diff mbox series

[RFC,DNM] perf: Add support for Qualcomm Last-Level Cache Controller PMU

Message ID 20230809-topic-llcc_pmu-v1-1-dd27bd1f44c9@linaro.org (mailing list archive)
State New, archived
Headers show
Series [RFC,DNM] perf: Add support for Qualcomm Last-Level Cache Controller PMU | expand

Commit Message

Konrad Dybcio Aug. 9, 2023, 8:09 p.m. UTC
Add support for the Qualcomm LLCC (Last-Level Cache Controller) PMU,
which provides a single event, expressing cache read misses.

Based on the vendor driver found in the msm-5.10 downstream kernel.

Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
---
Hi, I've been trying to get this driver going upstream by cleaning it
up and adding the necessary perf boilerplate (the original Qualcomm one
only pokes at the PMU from within the kernel itself) to use the
userspace tool.

I can not however get it to cooperate.. in this iteration I get a PMU
event registered (though with only a "raw" name - no "x OR y" like with
other PMUs on the system) as:

llcc_pmu/read_miss/                                [Kernel PMU event]

but the .read callback is never called when I run:

sudo perf stat -C 0 -a -e llcc_pmu/read_miss/ stress-ng -C 8 -c 8 -m 10

which always returns 0

if I add --always-kernel I get:
<not supported>      llcc_pmu/read_miss/

So, here's me asking for some help. It's probably missing some small
detail, as per usual..
---
 drivers/perf/Kconfig         |   8 ++
 drivers/perf/Makefile        |   1 +
 drivers/perf/qcom_llcc_pmu.c | 277 +++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 286 insertions(+)


---
base-commit: 21ef7b1e17d039053edaeaf41142423810572741
change-id: 20230809-topic-llcc_pmu-c6e9dbc36b12

Best regards,

Comments

Trilok Soni Aug. 24, 2023, 9:31 p.m. UTC | #1
On 8/9/2023 1:09 PM, Konrad Dybcio wrote:
> Add support for the Qualcomm LLCC (Last-Level Cache Controller) PMU,
> which provides a single event, expressing cache read misses.
> 
> Based on the vendor driver found in the msm-5.10 downstream kernel.
> 
> Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
> ---
> Hi, I've been trying to get this driver going upstream by cleaning it
> up and adding the necessary perf boilerplate (the original Qualcomm one
> only pokes at the PMU from within the kernel itself) to use the
> userspace tool.
> 
> I can not however get it to cooperate.. in this iteration I get a PMU
> event registered (though with only a "raw" name - no "x OR y" like with
> other PMUs on the system) as:
> 
> llcc_pmu/read_miss/                                [Kernel PMU event]
> 
> but the .read callback is never called when I run:
> 
> sudo perf stat -C 0 -a -e llcc_pmu/read_miss/ stress-ng -C 8 -c 8 -m 10
> 
> which always returns 0
> 
> if I add --always-kernel I get:
> <not supported>      llcc_pmu/read_miss/

Which SOC you are trying this on?
Konrad Dybcio Aug. 25, 2023, 11:50 a.m. UTC | #2
On 24.08.2023 23:31, Trilok Soni wrote:
> On 8/9/2023 1:09 PM, Konrad Dybcio wrote:
>> Add support for the Qualcomm LLCC (Last-Level Cache Controller) PMU,
>> which provides a single event, expressing cache read misses.
>>
>> Based on the vendor driver found in the msm-5.10 downstream kernel.
>>
>> Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
>> ---
>> Hi, I've been trying to get this driver going upstream by cleaning it
>> up and adding the necessary perf boilerplate (the original Qualcomm one
>> only pokes at the PMU from within the kernel itself) to use the
>> userspace tool.
>>
>> I can not however get it to cooperate.. in this iteration I get a PMU
>> event registered (though with only a "raw" name - no "x OR y" like with
>> other PMUs on the system) as:
>>
>> llcc_pmu/read_miss/                                [Kernel PMU event]
>>
>> but the .read callback is never called when I run:
>>
>> sudo perf stat -C 0 -a -e llcc_pmu/read_miss/ stress-ng -C 8 -c 8 -m 10
>>
>> which always returns 0
>>
>> if I add --always-kernel I get:
>> <not supported>      llcc_pmu/read_miss/
> 
> Which SOC you are trying this on?
8250

