Re: [PATCH v4 1/1] RISC-V: Create unique identification for SoC PMU
From: Nikita Shubin
Date: Tue Jun 21 2022 - 04:07:34 EST
Hello Anup!
On Tue, 21 Jun 2022 13:21:11 +0530
Anup Patel <apatel@xxxxxxxxxxxxxxxx> wrote:
> On Tue, Jun 21, 2022 at 1:13 PM Nikita Shubin
> <nikita.shubin@xxxxxxxxxxx> wrote:
> >
> > Hello Anup!
> >
> > On Mon, 20 Jun 2022 12:40:20 -0700
> > Atish Patra <atishp@xxxxxxxxxxxxxx> wrote:
> >
> > > On Mon, Jun 20, 2022 at 7:40 AM Nikita Shubin
> > > <nikita.shubin@xxxxxxxxxxx> wrote:
> > > >
> > > > On Mon, 20 Jun 2022 17:30:58 +0530
> > > > Anup Patel <apatel@xxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > > On Sun, Jun 19, 2022 at 4:41 PM Nikita Shubin
> > > > > <nikita.shubin@xxxxxxxxxxx> wrote:
> > > > > >
> > > > > > From: Nikita Shubin <n.shubin@xxxxxxxxx>
> > > > > >
> > > > > > Provide RISC-V SBI PMU id to distinguish different cores or
> > > > > > SoCs via "devices/platform/riscv-pmu/id" sysfs entry.
> > > > > >
> > > > > > The identification is generated as string of marchid,
> > > > > > mimpid, mvendorid in hex format separated by coma -
> > > > > > "0x70032,0x70032,0x0".
> > > > > >
> > > > > > The CSRs are detailed in the RISC-V privileged spec [1].
> > > > > > [1] https://github.com/riscv/riscv-isa-manual
> > > > > >
> > > > > > Inspired-by: João Mário Domingos
> > > > > > <joao.mario@xxxxxxxxxxxxxxxxxx> Signed-off-by: Nikita
> > > > > > Shubin <n.shubin@xxxxxxxxx>
> > > > >
> > > > > The mvendorid, marchid, and mimpid can be useful to apps other
> > > > > than perf tool.
> > > > >
> > > > > I have tried to extend /proc/cpuinfo with this information
> > > > > which can be parsed by perf tool:
> > > > > https://lore.kernel.org/all/20220620115549.1529597-1-apatel@xxxxxxxxxxxxxxxx/
> > > > >
> > > >
> > > > Atish, what do you think about this ?
> > > >
> > > > RISC-V perf can rely on "/proc/cpuinfo", in some similar manner
> > > > like "tools/perf/arch/s390/util/header.c" does.
> > > >
> > >
> > > Yes. We can expose these three values either in sysfs or procfs
> > > (/proc/cpuinfo). For perf tool, it shouldn't matter as the
> > > header.c will need to generate the unique cpuid
> > > string from either.
> > >
> > > I am not sure if any other userspace tool prefers to parse sysfs
> > > instead of cpuinfo.
> >
> > Okay - let's stick to /proc/cpuinfo.
>
> Sounds good.
>
> You might have to write /proc/cpuinfo parsing code in perf tool
> header.c. Do you plan to send v4 of perf tool patches ??
Yes, but i might split it into separate series, not to mix perf headers
and U74 pmu-events bindings, i ll also add some SBI firmware event
bindings to make things easier.
>
> Regards,
> Anup
>
> >
> > >
> > > > Can it create problems with pmu identification in case of
> > > > hetergenous harts ?
> > > >
> > >
> > > Does perf support hetergenous harts at all ? ARM64 code
> > > (tool/perf/arch/arm64/util/header.c)
> > > just breaks out of the loop after finding the first MIDR.
