Re: [PATCH] drivers/perf: riscv_pmu_sbi: add support for PMU variant on T-Head C9xx cores

[Anup Patel]

On Thu, Aug 25, 2022 at 7:34 AM Heiko Stuebner <heiko@xxxxxxxxx> wrote:
>
> Am Donnerstag, 18. August 2022, 10:24:33 CEST schrieb Anup Patel:
> > On Thu, Aug 18, 2022 at 1:03 AM Atish Patra <atishp@xxxxxxxxxxxxxx> wrote:
> > >
> > > On Wed, Aug 17, 2022 at 4:13 AM Heiko Stuebner <heiko@xxxxxxxxx> wrote:
> > > >
> > > > With the T-HEAD C9XX cores being designed before or during the ratification
> > > > to the SSCOFPMF extension, they implement functionality very similar but
> > > > not equal to it. So add some adaptions to allow the C9XX to still handle
> > > > its PMU through the regular SBI PMU interface instead of defining new
> > > > interfaces or drivers.
> > > >
> > >
> > > IMO, vendor specific workarounds in the generic implementation is not
> > > a good idea.
> > > If we have to support it, I think we should just put the IRQ number in
> > > the DT and parse from the DT.
> > > The initial sscofpmf series was based on the DT. It was removed later
> > > as there was no need for it at that time.
> > > We can always revive it.
> > >
> > > Regarding the CSR number difference and static key enablement, can we
> > > use code patching techniques here as well ?
> > > At least all the T-HEAD C9XX core erratas are in one place.
> > >
> > > The alternate would be just to say T-HEAD C9XX support SSCOFPMF but
> > > with erratas. I don't prefer this approach
> > > but it keeps the vendor specific checks out of the generic code.
> >
> > Whether to have a DT node (or not) was already discussed and concluded
> > in the past.
> >
> > We don't need a DT node just to get the IRQ number. The T-HEAD custom
> > IRQ number can be derived based on mvendorid.
>
> Yeah, I remember reading that discussion and thus went with the mvendorid
> way in this patch.
>
> > Also, all these T-HEAD specific changes in SBI PMU driver should be
> > implemented as erratas using ALTERNATIVE() macros.
>
> (1) "All these T-HEAD specific changes ..."
> Actually the only T-HEAD-specific change is reading that different CSR
> for the overflow information, the rest only makes the irq-number variable

If it is just overflow CSR then it is simpler to do instruction patching
in drivers/perf/riscv_pmu_sbi.c itself.

>
> (2) ALTERNATIVE macros are working on assembler instructions, so are you
> sugesting to replace the generic csr_read() for the overflow-csr with a
> custom copy like
>
> #define sbi_pmu_read_overflow(void)                                             \
> ({                                                              \
>         register unsigned long __v;                             \
>         ALT_THEAD_PMU_OVERFLOW(__v); \
>         __v;                                                    \
> })
>
> and then in errata_list.h
>
> #define ALT_THEAD_PMU_OVERFLOW(__ovl) \
> __asm__ __volatile__ (alternative(
>         "csrr %0, " __ASM_STR(CSR_SSCOUNTOVF),  \
>         "csrr %0, " __ASM_STR(THEAD_C9XX_CSR_SCOUNTEROF), THEAD_VENDOR_ID, \
>                 ERRATA_THEAD_PMU, CONFIG_ERRATA_THEAD_PMU) \
>                               : "=r" (__ovl) :                  \
>                               : "memory");
>
> I'm not yet seeing what you're gaining by going with this approach:
> - that the overflow-csr is the same bitwise but only at a different
>   address is specific to the c9xx, other deviants might implement things
>   completely different
> - you're not getting rid of the thead mention
> - and we're now duplicating the generic csr-read code
>
> Is this the preferred way or what am I overlooking?

Yes, better to have special sbi_pmu_read_overflow() in
drivers/perf/riscv_pmu_sbi.c itself which is based on ALTERNATIVES.

I am suggesting ALTERNATIVEs only because overflow CSR is
accessed in the interrupt handler which is in hot-path when we run
"perf record".

Regards,
Anup