Konrad
Trilok Soni Aug. 25, 2023, 5:07 p.m. UTC | #3
On 8/25/2023 4:50 AM, Konrad Dybcio wrote:
> On 24.08.2023 23:31, Trilok Soni wrote:
>> On 8/9/2023 1:09 PM, Konrad Dybcio wrote:
>>> Add support for the Qualcomm LLCC (Last-Level Cache Controller) PMU,
>>> which provides a single event, expressing cache read misses.
>>>
>>> Based on the vendor driver found in the msm-5.10 downstream kernel.
>>>
>>> Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
>>> ---
>>> Hi, I've been trying to get this driver going upstream by cleaning it
>>> up and adding the necessary perf boilerplate (the original Qualcomm one
>>> only pokes at the PMU from within the kernel itself) to use the
>>> userspace tool.
>>>
>>> I can not however get it to cooperate.. in this iteration I get a PMU
>>> event registered (though with only a "raw" name - no "x OR y" like with
>>> other PMUs on the system) as:
>>>
>>> llcc_pmu/read_miss/                                [Kernel PMU event]
>>>
>>> but the .read callback is never called when I run:
>>>
>>> sudo perf stat -C 0 -a -e llcc_pmu/read_miss/ stress-ng -C 8 -c 8 -m 10
>>>
>>> which always returns 0
>>>
>>> if I add --always-kernel I get:
>>> <not supported>      llcc_pmu/read_miss/
>>
>> Which SOC you are trying this on?
> 8250

Thanks. Let me see if my team can try this on latest SOCs and if it is 
the same behavior. Did you tried reading the counter by "printk" in the 
kernel and see the values are dumped from the register?
Konrad Dybcio Aug. 25, 2023, 5:09 p.m. UTC | #4
On 25.08.2023 19:07, Trilok Soni wrote:
> On 8/25/2023 4:50 AM, Konrad Dybcio wrote:
>> On 24.08.2023 23:31, Trilok Soni wrote:
>>> On 8/9/2023 1:09 PM, Konrad Dybcio wrote:
>>>> Add support for the Qualcomm LLCC (Last-Level Cache Controller) PMU,
>>>> which provides a single event, expressing cache read misses.
>>>>
>>>> Based on the vendor driver found in the msm-5.10 downstream kernel.
>>>>
>>>> Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
>>>> ---
>>>> Hi, I've been trying to get this driver going upstream by cleaning it
>>>> up and adding the necessary perf boilerplate (the original Qualcomm one
>>>> only pokes at the PMU from within the kernel itself) to use the
>>>> userspace tool.
>>>>
>>>> I can not however get it to cooperate.. in this iteration I get a PMU
>>>> event registered (though with only a "raw" name - no "x OR y" like with
>>>> other PMUs on the system) as:
>>>>
>>>> llcc_pmu/read_miss/                                [Kernel PMU event]
>>>>
>>>> but the .read callback is never called when I run:
>>>>
>>>> sudo perf stat -C 0 -a -e llcc_pmu/read_miss/ stress-ng -C 8 -c 8 -m 10
>>>>
>>>> which always returns 0
>>>>
>>>> if I add --always-kernel I get:
>>>> <not supported>      llcc_pmu/read_miss/
>>>
>>> Which SOC you are trying this on?
>> 8250
> 
> Thanks. Let me see if my team can try this on latest SOCs and if it is the same behavior. Did you tried reading the counter by "printk" in the kernel and see the values are dumped from the register?
Thanks for looking into this.

I debugbombed all the interesting functions and somehow the ones
which interface with the hardware are not even called as part of
executing `perf [...]` :/

Konrad
diff mbox series

Patch

diff --git a/drivers/perf/Kconfig b/drivers/perf/Kconfig
index 273d67ecf6d2..31d848c88d8a 100644
--- a/drivers/perf/Kconfig
+++ b/drivers/perf/Kconfig
@@ -155,6 +155,14 @@  config QCOM_L3_PMU
 	   Adds the L3 cache PMU into the perf events subsystem for
 	   monitoring L3 cache events.
 
