arch/x86/lib/delay_32.c

   1 /*
   2  *      Precise Delay Loops for i386
   3  *
   4  *      Copyright (C) 1993 Linus Torvalds
   5  *      Copyright (C) 1997 Martin Mares <mj@atrey.karlin.mff.cuni.cz>
   6  *      Copyright (C) 2008 Jiri Hladky <hladky _dot_ jiri _at_ gmail _dot_ com>
   7  *
   8  *      The __delay function must _NOT_ be inlined as its execution time
   9  *      depends wildly on alignment on many x86 processors. The additional
  10  *      jump magic is needed to get the timing stable on all the CPU's
  11  *      we have to worry about.
  12  */
  13
  14 #include <linux/module.h>
  15 #include <linux/sched.h>
  16 #include <linux/timex.h>
  17 #include <linux/preempt.h>
  18 #include <linux/delay.h>
  19 #include <linux/init.h>
  20
  21 #include <asm/processor.h>
  22 #include <asm/delay.h>
  23 #include <asm/timer.h>
  24
  25 #ifdef CONFIG_SMP
  26 # include <asm/smp.h>
  27 #endif
  28
  29 /* simple loop based delay: */
  30 static void delay_loop(unsigned long loops)
  31 {
  32         __asm__ __volatile__(
  33                 "       test %0,%0      \n"
  34                 "       jz 3f           \n"
  35                 "       jmp 1f          \n"
  36
  37                 ".align 16              \n"
  38                 "1:     jmp 2f          \n"
  39
  40                 ".align 16              \n"
  41                 "2:     dec %0          \n"
  42                 "       jnz 2b          \n"
  43                 "3:     dec %0          \n"
  44
  45                 : /* we don't need output */
  46                 :"a" (loops)
  47         );
  48 }
  49
  50 /* TSC based delay: */
  51 static void delay_tsc(unsigned long loops)
  52 {
  53         unsigned long bclock, now;
  54         int cpu;
  55
  56         preempt_disable();
  57         cpu = smp_processor_id();
  58         rdtscl(bclock);
  59         for (;;) {
  60                 rdtscl(now);
  61                 if ((now - bclock) >= loops)
  62                         break;
  63
  64                 /* Allow RT tasks to run */
  65                 preempt_enable();
  66                 rep_nop();
  67                 preempt_disable();
  68
  69                 /*
  70                  * It is possible that we moved to another CPU, and
  71                  * since TSC's are per-cpu we need to calculate
  72                  * that. The delay must guarantee that we wait "at
  73                  * least" the amount of time. Being moved to another
  74                  * CPU could make the wait longer but we just need to
  75                  * make sure we waited long enough. Rebalance the
  76                  * counter for this CPU.
  77                  */
  78                 if (unlikely(cpu != smp_processor_id())) {
  79                         loops -= (now - bclock);
  80                         cpu = smp_processor_id();
  81                         rdtscl(bclock);
  82                 }
  83         }
  84         preempt_enable();
  85 }
  86
  87 /*
  88  * Since we calibrate only once at boot, this
  89  * function should be set once at boot and not changed
  90  */
  91 static void (*delay_fn)(unsigned long) = delay_loop;
  92
  93 void use_tsc_delay(void)
  94 {
  95         delay_fn = delay_tsc;
  96 }
  97
  98 int __devinit read_current_timer(unsigned long *timer_val)
  99 {
 100         if (delay_fn == delay_tsc) {
 101                 rdtscl(*timer_val);
 102                 return 0;
 103         }
 104         return -1;
 105 }
 106
 107 void __delay(unsigned long loops)
 108 {
 109         delay_fn(loops);
 110 }
 111
 112 inline void __const_udelay(unsigned long xloops)
 113 {
 114         int d0;
 115
 116         xloops *= 4;
 117         __asm__("mull %0"
 118                 :"=d" (xloops), "=&a" (d0)
 119                 :"1" (xloops), "0"
 120                 (cpu_data(raw_smp_processor_id()).loops_per_jiffy * (HZ/4)));
 121
 122         __delay(++xloops);
 123 }
 124
 125 void __udelay(unsigned long usecs)
 126 {
 127         __const_udelay(usecs * 0x000010c7); /* 2**32 / 1000000 (rounded up) */
 128 }
 129
 130 void __ndelay(unsigned long nsecs)
 131 {
 132         __const_udelay(nsecs * 0x00005); /* 2**32 / 1000000000 (rounded up) */
 133 }
 134
 135 EXPORT_SYMBOL(__delay);
 136 EXPORT_SYMBOL(__const_udelay);
 137 EXPORT_SYMBOL(__udelay);
 138 EXPORT_SYMBOL(__ndelay);