When the "affinity=" cmdline parameter is configured, the interrupt
affinity displayed in the proc directory does not match with that of the
the percu interrupt, and the percu interrupt uses desc->percu_affinity.
So show desc->percu_affinity in show_irq_affinity() for percpu
interrupt.
Signed-off-by: yeyunfeng <[email protected]>
---
kernel/irq/proc.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/kernel/irq/proc.c b/kernel/irq/proc.c
index 32c071d7bc03..b9d0fa87b4b4 100644
--- a/kernel/irq/proc.c
+++ b/kernel/irq/proc.c
@@ -52,6 +52,8 @@ static int show_irq_affinity(int type, struct seq_file *m)
case AFFINITY:
case AFFINITY_LIST:
mask = desc->irq_common_data.affinity;
+ if (irqd_is_per_cpu(&desc->irq_data))
+ mask = desc->percpu_affinity;
#ifdef CONFIG_GENERIC_PENDING_IRQ
if (irqd_is_setaffinity_pending(&desc->irq_data))
mask = desc->pending_mask;
--
1.8.3.1
Yunfeng Ye <[email protected]> writes:
> When the "affinity=" cmdline parameter is configured,
There is no such parameter.
> the interrupt affinity displayed in the proc directory does not match
> with that of the the percu interrupt, and the percu interrupt uses
> desc->percu_affinity.
And when the non-existing parameter is not on the command line then
irq->affinity is showing the correct value magically?
Definitely not: It's unconditionally showing irq->affinity and that is
pretty unlikely to match irq->percpu_affinity in any case.
> diff --git a/kernel/irq/proc.c b/kernel/irq/proc.c
> index 32c071d7bc03..b9d0fa87b4b4 100644
> --- a/kernel/irq/proc.c
> +++ b/kernel/irq/proc.c
> @@ -52,6 +52,8 @@ static int show_irq_affinity(int type, struct seq_file *m)
> case AFFINITY:
> case AFFINITY_LIST:
> mask = desc->irq_common_data.affinity;
> + if (irqd_is_per_cpu(&desc->irq_data))
> + mask = desc->percpu_affinity;
This breaks all architecture which mark interrupts as per CPU without
using the partition mechanism resulting in a NULL pointer dereference.
Thanks,
tglx
On 2020/8/13 16:27, Thomas Gleixner wrote:
> Yunfeng Ye <[email protected]> writes:
>
>> When the "affinity=" cmdline parameter is configured,
>
> There is no such parameter.
>
>> the interrupt affinity displayed in the proc directory does not match
>> with that of the the percu interrupt, and the percu interrupt uses
>> desc->percu_affinity.
>
> And when the non-existing parameter is not on the command line then
> irq->affinity is showing the correct value magically?
>
> Definitely not: It's unconditionally showing irq->affinity and that is
> pretty unlikely to match irq->percpu_affinity in any case.
>
Sorry，it is "irqaffinity=" cmdline parameter. it will set irq_default_affinity
mask. if the interrupt is not the managed irq, the irq_desc will use
irq_default_affinity as default affinity.
For example, the cmdline "irqaffinity=0,1,126,127" on the 128 cores system：
[root@localhost ~]# cat /proc/irq/4/smp_affinity_list
0-1,126-127
The irq 4 is "arch_timer" interrupt, which is a percpu interrupt.
So is it necessary to show the percpu irq affinity correct?
thanks.
>> diff --git a/kernel/irq/proc.c b/kernel/irq/proc.c
>> index 32c071d7bc03..b9d0fa87b4b4 100644
>> --- a/kernel/irq/proc.c
>> +++ b/kernel/irq/proc.c
>> @@ -52,6 +52,8 @@ static int show_irq_affinity(int type, struct seq_file *m)
>> case AFFINITY:
>> case AFFINITY_LIST:
>> mask = desc->irq_common_data.affinity;
>> + if (irqd_is_per_cpu(&desc->irq_data))
>> + mask = desc->percpu_affinity;
>
> This breaks all architecture which mark interrupts as per CPU without
> using the partition mechanism resulting in a NULL pointer dereference.
>
Yes, it is a problem.
> Thanks,
>
> tglx
>
> .
>
On Sat, Aug 22 2020 at 17:33, Yunfeng Ye wrote:
> On 2020/8/13 16:27, Thomas Gleixner wrote:
> For example, the cmdline "irqaffinity=0,1,126,127" on the 128 cores system：
>
> [root@localhost ~]# cat /proc/irq/4/smp_affinity_list
> 0-1,126-127
>
> The irq 4 is "arch_timer" interrupt, which is a percpu interrupt.
>
> So is it necessary to show the percpu irq affinity correct?
Yes, it makes sense to do so, but you used the wrong check. The correct
one is:
irq_settings_is_per_cpu_devid()
which will not wreckage the output for other per cpu marked interrupts
which never set the percpu_affinity pointer with the obvious
consequences... The pointer would need a NULL check in any case, but it
might be more straight forward to update affinity when percpu_affinity
is initialized. Haven't looked in detail though.
Thanks,
tglx
On 2020/8/22 19:35, Thomas Gleixner wrote:
> On Sat, Aug 22 2020 at 17:33, Yunfeng Ye wrote:
>> On 2020/8/13 16:27, Thomas Gleixner wrote:
>> For example, the cmdline "irqaffinity=0,1,126,127" on the 128 cores system：
>>
>> [root@localhost ~]# cat /proc/irq/4/smp_affinity_list
>> 0-1,126-127
>>
>> The irq 4 is "arch_timer" interrupt, which is a percpu interrupt.
>>
>> So is it necessary to show the percpu irq affinity correct?
>
> Yes, it makes sense to do so, but you used the wrong check. The correct
> one is:
>
> irq_settings_is_per_cpu_devid()
>
> which will not wreckage the output for other per cpu marked interrupts
> which never set the percpu_affinity pointer with the obvious
> consequences... The pointer would need a NULL check in any case, but it
> might be more straight forward to update affinity when percpu_affinity
> is initialized. Haven't looked in detail though.
>
ok, I will send the patch v2, thanks.
> Thanks,
>
> tglx
>
> .
>