Re: [PATCH v9] xen/grant-table: Avoid m2p_override during mapping

[David Vrabel <david.vrabel@citrix.com>]

On 20/02/14 18:17, Stefano Stabellini wrote:
> On Thu, 20 Feb 2014, Zoltan Kiss wrote:
>> On 20/02/14 17:26, Stefano Stabellini wrote:
>>> On Mon, 17 Feb 2014, Zoltan Kiss wrote:
>>>> On 16/02/14 18:36, Stefano Stabellini wrote:
>>>>> On Wed, 12 Feb 2014, Zoltan Kiss wrote:
>>>>>> diff --git a/arch/arm/include/asm/xen/page.h
>>>>>> b/arch/arm/include/asm/xen/page.h
>>>>>> index e0965ab..4eaeb3f 100644
>>>>>> --- a/arch/arm/include/asm/xen/page.h
>>>>>> +++ b/arch/arm/include/asm/xen/page.h
>>>>>> @@ -97,16 +97,15 @@ static inline pte_t *lookup_address(unsigned long
>>>>>> address, unsigned int *level)
>>>>>>    	return NULL;
>>>>>>    }
>>>>>>
>>>>>> -static inline int m2p_add_override(unsigned long mfn, struct page
>>>>>> *page,
>>>>>> -		struct gnttab_map_grant_ref *kmap_op)
>>>>>> -{
>>>>>> -	return 0;
>>>>>> -}
>>>>>> -
>>>>>> -static inline int m2p_remove_override(struct page *page, bool
>>>>>> clear_pte)
>>>>>> -{
>>>>>> -	return 0;
>>>>>> -}
>>>>>> +extern int set_foreign_p2m_mapping(struct gnttab_map_grant_ref
>>>>>> *map_ops,
>>>>>> +				   struct gnttab_map_grant_ref
>>>>>> *kmap_ops,
>>>>>> +				   struct page **pages, unsigned int
>>>>>> count,
>>>>>> +				   bool m2p_override);
>>>>>> +
>>>>>> +extern int clear_foreign_p2m_mapping(struct gnttab_unmap_grant_ref
>>>>>> *unmap_ops,
>>>>>> +				     struct gnttab_map_grant_ref
>>>>>> *kmap_ops,
>>>>>> +				     struct page **pages, unsigned int
>>>>>> count,
>>>>>> +				     bool m2p_override);
>>>>>
>>>>> Much much better.
>>>>> The only comment I have is about this m2p_override boolean parameter.
>>>>> m2p_override is now meaningless in this context, what we really want to
>>>>> let the arch specific implementation know is whether the mapping is a
>>>>> kernel only mapping or a userspace mapping.
>>>>> Testing for kmap_ops != NULL might even be enough, but it would not
>>>>> improve the interface.
>>>> gntdev is the only user of this, the kmap_ops parameter there is:
>>>> use_ptemod ? map->kmap_ops + offset : NULL
>>>> where:
>>>> use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
>>>> So I think we can't rely on kmap_ops to decide whether we should
>>>> m2p_override
>>>> or not.
>>>>
>>>>> Is it possible to realize if the mapping is a userspace mapping by
>>>>> checking for GNTMAP_application_map in map_ops?
>>>>> Otherwise I would keep the boolean and rename it to user_mapping.
>>>> Sounds better, but as far as I see gntdev set that flag in
>>>> find_grant_ptes,
>>>> which is called only
>>>>
>>>> if (use_ptemod) {
>>>> 	err = apply_to_page_range(vma->vm_mm, vma->vm_start,
>>>> 				  vma->vm_end - vma->vm_start,
>>>> 				  find_grant_ptes, map);
>>>>
>>>> So if xen_feature(XENFEAT_auto_translated_physmap), we don't have
>>>> kmap_ops,
>>>> and GNTMAP_application_map is not set as well, but I guess we still need
>>>> m2p_override. Or not? I'm a bit confused, maybe because of Monday ...
>>>
>>> If xen_feature(XENFEAT_auto_translated_physmap) we shouldn't need the
>>> m2p_override.
>>>
>>
>> So it's safe to assume that we need m2p_override only if kmap_ops != NULL, and
>> we can avoid the extra bool parameter, is that correct? At least with the
>> current users of grant mapping it seems to be true.
>> In which case we don't need the wrappers for gnttab_[un]map_refs as well.
>> Actually the most of m2p_add/remove_override takes effect only if there is a
>> kmap_op parameter, but what about the rest of the code there?
> 
> It is safe to assume that we only need the m2p_override if
> !xen_feature(XENFEAT_auto_translated_physmap).
> I wouldn't make any assumptions on kmap_ops != NULL.

I think it is -- we only need the m2p override if we have userspace
mappings (where kmap_ops != 0).
> 
> I would remove the bool m2p_override parameter completely and determine
> whether we need to call the m2p_override functions from the x86
> implementation of set/clear_foreign_p2m_mapping by checking
> xen_feature(XENFEAT_auto_translated_physmap).
> 
> David, does it seem reasonable to you?

That would miss the point of this patch which is to avoid adding to the
m2p_override for kernel only mappings.

David