Hey Minda,
On Wed, Aug 02, 2023 at 02:42:15PM +0800, Minda Chen wrote:
> Actually it is a part of Conor's
> commit aae538cd03bc ("riscv: fix detection of toolchain
> Zihintpause support").
> It is looks like a merge issue.
Yup, spot on.
> Samuel's
> commit 0b1d60d6dd9e ("riscv: Fix build with
> CONFIG_CC_OPTIMIZE_FOR_SIZE=y") do not base on Conor's commit and
> revert to __riscv_zihintpause. So this patch can fix it.
>
> Signed-off-by: Minda Chen <[email protected]>
Did you actually manage to trigger this, or was this by inspection?
clang-15 + binutils 2.35 was, IIRC, how we spotted this because that's
what the clang-built-linux CI uses to test the LTS kernels from before
LLVM's IAS was supported for RISC-V. Seemingly all that needs to be
satisfied there is that zihintpause doesn't appear in -march so this has
gone unnoticed.
Fixes: 3c349eacc559 ("Merge patch "riscv: Fix build with CONFIG_CC_OPTIMIZE_FOR_SIZE=y"")
Reviewed-by: Conor Dooley <[email protected]>
Thanks,
Conor.
> ---
> arch/riscv/include/asm/vdso/processor.h | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/arch/riscv/include/asm/vdso/processor.h b/arch/riscv/include/asm/vdso/processor.h
> index 14f5d27783b8..96b65a5396df 100644
> --- a/arch/riscv/include/asm/vdso/processor.h
> +++ b/arch/riscv/include/asm/vdso/processor.h
> @@ -14,7 +14,7 @@ static inline void cpu_relax(void)
> __asm__ __volatile__ ("div %0, %0, zero" : "=r" (dummy));
> #endif
>
> -#ifdef __riscv_zihintpause
> +#ifdef CONFIG_TOOLCHAIN_HAS_ZIHINTPAUSE
> /*
> * Reduce instruction retirement.
> * This assumes the PC changes.
> --
> 2.17.1
>
On 2023/8/2 14:54, Conor Dooley wrote:
> Hey Minda,
>
> On Wed, Aug 02, 2023 at 02:42:15PM +0800, Minda Chen wrote:
>> Actually it is a part of Conor's
>> commit aae538cd03bc ("riscv: fix detection of toolchain
>> Zihintpause support").
>> It is looks like a merge issue.
>
> Yup, spot on.
>
>> Samuel's
>> commit 0b1d60d6dd9e ("riscv: Fix build with
>> CONFIG_CC_OPTIMIZE_FOR_SIZE=y") do not base on Conor's commit and
>> revert to __riscv_zihintpause. So this patch can fix it.
>>
>> Signed-off-by: Minda Chen <[email protected]>
>
> Did you actually manage to trigger this, or was this by inspection?
> clang-15 + binutils 2.35 was, IIRC, how we spotted this because that's
> what the clang-built-linux CI uses to test the LTS kernels from before
> LLVM's IAS was supported for RISC-V. Seemingly all that needs to be
> satisfied there is that zihintpause doesn't appear in -march so this has
> gone unnoticed.
>
> Fixes: 3c349eacc559 ("Merge patch "riscv: Fix build with CONFIG_CC_OPTIMIZE_FOR_SIZE=y"")
> Reviewed-by: Conor Dooley <[email protected]>
>
> Thanks,
> Conor.
>
Thanks, Conor. I found this just by inspection. I found a issue that vdso.so call cpu_relax
cause application core dump in kernel 6.1.31. I need Samuel'patch to fix this. And I search the log
of processor.h found this issue.
>> ---
>> arch/riscv/include/asm/vdso/processor.h | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/arch/riscv/include/asm/vdso/processor.h b/arch/riscv/include/asm/vdso/processor.h
>> index 14f5d27783b8..96b65a5396df 100644
>> --- a/arch/riscv/include/asm/vdso/processor.h
>> +++ b/arch/riscv/include/asm/vdso/processor.h
>> @@ -14,7 +14,7 @@ static inline void cpu_relax(void)
>> __asm__ __volatile__ ("div %0, %0, zero" : "=r" (dummy));
>> #endif
>>
>> -#ifdef __riscv_zihintpause
>> +#ifdef CONFIG_TOOLCHAIN_HAS_ZIHINTPAUSE
>> /*
>> * Reduce instruction retirement.
>> * This assumes the PC changes.
