Hi Sunil,
On Tue, Aug 12, 2025 at 10:06 PM Sunil V L <sunilvl@xxxxxxxxxxxxxxxx> wrote:
>
> On Tue, Aug 12, 2025 at 09:32:10PM +0800, yunhui cui wrote:
> > Hi Sunil,
> >
> >
> > On Tue, Aug 12, 2025 at 9:15 PM Sunil V L <sunilvl@xxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Tue, Aug 12, 2025 at 07:25:44PM +0800, yunhui cui wrote:
> > > > Hi Sunil,
> > > >
> > > > On Thu, May 15, 2025 at 5:44 PM Yunhui Cui <cuiyunhui@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > Add the read of CSR_CYCLE to cppc_ffh_csr_read() to fix the
> > > > > warning message: "CPPC Cpufreq: cppc_scale_freq_wokrfn: failed
> > > > > to read perf counters".
> > > > >
> > > > > Signed-off-by: Yunhui Cui <cuiyunhui@xxxxxxxxxxxxx>
> > > > > ---
> > > > > drivers/acpi/riscv/cppc.c | 5 ++++-
> > > > > 1 file changed, 4 insertions(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/drivers/acpi/riscv/cppc.c b/drivers/acpi/riscv/cppc.c
> > > > > index 4cdff387deff6..c1acaeb18eac3 100644
> > > > > --- a/drivers/acpi/riscv/cppc.c
> > > > > +++ b/drivers/acpi/riscv/cppc.c
> > > > > @@ -69,11 +69,14 @@ static void cppc_ffh_csr_read(void *read_data)
> > > > > struct sbi_cppc_data *data = (struct sbi_cppc_data *)read_data;
> > > > >
> > > > > switch (data->reg) {
> > > > > - /* Support only TIME CSR for now */
> > > > > case CSR_TIME:
> > > > > data->ret.value = csr_read(CSR_TIME);
> > > > > data->ret.error = 0;
> > > > > break;
> > > > > + case CSR_CYCLE:
> > > > > + data->ret.value = csr_read(CSR_CYCLE);
> > > > > + data->ret.error = 0;
> > > > > + break;
> > > > > default:
> > > > > data->ret.error = -EINVAL;
> > > > > break;
> > > > > --
> > > > > 2.39.2
> > > > >
> > > >
> > > > The purpose of cppc_ffh_csr_read() is to calculate the actual
> > > > frequency of the CPU, which is delta_CSR_CYCLE/delta_CSR_XXX.
> > > >
> > > > CSR_XXX should be a reference clock and does not count during WFI
> > > > (Wait For Interrupt).
> > > >
> > > > Similar solutions include: x86's aperf/mperf, and ARM64's AMU with
> > > > registers SYS_AMEVCNTR0_CORE_EL0/SYS_AMEVCNTR0_CONST_EL0.
> > > >
> > > > However, we know that CSR_TIME in the current code does count during
> > > > WFI. So, is this design unreasonable?
> > > >
> > > > Should we consider proposing an extension to support such a dedicated
> > > > counter (a reference clock that does not count during WFI)? This way,
> > > > the value can be obtained directly in S-mode without trapping to
> > > > M-mode, especially since reading this counter is very frequent.
> > > >
> > > Hi Yunhui,
> > >
> > > Yes, but we anticipated that vendors might define their own custom CSRs.
> > > So, we introduced FFH encoding to accommodate such cases.
> > >
> > > Thanks,
> > > Sunil
> >
> > As mentioned earlier, it is best to directly read CSR_XXX (a reference
> > clock that does not count during WFI) and CSR_CYCLE in S-mode, rather
> > than trapping to SBI.
> >
> No. I meant direct CSR access itself not SBI. Please take a look at
> Table 6 of RISC-V FFH spec.
>
> > drivers/acpi/riscv/cppc.c is a generic driver that is not specific to
> > any vendor. Currently, the upstream code already uses CSR_TIME, and
> > the logic of CSR_TIME is incorrect.
> >
> CSR_TIME is just an example. It is upto the vendor how _CPC objects are
> encoded using FFH. The linux code doesn't mean one should use CSR_TIME
> always.
First, the example of CSR_TIME is incorrect. What is needed is a
CSR_XXX (a reference clock that does not count during WFI).
Second, you mentioned that each vendor can customize their own
implementations. But should all vendors' CSR_XXX/YYY/... be added to
drivers/acpi/riscv/cppc.c? Shouldn’t drivers/acpi/riscv/cppc.c fall
under the scope defined by the RISC-V architecture?
>
> > It would be best to promote a specification to support CSR_XXX, just
> > like what has been done for x86 and arm64. What do you think?
> >
> Wouldn't above work? For a standard extension, you may have to provide
> more data with actual HW.
This won’t work. May I ask how the current upstream code can calculate
the actual CPU frequency using CSR_TIME without trapping to SBI?
This is a theoretical logical issue. Why is data needed here?
Could you take a look at the "AMU events and event numbers" chapter in
the ARM64 manual?
>
> Thanks,
> Sunil
Thanks,
Yunhui
