[PATCH][2.6][3/5]Support for HPET based timer

[Pallipadi, Venkatesh]

[Resend - The original mail hasn't yet appeared on lkml, even 5 hours 
after posting]

3/5 - hpet3.patch - All changes required to support timer services
                    (gettimeofday) with HPET.

diff -purN linux-2.6.0-test1/arch/i386/kernel/timers/timer.c linux-2.6.0-test1-hpet/arch/i386/kernel/timers/timer.c
--- linux-2.6.0-test1/arch/i386/kernel/timers/timer.c	2003-07-13 20:38:41.000000000 -0700
+++ linux-2.6.0-test1-hpet/arch/i386/kernel/timers/timer.c	2003-08-18 14:09:26.000000000 -0700
@@ -3,11 +3,22 @@
 #include <linux/string.h>
 #include <asm/timer.h>
 
+#ifdef CONFIG_HPET_TIMER
+/* 
+ * HPET memory read is slower than tsc reads, but is more dependable as it 
+ * always runs at constant frequency and reduces complexity due to
+ * cpufreq. So, we prefer HPET timer to tsc based one. Also, we cannot use
+ * timer_pit when HPET is active. So, we default to timer_tsc.
+ */
+#endif
 /* list of timers, ordered by preference, NULL terminated */
 static struct timer_opts* timers[] = {
 #ifdef CONFIG_X86_CYCLONE_TIMER
 	&timer_cyclone,
 #endif
+#ifdef CONFIG_HPET_TIMER
+	&timer_hpet,
+#endif
 	&timer_tsc,
 	&timer_pit,
 	NULL,
diff -purN linux-2.6.0-test1/arch/i386/kernel/timers/timer_hpet.c linux-2.6.0-test1-hpet/arch/i386/kernel/timers/timer_hpet.c
--- linux-2.6.0-test1/arch/i386/kernel/timers/timer_hpet.c	1969-12-31 16:00:00.000000000 -0800
+++ linux-2.6.0-test1-hpet/arch/i386/kernel/timers/timer_hpet.c	2003-08-18 17:20:36.000000000 -0700
@@ -0,0 +1,243 @@
+/*
+ * This code largely moved from arch/i386/kernel/time.c.
+ * See comments there for proper credits.
+ */
+
+#include <linux/spinlock.h>
+#include <linux/init.h>
+#include <linux/timex.h>
+#include <linux/errno.h>
+#include <linux/string.h>
+
+#include <asm/timer.h>
+#include <asm/io.h>
+#include <asm/processor.h>
+
+#include "io_ports.h"
+#include "mach_timer.h"
+#include <asm/hpet.h>
+
+extern volatile unsigned long jiffies;
+
+static unsigned long hpet_usec_quotient;	/* convert hpet clks to usec */
+static unsigned long tsc_hpet_quotient;		/* convert tsc to hpet clks */
+static unsigned long hpet_last; 	/* hpet counter value at last tick*/
+static unsigned long last_tsc_low;	/* lsb 32 bits of Time Stamp Counter */
+static unsigned long last_tsc_high; 	/* msb 32 bits of Time Stamp Counter */
+static unsigned long long monotonic_base;
+static rwlock_t monotonic_lock = RW_LOCK_UNLOCKED;
+
+/* convert from cycles(64bits) => nanoseconds (64bits)
+ *  basic equation:
+ *		ns = cycles / (freq / ns_per_sec)
+ *		ns = cycles * (ns_per_sec / freq)
+ *		ns = cycles * (10^9 / (cpu_mhz * 10^6))
+ *		ns = cycles * (10^3 / cpu_mhz)
+ *
+ *	Then we use scaling math (suggested by george@xxxxxxxxxx) to get:
+ *		ns = cycles * (10^3 * SC / cpu_mhz) / SC
+ *		ns = cycles * cyc2ns_scale / SC
+ *
+ *	And since SC is a constant power of two, we can convert the div
+ *  into a shift.   
+ *			-johnstul@xxxxxxxxxx "math is hard, lets go shopping!"
+ */
