xref: /linux/arch/x86/lib/delay.c (revision f7af616c632ee2ac3af0876fe33bf9e0232e665a)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *	Precise Delay Loops for i386
4  *
5  *	Copyright (C) 1993 Linus Torvalds
6  *	Copyright (C) 1997 Martin Mares <mj@atrey.karlin.mff.cuni.cz>
7  *	Copyright (C) 2008 Jiri Hladky <hladky _dot_ jiri _at_ gmail _dot_ com>
8  *
9  *	The __delay function must _NOT_ be inlined as its execution time
10  *	depends wildly on alignment on many x86 processors. The additional
11  *	jump magic is needed to get the timing stable on all the CPU's
12  *	we have to worry about.
13  */
14 
15 #include <linux/export.h>
16 #include <linux/sched.h>
17 #include <linux/timex.h>
18 #include <linux/preempt.h>
19 #include <linux/delay.h>
20 
21 #include <asm/processor.h>
22 #include <asm/delay.h>
23 #include <asm/timer.h>
24 #include <asm/mwait.h>
25 
26 #ifdef CONFIG_SMP
27 # include <asm/smp.h>
28 #endif
29 
30 static void delay_loop(u64 __loops);
31 
32 /*
33  * Calibration and selection of the delay mechanism happens only once
34  * during boot.
35  */
36 static void (*delay_fn)(u64) __ro_after_init = delay_loop;
37 static void (*delay_halt_fn)(u64 start, u64 cycles) __ro_after_init;
38 
39 /* simple loop based delay: */
40 static void delay_loop(u64 __loops)
41 {
42 	unsigned long loops = (unsigned long)__loops;
43 
44 	asm volatile(
45 		"	test %0,%0	\n"
46 		"	jz 3f		\n"
47 		"	jmp 1f		\n"
48 
49 		".align 16		\n"
50 		"1:	jmp 2f		\n"
51 
52 		".align 16		\n"
53 		"2:	dec %0		\n"
54 		"	jnz 2b		\n"
55 		"3:	dec %0		\n"
56 
57 		: /* we don't need output */
58 		:"a" (loops)
59 	);
60 }
61 
62 /* TSC based delay: */
63 static void delay_tsc(u64 cycles)
64 {
65 	u64 bclock, now;
66 	int cpu;
67 
68 	preempt_disable();
69 	cpu = smp_processor_id();
70 	bclock = rdtsc_ordered();
71 	for (;;) {
72 		now = rdtsc_ordered();
73 		if ((now - bclock) >= cycles)
74 			break;
75 
76 		/* Allow RT tasks to run */
77 		preempt_enable();
78 		rep_nop();
79 		preempt_disable();
80 
81 		/*
82 		 * It is possible that we moved to another CPU, and
83 		 * since TSC's are per-cpu we need to calculate
84 		 * that. The delay must guarantee that we wait "at
85 		 * least" the amount of time. Being moved to another
86 		 * CPU could make the wait longer but we just need to
87 		 * make sure we waited long enough. Rebalance the
88 		 * counter for this CPU.
89 		 */
90 		if (unlikely(cpu != smp_processor_id())) {
91 			cycles -= (now - bclock);
92 			cpu = smp_processor_id();
93 			bclock = rdtsc_ordered();
94 		}
95 	}
96 	preempt_enable();
97 }
98 
99 /*
100  * On Intel the TPAUSE instruction waits until any of:
101  * 1) the TSC counter exceeds the value provided in EDX:EAX
102  * 2) global timeout in IA32_UMWAIT_CONTROL is exceeded
103  * 3) an external interrupt occurs
104  */
105 static void delay_halt_tpause(u64 start, u64 cycles)
106 {
107 	u64 until = start + cycles;
108 	u32 eax, edx;
109 
110 	eax = lower_32_bits(until);
111 	edx = upper_32_bits(until);
112 
113 	/*
114 	 * Hard code the deeper (C0.2) sleep state because exit latency is
115 	 * small compared to the "microseconds" that usleep() will delay.