+config QCOM_LLCC_PMU
+	tristate "Qualcomm Technologies LLCC PMU"
+	depends on ARCH_QCOM || COMPILE_TEST
+	depends on OF
+	help
+	  Support for the last-level cache performance monitor unit found
+	  on some Qualcomm Snapdragon SoCs.
+
 config THUNDERX2_PMU
 	tristate "Cavium ThunderX2 SoC PMU UNCORE"
 	depends on ARCH_THUNDER2 || COMPILE_TEST
diff --git a/drivers/perf/Makefile b/drivers/perf/Makefile
index 16b3ec4db916..eb02574780b5 100644
--- a/drivers/perf/Makefile
+++ b/drivers/perf/Makefile
@@ -12,6 +12,7 @@  obj-$(CONFIG_FSL_IMX9_DDR_PMU) += fsl_imx9_ddr_perf.o
 obj-$(CONFIG_HISI_PMU) += hisilicon/
 obj-$(CONFIG_QCOM_L2_PMU)	+= qcom_l2_pmu.o
 obj-$(CONFIG_QCOM_L3_PMU) += qcom_l3_pmu.o
+obj-$(CONFIG_QCOM_LLCC_PMU) += qcom_llcc_pmu.o
 obj-$(CONFIG_RISCV_PMU) += riscv_pmu.o
 obj-$(CONFIG_RISCV_PMU_LEGACY) += riscv_pmu_legacy.o
 obj-$(CONFIG_RISCV_PMU_SBI) += riscv_pmu_sbi.o
diff --git a/drivers/perf/qcom_llcc_pmu.c b/drivers/perf/qcom_llcc_pmu.c
new file mode 100644
index 000000000000..db290ae141a7
--- /dev/null
+++ b/drivers/perf/qcom_llcc_pmu.c
@@ -0,0 +1,277 @@ 
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (c) 2017-2020, The Linux Foundation. All rights reserved.
+ * Copyright (c) 2023, Linaro Limited
+ */
+
+#include <linux/module.h>
+#include <linux/of_device.h>
+#include <linux/perf_event.h>
+#include <linux/platform_device.h>
+#include <linux/spinlock.h>
+
+struct llcc_pmu {
+	struct pmu		pmu;
+	struct hlist_node	node;
+	void __iomem		*base;
+	struct perf_event	event;
+	raw_spinlock_t		lock;
+	u64			*llcc_stats;
+};
+#define to_llcc_pmu(p) (container_of(p, struct llcc_pmu, pmu))
+
+#define LLCC_READ_MISS_EV 	0x1000
+
+#define CNT_SCALING_FACTOR	0x3
+
+#define MAX_NUM_CPUS		16
+
+#define MON_CFG(m)		((m)->base + 0x200)
+ #define MON_CFG_ENABLE(cpu)	BIT(cpu)
+ #define MON_CFG_CLEARn(cpu)	BIT(16 + cpu)
+
+#define MON_CNT(m)		((m)->base + 0x220)
+ #define MON_CNT_VAL		GENMASK(23, 0)
+#define MON_CNTn(m, cpu)	(MON_CNT(m) + 0x4 * cpu)
+
+static DEFINE_PER_CPU(unsigned int, users_alive);
+
+static void mon_enable(struct llcc_pmu *llcc_pmu, int cpu)
+{
+	u32 val;
+
+	val = readl_relaxed(MON_CFG(llcc_pmu));
+	val |= MON_CFG_ENABLE(cpu);
+	writel_relaxed(val, MON_CFG(llcc_pmu));
+}
+
+static void mon_disable(struct llcc_pmu *llcc_pmu, int cpu)
+{
+	u32 val;
+
+	val = readl_relaxed(MON_CFG(llcc_pmu));
+	val &= ~MON_CFG_ENABLE(cpu);
+	writel_relaxed(val, MON_CFG(llcc_pmu));
+}
+
+static void mon_clear(struct llcc_pmu *llcc_pmu, int cpu)
+{
+	u32 val;
+
+	val = readl_relaxed(MON_CFG(llcc_pmu));
+
+	val |= MON_CFG_CLEARn(cpu);
+	writel_relaxed(val, MON_CFG(llcc_pmu));
+
+	val &= ~MON_CFG_CLEARn(cpu);
+	writel_relaxed(val, MON_CFG(llcc_pmu));
+}
