2015-05-19 19:00:33

by Shreyas B. Prabhu

[permalink] [raw]
Subject: [PATCH] powerpc: Make doorbell check preemption safe

Doorbell can be used to cause ipi on cpus which are sibling threads on
the same core. So icp_native_cause_ipi checks if the destination cpu
is a sibling thread of the current cpu and uses doorbell in such cases.

But while running with CONFIG_PREEMPT=y, since this section is
preemtible, we can run into issues if after we check if the destination
cpu is a sibling cpu, the task gets migrated from a sibling cpu to a
cpu on another core.

Fix this by using get_cpu()/ put_cpu()

Signed-off-by: Shreyas B. Prabhu <[email protected]>
---
arch/powerpc/sysdev/xics/icp-native.c | 14 +++++++++-----
1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/arch/powerpc/sysdev/xics/icp-native.c b/arch/powerpc/sysdev/xics/icp-native.c
index 2fc4cf1..62fd478 100644
--- a/arch/powerpc/sysdev/xics/icp-native.c
+++ b/arch/powerpc/sysdev/xics/icp-native.c
@@ -147,12 +147,16 @@ static void icp_native_cause_ipi(int cpu, unsigned long data)
{
kvmppc_set_host_ipi(cpu, 1);
#ifdef CONFIG_PPC_DOORBELL
- if (cpu_has_feature(CPU_FTR_DBELL) &&
- (cpumask_test_cpu(cpu, cpu_sibling_mask(smp_processor_id()))))
- doorbell_cause_ipi(cpu, data);
- else
+ if (cpu_has_feature(CPU_FTR_DBELL)) {
+ if (cpumask_test_cpu(cpu, cpu_sibling_mask(get_cpu()))) {
+ doorbell_cause_ipi(cpu, data);
+ put_cpu();
+ return;
+ }
+ put_cpu();
+ }
#endif
- icp_native_set_qirr(cpu, IPI_PRIORITY);
+ icp_native_set_qirr(cpu, IPI_PRIORITY);
}

/*
--
1.9.3


2015-05-20 01:00:12

by Michael Neuling

[permalink] [raw]
Subject: Re: [PATCH] powerpc: Make doorbell check preemption safe

On Wed, 2015-05-20 at 00:30 +0530, Shreyas B. Prabhu wrote:
> Doorbell can be used to cause ipi on cpus which are sibling threads on
> the same core. So icp_native_cause_ipi checks if the destination cpu
> is a sibling thread of the current cpu and uses doorbell in such cases.
>
> But while running with CONFIG_PREEMPT=y, since this section is
> preemtible, we can run into issues if after we check if the destination
> cpu is a sibling cpu, the task gets migrated from a sibling cpu to a
> cpu on another core.
>
> Fix this by using get_cpu()/ put_cpu()

Thanks. Looks good and it's boots for me.

Signed-off-by: Michael Neuling <[email protected]>

> Signed-off-by: Shreyas B. Prabhu <[email protected]>
> ---
> arch/powerpc/sysdev/xics/icp-native.c | 14 +++++++++-----
> 1 file changed, 9 insertions(+), 5 deletions(-)
>
> diff --git a/arch/powerpc/sysdev/xics/icp-native.c b/arch/powerpc/sysdev/xics/icp-native.c
> index 2fc4cf1..62fd478 100644
> --- a/arch/powerpc/sysdev/xics/icp-native.c
> +++ b/arch/powerpc/sysdev/xics/icp-native.c
> @@ -147,12 +147,16 @@ static void icp_native_cause_ipi(int cpu, unsigned long data)
> {
> kvmppc_set_host_ipi(cpu, 1);
> #ifdef CONFIG_PPC_DOORBELL
> - if (cpu_has_feature(CPU_FTR_DBELL) &&
> - (cpumask_test_cpu(cpu, cpu_sibling_mask(smp_processor_id()))))
> - doorbell_cause_ipi(cpu, data);
> - else
> + if (cpu_has_feature(CPU_FTR_DBELL)) {
> + if (cpumask_test_cpu(cpu, cpu_sibling_mask(get_cpu()))) {
> + doorbell_cause_ipi(cpu, data);
> + put_cpu();
> + return;
> + }
> + put_cpu();
> + }
> #endif
> - icp_native_set_qirr(cpu, IPI_PRIORITY);
> + icp_native_set_qirr(cpu, IPI_PRIORITY);
> }
>
> /*

2015-06-04 06:46:15

by Shreyas B. Prabhu

[permalink] [raw]
Subject: Re: [PATCH] powerpc: Make doorbell check preemption safe



On Wednesday 20 May 2015 06:30 AM, Michael Neuling wrote:
> On Wed, 2015-05-20 at 00:30 +0530, Shreyas B. Prabhu wrote:
>> Doorbell can be used to cause ipi on cpus which are sibling threads on
>> the same core. So icp_native_cause_ipi checks if the destination cpu
>> is a sibling thread of the current cpu and uses doorbell in such cases.
>>
>> But while running with CONFIG_PREEMPT=y, since this section is
>> preemtible, we can run into issues if after we check if the destination
>> cpu is a sibling cpu, the task gets migrated from a sibling cpu to a
>> cpu on another core.
>>
>> Fix this by using get_cpu()/ put_cpu()
>
> Thanks. Looks good and it's boots for me.
>
> Signed-off-by: Michael Neuling <[email protected]>
>
mikey, Thanks!


mpe, if this looks ok, can you please pick it up?