[External] Re: [RFC 0/2] riscv: Idle thread using Zawrs extension

[Christoph Müllner]

On Thu, Apr 18, 2024 at 2:44 PM Xu Lu <luxu.kernel at bytedance.com> wrote:
>
> On Thu, Apr 18, 2024 at 8:26 PM Christoph Müllner
> <christoph.muellner at vrull.eu> wrote:
> >
> > On Thu, Apr 18, 2024 at 1:50 PM Xu Lu <luxu.kernel at bytedance.com> wrote:
> > >
> > > This patch series introduces a new implementation of idle thread using
> > > Zawrs extension.
> >
> > This overlaps with the following series:
> >   https://lore.kernel.org/all/20240315134009.580167-7-ajones@ventanamicro.com/
>
> Hi Christoph.
> Thanks for your reply!
> Actually our patch series is different from this. The work from your
> link focuses on providing support for Zawrs and implementing spinlock
> using it, while our work focuses on implementing idle thread using
> Zawrs and accelerating IPI to idle cpu. Of course, the ISA ZAWRS
> config part can be merged. We will refine our code in the next version
> to reduce code conflicts.

Yes, I've seen that this targets another optimization, but the basic
Zawrs support
would be identical to the other patchset (even if it is not).
I would propose that we work on a basic Zawrs support patchset that introduces
the Kconfig, DTS and hwprobe parts (a subset of Andrew's patchset).
Once this is merged, all other optimizations can be built upon it
(spinlocks, idle thread, glibc CPU spinning).
If this proposal is fine for the maintainers/reviewers, then Andrew could resend
these basic-support patches.

BR
Christoph


>
> >
> > BR
> > Christoph
> >
> > >
> > > The Zawrs[0] extension introduces two new instructions named WRS.STO and
> > > WRS.NTO in RISC-V. When software registers a reservation set using LR
> > > instruction, a subsequent WRS.STO or WRS.NTO instruction will cause the
> > > hart to stall in a low-power state until a store happens to the
> > > reservation set or an interrupt becomes pending. The difference between
> > > these two instructions is that WRS.STO will terminate stall after an
> > > implementation-defined timeout while WRS.NTO won't.
> > >
> > > This patch series implements idle thread using WRS.NTO instruction.
> > > Besides, we found there is no need to send a real IPI to wake up an idle
> > > CPU. Instead, we write IPI information to the reservation set of an idle
> > > CPU to wake it up and let it handle IPI quickly, without going through
> > > tranditional interrupt handling routine.
> > >
> > > [0] https://github.com/riscv/riscv-zawrs/blob/main/zawrs.adoc
> > >
> > > Xu Lu (2):
> > >   riscv: process: Introduce idle thread using Zawrs extension
> > >   riscv: Use Zawrs to accelerate IPI to idle cpu
> > >
> > >  arch/riscv/Kconfig                 |  24 +++++++
> > >  arch/riscv/include/asm/cpuidle.h   |  11 +---
> > >  arch/riscv/include/asm/hwcap.h     |   1 +
> > >  arch/riscv/include/asm/processor.h |  31 +++++++++
> > >  arch/riscv/include/asm/smp.h       |  14 ++++
> > >  arch/riscv/kernel/cpu.c            |   5 ++
> > >  arch/riscv/kernel/cpufeature.c     |   1 +
> > >  arch/riscv/kernel/process.c        | 102 ++++++++++++++++++++++++++++-
> > >  arch/riscv/kernel/smp.c            |  39 +++++++----
> > >  9 files changed, 205 insertions(+), 23 deletions(-)
> > >
> > > --
> > > 2.20.1
> > >
> > >
> > > _______________________________________________
> > > linux-riscv mailing list
> > > linux-riscv at lists.infradead.org
> > > http://lists.infradead.org/mailman/listinfo/linux-riscv