Re: [PATCH] riscv: Add RISC-V svpbmt extension

From: Guo Ren
Date: Thu Sep 23 2021 - 08:04:46 EST

On Thu, Sep 23, 2021 at 5:48 PM Nick Kossifidis <mick@xxxxxxxxxxxx> wrote:
>
> Hello Guo,
>
> Στις 2021-09-23 10:27, guoren@xxxxxxxxxx έγραψε:
> diff --git a/Documentation/devicetree/bindings/riscv/cpus.yaml
> b/Documentation/devicetree/bindings/riscv/cpus.yaml
> index e534f6a7cfa1..1825cd8db0de 100644
> --- a/Documentation/devicetree/bindings/riscv/cpus.yaml
> +++ b/Documentation/devicetree/bindings/riscv/cpus.yaml
> @@ -56,7 +56,9 @@ properties:
> enum:
> - riscv,sv32
> - riscv,sv39
> + - riscv,sv39,svpbmt
> - riscv,sv48
> + - riscv,sv48,svpbmt
> - riscv,none
>
> Isn't svpbmt orthogonal to the mmu type ? It's a functionality that can
> be present on either sv39/48/57 so why not have another "svpbmt"
> property directly on the cpu node ?
Agree, But that's another patch to redesign mmu-type in dts.

>
> > + * rv64 PTE format:
> > + * | 63 | 62 61 | 60 54 | 53 10 | 9 8 | 7 | 6 | 5 | 4 |
> > 3 | 2 | 1 | 0
> > + * N MT RSV PFN reserved for SW D A G U
> > X W R V
> > + * [62:61] Memory Type definitions:
> > + * 00 - PMA Normal Cacheable, No change to implied PMA memory type
> > + * 01 - NC Non-cacheable, idempotent, weakly-ordered Main Memory
> > + * 10 - IO Non-cacheable, non-idempotent, strongly-ordered I/O
> > memory
> > + * 11 - Rsvd Reserved for future standard use
> > + */
> > +#define _PAGE_MT_MASK ((u64)0x3 << 61)
> > +#define _PAGE_MT_PMA ((u64)0x0 << 61)
> > +#define _PAGE_MT_NC ((u64)0x1 << 61)
> > +#define _PAGE_MT_IO ((u64)0x2 << 61)
> > +
>
> It'd be cleaner IMHO if you defined _PAGE_MT_MASK as (_PAGE_MT_PMA |
> _PAGE_MT_NC | _PAGE_MT_IO), like other masks are defined (e.g.
> _PAGE_CHG_MASK on the same file). I also suggest you use unsigned long
> instead of u64 for consistency.
Okay

>
> > +enum {
> > + MT_PMA,
> > + MT_NC,
> > + MT_IO,
> > + MT_MAX
> > +};
> > +
> > +extern struct __riscv_svpbmt_struct {
> > + unsigned long mask;
> > + unsigned long mt[MT_MAX];
> > +} __riscv_svpbmt;
> > +
> > +#define _PAGE_DMA_MASK __riscv_svpbmt.mask
> > +#define _PAGE_DMA_PMA __riscv_svpbmt.mt[MT_PMA]
> > +#define _PAGE_DMA_NC __riscv_svpbmt.mt[MT_NC]
> > +#define _PAGE_DMA_IO __riscv_svpbmt.mt[MT_IO]
> > +#else
> > +#define _PAGE_DMA_MASK 0
> > +#define _PAGE_DMA_PMA 0
> > +#define _PAGE_DMA_NC 0
> > +#define _PAGE_DMA_IO 0
> > +#endif /* CONFIG_64BIT */
> > +#endif /* __ASSEMBLY__ */
> > +
> > #define _PAGE_SPECIAL _PAGE_SOFT
> > #define _PAGE_TABLE _PAGE_PRESENT
> >
>
> This struct is not useful as part of enabling the standard Svpbmt
> extension on Linux, we can set _PAGE_DMA_* macros directly on this patch
> and introduce the struct approach later on, when we also define
> alternative values for _PAGE_DMA_* flags. Also to someone reading the
> code the struct doesn't make sense without some documentation on why
> it's needed. Finally why the enum / array ? Why not just have different
> fields on the struct ?
Okay, I'll use different fields on the struct.

