[PATCH v3 1/2] drivers/perf: hisi: Support uncore ITS PMU

Tue Jul 14 05:56:12 PDT 2026

On 14/07/2026 1:32 pm, Uwe Kleine-König wrote:
> Hello,
> 
> On Mon, Jul 13, 2026 at 08:56:46PM +0800, Yushan Wang wrote:
>> Support uncore ITS PMU, which provides the capability of counting
>> the number of interrupts routed to ITS by interrupt catagories, and the
>> latency. It also supports collecting statistics of micro-ops of ITS.
>>
>> The driver adapts to HiSilicon uncore PMU framework. It does not support
>> overflow interruption, which is the same as NoC PMU, so a few dummy
>> functions or handling interrupts are left empty.
>>
>> Signed-off-by: Yushan Wang <wangyushan12 at huawei.com>
>> ---
>>   Documentation/admin-guide/perf/hisi-pmu.rst  |  13 +
>>   drivers/perf/hisilicon/Makefile              |   2 +-
>>   drivers/perf/hisilicon/hisi_uncore_its_pmu.c | 393 +++++++++++++++++++
>>   3 files changed, 407 insertions(+), 1 deletion(-)
>>   create mode 100644 drivers/perf/hisilicon/hisi_uncore_its_pmu.c
>>
>> diff --git a/Documentation/admin-guide/perf/hisi-pmu.rst b/Documentation/admin-guide/perf/hisi-pmu.rst
>> index d56b2d690709..278bd7e0ae60 100644
>> --- a/Documentation/admin-guide/perf/hisi-pmu.rst
>> +++ b/Documentation/admin-guide/perf/hisi-pmu.rst
>> @@ -128,6 +128,19 @@ channel with this option. The current supported channels are as follows:
>>   7. tt_en: NoC PMU supports counting only transactions that have tracetag set
>>   if this option is set. See the 2nd list for more information about tracetag.
>>   
>> +8. int_id: ITS PMU supports filtering by interrupt id, which is defined by
>> +hardware. Interrupt id takes up to 32 bits, and can be divided into 2 parts:
>> +
>> +- Upper 16 bits: DeviceID if counting LPI, PEID if counting SGI/PPI.
>> +- Lower 16 bits: EventID if counting LPI, IntID if counting SGI/PPI.
>> +
>> +int_id is a global configuration for each PMU instance. If multiple different
>> +int_id's are specified, the last came in will be effective. And if there are
>> +already filtered events running, new filtered events came in will be refused.
>> +
>> +9. int_en: A one-bit flag to tell if int_id is used to filter the statistics. It
>> +allows filtering 0 DeviceID and EventID.
>> +
>>   For HiSilicon uncore PMU v3 whose identifier is 0x40, some uncore PMUs are
>>   further divided into parts for finer granularity of tracing, each part has its
>>   own dedicated PMU, and all such PMUs together cover the monitoring job of events
>> diff --git a/drivers/perf/hisilicon/Makefile b/drivers/perf/hisilicon/Makefile
>> index 186be3d02238..5f28cfdb8a72 100644
>> --- a/drivers/perf/hisilicon/Makefile
>> +++ b/drivers/perf/hisilicon/Makefile
>> @@ -2,7 +2,7 @@
>>   obj-$(CONFIG_HISI_PMU) += hisi_uncore_pmu.o hisi_uncore_l3c_pmu.o \
>>   			  hisi_uncore_hha_pmu.o hisi_uncore_ddrc_pmu.o hisi_uncore_sllc_pmu.o \
>>   			  hisi_uncore_pa_pmu.o hisi_uncore_cpa_pmu.o hisi_uncore_uc_pmu.o \
>> -			  hisi_uncore_noc_pmu.o hisi_uncore_mn_pmu.o
>> +			  hisi_uncore_noc_pmu.o hisi_uncore_mn_pmu.o hisi_uncore_its_pmu.o
>>   
>>   obj-$(CONFIG_HISI_PCIE_PMU) += hisi_pcie_pmu.o
>>   obj-$(CONFIG_HNS3_PMU) += hns3_pmu.o
>> diff --git a/drivers/perf/hisilicon/hisi_uncore_its_pmu.c b/drivers/perf/hisilicon/hisi_uncore_its_pmu.c
>> new file mode 100644
>> index 000000000000..430e2b06cf4d
>> --- /dev/null
>> +++ b/drivers/perf/hisilicon/hisi_uncore_its_pmu.c
>> @@ -0,0 +1,393 @@
>> +// SPDX-License-Identifier: GPL-2.0
>> +/*
>> + * Driver for HiSilicon Uncore ITS PMU device
>> + *
>> + * Copyright (c) 2026 HiSilicon Technologies Co., Ltd.
>> + * Author: Yushan Wang <wangyushan12 at huawei.com>
>> + */
>> +#include <linux/bitops.h>
>> +#include <linux/cpuhotplug.h>
>> +#include <linux/device.h>
>> +#include <linux/io.h>
>> +#include <linux/mod_devicetable.h>
> 
> Please rely on linux/platform_device.h to provice acpi_device_id and
> drop the include for <linux/mod_devicetable.h>.
> 
>> +static int __init hisi_its_pmu_module_init(void)
>> +{
>> +	int ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
>> +					  "perf/hisi/its:online",
>> +					  hisi_uncore_pmu_online_cpu,
>> +					  hisi_uncore_pmu_offline_cpu);
> 
> <disclaimer>I don't know anything about the functions involved
> here</disclaimer>
> 
> This looks fishy. The module might be loaded on any machine, is it right
> to call hisilicon specific functions then?

This only registers the state with the hotplug mechanism - the callbacks 
themselves won't ever be invoked until the driver binds to something and 
calls cpuhp_state_add_instance() from probe.

It is true that the dynamic hotplug states are somewhat limited, so at 
worst it might consume one unnecessarily if the driver is built-in, but 
this is the status quo for system PMU drivers, and TBH there are enough 
*other* issues with this whole perf hotplug mess anyway that I wouldn't 
see this one as a major concern. (FWIW I have started another 
side-project to make it all go away, which is beginning to look 
promising...)

That said, in this particular case, HISI_PMU could seemingly do better 
than registering *multiple* states with the same exact callbacks - 
registering a single state from the common hisi_uncore_pmu.c framework, 
then just having all the subdrivers add their instances to that, would 
be something of an improvement for now.

Thanks,
Robin.

> 
>> +	if (ret < 0) {
>> +		pr_err("hisi_its_pmu: Fail to setup cpuhp callbacks, ret = %d\n", ret);
>> +		return ret;
>> +	}
>> +	hisi_its_pmu_cpuhp_state = ret;
>> +
>> +	ret = platform_driver_register(&hisi_its_pmu_driver);
>> +	if (ret)
>> +		cpuhp_remove_multi_state(hisi_its_pmu_cpuhp_state);
>> +
>> +	return ret;
>> +}
>> +module_init(hisi_its_pmu_module_init);
> 
> Best regards
> Uwe