Hi Jessica,
On Fri, Aug 15, 2025 at 12:57 AM Jessica Clarke <jrtc27@xxxxxxxxxx> wrote:
>
> On 14 Aug 2025, at 14:37, Anup Patel <apatel@xxxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, Aug 14, 2025 at 11:49 AM yunhui cui <cuiyunhui@xxxxxxxxxxxxx> wrote:
> >>
> >> Hi Anup,
> >>
> >> On Thu, Aug 14, 2025 at 1:48 PM Anup Patel <apatel@xxxxxxxxxxxxxxxx> wrote:
> >>>
> >>> On Thu, Aug 14, 2025 at 9:08 AM yunhui cui <cuiyunhui@xxxxxxxxxxxxx> wrote:
> >>>>
> >>>> Hi Anup,
> >>>>
> >>>> On Wed, Aug 13, 2025 at 7:12 PM Anup Patel <apatel@xxxxxxxxxxxxxxxx> wrote:
> >>>>>
> >>>>> On Wed, Aug 13, 2025 at 12:14 PM yunhui cui <cuiyunhui@xxxxxxxxxxxxx> wrote:
> >>>>>>
> >>>>>> Hi Sunil,
> >>>>>>
> >>>>>> On Wed, Aug 13, 2025 at 1:28 PM Sunil V L <sunilvl@xxxxxxxxxxxxxxxx> wrote:
> >>>>>>>
> >>>>>>> Hi Yunhui,
> >>>>>>>
> >>>>>>> On Wed, Aug 13, 2025 at 11:23:39AM +0800, yunhui cui wrote:
> >>>>>>>> Hi Sunil,
> >>>>>>>>
> >>>>>>>> On Tue, Aug 12, 2025 at 10:06 PM Sunil V L <sunilvl@xxxxxxxxxxxxxxxx> wrote:
> >>>>>>>>>
> >>>>>>> [...]
> >>>>>>>>>>>>
> >>>>>>>>>>>> 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.
> >>>>>>>>>>
> >>>>>>> ACPI spec for "Reference Performance Register" says,
> >>>>>>>
> >>>>>>> "The Reference Performance Counter Register counts at a fixed rate any
> >>>>>>> time the processor is active. It is not affected by changes to Desired
> >>>>>>> Performance, processor throttling, etc."
> >>>>>>>
> >>>>>>>>> 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?
> >>>>>>>>
> >>>>>>> No. One can implement similar to csr_read_num() in opensbi. We didn't
> >>>>>>> add it since there was no HW implementing such thing. What I am
> >>>>>>> saying is we have FFH encoding to support such case.
> >>>>>>>
> >>>>>>>>>
> >>>>>>>>>> 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?
> >>>>>>>>
> >>>>>>> As I mentioned above, one can implement a generic CSR read without
> >>>>>>> trapping to SBI.
> >>>>>>>
> >>>>>>>> Could you take a look at the "AMU events and event numbers" chapter in
> >>>>>>>> the ARM64 manual?
> >>>>>>>>
> >>>>>>> As-per ACPI spec reference performance counter is not affected by CPU
> >>>>>>> state. The RISC-V FFH encoding is sufficiently generic to support this
> >>>>>>> requirement, even if the standard CSR_TIME cannot be used. In such
> >>>>>>> cases, an alternative CSR can be encodeded, accessed via an OS-level
> >>>>>>> abstraction such as csr_read_num().
> >>>>>>
> >>>>>> So what you're saying is that we should submit a patch like this, right?
> >>>>>>
> >>>>>> diff --git a/drivers/acpi/riscv/cppc.c b/drivers/acpi/riscv/cppc.c
> >>>>>> index 440cf9fb91aab..953c259d46c69 100644
> >>>>>> --- a/drivers/acpi/riscv/cppc.c
> >>>>>> +++ b/drivers/acpi/riscv/cppc.c
> >>>>>> @@ -66,16 +66,8 @@ 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;
> >>>>>> - default:
> >>>>>> - data->ret.error = -EINVAL;
> >>>>>> - break;
> >>>>>> - }
> >>>>>> + data->ret.value = csr_read_num(data->reg);
> >>>>>> + data->ret.error = 0;
> >>>>>> }
> >>>>>>
> >>>>>> If that's the case, the robustness of the code cannot be guaranteed,
> >>>>>> because the range of CSRs from different vendors is unknown.
> >>>>>
> >>>>> ACPI FFH is allows mapping to any CSR.
> >>>>
> >>>> Yes, FFH can map any CSR, and this is not the point of contention.
> >>>>
> >>>> If that's the case, the CSR_TIME used in the current kernel code is
> >>>> inappropriate. Some vendors may design a counter that does not count
> >>>> during WFI, making CSR_TIME irrelevant. Even if counting continues
> >>>> during WFI, are you planning to have one counter operate in S-mode
> >>>> while the other traps to M-mode?
> >>>>
> >>>> In that case, the code would need to be modified as proposed above. Do
> >>>> you agree?
> >>>
> >>> I disagree.
> >>>
> >>> Like Sunil already explained, if an implementation has reference counter
> >>> which does not count during WFI state then for such implementation the
> >>> delivered performance counter should also not increment during WFI
> >>> to maintain the relative delta of increments. This means if an implementation
> >>> uses TIME CSR as reference counter then for such implementation
> >>> the delivered performance counter should increment accordingly. Ultimately,
> >>> what matters is OS being able to correctly compute the performance level
> >>> using reference and delivered performance counters.
> >>
> >>
> >> For calculating the actual CPU frequency, both implementations are
> >> acceptable where either both counters continue counting during WFI or
> >> both stop counting.
> >> In the current code, how do you read the other counter?
> >> Shouldn't it be modified like this to support it? This way, all
> >> counters can be read directly in S-mode without trapping to M-mode:
> >> + data->ret.value = csr_read_num(data->reg);
> >> + data->ret.error = 0;
> >
> > Yes, the current switch-case needs to replaced by common
> > csr_read_num() and csr_write_num() implemented in arch/riscv/
> >
> > The RISC-V CSR space is limited so with it is straightforward
> > to implement csr_read_num() and csr_write_num() using
> > macros where each CSR access will turn-out to be roughly
> > 3-4 instructions.
>
> 12 bits, or 4096 CSRs. Are you saying you want to have a jump table
> that dispatches to one of 4096 entry points?
>
> Maybe you can cut that down a bit for S-mode based on the encoding
> convention, but that only eliminates 1/4, so you’re still looking at
> 3072 entry points, perhaps also minus the few that are allocated and
> clearly not sensible things to use for this, like stval.
>
> But I think that’s not a reasonable approach to take, and if there is
> no CSR in the current RISC-V spec that fits the needs of ACPI then one
> needs to be defined so that we don’t need every vendor to invent their
> own. If there is a CSR already then that should be the only one that’s
> allowed to be used here. If you look at arm64, it hard-codes which
> counter to use for each of the two calls it supports. That’s a much
> better world to be in.
Agreed. Because the second operand of the csrr instruction must be a
constant, a switch-case conversion is therefore necessary.
>
> Jessica
>
Thanks,
Yunhui
