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

[Jan Beulich <jbeulich@xxxxxxxx>]

On 22.09.2021 11:26, Roger Pau Monné wrote:
> On Tue, Sep 21, 2021 at 12:12:05PM +0200, Jan Beulich wrote:
>> 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.
> 
> OK, sorry, it wasn't clear to me from the description. Could you
> explicitly mention in the description that the removal is moved into
> gnttab_set_frame_gfn on Arm in order to cope with the fact that
> xenmem_add_to_physmap_one doesn't perform it.

Well, it's not really "in order to cope" - that's true for the placement
prior to this change as well, so not a justification for the change.
Nevertheless I've tried to make this more clear by changing the 1st
paragraph to:

"Without holding appropriate locks, attempting to remove a prior mapping
 of the underlying page is pointless, as the same (or another) mapping
 could be re-established by a parallel request on another vCPU. Move the
 code to Arm's gnttab_set_frame_gfn(); it cannot be dropped there since
 xenmem_add_to_physmap_one() doesn't call it either (unlike on x86). Of
 course this new placement doesn't improve things in any way as far as
 the security of grant status frame mappings goes (see XSA-379). Proper
 locking would be needed here to allow status frames to be mapped
 securely."

> TBH I think it would be in our best interest to try to make
> xenmem_add_to_physmap_one behave as close as possible between arches.
> This discrepancy between x86 and Arm regarding page removal is just
> going to bring more trouble in the long term, and hiding the
> differences inside gnttab_set_frame_gfn just makes it even more
> obscure.

Stefano, Julien?

Jan