Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2012 Regents of the University of California
4 */
5
6#include <linux/delay.h>
7#include <linux/math.h>
8#include <linux/param.h>
9#include <linux/timex.h>
10#include <linux/types.h>
11#include <linux/export.h>
12
13#include <asm/processor.h>
14
15/*
16 * This is copies from arch/arm/include/asm/delay.h
17 *
18 * Loop (or tick) based delay:
19 *
20 * loops = loops_per_jiffy * jiffies_per_sec * delay_us / us_per_sec
21 *
22 * where:
23 *
24 * jiffies_per_sec = HZ
25 * us_per_sec = 1000000
26 *
27 * Therefore the constant part is HZ / 1000000 which is a small
28 * fractional number. To make this usable with integer math, we
29 * scale up this constant by 2^31, perform the actual multiplication,
30 * and scale the result back down by 2^31 with a simple shift:
31 *
32 * loops = (loops_per_jiffy * delay_us * UDELAY_MULT) >> 31
33 *
34 * where:
35 *
36 * UDELAY_MULT = 2^31 * HZ / 1000000
37 * = (2^31 / 1000000) * HZ
38 * = 2147.483648 * HZ
39 * = 2147 * HZ + 483648 * HZ / 1000000
40 *
41 * 31 is the biggest scale shift value that won't overflow 32 bits for
42 * delay_us * UDELAY_MULT assuming HZ <= 1000 and delay_us <= 2000.
43 */
44#define MAX_UDELAY_US 2000
45#define MAX_UDELAY_HZ 1000
46#define UDELAY_MULT (2147UL * HZ + 483648UL * HZ / 1000000UL)
47#define UDELAY_SHIFT 31
48
49#if HZ > MAX_UDELAY_HZ
50#error "HZ > MAX_UDELAY_HZ"
51#endif
52
53/*
54 * RISC-V supports both UDELAY and NDELAY. This is largely the same as above,
55 * but with different constants. I added 10 bits to the shift to get this, but
56 * the result is that I need a 64-bit multiply, which is slow on 32-bit
57 * platforms.
58 *
59 * NDELAY_MULT = 2^41 * HZ / 1000000000
60 * = (2^41 / 1000000000) * HZ
61 * = 2199.02325555 * HZ
62 * = 2199 * HZ + 23255550 * HZ / 1000000000
63 *
64 * The maximum here is to avoid 64-bit overflow, but it isn't checked as it
65 * won't happen.
66 */
67#define MAX_NDELAY_NS (1ULL << 42)
68#define MAX_NDELAY_HZ MAX_UDELAY_HZ
69#define NDELAY_MULT ((unsigned long long)(2199ULL * HZ + 23255550ULL * HZ / 1000000000ULL))
70#define NDELAY_SHIFT 41
71
72#if HZ > MAX_NDELAY_HZ
73#error "HZ > MAX_NDELAY_HZ"
74#endif
75
76void __delay(unsigned long cycles)
77{
78 u64 t0 = get_cycles();
79
80 while ((unsigned long)(get_cycles() - t0) < cycles)
81 cpu_relax();
82}
83EXPORT_SYMBOL(__delay);
84
85void udelay(unsigned long usecs)
86{
87 u64 ucycles = (u64)usecs * lpj_fine * UDELAY_MULT;
88 u64 n;
89
90 if (unlikely(usecs > MAX_UDELAY_US)) {
91 n = (u64)usecs * riscv_timebase;
92 do_div(n, 1000000);
93
94 __delay(n);
95 return;
96 }
97
98 __delay(ucycles >> UDELAY_SHIFT);
99}
100EXPORT_SYMBOL(udelay);
101
102void ndelay(unsigned long nsecs)
103{
104 /*
105 * This doesn't bother checking for overflow, as it won't happen (it's
106 * an hour) of delay.
