2 * Copyright (c) 2012 The Chromium OS Authors.
4 * TSC calibration codes are adapted from Linux kernel
5 * arch/x86/kernel/tsc_msr.c and arch/x86/kernel/tsc.c
7 * SPDX-License-Identifier: GPL-2.0+
15 #include <asm/i8254.h>
16 #include <asm/ibmpc.h>
18 #include <asm/u-boot-x86.h>
20 /* CPU reference clock frequency: in KHz */
22 #define FREQ_100 99840
23 #define FREQ_133 133200
24 #define FREQ_166 166400
26 #define MAX_NUM_FREQS 8
28 DECLARE_GLOBAL_DATA_PTR
;
31 * According to Intel 64 and IA-32 System Programming Guide,
32 * if MSR_PERF_STAT[31] is set, the maximum resolved bus ratio can be
33 * read in MSR_PLATFORM_ID[12:8], otherwise in MSR_PERF_STAT[44:40].
34 * Unfortunately some Intel Atom SoCs aren't quite compliant to this,
35 * so we need manually differentiate SoC families. This is what the
36 * field msr_plat does.
39 u8 x86_family
; /* CPU family */
40 u8 x86_model
; /* model */
41 /* 2: use 100MHz, 1: use MSR_PLATFORM_INFO, 0: MSR_IA32_PERF_STATUS */
43 u32 freqs
[MAX_NUM_FREQS
];
46 static struct freq_desc freq_desc_tables
[] = {
48 { 6, 0x27, 0, { 0, 0, 0, 0, 0, FREQ_100
, 0, FREQ_83
} },
50 { 6, 0x35, 0, { 0, FREQ_133
, 0, 0, 0, FREQ_100
, 0, FREQ_83
} },
52 { 6, 0x4a, 1, { 0, FREQ_100
, FREQ_133
, 0, 0, 0, 0, 0 } },
54 { 6, 0x37, 1, { FREQ_83
, FREQ_100
, FREQ_133
, FREQ_166
, 0, 0, 0, 0 } },
56 { 6, 0x3a, 2, { 0, 0, 0, 0, 0, 0, 0, 0 } },
58 { 6, 0x5a, 1, { FREQ_83
, FREQ_100
, FREQ_133
, FREQ_100
, 0, 0, 0, 0 } },
61 static int match_cpu(u8 family
, u8 model
)
65 for (i
= 0; i
< ARRAY_SIZE(freq_desc_tables
); i
++) {
66 if ((family
== freq_desc_tables
[i
].x86_family
) &&
67 (model
== freq_desc_tables
[i
].x86_model
))
74 /* Map CPU reference clock freq ID(0-7) to CPU reference clock freq(KHz) */
75 #define id_to_freq(cpu_index, freq_id) \
76 (freq_desc_tables[cpu_index].freqs[freq_id])
79 * Do MSR calibration only for known/supported CPUs.
81 * Returns the calibration value or 0 if MSR calibration failed.
83 static unsigned long __maybe_unused
try_msr_calibrate_tsc(void)
85 u32 lo
, hi
, ratio
, freq_id
, freq
;
89 cpu_index
= match_cpu(gd
->arch
.x86
, gd
->arch
.x86_model
);
93 if (freq_desc_tables
[cpu_index
].msr_plat
) {
94 rdmsr(MSR_PLATFORM_INFO
, lo
, hi
);
95 ratio
= (lo
>> 8) & 0x1f;
97 rdmsr(MSR_IA32_PERF_STATUS
, lo
, hi
);
98 ratio
= (hi
>> 8) & 0x1f;
100 debug("Maximum core-clock to bus-clock ratio: 0x%x\n", ratio
);
105 if (freq_desc_tables
[cpu_index
].msr_plat
== 2) {
106 /* TODO: Figure out how best to deal with this */
108 debug("Using frequency: %u KHz\n", freq
);
110 /* Get FSB FREQ ID */
111 rdmsr(MSR_FSB_FREQ
, lo
, hi
);
113 freq
= id_to_freq(cpu_index
, freq_id
);
114 debug("Resolved frequency ID: %u, frequency: %u KHz\n",
120 /* TSC frequency = maximum resolved freq * maximum resolved bus ratio */
121 res
= freq
* ratio
/ 1000;
122 debug("TSC runs at %lu MHz\n", res
);
127 debug("Fast TSC calibration using MSR failed\n");
132 * This reads the current MSB of the PIT counter, and
133 * checks if we are running on sufficiently fast and
134 * non-virtualized hardware.
