On Thu, Jun 26, 2025 at 11:34:21AM +0200, Andrew Jones wrote:
> On Thu, Jun 26, 2025 at 11:21:10AM +0200, Andrew Jones wrote:
> > On Wed, Jun 25, 2025 at 04:21:01PM +0200, Aleksa Paunovic via B4 Relay wrote:
> > > From: Aleksa Paunovic <aleksa.paunovic@xxxxxxxxxxxxx>
> > >
> > > Use the hwprobe syscall to decide which PAUSE instruction to execute in
> > > userspace code.
> > >
> > > Signed-off-by: Aleksa Paunovic <aleksa.paunovic@xxxxxxxxxxxxx>
> > > ---
> > > tools/arch/riscv/include/asm/vdso/processor.h | 27 +++++++++++++++++----------
> > > 1 file changed, 17 insertions(+), 10 deletions(-)
> > >
> > > diff --git a/tools/arch/riscv/include/asm/vdso/processor.h b/tools/arch/riscv/include/asm/vdso/processor.h
> > > index 662aca03984817f9c69186658b19e9dad9e4771c..027219a486b7b93814888190f8224af29498707c 100644
> > > --- a/tools/arch/riscv/include/asm/vdso/processor.h
> > > +++ b/tools/arch/riscv/include/asm/vdso/processor.h
> > > @@ -4,26 +4,33 @@
> > >
> > > #ifndef __ASSEMBLY__
> > >
> > > +#include <asm/hwprobe.h>
> > > +#include <sys/hwprobe.h>
> > > +#include <asm/vendor/mips.h>
> > > #include <asm-generic/barrier.h>
> > >
> > > static inline void cpu_relax(void)
> > > {
> > > + struct riscv_hwprobe pair;
> > > + bool has_mipspause;
> > > #ifdef __riscv_muldiv
> > > int dummy;
> > > /* In lieu of a halt instruction, induce a long-latency stall. */
> > > __asm__ __volatile__ ("div %0, %0, zero" : "=r" (dummy));
> > > #endif
> > >
> > > -#ifdef CONFIG_TOOLCHAIN_HAS_ZIHINTPAUSE
> > > - /*
> > > - * Reduce instruction retirement.
> > > - * This assumes the PC changes.
> > > - */
> > > - __asm__ __volatile__ ("pause");
> > > -#else
> > > - /* Encoding of the pause instruction */
> > > - __asm__ __volatile__ (".4byte 0x100000F");
> > > -#endif
> > > + pair.key = RISCV_HWPROBE_KEY_VENDOR_EXT_MIPS_0;
> > > + __riscv_hwprobe(&pair, 1, 0, NULL, 0);
> > > + has_mipspause = pair.value & RISCV_HWPROBE_VENDOR_EXT_XMIPSEXECTL;
> > > +
> > > + if (has_mipspause) {
> > > + /* Encoding of the mips pause instruction */
> > > + __asm__ __volatile__(".4byte 0x00501013");
> > > + } else {
> > > + /* Encoding of the pause instruction */
> > > + __asm__ __volatile__(".4byte 0x100000F");
> > > + }
> > > +
> >
> > cpu_relax() is used in places where we cannot afford the overhead nor call
> > arbitrary functions which may take locks, etc. We've even had trouble
> > using a static key here in the past since this is inlined and it bloated
> > the size too much. You'll need to use ALTERNATIVE().
>
> Oh, I see now that the next patch is handling the kernel cpu_relax with
> ALTERNATIVE and this was just the tools cpu_relax. We don't want to make
> a syscall inside cpu_relax though either, since it gets called in loops.
(Another follow up to myself...)
I guess with the vdso cached result it should only be a handful of
instructions, but it still seems odd to embed a call in cpu_relax.
Thanks,
drew
> It'd be better to just call the standard pause (0x100000F) even if it
> does nothing. Or maybe there's some define that can be added/used to
> select the correct instruction?
>
> Thanks,
> drew
