Re: [PATCH 1/2] gnttab: remove guest_physmap_remove_page() call from gnttab_map_frame()

[Jan Beulich <jbeulich@xxxxxxxx>]

On 21.09.2021 10:32, Roger Pau Monné wrote:
> On Mon, Sep 20, 2021 at 05:27:17PM +0200, Jan Beulich wrote:
>> On 20.09.2021 12:20, Roger Pau Monné wrote:
>>> On Mon, Sep 13, 2021 at 08:41:47AM +0200, Jan Beulich wrote:
>>>> --- a/xen/include/asm-arm/grant_table.h
>>>> +++ b/xen/include/asm-arm/grant_table.h
>>>> +        if ( gfn_eq(ogfn, INVALID_GFN) || gfn_eq(ogfn, gfn) ||           \
>>>
>>> I'm slightly confused by this checks, don't you need to check for
>>> gfn_eq(gfn, INVALID_GFN) (not ogfn) in order to call
>>> guest_physmap_remove_page?
>>
>> Why? It's ogfn which gets passed to the function. And it indeed is the
>> prior GFN's mapping that we want to remove here.
>>
>>> Or assuming that ogfn is not invalid can be used to imply a removal?
>>
>> That implication can be (and on x86 is) used for the incoming argument,
>> i.e. "gfn". I don't think "ogfn" can serve this purpose.
> 
> I guess I'm confused due to the ogfn checks done on the Arm side that
> are not performed on x86. So on Arm you always need to explicitly
> unhook the previous GFN before attempting to setup a new mapping,
> while on x86 you only need to do this when it's a removal in order to
> clear the entry?

The difference isn't with guest_physmap_add_entry() (both x86 and
Arm only insert a new mapping there), but with
xenmem_add_to_physmap_one(): Arm's variant doesn't care about prior
mappings. And gnttab_map_frame() gets called only from there. This
is effectively what the first paragraph of the description is about.

> So you are effectively only removing the call to
> guest_physmap_remove_page in gnttab_map_frame for x86, because Arm
> will still perform it in gnttab_set_frame_gfn.

Yes.

> This seems like a limitation of Arm's guest_physmap_add_entry.

As per above I'm viewing this as a limitation of Arm's
xenmem_add_to_physmap_one().

Jan