>> --
>> 2.17.1
>>
On Wed, Aug 02, 2023 at 03:32:15PM +0800, Minda Chen wrote:
>
>
> On 2023/8/2 14:54, Conor Dooley wrote:
> > Hey Minda,
> >
> > On Wed, Aug 02, 2023 at 02:42:15PM +0800, Minda Chen wrote:
> >> Actually it is a part of Conor's
> >> commit aae538cd03bc ("riscv: fix detection of toolchain
> >> Zihintpause support").
> >> It is looks like a merge issue.
> >
> > Yup, spot on.
> >
> >> Samuel's
> >> commit 0b1d60d6dd9e ("riscv: Fix build with
> >> CONFIG_CC_OPTIMIZE_FOR_SIZE=y") do not base on Conor's commit and
> >> revert to __riscv_zihintpause. So this patch can fix it.
> >>
> >> Signed-off-by: Minda Chen <[email protected]>
> >
> > Did you actually manage to trigger this, or was this by inspection?
> > clang-15 + binutils 2.35 was, IIRC, how we spotted this because that's
> > what the clang-built-linux CI uses to test the LTS kernels from before
> > LLVM's IAS was supported for RISC-V. Seemingly all that needs to be
> > satisfied there is that zihintpause doesn't appear in -march so this has
> > gone unnoticed.
> >
> > Fixes: 3c349eacc559 ("Merge patch "riscv: Fix build with CONFIG_CC_OPTIMIZE_FOR_SIZE=y"")
> > Reviewed-by: Conor Dooley <[email protected]>
> >
> > Thanks,
> > Conor.
> >
> Thanks, Conor. I found this just by inspection. I found a issue that vdso.so call cpu_relax
> cause application core dump in kernel 6.1.31. I need Samuel'patch to fix this. And I search the log
> of processor.h found this issue.
That doesn't look like it is fixed in later stable kernels (we are at
6.1.42-rcN right now I think). It sounds we should ask Greg to backport
0b1d60d6dd9e ("riscv: Fix build with CONFIG_CC_OPTIMIZE_FOR_SIZE=y")
to 6.1. Does that make sense to you?
On 2023/8/2 15:48, Conor Dooley wrote:
> On Wed, Aug 02, 2023 at 03:32:15PM +0800, Minda Chen wrote:
>>
>>
>> On 2023/8/2 14:54, Conor Dooley wrote:
>> > Hey Minda,
>> >
>> > On Wed, Aug 02, 2023 at 02:42:15PM +0800, Minda Chen wrote:
>> >> Actually it is a part of Conor's
>> >> commit aae538cd03bc ("riscv: fix detection of toolchain
>> >> Zihintpause support").
>> >> It is looks like a merge issue.
>> >
>> > Yup, spot on.
>> >
>> >> Samuel's
>> >> commit 0b1d60d6dd9e ("riscv: Fix build with
>> >> CONFIG_CC_OPTIMIZE_FOR_SIZE=y") do not base on Conor's commit and
>> >> revert to __riscv_zihintpause. So this patch can fix it.
>> >>
>> >> Signed-off-by: Minda Chen <[email protected]>
>> >
>> > Did you actually manage to trigger this, or was this by inspection?
>> > clang-15 + binutils 2.35 was, IIRC, how we spotted this because that's
>> > what the clang-built-linux CI uses to test the LTS kernels from before
>> > LLVM's IAS was supported for RISC-V. Seemingly all that needs to be
>> > satisfied there is that zihintpause doesn't appear in -march so this has
>> > gone unnoticed.
>> >
>> > Fixes: 3c349eacc559 ("Merge patch "riscv: Fix build with CONFIG_CC_OPTIMIZE_FOR_SIZE=y"")
>> > Reviewed-by: Conor Dooley <[email protected]>
>> >
>> > Thanks,
>> > Conor.
>> >
>> Thanks, Conor. I found this just by inspection. I found a issue that vdso.so call cpu_relax
>> cause application core dump in kernel 6.1.31. I need Samuel'patch to fix this. And I search the log
>> of processor.h found this issue.
>
> That doesn't look like it is fixed in later stable kernels (we are at
> 6.1.42-rcN right now I think). It sounds we should ask Greg to backport
> 0b1d60d6dd9e ("riscv: Fix build with CONFIG_CC_OPTIMIZE_FOR_SIZE=y")
> to 6.1. Does that make sense to you?
Yes. 6.1 is lts kernel. Starfive will use this kernel for a long time. Thanks.