>
> > diff --git a/arch/riscv/include/asm/pgtable.h
> > b/arch/riscv/include/asm/pgtable.h
> > index 39b550310ec6..d07ba586c866 100644
> > --- a/arch/riscv/include/asm/pgtable.h
> > +++ b/arch/riscv/include/asm/pgtable.h
> > @@ -136,7 +136,8 @@
> > | _PAGE_PRESENT \
> > | _PAGE_ACCESSED \
> > | _PAGE_DIRTY \
> > - | _PAGE_GLOBAL)
> > + | _PAGE_GLOBAL \
> > + | _PAGE_DMA_PMA)
> >
>
> That's a bit misleading, it's like marking the kernel pages as DMAable.
>
> -/*
> - * The RISC-V ISA doesn't yet specify how to query or modify PMAs, so
> we can't
> - * change the properties of memory regions.
> - */
> -#define _PAGE_IOREMAP _PAGE_KERNEL
> +#define _PAGE_IOREMAP ((_PAGE_KERNEL & ~_PAGE_DMA_MASK) |
> _PAGE_DMA_IO)
> +
> +#define PAGE_IOREMAP __pgprot(_PAGE_IOREMAP)
>
> This isn't used anywhere.

See:

arch/riscv/include/asm/fixmap.h
@@ -43,7 +43,7 @@ enum fixed_addresses {
__end_of_fixed_addresses
};

-#define FIXMAP_PAGE_IO PAGE_KERNEL
+#define FIXMAP_PAGE_IO PAGE_IOREMAP

>
> @@ -490,6 +489,28 @@ static inline int ptep_clear_flush_young(struct
> vm_area_struct *vma,
> return ptep_test_and_clear_young(vma, address, ptep);
> }
>
> +#define pgprot_noncached pgprot_noncached
> +static inline pgprot_t pgprot_noncached(pgprot_t _prot)
> +{
> + unsigned long prot = pgprot_val(_prot);
> +
> + prot &= ~_PAGE_DMA_MASK;
> + prot |= _PAGE_DMA_IO;
> +
> + return __pgprot(prot);
> +}
> +
> +#define pgprot_writecombine pgprot_writecombine
> +static inline pgprot_t pgprot_writecombine(pgprot_t _prot)
> +{
> + unsigned long prot = pgprot_val(_prot);
> +
> + prot &= ~_PAGE_DMA_MASK;
> + prot |= _PAGE_DMA_NC;
> +
> + return __pgprot(prot);
> +}
> +
>
> We also have the IO type, we should also define pgprot_device to also
> ensure ordering, or else it'll fallback to pgprot_noncached, which in
> our case won't work well due to RVWMO:
> https://elixir.bootlin.com/linux/latest/source/include/linux/pgtable.h#L930
Current is:
pgprot_noncached = pgprot_device = IO type
writecombine = NC type

I think it's proper.

>
> +void __init riscv_svpbmt(void)
> +{
> +#ifdef CONFIG_64BIT
> + 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 + 11, "svpbmt", 6)) {
> + __riscv_svpbmt.mask = _PAGE_MT_MASK;
> + __riscv_svpbmt.mt[MT_PMA] = _PAGE_MT_PMA;
> + __riscv_svpbmt.mt[MT_NC] = _PAGE_MT_NC;
> + __riscv_svpbmt.mt[MT_IO] = _PAGE_MT_IO;
> + break;
> + }
> + }
> +#endif
> +}
>
> You break; here the first time you find a cpu node with svpbmt enabled,
> shouldn't we make sure that all used cpu nodes support svpbmt before
> using the extension ?
Okay

>
> Regards,
> Nick



--
Best Regards
Guo Ren

ML: https://lore.kernel.org/linux-csky/