Re: [kvm-unit-tests PATCH] riscv: Make NR_CPUS configurable

[Andrew Jones <andrew.jones@linux.dev>]

On Tue, Aug 20, 2024 at 07:01:51PM GMT, Andrew Jones wrote:
> Unit tests would like to go nuts with the number of harts in order
> to help shake out issues with hart number assumptions. Rather than
> set a huge number that will only be used when a platform supports
> a huge number or when QEMU is told to exceed the recommended
> number of vcpus, make the number configurable. However, we do bump
> the default from 16 to 2*xlen since we would like to always force
> kvm-unit-tests to use cpumasks with more than one word in order to
> ensure that code stays maintained.
> 
> Signed-off-by: Andrew Jones <andrew.jones@linux.dev>
> ---
>  configure             | 8 ++++++++
>  lib/riscv/asm/setup.h | 3 ++-
>  2 files changed, 10 insertions(+), 1 deletion(-)
> 
> diff --git a/configure b/configure
> index 27ae9cc89657..6a8d6239a3cb 100755
> --- a/configure
> +++ b/configure
> @@ -77,6 +77,8 @@ usage() {
>  	                           Specify the page size (translation granule). PAGE_SIZE can be
>  	                           4k [default], 16k, 64k for arm64.
>  	                           4k [default], 64k for ppc64.
> +	    --max-cpus=MAX_CPUS
> +	                           Specify the maximum number of CPUs supported. (riscv64 only)
>  	    --earlycon=EARLYCON
>  	                           Specify the UART name, type and address (optional, arm and
>  	                           arm64 only). The specified address will overwrite the UART
> @@ -168,6 +170,9 @@ while [[ "$1" = -* ]]; do
>  	--page-size)
>  	    page_size="$arg"
>  	    ;;
> +	--max-cpus)
> +	    max_cpus="$arg"
> +	    ;;
>  	--earlycon)
>  	    earlycon="$arg"
>  	    ;;
> @@ -496,8 +501,11 @@ cat <<EOF >> lib/config.h
>  
>  EOF
>  elif [ "$arch" = "riscv32" ] || [ "$arch" = "riscv64" ]; then
> +	[ -z $max_cpus ] && max_cpus='(__riscv_xlen * 2)'
> +
>  cat <<EOF >> lib/config.h
>  
> +#define CONFIG_NR_CPUS $max_cpus
>  #define CONFIG_UART_EARLY_BASE 0x10000000
>  
>  EOF
> diff --git a/lib/riscv/asm/setup.h b/lib/riscv/asm/setup.h
> index a13159bfe395..43b63c56d96f 100644
> --- a/lib/riscv/asm/setup.h
> +++ b/lib/riscv/asm/setup.h
> @@ -2,9 +2,10 @@
>  #ifndef _ASMRISCV_SETUP_H_
>  #define _ASMRISCV_SETUP_H_
>  #include <libcflat.h>
> +#include <config.h>
>  #include <asm/processor.h>
>  
> -#define NR_CPUS 16
> +#define NR_CPUS CONFIG_NR_CPUS
>  extern struct thread_info cpus[NR_CPUS];
>  extern int nr_cpus;
>  extern uint64_t timebase_frequency;
> -- 
> 2.45.2
>

I started having second thoughts about this approach since we'll have to
keep adding configure command line options for every CONFIG option. I
posted an alternative approach now [1] and CC'ed everyone since other
architectures should chime in on the approach too.

[1] https://lore.kernel.org/all/20240903143946.834864-4-andrew.jones@linux.dev/

Thanks,
drew