Re: [PATCH 3/5 gnumach] smp: Delay setting spl_init flag until after smp bringup

[Samuel Thibault]

Damien Zammit via Bug reports for the GNU Hurd, le mar. 10 déc. 2024 07:29:56 
+0000, a ecrit:
> If this is not done, the spl code is called by kmsg_putchar
> on APs during printing of bringup debug, causing the only
> unnecessary early access of GS.
> ---
>  i386/i386/mp_desc.c     | 3 ++-
>  i386/i386at/model_dep.c | 3 ++-
>  2 files changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/i386/i386/mp_desc.c b/i386/i386/mp_desc.c
> index 16b1c865..7595eb1e 100644
> --- a/i386/i386/mp_desc.c
> +++ b/i386/i386/mp_desc.c
> @@ -325,7 +325,7 @@ start_other_cpus(void)
>  
>       unsigned cpu;
>  
> -     splhigh();
> +     asm volatile ("cli");
>  
>       /* Disable IOAPIC interrupts (IPIs not affected).
>        * Clearing this flag is similar to masking all
> @@ -357,5 +357,6 @@ start_other_cpus(void)
>  
>       /* Re-enable IOAPIC interrupts as per setup */
>       lapic_enable();
> +     spl_init = 1;
>  }
>  #endif       /* NCPUS > 1 */
> diff --git a/i386/i386at/model_dep.c b/i386/i386at/model_dep.c
> index 30449c37..7cbbf70b 100644
> --- a/i386/i386at/model_dep.c
> +++ b/i386/i386at/model_dep.c
> @@ -380,8 +380,9 @@ i386at_init(void)
>  #else        /* MACH_HYP */
>       hyp_intrinit();
>  #endif       /* MACH_HYP */
> +#if NCPUS == 1
>       spl_init = 1;
> -
> +#endif

But then kmsg_putchar becomes unsafe until very late, that will bite
us... And you never know who would introduce gs usage in whatever place
in the boot code.

I know that you don't like tinkering with assembler, but it's simpler to
fix one place than to have to maintain principles in a lot of places.

Samuel