Re: 2.6.9-rc2-mm2

[William Lee Irwin III <wli@holomorphy.com>]

William Lee Irwin III wrote:
>> Something's a tad off here. Should be easy enough to fix up.
[...]
>> TPC: <sched_clock+0xc/0x40>

On Thu, Sep 23, 2004 at 03:46:41PM +1000, Peter Williams wrote:
> This looks the problem of sched_clock() being called before it's ready 
> (that we experienced with 2.6.9-rc2 on IA32 systems) only this time it's 
> fatal :-(
> A quick workaround for this would be to initialize idle->sched_timestamp 
> in init_idle() and current->sched_timestamp in sched_init() to the 
> INITIAL_JIFFIES converted to nanoseconds instead of using sched_clock().
> Another solution would be to set them to a value much greater than the 
> nanosecond equivalent of INITIAL_JIFFIES (e.g. 1ULL << 63) and let the 
> code that handles the non monotonic behaviour of sched_clock() sort it 
> out later.

Well, I posted a quick hack to get it to tolerate being called so early.
Might be better if I statically initialized the thing to a dummy driver
so only the indirect call remains at runtime. e.g.:


-- wli

Index: mm2-2.6.9-rc2/arch/sparc64/kernel/time.c
===================================================================
--- mm2-2.6.9-rc2.orig/arch/sparc64/kernel/time.c	2004-09-22 21:33:03.000000000 -0700
+++ mm2-2.6.9-rc2/arch/sparc64/kernel/time.c	2004-09-22 22:59:35.980157226 -0700
@@ -64,7 +64,16 @@
 
 static int set_rtc_mmss(unsigned long);
 
-struct sparc64_tick_ops *tick_ops;
+static __init unsigned long dummy_get_tick(void)
+{
+	return 0;
+}
+
+static __initdata struct sparc64_tick_ops dummy_tick_ops = {
+	.get_tick	= dummy_get_tick,
+};
+
+struct sparc64_tick_ops *tick_ops = &dummy_tick_ops;
 
 #define TICK_PRIV_BIT	(1UL << 63)