[v3,1/3] ARM: rockchip: fix the CPU soft reset

Commit Message

Heiko Stübner June 5, 2015, 3:52 p.m. UTC

Hi Caesar,

Am Freitag, 5. Juni 2015, 23:11:42 schrieb Caesar Wang:
> We need different orderings when turning a core on and turning a core
> off.  In one case we need to assert reset before turning power off.
> In ther other case we need to turn power on and the deassert reset.
> 
> In general, the correct flow is:
> 
> CPU off:
>     reset_control_assert
>     regmap_update_bits(pmu, PMU_PWRDN_CON, BIT(pd), BIT(pd))
> CPU on:
>     regmap_update_bits(pmu, PMU_PWRDN_CON, BIT(pd), 0)
>     reset_control_deassert
> 
> This is needed for stressing CPU up/down, as per:
>     cd /sys/devices/system/cpu/
>     for i in $(seq 1000); do
>         echo "================= $i ============"
>         for j in $(seq 100); do
>             while [[ "$(cat cpu1/online)$(cat cpu2/online)$(cat
> cpu3/online)" != "000"" ]] echo 0 > cpu1/online
>                 echo 0 > cpu2/online
>                 echo 0 > cpu3/online
>             done
>             while [[ "$(cat cpu1/online)$(cat cpu2/online)$(cat
> cpu3/online)" != "111" ]]; do echo 1 > cpu1/online
>                 echo 1 > cpu2/online
>                 echo 1 > cpu3/online
>             done
>         done
>     done
> 
> The following is reproducile log:
>     [34466.186812] PM: noirq suspend of devices complete after 0.669 msecs
>     [34466.186824] Disabling non-boot CPUs ...
>     [34466.187509] CPU1: shutdown
>     [34466.188672] CPU2: shutdown
>     [34473.736627] Kernel panic - not syncing:Watchdog detected hard LOCKUP
> on cpu 0 .......
> 
> Signed-off-by: Caesar Wang <[e-mail]>
> ---

could we do this something like the shown below instead?
Here the deassertion of the reset really happens after we are sure the
power-domain is on


-------------- 8< ---------------
-------------- 8< ---------------

Comments

Heiko Stübner June 5, 2015, 4:17 p.m. UTC | #1

Hi Caesar,

Am Freitag, 5. Juni 2015, 17:52:48 schrieb Heiko Stübner:
> could we do this something like the shown below instead?
> Here the deassertion of the reset really happens after we are sure the
> power-domain is on
 
As you mention in the gerrit cl that the hang does happen when you
deassert the reset before the power-domain wait, ignore my last message please 
:-)

But please make sure to also provide a changelog to the list submissions in 
the future.


Thanks
Heiko

Patch

diff --git a/arch/arm/mach-rockchip/platsmp.c b/arch/arm/mach-rockchip/platsmp.c
index 5b4ca3c..ee5dbad6 100644
--- a/arch/arm/mach-rockchip/platsmp.c
+++ b/arch/arm/mach-rockchip/platsmp.c
@@ -72,6 +72,7 @@  static struct reset_control *rockchip_get_core_reset(int cpu)
 static int pmu_set_power_domain(int pd, bool on)
 {
 	u32 val = (on) ? 0 : BIT(pd);
+	struct reset_control *rstc = rockchip_get_core_reset(pd);
 	int ret;
 
 	/*
@@ -80,20 +81,16 @@  static int pmu_set_power_domain(int pd, bool on)
 	 * processor is powered down.
 	 */
 	if (read_cpuid_part() != ARM_CPU_PART_CORTEX_A9) {
-		struct reset_control *rstc = rockchip_get_core_reset(pd);
-
+		/* We only require the reset on the RK3288 at the moment */
 		if (IS_ERR(rstc)) {
 			pr_err("%s: could not get reset control for core %d\n",
 			       __func__, pd);
 			return PTR_ERR(rstc);
 		}
 
-		if (on)
-			reset_control_deassert(rstc);
-		else
+		if (!on)
 			reset_control_assert(rstc);
 
-		reset_control_put(rstc);
 	}
 
 	ret = regmap_update_bits(pmu, PMU_PWRDN_CON, BIT(pd), val);
@@ -112,6 +109,12 @@  static int pmu_set_power_domain(int pd, bool on)
 		}
 	}
 
+	if (read_cpuid_part() != ARM_CPU_PART_CORTEX_A9 && on)
+			reset_control_deassert(rstc);
+
+	if (!IS_ERR(rstc))
+		reset_control_put(rstc);
+
 	return 0;
 }