+static unsigned long cyc2ns_scale; 
+#define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
+
+static inline void set_cyc2ns_scale(unsigned long cpu_mhz)
+{
+	cyc2ns_scale = (1000 << CYC2NS_SCALE_FACTOR)/cpu_mhz;
+}
+
+static inline unsigned long long cycles_2_ns(unsigned long long cyc)
+{
+	return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
+}
+
+static unsigned long long monotonic_clock_hpet(void)
+{
+	unsigned long long last_offset, this_offset, base;
+	
+	/* atomically read monotonic base & last_offset */
+	read_lock_irq(&monotonic_lock);
+	last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
+	base = monotonic_base;
+	read_unlock_irq(&monotonic_lock);
+
+	/* Read the Time Stamp Counter */
+	rdtscll(this_offset);
+
+	/* return the value in ns */
+	return base + cycles_2_ns(this_offset - last_offset);
+}
+
+static unsigned long get_offset_hpet(void)
+{
+	register unsigned long eax, edx;
+
+	eax = hpet_readl(HPET_COUNTER);
+	eax -= hpet_last;	/* hpet delta */
+
+	/*
+         * Time offset = (hpet delta) * ( usecs per HPET clock )
+	 *             = (hpet delta) * ( usecs per tick / HPET clocks per tick)
+	 *             = (hpet delta) * ( hpet_usec_quotient ) / (2^32)
+	 *
+	 * Where,
+	 * hpet_usec_quotient = (2^32 * usecs per tick)/HPET clocks per tick
+	 * 
+	 * Using a mull instead of a divl saves some cycles in critical path.
+         */
+	ASM_MUL64_REG(eax, edx, hpet_usec_quotient, eax);
+
+	/* our adjusted time offset in microseconds */
+	return edx;
+}
+
+static void mark_offset_hpet(void)
+{
+	unsigned long long this_offset, last_offset;
+	unsigned long offset;
+
+	write_lock(&monotonic_lock);
+	last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
+	rdtsc(last_tsc_low, last_tsc_high);
+
+	offset = hpet_readl(HPET_T0_CMP) - hpet_tick;
+	if (unlikely(((offset - hpet_last) > hpet_tick) && (hpet_last != 0))) {
+		int lost_ticks = (offset - hpet_last) / hpet_tick;
+		jiffies += lost_ticks;
+	}
+	hpet_last = offset;
+
+	/* update the monotonic base value */
+	this_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
+	monotonic_base += cycles_2_ns(this_offset - last_offset);
+	write_unlock(&monotonic_lock);
+}
+
+void delay_hpet(unsigned long loops)
+{
+	unsigned long hpet_start, hpet_end;
+	unsigned long eax;
+
+	/* loops is the number of cpu cycles. Convert it to hpet clocks */
+	ASM_MUL64_REG(eax, loops, tsc_hpet_quotient, loops);
+
+	hpet_start = hpet_readl(HPET_COUNTER);
+	do {
+		rep_nop();
+		hpet_end = hpet_readl(HPET_COUNTER);
+	} while ((hpet_end - hpet_start) < (loops));
+}
+
+/* ------ Calibrate the TSC ------- 
+ * Return 2^32 * (1 / (TSC clocks per usec)) for getting the CPU freq.
+ * Set 2^32 * (1 / (tsc per HPET clk)) for delay_hpet().
+ * calibrate_tsc() calibrates the processor TSC by comparing
+ * it to the HPET timer of known frequency.
+ * Too much 64-bit arithmetic here to do this cleanly in C
+ */
+#define CALIBRATE_CNT_HPET 	(5 * hpet_tick)
+#define CALIBRATE_TIME_HPET 	(5 * KERNEL_TICK_USEC)
+