136 * Our expectations are:
138 * - the PIT is running at roughly 1.19MHz
140 * - each IO is going to take about 1us on real hardware,
141 * but we allow it to be much faster (by a factor of 10) or
142 * _slightly_ slower (ie we allow up to a 2us read+counter
143 * update - anything else implies a unacceptably slow CPU
144 * or PIT for the fast calibration to work.
146 * - with 256 PIT ticks to read the value, we have 214us to
147 * see the same MSB (and overhead like doing a single TSC
148 * read per MSB value etc).
150 * - We're doing 2 reads per loop (LSB, MSB), and we expect
151 * them each to take about a microsecond on real hardware.
152 * So we expect a count value of around 100. But we'll be
153 * generous, and accept anything over 50.
155 * - if the PIT is stuck, and we see *many* more reads, we
156 * return early (and the next caller of pit_expect_msb()
157 * then consider it a failure when they don't see the
158 * next expected value).
160 * These expectations mean that we know that we have seen the
161 * transition from one expected value to another with a fairly
162 * high accuracy, and we didn't miss any events. We can thus
163 * use the TSC value at the transitions to calculate a pretty
164 * good value for the TSC frequencty.
166 static inline int pit_verify_msb(unsigned char val
)
170 return inb(0x42) == val
;
173 static inline int pit_expect_msb(unsigned char val
, u64
*tscp
,
174 unsigned long *deltap
)
177 u64 tsc
= 0, prev_tsc
= 0;
179 for (count
= 0; count
< 50000; count
++) {
180 if (!pit_verify_msb(val
))
185 *deltap
= rdtsc() - prev_tsc
;
189 * We require _some_ success, but the quality control
190 * will be based on the error terms on the TSC values.
196 * How many MSB values do we want to see? We aim for
197 * a maximum error rate of 500ppm (in practice the
198 * real error is much smaller), but refuse to spend
199 * more than 50ms on it.
201 #define MAX_QUICK_PIT_MS 50
202 #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
204 static unsigned long __maybe_unused
quick_pit_calibrate(void)
208 unsigned long d1
, d2
;
210 /* Set the Gate high, disable speaker */
211 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
214 * Counter 2, mode 0 (one-shot), binary count
216 * NOTE! Mode 2 decrements by two (and then the
217 * output is flipped each time, giving the same
218 * final output frequency as a decrement-by-one),
219 * so mode 0 is much better when looking at the
224 /* Start at 0xffff */
229 * The PIT starts counting at the next edge, so we
230 * need to delay for a microsecond. The easiest way
231 * to do that is to just read back the 16-bit counter
236 if (pit_expect_msb(0xff, &tsc
, &d1
)) {
237 for (i
= 1; i
<= MAX_QUICK_PIT_ITERATIONS
; i
++) {
238 if (!pit_expect_msb(0xff-i
, &delta
, &d2
))
242 * Iterate until the error is less than 500 ppm
245 if (d1
+d2
>= delta
>> 11)
249 * Check the PIT one more time to verify that
250 * all TSC reads were stable wrt the PIT.
252 * This also guarantees serialization of the
253 * last cycle read ('d2') in pit_expect_msb.
255 if (!pit_verify_msb(0xfe - i
))
260 debug("Fast TSC calibration failed\n");
265 * Ok, if we get here, then we've seen the
266 * MSB of the PIT decrement 'i' times, and the
267 * error has shrunk to less than 500 ppm.
269 * As a result, we can depend on there not being
270 * any odd delays anywhere, and the TSC reads are
271 * reliable (within the error).
