162306a36Sopenharmony_ci// SPDX-License-Identifier: GPL-2.0-only 262306a36Sopenharmony_ci/* 362306a36Sopenharmony_ci * Copyright (C) 2012 Regents of the University of California 462306a36Sopenharmony_ci */ 562306a36Sopenharmony_ci 662306a36Sopenharmony_ci#include <linux/delay.h> 762306a36Sopenharmony_ci#include <linux/math.h> 862306a36Sopenharmony_ci#include <linux/param.h> 962306a36Sopenharmony_ci#include <linux/timex.h> 1062306a36Sopenharmony_ci#include <linux/types.h> 1162306a36Sopenharmony_ci#include <linux/export.h> 1262306a36Sopenharmony_ci 1362306a36Sopenharmony_ci#include <asm/processor.h> 1462306a36Sopenharmony_ci 1562306a36Sopenharmony_ci/* 1662306a36Sopenharmony_ci * This is copies from arch/arm/include/asm/delay.h 1762306a36Sopenharmony_ci * 1862306a36Sopenharmony_ci * Loop (or tick) based delay: 1962306a36Sopenharmony_ci * 2062306a36Sopenharmony_ci * loops = loops_per_jiffy * jiffies_per_sec * delay_us / us_per_sec 2162306a36Sopenharmony_ci * 2262306a36Sopenharmony_ci * where: 2362306a36Sopenharmony_ci * 2462306a36Sopenharmony_ci * jiffies_per_sec = HZ 2562306a36Sopenharmony_ci * us_per_sec = 1000000 2662306a36Sopenharmony_ci * 2762306a36Sopenharmony_ci * Therefore the constant part is HZ / 1000000 which is a small 2862306a36Sopenharmony_ci * fractional number. To make this usable with integer math, we 2962306a36Sopenharmony_ci * scale up this constant by 2^31, perform the actual multiplication, 3062306a36Sopenharmony_ci * and scale the result back down by 2^31 with a simple shift: 3162306a36Sopenharmony_ci * 3262306a36Sopenharmony_ci * loops = (loops_per_jiffy * delay_us * UDELAY_MULT) >> 31 3362306a36Sopenharmony_ci * 3462306a36Sopenharmony_ci * where: 3562306a36Sopenharmony_ci * 3662306a36Sopenharmony_ci * UDELAY_MULT = 2^31 * HZ / 1000000 3762306a36Sopenharmony_ci * = (2^31 / 1000000) * HZ 3862306a36Sopenharmony_ci * = 2147.483648 * HZ 3962306a36Sopenharmony_ci * = 2147 * HZ + 483648 * HZ / 1000000 4062306a36Sopenharmony_ci * 4162306a36Sopenharmony_ci * 31 is the biggest scale shift value that won't overflow 32 bits for 4262306a36Sopenharmony_ci * delay_us * UDELAY_MULT assuming HZ <= 1000 and delay_us <= 2000. 4362306a36Sopenharmony_ci */ 4462306a36Sopenharmony_ci#define MAX_UDELAY_US 2000 4562306a36Sopenharmony_ci#define MAX_UDELAY_HZ 1000 4662306a36Sopenharmony_ci#define UDELAY_MULT (2147UL * HZ + 483648UL * HZ / 1000000UL) 4762306a36Sopenharmony_ci#define UDELAY_SHIFT 31 4862306a36Sopenharmony_ci 4962306a36Sopenharmony_ci#if HZ > MAX_UDELAY_HZ 5062306a36Sopenharmony_ci#error "HZ > MAX_UDELAY_HZ" 5162306a36Sopenharmony_ci#endif 5262306a36Sopenharmony_ci 5362306a36Sopenharmony_ci/* 5462306a36Sopenharmony_ci * RISC-V supports both UDELAY and NDELAY. This is largely the same as above, 5562306a36Sopenharmony_ci * but with different constants. I added 10 bits to the shift to get this, but 5662306a36Sopenharmony_ci * the result is that I need a 64-bit multiply, which is slow on 32-bit 5762306a36Sopenharmony_ci * platforms. 