Commit 1e266629 authored by Vineet Gupta's avatar Vineet Gupta

ARC: 64bit RTSC timestamp hardware issue

The 64bit RTSC is not reliable, causing spurious "jumps" in higher word,
making Linux timekeeping go bonkers. So as of now just use the lower
32bit timestamp.

A cleaner approach would have been removing RTSC support altogether as the
32bit RTSC is equivalent to old TIMER1 based solution, but some customers
can use the 32bit RTSC in SMP syn fashion (vs. TIMER1 which being incore
can't be done easily).

A fallout of this is sched_clock()'s hardware assisted version needs to
go away since it can't use 32bit wrapping counter - instead we use the
generic "weak" jiffies based version.
Signed-off-by: default avatarVineet Gupta <vgupta@synopsys.com>
parent d626f547
......@@ -76,7 +76,7 @@ static cycle_t arc_counter_read(struct clocksource *cs)
__asm__ __volatile(
" .extCoreRegister tsch, 58, r, cannot_shortcut \n"
" rtsc %0, 0 \n"
" mov %1, tsch \n" /* TSCH is extn core reg 58 */
" mov %1, 0 \n"
: "=r" (stamp.low), "=r" (stamp.high));
arch_local_irq_restore(flags);
......@@ -88,7 +88,7 @@ static struct clocksource arc_counter = {
.name = "ARC RTSC",
.rating = 300,
.read = arc_counter_read,
.mask = CLOCKSOURCE_MASK(64),
.mask = CLOCKSOURCE_MASK(32),
.flags = CLOCK_SOURCE_IS_CONTINUOUS,
};
......@@ -263,37 +263,3 @@ void __init time_init(void)
if (machine_desc->init_time)
machine_desc->init_time();
}
#ifdef CONFIG_ARC_HAS_RTSC
/*
* sched_clock math assist
* ns = cycles * (ns_per_sec / cpu_freq_hz)
* ns = cycles * (10^6 / cpu_freq_khz)
* ns = cycles * (10^6 * 2^SF / cpu_freq_khz) / 2^SF
* ns = cycles * cyc2ns_scale >> SF
*/
#define CYC2NS_SF 10 /* 2^10, carefully chosen */
#define CYC2NS_SCALE ((1000000 << CYC2NS_SF) / (arc_get_core_freq() / 1000))
static unsigned long long cycles2ns(unsigned long long cyc)
{
return (cyc * CYC2NS_SCALE ) >> CYC2NS_SF;
}
/*
* Scheduler clock - a monotonically increasing clock in nanosec units.
* It's return value must NOT wrap around.
*
* - Since 32bit TIMER1 will overflow almost immediately (53sec @ 80MHz), it
* can't be used directly.
* - Using getrawmonotonic (TIMER1 based, but with state for last + current
* snapshots), is no-good either because of seqlock deadlock possibilities
* - So only with native 64bit timer we do this, otherwise fallback to generic
* jiffies based version - which despite not being fine grained gaurantees
* the monotonically increasing semantics.
*/
unsigned long long sched_clock(void)
{
return cycles2ns(arc_counter_read(NULL));
}
#endif
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment