On 04/16/2012 08:08 PM, Yong Zhang wrote:
On Mon, Apr 16, 2012 at 09:48:14AM -0700, David Daney wrote:
On 04/16/2012 12:25 AM, Yong Zhang wrote:
From: Yong Zhang<email@example.com>
Too early to enable irq will break some following action,
such as notify_cpu_starting().
Can you be more specific about what breaks?
wake up softirqd;
/* select wrong cpu */
Yeah, that looks broken to me too.
I don't get side effect with this patch.
Without this, where do irqs get enabled on the secondary CPUs?
cpu_idle() will handle it. But in fact we should not depend on
It is not done in cpu_idle() itself. If irqs are disabled upon entry to
cpu_idle() *and* need_resched(), then we call into schedule(). The irqs
are then enabled by the raw_spin_unlock_irq(&rq->lock) at the end of
But it seems there is not suitable place to put local_irq_enable(),
though ->smp_finish() looks like a more suitable place.
It would be better, but it seems like really it should be done in
cpu_idle() immediately after rcu_idle_enter().
When looking more at smp support on MIPS, there is more things I find.
Such as set_cpu_online() is called on CPU1, so there will be another race
window like above scenario. Please take a look at what commit 2baab4e9
intend to resolve.
Signed-off-by: Yong Zhang<firstname.lastname@example.org>
Cc: David Daney<email@example.com>
Cc: Ralf Baechle<firstname.lastname@example.org>
arch/mips/cavium-octeon/smp.c | 1 -
1 files changed, 0 insertions(+), 1 deletions(-)
diff --git a/arch/mips/cavium-octeon/smp.c b/arch/mips/cavium-octeon/smp.c
index 97e7ce9..7e65c88 100644
@@ -185,7 +185,6 @@ static void __cpuinit octeon_init_secondary(void)