[PATCH] RISC-V: Do not issue remote fences until smp is available

[Palmer Dabbelt]

On Mon, 31 Oct 2022 12:26:05 PDT (-0700), atishp at atishpatra.org wrote:
> On Mon, Oct 31, 2022 at 12:12 PM Conor Dooley <conor at kernel.org> wrote:
>>
>> On Fri, Oct 28, 2022 at 04:19:29PM -0700, Atish Patra wrote:
>> > It is useless to issue remote fences if there is a single core
>> > available. It becomes a bottleneck for sbi based rfences where
>> > we will be making those ECALLs for no reason. Early code patching
>> > because of static calls end up in this path.
>> >
>> > Signed-off-by: Atish Patra <atishp at rivosinc.com>
>>
>> Hey Atish,
>> This doesn't apply for me to either fixes or for-next. What branch does
>> it apply to?
>> Thanks,
>> Conor.
>>
>> > ---
>> >  arch/riscv/mm/cacheflush.c | 4 ++++
>> >  1 file changed, 4 insertions(+)
>> >
>> > diff --git a/arch/riscv/mm/cacheflush.c b/arch/riscv/mm/cacheflush.c
>> > index f10cb47eac3a..7fafc8c26505 100644
>> > --- a/arch/riscv/mm/cacheflush.c
>> > +++ b/arch/riscv/mm/cacheflush.c
>> > @@ -19,6 +19,10 @@ void flush_icache_all(void)
>> >  {
>> >       local_flush_icache_all();
>> >
>> > +     /* No need to issue remote fence if only 1 cpu is online */
>> > +     if (num_online_cpus() == 1)
>> > +             return;
>> > +
>> >       if (IS_ENABLED(CONFIG_RISCV_SBI) && !riscv_use_ipi_for_rfence())
>> >               sbi_remote_fence_i(NULL);
>> >       else
>> > --
>> > 2.34.1
>> >
>
> Sorry I forgot to specify the dependencies for this patch. This patch
> is based on Anup's IPI series [1] as
> I assumed the IPI series would go first. I can rebase on top of the
> master if required.
> However, the issue will manifest only after Jisheng's patch[2] which
> moved the sbi_init to earlier and introduced the
> static key in the paging_init path.
>
> [1] https://patchwork.kernel.org/project/linux-riscv/patch/20220820065446.389788-8-apatel@ventanamicro.com/
> [2] https://lore.kernel.org/lkml/20220716115059.3509-1-jszhang@kernel.org/

IMO we should just stop issuing the SBI remote fences at all, with the 
code to do IPI-based fences we're just adding complexity for the slow 
case.