+
+static int qcom_llcc_event_init(struct perf_event *event)
+{
+	struct llcc_pmu *llcc_pmu = to_llcc_pmu(event->pmu);
+
+	if (event->attr.type != event->pmu->type)
+		return -ENOENT;
+
+	if (event->attach_state & PERF_ATTACH_TASK)
+		return -EINVAL;
+
+	if (is_sampling_event(event)) {
+		dev_dbg(llcc_pmu->pmu.dev, "Per-task counters are unsupported\n");
+		return -EOPNOTSUPP;
+	}
+
+	if (has_branch_stack(event)) {
+		dev_dbg(llcc_pmu->pmu.dev, "Filtering is unsupported\n");
+		return -EINVAL;
+	}
+
+	if (event->cpu < 0) {
+		dev_warn(llcc_pmu->pmu.dev, "Can't provide per-task data!\n");
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
+static void qcom_llcc_event_read(struct perf_event *event)
+{
+	struct llcc_pmu *llcc_pmu = to_llcc_pmu(event->pmu);
+	unsigned long irq_flags;
+	int cpu = event->cpu;
+	u64 readout;
+
+	raw_spin_lock_irqsave(&llcc_pmu->lock, irq_flags);
+
+	mon_disable(llcc_pmu, cpu);
+
+	readout = FIELD_GET(MON_CNT_VAL, readl_relaxed(MON_CNTn(llcc_pmu, cpu)));
+	readout <<= CNT_SCALING_FACTOR;
+
+	llcc_pmu->llcc_stats[cpu] += readout;
+
+	mon_clear(llcc_pmu, cpu);
+	mon_enable(llcc_pmu, cpu);
+
+	if (!(event->hw.state & PERF_HES_STOPPED))
+		local64_set(&event->count, llcc_pmu->llcc_stats[cpu]);
+
+	raw_spin_unlock_irqrestore(&llcc_pmu->lock, irq_flags);
+}
+
+static void qcom_llcc_pmu_start(struct perf_event *event, int flags)
+{
+	if (flags & PERF_EF_RELOAD)
+		WARN_ON(!(event->hw.state & PERF_HES_UPTODATE));
+
+	event->hw.state = 0;
+}
+
+static void qcom_llcc_event_stop(struct perf_event *event, int flags)
+{
+	qcom_llcc_event_read(event);
+	event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
+}
+
+static int qcom_llcc_event_add(struct perf_event *event, int flags)
+{
+	struct llcc_pmu *llcc_pmu = to_llcc_pmu(event->pmu);
+	unsigned int cpu_users;
+
+	raw_spin_lock(&llcc_pmu->lock);
+
+	cpu_users = per_cpu(users_alive, event->cpu);
+	if (!cpu_users)
+		mon_enable(llcc_pmu, event->cpu);
+
+	cpu_users++;
+	per_cpu(users_alive, event->cpu) = cpu_users;
+
+	raw_spin_unlock(&llcc_pmu->lock);
+
+	event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
+
+	if (flags & PERF_EF_START)
+		qcom_llcc_pmu_start(event, PERF_EF_RELOAD);
+
+	return 0;
+}
+
+static void qcom_llcc_event_del(struct perf_event *event, int flags)
+{
+	struct llcc_pmu *llcc_pmu = to_llcc_pmu(event->pmu);
+	unsigned int cpu_users;
+
+	raw_spin_lock(&llcc_pmu->lock);
+
+	cpu_users = per_cpu(users_alive, event->cpu);
+	cpu_users--;
+	if (!cpu_users)
+		mon_disable(llcc_pmu, event->cpu);
+
+	per_cpu(users_alive, event->cpu) = cpu_users;
+
+	raw_spin_unlock(&llcc_pmu->lock);
+}
+
+static ssize_t llcc_pmu_event_show(struct device *dev, struct device_attribute *attr, char *page)
