Re: [PATCH v5 1/2] RISC-V: Cache SBI vendor values

From: Anup Patel
Date: Mon Oct 10 2022 - 09:14:44 EST


On Mon, Oct 10, 2022 at 6:25 PM Heiko Stuebner <heiko@xxxxxxxxx> wrote:
>
> Am Montag, 10. Oktober 2022, 14:45:45 CEST schrieb Anup Patel:
> > On Mon, Oct 10, 2022 at 5:57 PM Heiko Stuebner <heiko@xxxxxxxxx> wrote:
> > >
> > > sbi_get_mvendorid(), sbi_get_marchid() and sbi_get_mimpid() might get
> > > called multiple times, though the values of these CSRs should not change
> > > during the runtime of a specific machine.
> > >
> > > So cache the values in the functions and prevent multiple ecalls
> > > to read these values.
> > >
> > > As Andrew Jones noted, at least marchid and mimpid may be negative
> > > values when viewed as a long, so we use a separate static bool to
> > > indiciate the cached status.
> > >
> > > Suggested-by: Atish Patra <atishp@xxxxxxxxxxxxxx>
> > > Signed-off-by: Heiko Stuebner <heiko@xxxxxxxxx>
> > > ---
> > > arch/riscv/kernel/sbi.c | 30 +++++++++++++++++++++++++++---
> > > 1 file changed, 27 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/arch/riscv/kernel/sbi.c b/arch/riscv/kernel/sbi.c
> > > index 775d3322b422..cc618aaa9d11 100644
> > > --- a/arch/riscv/kernel/sbi.c
> > > +++ b/arch/riscv/kernel/sbi.c
> > > @@ -625,17 +625,41 @@ static inline long sbi_get_firmware_version(void)
> > >
> > > long sbi_get_mvendorid(void)
> > > {
> > > - return __sbi_base_ecall(SBI_EXT_BASE_GET_MVENDORID);
> > > + static long id;
> > > + static bool cached;
> > > +
> > > + if (!cached) {
> > > + id = __sbi_base_ecall(SBI_EXT_BASE_GET_MVENDORID);
> > > + cached = true;
> > > + }
> > > +
> > > + return id;
> > > }
> > >
> > > long sbi_get_marchid(void)
> > > {
> > > - return __sbi_base_ecall(SBI_EXT_BASE_GET_MARCHID);
> > > + static long id;
> > > + static bool cached;
> >
> > This breaks for heterogeneous SMP systems (similar to big.LITTLE)
> > where HARTs will have different marchid even though they belong to
> > same CPU Vendor.
> >
> > Due to the above rationale, the patch adding marchid, mvendorid, and
> > mimpid in /proc/cpuinfo caches these values on a per-CPU basis.
>
> For people reading along, I think you mean
> https://lore.kernel.org/r/20220727043829.151794-1-apatel@xxxxxxxxxxxxxxxx

Yes, this is the patch I am referring.

>
> For my understanding, was there a reason in the past for doing the caching
> only for cpuinfo and not for every invocation of the ecalls?

The caching was done only for /proc/cpuinfo because at that point
time only "cat /proc/cpuinfo" would need these value at runtime.

Now that we have more uses of marchid, mvendorid, and mimpid,
at runtime, we should definitely have the cached values available
to other parts of kernel.

My suggestion is to keep sbi_get_xyz() functions unmodified and
instead add new functions in arch/riscv/kernel/cpu.c (on-top-of
/proc/cpuinfo patch) which allows users to read cached values of
any CPU.

For example, we could export following functions from
arch/riscv/kernel/cpu.c:
unsigned long riscv_cached_mvendorid(unsigned int cpu);
unsigned long riscv_cached_marchid(unsigned int cpu);
unsigned long riscv_cached_mimpid(unsigned int cpu);

Regards,
Anup