+static unsigned long __init calibrate_tsc(void)
+{
+	unsigned long tsc_startlow, tsc_starthigh;
+	unsigned long tsc_endlow, tsc_endhigh;
+	unsigned long hpet_start, hpet_end;
+	unsigned long result, remain;
+
+	hpet_start = hpet_readl(HPET_COUNTER);
+	rdtsc(tsc_startlow, tsc_starthigh);
+	do {
+		hpet_end = hpet_readl(HPET_COUNTER);
+	} while ((hpet_end - hpet_start) < CALIBRATE_CNT_HPET);
+	rdtsc(tsc_endlow, tsc_endhigh);
+
+	/* 64-bit subtract - gcc just messes up with long longs */
+	__asm__("subl %2,%0\n\t"
+		"sbbl %3,%1"
+		:"=a" (tsc_endlow), "=d" (tsc_endhigh)
+		:"g" (tsc_startlow), "g" (tsc_starthigh),
+		 "0" (tsc_endlow), "1" (tsc_endhigh));
+
+	/* Error: ECPUTOOFAST */
+	if (tsc_endhigh)
+		goto bad_calibration;
+
+	/* Error: ECPUTOOSLOW */
+	if (tsc_endlow <= CALIBRATE_TIME_HPET)
+		goto bad_calibration;
+
+	ASM_DIV64_REG(result, remain, tsc_endlow, 0, CALIBRATE_TIME_HPET);
+	if (remain > (tsc_endlow >> 1))
+		result++; /* rounding the result */
+
+	ASM_DIV64_REG(tsc_hpet_quotient, remain, tsc_endlow, 0, 
+			CALIBRATE_CNT_HPET);
+	if (remain > (tsc_endlow >> 1))
+		tsc_hpet_quotient++; /* rounding the result */
+
+	return result;
+bad_calibration:
+	/*
+	 * the CPU was so fast/slow that the quotient wouldn't fit in
+	 * 32 bits..
+	 */
+	return 0;
+}
+
+static int __init init_hpet(char* override)
+{
+
+	/* check clock override */
+	if (override[0] && strncmp(override,"hpet",4))
+		return -ENODEV;
+
+	if (!is_hpet_enabled())
+		return -ENODEV;
+
+	printk("Using HPET for gettimeofday\n");
+	if (cpu_has_tsc) {
+		unsigned long tsc_quotient = calibrate_tsc();
+		if (tsc_quotient) {
+			/* report CPU clock rate in Hz.
+			 * The formula is (10^6 * 2^32) / (2^32 * 1 / (clocks/us)) =
+			 * clock/second. Our precision is about 100 ppm.
+			 */
+			{	unsigned long eax=0, edx=1000;
+				ASM_DIV64_REG(cpu_khz, edx, tsc_quotient, 
+						eax, edx);
+				printk("Detected %lu.%03lu MHz processor.\n", cpu_khz / 1000, cpu_khz % 1000);
+			}
+			set_cyc2ns_scale(cpu_khz/1000);
+		}
+	}
+	{
+		/*
+		 * Math to calculate hpet to usec multiplier 
+		 * Look for the comments at get_offset_hpet()
+		 */
+		unsigned long result, remain;
+		ASM_DIV64_REG(result, remain, hpet_tick, 0, KERNEL_TICK_USEC);
+		if (remain > (hpet_tick >> 1))
+			result++; /* rounding the result */
+
+		hpet_usec_quotient = result;
+	}
+	return 0;
+}
+
+/************************************************************/
+
+/* tsc timer_opts struct */
+struct timer_opts timer_hpet = {
+	.init =			init_hpet,
+	.mark_offset =		mark_offset_hpet, 
+	.get_offset =		get_offset_hpet,
+	.monotonic_clock =	monotonic_clock_hpet,
+	.delay = 		delay_hpet,
+};
+
diff -purN linux-2.6.0-test1/arch/i386/kernel/timers/timer_tsc.c linux-2.6.0-test1-hpet/arch/i386/kernel/timers/timer_tsc.c
--- linux-2.6.0-test1/arch/i386/kernel/timers/timer_tsc.c	2003-07-13 20:35:57.000000000 -0700
+++ linux-2.6.0-test1-hpet/arch/i386/kernel/timers/timer_tsc.c	2003-08-18 17:25:50.000000000 -0700
@@ -19,9 +19,18 @@
 #include "io_ports.h"
 #include "mach_timer.h"
 
