Re: [PATCH v3 3/6] RISC-V: add f & d extension validation checks

[Conor Dooley <conor@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 2602 bytes --]

On Tue, Feb 11, 2025 at 11:22:13AM +0100, Clément Léger wrote:
> 
> 
> On 05/02/2025 17:05, Conor Dooley wrote:
> > From: Conor Dooley <conor.dooley@microchip.com>
> > 
> > Using Clement's new validation callbacks, support checking that
> > dependencies have been satisfied for the floating point extensions.
> > 
> > The check for "d" might be slightly confusingly shorter than that of "f",
> > despite "d" depending on "f". This is because the requirement that a
> > hart supporting double precision must also support single precision,
> > should be validated by dt-bindings etc, not the kernel but lack of
> > support for single precision only is a limitation of the kernel.
> > 
> > Since vector will now be disabled proactively, there's no need to clear
> > the bit in elf_hwcap in riscv_fill_hwcap() any longer.
> > 
> > Signed-off-by: Conor Dooley <conor.dooley@microchip.com>
> > ---
> >  arch/riscv/kernel/cpufeature.c | 27 +++++++++++++++++++++++++--
> >  1 file changed, 25 insertions(+), 2 deletions(-)
> > 
> > diff --git a/arch/riscv/kernel/cpufeature.c b/arch/riscv/kernel/cpufeature.c
> > index 1c148ecea612..ad4fbaa4ff0d 100644
> > --- a/arch/riscv/kernel/cpufeature.c
> > +++ b/arch/riscv/kernel/cpufeature.c
> > @@ -109,6 +109,29 @@ static int riscv_ext_zicboz_validate(const struct riscv_isa_ext_data *data,
> >  	return 0;
> >  }
> >  
> > +static int riscv_ext_f_validate(const struct riscv_isa_ext_data *data,
> > +				const unsigned long *isa_bitmap)
> > +{
> > +	if (!__riscv_isa_extension_available(isa_bitmap, RISCV_ISA_EXT_d)) {
> > +		pr_warn_once("This kernel does not support systems with F but not D\n");
> > +		return -EINVAL;
> > +	}
> 
> While I tested to remove the RISCV_ISA_EXT_d from the input isa bitmap
> and it worked, I didn't realized that it was due to the probe order of
> single letter extensions. D is probed before F so that works as
> expected. But returning -EPROBEDEFER would not allow to display the
> warn_once or wrongly display it if D was not yet probed. So I'm inclined
> to keep it as is and rely on probe order (a bit fragile but for single
> letter extensions, that seems acceptable).

I guess it's worth adding a comment to that effect.

> > +
> > +	if (!IS_ENABLED(CONFIG_FPU))
> > +		return -EINVAL;
> 
> I would have actually move that chunk before the
> __riscv_isa_extension_available() check so that the whole function body
> is elided if FPU is disabled.

I think you're right, but for another reason too - warning when someone
has turned off CONIFG_FPU doesn't make sense.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]