Re: [PATCH V4 2/2] riscv: add RISC-V Svpbmt extension supports

[Jisheng Zhang <jszhang@kernel.org>]

On Wed, 1 Dec 2021 11:48:44 +0530
Anup Patel <anup@brainfault.org> wrote:


> > > >   */
> > > > diff --git a/arch/riscv/kernel/cpufeature.c b/arch/riscv/kernel/cpufeature.c
> > > > index d959d207a40d..fa7480cb8b87 100644
> > > > --- a/arch/riscv/kernel/cpufeature.c
> > > > +++ b/arch/riscv/kernel/cpufeature.c
> > > > @@ -8,6 +8,7 @@
> > > >
> > > >  #include <linux/bitmap.h>
> > > >  #include <linux/of.h>
> > > > +#include <linux/pgtable.h>
> > > >  #include <asm/processor.h>
> > > >  #include <asm/hwcap.h>
> > > >  #include <asm/smp.h>
> > > > @@ -59,6 +60,38 @@ bool __riscv_isa_extension_available(const unsigned long *isa_bitmap, int bit)
> > > >  }
> > > >  EXPORT_SYMBOL_GPL(__riscv_isa_extension_available);
> > > >
> > > > +static void __init mmu_supports_svpbmt(void)
> > > > +{
> > > > +#if defined(CONFIG_MMU) && defined(CONFIG_64BIT)  
> > >
> > > IIRC, Christoph suggested a CONFIG_RISCV_SVPBMT when reviewing v3. What
> > > about that idea?  
> >
> > Yes, sorry for missing it, yes, I think we can have something like this
> >
> > config ARCH_HAS_RISCV_SVPBMT
> > bool
> > default n
> >
> > any platform which needs this support, can just
> >
> > select ARCH_HAS_RISCV_SVPBMT
> >
> > and which is the best name? ARCH_HAS_RISCV_SVPBMT or just ARCH_HAS_SVPBMT ?
> >  
> > >  
> > > > +     struct device_node *node;
> > > > +     const char *str;
> > > > +
> > > > +     for_each_of_cpu_node(node) {
> > > > +             if (of_property_read_string(node, "mmu-type", &str))
> > > > +                     continue;
> > > > +
> > > > +             if (!strncmp(str + 6, "none", 4))
> > > > +                     continue;
> > > > +
> > > > +             if (of_property_read_string(node, "mmu", &str))
> > > > +                     continue;
> > > > +
> > > > +             if (strncmp(str + 6, "svpmbt", 6))
> > > > +                     continue;
> > > > +     }
> > > > +
> > > > +     __svpbmt.pma            = _SVPBMT_PMA;
> > > > +     __svpbmt.nocache        = _SVPBMT_NC;
> > > > +     __svpbmt.io             = _SVPBMT_IO;
> > > > +     __svpbmt.mask           = _SVPBMT_MASK;
> > > > +#endif
> > > > +}
> > > > +
> > > > +static void __init mmu_supports(void)  
> > >
> > > can we remove this function currently? Instead, directly call
> > > mmu_supports_svpbmt()?
> > >  
> > > > +{
> > > > +     mmu_supports_svpbmt();
> > > > +}
> > > > +
> > > >  void __init riscv_fill_hwcap(void)
> > > >  {
> > > >       struct device_node *node;
> > > > @@ -67,6 +100,8 @@ void __init riscv_fill_hwcap(void)
> > > >       size_t i, j, isa_len;
> > > >       static unsigned long isa2hwcap[256] = {0};
> > > >
> > > > +     mmu_supports();
> > > > +
> > > >       isa2hwcap['i'] = isa2hwcap['I'] = COMPAT_HWCAP_ISA_I;
> > > >       isa2hwcap['m'] = isa2hwcap['M'] = COMPAT_HWCAP_ISA_M;
> > > >       isa2hwcap['a'] = isa2hwcap['A'] = COMPAT_HWCAP_ISA_A;
> > > > diff --git a/arch/riscv/mm/init.c b/arch/riscv/mm/init.c
> > > > index 24b2b8044602..e4e658165ee1 100644
> > > > --- a/arch/riscv/mm/init.c
> > > > +++ b/arch/riscv/mm/init.c
> > > > @@ -854,3 +854,8 @@ int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
> > > >       return vmemmap_populate_basepages(start, end, node, NULL);
> > > >  }
> > > >  #endif
> > > > +
> > > > +#if defined(CONFIG_64BIT)
> > > > +struct __svpbmt_struct __svpbmt __ro_after_init;  
> > >
> > > Added the structure for all RV64 including NOMMU case and those platforms
> > > which doen't want SVPBMT at all, I believe Christoph's CONFIG_RISCV_SVPBMT
> > > suggestion can solve this problem.  
> >
> > see ARCH_HAS_RISCV_SVPBMT above . :-)  
> 
> This config option will not align with the goal of having a unified
> kernel image which works on HW with/without Svpmbt.

Just my thoughts: 

If disable this option, HW without Svpbmt can work as before; Hw with
Svpbmt will only have a basic working, those DMA etc can't work.

If enable this option, HW without Svpbmt can work as well, but with
a bit overhead and waste. HW with Svpbmt can work. So this option gives
those platforms which doesn't need Svpbmt a chance to totally disable it.

But linux distributions which want a uniified Image usually enable features as
much as possible, so IMHO, this config option can still meet unified kernel
image requirement.

> 
> Better to explore code patching approaches which have zero
> overhead.

It would be nice if the Svpbmt can be supported via. coding patching tech.

Thanks