+#include <asm/hpet.h>
+
+#ifdef CONFIG_HPET_TIMER
+static unsigned long hpet_usec_quotient;
+static unsigned long hpet_last;
+struct timer_opts timer_tsc;
+#endif
+
 int tsc_disable __initdata = 0;
 
 extern spinlock_t i8253_lock;
+extern volatile unsigned long jiffies;
 
 static int use_tsc;
 /* Number of usecs that the last interrupt was delayed */
@@ -232,7 +241,7 @@ static void delay_tsc(unsigned long loop
 
 #define CALIBRATE_TIME	(5 * 1000020/HZ)
 
-unsigned long __init calibrate_tsc(void)
+static unsigned long __init calibrate_tsc(void)
 {
 	mach_prepare_counter();
 
@@ -282,6 +291,107 @@ bad_ctc:
 	return 0;
 }
 
+#ifdef CONFIG_HPET_TIMER
+static void mark_offset_tsc_hpet(void)
+{
+	unsigned long long this_offset, last_offset;
+ 	unsigned long offset, temp, hpet_current;
+
+	write_lock(&monotonic_lock);
+	last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
+	/*
+	 * It is important that these two operations happen almost at
+	 * the same time. We do the RDTSC stuff first, since it's
+	 * faster. To avoid any inconsistencies, we need interrupts
+	 * disabled locally.
+	 */
+	/*
+	 * Interrupts are just disabled locally since the timer irq
+	 * has the SA_INTERRUPT flag set. -arca
+	 */
+	/* read Pentium cycle counter */
+
+	hpet_current = hpet_readl(HPET_COUNTER);
+	rdtsc(last_tsc_low, last_tsc_high);
+
+	/* lost tick compensation */
+	offset = hpet_readl(HPET_T0_CMP) - hpet_tick;
+	if (unlikely(((offset - hpet_last) > hpet_tick) && (hpet_last != 0))) {
+		int lost_ticks = (offset - hpet_last) / hpet_tick;
+		jiffies += lost_ticks;
+	}
+	hpet_last = hpet_current;
+
+	/* update the monotonic base value */
+	this_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
+	monotonic_base += cycles_2_ns(this_offset - last_offset);
+	write_unlock(&monotonic_lock);
+
+	/* calculate delay_at_last_interrupt */
+	/*
+	 * Time offset = (hpet delta) * ( usecs per HPET clock )
+	 *             = (hpet delta) * ( usecs per tick / HPET clocks per tick)
+	 *             = (hpet delta) * ( hpet_usec_quotient ) / (2^32)
+	 * Where,
+	 * hpet_usec_quotient = (2^32 * usecs per tick)/HPET clocks per tick
+	 */
+	delay_at_last_interrupt = hpet_current - offset;
+	ASM_MUL64_REG(temp, delay_at_last_interrupt, 
+			hpet_usec_quotient, delay_at_last_interrupt);
+}
+
+/* ------ Calibrate the TSC based on HPET timer ------- 
+ * Return 2^32 * (1 / (TSC clocks per usec)) for do_fast_gettimeoffset().
+ * calibrate_tsc() calibrates the processor TSC by comparing
+ * it to the HPET timer of known frequency.
+ * Too much 64-bit arithmetic here to do this cleanly in C
+ */
+
+#define CALIBRATE_CNT_HPET 	(5 * hpet_tick)
+#define CALIBRATE_TIME_HPET 	(5 * KERNEL_TICK_USEC)
+
+unsigned long __init calibrate_tsc_hpet(void)
+{
+	unsigned long tsc_startlow, tsc_starthigh;
+	unsigned long tsc_endlow, tsc_endhigh;
+	unsigned long hpet_start, hpet_end;
+	unsigned long result, remain;
+
+	hpet_start = hpet_readl(HPET_COUNTER);
+	rdtsc(tsc_startlow, tsc_starthigh);
+	do {
+		hpet_end = hpet_readl(HPET_COUNTER);
+	} while ((hpet_end - hpet_start) < CALIBRATE_CNT_HPET);
+	rdtsc(tsc_endlow, tsc_endhigh);
+
+	/* 64-bit subtract - gcc just messes up with long longs */
+	__asm__("subl %2,%0\n\t"
+		"sbbl %3,%1"
+		:"=a" (tsc_endlow), "=d" (tsc_endhigh)
+		:"g" (tsc_startlow), "g" (tsc_starthigh),
+		 "0" (tsc_endlow), "1" (tsc_endhigh));
+
+	/* Error: ECPUTOOFAST */
+	if (tsc_endhigh)
+		goto bad_calibration;
+
+	/* Error: ECPUTOOSLOW */
+	if (tsc_endlow <= CALIBRATE_TIME_HPET)
+		goto bad_calibration;
+
+	ASM_DIV64_REG(result, remain, tsc_endlow, 0, CALIBRATE_TIME_HPET);
+	if (remain > (tsc_endlow >> 1))
+		result++; /* rounding the result */
+
+	return result;
+bad_calibration:
+	/*
+	 * the CPU was so fast/slow that the quotient wouldn't fit in
+	 * 32 bits..
+	 */
+	return 0;
+}
+#endif
 
