arch/x86/lib/delay.c

   1 /*
   2  *      Precise Delay Loops for i386
   3  *
   4  *      Copyright (C) 1993 Linus Torvalds
   5  *      Copyright (C) 1997 Martin Mares <[email protected]>
   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
  20 #include <asm/processor.h>
  21 #include <asm/delay.h>
  22 #include <asm/timer.h>
  23
  24 #ifdef CONFIG_SMP
  25 # include <asm/smp.h>
  26 #endif
  27
  28 /* simple loop based delay: */
  29 static void delay_loop(unsigned long loops)
  30 {
  31         asm volatile(
  32                 "       test %0,%0      \n"
  33                 "       jz 3f           \n"
  34                 "       jmp 1f          \n"
  35
  36                 ".align 16              \n"
  37                 "1:     jmp 2f          \n"
  38
  39                 ".align 16              \n"
  40                 "2:     dec %0          \n"
  41                 "       jnz 2b          \n"
  42                 "3:     dec %0          \n"
  43
  44                 : /* we don't need output */
  45                 :"a" (loops)
  46         );
  47 }
  48
  49 /* TSC based delay: */
  50 static void delay_tsc(unsigned long __loops)
  51 {
  52         u32 bclock, now, loops = __loops;
  53         int cpu;
  54
  55         preempt_disable();
  56         cpu = smp_processor_id();
  57         rdtsc_barrier();
  58         rdtscl(bclock);
  59         for (;;) {
  60                 rdtsc_barrier();
  61                 rdtscl(now);
  62                 if ((now - bclock) >= loops)
  63                         break;
  64
  65                 /* Allow RT tasks to run */
  66                 preempt_enable();
  67                 rep_nop();
  68                 preempt_disable();
  69
  70                 /*
  71                  * It is possible that we moved to another CPU, and
  72                  * since TSC's are per-cpu we need to calculate
  73                  * that. The delay must guarantee that we wait "at
  74                  * least" the amount of time. Being moved to another
  75                  * CPU could make the wait longer but we just need to
  76                  * make sure we waited long enough. Rebalance the
  77                  * counter for this CPU.
  78                  */
  79                 if (unlikely(cpu != smp_processor_id())) {
  80                         loops -= (now - bclock);
  81                         cpu = smp_processor_id();
  82                         rdtsc_barrier();
  83                         rdtscl(bclock);
  84                 }
  85         }
  86         preempt_enable();
  87 }
  88
  89 /*
  90  * Since we calibrate only once at boot, this
  91  * function should be set once at boot and not changed
  92  */
  93 static void (*delay_fn)(unsigned long) = delay_loop;
  94
  95 void use_tsc_delay(void)
  96 {
  97         delay_fn = delay_tsc;
  98 }
  99
 100 int read_current_timer(unsigned long *timer_val)
 101 {
 102         if (delay_fn == delay_tsc) {
 103                 rdtscll(*timer_val);
 104                 return 0;
 105         }
 106         return -1;
 107 }
 108
 109 void __delay(unsigned long loops)
 110 {
 111         delay_fn(loops);
 112 }
 113 EXPORT_SYMBOL(__delay);
 114
 115 inline void __const_udelay(unsigned long xloops)
 116 {
 117         int d0;
 118
 119         xloops *= 4;
 120         asm("mull %%edx"
 121                 :"=d" (xloops), "=&a" (d0)
 122                 :"1" (xloops), "0"
 123                 (this_cpu_read(cpu_info.loops_per_jiffy) * (HZ/4)));
 124
 125         __delay(++xloops);
 126 }
 127 EXPORT_SYMBOL(__const_udelay);
 128
 129 void __udelay(unsigned long usecs)
 130 {
 131         __const_udelay(usecs * 0x000010c7); /* 2**32 / 1000000 (rounded up) */
 132 }
 133 EXPORT_SYMBOL(__udelay);
 134
 135 void __ndelay(unsigned long nsecs)
 136 {
 137         __const_udelay(nsecs * 0x00005); /* 2**32 / 1000000000 (rounded up) */
 138 }
 139 EXPORT_SYMBOL(__ndelay);