+{
+	struct perf_pmu_events_attr *pmu_attr;
+
+	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
+
+	return sysfs_emit(page, "event=0x%04llx\n", pmu_attr->id);
+}
+
+static struct attribute *qcom_llcc_pmu_events[] = {
+	PMU_EVENT_ATTR_ID(read_miss, llcc_pmu_event_show, LLCC_READ_MISS_EV),
+	NULL,
+};
+
+static const struct attribute_group qcom_llcc_pmu_events_group = {
+	.name = "events",
+	.attrs = qcom_llcc_pmu_events,
+};
+
+PMU_FORMAT_ATTR(event, "config:0-15");
+static struct attribute *qcom_llcc_pmu_format_attrs[] = {
+	&format_attr_event.attr,
+	NULL,
+};
+
+static const struct attribute_group qcom_llcc_pmu_format_group = {
+	.name = "format",
+	.attrs = qcom_llcc_pmu_format_attrs,
+};
+
+static const struct attribute_group *qcom_llcc_pmu_attr_groups[] = {
+	&qcom_llcc_pmu_format_group,
+	&qcom_llcc_pmu_events_group,
+	NULL,
+};
+
+static int qcom_llcc_pmu_probe(struct platform_device *pdev)
+{
+	static struct llcc_pmu *llcc_pmu;
+	int ret;
+
+	if (num_possible_cpus() > MAX_NUM_CPUS)
+		return dev_err_probe(&pdev->dev, -EINVAL,
+				     "LLCC PMU only supports <=%u CPUs\n",
+				     MAX_NUM_CPUS);
+
+	llcc_pmu = devm_kzalloc(&pdev->dev, sizeof(*llcc_pmu), GFP_KERNEL);
+	if (!llcc_pmu)
+		return -ENOMEM;
+
+	llcc_pmu->llcc_stats = devm_kcalloc(&pdev->dev, num_possible_cpus(),
+					    sizeof(*llcc_pmu->llcc_stats), GFP_KERNEL);
+
+	llcc_pmu->base = devm_platform_ioremap_resource(pdev, 0);
+	if (IS_ERR(llcc_pmu->base))
+		return dev_err_probe(&pdev->dev, PTR_ERR(llcc_pmu->base),
+				     "Failed to register LLCC PMU\n");
+
+	llcc_pmu->pmu = (struct pmu) {
+		.event_init	= qcom_llcc_event_init,
+		.add		= qcom_llcc_event_add,
+		.del		= qcom_llcc_event_del,
+		.start		= qcom_llcc_pmu_start,
+		.stop		= qcom_llcc_event_stop,
+		.read		= qcom_llcc_event_read,
+
+		.attr_groups	= qcom_llcc_pmu_attr_groups,
+		.capabilities	= PERF_PMU_CAP_NO_EXCLUDE,
+		.task_ctx_nr	= perf_invalid_context,
+
+		.module		= THIS_MODULE,
+	};
+
+	raw_spin_lock_init(&llcc_pmu->lock);
+
+	ret = perf_pmu_register(&llcc_pmu->pmu, "llcc_pmu", -1);
+	if (ret)
+		return dev_err_probe(&pdev->dev, ret, "Failed to register LLCC PMU\n");
+
+	return 0;
+}
+
+static const struct of_device_id qcom_llcc_pmu_match_table[] = {
+	{ .compatible = "qcom,llcc-pmu-v2" },
+	{ }
+};
+
+static struct platform_driver qcom_llcc_pmu_driver = {
+	.probe = qcom_llcc_pmu_probe,
+	.driver = {
+		.name = "qcom-llcc-pmu",
+		.of_match_table = qcom_llcc_pmu_match_table,
+		.suppress_bind_attrs = true,
+	},
+};
+module_platform_driver(qcom_llcc_pmu_driver);
+
+MODULE_DEVICE_TABLE(of, qcom_llcc_pmu_match_table);
+MODULE_DESCRIPTION("QCOM LLCC PMU");
+MODULE_LICENSE("GPL");