2 * This code largely moved from arch/i386/kernel/time.c.
3 * See comments there for proper credits.
6 #include <linux/spinlock.h>
7 #include <linux/init.h>
8 #include <linux/timex.h>
9 #include <linux/errno.h>
10 #include <linux/string.h>
11 #include <linux/jiffies.h>
13 #include <asm/timer.h>
15 #include <asm/processor.h>
18 #include "mach_timer.h"
21 static unsigned long hpet_usec_quotient __read_mostly; /* convert hpet clks to usec */
22 static unsigned long tsc_hpet_quotient __read_mostly; /* convert tsc to hpet clks */
23 static unsigned long hpet_last; /* hpet counter value at last tick*/
24 static unsigned long last_tsc_low; /* lsb 32 bits of Time Stamp Counter */
25 static unsigned long last_tsc_high; /* msb 32 bits of Time Stamp Counter */
26 static unsigned long long monotonic_base;
27 static seqlock_t monotonic_lock = SEQLOCK_UNLOCKED;
29 /* convert from cycles(64bits) => nanoseconds (64bits)
31 * ns = cycles / (freq / ns_per_sec)
32 * ns = cycles * (ns_per_sec / freq)
33 * ns = cycles * (10^9 / (cpu_khz * 10^3))
34 * ns = cycles * (10^6 / cpu_khz)
36 * Then we use scaling math (suggested by george@mvista.com) to get:
37 * ns = cycles * (10^6 * SC / cpu_khz) / SC
38 * ns = cycles * cyc2ns_scale / SC
40 * And since SC is a constant power of two, we can convert the div
43 * We can use khz divisor instead of mhz to keep a better percision, since
44 * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
45 * (mathieu.desnoyers@polymtl.ca)
47 * -johnstul@us.ibm.com "math is hard, lets go shopping!"
49 static unsigned long cyc2ns_scale;
50 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
52 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
54 cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
57 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
59 return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
62 static unsigned long long monotonic_clock_hpet(void)
64 unsigned long long last_offset, this_offset, base;
67 /* atomically read monotonic base & last_offset */
69 seq = read_seqbegin(&monotonic_lock);
70 last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
71 base = monotonic_base;
72 } while (read_seqretry(&monotonic_lock, seq));
74 /* Read the Time Stamp Counter */
77 /* return the value in ns */
78 return base + cycles_2_ns(this_offset - last_offset);
81 static unsigned long get_offset_hpet(void)
83 register unsigned long eax, edx;
85 eax = hpet_readl(HPET_COUNTER);
86 eax -= hpet_last; /* hpet delta */
87 eax = min(hpet_tick, eax);
89 * Time offset = (hpet delta) * ( usecs per HPET clock )
90 * = (hpet delta) * ( usecs per tick / HPET clocks per tick)
91 * = (hpet delta) * ( hpet_usec_quotient ) / (2^32)
94 * hpet_usec_quotient = (2^32 * usecs per tick)/HPET clocks per tick
96 * Using a mull instead of a divl saves some cycles in critical path.
98 ASM_MUL64_REG(eax, edx, hpet_usec_quotient, eax);
100 /* our adjusted time offset in microseconds */
104 static void mark_offset_hpet(void)
106 unsigned long long this_offset, last_offset;
107 unsigned long offset;
109 write_seqlock(&monotonic_lock);
110 last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
111 rdtsc(last_tsc_low, last_tsc_high);
114 offset = hpet_readl(HPET_T0_CMP) - hpet_tick;
116 offset = hpet_readl(HPET_COUNTER);
117 if (unlikely(((offset - hpet_last) >= (2*hpet_tick)) && (hpet_last != 0))) {
118 int lost_ticks = ((offset - hpet_last) / hpet_tick) - 1;
119 jiffies_64 += lost_ticks;
123 /* update the monotonic base value */
124 this_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
125 monotonic_base += cycles_2_ns(this_offset - last_offset);
126 write_sequnlock(&monotonic_lock);
129 static void delay_hpet(unsigned long loops)
131 unsigned long hpet_start, hpet_end;
134 /* loops is the number of cpu cycles. Convert it to hpet clocks */
135 ASM_MUL64_REG(eax, loops, tsc_hpet_quotient, loops);
137 hpet_start = hpet_readl(HPET_COUNTER);
140 hpet_end = hpet_readl(HPET_COUNTER);
141 } while ((hpet_end - hpet_start) < (loops));
144 static struct timer_opts timer_hpet;
146 static int __init init_hpet(char* override)
148 unsigned long result, remain;
150 /* check clock override */
151 if (override[0] && strncmp(override,"hpet",4))
154 if (!is_hpet_enabled())
157 printk("Using HPET for gettimeofday\n");
159 unsigned long tsc_quotient = calibrate_tsc_hpet(&tsc_hpet_quotient);
161 /* report CPU clock rate in Hz.
162 * The formula is (10^6 * 2^32) / (2^32 * 1 / (clocks/us)) =
163 * clock/second. Our precision is about 100 ppm.
165 { unsigned long eax=0, edx=1000;
166 ASM_DIV64_REG(cpu_khz, edx, tsc_quotient,
168 printk("Detected %u.%03u MHz processor.\n",
169 cpu_khz / 1000, cpu_khz % 1000);
171 set_cyc2ns_scale(cpu_khz);
173 /* set this only when cpu_has_tsc */
174 timer_hpet.read_timer = read_timer_tsc;
178 * Math to calculate hpet to usec multiplier
179 * Look for the comments at get_offset_hpet()
181 ASM_DIV64_REG(result, remain, hpet_tick, 0, KERNEL_TICK_USEC);
182 if (remain > (hpet_tick >> 1))
183 result++; /* rounding the result */
184 hpet_usec_quotient = result;
189 static int hpet_resume(void)
191 write_seqlock(&monotonic_lock);
192 /* Assume this is the last mark offset time */
193 rdtsc(last_tsc_low, last_tsc_high);
196 hpet_last = hpet_readl(HPET_T0_CMP) - hpet_tick;
198 hpet_last = hpet_readl(HPET_COUNTER);
199 write_sequnlock(&monotonic_lock);
202 /************************************************************/
204 /* tsc timer_opts struct */
205 static struct timer_opts timer_hpet __read_mostly = {
207 .mark_offset = mark_offset_hpet,
208 .get_offset = get_offset_hpet,
209 .monotonic_clock = monotonic_clock_hpet,
211 .resume = hpet_resume,
214 struct init_timer_opts __initdata timer_hpet_init = {