107 */
108 unsigned long long ncycles = nsecs * lpj_fine * NDELAY_MULT;
109 __delay(ncycles >> NDELAY_SHIFT);
110}
111EXPORT_SYMBOL(ndelay);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2012 Regents of the University of California
4 */
5
6#include <linux/delay.h>
7#include <linux/math.h>
8#include <linux/param.h>
9#include <linux/timex.h>
10#include <linux/types.h>
11#include <linux/export.h>
12
13#include <asm/processor.h>
14
15/*
16 * This is copies from arch/arm/include/asm/delay.h
17 *
18 * Loop (or tick) based delay:
19 *
20 * loops = loops_per_jiffy * jiffies_per_sec * delay_us / us_per_sec
21 *
22 * where:
23 *
24 * jiffies_per_sec = HZ
25 * us_per_sec = 1000000
26 *
27 * Therefore the constant part is HZ / 1000000 which is a small
28 * fractional number. To make this usable with integer math, we
29 * scale up this constant by 2^31, perform the actual multiplication,
30 * and scale the result back down by 2^31 with a simple shift:
31 *
32 * loops = (loops_per_jiffy * delay_us * UDELAY_MULT) >> 31
33 *
34 * where:
35 *
36 * UDELAY_MULT = 2^31 * HZ / 1000000
37 * = (2^31 / 1000000) * HZ
38 * = 2147.483648 * HZ
39 * = 2147 * HZ + 483648 * HZ / 1000000
40 *
41 * 31 is the biggest scale shift value that won't overflow 32 bits for
42 * delay_us * UDELAY_MULT assuming HZ <= 1000 and delay_us <= 2000.
43 */
44#define MAX_UDELAY_US 2000
45#define MAX_UDELAY_HZ 1000
46#define UDELAY_MULT (2147UL * HZ + 483648UL * HZ / 1000000UL)
47#define UDELAY_SHIFT 31
48
49#if HZ > MAX_UDELAY_HZ
50#error "HZ > MAX_UDELAY_HZ"
51#endif
52
53/*
54 * RISC-V supports both UDELAY and NDELAY. This is largely the same as above,
55 * but with different constants. I added 10 bits to the shift to get this, but
56 * the result is that I need a 64-bit multiply, which is slow on 32-bit
57 * platforms.
58 *
59 * NDELAY_MULT = 2^41 * HZ / 1000000000
60 * = (2^41 / 1000000000) * HZ
61 * = 2199.02325555 * HZ
62 * = 2199 * HZ + 23255550 * HZ / 1000000000
63 *
64 * The maximum here is to avoid 64-bit overflow, but it isn't checked as it
65 * won't happen.
66 */
67#define MAX_NDELAY_NS (1ULL << 42)
68#define MAX_NDELAY_HZ MAX_UDELAY_HZ
69#define NDELAY_MULT ((unsigned long long)(2199ULL * HZ + 23255550ULL * HZ / 1000000000ULL))
70#define NDELAY_SHIFT 41
71
72#if HZ > MAX_NDELAY_HZ
73#error "HZ > MAX_NDELAY_HZ"
74#endif
75
76void __delay(unsigned long cycles)
77{
78 u64 t0 = get_cycles();
79
80 while ((unsigned long)(get_cycles() - t0) < cycles)
81 cpu_relax();
82}
83EXPORT_SYMBOL(__delay);
84
85void udelay(unsigned long usecs)
86{
87 u64 ucycles = (u64)usecs * lpj_fine * UDELAY_MULT;
88 u64 n;
89
90 if (unlikely(usecs > MAX_UDELAY_US)) {
91 n = (u64)usecs * riscv_timebase;
92 do_div(n, 1000000);
93
94 __delay(n);
95 return;
96 }
97
98 __delay(ucycles >> UDELAY_SHIFT);
99}
100EXPORT_SYMBOL(udelay);
101
102void ndelay(unsigned long nsecs)
103{
104 /*
105 * This doesn't bother checking for overflow, as it won't happen (it's
106 * an hour) of delay.
107 */
108 unsigned long long ncycles = nsecs * lpj_fine * NDELAY_MULT;
109 __delay(ncycles >> NDELAY_SHIFT);
110}
111EXPORT_SYMBOL(ndelay);