[PATCH] riscv: Add RISC-V svpbmt extension

[Nick Kossifidis]

Hello Guo,

Στις 2021-09-23 10:27, guoren at kernel.org 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 ?

> + * 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.

> +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 ?

> 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.

@@ -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

+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 ?

Regards,
Nick