273 * kHz = ticks / time-in-seconds / 1000;
274 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
275 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
277 delta
*= PIT_TICK_RATE
;
278 delta
/= (i
*256*1000);
279 debug("Fast TSC calibration using PIT\n");
284 void timer_set_base(u64 base
)
286 gd
->arch
.tsc_base
= base
;
290 * Get the number of CPU time counter ticks since it was read first time after
291 * restart. This yields a free running counter guaranteed to take almost 6
292 * years to wrap around even at 100GHz clock rate.
294 u64 notrace
get_ticks(void)
296 u64 now_tick
= rdtsc();
298 /* We assume that 0 means the base hasn't been set yet */
299 if (!gd
->arch
.tsc_base
)
300 panic("No tick base available");
301 return now_tick
- gd
->arch
.tsc_base
;
303 #endif /* CONFIG_TIMER */
305 /* Get the speed of the TSC timer in MHz */
306 unsigned notrace
long get_tbclk_mhz(void)
309 return get_tbclk() / 1000000;
311 unsigned long fast_calibrate
;
313 if (gd
->arch
.tsc_mhz
)
314 return gd
->arch
.tsc_mhz
;
316 #ifdef CONFIG_TSC_CALIBRATION_BYPASS
317 fast_calibrate
= CONFIG_TSC_FREQ_IN_MHZ
;
319 fast_calibrate
= try_msr_calibrate_tsc();
320 if (!fast_calibrate
) {
322 fast_calibrate
= quick_pit_calibrate();
324 panic("TSC frequency is ZERO");
328 gd
->arch
.tsc_mhz
= fast_calibrate
;
329 return fast_calibrate
;
334 unsigned long get_tbclk(void)
336 return get_tbclk_mhz() * 1000 * 1000;
340 static ulong
get_ms_timer(void)
342 return (get_ticks() * 1000) / get_tbclk();
345 ulong
get_timer(ulong base
)
347 return get_ms_timer() - base
;
350 ulong notrace
timer_get_us(void)
352 return get_ticks() / get_tbclk_mhz();
355 ulong
timer_get_boot_us(void)
357 return timer_get_us();
360 void __udelay(unsigned long usec
)
362 u64 now
= get_ticks();
365 stop
= now
+ usec
* get_tbclk_mhz();
367 while ((int64_t)(stop
- get_ticks()) > 0)
368 #if defined(CONFIG_QEMU) && defined(CONFIG_SMP)
370 * Add a 'pause' instruction on qemu target,
371 * to give other VCPUs a chance to run.
373 asm volatile("pause");
381 #ifdef CONFIG_I8254_TIMER
382 /* Set up the i8254 timer if required */
390 static int tsc_timer_get_count(struct udevice
*dev
, u64
*count
)
392 u64 now_tick
= rdtsc();
394 *count
= now_tick
- gd
->arch
.tsc_base
;
399 static int tsc_timer_probe(struct udevice
*dev
)
401 struct timer_dev_priv
*uc_priv
= dev_get_uclass_priv(dev
);
403 gd
->arch
.tsc_base
= rdtsc();
406 * If there is no clock frequency specified in the device tree,
407 * calibrate it by ourselves.
409 if (!uc_priv
->clock_rate
) {
410 unsigned long fast_calibrate
;
412 fast_calibrate
= try_msr_calibrate_tsc();
413 if (!fast_calibrate
) {
414 fast_calibrate
= quick_pit_calibrate();
416 panic("TSC frequency is ZERO");
419 uc_priv
->clock_rate
= fast_calibrate
* 1000000;
425 static const struct timer_ops tsc_timer_ops
= {
426 .get_count
= tsc_timer_get_count
,
429 static const struct udevice_id tsc_timer_ids
[] = {
430 { .compatible
= "x86,tsc-timer", },
434 U_BOOT_DRIVER(tsc_timer
) = {
437 .of_match
= tsc_timer_ids
,
438 .probe
= tsc_timer_probe
,
439 .ops
= &tsc_timer_ops
,
440 .flags
= DM_FLAG_PRE_RELOC
,
442 #endif /* CONFIG_TIMER */