 #ifdef CONFIG_CPU_FREQ
 static unsigned int  ref_freq = 0;
@@ -333,8 +443,16 @@ static int __init init_tsc(char* overrid
 {
 
 	/* check clock override */
-	if (override[0] && strncmp(override,"tsc",3))
+	if (override[0] && strncmp(override,"tsc",3)) {
+#ifdef CONFIG_HPET_TIMER
+		if (is_hpet_enabled()) {
+			printk(KERN_ERR "Warning: clock= override failed. Defaulting to tsc\n");
+		} else
+#endif
+		{
 			return -ENODEV;
+		}
+	}
 
 	/*
 	 * If we have APM enabled or the CPU clock speed is variable
@@ -368,7 +486,29 @@ static int __init init_tsc(char* overrid
 	count2 = LATCH; /* initialize counter for mark_offset_tsc() */
 
 	if (cpu_has_tsc) {
-		unsigned long tsc_quotient = calibrate_tsc();
+		unsigned long tsc_quotient;
+#ifdef CONFIG_HPET_TIMER
+		if (is_hpet_enabled()){
+			unsigned long result, remain;
+			printk("Using TSC for gettimeofday\n");
+			tsc_quotient = calibrate_tsc_hpet();
+			timer_tsc.mark_offset = &mark_offset_tsc_hpet;
+			/*
+			 * Math to calculate hpet to usec multiplier 
+			 * Look for the comments at get_offset_tsc_hpet()
+			 */
+			ASM_DIV64_REG(result, remain, hpet_tick, 
+					0, KERNEL_TICK_USEC);
+			if (remain > (hpet_tick >> 1))
+				result++; /* rounding the result */
+
+			hpet_usec_quotient = result;
+		} else
+#endif
+		{
+			tsc_quotient = calibrate_tsc();
+		}
+
 		if (tsc_quotient) {
 			fast_gettimeoffset_quotient = tsc_quotient;
 			use_tsc = 1;
diff -purN linux-2.6.0-test1/include/asm-i386/timer.h linux-2.6.0-test1-hpet/include/asm-i386/timer.h
--- linux-2.6.0-test1/include/asm-i386/timer.h	2003-07-13 20:32:28.000000000 -0700
+++ linux-2.6.0-test1-hpet/include/asm-i386/timer.h	2003-08-18 13:52:52.000000000 -0700
@@ -38,4 +38,8 @@ extern struct timer_opts timer_tsc;
 extern struct timer_opts timer_cyclone;
 #endif
 
+#ifdef CONFIG_HPET_TIMER
+extern struct timer_opts timer_hpet;
+#endif
+
 #endif

Attachment: hpet3.ZIP
Description: hpet3.ZIP