> > >
> > > > >
> > > > > Regards,
> > > > > Anup
> > > > >
> > > > > > ---
> > > > > > v3->v4:
> > > > > > - use string for pmuid
> > > > > > - rename pmu_sbi_id_show to id_show
> > > > > > - fix error print message in id_show
> > > > > > - fix DEVICE_ATTR to use octal permissions
> > > > > > ---
> > > > > > arch/riscv/kernel/sbi.c | 3 +++
> > > > > > drivers/perf/riscv_pmu_sbi.c | 41
> > > > > > ++++++++++++++++++++++++++++++++++
> > > > > > include/linux/perf/riscv_pmu.h | 1 + 3 files changed, 45
> > > > > > insertions(+)
> > > > > >
> > > > > > diff --git a/arch/riscv/kernel/sbi.c
> > > > > > b/arch/riscv/kernel/sbi.c index 775d3322b422..50dd9b6ecc9e
> > > > > > 100644 --- a/arch/riscv/kernel/sbi.c
> > > > > > +++ b/arch/riscv/kernel/sbi.c
> > > > > > @@ -627,16 +627,19 @@ long sbi_get_mvendorid(void)
> > > > > > {
> > > > > > return __sbi_base_ecall(SBI_EXT_BASE_GET_MVENDORID);
> > > > > > }
> > > > > > +EXPORT_SYMBOL(sbi_get_mvendorid);
> > > > > >
> > > > > > long sbi_get_marchid(void)
> > > > > > {
> > > > > > return __sbi_base_ecall(SBI_EXT_BASE_GET_MARCHID);
> > > > > > }
> > > > > > +EXPORT_SYMBOL(sbi_get_marchid);
> > > > > >
> > > > > > long sbi_get_mimpid(void)
> > > > > > {
> > > > > > return __sbi_base_ecall(SBI_EXT_BASE_GET_MIMPID);
> > > > > > }
> > > > > > +EXPORT_SYMBOL(sbi_get_mimpid);
> > > > > >
> > > > > > static void sbi_send_cpumask_ipi(const struct cpumask
> > > > > > *target) {
> > > > > > diff --git a/drivers/perf/riscv_pmu_sbi.c
> > > > > > b/drivers/perf/riscv_pmu_sbi.c index
> > > > > > dca3537a8dcc..be812f855617 100644 ---
> > > > > > a/drivers/perf/riscv_pmu_sbi.c +++
> > > > > > b/drivers/perf/riscv_pmu_sbi.c @@ -693,6 +693,28 @@ static
> > > > > > int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct
> > > > > > platform_device *pde return 0; }
> > > > > >
> > > > > > +static ssize_t id_show(struct device *dev,
> > > > > > + struct device_attribute
> > > > > > *attr, char *buf) +{
> > > > > > + int len;
> > > > > > + struct riscv_pmu *pmu =
> > > > > > container_of(dev_get_drvdata(dev), struct riscv_pmu, pmu); +
> > > > > > + len = sprintf(buf, "%s\n", pmu->pmuid);
> > > > > > + if (len <= 0)
> > > > > > + dev_err(dev, "invalid sprintf len: %d\n",
> > > > > > len); +
> > > > > > + return len;
> > > > > > +}
> > > > > > +
> > > > > > +static DEVICE_ATTR(id, 0644, id_show, NULL);
> > > > > > +
> > > > > > +static struct attribute *pmu_sbi_attrs[] = {
> > > > > > + &dev_attr_id.attr,
> > > > > > + NULL
> > > > > > +};
> > > > > > +
> > > > > > +ATTRIBUTE_GROUPS(pmu_sbi);
> > > > > > +
> > > > > > static int pmu_sbi_device_probe(struct platform_device
> > > > > > *pdev) {
> > > > > > struct riscv_pmu *pmu = NULL;
> > > > > > @@ -714,6 +736,14 @@ static int pmu_sbi_device_probe(struct
> > > > > > platform_device *pdev) if
> > > > > > (pmu_sbi_get_ctrinfo(num_counters)) goto out_free;
> > > > > >
> > > > > > + /* fill pmuid */
> > > > > > + pmu->pmuid = kasprintf(GFP_KERNEL,
> > > > > > "0x%lx,0x%lx,0x%lx",
> > > > > > + sbi_get_marchid(),
> > > > > > + sbi_get_mimpid(),
> > > > > > + sbi_get_mvendorid());
> > > > > > + if (!pmu->pmuid)
> > > > > > + goto out_free_pmuid;
> > > > > > +
> > > > > > ret = pmu_sbi_setup_irqs(pmu, pdev);
> > > > > > if (ret < 0) {
> > > > > > pr_info("Perf sampling/filtering is not
> > > > > > supported as sscof extension is not available\n"); @@ -739,8
> > > > > > +769,19 @@ static int pmu_sbi_device_probe(struct
> > > > > > platform_device *pdev) return ret; }
> > > > > >
> > > > > > + ret = sysfs_create_group(&pdev->dev.kobj,
> > > > > > &pmu_sbi_group);
> > > > > > + if (ret) {
> > > > > > + dev_err(&pdev->dev, "sysfs creation
> > > > > > failed\n");
> > > > > > + return ret;
> > > > > > + }
> > > > > > +
> > > > > > + pdev->dev.groups = pmu_sbi_groups;
> > > > > > + dev_set_drvdata(&pdev->dev, pmu);
> > > > > > +
> > > > > > return 0;
> > > > > >
> > > > > > +out_free_pmuid:
> > > > > > + kfree(pmu->pmuid);
> > > > > > out_free:
> > > > > > kfree(pmu);
> > > > > > return ret;
> > > > > > diff --git a/include/linux/perf/riscv_pmu.h
> > > > > > b/include/linux/perf/riscv_pmu.h index
> > > > > > 46f9b6fe306e..cf3557b77fb8 100644 ---
> > > > > > a/include/linux/perf/riscv_pmu.h +++
> > > > > > b/include/linux/perf/riscv_pmu.h @@ -42,6 +42,7 @@ struct
> > > > > > cpu_hw_events { struct riscv_pmu {
> > > > > > struct pmu pmu;
> > > > > > char *name;
> > > > > > + char *pmuid;
> > > > > >
> > > > > > irqreturn_t (*handle_irq)(int irq_num, void
> > > > > > *dev);
> > > > > >
> > > > > > --
> > > > > > 2.35.1
> > > > > >
> > > >
> > >
> > >
> >