Hello,
First of all, altough I added Patch in the subject and post it here, I don't think that this patch should be applied to rt tree... Read the rest of the email to see why...
On my ARM SoC with a marvell orion chip I get weird cpu usage values in top when I load my kernel module (different processes shows very high cpu usage). My kernel module sends ethernet frames every 200us. Therefor I use hrtimers and send this network frames in the callback method of the timer. I use 2.6.29.1-rt9. All the involved processes (IRQ-21 (irq of the network interface), sirq-net-rx, sirq-hrtimer) gets randomly high cpu usage. Sometimes I get no cpu usage at all. Its different each time I load my module. Sometimes IRQ-21 gets high cpu load, sometimes sirq-net-rx... Load average seems to show realistic values (~0.6).
This have a unpleasant side effect: Due to the limited runtime for rt tasks (in /proc/sys/kernel/sched_rt_runtime_us), i get a delay of 50ms once every second. When i disable the limit (echo -1 >/proc/sys/kernel/sched_rt_runtime_us) i don't get this delay. However, it would be great if the cpu usage per process would be fixed anyway....
Therefor I implemented a sched_clock implementation for my orion platform (see attached patch). I tested the values that sched_clock returns, it should be ok! But it still doesn't work...
Any ideas?
Thanks
Stefan aka falstaff
From 9ab3c6030dc01ffe4065c334cac1a07c0ff6af5d Mon Sep 17 00:00:00 2001
From: Stefan Agner <stefan.agner@xxxxxxxxx>
Date: Mon, 27 Apr 2009 11:11:18 +0200
Subject: [PATCH] sched_clock implementation for orion platforms
---
arch/arm/plat-orion/time.c | 35 +++++++++++++++++++++++++++++++++++
1 files changed, 35 insertions(+), 0 deletions(-)
diff --git a/arch/arm/plat-orion/time.c b/arch/arm/plat-orion/time.c
index 6fa2923..f873f67 100644
--- a/arch/arm/plat-orion/time.c
+++ b/arch/arm/plat-orion/time.c
@@ -17,6 +17,8 @@
#include <linux/irq.h>
#include <asm/mach/time.h>
#include <mach/hardware.h>
+#include <linux/sched.h>
+#include <linux/cnt32_to_63.h>
/*
* Number of timer ticks per jiffy.
@@ -39,6 +41,35 @@ static u32 ticks_per_jiffy;
/*
+ * Orion's sched_clock implementation. It has a resolution
+ * of at least 7.5ns (133MHz TCLK) and a maximum value of
+ * 32 seconds.
+ */
+#define ORIONCLK2NS_SCALE_FACTOR 20
+
+static unsigned long orion2ns_scale;
+
+static void __init set_orion2ns_scale(unsigned long orion_rate)
+{
+ unsigned long long v = ((unsigned long long)NSEC_PER_SEC) << ORIONCLK2NS_SCALE_FACTOR;
+ do_div(v, orion_rate);
+ /*
+ * We want an even value to automatically clear the top bit
+ * returned by cnt32_to_63() without an additional run time
+ * instruction. So if the LSB is 1 then round it up.
+ */
+ if(v & 1)
+ v++;
+ orion2ns_scale = v;
+}
+
+unsigned long long sched_clock(void)
+{
+ unsigned long long v = cnt32_to_63(0xffffffff - readl(TIMER0_VAL));
+ return (v * orion2ns_scale) >> ORIONCLK2NS_SCALE_FACTOR;
+}
+
+/*
* Clocksource handling.
*/
static cycle_t orion_clksrc_read(void)
@@ -176,6 +207,10 @@ void __init orion_time_init(unsigned int irq, unsigned int tclk)
ticks_per_jiffy = (tclk + HZ/2) / HZ;
+ /*
+ * Set scale for sched_clock
+ */
+ set_orion2ns_scale(tclk);
/*
* Setup free-running clocksource timer (interrupts
--
1.5.6.3