5862306a36Sopenharmony_ci * 5962306a36Sopenharmony_ci * NDELAY_MULT = 2^41 * HZ / 1000000000 6062306a36Sopenharmony_ci * = (2^41 / 1000000000) * HZ 6162306a36Sopenharmony_ci * = 2199.02325555 * HZ 6262306a36Sopenharmony_ci * = 2199 * HZ + 23255550 * HZ / 1000000000 6362306a36Sopenharmony_ci * 6462306a36Sopenharmony_ci * The maximum here is to avoid 64-bit overflow, but it isn't checked as it 6562306a36Sopenharmony_ci * won't happen. 6662306a36Sopenharmony_ci */ 6762306a36Sopenharmony_ci#define MAX_NDELAY_NS (1ULL << 42) 6862306a36Sopenharmony_ci#define MAX_NDELAY_HZ MAX_UDELAY_HZ 6962306a36Sopenharmony_ci#define NDELAY_MULT ((unsigned long long)(2199ULL * HZ + 23255550ULL * HZ / 1000000000ULL)) 7062306a36Sopenharmony_ci#define NDELAY_SHIFT 41 7162306a36Sopenharmony_ci 7262306a36Sopenharmony_ci#if HZ > MAX_NDELAY_HZ 7362306a36Sopenharmony_ci#error "HZ > MAX_NDELAY_HZ" 7462306a36Sopenharmony_ci#endif 7562306a36Sopenharmony_ci 7662306a36Sopenharmony_civoid __delay(unsigned long cycles) 7762306a36Sopenharmony_ci{ 7862306a36Sopenharmony_ci u64 t0 = get_cycles(); 7962306a36Sopenharmony_ci 8062306a36Sopenharmony_ci while ((unsigned long)(get_cycles() - t0) < cycles) 8162306a36Sopenharmony_ci cpu_relax(); 8262306a36Sopenharmony_ci} 8362306a36Sopenharmony_ciEXPORT_SYMBOL(__delay); 8462306a36Sopenharmony_ci 8562306a36Sopenharmony_civoid udelay(unsigned long usecs) 8662306a36Sopenharmony_ci{ 8762306a36Sopenharmony_ci u64 ucycles = (u64)usecs * lpj_fine * UDELAY_MULT; 8862306a36Sopenharmony_ci u64 n; 8962306a36Sopenharmony_ci 9062306a36Sopenharmony_ci if (unlikely(usecs > MAX_UDELAY_US)) { 9162306a36Sopenharmony_ci n = (u64)usecs * riscv_timebase; 9262306a36Sopenharmony_ci do_div(n, 1000000); 9362306a36Sopenharmony_ci 9462306a36Sopenharmony_ci __delay(n); 9562306a36Sopenharmony_ci return; 9662306a36Sopenharmony_ci } 9762306a36Sopenharmony_ci 9862306a36Sopenharmony_ci __delay(ucycles >> UDELAY_SHIFT); 9962306a36Sopenharmony_ci} 10062306a36Sopenharmony_ciEXPORT_SYMBOL(udelay); 10162306a36Sopenharmony_ci 10262306a36Sopenharmony_civoid ndelay(unsigned long nsecs) 10362306a36Sopenharmony_ci{ 10462306a36Sopenharmony_ci /* 10562306a36Sopenharmony_ci * This doesn't bother checking for overflow, as it won't happen (it's 10662306a36Sopenharmony_ci * an hour) of delay. 10762306a36Sopenharmony_ci */ 10862306a36Sopenharmony_ci unsigned long long ncycles = nsecs * lpj_fine * NDELAY_MULT; 10962306a36Sopenharmony_ci __delay(ncycles >> NDELAY_SHIFT); 11062306a36Sopenharmony_ci} 11162306a36Sopenharmony_ciEXPORT_SYMBOL(ndelay); 112