2023-08-08 18:49:47

by Yicong Yang

Subject: [PATCH] drivers/perf: hisi: Schedule perf session according to locality

From: Yicong Yang <[email protected]>

The PCIe PMUs locate on different NUMA node but currently we don't
consider it and likely stack all the sessions on the same CPU:

[root@localhost tmp]# cat /sys/devices/hisi_pcie*/cpumask
0
0
0
0
0
0

This can be optimize a bit to use a local CPU for the PMU.

Signed-off-by: Yicong Yang <[email protected]>
---
drivers/perf/hisilicon/hisi_pcie_pmu.c | 15 ++++++++++++---
1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/drivers/perf/hisilicon/hisi_pcie_pmu.c b/drivers/perf/hisilicon/hisi_pcie_pmu.c
index e10fc7cb9493..60ecf469782b 100644
--- a/drivers/perf/hisilicon/hisi_pcie_pmu.c
+++ b/drivers/perf/hisilicon/hisi_pcie_pmu.c
@@ -665,7 +665,7 @@ static int hisi_pcie_pmu_online_cpu(unsigned int cpu, struct hlist_node *node)
struct hisi_pcie_pmu *pcie_pmu = hlist_entry_safe(node, struct hisi_pcie_pmu, node);

if (pcie_pmu->on_cpu == -1) {
- pcie_pmu->on_cpu = cpu;
+ pcie_pmu->on_cpu = cpumask_local_spread(0, dev_to_node(&pcie_pmu->pdev->dev));
WARN_ON(irq_set_affinity(pcie_pmu->irq, cpumask_of(cpu)));
}

@@ -676,14 +676,23 @@ static int hisi_pcie_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
{
struct hisi_pcie_pmu *pcie_pmu = hlist_entry_safe(node, struct hisi_pcie_pmu, node);
unsigned int target;
+ cpumask_t mask;
+ int numa_node;

/* Nothing to do if this CPU doesn't own the PMU */
if (pcie_pmu->on_cpu != cpu)
return 0;

pcie_pmu->on_cpu = -1;
- /* Choose a new CPU from all online cpus. */
- target = cpumask_any_but(cpu_online_mask, cpu);
+
+ /* Choose a local CPU from all online cpus. */
+ numa_node = dev_to_node(&pcie_pmu->pdev->dev);
+ if (cpumask_and(&mask, cpumask_of_node(numa_node), cpu_online_mask) &&
+ cpumask_andnot(&mask, &mask, cpumask_of(cpu)))
+ target = cpumask_any(&mask);
+ else
+ target = cpumask_any_but(cpu_online_mask, cpu);
+
if (target >= nr_cpu_ids) {
pci_err(pcie_pmu->pdev, "There is no CPU to set\n");
return 0;
--
2.24.0

2023-08-11 12:03:15

by Will Deacon

Subject: Re: [PATCH] drivers/perf: hisi: Schedule perf session according to locality

On Tue, Aug 08, 2023 at 08:51:47PM +0800, Yicong Yang wrote:
> From: Yicong Yang <[email protected]>
>
> The PCIe PMUs locate on different NUMA node but currently we don't
> consider it and likely stack all the sessions on the same CPU:
>
> [root@localhost tmp]# cat /sys/devices/hisi_pcie*/cpumask
> 0
> 0
> 0
> 0
> 0
> 0
>
> This can be optimize a bit to use a local CPU for the PMU.
>
> Signed-off-by: Yicong Yang <[email protected]>
> ---
> drivers/perf/hisilicon/hisi_pcie_pmu.c | 15 ++++++++++++---
> 1 file changed, 12 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/perf/hisilicon/hisi_pcie_pmu.c b/drivers/perf/hisilicon/hisi_pcie_pmu.c
> index e10fc7cb9493..60ecf469782b 100644
> --- a/drivers/perf/hisilicon/hisi_pcie_pmu.c
> +++ b/drivers/perf/hisilicon/hisi_pcie_pmu.c
> @@ -665,7 +665,7 @@ static int hisi_pcie_pmu_online_cpu(unsigned int cpu, struct hlist_node *node)
> struct hisi_pcie_pmu *pcie_pmu = hlist_entry_safe(node, struct hisi_pcie_pmu, node);
>
> if (pcie_pmu->on_cpu == -1) {
> - pcie_pmu->on_cpu = cpu;
> + pcie_pmu->on_cpu = cpumask_local_spread(0, dev_to_node(&pcie_pmu->pdev->dev));
> WARN_ON(irq_set_affinity(pcie_pmu->irq, cpumask_of(cpu)));

Hmm, this is a bit weird now, because the interrupt is affine to a different
CPU from the one you've chosen. Are you sure that's ok? When the offline
notifier picks a new target, it moves the irq as well.

Will

2023-08-15 18:53:13

by Yicong Yang

Subject: Re: [PATCH] drivers/perf: hisi: Schedule perf session according to locality

On 2023/8/11 19:14, Will Deacon wrote:
> On Tue, Aug 08, 2023 at 08:51:47PM +0800, Yicong Yang wrote:
>> From: Yicong Yang <[email protected]>
>>
>> The PCIe PMUs locate on different NUMA node but currently we don't
>> consider it and likely stack all the sessions on the same CPU:
>>
>> [root@localhost tmp]# cat /sys/devices/hisi_pcie*/cpumask
>> 0
>> 0
>> 0
>> 0
>> 0
>> 0
>>
>> This can be optimize a bit to use a local CPU for the PMU.
>>
>> Signed-off-by: Yicong Yang <[email protected]>
>> ---
>> drivers/perf/hisilicon/hisi_pcie_pmu.c | 15 ++++++++++++---
>> 1 file changed, 12 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/perf/hisilicon/hisi_pcie_pmu.c b/drivers/perf/hisilicon/hisi_pcie_pmu.c
>> index e10fc7cb9493..60ecf469782b 100644
>> --- a/drivers/perf/hisilicon/hisi_pcie_pmu.c
>> +++ b/drivers/perf/hisilicon/hisi_pcie_pmu.c
>> @@ -665,7 +665,7 @@ static int hisi_pcie_pmu_online_cpu(unsigned int cpu, struct hlist_node *node)
>> struct hisi_pcie_pmu *pcie_pmu = hlist_entry_safe(node, struct hisi_pcie_pmu, node);
>>
>> if (pcie_pmu->on_cpu == -1) {
>> - pcie_pmu->on_cpu = cpu;
>> + pcie_pmu->on_cpu = cpumask_local_spread(0, dev_to_node(&pcie_pmu->pdev->dev));
>> WARN_ON(irq_set_affinity(pcie_pmu->irq, cpumask_of(cpu)));
>
> Hmm, this is a bit weird now, because the interrupt is affine to a different
> CPU from the one you've chosen. Are you sure that's ok? When the offline
> notifier picks a new target, it moves the irq as well.
>

Thanks for pointing it out. This is indeed a problem. Will fix this.

Thanks.