This came to light in some internal discussions and it is nice to have this documented rather than digging up the PRM (Prog Ref Manual) again. Signed-off-by: Eugeniy Paltsev <Eugeniy.Paltsev at synopsys.com> --- drivers/clocksource/arc_timer.c | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/drivers/clocksource/arc_timer.c b/drivers/clocksource/arc_timer.c index 4927355f9cbe..b594c373debc 100644 --- a/drivers/clocksource/arc_timer.c +++ b/drivers/clocksource/arc_timer.c @@ -61,6 +61,19 @@ static u64 arc_read_gfrc(struct clocksource *cs) unsigned long flags; u32 l, h; + /* + * MCIP_CMD/MCIP_READBACK registers are allocated + * PER ARC CORE (not per cluster), and there are dedicated hardware + * decode logic (per core) inside ARConnect to handle simultaneous + * read/write accesses from cores via those two registers. + * So several concurrent commands to ARConnect are OK if they are + * trying to access two different sub-components (like GFRC, + * inter-core interrupt, etc...). HW also support simultaneously + * accessing GFRC by multiple cores. + * That's why it is safe to disable hard interrupts on the local CPU + * before access to GFRC instead of taking global MCIP spinlock + * defined in arch/arc/kernel/mcip.c + */ local_irq_save(flags); __mcip_cmd(CMD_GFRC_READ_LO, 0); -- 2.14.3
