From 8ea66915f29c2b8abb1de57c34b0b857fb4af756 Mon Sep 17 00:00:00 2001
From: Ben Gras <ben@minix3.org>
Date: Sun, 3 Feb 2013 19:28:24 +0100
Subject: [PATCH] kernel: scheduling fix for ARM

	. make read_tsc_64 use the free-running clock, significantly
	  improving scheduling behaviour

Change-Id: Idf6a12f6e26be7fe3b3664c278cae846d8b2a442
---
 kernel/arch/earm/arch_clock.c | 23 ++++++++---------------
 kernel/arch/earm/memory.c     |  4 +++-
 kernel/arch/earm/omap_timer.c | 33 ++++++++++++++++++++++++++-------
 3 files changed, 37 insertions(+), 23 deletions(-)

diff --git a/kernel/arch/earm/arch_clock.c b/kernel/arch/earm/arch_clock.c
index 96d8811b2..67fc0ed45 100644
--- a/kernel/arch/earm/arch_clock.c
+++ b/kernel/arch/earm/arch_clock.c
@@ -52,13 +52,13 @@ void cycles_accounting_init(void)
 
 void context_stop(struct proc * p)
 {
-	u64_t tsc, tsc_delta;
+	u64_t tsc;
+	u32_t tsc_delta;
 	u64_t * __tsc_ctr_switch = get_cpulocal_var_ptr(tsc_ctr_switch);
 
 	read_tsc_64(&tsc);
-	p->p_cycles = add64(p->p_cycles, sub64(tsc, *__tsc_ctr_switch));
-
-	tsc_delta = sub64(tsc, *__tsc_ctr_switch);
+	tsc_delta = tsc - *__tsc_ctr_switch;
+	p->p_cycles += tsc_delta;
 
 	if(kbill_ipc) {
 		kbill_ipc->p_kipc_cycles =
@@ -79,18 +79,11 @@ void context_stop(struct proc * p)
 	 */
 	if (p->p_endpoint >= 0) {
 #if DEBUG_RACE
-		make_zero64(p->p_cpu_time_left);
+		p->p_cpu_time_left = 0;
 #else
-		/* if (tsc_delta < p->p_cpu_time_left) in 64bit */
-		if (ex64hi(tsc_delta) < ex64hi(p->p_cpu_time_left) ||
-				(ex64hi(tsc_delta) == ex64hi(p->p_cpu_time_left) &&
-				 ex64lo(tsc_delta) < ex64lo(p->p_cpu_time_left)))
-		{
-			p->p_cpu_time_left = sub64(p->p_cpu_time_left, tsc_delta);
-		}
-		else {
-			make_zero64(p->p_cpu_time_left);
-		}
+		if (tsc_delta < p->p_cpu_time_left) {
+			p->p_cpu_time_left -= tsc_delta;
+		} else p->p_cpu_time_left = 0;
 #endif
 	}
 
diff --git a/kernel/arch/earm/memory.c b/kernel/arch/earm/memory.c
index f422d43be..9608868a8 100644
--- a/kernel/arch/earm/memory.c
+++ b/kernel/arch/earm/memory.c
@@ -29,6 +29,8 @@ static int freepdes[MAXFREEPDES];
 
 static u32_t phys_get32(phys_bytes v);
 
+extern vir_bytes omap3_gptimer10_base = OMAP3_GPTIMER10_BASE;
+
 void mem_clear_mapcache(void)
 {
 	int i;
@@ -763,7 +765,7 @@ int arch_phys_map_reply(const int index, const vir_bytes addr)
 		return OK;
 	}
 	else if (index == frclock_index) {
-		minix_kerninfo.minix_frclock = addr;
+		omap3_gptimer10_base = minix_kerninfo.minix_frclock = addr;
 		return OK;
 	}
 
diff --git a/kernel/arch/earm/omap_timer.c b/kernel/arch/earm/omap_timer.c
index a5bac65a9..cc7dbb402 100644
--- a/kernel/arch/earm/omap_timer.c
+++ b/kernel/arch/earm/omap_timer.c
@@ -8,7 +8,9 @@
 #include "omap_intr.h"
 
 static irq_hook_t omap3_timer_hook;		/* interrupt handler hook */
-static u64_t tsc;
+static u64_t high_frc;
+
+vir_bytes omap3_gptimer10_base;
 
 int omap3_register_timer_handler(const irq_handler_t handler)
 {
@@ -90,6 +92,23 @@ void omap3_timer_stop()
     mmio_clear(OMAP3_GPTIMER1_TCLR, OMAP3_TCLR_ST);
 }
 
+static u32_t read_frc(void)
+{
+	u32_t frc = *(u32_t *) ((char *) omap3_gptimer10_base + OMAP3_TCRR);
+	return frc;
+}
+
+static void frc_overflow_check(void)
+{
+	static int prev_frc_valid;
+	static u32_t prev_frc;
+	u32_t cur_frc = read_frc();
+	if(prev_frc_valid && prev_frc > cur_frc)
+		high_frc++;
+	prev_frc = cur_frc;
+	prev_frc_valid = 1;
+}
+
 void omap3_timer_int_handler()
 {
     /* Clear all interrupts */
@@ -98,15 +117,15 @@ void omap3_timer_int_handler()
     tisr = OMAP3_TISR_MAT_IT_FLAG | OMAP3_TISR_OVF_IT_FLAG |
            OMAP3_TISR_TCAR_IT_FLAG;
     mmio_write(OMAP3_GPTIMER1_TISR, tisr);
-    tsc++;
 
+   frc_overflow_check();
 }
 
-/* Don't use libminlib's read_tsc_64, but our own version instead. We emulate
- * the ARM Cycle Counter (CCNT) with 1 cycle per ms. We can't rely on the
- * actual counter hardware to be working (i.e., qemu doesn't emulate it at all)
- */
+/* Use the free running clock as TSC */
 void read_tsc_64(u64_t *t)
 {
-    *t = tsc;
+	u32_t now;
+   	frc_overflow_check();
+	now = read_frc();
+	*t = (u64_t) now + (high_frc << 32);
 }
-- 
2.44.0