116 	 */
117 	__tpause(TPAUSE_C02_STATE, edx, eax);
118 }
119 
120 /*
121  * On some AMD platforms, MWAITX has a configurable 32-bit timer, that
122  * counts with TSC frequency. The input value is the number of TSC cycles
123  * to wait. MWAITX will also exit when the timer expires.
124  */
125 static void delay_halt_mwaitx(u64 unused, u64 cycles)
126 {
127 	u64 delay;
128 
129 	delay = min_t(u64, MWAITX_MAX_WAIT_CYCLES, cycles);
130 	/*
131 	 * Use cpu_tss_rw as a cacheline-aligned, seldomly accessed per-cpu
132 	 * variable as the monitor target.
133 	 */
134 	 __monitorx(raw_cpu_ptr(&cpu_tss_rw), 0, 0);
135 
136 	/*
137 	 * AMD, like Intel, supports the EAX hint and EAX=0xf means, do not
138 	 * enter any deep C-state and we use it here in delay() to minimize
139 	 * wakeup latency.
140 	 */
141 	__mwaitx(MWAITX_DISABLE_CSTATES, delay, MWAITX_ECX_TIMER_ENABLE);
142 }
143 
144 /*
145  * Call a vendor specific function to delay for a given amount of time. Because
146  * these functions may return earlier than requested, check for actual elapsed
147  * time and call again until done.
148  */
149 static void delay_halt(u64 __cycles)
150 {
151 	u64 start, end, cycles = __cycles;
152 
153 	/*
154 	 * Timer value of 0 causes MWAITX to wait indefinitely, unless there
155 	 * is a store on the memory monitored by MONITORX.
156 	 */
157 	if (!cycles)
158 		return;
159 
160 	start = rdtsc_ordered();
161 
162 	for (;;) {
163 		delay_halt_fn(start, cycles);
164 		end = rdtsc_ordered();
165 
166 		if (cycles <= end - start)
167 			break;
168 
169 		cycles -= end - start;
170 		start = end;
171 	}
172 }
173 
174 void __init use_tsc_delay(void)
175 {
176 	if (delay_fn == delay_loop)
177 		delay_fn = delay_tsc;
178 }
179 
180 void __init use_tpause_delay(void)
181 {
182 	delay_halt_fn = delay_halt_tpause;
183 	delay_fn = delay_halt;
184 }
185 
186 void use_mwaitx_delay(void)
187 {
188 	delay_halt_fn = delay_halt_mwaitx;
189 	delay_fn = delay_halt;
190 }
191 
192 int read_current_timer(unsigned long *timer_val)
193 {
194 	if (delay_fn == delay_tsc) {
195 		*timer_val = rdtsc();
196 		return 0;
197 	}
198 	return -1;
199 }
200 
201 void __delay(unsigned long loops)
202 {
203 	delay_fn(loops);
204 }
205 EXPORT_SYMBOL(__delay);
206 
207 noinline void __const_udelay(unsigned long xloops)
208 {
209 	unsigned long lpj = this_cpu_read(cpu_info.loops_per_jiffy) ? : loops_per_jiffy;
210 	int d0;
211 
212 	xloops *= 4;
213 	asm("mull %%edx"
214 		:"=d" (xloops), "=&a" (d0)
215 		:"1" (xloops), "0" (lpj * (HZ / 4)));
216 
217 	__delay(++xloops);
218 }
219 EXPORT_SYMBOL(__const_udelay);
220 
221 void __udelay(unsigned long usecs)
222 {
223 	__const_udelay(usecs * 0x000010c7); /* 2**32 / 1000000 (rounded up) */
224 }
225 EXPORT_SYMBOL(__udelay);
226 
227 void __ndelay(unsigned long nsecs)
228 {
229 	__const_udelay(nsecs * 0x00005); /* 2**32 / 1000000000 (rounded up) */
230 }
231 EXPORT_SYMBOL(__ndelay);
232