Paul Mackerras <pau...@ozlabs.org> writes:
> [ text/plain ]
> On Fri, Feb 26, 2016 at 08:50:50AM +0530, Aneesh Kumar K.V wrote:
>> _PAGE_PRIV means the page can be accessed only by kernel. This is done
>> to keep pte bits similar to PowerISA 3.0 radix PTE format. User
>> pages are now makred by clearing _PAGE_PRIV bit.
>
>
.....
>> @@ -40,6 +40,11 @@ int __hash_page_4K(unsigned long ea, unsigned long
>> access, unsigned long vsid,
>> if (unlikely(access & ~old_pte))
>> return 1;
>
> This check is going to do a different thing now as far as
> _PAGE_USER/_PAGE_PRIV is concerned: previously it would prevent a
> non-privileged access to a privileged page from creating a HPTE, now
> it prevents a privileged access to a non-privileged page from creating
> a HPTE. A privileged access means an access by the kernel to a high
> address, and arguably we would never have a non-privileged PTE at a
> high (i.e. kernel) address, but it's still a semantic change that
> should have been flagged in the patch description.
We don't set _PAGE_PRIVILGED when we have a privilged acess to a non
privilged page. We set it as below (with updated comments)
/*
* We set _PAGE_PRIVILEGED only when
* kernel mode access kernel space.
*
* _PAGE_PRIVILGED is NOT set
* 1) when kernel mode access user space
* 2) user space access kernel space.
*/
if (!(msr & MSR_PR) && !(REGION_ID(ea) == USER_REGION_ID))
access |= _PAGE_PRIVILEGED;
>
> In fact it wouldn't really matter if we didn't check the privilege
> here and went ahead and created the HPTE - if it's a non-privileged
> access to a privileged page, the HPTE will get its permission bits set
> so as to prevent non-privileged access anyway.
>
>> /*
>> + * access from user, but pte in _PAGE_PRIV
>> + */
>> + if (unlikely((access & _PAGE_PRIV) != (old_pte & _PAGE_PRIV)))
>> + return 1;
>
> And this catches both the privileged access to non-privileged page
> case (which the previous statement already caught) and the
> non-privileged access to privileged page case.
The actual PRIVILEGED check is done by the above conditional.
-aneesh
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev
