, disable speaker */
273 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
276 * Counter 2, mode 0 (one-shot), binary count
278 * NOTE! Mode 2 decrements by two (and then the
279 * output is flipped each time, giving the same
280 * final output frequency as a decrement-by-one),
281 * so mode 0 is much better when looking at the
286 /* Start at 0xffff */
291 * The PIT starts counting at the next edge, so we
292 * need to delay for a microsecond. The easiest way
293 * to do that is to just read back the 16-bit counter
298 if (pit_expect_msb(0xff, &tsc, &d1)) {
299 for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
300 if (!pit_expect_msb(0xff-i, &delta, &d2))
304 * Iterate until the error is less than 500 ppm
307 if (d1+d2 >= delta >> 11)
311 * Check the PIT one more time to verify that
312 * all TSC reads were stable wrt the PIT.
314 * This also guarantees serialization of the
315 * last cycle read ('d2') in pit_expect_msb.
317 if (!pit_verify_msb(0xfe - i))
322 debug("Fast TSC calibration failed\n");
327 * Ok, if we get here, then we've seen the
328 * MSB of the PIT decrement 'i' times, and the
329 * error has shrunk to less than 500 ppm.
331 * As a result, we can depend on there not being
332 * any odd delays anywhere, and the TSC reads are
333 * reliable (within the error).
335 * kHz = ticks / time-in-seconds / 1000;
336 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
337 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
339 delta *= PIT_TICK_RATE;
340 delta /= (i*256*1000);
341 debug("Fast TSC calibration using PIT\n");
345 /* Get the speed of the TSC timer in MHz */
346 unsigned notrace long get_tbclk_mhz(void)
348 return get_tbclk() / 1000000;
351 static ulong get_ms_timer(void)
353 return (get_ticks() * 1000) / get_tbclk();
356 ulong get_timer(ulong base)
358 return get_ms_timer() - base;
361 ulong notrace timer_get_us(void)
363 return get_ticks() / get_tbclk_mhz();
366 ulong timer_get_boot_us(void)
368 return timer_get_us();
371 void __udelay(unsigned long usec)
373 u64 now = get_ticks();
376 stop = now + (u64)usec * get_tbclk_mhz();
378 while ((int64_t)(stop - get_ticks()) > 0)
379 #if defined(CONFIG_QEMU) && defined(CONFIG_SMP)
381 * Add a 'pause' instruction on qemu target,
382 * to give other VCPUs a chance to run.
384 asm volatile("pause");
390 static u64 tsc_timer_get_count(struct udevice *dev)
392 u64 now_tick = rdtsc();
394 return now_tick - gd->arch.tsc_base;
397 static void tsc_timer_ensure_setup(bool early)
399 if (gd->arch.tsc_inited)
401 if (IS_ENABLED(CONFIG_X86_TSC_READ_BASE))
402 gd->arch.tsc_base = rdtsc();
404 if (!gd->arch.clock_rate) {
405 unsigned long fast_calibrate;
407 fast_calibrate = native_calibrate_tsc();
411 /* Reduce code size by dropping other methods */
412 if (CONFIG_IS_ENABLED(X86_TSC_TIMER_NATIVE))
415 fast_calibrate = cpu_mhz_from_cpuid();
419 fast_calibrate = cpu_mhz_from_msr();
423 fast_calibrate = quick_pit_calibrate();
428 gd->arch.clock_rate = CONFIG_X86_TSC_TIMER_FREQ;
433 if (!gd->arch.clock_rate)
434 gd->arch.clock_rate = fast_calibrate * 1000000;
436 gd->arch.tsc_inited = true;
439 static int tsc_timer_probe(struct udevice *dev)
441 struct timer_dev_priv *uc_priv = dev_get_uclass_priv(dev);
443 /* Try hardware calibration first */
444 tsc_timer_ensure_setup(false);
445 if (!gd->arch.clock_rate) {
447 * Use the clock frequency specified in the
448 * device tree as last resort
450 if (!uc_priv->clock_rate)
451 panic("TSC frequency is ZERO");
453 uc_priv->clock_rate = gd->arch.clock_rate;
459 unsigned long notrace timer_early_get_rate(void)
462 * When TSC timer is used as the early timer, be warned that the timer
463 * clock rate can only be calibrated via some hardware ways. Specifying
464 * it in the device tree won't work for the early timer.
466 tsc_timer_ensure_setup(true);
468 return gd->arch.clock_rate;
471 u64 notrace timer_early_get_count(void)
473 tsc_timer_ensure_setup(true);
475 return rdtsc() - gd->arch.tsc_base;
478 static const struct timer_ops tsc_timer_ops = {
479 .get_count = tsc_timer_get_count,
482 #if !CONFIG_IS_ENABLED(OF_PLATDATA)
483 static const struct udevice_id tsc_timer_ids[] = {
484 { .compatible = "x86,tsc-timer", },
489 U_BOOT_DRIVER(x86_tsc_timer) = {
490 .name = "x86_tsc_timer",
492 .of_match = of_match_ptr(tsc_timer_ids),
493 .probe = tsc_timer_probe,
494 .ops = &tsc_timer_ops,