Re: [PATCH v3 2/4] powerpc: Fixes for CONFIG_PTE_64BIT for SMP support

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Wed, 2008-09-03 at 13:09 -0500, Kumar Gala wrote:
> There are some minor issues with support 64-bit PTEs on a 32-bit processor
> when dealing with SMP.
> 
> * We need to order the stores in set_pte_at to make sure the flag word
>   is set second.
> * Change pte_clear to use pte_update so only the flag word is cleared
> 
> Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
> ---
> 
> Fixed pte_clear to not break on 6xx.

Thanks :-)

>  static inline unsigned long long pte_update(pte_t *p,
>  					    unsigned long clr,
>  					    unsigned long set)
> @@ -658,8 +656,17 @@ static inline void set_pte_at(struct mm_struct *mm, unsigned long addr,
>  #if _PAGE_HASHPTE != 0
>  	pte_update(ptep, ~_PAGE_HASHPTE, pte_val(pte) & ~_PAGE_HASHPTE);
>  #else
> +#if defined(CONFIG_PTE_64BIT) && defined(CONFIG_SMP)

Minor nit, but there's #elif :-)

> +	__asm__ __volatile__("\
> +		stw%U0%X0 %2,%0\n\
> +		eieio\n\
> +		stw%U0%X0 %L2,%1"
> +	: "=m" (*ptep), "=m" (*((unsigned char *)ptep+4))
> +	: "r" (pte) : "memory");

Any reason why it has to be in assembly ? Can gcc re-order stores around
eieio() if written in C ? I would hope not but heh ... it's gcc :-)

I also wonder if you should first ensure that the PTE is invalid and
if not, clear it and flush the TLB page ... Or at least add a
WARN_ON(pte_valid()) in case we get that wrong ...

> +#else
>  	*ptep = pte;
>  #endif
> +#endif
>  }
>  
>  /*